release: OpenCV 4.7.0

Merge pull request #22986 from AleksandrPanov:move_contrib_charuco_to_main_objdetect
merge with https://github.com/opencv/opencv_contrib/pull/3394 move Charuco API from contrib to main repo: - add CharucoDetector: ``` CharucoDetector::detectBoard(InputArray image, InputOutputArrayOfArrays markerCorners, InputOutputArray markerIds, OutputArray charucoCorners, OutputArray charucoIds) const // detect charucoCorners and/or markerCorners CharucoDetector::detectDiamonds(InputArray image, InputOutputArrayOfArrays _markerCorners, InputOutputArrayOfArrays _markerIds, OutputArrayOfArrays _diamondCorners, OutputArray _diamondIds) const ``` - add `matchImagePoints()` for `CharucoBoard` - remove contrib aruco dependencies from interactive-calibration tool - move almost all aruco tests to objdetect ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2026-07-23 12:23:02 +04:00 · 2022-12-28 17:31:52 +03:00 · 2022-12-28 17:28:59 +03:00 · 2022-12-28 10:04:25 +00:00 · 2022-12-28 09:54:16 +00:00 · 2022-12-28 17:16:11 +08:00
2117 changed files with 362341 additions and 172036 deletions
@@ -34,11 +34,11 @@ This is a template helping you to create an issue which can be processed as quic
 - [ ] I report the issue, it's not a question
   <!--
   OpenCV team works with forum.opencv.org, Stack Overflow and other communities
-   to discuss problems. Tickets with question without real issue statement will be
+   to discuss problems. Tickets with questions without a real issue statement will be
   closed.
   -->
 - [ ] I checked the problem with documentation, FAQ, open issues,
-       forum.opencv.org, Stack Overflow, etc and have not found solution
+       forum.opencv.org, Stack Overflow, etc and have not found any solution
   <!--
   Places to check:
   * OpenCV documentation: https://docs.opencv.org
@@ -47,11 +47,11 @@ This is a template helping you to create an issue which can be processed as quic
   * OpenCV issue tracker: https://github.com/opencv/opencv/issues?q=is%3Aissue
   * Stack Overflow branch: https://stackoverflow.com/questions/tagged/opencv
   -->
- - [ ] I updated to latest OpenCV version and the issue is still there
+ - [ ] I updated to the latest OpenCV version and the issue is still there
   <!--
   master branch for OpenCV 4.x and 3.4 branch for OpenCV 3.x releases.
-   OpenCV team supports only latest release for each branch.
-   The ticket is closed, if the problem is not reproduced with modern version.
+   OpenCV team supports only the latest release for each branch.
+   The ticket is closed if the problem is not reproduced with the modern version.
   -->
 - [ ] There is reproducer code and related data files: videos, images, onnx, etc
   <!--
@@ -61,9 +61,9 @@ This is a template helping you to create an issue which can be processed as quic
     to reduce attachment size
   * Use PNG for images, if you report some CV related bug, but not image reader
     issue
-   * Attach the image as archive to the ticket, if you report some reader issue.
+   * Attach the image as an archive to the ticket, if you report some reader issue.
     Image hosting services compress images and it breaks the repro code.
-   * Provide ONNX file for some public model or ONNX file with with random weights,
+   * Provide ONNX file for some public model or ONNX file with random weights,
     if you report ONNX parsing or handling issue. Architecture details diagram
     from netron tool can be very useful too. See https://lutzroeder.github.io/netron/
   -->
@@ -0,0 +1,64 @@
+name: Bug Report
+description: Create a report to help us reproduce and fix the bug
+labels: ["bug"]
+
+body:
+- type: markdown
+  attributes:
+    value: >
+      #### Thank you for contributing! Before reporting a bug, please have a look at the [FAQ](https://github.com/opencv/opencv/wiki/FAQ), make sure the issue has no duplicate and hasn't been already addressed by searching through [the existing and past issues](https://github.com/opencv/opencv/issues?page=1&q=is%3Aissue+sort%3Acreated-desc).
+
+- type: textarea
+  attributes:
+    label: System Information
+    description: |
+      Please provide the following system information to help us diagnose the bug. For example:
+
+      // example for c++ user
+      OpenCV version: 4.6.0
+      Operating System / Platform: Ubuntu 20.04
+      Compiler & compiler version: GCC 9.3.0
+
+      // example for python user
+      OpenCV python version: 4.6.0.66
+      Operating System / Platform: Ubuntu 20.04
+      Python version: 3.9.6
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: Detailed description
+    description: |
+      Please provide a clear and concise description of what the bug is and paste the error log below. It helps improving readability if the error log is wrapped in ```` ```triple quotes blocks``` ````.
+    placeholder: |
+      A clear and concise description of what the bug is.
+
+      ```
+      # error log
+      ```
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: Steps to reproduce
+    description: |
+      Please provide a minimal example to help us reproduce the bug. Code should be wrapped with ```` ```triple quotes blocks``` ```` to improve readability. If the code is too long, please attach as a file or create and link a public gist: https://gist.github.com.
+
+      Related data files (images, onnx, etc) should be attached below as well. If the data files are too big, feel free to upload them to a online drive, share them and put the link below.
+    placeholder: |
+      ```cpp (replace cpp with python if python code)
+      # sample code to reproduce the bug
+      ```
+
+      Test data: [image](https://link/to/the/image), [model.onnx](htts://link/to/the/onnx/model)
+  validations:
+    required: true
+- type: checkboxes
+  attributes:
+    label: Issue submission checklist
+    options:
+      - label: I report the issue, it's not a question
+        required: true
+      - label: I checked the problem with documentation, FAQ, open issues, forum.opencv.org, Stack Overflow, etc and have not found any solution
+      - label: I updated to the latest OpenCV version and the issue is still there
+      - label: There is reproducer code and related data files (videos, images, onnx, etc)
@@ -0,0 +1,5 @@
+blank_issues_enabled: true
+contact_links:
+  - name: Questions
+    url: https://forum.opencv.org/
+    about: Ask questions and discuss with OpenCV community members
@@ -0,0 +1,26 @@
+name: Documentation
+description: Report an issue related to https://docs.opencv.org/
+labels: ["category: documentation"]
+
+body:
+- type: markdown
+  attributes:
+    value: >
+      #### Thank you for contributing! Before submitting a doc issue, please make sure it has no duplicate by searching through [the existing and past issues](https://github.com/opencv/opencv/issues?page=1&q=is%3Aissue+sort%3Acreated-desc)
+
+- type: textarea
+  attributes:
+    label: Descripe the doc issue
+    description: >
+      Please provide a clear and concise description of what content in https://docs.opencv.org/ is an issue. Note that there are multiple active branches, such as 3.4, 4.x and 5.x, so please specify the branch with the problem.
+    placeholder: |
+      A clear and concise description of what content in https://docs.opencv.org/ is an issue.
+
+      Link to the doc: https://docs.opencv.org/4.x/d3/d63/classcv_1_1Mat.html
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: Fix suggestion
+    description: >
+      Tell us how we could improve the documentation in this regard.
@@ -0,0 +1,22 @@
+name: Feature request
+description: Submit a request for a new OpenCV feature
+labels: ["feature"]
+
+body:
+- type: markdown
+  attributes:
+    value: >
+      #### Thank you for contributing! Before submitting a feature request, please make sure the request has no duplicate by searching through [the existing and past issues](https://github.com/opencv/opencv/issues?page=1&q=is%3Aissue+sort%3Acreated-desc)
+
+- type: textarea
+  attributes:
+    label: Descripe the feature and motivation
+    description: |
+      Please provide a clear and concise proposal of the feature and outline the motivation.
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: Additional context
+    description: |
+      Add any other context, such as pseudo code, links, diagram, screenshots, to help the community better understand the feature request.
@@ -3,9 +3,9 @@
 See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

 - [ ] I agree to contribute to the project under Apache 2 License.
- [ ] To the best of my knowledge, the proposed patch is not based on a code under GPL or other license that is incompatible with OpenCV
- [ ] The PR is proposed to proper branch
- [ ] There is reference to original bug report and related work
+- [ ] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
+- [ ] The PR is proposed to the proper branch
+- [ ] There is a reference to the original bug report and related work
 - [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
      Patch to opencv_extra has the same branch name.
 - [ ] The feature is well documented and sample code can be built with the project CMake
@@ -0,0 +1,41 @@
+name: PR:4.x
+
+on:
+  pull_request:
+    branches:
+      - 4.x
+
+jobs:
+  Ubuntu2004-ARM64:
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-ARM64.yaml@main
+
+  Ubuntu2004-ARM64-Debug:
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-ARM64-Debug.yaml@main
+
+  Ubuntu2004-x64:
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-U20.yaml@main
+
+  Ubuntu2004-x64-CUDA:
+    if: "${{ contains(github.event.pull_request.labels.*.name, 'category: dnn') }} || ${{ contains(github.event.pull_request.labels.*.name, 'category: dnn (onnx)') }}"
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-U20-Cuda.yaml@main
+
+  Windows10-x64:
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-W10.yaml@main
+
+  macOS-ARM64:
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-macOS-ARM64.yaml@main
+
+  macOS-x64:
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-macOS-x86_64.yaml@main
+
+  iOS:
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-iOS.yaml@main
+
+  Android:
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-Android.yaml@main
+
+  TIM-VX:
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-timvx-backend-tests-4.x.yml@main
+
+  docs:
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-docs.yaml@main
@@ -2,6 +2,9 @@ name: arm64 build checks

 on: workflow_dispatch

+permissions:
+  contents: read # to fetch code (actions/checkout)
+
 jobs:
  build:

@@ -0,0 +1,27 @@
+name: lint_python
+on: workflow_dispatch
+permissions:
+  contents: read # to fetch code (actions/checkout)
+jobs:
+  lint_python:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v2
+      - uses: actions/setup-python@v2
+      - run: pip install --upgrade pip wheel
+      - run: pip install bandit black codespell flake8 flake8-2020 flake8-bugbear
+                         flake8-comprehensions isort mypy pytest pyupgrade safety
+      - run: bandit --recursive --skip B101 . || true  # B101 is assert statements
+      - run: black --check . || true
+      - run: codespell || true  # --ignore-words-list="" --skip="*.css,*.js,*.lock"
+      - run: flake8 . --count --select=E9,F63,F7 --show-source --statistics
+      - run: flake8 . --count --exit-zero --max-complexity=10 --max-line-length=88
+                      --show-source --statistics
+      - run: isort --check-only --profile black . || true
+      - run: pip install -r requirements.txt || pip install --editable . || true
+      - run: mkdir --parents --verbose .mypy_cache
+      - run: mypy --ignore-missing-imports --install-types --non-interactive . || true
+      - run: pytest . || true
+      - run: pytest --doctest-modules . || true
+      - run: shopt -s globstar && pyupgrade --py36-plus **/*.py || true
+      - run: safety check
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 2.8.11 FATAL_ERROR)
+cmake_minimum_required(VERSION ${MIN_VER_CMAKE} FATAL_ERROR)

 project(Carotene)

@@ -27,6 +27,10 @@ if(CMAKE_COMPILER_IS_GNUCC)
    endif()
 endif()

+if(APPLE AND CV_CLANG AND WITH_NEON)
+    ocv_warnings_disable(CMAKE_CXX_FLAGS -Wno-unused-function)
+endif()
+
 add_library(carotene_objs OBJECT EXCLUDE_FROM_ALL
  ${carotene_headers}
  ${carotene_sources}
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 2.8.8 FATAL_ERROR)
+cmake_minimum_required(VERSION ${MIN_VER_CMAKE} FATAL_ERROR)

 include(CheckCCompilerFlag)
 include(CheckCXXCompilerFlag)
@@ -1296,13 +1296,13 @@ struct MorphCtx
    CAROTENE_NS::BORDER_MODE border;
    uchar borderValues[4];
 };
-inline int TEGRA_MORPHINIT(cvhalFilter2D **context, int operation, int src_type, int dst_type, int, int,
+inline int TEGRA_MORPHINIT(cvhalFilter2D **context, int operation, int src_type, int dst_type, int width, int height,
                           int kernel_type, uchar *kernel_data, size_t kernel_step, int kernel_width, int kernel_height, int anchor_x, int anchor_y,
                           int borderType, const double borderValue[4], int iterations, bool allowSubmatrix, bool allowInplace)
 {
    if(!context || !kernel_data || src_type != dst_type ||
       CV_MAT_DEPTH(src_type) != CV_8U || src_type < 0 || (src_type >> CV_CN_SHIFT) > 3 ||
-
+       width < kernel_width || height < kernel_height ||
       allowSubmatrix || allowInplace || iterations != 1 ||
       !CAROTENE_NS::isSupportedConfiguration())
        return CV_HAL_ERROR_NOT_IMPLEMENTED;
@@ -1844,14 +1844,18 @@ TegraCvtColor_Invoker(bgrx2hsvf, bgrx2hsv, src_data + static_cast<size_t>(range.
 #define cv_hal_cvtBGRtoGray TEGRA_CVTBGRTOGRAY
 #undef cv_hal_cvtGraytoBGR
 #define cv_hal_cvtGraytoBGR TEGRA_CVTGRAYTOBGR
+#if 0  // bit-exact tests are failed
 #undef cv_hal_cvtBGRtoYUV
 #define cv_hal_cvtBGRtoYUV TEGRA_CVTBGRTOYUV
+#endif
 #undef cv_hal_cvtBGRtoHSV
 #define cv_hal_cvtBGRtoHSV TEGRA_CVTBGRTOHSV
+#if 0  // bit-exact tests are failed
 #undef cv_hal_cvtTwoPlaneYUVtoBGR
 #define cv_hal_cvtTwoPlaneYUVtoBGR TEGRA_CVT2PYUVTOBGR
 #undef cv_hal_cvtTwoPlaneYUVtoBGREx
 #define cv_hal_cvtTwoPlaneYUVtoBGREx TEGRA_CVT2PYUVTOBGR_EX
+#endif

 #endif // OPENCV_IMGPROC_HAL_INTERFACE_H

@@ -109,9 +109,9 @@ template <> struct wAdd<s32>
        vgamma = vdupq_n_f32(_gamma + 0.5);
    }

-    void operator() (const typename VecTraits<s32>::vec128 & v_src0,
-                     const typename VecTraits<s32>::vec128 & v_src1,
-                     typename VecTraits<s32>::vec128 & v_dst) const
+    void operator() (const VecTraits<s32>::vec128 & v_src0,
+                     const VecTraits<s32>::vec128 & v_src1,
+                     VecTraits<s32>::vec128 & v_dst) const
    {
        float32x4_t vs1 = vcvtq_f32_s32(v_src0);
        float32x4_t vs2 = vcvtq_f32_s32(v_src1);
@@ -121,9 +121,9 @@ template <> struct wAdd<s32>
        v_dst = vcvtq_s32_f32(vs1);
    }

-    void operator() (const typename VecTraits<s32>::vec64 & v_src0,
-                     const typename VecTraits<s32>::vec64 & v_src1,
-                     typename VecTraits<s32>::vec64 & v_dst) const
+    void operator() (const VecTraits<s32>::vec64 & v_src0,
+                     const VecTraits<s32>::vec64 & v_src1,
+                     VecTraits<s32>::vec64 & v_dst) const
    {
        float32x2_t vs1 = vcvt_f32_s32(v_src0);
        float32x2_t vs2 = vcvt_f32_s32(v_src1);
@@ -153,9 +153,9 @@ template <> struct wAdd<u32>
        vgamma = vdupq_n_f32(_gamma + 0.5);
    }

-    void operator() (const typename VecTraits<u32>::vec128 & v_src0,
-                     const typename VecTraits<u32>::vec128 & v_src1,
-                     typename VecTraits<u32>::vec128 & v_dst) const
+    void operator() (const VecTraits<u32>::vec128 & v_src0,
+                     const VecTraits<u32>::vec128 & v_src1,
+                     VecTraits<u32>::vec128 & v_dst) const
    {
        float32x4_t vs1 = vcvtq_f32_u32(v_src0);
        float32x4_t vs2 = vcvtq_f32_u32(v_src1);
@@ -165,9 +165,9 @@ template <> struct wAdd<u32>
        v_dst = vcvtq_u32_f32(vs1);
    }

-    void operator() (const typename VecTraits<u32>::vec64 & v_src0,
-                     const typename VecTraits<u32>::vec64 & v_src1,
-                     typename VecTraits<u32>::vec64 & v_dst) const
+    void operator() (const VecTraits<u32>::vec64 & v_src0,
+                     const VecTraits<u32>::vec64 & v_src1,
+                     VecTraits<u32>::vec64 & v_dst) const
    {
        float32x2_t vs1 = vcvt_f32_u32(v_src0);
        float32x2_t vs2 = vcvt_f32_u32(v_src1);
@@ -197,17 +197,17 @@ template <> struct wAdd<f32>
        vgamma = vdupq_n_f32(_gamma + 0.5);
    }

-    void operator() (const typename VecTraits<f32>::vec128 & v_src0,
-                     const typename VecTraits<f32>::vec128 & v_src1,
-                     typename VecTraits<f32>::vec128 & v_dst) const
+    void operator() (const VecTraits<f32>::vec128 & v_src0,
+                     const VecTraits<f32>::vec128 & v_src1,
+                     VecTraits<f32>::vec128 & v_dst) const
    {
        float32x4_t vs1 = vmlaq_f32(vgamma, v_src0, valpha);
        v_dst = vmlaq_f32(vs1, v_src1, vbeta);
    }

-    void operator() (const typename VecTraits<f32>::vec64 & v_src0,
-                     const typename VecTraits<f32>::vec64 & v_src1,
-                     typename VecTraits<f32>::vec64 & v_dst) const
+    void operator() (const VecTraits<f32>::vec64 & v_src0,
+                     const VecTraits<f32>::vec64 & v_src1,
+                     VecTraits<f32>::vec64 & v_dst) const
    {
        float32x2_t vs1 = vmla_f32(vget_low(vgamma), v_src0, vget_low(valpha));
        v_dst = vmla_f32(vs1, v_src1, vget_low(vbeta));
@@ -391,9 +391,9 @@ void blur3x3(const Size2D &size, s32 cn,
                        }
                        else if (borderType == BORDER_MODE_REFLECT101)
                        {
-                            tcurr = vsetq_lane_u16(vgetq_lane_u16(tcurr, 3),tcurr, 5);
                            tcurr = vsetq_lane_u16(vgetq_lane_u16(tcurr, 4),tcurr, 6);
                            tcurr = vsetq_lane_u16(vgetq_lane_u16(tcurr, 5),tcurr, 7);
+                            tcurr = vsetq_lane_u16(vgetq_lane_u16(tcurr, 3),tcurr, 5);
                        }
                        else
                        {
@@ -1,8 +1,8 @@
-# Binaries branch name: ffmpeg/master_20211005
-# Binaries were created for OpenCV: 672399c751c431bbe52818b33fd3ca17b51e0e16
-ocv_update(FFMPEG_BINARIES_COMMIT "40b4666d1aa374205fd61373496e15d92ecd5313")
-ocv_update(FFMPEG_FILE_HASH_BIN32 "c2f9a897d464a2dce2286f8067ad9d90")
-ocv_update(FFMPEG_FILE_HASH_BIN64 "878a4e8fe5a4d68f18c9cdde543b9ead")
+# Binaries branch name: ffmpeg/4.x_20221225
+# Binaries were created for OpenCV: 4abe6dc48d4ec6229f332cc6cf6c7e234ac8027e
+ocv_update(FFMPEG_BINARIES_COMMIT "7dd0d4f1d6fe75f05f3d3b5e38cbc96c1a2d2809")
+ocv_update(FFMPEG_FILE_HASH_BIN32 "e598ae2ece1ddf310bc49b58202fd87a")
+ocv_update(FFMPEG_FILE_HASH_BIN64 "b2a40c142c20aef9fd663fc8f85c2971")
 ocv_update(FFMPEG_FILE_HASH_CMAKE "8862c87496e2e8c375965e1277dee1c7")

 function(download_win_ffmpeg script_var)
@@ -54,6 +54,7 @@ set_target_properties(${ITT_LIBRARY} PROPERTIES
    )

 ocv_warnings_disable(CMAKE_C_FLAGS -Wundef -Wsign-compare)
+ocv_warnings_disable(CMAKE_C_FLAGS -Wstrict-prototypes) # clang15

 if(ENABLE_SOLUTION_FOLDERS)
  set_target_properties(${ITT_LIBRARY} PROPERTIES FOLDER "3rdparty")
@@ -4,9 +4,9 @@ ocv_warnings_disable(CMAKE_C_FLAGS -Wunused-parameter -Wsign-compare -Wshorten-6

 set(VERSION_MAJOR 2)
 set(VERSION_MINOR 1)
-set(VERSION_REVISION 0)
+set(VERSION_REVISION 3)
 set(VERSION ${VERSION_MAJOR}.${VERSION_MINOR}.${VERSION_REVISION})
-set(LIBJPEG_TURBO_VERSION_NUMBER 2001000)
+set(LIBJPEG_TURBO_VERSION_NUMBER 2001003)

 string(TIMESTAMP BUILD "opencv-${OPENCV_VERSION}-libjpeg-turbo")
 if(CMAKE_BUILD_TYPE STREQUAL "Debug")
@@ -15,8 +15,53 @@ endif()

 message(STATUS "libjpeg-turbo: VERSION = ${VERSION}, BUILD = ${BUILD}")

+math(EXPR BITS "${CMAKE_SIZEOF_VOID_P} * 8")
+string(TOLOWER ${CMAKE_SYSTEM_PROCESSOR} CMAKE_SYSTEM_PROCESSOR_LC)
+
+if(CMAKE_SYSTEM_PROCESSOR_LC MATCHES "x86_64" OR
+  CMAKE_SYSTEM_PROCESSOR_LC MATCHES "amd64" OR
+  CMAKE_SYSTEM_PROCESSOR_LC MATCHES "i[0-9]86" OR
+  CMAKE_SYSTEM_PROCESSOR_LC MATCHES "x86" OR
+  CMAKE_SYSTEM_PROCESSOR_LC MATCHES "ia32")
+  if(BITS EQUAL 64 OR CMAKE_C_COMPILER_ABI MATCHES "ELF X32")
+    set(CPU_TYPE x86_64)
+  else()
+    set(CPU_TYPE i386)
+  endif()
+  if(NOT CMAKE_SYSTEM_PROCESSOR STREQUAL ${CPU_TYPE})
+    set(CMAKE_SYSTEM_PROCESSOR ${CPU_TYPE})
+  endif()
+elseif(CMAKE_SYSTEM_PROCESSOR_LC STREQUAL "aarch64" OR
+  CMAKE_SYSTEM_PROCESSOR_LC MATCHES "^arm")
+  if(BITS EQUAL 64)
+    set(CPU_TYPE arm64)
+  else()
+    set(CPU_TYPE arm)
+  endif()
+elseif(CMAKE_SYSTEM_PROCESSOR_LC MATCHES "^ppc" OR
+  CMAKE_SYSTEM_PROCESSOR_LC MATCHES "^powerpc")
+  set(CPU_TYPE powerpc)
+else()
+  set(CPU_TYPE ${CMAKE_SYSTEM_PROCESSOR_LC})
+endif()
+if(CMAKE_OSX_ARCHITECTURES MATCHES "x86_64" OR
+  CMAKE_OSX_ARCHITECTURES MATCHES "arm64" OR
+  CMAKE_OSX_ARCHITECTURES MATCHES "i386")
+  set(CPU_TYPE ${CMAKE_OSX_ARCHITECTURES})
+endif()
+if(CMAKE_OSX_ARCHITECTURES MATCHES "ppc")
+  set(CPU_TYPE powerpc)
+endif()
+if(MSVC_IDE AND CMAKE_GENERATOR_PLATFORM MATCHES "arm64")
+  set(CPU_TYPE arm64)
+endif()
+
+OCV_OPTION(ENABLE_LIBJPEG_TURBO_SIMD "Include SIMD extensions for libjpeg-turbo, if available for this platform" (NOT CV_DISABLE_OPTIMIZATION)
+  VISIBLE_IF BUILD_JPEG)
 option(WITH_ARITH_ENC "Include arithmetic encoding support when emulating the libjpeg v6b API/ABI" TRUE)
 option(WITH_ARITH_DEC "Include arithmetic decoding support when emulating the libjpeg v6b API/ABI" TRUE)
+set(WITH_SIMD 1)
+set(HAVE_LIBJPEG_TURBO_SIMD 0 PARENT_SCOPE)

 include(CheckCSourceCompiles)
 include(CheckIncludeFiles)
@@ -79,14 +124,13 @@ configure_file(jconfigint.h.in jconfigint.h)

 include_directories(${CMAKE_CURRENT_BINARY_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/src)

-set(JPEG_SOURCES
-  jcapimin.c jcapistd.c jccoefct.c jccolor.c jcdctmgr.c jchuff.c jcicc.c
-  jcinit.c jcmainct.c jcmarker.c jcmaster.c jcomapi.c jcparam.c jcphuff.c
-  jcprepct.c jcsample.c jctrans.c jdapimin.c jdapistd.c jdatadst.c jdatasrc.c
-  jdcoefct.c jdcolor.c jddctmgr.c jdhuff.c jdicc.c jdinput.c jdmainct.c jdmarker.c
-  jdmaster.c jdmerge.c jdphuff.c jdpostct.c jdsample.c jdtrans.c jerror.c
-  jfdctflt.c jfdctfst.c jfdctint.c jidctflt.c jidctfst.c jidctint.c jidctred.c
-  jquant1.c jquant2.c jutils.c jmemmgr.c jmemnobs.c)
+set(JPEG_SOURCES jcapimin.c jcapistd.c jccoefct.c jccolor.c jcdctmgr.c jchuff.c
+        jcicc.c jcinit.c jcmainct.c jcmarker.c jcmaster.c jcomapi.c jcparam.c
+        jcphuff.c jcprepct.c jcsample.c jctrans.c jdapimin.c jdapistd.c jdatadst.c
+        jdatasrc.c jdcoefct.c jdcolor.c jddctmgr.c jdhuff.c jdicc.c jdinput.c
+        jdmainct.c jdmarker.c jdmaster.c jdmerge.c jdphuff.c jdpostct.c jdsample.c
+        jdtrans.c jerror.c jfdctflt.c jfdctfst.c jfdctint.c jidctflt.c jidctfst.c
+        jidctint.c jidctred.c jquant1.c jquant2.c jutils.c jmemmgr.c jmemnobs.c)

 if(WITH_ARITH_ENC OR WITH_ARITH_DEC)
  set(JPEG_SOURCES ${JPEG_SOURCES} jaricom.c)
@@ -100,12 +144,57 @@ if(WITH_ARITH_DEC)
  set(JPEG_SOURCES ${JPEG_SOURCES} jdarith.c)
 endif()

-# No SIMD
-set(JPEG_SOURCES ${JPEG_SOURCES} jsimd_none.c)
+if(CMAKE_COMPILER_IS_GNUCC OR CMAKE_C_COMPILER_ID STREQUAL "Clang")
+  # Use the maximum optimization level for release builds
+  foreach(var CMAKE_C_FLAGS_RELEASE CMAKE_C_FLAGS_RELWITHDEBINFO)
+    if(${var} MATCHES "-O2")
+      string(REGEX REPLACE "-O2" "-O3" ${var} "${${var}}")
+    endif()
+  endforeach()
+endif()
+
+if(CMAKE_SYSTEM_NAME STREQUAL "SunOS")
+  if(CMAKE_C_COMPILER_ID MATCHES "SunPro")
+    # Use the maximum optimization level for release builds
+    foreach(var CMAKE_C_FLAGS_RELEASE CMAKE_C_FLAGS_RELWITHDEBINFO)
+      if(${var} MATCHES "-xO3")
+        string(REGEX REPLACE "-xO3" "-xO5" ${var} "${${var}}")
+      endif()
+      if(${var} MATCHES "-xO2")
+        string(REGEX REPLACE "-xO2" "-xO5" ${var} "${${var}}")
+      endif()
+    endforeach()
+  endif()
+endif()
+
+if(ENABLE_LIBJPEG_TURBO_SIMD)
+  add_subdirectory(src/simd)
+  if(NEON_INTRINSICS)
+    add_definitions(-DNEON_INTRINSICS)
+  endif()
+else()
+  set(WITH_SIMD 0)
+endif()
+
+if(WITH_SIMD)
+  message(STATUS "SIMD extensions: ${CPU_TYPE} (WITH_SIMD = ${WITH_SIMD})")
+  set(HAVE_LIBJPEG_TURBO_SIMD 1 PARENT_SCOPE)
+  if(MSVC_IDE OR XCODE)
+    set_source_files_properties(${SIMD_OBJS} PROPERTIES GENERATED 1)
+  endif()
+else()
+  add_library(jsimd OBJECT src/jsimd_none.c)
+  set_target_properties(jsimd PROPERTIES FOLDER "3rdparty")
+  if(NOT WIN32 AND (CMAKE_POSITION_INDEPENDENT_CODE OR ENABLE_SHARED))
+    set_target_properties(jsimd PROPERTIES POSITION_INDEPENDENT_CODE 1)
+  endif()
+endif()

 ocv_list_add_prefix(JPEG_SOURCES src/)

-add_library(${JPEG_LIBRARY} STATIC ${OPENCV_3RDPARTY_EXCLUDE_FROM_ALL} ${JPEG_SOURCES} ${SIMD_OBJS})
+set(JPEG_SOURCES ${JPEG_SOURCES} ${SIMD_OBJS})
+
+add_library(${JPEG_LIBRARY} STATIC ${OPENCV_3RDPARTY_EXCLUDE_FROM_ALL} ${JPEG_SOURCES} $<TARGET_OBJECTS:jsimd> ${SIMD_OBJS})

 set_target_properties(${JPEG_LIBRARY}
  PROPERTIES OUTPUT_NAME ${JPEG_LIBRARY}
@@ -91,7 +91,7 @@ best of our understanding.
 The Modified (3-clause) BSD License
 ===================================

-Copyright (C)2009-2021 D. R. Commander.  All Rights Reserved.<br>
+Copyright (C)2009-2022 D. R. Commander.  All Rights Reserved.<br>
 Copyright (C)2015 Viktor Szathmáry.  All Rights Reserved.

 Redistribution and use in source and binary forms, with or without
@@ -40,3 +40,13 @@
 #define HAVE_BITSCANFORWARD
 #endif
 #endif
+
+#if defined(__has_attribute)
+#if __has_attribute(fallthrough)
+#define FALLTHROUGH  __attribute__((fallthrough));
+#else
+#define FALLTHROUGH
+#endif
+#else
+#define FALLTHROUGH
+#endif
@@ -4,8 +4,8 @@
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1994-1998, Thomas G. Lane.
 * Modified 2003-2010 by Guido Vollbeding.
- * It was modified by The libjpeg-turbo Project to include only code relevant
- * to libjpeg-turbo.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -52,7 +52,7 @@ jpeg_CreateCompress(j_compress_ptr cinfo, int version, size_t structsize)
  {
    struct jpeg_error_mgr *err = cinfo->err;
    void *client_data = cinfo->client_data; /* ignore Purify complaint here */
-    MEMZERO(cinfo, sizeof(struct jpeg_compress_struct));
+    memset(cinfo, 0, sizeof(struct jpeg_compress_struct));
    cinfo->err = err;
    cinfo->client_data = client_data;
  }
@@ -4,7 +4,7 @@
 * This file was part of the Independent JPEG Group's software:
 * Developed 1997-2009 by Guido Vollbeding.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2015, 2018, D. R. Commander.
+ * Copyright (C) 2015, 2018, 2021-2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -338,14 +338,14 @@ emit_restart(j_compress_ptr cinfo, int restart_num)
    compptr = cinfo->cur_comp_info[ci];
    /* DC needs no table for refinement scan */
    if (cinfo->progressive_mode == 0 || (cinfo->Ss == 0 && cinfo->Ah == 0)) {
-      MEMZERO(entropy->dc_stats[compptr->dc_tbl_no], DC_STAT_BINS);
+      memset(entropy->dc_stats[compptr->dc_tbl_no], 0, DC_STAT_BINS);
      /* Reset DC predictions to 0 */
      entropy->last_dc_val[ci] = 0;
      entropy->dc_context[ci] = 0;
    }
    /* AC needs no table when not present */
    if (cinfo->progressive_mode == 0 || cinfo->Se) {
-      MEMZERO(entropy->ac_stats[compptr->ac_tbl_no], AC_STAT_BINS);
+      memset(entropy->ac_stats[compptr->ac_tbl_no], 0, AC_STAT_BINS);
    }
  }

@@ -836,7 +836,7 @@ start_pass(j_compress_ptr cinfo, boolean gather_statistics)
     * We are fully adaptive here and need no extra
     * statistics gathering pass!
     */
-    ERREXIT(cinfo, JERR_NOT_COMPILED);
+    ERREXIT(cinfo, JERR_NOTIMPL);

  /* We assume jcmaster.c already validated the progressive scan parameters. */

@@ -867,7 +867,7 @@ start_pass(j_compress_ptr cinfo, boolean gather_statistics)
      if (entropy->dc_stats[tbl] == NULL)
        entropy->dc_stats[tbl] = (unsigned char *)(*cinfo->mem->alloc_small)
          ((j_common_ptr)cinfo, JPOOL_IMAGE, DC_STAT_BINS);
-      MEMZERO(entropy->dc_stats[tbl], DC_STAT_BINS);
+      memset(entropy->dc_stats[tbl], 0, DC_STAT_BINS);
      /* Initialize DC predictions to 0 */
      entropy->last_dc_val[ci] = 0;
      entropy->dc_context[ci] = 0;
@@ -880,7 +880,7 @@ start_pass(j_compress_ptr cinfo, boolean gather_statistics)
      if (entropy->ac_stats[tbl] == NULL)
        entropy->ac_stats[tbl] = (unsigned char *)(*cinfo->mem->alloc_small)
          ((j_common_ptr)cinfo, JPOOL_IMAGE, AC_STAT_BINS);
-      MEMZERO(entropy->ac_stats[tbl], AC_STAT_BINS);
+      memset(entropy->ac_stats[tbl], 0, AC_STAT_BINS);
 #ifdef CALCULATE_SPECTRAL_CONDITIONING
      if (cinfo->progressive_mode)
        /* Section G.1.3.2: Set appropriate arithmetic conditioning value Kx */
@@ -4,7 +4,7 @@
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1991-1997, Thomas G. Lane.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2009-2011, 2014-2016, 2018-2021, D. R. Commander.
+ * Copyright (C) 2009-2011, 2014-2016, 2018-2022, D. R. Commander.
 * Copyright (C) 2015, Matthieu Darbois.
 * Copyright (C) 2018, Matthias Räncker.
 * Copyright (C) 2020, Arm Limited.
@@ -44,8 +44,9 @@
 * flags (this defines __thumb__).
 */

-#if defined(__arm__) || defined(__aarch64__) || defined(_M_ARM) || \
-    defined(_M_ARM64)
+/* NOTE: Both GCC and Clang define __GNUC__ */
+#if (defined(__GNUC__) && (defined(__arm__) || defined(__aarch64__))) || \
+    defined(_M_ARM) || defined(_M_ARM64)
 #if !defined(__thumb__) || defined(__thumb2__)
 #define USE_CLZ_INTRINSIC
 #endif
@@ -199,12 +200,12 @@ start_pass_huff(j_compress_ptr cinfo, boolean gather_statistics)
        entropy->dc_count_ptrs[dctbl] = (long *)
          (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
                                      257 * sizeof(long));
-      MEMZERO(entropy->dc_count_ptrs[dctbl], 257 * sizeof(long));
+      memset(entropy->dc_count_ptrs[dctbl], 0, 257 * sizeof(long));
      if (entropy->ac_count_ptrs[actbl] == NULL)
        entropy->ac_count_ptrs[actbl] = (long *)
          (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
                                      257 * sizeof(long));
-      MEMZERO(entropy->ac_count_ptrs[actbl], 257 * sizeof(long));
+      memset(entropy->ac_count_ptrs[actbl], 0, 257 * sizeof(long));
 #endif
    } else {
      /* Compute derived values for Huffman tables */
@@ -314,8 +315,8 @@ jpeg_make_c_derived_tbl(j_compress_ptr cinfo, boolean isDC, int tblno,
   * this lets us detect duplicate VAL entries here, and later
   * allows emit_bits to detect any attempt to emit such symbols.
   */
-  MEMZERO(dtbl->ehufco, sizeof(dtbl->ehufco));
-  MEMZERO(dtbl->ehufsi, sizeof(dtbl->ehufsi));
+  memset(dtbl->ehufco, 0, sizeof(dtbl->ehufco));
+  memset(dtbl->ehufsi, 0, sizeof(dtbl->ehufsi));

  /* This is also a convenient place to check for out-of-range
   * and duplicated VAL entries.  We allow 0..255 for AC symbols
@@ -477,7 +478,7 @@ dump_buffer(working_state *state)
    buffer = _buffer; \
    while (bytes > 0) { \
      bytestocopy = MIN(bytes, state->free_in_buffer); \
-      MEMCOPY(state->next_output_byte, buffer, bytestocopy); \
+      memcpy(state->next_output_byte, buffer, bytestocopy); \
      state->next_output_byte += bytestocopy; \
      buffer += bytestocopy; \
      state->free_in_buffer -= bytestocopy; \
@@ -940,8 +941,8 @@ jpeg_gen_optimal_table(j_compress_ptr cinfo, JHUFF_TBL *htbl, long freq[])

  /* This algorithm is explained in section K.2 of the JPEG standard */

-  MEMZERO(bits, sizeof(bits));
-  MEMZERO(codesize, sizeof(codesize));
+  memset(bits, 0, sizeof(bits));
+  memset(codesize, 0, sizeof(codesize));
  for (i = 0; i < 257; i++)
    others[i] = -1;             /* init links to empty */

@@ -1043,7 +1044,7 @@ jpeg_gen_optimal_table(j_compress_ptr cinfo, JHUFF_TBL *htbl, long freq[])
  bits[i]--;

  /* Return final symbol counts (only for lengths 0..16) */
-  MEMCOPY(htbl->bits, bits, sizeof(htbl->bits));
+  memcpy(htbl->bits, bits, sizeof(htbl->bits));

  /* Return a list of the symbols sorted by code length */
  /* It's not real clear to me why we don't need to consider the codelength
@@ -1082,8 +1083,8 @@ finish_pass_gather(j_compress_ptr cinfo)
  /* It's important not to apply jpeg_gen_optimal_table more than once
   * per table, because it clobbers the input frequency counts!
   */
-  MEMZERO(did_dc, sizeof(did_dc));
-  MEMZERO(did_ac, sizeof(did_ac));
+  memset(did_dc, 0, sizeof(did_dc));
+  memset(did_ac, 0, sizeof(did_ac));

  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
    compptr = cinfo->cur_comp_info[ci];
@@ -493,7 +493,7 @@ prepare_for_pass(j_compress_ptr cinfo)
    master->pass_type = output_pass;
    master->pass_number++;
 #endif
-    /*FALLTHROUGH*/
+    FALLTHROUGH                 /*FALLTHROUGH*/
  case output_pass:
    /* Do a data-output pass. */
    /* We need not repeat per-scan setup if prior optimization pass did it. */
@@ -4,9 +4,10 @@
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1995-1997, Thomas G. Lane.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2011, 2015, 2018, 2021, D. R. Commander.
+ * Copyright (C) 2011, 2015, 2018, 2021-2022, D. R. Commander.
 * Copyright (C) 2016, 2018, Matthieu Darbois.
 * Copyright (C) 2020, Arm Limited.
+ * Copyright (C) 2021, Alex Richardson.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -52,8 +53,9 @@
 * flags (this defines __thumb__).
 */

-#if defined(__arm__) || defined(__aarch64__) || defined(_M_ARM) || \
-    defined(_M_ARM64)
+/* NOTE: Both GCC and Clang define __GNUC__ */
+#if (defined(__GNUC__) && (defined(__arm__) || defined(__aarch64__))) || \
+    defined(_M_ARM) || defined(_M_ARM64)
 #if !defined(__thumb__) || defined(__thumb2__)
 #define USE_CLZ_INTRINSIC
 #endif
@@ -273,7 +275,7 @@ start_pass_phuff(j_compress_ptr cinfo, boolean gather_statistics)
        entropy->count_ptrs[tbl] = (long *)
          (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
                                      257 * sizeof(long));
-      MEMZERO(entropy->count_ptrs[tbl], 257 * sizeof(long));
+      memset(entropy->count_ptrs[tbl], 0, 257 * sizeof(long));
    } else {
      /* Compute derived values for Huffman table */
      /* We may do this more than once for a table, but it's not expensive */
@@ -582,8 +584,8 @@ encode_mcu_DC_first(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
      continue; \
    /* For a negative coef, want temp2 = bitwise complement of abs(coef) */ \
    temp2 ^= temp; \
-    values[k] = temp; \
-    values[k + DCTSIZE2] = temp2; \
+    values[k] = (JCOEF)temp; \
+    values[k + DCTSIZE2] = (JCOEF)temp2; \
    zerobits |= ((size_t)1U) << k; \
  } \
 }
@@ -679,7 +681,7 @@ encode_mcu_AC_first(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
      emit_restart(entropy, entropy->next_restart_num);

 #ifdef WITH_SIMD
-  cvalue = values = (JCOEF *)PAD((size_t)values_unaligned, 16);
+  cvalue = values = (JCOEF *)PAD((JUINTPTR)values_unaligned, 16);
 #else
  /* Not using SIMD, so alignment is not needed */
  cvalue = values = values_unaligned;
@@ -944,7 +946,7 @@ encode_mcu_AC_refine(j_compress_ptr cinfo, JBLOCKROW *MCU_data)
      emit_restart(entropy, entropy->next_restart_num);

 #ifdef WITH_SIMD
-  cabsvalue = absvalues = (JCOEF *)PAD((size_t)absvalues_unaligned, 16);
+  cabsvalue = absvalues = (JCOEF *)PAD((JUINTPTR)absvalues_unaligned, 16);
 #else
  /* Not using SIMD, so alignment is not needed */
  cabsvalue = absvalues = absvalues_unaligned;
@@ -1060,7 +1062,7 @@ finish_pass_gather_phuff(j_compress_ptr cinfo)
  /* It's important not to apply jpeg_gen_optimal_table more than once
   * per table, because it clobbers the input frequency counts!
   */
-  MEMZERO(did, sizeof(did));
+  memset(did, 0, sizeof(did));

  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
    compptr = cinfo->cur_comp_info[ci];
@@ -3,8 +3,8 @@
 *
 * This file is part of the Independent JPEG Group's software:
 * Copyright (C) 1994-1996, Thomas G. Lane.
- * It was modified by The libjpeg-turbo Project to include only code relevant
- * to libjpeg-turbo.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -289,8 +289,8 @@ create_context_buffer(j_compress_ptr cinfo)
                     cinfo->max_h_samp_factor) / compptr->h_samp_factor),
       (JDIMENSION)(3 * rgroup_height));
    /* Copy true buffer row pointers into the middle of the fake row array */
-    MEMCOPY(fake_buffer + rgroup_height, true_buffer,
-            3 * rgroup_height * sizeof(JSAMPROW));
+    memcpy(fake_buffer + rgroup_height, true_buffer,
+           3 * rgroup_height * sizeof(JSAMPROW));
    /* Fill in the above and below wraparound pointers */
    for (i = 0; i < rgroup_height; i++) {
      fake_buffer[i] = true_buffer[2 * rgroup_height + i];
@@ -5,7 +5,7 @@
 * Copyright (C) 1995-1998, Thomas G. Lane.
 * Modified 2000-2009 by Guido Vollbeding.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2020, D. R. Commander.
+ * Copyright (C) 2020, 2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -100,8 +100,8 @@ jpeg_copy_critical_parameters(j_decompress_ptr srcinfo, j_compress_ptr dstinfo)
      qtblptr = &dstinfo->quant_tbl_ptrs[tblno];
      if (*qtblptr == NULL)
        *qtblptr = jpeg_alloc_quant_table((j_common_ptr)dstinfo);
-      MEMCOPY((*qtblptr)->quantval, srcinfo->quant_tbl_ptrs[tblno]->quantval,
-              sizeof((*qtblptr)->quantval));
+      memcpy((*qtblptr)->quantval, srcinfo->quant_tbl_ptrs[tblno]->quantval,
+             sizeof((*qtblptr)->quantval));
      (*qtblptr)->sent_table = FALSE;
    }
  }
@@ -4,7 +4,7 @@
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1994-1998, Thomas G. Lane.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2016, D. R. Commander.
+ * Copyright (C) 2016, 2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -23,6 +23,7 @@
 #include "jinclude.h"
 #include "jpeglib.h"
 #include "jdmaster.h"
+#include "jconfigint.h"


 /*
@@ -52,7 +53,7 @@ jpeg_CreateDecompress(j_decompress_ptr cinfo, int version, size_t structsize)
  {
    struct jpeg_error_mgr *err = cinfo->err;
    void *client_data = cinfo->client_data; /* ignore Purify complaint here */
-    MEMZERO(cinfo, sizeof(struct jpeg_decompress_struct));
+    memset(cinfo, 0, sizeof(struct jpeg_decompress_struct));
    cinfo->err = err;
    cinfo->client_data = client_data;
  }
@@ -91,7 +92,7 @@ jpeg_CreateDecompress(j_decompress_ptr cinfo, int version, size_t structsize)
  cinfo->master = (struct jpeg_decomp_master *)
    (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_PERMANENT,
                                sizeof(my_decomp_master));
-  MEMZERO(cinfo->master, sizeof(my_decomp_master));
+  memset(cinfo->master, 0, sizeof(my_decomp_master));
 }


@@ -308,7 +309,7 @@ jpeg_consume_input(j_decompress_ptr cinfo)
    /* Initialize application's data source module */
    (*cinfo->src->init_source) (cinfo);
    cinfo->global_state = DSTATE_INHEADER;
-    /*FALLTHROUGH*/
+    FALLTHROUGH                 /*FALLTHROUGH*/
  case DSTATE_INHEADER:
    retcode = (*cinfo->inputctl->consume_input) (cinfo);
    if (retcode == JPEG_REACHED_SOS) { /* Found SOS, prepare to decompress */
@@ -4,7 +4,7 @@
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1994-1996, Thomas G. Lane.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2010, 2015-2020, D. R. Commander.
+ * Copyright (C) 2010, 2015-2020, 2022, D. R. Commander.
 * Copyright (C) 2015, Google, Inc.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
@@ -159,6 +159,7 @@ jpeg_crop_scanline(j_decompress_ptr cinfo, JDIMENSION *xoffset,
  JDIMENSION input_xoffset;
  boolean reinit_upsampler = FALSE;
  jpeg_component_info *compptr;
+  my_master_ptr master = (my_master_ptr)cinfo->master;

  if (cinfo->global_state != DSTATE_SCANNING || cinfo->output_scanline != 0)
    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
@@ -208,6 +209,11 @@ jpeg_crop_scanline(j_decompress_ptr cinfo, JDIMENSION *xoffset,
   */
  *width = *width + input_xoffset - *xoffset;
  cinfo->output_width = *width;
+  if (master->using_merged_upsample && cinfo->max_v_samp_factor == 2) {
+    my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+    upsample->out_row_width =
+      cinfo->output_width * cinfo->out_color_components;
+  }

  /* Set the first and last iMCU columns that we must decompress.  These values
   * will be used in single-scan decompressions.
@@ -4,7 +4,7 @@
 * This file was part of the Independent JPEG Group's software:
 * Developed 1997-2015 by Guido Vollbeding.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2015-2020, D. R. Commander.
+ * Copyright (C) 2015-2020, 2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -210,13 +210,13 @@ process_restart(j_decompress_ptr cinfo)
  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
    compptr = cinfo->cur_comp_info[ci];
    if (!cinfo->progressive_mode || (cinfo->Ss == 0 && cinfo->Ah == 0)) {
-      MEMZERO(entropy->dc_stats[compptr->dc_tbl_no], DC_STAT_BINS);
+      memset(entropy->dc_stats[compptr->dc_tbl_no], 0, DC_STAT_BINS);
      /* Reset DC predictions to 0 */
      entropy->last_dc_val[ci] = 0;
      entropy->dc_context[ci] = 0;
    }
    if (!cinfo->progressive_mode || cinfo->Ss) {
-      MEMZERO(entropy->ac_stats[compptr->ac_tbl_no], AC_STAT_BINS);
+      memset(entropy->ac_stats[compptr->ac_tbl_no], 0, AC_STAT_BINS);
    }
  }

@@ -471,17 +471,17 @@ decode_mcu_AC_refine(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
      if (*thiscoef) {                          /* previously nonzero coef */
        if (arith_decode(cinfo, st + 2)) {
          if (*thiscoef < 0)
-            *thiscoef += m1;
+            *thiscoef += (JCOEF)m1;
          else
-            *thiscoef += p1;
+            *thiscoef += (JCOEF)p1;
        }
        break;
      }
      if (arith_decode(cinfo, st + 1)) {        /* newly nonzero coef */
        if (arith_decode(cinfo, entropy->fixed_bin))
-          *thiscoef = m1;
+          *thiscoef = (JCOEF)m1;
        else
-          *thiscoef = p1;
+          *thiscoef = (JCOEF)p1;
        break;
      }
      st += 3;  k++;
@@ -698,8 +698,8 @@ bad:
    /* Check that the scan parameters Ss, Se, Ah/Al are OK for sequential JPEG.
     * This ought to be an error condition, but we make it a warning.
     */
-    if (cinfo->Ss != 0 || cinfo->Ah != 0 || cinfo->Al != 0 ||
-        (cinfo->Se < DCTSIZE2 && cinfo->Se != DCTSIZE2 - 1))
+    if (cinfo->Ss != 0 || cinfo->Se != DCTSIZE2 - 1 ||
+        cinfo->Ah != 0 || cinfo->Al != 0)
      WARNMS(cinfo, JWRN_NOT_SEQUENTIAL);
    /* Select MCU decoding routine */
    entropy->pub.decode_mcu = decode_mcu;
@@ -715,7 +715,7 @@ bad:
      if (entropy->dc_stats[tbl] == NULL)
        entropy->dc_stats[tbl] = (unsigned char *)(*cinfo->mem->alloc_small)
          ((j_common_ptr)cinfo, JPOOL_IMAGE, DC_STAT_BINS);
-      MEMZERO(entropy->dc_stats[tbl], DC_STAT_BINS);
+      memset(entropy->dc_stats[tbl], 0, DC_STAT_BINS);
      /* Initialize DC predictions to 0 */
      entropy->last_dc_val[ci] = 0;
      entropy->dc_context[ci] = 0;
@@ -727,7 +727,7 @@ bad:
      if (entropy->ac_stats[tbl] == NULL)
        entropy->ac_stats[tbl] = (unsigned char *)(*cinfo->mem->alloc_small)
          ((j_common_ptr)cinfo, JPOOL_IMAGE, AC_STAT_BINS);
-      MEMZERO(entropy->ac_stats[tbl], AC_STAT_BINS);
+      memset(entropy->ac_stats[tbl], 0, AC_STAT_BINS);
    }
  }

@@ -5,7 +5,7 @@
 * Copyright (C) 1994-1996, Thomas G. Lane.
 * Modified 2009-2012 by Guido Vollbeding.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2013, 2016, D. R. Commander.
+ * Copyright (C) 2013, 2016, 2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -23,11 +23,6 @@
 #include "jpeglib.h"
 #include "jerror.h"

-#ifndef HAVE_STDLIB_H           /* <stdlib.h> should declare malloc(),free() */
-extern void *malloc(size_t size);
-extern void free(void *ptr);
-#endif
-

 /* Expanded data destination object for stdio output */

@@ -116,7 +111,7 @@ empty_output_buffer(j_compress_ptr cinfo)
 {
  my_dest_ptr dest = (my_dest_ptr)cinfo->dest;

-  if (JFWRITE(dest->outfile, dest->buffer, OUTPUT_BUF_SIZE) !=
+  if (fwrite(dest->buffer, 1, OUTPUT_BUF_SIZE, dest->outfile) !=
      (size_t)OUTPUT_BUF_SIZE)
    ERREXIT(cinfo, JERR_FILE_WRITE);

@@ -141,7 +136,7 @@ empty_mem_output_buffer(j_compress_ptr cinfo)
  if (nextbuffer == NULL)
    ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 10);

-  MEMCOPY(nextbuffer, dest->buffer, dest->bufsize);
+  memcpy(nextbuffer, dest->buffer, dest->bufsize);

  free(dest->newbuffer);

@@ -175,7 +170,7 @@ term_destination(j_compress_ptr cinfo)

  /* Write any data remaining in the buffer */
  if (datacount > 0) {
-    if (JFWRITE(dest->outfile, dest->buffer, datacount) != datacount)
+    if (fwrite(dest->buffer, 1, datacount, dest->outfile) != datacount)
      ERREXIT(cinfo, JERR_FILE_WRITE);
  }
  fflush(dest->outfile);
@@ -5,7 +5,7 @@
 * Copyright (C) 1994-1996, Thomas G. Lane.
 * Modified 2009-2011 by Guido Vollbeding.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2013, 2016, D. R. Commander.
+ * Copyright (C) 2013, 2016, 2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -104,7 +104,7 @@ fill_input_buffer(j_decompress_ptr cinfo)
  my_src_ptr src = (my_src_ptr)cinfo->src;
  size_t nbytes;

-  nbytes = JFREAD(src->infile, src->buffer, INPUT_BUF_SIZE);
+  nbytes = fread(src->buffer, 1, INPUT_BUF_SIZE, src->infile);

  if (nbytes <= 0) {
    if (src->start_of_file)     /* Treat empty input file as fatal error */
@@ -6,7 +6,7 @@
 * Modified 2002-2010 by Guido Vollbeding.
 * libjpeg-turbo Modifications:
 * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
- * Copyright (C) 2010, 2015, D. R. Commander.
+ * Copyright (C) 2010, 2015, 2022, D. R. Commander.
 * Copyright (C) 2013, MIPS Technologies, Inc., California.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
@@ -345,7 +345,7 @@ jinit_inverse_dct(j_decompress_ptr cinfo)
    compptr->dct_table =
      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
                                  sizeof(multiplier_table));
-    MEMZERO(compptr->dct_table, sizeof(multiplier_table));
+    memset(compptr->dct_table, 0, sizeof(multiplier_table));
    /* Mark multiplier table not yet set up for any method */
    idct->cur_method[ci] = -1;
  }
@@ -584,7 +584,7 @@ decode_mcu_slow(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
       * behavior is, to the best of our understanding, innocuous, and it is
       * unclear how to work around it without potentially affecting
       * performance.  Thus, we (hopefully temporarily) suppress UBSan integer
-       * overflow errors for this function.
+       * overflow errors for this function and decode_mcu_fast().
       */
      s += state.last_dc_val[ci];
      state.last_dc_val[ci] = s;
@@ -651,6 +651,12 @@ decode_mcu_slow(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
 }


+#if defined(__has_feature)
+#if __has_feature(undefined_behavior_sanitizer)
+__attribute__((no_sanitize("signed-integer-overflow"),
+               no_sanitize("unsigned-integer-overflow")))
+#endif
+#endif
 LOCAL(boolean)
 decode_mcu_fast(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
 {
@@ -681,6 +687,9 @@ decode_mcu_fast(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)

    if (entropy->dc_needed[blkn]) {
      int ci = cinfo->MCU_membership[blkn];
+      /* Refer to the comment in decode_mcu_slow() regarding the supression of
+       * a UBSan integer overflow error in this line of code.
+       */
      s += state.last_dc_val[ci];
      state.last_dc_val[ci] = s;
      if (block)
@@ -18,10 +18,6 @@
 #include "jpeglib.h"
 #include "jerror.h"

-#ifndef HAVE_STDLIB_H           /* <stdlib.h> should declare malloc() */
-extern void *malloc(size_t size);
-#endif
-

 #define ICC_MARKER  (JPEG_APP0 + 2)     /* JPEG marker code for ICC */
 #define ICC_OVERHEAD_LEN  14            /* size of non-profile data in APP2 */
@@ -4,7 +4,7 @@
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1991-1997, Thomas G. Lane.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2010, 2016, 2018, D. R. Commander.
+ * Copyright (C) 2010, 2016, 2018, 2022, D. R. Commander.
 * Copyright (C) 2015, Google, Inc.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
@@ -264,7 +264,7 @@ latch_quant_tables(j_decompress_ptr cinfo)
    qtbl = (JQUANT_TBL *)
      (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
                                  sizeof(JQUANT_TBL));
-    MEMCOPY(qtbl, cinfo->quant_tbl_ptrs[qtblno], sizeof(JQUANT_TBL));
+    memcpy(qtbl, cinfo->quant_tbl_ptrs[qtblno], sizeof(JQUANT_TBL));
    compptr->quant_table = qtbl;
  }
 }
@@ -18,6 +18,7 @@

 #include "jinclude.h"
 #include "jdmainct.h"
+#include "jconfigint.h"


 /*
@@ -360,7 +361,7 @@ process_data_context_main(j_decompress_ptr cinfo, JSAMPARRAY output_buf,
    main_ptr->context_state = CTX_PREPARE_FOR_IMCU;
    if (*out_row_ctr >= out_rows_avail)
      return;                   /* Postprocessor exactly filled output buf */
-    /*FALLTHROUGH*/
+    FALLTHROUGH                 /*FALLTHROUGH*/
  case CTX_PREPARE_FOR_IMCU:
    /* Prepare to process first M-1 row groups of this iMCU row */
    main_ptr->rowgroup_ctr = 0;
@@ -371,7 +372,7 @@ process_data_context_main(j_decompress_ptr cinfo, JSAMPARRAY output_buf,
    if (main_ptr->iMCU_row_ctr == cinfo->total_iMCU_rows)
      set_bottom_pointers(cinfo);
    main_ptr->context_state = CTX_PROCESS_IMCU;
-    /*FALLTHROUGH*/
+    FALLTHROUGH                 /*FALLTHROUGH*/
  case CTX_PROCESS_IMCU:
    /* Call postprocessor using previously set pointers */
    (*cinfo->post->post_process_data) (cinfo,
@@ -4,7 +4,7 @@
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1991-1998, Thomas G. Lane.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2012, 2015, D. R. Commander.
+ * Copyright (C) 2012, 2015, 2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -473,7 +473,7 @@ get_dht(j_decompress_ptr cinfo)
    for (i = 0; i < count; i++)
      INPUT_BYTE(cinfo, huffval[i], return FALSE);

-    MEMZERO(&huffval[count], (256 - count) * sizeof(UINT8));
+    memset(&huffval[count], 0, (256 - count) * sizeof(UINT8));

    length -= count;

@@ -491,8 +491,8 @@ get_dht(j_decompress_ptr cinfo)
    if (*htblptr == NULL)
      *htblptr = jpeg_alloc_huff_table((j_common_ptr)cinfo);

-    MEMCOPY((*htblptr)->bits, bits, sizeof((*htblptr)->bits));
-    MEMCOPY((*htblptr)->huffval, huffval, sizeof((*htblptr)->huffval));
+    memcpy((*htblptr)->bits, bits, sizeof((*htblptr)->bits));
+    memcpy((*htblptr)->huffval, huffval, sizeof((*htblptr)->huffval));
  }

  if (length != 0)
@@ -5,7 +5,7 @@
 * Copyright (C) 1991-1997, Thomas G. Lane.
 * Modified 2002-2009 by Guido Vollbeding.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2009-2011, 2016, 2019, D. R. Commander.
+ * Copyright (C) 2009-2011, 2016, 2019, 2022, D. R. Commander.
 * Copyright (C) 2013, Linaro Limited.
 * Copyright (C) 2015, Google, Inc.
 * For conditions of distribution and use, see the accompanying README.ijg
@@ -417,7 +417,7 @@ prepare_range_limit_table(j_decompress_ptr cinfo)
  table += (MAXJSAMPLE + 1);    /* allow negative subscripts of simple table */
  cinfo->sample_range_limit = table;
  /* First segment of "simple" table: limit[x] = 0 for x < 0 */
-  MEMZERO(table - (MAXJSAMPLE + 1), (MAXJSAMPLE + 1) * sizeof(JSAMPLE));
+  memset(table - (MAXJSAMPLE + 1), 0, (MAXJSAMPLE + 1) * sizeof(JSAMPLE));
  /* Main part of "simple" table: limit[x] = x */
  for (i = 0; i <= MAXJSAMPLE; i++)
    table[i] = (JSAMPLE)i;
@@ -426,10 +426,10 @@ prepare_range_limit_table(j_decompress_ptr cinfo)
  for (i = CENTERJSAMPLE; i < 2 * (MAXJSAMPLE + 1); i++)
    table[i] = MAXJSAMPLE;
  /* Second half of post-IDCT table */
-  MEMZERO(table + (2 * (MAXJSAMPLE + 1)),
-          (2 * (MAXJSAMPLE + 1) - CENTERJSAMPLE) * sizeof(JSAMPLE));
-  MEMCOPY(table + (4 * (MAXJSAMPLE + 1) - CENTERJSAMPLE),
-          cinfo->sample_range_limit, CENTERJSAMPLE * sizeof(JSAMPLE));
+  memset(table + (2 * (MAXJSAMPLE + 1)), 0,
+         (2 * (MAXJSAMPLE + 1) - CENTERJSAMPLE) * sizeof(JSAMPLE));
+  memcpy(table + (4 * (MAXJSAMPLE + 1) - CENTERJSAMPLE),
+         cinfo->sample_range_limit, CENTERJSAMPLE * sizeof(JSAMPLE));
 }


@@ -4,7 +4,7 @@
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1995-1997, Thomas G. Lane.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2015-2016, 2018-2021, D. R. Commander.
+ * Copyright (C) 2015-2016, 2018-2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -578,9 +578,9 @@ decode_mcu_AC_refine(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
            if (GET_BITS(1)) {
              if ((*thiscoef & p1) == 0) { /* do nothing if already set it */
                if (*thiscoef >= 0)
-                  *thiscoef += p1;
+                  *thiscoef += (JCOEF)p1;
                else
-                  *thiscoef += m1;
+                  *thiscoef += (JCOEF)m1;
              }
            }
          } else {
@@ -612,9 +612,9 @@ decode_mcu_AC_refine(j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
          if (GET_BITS(1)) {
            if ((*thiscoef & p1) == 0) { /* do nothing if already changed it */
              if (*thiscoef >= 0)
-                *thiscoef += p1;
+                *thiscoef += (JCOEF)p1;
              else
-                *thiscoef += m1;
+                *thiscoef += (JCOEF)m1;
            }
          }
        }
@@ -3,8 +3,8 @@
 *
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1991-1998, Thomas G. Lane.
- * It was modified by The libjpeg-turbo Project to include only code relevant
- * to libjpeg-turbo.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -189,13 +189,13 @@ format_message(j_common_ptr cinfo, char *buffer)

  /* Format the message into the passed buffer */
  if (isstring)
-    sprintf(buffer, msgtext, err->msg_parm.s);
+    snprintf(buffer, JMSG_LENGTH_MAX, msgtext, err->msg_parm.s);
  else
-    sprintf(buffer, msgtext,
-            err->msg_parm.i[0], err->msg_parm.i[1],
-            err->msg_parm.i[2], err->msg_parm.i[3],
-            err->msg_parm.i[4], err->msg_parm.i[5],
-            err->msg_parm.i[6], err->msg_parm.i[7]);
+    snprintf(buffer, JMSG_LENGTH_MAX, msgtext,
+             err->msg_parm.i[0], err->msg_parm.i[1],
+             err->msg_parm.i[2], err->msg_parm.i[3],
+             err->msg_parm.i[4], err->msg_parm.i[5],
+             err->msg_parm.i[6], err->msg_parm.i[7]);
 }


@@ -5,7 +5,7 @@
 * Copyright (C) 1994-1997, Thomas G. Lane.
 * Modified 1997-2009 by Guido Vollbeding.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2014, 2017, D. R. Commander.
+ * Copyright (C) 2014, 2017, 2021-2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -103,7 +103,7 @@ JMESSAGE(JERR_MISMATCHED_QUANT_TABLE,
         "Cannot transcode due to multiple use of quantization table %d")
 JMESSAGE(JERR_MISSING_DATA, "Scan script does not transmit all data")
 JMESSAGE(JERR_MODE_CHANGE, "Invalid color quantization mode change")
-JMESSAGE(JERR_NOTIMPL, "Not implemented yet")
+JMESSAGE(JERR_NOTIMPL, "Requested features are incompatible")
 JMESSAGE(JERR_NOT_COMPILED, "Requested feature was omitted at compile time")
 #if JPEG_LIB_VERSION >= 70
 JMESSAGE(JERR_NO_ARITH_TABLE, "Arithmetic table 0x%02x was not defined")
@@ -268,6 +268,7 @@ JMESSAGE(JERR_BAD_DROP_SAMPLING,
 #define ERREXITS(cinfo, code, str) \
  ((cinfo)->err->msg_code = (code), \
   strncpy((cinfo)->err->msg_parm.s, (str), JMSG_STR_PARM_MAX), \
+   (cinfo)->err->msg_parm.s[JMSG_STR_PARM_MAX - 1] = '\0', \
   (*(cinfo)->err->error_exit) ((j_common_ptr)(cinfo)))

 #define MAKESTMT(stuff)         do { stuff } while (0)
@@ -324,6 +325,7 @@ JMESSAGE(JERR_BAD_DROP_SAMPLING,
 #define TRACEMSS(cinfo, lvl, code, str) \
  ((cinfo)->err->msg_code = (code), \
   strncpy((cinfo)->err->msg_parm.s, (str), JMSG_STR_PARM_MAX), \
+   (cinfo)->err->msg_parm.s[JMSG_STR_PARM_MAX - 1] = '\0', \
   (*(cinfo)->err->emit_message) ((j_common_ptr)(cinfo), (lvl)))

 #endif /* JERROR_H */
@@ -3,8 +3,8 @@
 *
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1991-1994, Thomas G. Lane.
- * It was modified by The libjpeg-turbo Project to include only code relevant
- * to libjpeg-turbo.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -17,72 +17,117 @@
 * JPEG library.  Most applications need only include jpeglib.h.
 */

+#ifndef __JINCLUDE_H__
+#define __JINCLUDE_H__

 /* Include auto-config file to find out which system include files we need. */

 #include "jconfig.h"            /* auto configuration options */
+#include "jconfigint.h"
 #define JCONFIG_INCLUDED        /* so that jpeglib.h doesn't do it again */

 /*
- * We need the NULL macro and size_t typedef.
- * On an ANSI-conforming system it is sufficient to include <stddef.h>.
- * Otherwise, we get them from <stdlib.h> or <stdio.h>; we may have to
- * pull in <sys/types.h> as well.
 * Note that the core JPEG library does not require <stdio.h>;
 * only the default error handler and data source/destination modules do.
 * But we must pull it in because of the references to FILE in jpeglib.h.
 * You can remove those references if you want to compile without <stdio.h>.
 */

-#ifdef HAVE_STDDEF_H
 #include <stddef.h>
-#endif
-
-#ifdef HAVE_STDLIB_H
 #include <stdlib.h>
-#endif
-
-#ifdef NEED_SYS_TYPES_H
-#include <sys/types.h>
-#endif
-
 #include <stdio.h>
-
-/*
- * We need memory copying and zeroing functions, plus strncpy().
- * ANSI and System V implementations declare these in <string.h>.
- * BSD doesn't have the mem() functions, but it does have bcopy()/bzero().
- * Some systems may declare memset and memcpy in <memory.h>.
- *
- * NOTE: we assume the size parameters to these functions are of type size_t.
- * Change the casts in these macros if not!
- */
-
-#ifdef NEED_BSD_STRINGS
-
-#include <strings.h>
-#define MEMZERO(target, size) \
-  bzero((void *)(target), (size_t)(size))
-#define MEMCOPY(dest, src, size) \
-  bcopy((const void *)(src), (void *)(dest), (size_t)(size))
-
-#else /* not BSD, assume ANSI/SysV string lib */
-
 #include <string.h>
-#define MEMZERO(target, size) \
-  memset((void *)(target), 0, (size_t)(size))
-#define MEMCOPY(dest, src, size) \
-  memcpy((void *)(dest), (const void *)(src), (size_t)(size))
-
-#endif

 /*
- * The modules that use fread() and fwrite() always invoke them through
- * these macros.  On some systems you may need to twiddle the argument casts.
- * CAUTION: argument order is different from underlying functions!
+ * These macros/inline functions facilitate using Microsoft's "safe string"
+ * functions with Visual Studio builds without the need to scatter #ifdefs
+ * throughout the code base.
 */

-#define JFREAD(file, buf, sizeofbuf) \
-  ((size_t)fread((void *)(buf), (size_t)1, (size_t)(sizeofbuf), (file)))
-#define JFWRITE(file, buf, sizeofbuf) \
-  ((size_t)fwrite((const void *)(buf), (size_t)1, (size_t)(sizeofbuf), (file)))
+
+#ifndef NO_GETENV
+
+#ifdef _MSC_VER
+
+static INLINE int GETENV_S(char *buffer, size_t buffer_size, const char *name)
+{
+  size_t required_size;
+
+  return (int)getenv_s(&required_size, buffer, buffer_size, name);
+}
+
+#else /* _MSC_VER */
+
+#include <errno.h>
+
+/* This provides a similar interface to the Microsoft/C11 getenv_s() function,
+ * but other than parameter validation, it has no advantages over getenv().
+ */
+
+static INLINE int GETENV_S(char *buffer, size_t buffer_size, const char *name)
+{
+  char *env;
+
+  if (!buffer) {
+    if (buffer_size == 0)
+      return 0;
+    else
+      return (errno = EINVAL);
+  }
+  if (buffer_size == 0)
+    return (errno = EINVAL);
+  if (!name) {
+    *buffer = 0;
+    return 0;
+  }
+
+  env = getenv(name);
+  if (!env)
+  {
+    *buffer = 0;
+    return 0;
+  }
+
+  if (strlen(env) + 1 > buffer_size) {
+    *buffer = 0;
+    return ERANGE;
+  }
+
+  strncpy(buffer, env, buffer_size);
+
+  return 0;
+}
+
+#endif /* _MSC_VER */
+
+#endif /* NO_GETENV */
+
+
+#ifndef NO_PUTENV
+
+#ifdef _WIN32
+
+#define PUTENV_S(name, value)  _putenv_s(name, value)
+
+#else
+
+/* This provides a similar interface to the Microsoft _putenv_s() function, but
+ * other than parameter validation, it has no advantages over setenv().
+ */
+
+static INLINE int PUTENV_S(const char *name, const char *value)
+{
+  if (!name || !value)
+    return (errno = EINVAL);
+
+  setenv(name, value, 1);
+
+  return errno;
+}
+
+#endif /* _WIN32 */
+
+#endif /* NO_PUTENV */
+
+
+#endif /* JINCLUDE_H */
@@ -4,7 +4,7 @@
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1991-1997, Thomas G. Lane.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2016, D. R. Commander.
+ * Copyright (C) 2016, 2021-2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -37,12 +37,6 @@
 #endif
 #include <limits.h>

-#ifndef NO_GETENV
-#ifndef HAVE_STDLIB_H           /* <stdlib.h> should declare getenv() */
-extern char *getenv(const char *name);
-#endif
-#endif
-

 LOCAL(size_t)
 round_up_pow2(size_t a, size_t b)
@@ -1032,7 +1026,7 @@ free_pool(j_common_ptr cinfo, int pool_id)
    large_pool_ptr next_lhdr_ptr = lhdr_ptr->next;
    space_freed = lhdr_ptr->bytes_used +
                  lhdr_ptr->bytes_left +
-                  sizeof(large_pool_hdr);
+                  sizeof(large_pool_hdr) + ALIGN_SIZE - 1;
    jpeg_free_large(cinfo, (void *)lhdr_ptr, space_freed);
    mem->total_space_allocated -= space_freed;
    lhdr_ptr = next_lhdr_ptr;
@@ -1045,7 +1039,7 @@ free_pool(j_common_ptr cinfo, int pool_id)
  while (shdr_ptr != NULL) {
    small_pool_ptr next_shdr_ptr = shdr_ptr->next;
    space_freed = shdr_ptr->bytes_used + shdr_ptr->bytes_left +
-                  sizeof(small_pool_hdr);
+                  sizeof(small_pool_hdr) + ALIGN_SIZE - 1;
    jpeg_free_small(cinfo, (void *)shdr_ptr, space_freed);
    mem->total_space_allocated -= space_freed;
    shdr_ptr = next_shdr_ptr;
@@ -1162,12 +1156,16 @@ jinit_memory_mgr(j_common_ptr cinfo)
   */
 #ifndef NO_GETENV
  {
-    char *memenv;
+    char memenv[30] = { 0 };

-    if ((memenv = getenv("JPEGMEM")) != NULL) {
+    if (!GETENV_S(memenv, 30, "JPEGMEM") && strlen(memenv) > 0) {
      char ch = 'x';

+#ifdef _MSC_VER
+      if (sscanf_s(memenv, "%ld%c", &max_to_use, &ch, 1) > 0) {
+#else
      if (sscanf(memenv, "%ld%c", &max_to_use, &ch) > 0) {
+#endif
        if (ch == 'm' || ch == 'M')
          max_to_use *= 1000L;
        mem->pub.max_memory_to_use = max_to_use * 1000L;
@@ -22,11 +22,6 @@
 #include "jpeglib.h"
 #include "jmemsys.h"            /* import the system-dependent declarations */

-#ifndef HAVE_STDLIB_H           /* <stdlib.h> should declare malloc(),free() */
-extern void *malloc(size_t size);
-extern void free(void *ptr);
-#endif
-

 /*
 * Memory allocation and freeing are controlled by the regular library
@@ -100,11 +100,7 @@ typedef unsigned char UINT8;

 /* UINT16 must hold at least the values 0..65535. */

-#ifdef HAVE_UNSIGNED_SHORT
 typedef unsigned short UINT16;
-#else /* not HAVE_UNSIGNED_SHORT */
-typedef unsigned int UINT16;
-#endif /* HAVE_UNSIGNED_SHORT */

 /* INT16 must hold at least the values -32768..32767. */

@@ -5,8 +5,9 @@
 * Copyright (C) 1991-1997, Thomas G. Lane.
 * Modified 1997-2009 by Guido Vollbeding.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2015-2016, 2019, D. R. Commander.
+ * Copyright (C) 2015-2016, 2019, 2021, D. R. Commander.
 * Copyright (C) 2015, Google, Inc.
+ * Copyright (C) 2021, Alex Richardson.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -47,6 +48,18 @@ typedef enum {            /* Operating modes for buffer controllers */
 /* JLONG must hold at least signed 32-bit values. */
 typedef long JLONG;

+/* JUINTPTR must hold pointer values. */
+#ifdef __UINTPTR_TYPE__
+/*
+ * __UINTPTR_TYPE__ is GNU-specific and available in GCC 4.6+ and Clang 3.0+.
+ * Fortunately, that is sufficient to support the few architectures for which
+ * sizeof(void *) != sizeof(size_t).  The only other options would require C99
+ * or Clang-specific builtins.
+ */
+typedef __UINTPTR_TYPE__ JUINTPTR;
+#else
+typedef size_t JUINTPTR;
+#endif

 /*
 * Left shift macro that handles a negative operand without causing any
@@ -360,12 +373,3 @@ extern const int jpeg_natural_order[]; /* zigzag coef order to natural order */

 /* Arithmetic coding probability estimation tables in jaricom.c */
 extern const JLONG jpeg_aritab[];
-
-/* Suppress undefined-structure complaints if necessary. */
-
-#ifdef INCOMPLETE_TYPES_BROKEN
-#ifndef AM_MEMORY_MANAGER       /* only jmemmgr.c defines these */
-struct jvirt_sarray_control { long dummy; };
-struct jvirt_barray_control { long dummy; };
-#endif
-#endif /* INCOMPLETE_TYPES_BROKEN */
@@ -4,7 +4,7 @@
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1991-1998, Thomas G. Lane.
 * libjpeg-turbo Modifications:
- * Copyright (C) 2013, D. R. Commander.
+ * Copyright (C) 2013, 2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -29,7 +29,7 @@ add_huff_table(j_common_ptr cinfo, JHUFF_TBL **htblptr, const UINT8 *bits,
    return;

  /* Copy the number-of-symbols-of-each-code-length counts */
-  MEMCOPY((*htblptr)->bits, bits, sizeof((*htblptr)->bits));
+  memcpy((*htblptr)->bits, bits, sizeof((*htblptr)->bits));

  /* Validate the counts.  We do this here mainly so we can copy the right
   * number of symbols from the val[] array, without risking marching off
@@ -41,8 +41,9 @@ add_huff_table(j_common_ptr cinfo, JHUFF_TBL **htblptr, const UINT8 *bits,
  if (nsymbols < 1 || nsymbols > 256)
    ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);

-  MEMCOPY((*htblptr)->huffval, val, nsymbols * sizeof(UINT8));
-  MEMZERO(&((*htblptr)->huffval[nsymbols]), (256 - nsymbols) * sizeof(UINT8));
+  memcpy((*htblptr)->huffval, val, nsymbols * sizeof(UINT8));
+  memset(&((*htblptr)->huffval[nsymbols]), 0,
+         (256 - nsymbols) * sizeof(UINT8));

  /* Initialize sent_table FALSE so table will be written to JPEG file. */
  (*htblptr)->sent_table = FALSE;
@@ -3,8 +3,8 @@
 *
 * This file was part of the Independent JPEG Group's software:
 * Copyright (C) 1991-1996, Thomas G. Lane.
- * It was modified by The libjpeg-turbo Project to include only code
- * relevant to libjpeg-turbo.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2022, D. R. Commander.
 * For conditions of distribution and use, see the accompanying README.ijg
 * file.
 *
@@ -110,7 +110,7 @@ jcopy_sample_rows(JSAMPARRAY input_array, int source_row,
  for (row = num_rows; row > 0; row--) {
    inptr = *input_array++;
    outptr = *output_array++;
-    MEMCOPY(outptr, inptr, count);
+    memcpy(outptr, inptr, count);
  }
 }

@@ -120,7 +120,7 @@ jcopy_block_row(JBLOCKROW input_row, JBLOCKROW output_row,
                JDIMENSION num_blocks)
 /* Copy a row of coefficient blocks from one place to another. */
 {
-  MEMCOPY(output_row, input_row, num_blocks * (DCTSIZE2 * sizeof(JCOEF)));
+  memcpy(output_row, input_row, num_blocks * (DCTSIZE2 * sizeof(JCOEF)));
 }


@@ -129,5 +129,5 @@ jzero_far(void *target, size_t bytestozero)
 /* Zero out a chunk of memory. */
 /* This might be sample-array data, block-array data, or alloc_large data. */
 {
-  MEMZERO(target, bytestozero);
+  memset(target, 0, bytestozero);
 }
@@ -0,0 +1,54 @@
+/*
+ * jversion.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-2020, Thomas G. Lane, Guido Vollbeding.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2010, 2012-2022, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ *
+ * This file contains software version identification.
+ */
+
+
+#if JPEG_LIB_VERSION >= 80
+
+#define JVERSION        "8d  15-Jan-2012"
+
+#elif JPEG_LIB_VERSION >= 70
+
+#define JVERSION        "7  27-Jun-2009"
+
+#else
+
+#define JVERSION        "6b  27-Mar-1998"
+
+#endif
+
+/*
+ * NOTE: It is our convention to place the authors in the following order:
+ * - libjpeg-turbo authors (2009-) in descending order of the date of their
+ *   most recent contribution to the project, then in ascending order of the
+ *   date of their first contribution to the project, then in alphabetical
+ *   order
+ * - Upstream authors in descending order of the date of the first inclusion of
+ *   their code
+ */
+
+#define JCOPYRIGHT \
+  "Copyright (C) 2009-2022 D. R. Commander\n" \
+  "Copyright (C) 2015, 2020 Google, Inc.\n" \
+  "Copyright (C) 2019-2020 Arm Limited\n" \
+  "Copyright (C) 2015-2016, 2018 Matthieu Darbois\n" \
+  "Copyright (C) 2011-2016 Siarhei Siamashka\n" \
+  "Copyright (C) 2015 Intel Corporation\n" \
+  "Copyright (C) 2013-2014 Linaro Limited\n" \
+  "Copyright (C) 2013-2014 MIPS Technologies, Inc.\n" \
+  "Copyright (C) 2009, 2012 Pierre Ossman for Cendio AB\n" \
+  "Copyright (C) 2009-2011 Nokia Corporation and/or its subsidiary(-ies)\n" \
+  "Copyright (C) 1999-2006 MIYASAKA Masaru\n" \
+  "Copyright (C) 1991-2020 Thomas G. Lane, Guido Vollbeding"
+
+#define JCOPYRIGHT_SHORT \
+  "Copyright (C) @COPYRIGHT_YEAR@ The libjpeg-turbo Project and many others"
@@ -0,0 +1,545 @@
+macro(simd_fail message)
+    message(STATUS "libjpeg-turbo(SIMD): ${message}.  Performance will suffer.")
+    set(WITH_SIMD 0 PARENT_SCOPE)
+endmacro()
+
+macro(boolean_number var)
+  if(${var})
+    set(${var} 1 ${ARGN})
+  else()
+    set(${var} 0 ${ARGN})
+  endif()
+endmacro()
+
+
+###############################################################################
+# x86[-64] (NASM)
+###############################################################################
+
+if(CPU_TYPE STREQUAL "x86_64" OR CPU_TYPE STREQUAL "i386")
+
+set(CMAKE_ASM_NASM_FLAGS_DEBUG_INIT "-g")
+set(CMAKE_ASM_NASM_FLAGS_RELWITHDEBINFO_INIT "-g")
+
+# Allow the location of the NASM executable to be specified using the ASM_NASM
+# environment variable.  This should happen automatically, but unfortunately
+# enable_language(ASM_NASM) doesn't parse the ASM_NASM environment variable
+# until after CMAKE_ASM_NASM_COMPILER has been populated with the results of
+# searching for NASM or Yasm in the PATH.
+if(NOT DEFINED CMAKE_ASM_NASM_COMPILER AND DEFINED ENV{ASM_NASM})
+  set(CMAKE_ASM_NASM_COMPILER $ENV{ASM_NASM})
+endif()
+
+if(CPU_TYPE STREQUAL "x86_64")
+  if(CYGWIN)
+    set(CMAKE_ASM_NASM_OBJECT_FORMAT win64)
+  endif()
+  if(CMAKE_C_COMPILER_ABI MATCHES "ELF X32")
+    set(CMAKE_ASM_NASM_OBJECT_FORMAT elfx32)
+  endif()
+elseif(CPU_TYPE STREQUAL "i386")
+  if(BORLAND)
+    set(CMAKE_ASM_NASM_OBJECT_FORMAT obj)
+  elseif(CYGWIN)
+    set(CMAKE_ASM_NASM_OBJECT_FORMAT win32)
+  endif()
+endif()
+
+
+include(CheckLanguage)
+check_language(ASM_NASM)
+if(NOT CMAKE_ASM_NASM_COMPILER)
+  simd_fail("SIMD extensions disabled: could not find NASM compiler")
+  return()
+endif()
+
+enable_language(ASM_NASM)
+message(STATUS "CMAKE_ASM_NASM_COMPILER = ${CMAKE_ASM_NASM_COMPILER}")
+
+if(CMAKE_ASM_NASM_OBJECT_FORMAT MATCHES "^macho")
+  set(CMAKE_ASM_NASM_FLAGS "${CMAKE_ASM_NASM_FLAGS} -DMACHO")
+elseif(CMAKE_ASM_NASM_OBJECT_FORMAT MATCHES "^elf")
+  set(CMAKE_ASM_NASM_FLAGS "${CMAKE_ASM_NASM_FLAGS} -DELF")
+  set(CMAKE_ASM_NASM_DEBUG_FORMAT "dwarf2")
+endif()
+if(CPU_TYPE STREQUAL "x86_64")
+  if(WIN32 OR CYGWIN)
+    set(CMAKE_ASM_NASM_FLAGS "${CMAKE_ASM_NASM_FLAGS} -DWIN64")
+  endif()
+  set(CMAKE_ASM_NASM_FLAGS "${CMAKE_ASM_NASM_FLAGS} -D__x86_64__")
+elseif(CPU_TYPE STREQUAL "i386")
+  if(BORLAND)
+    set(CMAKE_ASM_NASM_FLAGS "${CMAKE_ASM_NASM_FLAGS} -DOBJ32")
+  elseif(WIN32 OR CYGWIN)
+    set(CMAKE_ASM_NASM_FLAGS "${CMAKE_ASM_NASM_FLAGS} -DWIN32")
+  endif()
+endif()
+
+if(NOT CMAKE_ASM_NASM_OBJECT_FORMAT)
+  simd_fail("SIMD extensions disabled: could not determine NASM object format")
+  return()
+endif()
+
+get_filename_component(CMAKE_ASM_NASM_COMPILER_TYPE
+  "${CMAKE_ASM_NASM_COMPILER}" NAME_WE)
+if(CMAKE_ASM_NASM_COMPILER_TYPE MATCHES "yasm")
+  foreach(var CMAKE_ASM_NASM_FLAGS_DEBUG CMAKE_ASM_NASM_FLAGS_RELWITHDEBINFO)
+    if(${var} STREQUAL "-g")
+      if(CMAKE_ASM_NASM_DEBUG_FORMAT)
+        set_property(CACHE ${var} PROPERTY VALUE "-g ${CMAKE_ASM_NASM_DEBUG_FORMAT}")
+      else()
+        set_property(CACHE ${var} PROPERTY VALUE "")
+      endif()
+    endif()
+  endforeach()
+endif()
+
+if(NOT WIN32 AND (CMAKE_POSITION_INDEPENDENT_CODE OR ENABLE_SHARED))
+  set(CMAKE_ASM_NASM_FLAGS "${CMAKE_ASM_NASM_FLAGS} -DPIC")
+endif()
+
+string(TOUPPER ${CMAKE_BUILD_TYPE} CMAKE_BUILD_TYPE_UC)
+set(EFFECTIVE_ASM_NASM_FLAGS "${CMAKE_ASM_NASM_FLAGS} ${CMAKE_ASM_NASM_FLAGS_${CMAKE_BUILD_TYPE_UC}}")
+
+set(CMAKE_ASM_NASM_FLAGS "${CMAKE_ASM_NASM_FLAGS} -I\"${CMAKE_CURRENT_SOURCE_DIR}/nasm/\" -I\"${CMAKE_CURRENT_SOURCE_DIR}/${CPU_TYPE}/\"")
+
+set(GREP grep)
+if(CMAKE_SYSTEM_NAME STREQUAL "SunOS")
+  set(GREP ggrep)
+endif()
+add_custom_target(jsimdcfg COMMAND
+  ${CMAKE_C_COMPILER} -E -I${CMAKE_BINARY_DIR} -I${CMAKE_CURRENT_BINARY_DIR}
+    -I${CMAKE_CURRENT_SOURCE_DIR}
+    ${CMAKE_CURRENT_SOURCE_DIR}/nasm/jsimdcfg.inc.h |
+  ${GREP} -E '^[\;%]|^\ %' | sed 's%_cpp_protection_%%' |
+  sed 's@% define@%define@g' >${CMAKE_CURRENT_SOURCE_DIR}/nasm/jsimdcfg.inc)
+set_target_properties(jsimdcfg PROPERTIES FOLDER "3rdparty")
+
+if(CPU_TYPE STREQUAL "x86_64")
+  set(SIMD_SOURCES x86_64/jsimdcpu.asm x86_64/jfdctflt-sse.asm
+    x86_64/jccolor-sse2.asm x86_64/jcgray-sse2.asm x86_64/jchuff-sse2.asm
+    x86_64/jcphuff-sse2.asm x86_64/jcsample-sse2.asm x86_64/jdcolor-sse2.asm
+    x86_64/jdmerge-sse2.asm x86_64/jdsample-sse2.asm x86_64/jfdctfst-sse2.asm
+    x86_64/jfdctint-sse2.asm x86_64/jidctflt-sse2.asm x86_64/jidctfst-sse2.asm
+    x86_64/jidctint-sse2.asm x86_64/jidctred-sse2.asm x86_64/jquantf-sse2.asm
+    x86_64/jquanti-sse2.asm
+    x86_64/jccolor-avx2.asm x86_64/jcgray-avx2.asm x86_64/jcsample-avx2.asm
+    x86_64/jdcolor-avx2.asm x86_64/jdmerge-avx2.asm x86_64/jdsample-avx2.asm
+    x86_64/jfdctint-avx2.asm x86_64/jidctint-avx2.asm x86_64/jquanti-avx2.asm)
+else()
+  set(SIMD_SOURCES i386/jsimdcpu.asm i386/jfdctflt-3dn.asm
+    i386/jidctflt-3dn.asm i386/jquant-3dn.asm
+    i386/jccolor-mmx.asm i386/jcgray-mmx.asm i386/jcsample-mmx.asm
+    i386/jdcolor-mmx.asm i386/jdmerge-mmx.asm i386/jdsample-mmx.asm
+    i386/jfdctfst-mmx.asm i386/jfdctint-mmx.asm i386/jidctfst-mmx.asm
+    i386/jidctint-mmx.asm i386/jidctred-mmx.asm i386/jquant-mmx.asm
+    i386/jfdctflt-sse.asm i386/jidctflt-sse.asm i386/jquant-sse.asm
+    i386/jccolor-sse2.asm i386/jcgray-sse2.asm i386/jchuff-sse2.asm
+    i386/jcphuff-sse2.asm i386/jcsample-sse2.asm i386/jdcolor-sse2.asm
+    i386/jdmerge-sse2.asm i386/jdsample-sse2.asm i386/jfdctfst-sse2.asm
+    i386/jfdctint-sse2.asm i386/jidctflt-sse2.asm i386/jidctfst-sse2.asm
+    i386/jidctint-sse2.asm i386/jidctred-sse2.asm i386/jquantf-sse2.asm
+    i386/jquanti-sse2.asm
+    i386/jccolor-avx2.asm i386/jcgray-avx2.asm i386/jcsample-avx2.asm
+    i386/jdcolor-avx2.asm i386/jdmerge-avx2.asm i386/jdsample-avx2.asm
+    i386/jfdctint-avx2.asm i386/jidctint-avx2.asm i386/jquanti-avx2.asm)
+endif()
+
+if(MSVC_IDE)
+  set(OBJDIR "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_CFG_INTDIR}")
+  string(REGEX REPLACE " " ";" CMAKE_ASM_NASM_FLAGS "${CMAKE_ASM_NASM_FLAGS}")
+elseif(XCODE)
+  set(OBJDIR "${CMAKE_CURRENT_BINARY_DIR}")
+  string(REGEX REPLACE " " ";" CMAKE_ASM_NASM_FLAGS "${CMAKE_ASM_NASM_FLAGS}")
+endif()
+
+file(GLOB INC_FILES nasm/*.inc)
+
+foreach(file ${SIMD_SOURCES})
+  set(OBJECT_DEPENDS "")
+  if(${file} MATCHES jccolor)
+    string(REGEX REPLACE "jccolor" "jccolext" DEPFILE ${file})
+    set(OBJECT_DEPENDS ${OBJECT_DEPENDS}
+      ${CMAKE_CURRENT_SOURCE_DIR}/${DEPFILE})
+  endif()
+  if(${file} MATCHES jcgray)
+    string(REGEX REPLACE "jcgray" "jcgryext" DEPFILE ${file})
+    set(OBJECT_DEPENDS ${OBJECT_DEPENDS}
+      ${CMAKE_CURRENT_SOURCE_DIR}/${DEPFILE})
+  endif()
+  if(${file} MATCHES jdcolor)
+    string(REGEX REPLACE "jdcolor" "jdcolext" DEPFILE ${file})
+    set(OBJECT_DEPENDS ${OBJECT_DEPENDS}
+      ${CMAKE_CURRENT_SOURCE_DIR}/${DEPFILE})
+  endif()
+  if(${file} MATCHES jdmerge)
+    string(REGEX REPLACE "jdmerge" "jdmrgext" DEPFILE ${file})
+    set(OBJECT_DEPENDS ${OBJECT_DEPENDS}
+      ${CMAKE_CURRENT_SOURCE_DIR}/${DEPFILE})
+  endif()
+  set(OBJECT_DEPENDS ${OBJECT_DEPENDS} ${INC_FILES})
+  if(MSVC_IDE OR XCODE)
+    # The CMake Visual Studio generators do not work properly with the ASM_NASM
+    # language, so we have to go rogue here and use a custom command like we
+    # did in prior versions of libjpeg-turbo.  (This is why we can't have nice
+    # things.)
+    string(REGEX REPLACE "${CPU_TYPE}/" "" filename ${file})
+    set(SIMD_OBJ ${OBJDIR}/${filename}${CMAKE_C_OUTPUT_EXTENSION})
+    add_custom_command(OUTPUT ${SIMD_OBJ} DEPENDS ${file} ${OBJECT_DEPENDS}
+      COMMAND ${CMAKE_ASM_NASM_COMPILER} -f${CMAKE_ASM_NASM_OBJECT_FORMAT}
+        ${CMAKE_ASM_NASM_FLAGS} ${CMAKE_CURRENT_SOURCE_DIR}/${file}
+        -o${SIMD_OBJ})
+    set(SIMD_OBJS ${SIMD_OBJS} ${SIMD_OBJ})
+  else()
+    set_source_files_properties(${file} PROPERTIES OBJECT_DEPENDS
+      "${OBJECT_DEPENDS}")
+  endif()
+endforeach()
+
+if(MSVC_IDE OR XCODE)
+  set(SIMD_OBJS ${SIMD_OBJS} PARENT_SCOPE)
+  add_library(jsimd OBJECT ${CPU_TYPE}/jsimd.c)
+  add_custom_target(jsimd-objs DEPENDS ${SIMD_OBJS})
+  add_dependencies(jsimd jsimd-objs)
+  set_target_properties(jsimd PROPERTIES FOLDER "3rdparty")
+  set_target_properties(jsimd-objs PROPERTIES FOLDER "3rdparty")
+else()
+  add_library(jsimd OBJECT ${SIMD_SOURCES} ${CPU_TYPE}/jsimd.c)
+endif()
+if(NOT WIN32 AND (CMAKE_POSITION_INDEPENDENT_CODE OR ENABLE_SHARED))
+  set_target_properties(jsimd PROPERTIES POSITION_INDEPENDENT_CODE 1)
+endif()
+
+
+###############################################################################
+# Arm (Intrinsics or GAS)
+###############################################################################
+
+elseif(CPU_TYPE STREQUAL "arm64" OR CPU_TYPE STREQUAL "arm")
+
+# If Neon instructions are not explicitly enabled at compile time (e.g. using
+# -mfpu=neon) with an AArch32 Linux or Android build, then the AArch32 SIMD
+# dispatcher will parse /proc/cpuinfo to determine whether the Neon SIMD
+# extensions can be enabled at run time.  In order to support all AArch32 CPUs
+# using the same code base, i.e. to support run-time FPU and Neon
+# auto-detection, it is necessary to compile the scalar C source code using
+# -mfloat-abi=soft (which is usually the default) but compile the intrinsics
+# implementation of the Neon SIMD extensions using -mfloat-abi=softfp.  The
+# following test determines whether -mfloat-abi=softfp should be explicitly
+# added to the compile flags for the intrinsics implementation of the Neon SIMD
+# extensions.
+
+if(BITS EQUAL 32)
+  check_c_source_compiles("
+    #if defined(__ARM_NEON__) || (!defined(__linux__) && !defined(ANDROID) && !defined(__ANDROID__))
+    #error \"Neon run-time auto-detection will not be used\"
+    #endif
+    #if __ARM_PCS_VFP == 1
+    #error \"float ABI = hard\"
+    #endif
+    #if __SOFTFP__ != 1
+    #error \"float ABI = softfp\"
+    #endif
+    int main(void) { return 0; }" NEED_SOFTFP_FOR_INTRINSICS)
+  if(NEED_SOFTFP_FOR_INTRINSICS)
+    set(SOFTFP_FLAG -mfloat-abi=softfp)
+  endif()
+endif()
+
+if(BITS EQUAL 32)
+  set(CMAKE_REQUIRED_FLAGS "-mfpu=neon ${SOFTFP_FLAG}")
+  check_c_source_compiles("
+    #include <arm_neon.h>
+    int main(int argc, char **argv) {
+      uint16x8_t input = vdupq_n_u16((uint16_t)argc);
+      uint8x8_t output = vmovn_u16(input);
+      return (int)output[0];
+    }" HAVE_NEON)
+  if(NOT HAVE_NEON)
+    simd_fail("SIMD extensions not available for this architecture")
+    return()
+  endif()
+endif()
+check_c_source_compiles("
+  #include <arm_neon.h>
+  int main(int argc, char **argv) {
+    int16_t input[] = {
+      (int16_t)argc, (int16_t)argc, (int16_t)argc, (int16_t)argc,
+      (int16_t)argc, (int16_t)argc, (int16_t)argc, (int16_t)argc,
+      (int16_t)argc, (int16_t)argc, (int16_t)argc, (int16_t)argc
+    };
+    int16x4x3_t output = vld1_s16_x3(input);
+    vst3_s16(input, output);
+    return (int)input[0];
+  }" HAVE_VLD1_S16_X3)
+check_c_source_compiles("
+  #include <arm_neon.h>
+  int main(int argc, char **argv) {
+    uint16_t input[] = {
+      (uint16_t)argc, (uint16_t)argc, (uint16_t)argc, (uint16_t)argc,
+      (uint16_t)argc, (uint16_t)argc, (uint16_t)argc, (uint16_t)argc
+    };
+    uint16x4x2_t output = vld1_u16_x2(input);
+    vst2_u16(input, output);
+    return (int)input[0];
+  }" HAVE_VLD1_U16_X2)
+check_c_source_compiles("
+  #include <arm_neon.h>
+  int main(int argc, char **argv) {
+    uint8_t input[] = {
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc,
+      (uint8_t)argc, (uint8_t)argc, (uint8_t)argc, (uint8_t)argc
+    };
+    uint8x16x4_t output = vld1q_u8_x4(input);
+    vst4q_u8(input, output);
+    return (int)input[0];
+  }" HAVE_VLD1Q_U8_X4)
+if(BITS EQUAL 32)
+  unset(CMAKE_REQUIRED_FLAGS)
+endif()
+configure_file(arm/neon-compat.h.in arm/neon-compat.h @ONLY)
+include_directories(${CMAKE_CURRENT_BINARY_DIR}/arm)
+
+# GCC 11 and earlier and some older versions of Clang do not have a full or
+# optimal set of Neon intrinsics, so for performance reasons, when using those
+# compilers, we default to using the older GAS implementation of the Neon SIMD
+# extensions for certain algorithms.  The presence or absence of the three
+# intrinsics we tested above is a reasonable proxy for this, except with GCC 10
+# and 11.
+if((HAVE_VLD1_S16_X3 AND HAVE_VLD1_U16_X2 AND HAVE_VLD1Q_U8_X4 AND
+  (NOT CMAKE_COMPILER_IS_GNUCC OR
+  CMAKE_C_COMPILER_VERSION VERSION_EQUAL 12.0.0 OR
+  CMAKE_C_COMPILER_VERSION VERSION_GREATER 12.0.0)))
+  set(DEFAULT_NEON_INTRINSICS 1)
+else()
+  set(DEFAULT_NEON_INTRINSICS 0)
+endif()
+option(NEON_INTRINSICS
+  "Because GCC (as of this writing) and some older versions of Clang do not have a full or optimal set of Neon intrinsics, for performance reasons, the default when building libjpeg-turbo with those compilers is to continue using the older GAS implementation of the Neon SIMD extensions for certain algorithms.  Setting this option forces the full Neon intrinsics implementation to be used with all compilers.  Unsetting this option forces the hybrid GAS/intrinsics implementation to be used with all compilers."
+  ${DEFAULT_NEON_INTRINSICS})
+if(NOT NEON_INTRINSICS)
+  enable_language(ASM)
+
+  set(CMAKE_ASM_FLAGS "${CMAKE_C_FLAGS} ${CMAKE_ASM_FLAGS}")
+
+  # Test whether gas-preprocessor.pl would be needed to build the GAS
+  # implementation of the Neon SIMD extensions.  If so, then automatically
+  # enable the full Neon intrinsics implementation.
+  if(CPU_TYPE STREQUAL "arm")
+    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/gastest.S "
+      .text
+      .fpu neon
+      .arch armv7a
+      .object_arch armv4
+      .arm
+      pld [r0]
+      vmovn.u16 d0, q0")
+  else()
+    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/gastest.S "
+      .text
+      MYVAR .req x0
+      movi v0.16b, #100
+      mov MYVAR, #100
+      .unreq MYVAR")
+  endif()
+  separate_arguments(CMAKE_ASM_FLAGS_SEP UNIX_COMMAND "${CMAKE_ASM_FLAGS}")
+  execute_process(COMMAND ${CMAKE_ASM_COMPILER} ${CMAKE_ASM_FLAGS_SEP}
+      -x assembler-with-cpp -c ${CMAKE_CURRENT_BINARY_DIR}/gastest.S
+    RESULT_VARIABLE RESULT OUTPUT_VARIABLE OUTPUT ERROR_VARIABLE ERROR)
+  if(NOT RESULT EQUAL 0)
+    message(STATUS "libjpeg-turbo(SIMD): GAS appears to be broken.  Using the full Neon SIMD intrinsics implementation.")
+    set(NEON_INTRINSICS 1 CACHE INTERNAL "" FORCE)
+  endif()
+endif()
+boolean_number(NEON_INTRINSICS PARENT_SCOPE)
+if(NEON_INTRINSICS)
+  add_definitions(-DNEON_INTRINSICS)
+  message(STATUS "Use full Neon SIMD intrinsics implementation (NEON_INTRINSICS = ${NEON_INTRINSICS})")
+else()
+  message(STATUS "Use partial Neon SIMD intrinsics implementation (NEON_INTRINSICS = ${NEON_INTRINSICS})")
+endif()
+
+set(SIMD_SOURCES arm/jcgray-neon.c arm/jcphuff-neon.c arm/jcsample-neon.c
+  arm/jdmerge-neon.c arm/jdsample-neon.c arm/jfdctfst-neon.c
+  arm/jidctred-neon.c arm/jquanti-neon.c)
+if(NEON_INTRINSICS)
+  set(SIMD_SOURCES ${SIMD_SOURCES} arm/jccolor-neon.c arm/jidctint-neon.c)
+endif()
+if(NEON_INTRINSICS OR BITS EQUAL 64)
+  set(SIMD_SOURCES ${SIMD_SOURCES} arm/jidctfst-neon.c)
+endif()
+if(NEON_INTRINSICS OR BITS EQUAL 32)
+  set(SIMD_SOURCES ${SIMD_SOURCES} arm/aarch${BITS}/jchuff-neon.c
+    arm/jdcolor-neon.c arm/jfdctint-neon.c)
+endif()
+if(BITS EQUAL 32)
+  set_source_files_properties(${SIMD_SOURCES} COMPILE_FLAGS "-mfpu=neon ${SOFTFP_FLAG}")
+endif()
+if(NOT NEON_INTRINSICS)
+  string(TOUPPER ${CMAKE_BUILD_TYPE} CMAKE_BUILD_TYPE_UC)
+  set(EFFECTIVE_ASM_FLAGS "${CMAKE_ASM_FLAGS} ${CMAKE_ASM_FLAGS_${CMAKE_BUILD_TYPE_UC}}")
+  message(STATUS "CMAKE_ASM_FLAGS = ${EFFECTIVE_ASM_FLAGS}")
+
+  set(SIMD_SOURCES ${SIMD_SOURCES} arm/aarch${BITS}/jsimd_neon.S)
+endif()
+
+add_library(jsimd OBJECT ${SIMD_SOURCES} arm/aarch${BITS}/jsimd.c)
+
+if(CMAKE_POSITION_INDEPENDENT_CODE OR ENABLE_SHARED)
+  set_target_properties(jsimd PROPERTIES POSITION_INDEPENDENT_CODE 1)
+endif()
+
+
+###############################################################################
+# MIPS (GAS)
+###############################################################################
+
+elseif(CPU_TYPE STREQUAL "mips" OR CPU_TYPE STREQUAL "mipsel")
+
+enable_language(ASM)
+
+string(TOUPPER ${CMAKE_BUILD_TYPE} CMAKE_BUILD_TYPE_UC)
+set(EFFECTIVE_ASM_FLAGS "${CMAKE_ASM_FLAGS} ${CMAKE_ASM_FLAGS_${CMAKE_BUILD_TYPE_UC}}")
+message(STATUS "CMAKE_ASM_FLAGS = ${EFFECTIVE_ASM_FLAGS}")
+
+set(CMAKE_REQUIRED_FLAGS -mdspr2)
+
+check_c_source_compiles("
+  #if !(defined(__mips__) && __mips_isa_rev >= 2)
+  #error MIPS DSPr2 is currently only available on MIPS32r2 platforms.
+  #endif
+  int main(void) {
+    int c = 0, a = 0, b = 0;
+    __asm__ __volatile__ (
+      \"precr.qb.ph %[c], %[a], %[b]\"
+      : [c] \"=r\" (c)
+      : [a] \"r\" (a), [b] \"r\" (b)
+    );
+    return c;
+  }" HAVE_DSPR2)
+
+unset(CMAKE_REQUIRED_FLAGS)
+
+if(NOT HAVE_DSPR2)
+  simd_fail("SIMD extensions not available for this CPU")
+  return()
+endif()
+
+add_library(jsimd OBJECT mips/jsimd_dspr2.S mips/jsimd.c)
+
+if(CMAKE_POSITION_INDEPENDENT_CODE OR ENABLE_SHARED)
+  set_target_properties(jsimd PROPERTIES POSITION_INDEPENDENT_CODE 1)
+endif()
+
+###############################################################################
+# MIPS64 (Intrinsics)
+###############################################################################
+
+elseif(CPU_TYPE STREQUAL "loongson" OR CPU_TYPE MATCHES "^mips64")
+
+set(CMAKE_REQUIRED_FLAGS -Wa,-mloongson-mmi,-mloongson-ext)
+
+check_c_source_compiles("
+  int main(void) {
+    int c = 0, a = 0, b = 0;
+    asm (
+      \"paddb %0, %1, %2\"
+      : \"=f\" (c)
+      : \"f\" (a), \"f\" (b)
+    );
+    return c;
+  }" HAVE_MMI)
+
+unset(CMAKE_REQUIRED_FLAGS)
+
+if(NOT HAVE_MMI)
+  simd_fail("SIMD extensions not available for this CPU")
+  return()
+endif()
+
+set(SIMD_SOURCES mips64/jccolor-mmi.c mips64/jcgray-mmi.c mips64/jcsample-mmi.c
+  mips64/jdcolor-mmi.c mips64/jdmerge-mmi.c mips64/jdsample-mmi.c
+  mips64/jfdctfst-mmi.c mips64/jfdctint-mmi.c mips64/jidctfst-mmi.c
+  mips64/jidctint-mmi.c mips64/jquanti-mmi.c)
+
+if(CMAKE_COMPILER_IS_GNUCC)
+  foreach(file ${SIMD_SOURCES})
+    set_property(SOURCE ${file} APPEND_STRING PROPERTY COMPILE_FLAGS
+      " -fno-strict-aliasing")
+  endforeach()
+endif()
+foreach(file ${SIMD_SOURCES})
+  set_property(SOURCE ${file} APPEND_STRING PROPERTY COMPILE_FLAGS
+    " -Wa,-mloongson-mmi,-mloongson-ext")
+endforeach()
+
+add_library(jsimd OBJECT ${SIMD_SOURCES} mips64/jsimd.c)
+
+if(CMAKE_POSITION_INDEPENDENT_CODE OR ENABLE_SHARED)
+  set_target_properties(jsimd PROPERTIES POSITION_INDEPENDENT_CODE 1)
+endif()
+
+###############################################################################
+# PowerPC (Intrinsics)
+###############################################################################
+
+elseif(CPU_TYPE STREQUAL "powerpc")
+
+set(CMAKE_REQUIRED_FLAGS -maltivec)
+
+check_c_source_compiles("
+  #include <altivec.h>
+  int main(void) {
+    __vector int vi = { 0, 0, 0, 0 };
+    int i[4];
+    vec_st(vi, 0, i);
+    return i[0];
+  }" HAVE_ALTIVEC)
+
+unset(CMAKE_REQUIRED_FLAGS)
+
+if(NOT HAVE_ALTIVEC)
+  simd_fail("SIMD extensions not available for this CPU (PowerPC SPE)")
+  return()
+endif()
+
+set(SIMD_SOURCES powerpc/jccolor-altivec.c powerpc/jcgray-altivec.c
+  powerpc/jcsample-altivec.c powerpc/jdcolor-altivec.c
+  powerpc/jdmerge-altivec.c powerpc/jdsample-altivec.c
+  powerpc/jfdctfst-altivec.c powerpc/jfdctint-altivec.c
+  powerpc/jidctfst-altivec.c powerpc/jidctint-altivec.c
+  powerpc/jquanti-altivec.c)
+
+set_source_files_properties(${SIMD_SOURCES} PROPERTIES
+  COMPILE_FLAGS -maltivec)
+
+add_library(jsimd OBJECT ${SIMD_SOURCES} powerpc/jsimd.c)
+
+if(CMAKE_POSITION_INDEPENDENT_CODE OR ENABLE_SHARED)
+  set_target_properties(jsimd PROPERTIES POSITION_INDEPENDENT_CODE 1)
+endif()
+
+
+###############################################################################
+# None
+###############################################################################
+
+else()
+
+simd_fail("SIMD extensions not available for this CPU (${CMAKE_SYSTEM_PROCESSOR})")
+
+endif() # CPU_TYPE
@@ -0,0 +1,148 @@
+/*
+ * jccolext-neon.c - colorspace conversion (32-bit Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* This file is included by jccolor-neon.c */
+
+
+/* RGB -> YCbCr conversion is defined by the following equations:
+ *    Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+ *    Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + 128
+ *    Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B  + 128
+ *
+ * Avoid floating point arithmetic by using shifted integer constants:
+ *    0.29899597 = 19595 * 2^-16
+ *    0.58700561 = 38470 * 2^-16
+ *    0.11399841 =  7471 * 2^-16
+ *    0.16874695 = 11059 * 2^-16
+ *    0.33125305 = 21709 * 2^-16
+ *    0.50000000 = 32768 * 2^-16
+ *    0.41868592 = 27439 * 2^-16
+ *    0.08131409 =  5329 * 2^-16
+ * These constants are defined in jccolor-neon.c
+ *
+ * We add the fixed-point equivalent of 0.5 to Cb and Cr, which effectively
+ * rounds up or down the result via integer truncation.
+ */
+
+void jsimd_rgb_ycc_convert_neon(JDIMENSION image_width, JSAMPARRAY input_buf,
+                                JSAMPIMAGE output_buf, JDIMENSION output_row,
+                                int num_rows)
+{
+  /* Pointer to RGB(X/A) input data */
+  JSAMPROW inptr;
+  /* Pointers to Y, Cb, and Cr output data */
+  JSAMPROW outptr0, outptr1, outptr2;
+  /* Allocate temporary buffer for final (image_width % 8) pixels in row. */
+  ALIGN(16) uint8_t tmp_buf[8 * RGB_PIXELSIZE];
+
+  /* Set up conversion constants. */
+#ifdef HAVE_VLD1_U16_X2
+  const uint16x4x2_t consts = vld1_u16_x2(jsimd_rgb_ycc_neon_consts);
+#else
+  /* GCC does not currently support the intrinsic vld1_<type>_x2(). */
+  const uint16x4_t consts1 = vld1_u16(jsimd_rgb_ycc_neon_consts);
+  const uint16x4_t consts2 = vld1_u16(jsimd_rgb_ycc_neon_consts + 4);
+  const uint16x4x2_t consts = { { consts1, consts2 } };
+#endif
+  const uint32x4_t scaled_128_5 = vdupq_n_u32((128 << 16) + 32767);
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    output_row++;
+
+    int cols_remaining = image_width;
+    for (; cols_remaining > 0; cols_remaining -= 8) {
+
+      /* To prevent buffer overread by the vector load instructions, the last
+       * (image_width % 8) columns of data are first memcopied to a temporary
+       * buffer large enough to accommodate the vector load.
+       */
+      if (cols_remaining < 8) {
+        memcpy(tmp_buf, inptr, cols_remaining * RGB_PIXELSIZE);
+        inptr = tmp_buf;
+      }
+
+#if RGB_PIXELSIZE == 4
+      uint8x8x4_t input_pixels = vld4_u8(inptr);
+#else
+      uint8x8x3_t input_pixels = vld3_u8(inptr);
+#endif
+      uint16x8_t r = vmovl_u8(input_pixels.val[RGB_RED]);
+      uint16x8_t g = vmovl_u8(input_pixels.val[RGB_GREEN]);
+      uint16x8_t b = vmovl_u8(input_pixels.val[RGB_BLUE]);
+
+      /* Compute Y = 0.29900 * R + 0.58700 * G + 0.11400 * B */
+      uint32x4_t y_low = vmull_lane_u16(vget_low_u16(r), consts.val[0], 0);
+      y_low = vmlal_lane_u16(y_low, vget_low_u16(g), consts.val[0], 1);
+      y_low = vmlal_lane_u16(y_low, vget_low_u16(b), consts.val[0], 2);
+      uint32x4_t y_high = vmull_lane_u16(vget_high_u16(r), consts.val[0], 0);
+      y_high = vmlal_lane_u16(y_high, vget_high_u16(g), consts.val[0], 1);
+      y_high = vmlal_lane_u16(y_high, vget_high_u16(b), consts.val[0], 2);
+
+      /* Compute Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + 128 */
+      uint32x4_t cb_low = scaled_128_5;
+      cb_low = vmlsl_lane_u16(cb_low, vget_low_u16(r), consts.val[0], 3);
+      cb_low = vmlsl_lane_u16(cb_low, vget_low_u16(g), consts.val[1], 0);
+      cb_low = vmlal_lane_u16(cb_low, vget_low_u16(b), consts.val[1], 1);
+      uint32x4_t cb_high = scaled_128_5;
+      cb_high = vmlsl_lane_u16(cb_high, vget_high_u16(r), consts.val[0], 3);
+      cb_high = vmlsl_lane_u16(cb_high, vget_high_u16(g), consts.val[1], 0);
+      cb_high = vmlal_lane_u16(cb_high, vget_high_u16(b), consts.val[1], 1);
+
+      /* Compute Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B  + 128 */
+      uint32x4_t cr_low = scaled_128_5;
+      cr_low = vmlal_lane_u16(cr_low, vget_low_u16(r), consts.val[1], 1);
+      cr_low = vmlsl_lane_u16(cr_low, vget_low_u16(g), consts.val[1], 2);
+      cr_low = vmlsl_lane_u16(cr_low, vget_low_u16(b), consts.val[1], 3);
+      uint32x4_t cr_high = scaled_128_5;
+      cr_high = vmlal_lane_u16(cr_high, vget_high_u16(r), consts.val[1], 1);
+      cr_high = vmlsl_lane_u16(cr_high, vget_high_u16(g), consts.val[1], 2);
+      cr_high = vmlsl_lane_u16(cr_high, vget_high_u16(b), consts.val[1], 3);
+
+      /* Descale Y values (rounding right shift) and narrow to 16-bit. */
+      uint16x8_t y_u16 = vcombine_u16(vrshrn_n_u32(y_low, 16),
+                                      vrshrn_n_u32(y_high, 16));
+      /* Descale Cb values (right shift) and narrow to 16-bit. */
+      uint16x8_t cb_u16 = vcombine_u16(vshrn_n_u32(cb_low, 16),
+                                       vshrn_n_u32(cb_high, 16));
+      /* Descale Cr values (right shift) and narrow to 16-bit. */
+      uint16x8_t cr_u16 = vcombine_u16(vshrn_n_u32(cr_low, 16),
+                                       vshrn_n_u32(cr_high, 16));
+      /* Narrow Y, Cb, and Cr values to 8-bit and store to memory.  Buffer
+       * overwrite is permitted up to the next multiple of ALIGN_SIZE bytes.
+       */
+      vst1_u8(outptr0, vmovn_u16(y_u16));
+      vst1_u8(outptr1, vmovn_u16(cb_u16));
+      vst1_u8(outptr2, vmovn_u16(cr_u16));
+
+      /* Increment pointers. */
+      inptr += (8 * RGB_PIXELSIZE);
+      outptr0 += 8;
+      outptr1 += 8;
+      outptr2 += 8;
+    }
+  }
+}
@@ -0,0 +1,334 @@
+/*
+ * jchuff-neon.c - Huffman entropy encoding (32-bit Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ *
+ * NOTE: All referenced figures are from
+ * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994.
+ */
+
+#define JPEG_INTERNALS
+#include "../../../jinclude.h"
+#include "../../../jpeglib.h"
+#include "../../../jsimd.h"
+#include "../../../jdct.h"
+#include "../../../jsimddct.h"
+#include "../../jsimd.h"
+#include "../jchuff.h"
+#include "neon-compat.h"
+
+#include <limits.h>
+
+#include <arm_neon.h>
+
+
+JOCTET *jsimd_huff_encode_one_block_neon(void *state, JOCTET *buffer,
+                                         JCOEFPTR block, int last_dc_val,
+                                         c_derived_tbl *dctbl,
+                                         c_derived_tbl *actbl)
+{
+  uint8_t block_nbits[DCTSIZE2];
+  uint16_t block_diff[DCTSIZE2];
+
+  /* Load rows of coefficients from DCT block in zig-zag order. */
+
+  /* Compute DC coefficient difference value. (F.1.1.5.1) */
+  int16x8_t row0 = vdupq_n_s16(block[0] - last_dc_val);
+  row0 = vld1q_lane_s16(block +  1, row0, 1);
+  row0 = vld1q_lane_s16(block +  8, row0, 2);
+  row0 = vld1q_lane_s16(block + 16, row0, 3);
+  row0 = vld1q_lane_s16(block +  9, row0, 4);
+  row0 = vld1q_lane_s16(block +  2, row0, 5);
+  row0 = vld1q_lane_s16(block +  3, row0, 6);
+  row0 = vld1q_lane_s16(block + 10, row0, 7);
+
+  int16x8_t row1 = vld1q_dup_s16(block + 17);
+  row1 = vld1q_lane_s16(block + 24, row1, 1);
+  row1 = vld1q_lane_s16(block + 32, row1, 2);
+  row1 = vld1q_lane_s16(block + 25, row1, 3);
+  row1 = vld1q_lane_s16(block + 18, row1, 4);
+  row1 = vld1q_lane_s16(block + 11, row1, 5);
+  row1 = vld1q_lane_s16(block +  4, row1, 6);
+  row1 = vld1q_lane_s16(block +  5, row1, 7);
+
+  int16x8_t row2 = vld1q_dup_s16(block + 12);
+  row2 = vld1q_lane_s16(block + 19, row2, 1);
+  row2 = vld1q_lane_s16(block + 26, row2, 2);
+  row2 = vld1q_lane_s16(block + 33, row2, 3);
+  row2 = vld1q_lane_s16(block + 40, row2, 4);
+  row2 = vld1q_lane_s16(block + 48, row2, 5);
+  row2 = vld1q_lane_s16(block + 41, row2, 6);
+  row2 = vld1q_lane_s16(block + 34, row2, 7);
+
+  int16x8_t row3 = vld1q_dup_s16(block + 27);
+  row3 = vld1q_lane_s16(block + 20, row3, 1);
+  row3 = vld1q_lane_s16(block + 13, row3, 2);
+  row3 = vld1q_lane_s16(block +  6, row3, 3);
+  row3 = vld1q_lane_s16(block +  7, row3, 4);
+  row3 = vld1q_lane_s16(block + 14, row3, 5);
+  row3 = vld1q_lane_s16(block + 21, row3, 6);
+  row3 = vld1q_lane_s16(block + 28, row3, 7);
+
+  int16x8_t abs_row0 = vabsq_s16(row0);
+  int16x8_t abs_row1 = vabsq_s16(row1);
+  int16x8_t abs_row2 = vabsq_s16(row2);
+  int16x8_t abs_row3 = vabsq_s16(row3);
+
+  int16x8_t row0_lz = vclzq_s16(abs_row0);
+  int16x8_t row1_lz = vclzq_s16(abs_row1);
+  int16x8_t row2_lz = vclzq_s16(abs_row2);
+  int16x8_t row3_lz = vclzq_s16(abs_row3);
+
+  /* Compute number of bits required to represent each coefficient. */
+  uint8x8_t row0_nbits = vsub_u8(vdup_n_u8(16),
+                                 vmovn_u16(vreinterpretq_u16_s16(row0_lz)));
+  uint8x8_t row1_nbits = vsub_u8(vdup_n_u8(16),
+                                 vmovn_u16(vreinterpretq_u16_s16(row1_lz)));
+  uint8x8_t row2_nbits = vsub_u8(vdup_n_u8(16),
+                                 vmovn_u16(vreinterpretq_u16_s16(row2_lz)));
+  uint8x8_t row3_nbits = vsub_u8(vdup_n_u8(16),
+                                 vmovn_u16(vreinterpretq_u16_s16(row3_lz)));
+
+  vst1_u8(block_nbits + 0 * DCTSIZE, row0_nbits);
+  vst1_u8(block_nbits + 1 * DCTSIZE, row1_nbits);
+  vst1_u8(block_nbits + 2 * DCTSIZE, row2_nbits);
+  vst1_u8(block_nbits + 3 * DCTSIZE, row3_nbits);
+
+  uint16x8_t row0_mask =
+    vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row0, 15)),
+              vnegq_s16(row0_lz));
+  uint16x8_t row1_mask =
+    vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row1, 15)),
+              vnegq_s16(row1_lz));
+  uint16x8_t row2_mask =
+    vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row2, 15)),
+              vnegq_s16(row2_lz));
+  uint16x8_t row3_mask =
+    vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row3, 15)),
+              vnegq_s16(row3_lz));
+
+  uint16x8_t row0_diff = veorq_u16(vreinterpretq_u16_s16(abs_row0), row0_mask);
+  uint16x8_t row1_diff = veorq_u16(vreinterpretq_u16_s16(abs_row1), row1_mask);
+  uint16x8_t row2_diff = veorq_u16(vreinterpretq_u16_s16(abs_row2), row2_mask);
+  uint16x8_t row3_diff = veorq_u16(vreinterpretq_u16_s16(abs_row3), row3_mask);
+
+  /* Store diff values for rows 0, 1, 2, and 3. */
+  vst1q_u16(block_diff + 0 * DCTSIZE, row0_diff);
+  vst1q_u16(block_diff + 1 * DCTSIZE, row1_diff);
+  vst1q_u16(block_diff + 2 * DCTSIZE, row2_diff);
+  vst1q_u16(block_diff + 3 * DCTSIZE, row3_diff);
+
+  /* Load last four rows of coefficients from DCT block in zig-zag order. */
+  int16x8_t row4 = vld1q_dup_s16(block + 35);
+  row4 = vld1q_lane_s16(block + 42, row4, 1);
+  row4 = vld1q_lane_s16(block + 49, row4, 2);
+  row4 = vld1q_lane_s16(block + 56, row4, 3);
+  row4 = vld1q_lane_s16(block + 57, row4, 4);
+  row4 = vld1q_lane_s16(block + 50, row4, 5);
+  row4 = vld1q_lane_s16(block + 43, row4, 6);
+  row4 = vld1q_lane_s16(block + 36, row4, 7);
+
+  int16x8_t row5 = vld1q_dup_s16(block + 29);
+  row5 = vld1q_lane_s16(block + 22, row5, 1);
+  row5 = vld1q_lane_s16(block + 15, row5, 2);
+  row5 = vld1q_lane_s16(block + 23, row5, 3);
+  row5 = vld1q_lane_s16(block + 30, row5, 4);
+  row5 = vld1q_lane_s16(block + 37, row5, 5);
+  row5 = vld1q_lane_s16(block + 44, row5, 6);
+  row5 = vld1q_lane_s16(block + 51, row5, 7);
+
+  int16x8_t row6 = vld1q_dup_s16(block + 58);
+  row6 = vld1q_lane_s16(block + 59, row6, 1);
+  row6 = vld1q_lane_s16(block + 52, row6, 2);
+  row6 = vld1q_lane_s16(block + 45, row6, 3);
+  row6 = vld1q_lane_s16(block + 38, row6, 4);
+  row6 = vld1q_lane_s16(block + 31, row6, 5);
+  row6 = vld1q_lane_s16(block + 39, row6, 6);
+  row6 = vld1q_lane_s16(block + 46, row6, 7);
+
+  int16x8_t row7 = vld1q_dup_s16(block + 53);
+  row7 = vld1q_lane_s16(block + 60, row7, 1);
+  row7 = vld1q_lane_s16(block + 61, row7, 2);
+  row7 = vld1q_lane_s16(block + 54, row7, 3);
+  row7 = vld1q_lane_s16(block + 47, row7, 4);
+  row7 = vld1q_lane_s16(block + 55, row7, 5);
+  row7 = vld1q_lane_s16(block + 62, row7, 6);
+  row7 = vld1q_lane_s16(block + 63, row7, 7);
+
+  int16x8_t abs_row4 = vabsq_s16(row4);
+  int16x8_t abs_row5 = vabsq_s16(row5);
+  int16x8_t abs_row6 = vabsq_s16(row6);
+  int16x8_t abs_row7 = vabsq_s16(row7);
+
+  int16x8_t row4_lz = vclzq_s16(abs_row4);
+  int16x8_t row5_lz = vclzq_s16(abs_row5);
+  int16x8_t row6_lz = vclzq_s16(abs_row6);
+  int16x8_t row7_lz = vclzq_s16(abs_row7);
+
+  /* Compute number of bits required to represent each coefficient. */
+  uint8x8_t row4_nbits = vsub_u8(vdup_n_u8(16),
+                                 vmovn_u16(vreinterpretq_u16_s16(row4_lz)));
+  uint8x8_t row5_nbits = vsub_u8(vdup_n_u8(16),
+                                 vmovn_u16(vreinterpretq_u16_s16(row5_lz)));
+  uint8x8_t row6_nbits = vsub_u8(vdup_n_u8(16),
+                                 vmovn_u16(vreinterpretq_u16_s16(row6_lz)));
+  uint8x8_t row7_nbits = vsub_u8(vdup_n_u8(16),
+                                 vmovn_u16(vreinterpretq_u16_s16(row7_lz)));
+
+  vst1_u8(block_nbits + 4 * DCTSIZE, row4_nbits);
+  vst1_u8(block_nbits + 5 * DCTSIZE, row5_nbits);
+  vst1_u8(block_nbits + 6 * DCTSIZE, row6_nbits);
+  vst1_u8(block_nbits + 7 * DCTSIZE, row7_nbits);
+
+  uint16x8_t row4_mask =
+    vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row4, 15)),
+              vnegq_s16(row4_lz));
+  uint16x8_t row5_mask =
+    vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row5, 15)),
+              vnegq_s16(row5_lz));
+  uint16x8_t row6_mask =
+    vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row6, 15)),
+              vnegq_s16(row6_lz));
+  uint16x8_t row7_mask =
+    vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row7, 15)),
+              vnegq_s16(row7_lz));
+
+  uint16x8_t row4_diff = veorq_u16(vreinterpretq_u16_s16(abs_row4), row4_mask);
+  uint16x8_t row5_diff = veorq_u16(vreinterpretq_u16_s16(abs_row5), row5_mask);
+  uint16x8_t row6_diff = veorq_u16(vreinterpretq_u16_s16(abs_row6), row6_mask);
+  uint16x8_t row7_diff = veorq_u16(vreinterpretq_u16_s16(abs_row7), row7_mask);
+
+  /* Store diff values for rows 4, 5, 6, and 7. */
+  vst1q_u16(block_diff + 4 * DCTSIZE, row4_diff);
+  vst1q_u16(block_diff + 5 * DCTSIZE, row5_diff);
+  vst1q_u16(block_diff + 6 * DCTSIZE, row6_diff);
+  vst1q_u16(block_diff + 7 * DCTSIZE, row7_diff);
+
+  /* Construct bitmap to accelerate encoding of AC coefficients.  A set bit
+   * means that the corresponding coefficient != 0.
+   */
+  uint8x8_t row0_nbits_gt0 = vcgt_u8(row0_nbits, vdup_n_u8(0));
+  uint8x8_t row1_nbits_gt0 = vcgt_u8(row1_nbits, vdup_n_u8(0));
+  uint8x8_t row2_nbits_gt0 = vcgt_u8(row2_nbits, vdup_n_u8(0));
+  uint8x8_t row3_nbits_gt0 = vcgt_u8(row3_nbits, vdup_n_u8(0));
+  uint8x8_t row4_nbits_gt0 = vcgt_u8(row4_nbits, vdup_n_u8(0));
+  uint8x8_t row5_nbits_gt0 = vcgt_u8(row5_nbits, vdup_n_u8(0));
+  uint8x8_t row6_nbits_gt0 = vcgt_u8(row6_nbits, vdup_n_u8(0));
+  uint8x8_t row7_nbits_gt0 = vcgt_u8(row7_nbits, vdup_n_u8(0));
+
+  /* { 0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01 } */
+  const uint8x8_t bitmap_mask =
+    vreinterpret_u8_u64(vmov_n_u64(0x0102040810204080));
+
+  row0_nbits_gt0 = vand_u8(row0_nbits_gt0, bitmap_mask);
+  row1_nbits_gt0 = vand_u8(row1_nbits_gt0, bitmap_mask);
+  row2_nbits_gt0 = vand_u8(row2_nbits_gt0, bitmap_mask);
+  row3_nbits_gt0 = vand_u8(row3_nbits_gt0, bitmap_mask);
+  row4_nbits_gt0 = vand_u8(row4_nbits_gt0, bitmap_mask);
+  row5_nbits_gt0 = vand_u8(row5_nbits_gt0, bitmap_mask);
+  row6_nbits_gt0 = vand_u8(row6_nbits_gt0, bitmap_mask);
+  row7_nbits_gt0 = vand_u8(row7_nbits_gt0, bitmap_mask);
+
+  uint8x8_t bitmap_rows_10 = vpadd_u8(row1_nbits_gt0, row0_nbits_gt0);
+  uint8x8_t bitmap_rows_32 = vpadd_u8(row3_nbits_gt0, row2_nbits_gt0);
+  uint8x8_t bitmap_rows_54 = vpadd_u8(row5_nbits_gt0, row4_nbits_gt0);
+  uint8x8_t bitmap_rows_76 = vpadd_u8(row7_nbits_gt0, row6_nbits_gt0);
+  uint8x8_t bitmap_rows_3210 = vpadd_u8(bitmap_rows_32, bitmap_rows_10);
+  uint8x8_t bitmap_rows_7654 = vpadd_u8(bitmap_rows_76, bitmap_rows_54);
+  uint8x8_t bitmap = vpadd_u8(bitmap_rows_7654, bitmap_rows_3210);
+
+  /* Shift left to remove DC bit. */
+  bitmap = vreinterpret_u8_u64(vshl_n_u64(vreinterpret_u64_u8(bitmap), 1));
+  /* Move bitmap to 32-bit scalar registers. */
+  uint32_t bitmap_1_32 = vget_lane_u32(vreinterpret_u32_u8(bitmap), 1);
+  uint32_t bitmap_33_63 = vget_lane_u32(vreinterpret_u32_u8(bitmap), 0);
+
+  /* Set up state and bit buffer for output bitstream. */
+  working_state *state_ptr = (working_state *)state;
+  int free_bits = state_ptr->cur.free_bits;
+  size_t put_buffer = state_ptr->cur.put_buffer;
+
+  /* Encode DC coefficient. */
+
+  unsigned int nbits = block_nbits[0];
+  /* Emit Huffman-coded symbol and additional diff bits. */
+  unsigned int diff = block_diff[0];
+  PUT_CODE(dctbl->ehufco[nbits], dctbl->ehufsi[nbits], diff)
+
+  /* Encode AC coefficients. */
+
+  unsigned int r = 0;  /* r = run length of zeros */
+  unsigned int i = 1;  /* i = number of coefficients encoded */
+  /* Code and size information for a run length of 16 zero coefficients */
+  const unsigned int code_0xf0 = actbl->ehufco[0xf0];
+  const unsigned int size_0xf0 = actbl->ehufsi[0xf0];
+
+  while (bitmap_1_32 != 0) {
+    r = BUILTIN_CLZ(bitmap_1_32);
+    i += r;
+    bitmap_1_32 <<= r;
+    nbits = block_nbits[i];
+    diff = block_diff[i];
+    while (r > 15) {
+      /* If run length > 15, emit special run-length-16 codes. */
+      PUT_BITS(code_0xf0, size_0xf0)
+      r -= 16;
+    }
+    /* Emit Huffman symbol for run length / number of bits. (F.1.2.2.1) */
+    unsigned int rs = (r << 4) + nbits;
+    PUT_CODE(actbl->ehufco[rs], actbl->ehufsi[rs], diff)
+    i++;
+    bitmap_1_32 <<= 1;
+  }
+
+  r = 33 - i;
+  i = 33;
+
+  while (bitmap_33_63 != 0) {
+    unsigned int leading_zeros = BUILTIN_CLZ(bitmap_33_63);
+    r += leading_zeros;
+    i += leading_zeros;
+    bitmap_33_63 <<= leading_zeros;
+    nbits = block_nbits[i];
+    diff = block_diff[i];
+    while (r > 15) {
+      /* If run length > 15, emit special run-length-16 codes. */
+      PUT_BITS(code_0xf0, size_0xf0)
+      r -= 16;
+    }
+    /* Emit Huffman symbol for run length / number of bits. (F.1.2.2.1) */
+    unsigned int rs = (r << 4) + nbits;
+    PUT_CODE(actbl->ehufco[rs], actbl->ehufsi[rs], diff)
+    r = 0;
+    i++;
+    bitmap_33_63 <<= 1;
+  }
+
+  /* If the last coefficient(s) were zero, emit an end-of-block (EOB) code.
+   * The value of RS for the EOB code is 0.
+   */
+  if (i != 64) {
+    PUT_BITS(actbl->ehufco[0], actbl->ehufsi[0])
+  }
+
+  state_ptr->cur.put_buffer = put_buffer;
+  state_ptr->cur.free_bits = free_bits;
+
+  return buffer;
+}
@@ -0,0 +1,980 @@
+/*
+ * jsimd_arm.c
+ *
+ * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright (C) 2011, Nokia Corporation and/or its subsidiary(-ies).
+ * Copyright (C) 2009-2011, 2013-2014, 2016, 2018, 2022, D. R. Commander.
+ * Copyright (C) 2015-2016, 2018, Matthieu Darbois.
+ * Copyright (C) 2019, Google LLC.
+ * Copyright (C) 2020, Arm Limited.
+ *
+ * Based on the x86 SIMD extension for IJG JPEG library,
+ * Copyright (C) 1999-2006, MIYASAKA Masaru.
+ * For conditions of distribution and use, see copyright notice in jsimdext.inc
+ *
+ * This file contains the interface between the "normal" portions
+ * of the library and the SIMD implementations when running on a
+ * 32-bit Arm architecture.
+ */
+
+#define JPEG_INTERNALS
+#include "../../../jinclude.h"
+#include "../../../jpeglib.h"
+#include "../../../jsimd.h"
+#include "../../../jdct.h"
+#include "../../../jsimddct.h"
+#include "../../jsimd.h"
+
+#include <stdio.h>
+#include <string.h>
+#include <ctype.h>
+
+static unsigned int simd_support = ~0;
+static unsigned int simd_huffman = 1;
+
+#if !defined(__ARM_NEON__) && (defined(__linux__) || defined(ANDROID) || defined(__ANDROID__))
+
+#define SOMEWHAT_SANE_PROC_CPUINFO_SIZE_LIMIT  (1024 * 1024)
+
+LOCAL(int)
+check_feature(char *buffer, char *feature)
+{
+  char *p;
+
+  if (*feature == 0)
+    return 0;
+  if (strncmp(buffer, "Features", 8) != 0)
+    return 0;
+  buffer += 8;
+  while (isspace(*buffer))
+    buffer++;
+
+  /* Check if 'feature' is present in the buffer as a separate word */
+  while ((p = strstr(buffer, feature))) {
+    if (p > buffer && !isspace(*(p - 1))) {
+      buffer++;
+      continue;
+    }
+    p += strlen(feature);
+    if (*p != 0 && !isspace(*p)) {
+      buffer++;
+      continue;
+    }
+    return 1;
+  }
+  return 0;
+}
+
+LOCAL(int)
+parse_proc_cpuinfo(int bufsize)
+{
+  char *buffer = (char *)malloc(bufsize);
+  FILE *fd;
+
+  simd_support = 0;
+
+  if (!buffer)
+    return 0;
+
+  fd = fopen("/proc/cpuinfo", "r");
+  if (fd) {
+    while (fgets(buffer, bufsize, fd)) {
+      if (!strchr(buffer, '\n') && !feof(fd)) {
+        /* "impossible" happened - insufficient size of the buffer! */
+        fclose(fd);
+        free(buffer);
+        return 0;
+      }
+      if (check_feature(buffer, "neon"))
+        simd_support |= JSIMD_NEON;
+    }
+    fclose(fd);
+  }
+  free(buffer);
+  return 1;
+}
+
+#endif
+
+/*
+ * Check what SIMD accelerations are supported.
+ *
+ * FIXME: This code is racy under a multi-threaded environment.
+ */
+LOCAL(void)
+init_simd(void)
+{
+#ifndef NO_GETENV
+  char env[2] = { 0 };
+#endif
+#if !defined(__ARM_NEON__) && (defined(__linux__) || defined(ANDROID) || defined(__ANDROID__))
+  int bufsize = 1024; /* an initial guess for the line buffer size limit */
+#endif
+
+  if (simd_support != ~0U)
+    return;
+
+  simd_support = 0;
+
+#if defined(__ARM_NEON__)
+  simd_support |= JSIMD_NEON;
+#elif defined(__linux__) || defined(ANDROID) || defined(__ANDROID__)
+  /* We still have a chance to use Neon regardless of globally used
+   * -mcpu/-mfpu options passed to gcc by performing runtime detection via
+   * /proc/cpuinfo parsing on linux/android */
+  while (!parse_proc_cpuinfo(bufsize)) {
+    bufsize *= 2;
+    if (bufsize > SOMEWHAT_SANE_PROC_CPUINFO_SIZE_LIMIT)
+      break;
+  }
+#endif
+
+#ifndef NO_GETENV
+  /* Force different settings through environment variables */
+  if (!GETENV_S(env, 2, "JSIMD_FORCENEON") && !strcmp(env, "1"))
+    simd_support = JSIMD_NEON;
+  if (!GETENV_S(env, 2, "JSIMD_FORCENONE") && !strcmp(env, "1"))
+    simd_support = 0;
+  if (!GETENV_S(env, 2, "JSIMD_NOHUFFENC") && !strcmp(env, "1"))
+    simd_huffman = 0;
+#endif
+}
+
+GLOBAL(int)
+jsimd_can_rgb_ycc(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_rgb_gray(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_ycc_rgb(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_ycc_rgb565(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_rgb_ycc_convert(j_compress_ptr cinfo, JSAMPARRAY input_buf,
+                      JSAMPIMAGE output_buf, JDIMENSION output_row,
+                      int num_rows)
+{
+  void (*neonfct) (JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
+
+  switch (cinfo->in_color_space) {
+  case JCS_EXT_RGB:
+    neonfct = jsimd_extrgb_ycc_convert_neon;
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    neonfct = jsimd_extrgbx_ycc_convert_neon;
+    break;
+  case JCS_EXT_BGR:
+    neonfct = jsimd_extbgr_ycc_convert_neon;
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    neonfct = jsimd_extbgrx_ycc_convert_neon;
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    neonfct = jsimd_extxbgr_ycc_convert_neon;
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    neonfct = jsimd_extxrgb_ycc_convert_neon;
+    break;
+  default:
+    neonfct = jsimd_extrgb_ycc_convert_neon;
+    break;
+  }
+
+  neonfct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
+}
+
+GLOBAL(void)
+jsimd_rgb_gray_convert(j_compress_ptr cinfo, JSAMPARRAY input_buf,
+                       JSAMPIMAGE output_buf, JDIMENSION output_row,
+                       int num_rows)
+{
+  void (*neonfct) (JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
+
+  switch (cinfo->in_color_space) {
+  case JCS_EXT_RGB:
+    neonfct = jsimd_extrgb_gray_convert_neon;
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    neonfct = jsimd_extrgbx_gray_convert_neon;
+    break;
+  case JCS_EXT_BGR:
+    neonfct = jsimd_extbgr_gray_convert_neon;
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    neonfct = jsimd_extbgrx_gray_convert_neon;
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    neonfct = jsimd_extxbgr_gray_convert_neon;
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    neonfct = jsimd_extxrgb_gray_convert_neon;
+    break;
+  default:
+    neonfct = jsimd_extrgb_gray_convert_neon;
+    break;
+  }
+
+  neonfct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
+}
+
+GLOBAL(void)
+jsimd_ycc_rgb_convert(j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
+                      JDIMENSION input_row, JSAMPARRAY output_buf,
+                      int num_rows)
+{
+  void (*neonfct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
+
+  switch (cinfo->out_color_space) {
+  case JCS_EXT_RGB:
+    neonfct = jsimd_ycc_extrgb_convert_neon;
+    break;
+  case JCS_EXT_RGBX:
+  case JCS_EXT_RGBA:
+    neonfct = jsimd_ycc_extrgbx_convert_neon;
+    break;
+  case JCS_EXT_BGR:
+    neonfct = jsimd_ycc_extbgr_convert_neon;
+    break;
+  case JCS_EXT_BGRX:
+  case JCS_EXT_BGRA:
+    neonfct = jsimd_ycc_extbgrx_convert_neon;
+    break;
+  case JCS_EXT_XBGR:
+  case JCS_EXT_ABGR:
+    neonfct = jsimd_ycc_extxbgr_convert_neon;
+    break;
+  case JCS_EXT_XRGB:
+  case JCS_EXT_ARGB:
+    neonfct = jsimd_ycc_extxrgb_convert_neon;
+    break;
+  default:
+    neonfct = jsimd_ycc_extrgb_convert_neon;
+    break;
+  }
+
+  neonfct(cinfo->output_width, input_buf, input_row, output_buf, num_rows);
+}
+
+GLOBAL(void)
+jsimd_ycc_rgb565_convert(j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
+                         JDIMENSION input_row, JSAMPARRAY output_buf,
+                         int num_rows)
+{
+  jsimd_ycc_rgb565_convert_neon(cinfo->output_width, input_buf, input_row,
+                                output_buf, num_rows);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_downsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_downsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
+                      JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  jsimd_h2v2_downsample_neon(cinfo->image_width, cinfo->max_v_samp_factor,
+                             compptr->v_samp_factor, compptr->width_in_blocks,
+                             input_data, output_data);
+}
+
+GLOBAL(void)
+jsimd_h2v1_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
+                      JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  jsimd_h2v1_downsample_neon(cinfo->image_width, cinfo->max_v_samp_factor,
+                             compptr->v_samp_factor, compptr->width_in_blocks,
+                             input_data, output_data);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                    JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
+{
+  jsimd_h2v2_upsample_neon(cinfo->max_v_samp_factor, cinfo->output_width,
+                           input_data, output_data_ptr);
+}
+
+GLOBAL(void)
+jsimd_h2v1_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                    JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
+{
+  jsimd_h2v1_upsample_neon(cinfo->max_v_samp_factor, cinfo->output_width,
+                           input_data, output_data_ptr);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_fancy_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_fancy_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h1v2_fancy_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
+{
+  jsimd_h2v2_fancy_upsample_neon(cinfo->max_v_samp_factor,
+                                 compptr->downsampled_width, input_data,
+                                 output_data_ptr);
+}
+
+GLOBAL(void)
+jsimd_h2v1_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
+{
+  jsimd_h2v1_fancy_upsample_neon(cinfo->max_v_samp_factor,
+                                 compptr->downsampled_width, input_data,
+                                 output_data_ptr);
+}
+
+GLOBAL(void)
+jsimd_h1v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
+{
+  jsimd_h1v2_fancy_upsample_neon(cinfo->max_v_samp_factor,
+                                 compptr->downsampled_width, input_data,
+                                 output_data_ptr);
+}
+
+GLOBAL(int)
+jsimd_can_h2v2_merged_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_h2v1_merged_upsample(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_h2v2_merged_upsample(j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
+                           JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf)
+{
+  void (*neonfct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+
+  switch (cinfo->out_color_space) {
+    case JCS_EXT_RGB:
+      neonfct = jsimd_h2v2_extrgb_merged_upsample_neon;
+      break;
+    case JCS_EXT_RGBX:
+    case JCS_EXT_RGBA:
+      neonfct = jsimd_h2v2_extrgbx_merged_upsample_neon;
+      break;
+    case JCS_EXT_BGR:
+      neonfct = jsimd_h2v2_extbgr_merged_upsample_neon;
+      break;
+    case JCS_EXT_BGRX:
+    case JCS_EXT_BGRA:
+      neonfct = jsimd_h2v2_extbgrx_merged_upsample_neon;
+      break;
+    case JCS_EXT_XBGR:
+    case JCS_EXT_ABGR:
+      neonfct = jsimd_h2v2_extxbgr_merged_upsample_neon;
+      break;
+    case JCS_EXT_XRGB:
+    case JCS_EXT_ARGB:
+      neonfct = jsimd_h2v2_extxrgb_merged_upsample_neon;
+      break;
+    default:
+      neonfct = jsimd_h2v2_extrgb_merged_upsample_neon;
+      break;
+  }
+
+  neonfct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
+}
+
+GLOBAL(void)
+jsimd_h2v1_merged_upsample(j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
+                           JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf)
+{
+  void (*neonfct) (JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
+
+  switch (cinfo->out_color_space) {
+    case JCS_EXT_RGB:
+      neonfct = jsimd_h2v1_extrgb_merged_upsample_neon;
+      break;
+    case JCS_EXT_RGBX:
+    case JCS_EXT_RGBA:
+      neonfct = jsimd_h2v1_extrgbx_merged_upsample_neon;
+      break;
+    case JCS_EXT_BGR:
+      neonfct = jsimd_h2v1_extbgr_merged_upsample_neon;
+      break;
+    case JCS_EXT_BGRX:
+    case JCS_EXT_BGRA:
+      neonfct = jsimd_h2v1_extbgrx_merged_upsample_neon;
+      break;
+    case JCS_EXT_XBGR:
+    case JCS_EXT_ABGR:
+      neonfct = jsimd_h2v1_extxbgr_merged_upsample_neon;
+      break;
+    case JCS_EXT_XRGB:
+    case JCS_EXT_ARGB:
+      neonfct = jsimd_h2v1_extxrgb_merged_upsample_neon;
+      break;
+    default:
+      neonfct = jsimd_h2v1_extrgb_merged_upsample_neon;
+      break;
+  }
+
+  neonfct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
+}
+
+GLOBAL(int)
+jsimd_can_convsamp(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_convsamp_float(void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_convsamp(JSAMPARRAY sample_data, JDIMENSION start_col,
+               DCTELEM *workspace)
+{
+  jsimd_convsamp_neon(sample_data, start_col, workspace);
+}
+
+GLOBAL(void)
+jsimd_convsamp_float(JSAMPARRAY sample_data, JDIMENSION start_col,
+                     FAST_FLOAT *workspace)
+{
+}
+
+GLOBAL(int)
+jsimd_can_fdct_islow(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_fdct_ifast(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_fdct_float(void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_fdct_islow(DCTELEM *data)
+{
+  jsimd_fdct_islow_neon(data);
+}
+
+GLOBAL(void)
+jsimd_fdct_ifast(DCTELEM *data)
+{
+  jsimd_fdct_ifast_neon(data);
+}
+
+GLOBAL(void)
+jsimd_fdct_float(FAST_FLOAT *data)
+{
+}
+
+GLOBAL(int)
+jsimd_can_quantize(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (sizeof(DCTELEM) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_quantize_float(void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_quantize(JCOEFPTR coef_block, DCTELEM *divisors, DCTELEM *workspace)
+{
+  jsimd_quantize_neon(coef_block, divisors, workspace);
+}
+
+GLOBAL(void)
+jsimd_quantize_float(JCOEFPTR coef_block, FAST_FLOAT *divisors,
+                     FAST_FLOAT *workspace)
+{
+}
+
+GLOBAL(int)
+jsimd_can_idct_2x2(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(ISLOW_MULT_TYPE) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_idct_4x4(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(ISLOW_MULT_TYPE) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_idct_2x2(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+               JCOEFPTR coef_block, JSAMPARRAY output_buf,
+               JDIMENSION output_col)
+{
+  jsimd_idct_2x2_neon(compptr->dct_table, coef_block, output_buf, output_col);
+}
+
+GLOBAL(void)
+jsimd_idct_4x4(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+               JCOEFPTR coef_block, JSAMPARRAY output_buf,
+               JDIMENSION output_col)
+{
+  jsimd_idct_4x4_neon(compptr->dct_table, coef_block, output_buf, output_col);
+}
+
+GLOBAL(int)
+jsimd_can_idct_islow(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(ISLOW_MULT_TYPE) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_idct_ifast(void)
+{
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if (sizeof(IFAST_MULT_TYPE) != 2)
+    return 0;
+  if (IFAST_SCALE_BITS != 2)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_can_idct_float(void)
+{
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_idct_islow(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  jsimd_idct_islow_neon(compptr->dct_table, coef_block, output_buf,
+                        output_col);
+}
+
+GLOBAL(void)
+jsimd_idct_ifast(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+  jsimd_idct_ifast_neon(compptr->dct_table, coef_block, output_buf,
+                        output_col);
+}
+
+GLOBAL(void)
+jsimd_idct_float(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+                 JCOEFPTR coef_block, JSAMPARRAY output_buf,
+                 JDIMENSION output_col)
+{
+}
+
+GLOBAL(int)
+jsimd_can_huff_encode_one_block(void)
+{
+  init_simd();
+
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_NEON && simd_huffman)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(JOCTET *)
+jsimd_huff_encode_one_block(void *state, JOCTET *buffer, JCOEFPTR block,
+                            int last_dc_val, c_derived_tbl *dctbl,
+                            c_derived_tbl *actbl)
+{
+  return jsimd_huff_encode_one_block_neon(state, buffer, block, last_dc_val,
+                                          dctbl, actbl);
+}
+
+GLOBAL(int)
+jsimd_can_encode_mcu_AC_first_prepare(void)
+{
+  init_simd();
+
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(void)
+jsimd_encode_mcu_AC_first_prepare(const JCOEF *block,
+                                  const int *jpeg_natural_order_start, int Sl,
+                                  int Al, JCOEF *values, size_t *zerobits)
+{
+  jsimd_encode_mcu_AC_first_prepare_neon(block, jpeg_natural_order_start,
+                                         Sl, Al, values, zerobits);
+}
+
+GLOBAL(int)
+jsimd_can_encode_mcu_AC_refine_prepare(void)
+{
+  init_simd();
+
+  if (DCTSIZE != 8)
+    return 0;
+  if (sizeof(JCOEF) != 2)
+    return 0;
+
+  if (simd_support & JSIMD_NEON)
+    return 1;
+
+  return 0;
+}
+
+GLOBAL(int)
+jsimd_encode_mcu_AC_refine_prepare(const JCOEF *block,
+                                   const int *jpeg_natural_order_start, int Sl,
+                                   int Al, JCOEF *absvalues, size_t *bits)
+{
+  return jsimd_encode_mcu_AC_refine_prepare_neon(block,
+                                                 jpeg_natural_order_start, Sl,
+                                                 Al, absvalues, bits);
+}
@@ -0,0 +1,316 @@
+/*
+ * jccolext-neon.c - colorspace conversion (64-bit Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* This file is included by jccolor-neon.c */
+
+
+/* RGB -> YCbCr conversion is defined by the following equations:
+ *    Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+ *    Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + 128
+ *    Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B  + 128
+ *
+ * Avoid floating point arithmetic by using shifted integer constants:
+ *    0.29899597 = 19595 * 2^-16
+ *    0.58700561 = 38470 * 2^-16
+ *    0.11399841 =  7471 * 2^-16
+ *    0.16874695 = 11059 * 2^-16
+ *    0.33125305 = 21709 * 2^-16
+ *    0.50000000 = 32768 * 2^-16
+ *    0.41868592 = 27439 * 2^-16
+ *    0.08131409 =  5329 * 2^-16
+ * These constants are defined in jccolor-neon.c
+ *
+ * We add the fixed-point equivalent of 0.5 to Cb and Cr, which effectively
+ * rounds up or down the result via integer truncation.
+ */
+
+void jsimd_rgb_ycc_convert_neon(JDIMENSION image_width, JSAMPARRAY input_buf,
+                                JSAMPIMAGE output_buf, JDIMENSION output_row,
+                                int num_rows)
+{
+  /* Pointer to RGB(X/A) input data */
+  JSAMPROW inptr;
+  /* Pointers to Y, Cb, and Cr output data */
+  JSAMPROW outptr0, outptr1, outptr2;
+  /* Allocate temporary buffer for final (image_width % 16) pixels in row. */
+  ALIGN(16) uint8_t tmp_buf[16 * RGB_PIXELSIZE];
+
+  /* Set up conversion constants. */
+  const uint16x8_t consts = vld1q_u16(jsimd_rgb_ycc_neon_consts);
+  const uint32x4_t scaled_128_5 = vdupq_n_u32((128 << 16) + 32767);
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    output_row++;
+
+    int cols_remaining = image_width;
+    for (; cols_remaining >= 16; cols_remaining -= 16) {
+
+#if RGB_PIXELSIZE == 4
+      uint8x16x4_t input_pixels = vld4q_u8(inptr);
+#else
+      uint8x16x3_t input_pixels = vld3q_u8(inptr);
+#endif
+      uint16x8_t r_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_RED]));
+      uint16x8_t g_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_GREEN]));
+      uint16x8_t b_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_BLUE]));
+      uint16x8_t r_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_RED]));
+      uint16x8_t g_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_GREEN]));
+      uint16x8_t b_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_BLUE]));
+
+      /* Compute Y = 0.29900 * R + 0.58700 * G + 0.11400 * B */
+      uint32x4_t y_ll = vmull_laneq_u16(vget_low_u16(r_l), consts, 0);
+      y_ll = vmlal_laneq_u16(y_ll, vget_low_u16(g_l), consts, 1);
+      y_ll = vmlal_laneq_u16(y_ll, vget_low_u16(b_l), consts, 2);
+      uint32x4_t y_lh = vmull_laneq_u16(vget_high_u16(r_l), consts, 0);
+      y_lh = vmlal_laneq_u16(y_lh, vget_high_u16(g_l), consts, 1);
+      y_lh = vmlal_laneq_u16(y_lh, vget_high_u16(b_l), consts, 2);
+      uint32x4_t y_hl = vmull_laneq_u16(vget_low_u16(r_h), consts, 0);
+      y_hl = vmlal_laneq_u16(y_hl, vget_low_u16(g_h), consts, 1);
+      y_hl = vmlal_laneq_u16(y_hl, vget_low_u16(b_h), consts, 2);
+      uint32x4_t y_hh = vmull_laneq_u16(vget_high_u16(r_h), consts, 0);
+      y_hh = vmlal_laneq_u16(y_hh, vget_high_u16(g_h), consts, 1);
+      y_hh = vmlal_laneq_u16(y_hh, vget_high_u16(b_h), consts, 2);
+
+      /* Compute Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + 128 */
+      uint32x4_t cb_ll = scaled_128_5;
+      cb_ll = vmlsl_laneq_u16(cb_ll, vget_low_u16(r_l), consts, 3);
+      cb_ll = vmlsl_laneq_u16(cb_ll, vget_low_u16(g_l), consts, 4);
+      cb_ll = vmlal_laneq_u16(cb_ll, vget_low_u16(b_l), consts, 5);
+      uint32x4_t cb_lh = scaled_128_5;
+      cb_lh = vmlsl_laneq_u16(cb_lh, vget_high_u16(r_l), consts, 3);
+      cb_lh = vmlsl_laneq_u16(cb_lh, vget_high_u16(g_l), consts, 4);
+      cb_lh = vmlal_laneq_u16(cb_lh, vget_high_u16(b_l), consts, 5);
+      uint32x4_t cb_hl = scaled_128_5;
+      cb_hl = vmlsl_laneq_u16(cb_hl, vget_low_u16(r_h), consts, 3);
+      cb_hl = vmlsl_laneq_u16(cb_hl, vget_low_u16(g_h), consts, 4);
+      cb_hl = vmlal_laneq_u16(cb_hl, vget_low_u16(b_h), consts, 5);
+      uint32x4_t cb_hh = scaled_128_5;
+      cb_hh = vmlsl_laneq_u16(cb_hh, vget_high_u16(r_h), consts, 3);
+      cb_hh = vmlsl_laneq_u16(cb_hh, vget_high_u16(g_h), consts, 4);
+      cb_hh = vmlal_laneq_u16(cb_hh, vget_high_u16(b_h), consts, 5);
+
+      /* Compute Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B  + 128 */
+      uint32x4_t cr_ll = scaled_128_5;
+      cr_ll = vmlal_laneq_u16(cr_ll, vget_low_u16(r_l), consts, 5);
+      cr_ll = vmlsl_laneq_u16(cr_ll, vget_low_u16(g_l), consts, 6);
+      cr_ll = vmlsl_laneq_u16(cr_ll, vget_low_u16(b_l), consts, 7);
+      uint32x4_t cr_lh = scaled_128_5;
+      cr_lh = vmlal_laneq_u16(cr_lh, vget_high_u16(r_l), consts, 5);
+      cr_lh = vmlsl_laneq_u16(cr_lh, vget_high_u16(g_l), consts, 6);
+      cr_lh = vmlsl_laneq_u16(cr_lh, vget_high_u16(b_l), consts, 7);
+      uint32x4_t cr_hl = scaled_128_5;
+      cr_hl = vmlal_laneq_u16(cr_hl, vget_low_u16(r_h), consts, 5);
+      cr_hl = vmlsl_laneq_u16(cr_hl, vget_low_u16(g_h), consts, 6);
+      cr_hl = vmlsl_laneq_u16(cr_hl, vget_low_u16(b_h), consts, 7);
+      uint32x4_t cr_hh = scaled_128_5;
+      cr_hh = vmlal_laneq_u16(cr_hh, vget_high_u16(r_h), consts, 5);
+      cr_hh = vmlsl_laneq_u16(cr_hh, vget_high_u16(g_h), consts, 6);
+      cr_hh = vmlsl_laneq_u16(cr_hh, vget_high_u16(b_h), consts, 7);
+
+      /* Descale Y values (rounding right shift) and narrow to 16-bit. */
+      uint16x8_t y_l = vcombine_u16(vrshrn_n_u32(y_ll, 16),
+                                    vrshrn_n_u32(y_lh, 16));
+      uint16x8_t y_h = vcombine_u16(vrshrn_n_u32(y_hl, 16),
+                                    vrshrn_n_u32(y_hh, 16));
+      /* Descale Cb values (right shift) and narrow to 16-bit. */
+      uint16x8_t cb_l = vcombine_u16(vshrn_n_u32(cb_ll, 16),
+                                     vshrn_n_u32(cb_lh, 16));
+      uint16x8_t cb_h = vcombine_u16(vshrn_n_u32(cb_hl, 16),
+                                     vshrn_n_u32(cb_hh, 16));
+      /* Descale Cr values (right shift) and narrow to 16-bit. */
+      uint16x8_t cr_l = vcombine_u16(vshrn_n_u32(cr_ll, 16),
+                                     vshrn_n_u32(cr_lh, 16));
+      uint16x8_t cr_h = vcombine_u16(vshrn_n_u32(cr_hl, 16),
+                                     vshrn_n_u32(cr_hh, 16));
+      /* Narrow Y, Cb, and Cr values to 8-bit and store to memory.  Buffer
+       * overwrite is permitted up to the next multiple of ALIGN_SIZE bytes.
+       */
+      vst1q_u8(outptr0, vcombine_u8(vmovn_u16(y_l), vmovn_u16(y_h)));
+      vst1q_u8(outptr1, vcombine_u8(vmovn_u16(cb_l), vmovn_u16(cb_h)));
+      vst1q_u8(outptr2, vcombine_u8(vmovn_u16(cr_l), vmovn_u16(cr_h)));
+
+      /* Increment pointers. */
+      inptr += (16 * RGB_PIXELSIZE);
+      outptr0 += 16;
+      outptr1 += 16;
+      outptr2 += 16;
+    }
+
+    if (cols_remaining > 8) {
+      /* To prevent buffer overread by the vector load instructions, the last
+       * (image_width % 16) columns of data are first memcopied to a temporary
+       * buffer large enough to accommodate the vector load.
+       */
+      memcpy(tmp_buf, inptr, cols_remaining * RGB_PIXELSIZE);
+      inptr = tmp_buf;
+
+#if RGB_PIXELSIZE == 4
+      uint8x16x4_t input_pixels = vld4q_u8(inptr);
+#else
+      uint8x16x3_t input_pixels = vld3q_u8(inptr);
+#endif
+      uint16x8_t r_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_RED]));
+      uint16x8_t g_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_GREEN]));
+      uint16x8_t b_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_BLUE]));
+      uint16x8_t r_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_RED]));
+      uint16x8_t g_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_GREEN]));
+      uint16x8_t b_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_BLUE]));
+
+      /* Compute Y = 0.29900 * R + 0.58700 * G + 0.11400 * B */
+      uint32x4_t y_ll = vmull_laneq_u16(vget_low_u16(r_l), consts, 0);
+      y_ll = vmlal_laneq_u16(y_ll, vget_low_u16(g_l), consts, 1);
+      y_ll = vmlal_laneq_u16(y_ll, vget_low_u16(b_l), consts, 2);
+      uint32x4_t y_lh = vmull_laneq_u16(vget_high_u16(r_l), consts, 0);
+      y_lh = vmlal_laneq_u16(y_lh, vget_high_u16(g_l), consts, 1);
+      y_lh = vmlal_laneq_u16(y_lh, vget_high_u16(b_l), consts, 2);
+      uint32x4_t y_hl = vmull_laneq_u16(vget_low_u16(r_h), consts, 0);
+      y_hl = vmlal_laneq_u16(y_hl, vget_low_u16(g_h), consts, 1);
+      y_hl = vmlal_laneq_u16(y_hl, vget_low_u16(b_h), consts, 2);
+      uint32x4_t y_hh = vmull_laneq_u16(vget_high_u16(r_h), consts, 0);
+      y_hh = vmlal_laneq_u16(y_hh, vget_high_u16(g_h), consts, 1);
+      y_hh = vmlal_laneq_u16(y_hh, vget_high_u16(b_h), consts, 2);
+
+      /* Compute Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + 128 */
+      uint32x4_t cb_ll = scaled_128_5;
+      cb_ll = vmlsl_laneq_u16(cb_ll, vget_low_u16(r_l), consts, 3);
+      cb_ll = vmlsl_laneq_u16(cb_ll, vget_low_u16(g_l), consts, 4);
+      cb_ll = vmlal_laneq_u16(cb_ll, vget_low_u16(b_l), consts, 5);
+      uint32x4_t cb_lh = scaled_128_5;
+      cb_lh = vmlsl_laneq_u16(cb_lh, vget_high_u16(r_l), consts, 3);
+      cb_lh = vmlsl_laneq_u16(cb_lh, vget_high_u16(g_l), consts, 4);
+      cb_lh = vmlal_laneq_u16(cb_lh, vget_high_u16(b_l), consts, 5);
+      uint32x4_t cb_hl = scaled_128_5;
+      cb_hl = vmlsl_laneq_u16(cb_hl, vget_low_u16(r_h), consts, 3);
+      cb_hl = vmlsl_laneq_u16(cb_hl, vget_low_u16(g_h), consts, 4);
+      cb_hl = vmlal_laneq_u16(cb_hl, vget_low_u16(b_h), consts, 5);
+      uint32x4_t cb_hh = scaled_128_5;
+      cb_hh = vmlsl_laneq_u16(cb_hh, vget_high_u16(r_h), consts, 3);
+      cb_hh = vmlsl_laneq_u16(cb_hh, vget_high_u16(g_h), consts, 4);
+      cb_hh = vmlal_laneq_u16(cb_hh, vget_high_u16(b_h), consts, 5);
+
+      /* Compute Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B  + 128 */
+      uint32x4_t cr_ll = scaled_128_5;
+      cr_ll = vmlal_laneq_u16(cr_ll, vget_low_u16(r_l), consts, 5);
+      cr_ll = vmlsl_laneq_u16(cr_ll, vget_low_u16(g_l), consts, 6);
+      cr_ll = vmlsl_laneq_u16(cr_ll, vget_low_u16(b_l), consts, 7);
+      uint32x4_t cr_lh = scaled_128_5;
+      cr_lh = vmlal_laneq_u16(cr_lh, vget_high_u16(r_l), consts, 5);
+      cr_lh = vmlsl_laneq_u16(cr_lh, vget_high_u16(g_l), consts, 6);
+      cr_lh = vmlsl_laneq_u16(cr_lh, vget_high_u16(b_l), consts, 7);
+      uint32x4_t cr_hl = scaled_128_5;
+      cr_hl = vmlal_laneq_u16(cr_hl, vget_low_u16(r_h), consts, 5);
+      cr_hl = vmlsl_laneq_u16(cr_hl, vget_low_u16(g_h), consts, 6);
+      cr_hl = vmlsl_laneq_u16(cr_hl, vget_low_u16(b_h), consts, 7);
+      uint32x4_t cr_hh = scaled_128_5;
+      cr_hh = vmlal_laneq_u16(cr_hh, vget_high_u16(r_h), consts, 5);
+      cr_hh = vmlsl_laneq_u16(cr_hh, vget_high_u16(g_h), consts, 6);
+      cr_hh = vmlsl_laneq_u16(cr_hh, vget_high_u16(b_h), consts, 7);
+
+      /* Descale Y values (rounding right shift) and narrow to 16-bit. */
+      uint16x8_t y_l = vcombine_u16(vrshrn_n_u32(y_ll, 16),
+                                    vrshrn_n_u32(y_lh, 16));
+      uint16x8_t y_h = vcombine_u16(vrshrn_n_u32(y_hl, 16),
+                                    vrshrn_n_u32(y_hh, 16));
+      /* Descale Cb values (right shift) and narrow to 16-bit. */
+      uint16x8_t cb_l = vcombine_u16(vshrn_n_u32(cb_ll, 16),
+                                     vshrn_n_u32(cb_lh, 16));
+      uint16x8_t cb_h = vcombine_u16(vshrn_n_u32(cb_hl, 16),
+                                     vshrn_n_u32(cb_hh, 16));
+      /* Descale Cr values (right shift) and narrow to 16-bit. */
+      uint16x8_t cr_l = vcombine_u16(vshrn_n_u32(cr_ll, 16),
+                                     vshrn_n_u32(cr_lh, 16));
+      uint16x8_t cr_h = vcombine_u16(vshrn_n_u32(cr_hl, 16),
+                                     vshrn_n_u32(cr_hh, 16));
+      /* Narrow Y, Cb, and Cr values to 8-bit and store to memory.  Buffer
+       * overwrite is permitted up to the next multiple of ALIGN_SIZE bytes.
+       */
+      vst1q_u8(outptr0, vcombine_u8(vmovn_u16(y_l), vmovn_u16(y_h)));
+      vst1q_u8(outptr1, vcombine_u8(vmovn_u16(cb_l), vmovn_u16(cb_h)));
+      vst1q_u8(outptr2, vcombine_u8(vmovn_u16(cr_l), vmovn_u16(cr_h)));
+
+    } else if (cols_remaining > 0) {
+      /* To prevent buffer overread by the vector load instructions, the last
+       * (image_width % 8) columns of data are first memcopied to a temporary
+       * buffer large enough to accommodate the vector load.
+       */
+      memcpy(tmp_buf, inptr, cols_remaining * RGB_PIXELSIZE);
+      inptr = tmp_buf;
+
+#if RGB_PIXELSIZE == 4
+      uint8x8x4_t input_pixels = vld4_u8(inptr);
+#else
+      uint8x8x3_t input_pixels = vld3_u8(inptr);
+#endif
+      uint16x8_t r = vmovl_u8(input_pixels.val[RGB_RED]);
+      uint16x8_t g = vmovl_u8(input_pixels.val[RGB_GREEN]);
+      uint16x8_t b = vmovl_u8(input_pixels.val[RGB_BLUE]);
+
+      /* Compute Y = 0.29900 * R + 0.58700 * G + 0.11400 * B */
+      uint32x4_t y_l = vmull_laneq_u16(vget_low_u16(r), consts, 0);
+      y_l = vmlal_laneq_u16(y_l, vget_low_u16(g), consts, 1);
+      y_l = vmlal_laneq_u16(y_l, vget_low_u16(b), consts, 2);
+      uint32x4_t y_h = vmull_laneq_u16(vget_high_u16(r), consts, 0);
+      y_h = vmlal_laneq_u16(y_h, vget_high_u16(g), consts, 1);
+      y_h = vmlal_laneq_u16(y_h, vget_high_u16(b), consts, 2);
+
+      /* Compute Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + 128 */
+      uint32x4_t cb_l = scaled_128_5;
+      cb_l = vmlsl_laneq_u16(cb_l, vget_low_u16(r), consts, 3);
+      cb_l = vmlsl_laneq_u16(cb_l, vget_low_u16(g), consts, 4);
+      cb_l = vmlal_laneq_u16(cb_l, vget_low_u16(b), consts, 5);
+      uint32x4_t cb_h = scaled_128_5;
+      cb_h = vmlsl_laneq_u16(cb_h, vget_high_u16(r), consts, 3);
+      cb_h = vmlsl_laneq_u16(cb_h, vget_high_u16(g), consts, 4);
+      cb_h = vmlal_laneq_u16(cb_h, vget_high_u16(b), consts, 5);
+
+      /* Compute Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B  + 128 */
+      uint32x4_t cr_l = scaled_128_5;
+      cr_l = vmlal_laneq_u16(cr_l, vget_low_u16(r), consts, 5);
+      cr_l = vmlsl_laneq_u16(cr_l, vget_low_u16(g), consts, 6);
+      cr_l = vmlsl_laneq_u16(cr_l, vget_low_u16(b), consts, 7);
+      uint32x4_t cr_h = scaled_128_5;
+      cr_h = vmlal_laneq_u16(cr_h, vget_high_u16(r), consts, 5);
+      cr_h = vmlsl_laneq_u16(cr_h, vget_high_u16(g), consts, 6);
+      cr_h = vmlsl_laneq_u16(cr_h, vget_high_u16(b), consts, 7);
+
+      /* Descale Y values (rounding right shift) and narrow to 16-bit. */
+      uint16x8_t y_u16 = vcombine_u16(vrshrn_n_u32(y_l, 16),
+                                      vrshrn_n_u32(y_h, 16));
+      /* Descale Cb values (right shift) and narrow to 16-bit. */
+      uint16x8_t cb_u16 = vcombine_u16(vshrn_n_u32(cb_l, 16),
+                                       vshrn_n_u32(cb_h, 16));
+      /* Descale Cr values (right shift) and narrow to 16-bit. */
+      uint16x8_t cr_u16 = vcombine_u16(vshrn_n_u32(cr_l, 16),
+                                       vshrn_n_u32(cr_h, 16));
+      /* Narrow Y, Cb, and Cr values to 8-bit and store to memory.  Buffer
+       * overwrite is permitted up to the next multiple of ALIGN_SIZE bytes.
+       */
+      vst1_u8(outptr0, vmovn_u16(y_u16));
+      vst1_u8(outptr1, vmovn_u16(cb_u16));
+      vst1_u8(outptr2, vmovn_u16(cr_u16));
+    }
+  }
+}
@@ -0,0 +1,411 @@
+/*
+ * jchuff-neon.c - Huffman entropy encoding (64-bit Arm Neon)
+ *
+ * Copyright (C) 2020-2021, Arm Limited.  All Rights Reserved.
+ * Copyright (C) 2020, 2022, D. R. Commander.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ *
+ * NOTE: All referenced figures are from
+ * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994.
+ */
+
+#define JPEG_INTERNALS
+#include "../../../jinclude.h"
+#include "../../../jpeglib.h"
+#include "../../../jsimd.h"
+#include "../../../jdct.h"
+#include "../../../jsimddct.h"
+#include "../../jsimd.h"
+#include "../align.h"
+#include "../jchuff.h"
+#include "neon-compat.h"
+
+#include <limits.h>
+
+#include <arm_neon.h>
+
+
+ALIGN(16) static const uint8_t jsimd_huff_encode_one_block_consts[] = {
+    0,   1,   2,   3,  16,  17,  32,  33,
+   18,  19,   4,   5,   6,   7,  20,  21,
+   34,  35,  48,  49, 255, 255,  50,  51,
+   36,  37,  22,  23,   8,   9,  10,  11,
+  255, 255,   6,   7,  20,  21,  34,  35,
+   48,  49, 255, 255,  50,  51,  36,  37,
+   54,  55,  40,  41,  26,  27,  12,  13,
+   14,  15,  28,  29,  42,  43,  56,  57,
+    6,   7,  20,  21,  34,  35,  48,  49,
+   50,  51,  36,  37,  22,  23,   8,   9,
+   26,  27,  12,  13, 255, 255,  14,  15,
+   28,  29,  42,  43,  56,  57, 255, 255,
+   52,  53,  54,  55,  40,  41,  26,  27,
+   12,  13, 255, 255,  14,  15,  28,  29,
+   26,  27,  40,  41,  42,  43,  28,  29,
+   14,  15,  30,  31,  44,  45,  46,  47
+};
+
+/* The AArch64 implementation of the FLUSH() macro triggers a UBSan misaligned
+ * address warning because the macro sometimes writes a 64-bit value to a
+ * non-64-bit-aligned address.  That behavior is technically undefined per
+ * the C specification, but it is supported by the AArch64 architecture and
+ * compilers.
+ */
+#if defined(__has_feature)
+#if __has_feature(undefined_behavior_sanitizer)
+__attribute__((no_sanitize("alignment")))
+#endif
+#endif
+JOCTET *jsimd_huff_encode_one_block_neon(void *state, JOCTET *buffer,
+                                         JCOEFPTR block, int last_dc_val,
+                                         c_derived_tbl *dctbl,
+                                         c_derived_tbl *actbl)
+{
+  uint16_t block_diff[DCTSIZE2];
+
+  /* Load lookup table indices for rows of zig-zag ordering. */
+#ifdef HAVE_VLD1Q_U8_X4
+  const uint8x16x4_t idx_rows_0123 =
+    vld1q_u8_x4(jsimd_huff_encode_one_block_consts + 0 * DCTSIZE);
+  const uint8x16x4_t idx_rows_4567 =
+    vld1q_u8_x4(jsimd_huff_encode_one_block_consts + 8 * DCTSIZE);
+#else
+  /* GCC does not currently support intrinsics vl1dq_<type>_x4(). */
+  const uint8x16x4_t idx_rows_0123 = { {
+    vld1q_u8(jsimd_huff_encode_one_block_consts + 0 * DCTSIZE),
+    vld1q_u8(jsimd_huff_encode_one_block_consts + 2 * DCTSIZE),
+    vld1q_u8(jsimd_huff_encode_one_block_consts + 4 * DCTSIZE),
+    vld1q_u8(jsimd_huff_encode_one_block_consts + 6 * DCTSIZE)
+  } };
+  const uint8x16x4_t idx_rows_4567 = { {
+    vld1q_u8(jsimd_huff_encode_one_block_consts + 8 * DCTSIZE),
+    vld1q_u8(jsimd_huff_encode_one_block_consts + 10 * DCTSIZE),
+    vld1q_u8(jsimd_huff_encode_one_block_consts + 12 * DCTSIZE),
+    vld1q_u8(jsimd_huff_encode_one_block_consts + 14 * DCTSIZE)
+  } };
+#endif
+
+  /* Load 8x8 block of DCT coefficients. */
+#ifdef HAVE_VLD1Q_U8_X4
+  const int8x16x4_t tbl_rows_0123 =
+    vld1q_s8_x4((int8_t *)(block + 0 * DCTSIZE));
+  const int8x16x4_t tbl_rows_4567 =
+    vld1q_s8_x4((int8_t *)(block + 4 * DCTSIZE));
+#else
+  const int8x16x4_t tbl_rows_0123 = { {
+    vld1q_s8((int8_t *)(block + 0 * DCTSIZE)),
+    vld1q_s8((int8_t *)(block + 1 * DCTSIZE)),
+    vld1q_s8((int8_t *)(block + 2 * DCTSIZE)),
+    vld1q_s8((int8_t *)(block + 3 * DCTSIZE))
+  } };
+  const int8x16x4_t tbl_rows_4567 = { {
+    vld1q_s8((int8_t *)(block + 4 * DCTSIZE)),
+    vld1q_s8((int8_t *)(block + 5 * DCTSIZE)),
+    vld1q_s8((int8_t *)(block + 6 * DCTSIZE)),
+    vld1q_s8((int8_t *)(block + 7 * DCTSIZE))
+  } };
+#endif
+
+  /* Initialise extra lookup tables. */
+  const int8x16x4_t tbl_rows_2345 = { {
+    tbl_rows_0123.val[2], tbl_rows_0123.val[3],
+    tbl_rows_4567.val[0], tbl_rows_4567.val[1]
+  } };
+  const int8x16x3_t tbl_rows_567 =
+    { { tbl_rows_4567.val[1], tbl_rows_4567.val[2], tbl_rows_4567.val[3] } };
+
+  /* Shuffle coefficients into zig-zag order. */
+  int16x8_t row0 =
+    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_0123, idx_rows_0123.val[0]));
+  int16x8_t row1 =
+    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_0123, idx_rows_0123.val[1]));
+  int16x8_t row2 =
+    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_2345, idx_rows_0123.val[2]));
+  int16x8_t row3 =
+    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_0123, idx_rows_0123.val[3]));
+  int16x8_t row4 =
+    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_4567, idx_rows_4567.val[0]));
+  int16x8_t row5 =
+    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_2345, idx_rows_4567.val[1]));
+  int16x8_t row6 =
+    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_4567, idx_rows_4567.val[2]));
+  int16x8_t row7 =
+    vreinterpretq_s16_s8(vqtbl3q_s8(tbl_rows_567, idx_rows_4567.val[3]));
+
+  /* Compute DC coefficient difference value (F.1.1.5.1). */
+  row0 = vsetq_lane_s16(block[0] - last_dc_val, row0, 0);
+  /* Initialize AC coefficient lanes not reachable by lookup tables. */
+  row1 =
+    vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_4567.val[0]),
+                                  0), row1, 2);
+  row2 =
+    vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_0123.val[1]),
+                                  4), row2, 0);
+  row2 =
+    vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_4567.val[2]),
+                                  0), row2, 5);
+  row5 =
+    vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_0123.val[1]),
+                                  7), row5, 2);
+  row5 =
+    vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_4567.val[2]),
+                                  3), row5, 7);
+  row6 =
+    vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_0123.val[3]),
+                                  7), row6, 5);
+
+  /* DCT block is now in zig-zag order; start Huffman encoding process. */
+
+  /* Construct bitmap to accelerate encoding of AC coefficients.  A set bit
+   * means that the corresponding coefficient != 0.
+   */
+  uint16x8_t row0_ne_0 = vtstq_s16(row0, row0);
+  uint16x8_t row1_ne_0 = vtstq_s16(row1, row1);
+  uint16x8_t row2_ne_0 = vtstq_s16(row2, row2);
+  uint16x8_t row3_ne_0 = vtstq_s16(row3, row3);
+  uint16x8_t row4_ne_0 = vtstq_s16(row4, row4);
+  uint16x8_t row5_ne_0 = vtstq_s16(row5, row5);
+  uint16x8_t row6_ne_0 = vtstq_s16(row6, row6);
+  uint16x8_t row7_ne_0 = vtstq_s16(row7, row7);
+
+  uint8x16_t row10_ne_0 = vuzp1q_u8(vreinterpretq_u8_u16(row1_ne_0),
+                                    vreinterpretq_u8_u16(row0_ne_0));
+  uint8x16_t row32_ne_0 = vuzp1q_u8(vreinterpretq_u8_u16(row3_ne_0),
+                                    vreinterpretq_u8_u16(row2_ne_0));
+  uint8x16_t row54_ne_0 = vuzp1q_u8(vreinterpretq_u8_u16(row5_ne_0),
+                                    vreinterpretq_u8_u16(row4_ne_0));
+  uint8x16_t row76_ne_0 = vuzp1q_u8(vreinterpretq_u8_u16(row7_ne_0),
+                                    vreinterpretq_u8_u16(row6_ne_0));
+
+  /* { 0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01 } */
+  const uint8x16_t bitmap_mask =
+    vreinterpretq_u8_u64(vdupq_n_u64(0x0102040810204080));
+
+  uint8x16_t bitmap_rows_10 = vandq_u8(row10_ne_0, bitmap_mask);
+  uint8x16_t bitmap_rows_32 = vandq_u8(row32_ne_0, bitmap_mask);
+  uint8x16_t bitmap_rows_54 = vandq_u8(row54_ne_0, bitmap_mask);
+  uint8x16_t bitmap_rows_76 = vandq_u8(row76_ne_0, bitmap_mask);
+
+  uint8x16_t bitmap_rows_3210 = vpaddq_u8(bitmap_rows_32, bitmap_rows_10);
+  uint8x16_t bitmap_rows_7654 = vpaddq_u8(bitmap_rows_76, bitmap_rows_54);
+  uint8x16_t bitmap_rows_76543210 = vpaddq_u8(bitmap_rows_7654,
+                                              bitmap_rows_3210);
+  uint8x8_t bitmap_all = vpadd_u8(vget_low_u8(bitmap_rows_76543210),
+                                  vget_high_u8(bitmap_rows_76543210));
+
+  /* Shift left to remove DC bit. */
+  bitmap_all =
+    vreinterpret_u8_u64(vshl_n_u64(vreinterpret_u64_u8(bitmap_all), 1));
+  /* Count bits set (number of non-zero coefficients) in bitmap. */
+  unsigned int non_zero_coefficients = vaddv_u8(vcnt_u8(bitmap_all));
+  /* Move bitmap to 64-bit scalar register. */
+  uint64_t bitmap = vget_lane_u64(vreinterpret_u64_u8(bitmap_all), 0);
+
+  /* Set up state and bit buffer for output bitstream. */
+  working_state *state_ptr = (working_state *)state;
+  int free_bits = state_ptr->cur.free_bits;
+  size_t put_buffer = state_ptr->cur.put_buffer;
+
+  /* Encode DC coefficient. */
+
+  /* For negative coeffs: diff = abs(coeff) -1 = ~abs(coeff) */
+  int16x8_t abs_row0 = vabsq_s16(row0);
+  int16x8_t row0_lz = vclzq_s16(abs_row0);
+  uint16x8_t row0_mask = vshlq_u16(vcltzq_s16(row0), vnegq_s16(row0_lz));
+  uint16x8_t row0_diff = veorq_u16(vreinterpretq_u16_s16(abs_row0), row0_mask);
+  /* Find nbits required to specify sign and amplitude of coefficient. */
+  unsigned int lz = vgetq_lane_u16(vreinterpretq_u16_s16(row0_lz), 0);
+  unsigned int nbits = 16 - lz;
+  /* Emit Huffman-coded symbol and additional diff bits. */
+  unsigned int diff = vgetq_lane_u16(row0_diff, 0);
+  PUT_CODE(dctbl->ehufco[nbits], dctbl->ehufsi[nbits], diff)
+
+  /* Encode AC coefficients. */
+
+  unsigned int r = 0;  /* r = run length of zeros */
+  unsigned int i = 1;  /* i = number of coefficients encoded */
+  /* Code and size information for a run length of 16 zero coefficients */
+  const unsigned int code_0xf0 = actbl->ehufco[0xf0];
+  const unsigned int size_0xf0 = actbl->ehufsi[0xf0];
+
+  /* The most efficient method of computing nbits and diff depends on the
+   * number of non-zero coefficients.  If the bitmap is not too sparse (> 8
+   * non-zero AC coefficients), it is beneficial to do all of the work using
+   * Neon; else we do some of the work using Neon and the rest on demand using
+   * scalar code.
+   */
+  if (non_zero_coefficients > 8) {
+    uint8_t block_nbits[DCTSIZE2];
+
+    int16x8_t abs_row1 = vabsq_s16(row1);
+    int16x8_t abs_row2 = vabsq_s16(row2);
+    int16x8_t abs_row3 = vabsq_s16(row3);
+    int16x8_t abs_row4 = vabsq_s16(row4);
+    int16x8_t abs_row5 = vabsq_s16(row5);
+    int16x8_t abs_row6 = vabsq_s16(row6);
+    int16x8_t abs_row7 = vabsq_s16(row7);
+    int16x8_t row1_lz = vclzq_s16(abs_row1);
+    int16x8_t row2_lz = vclzq_s16(abs_row2);
+    int16x8_t row3_lz = vclzq_s16(abs_row3);
+    int16x8_t row4_lz = vclzq_s16(abs_row4);
+    int16x8_t row5_lz = vclzq_s16(abs_row5);
+    int16x8_t row6_lz = vclzq_s16(abs_row6);
+    int16x8_t row7_lz = vclzq_s16(abs_row7);
+    /* Narrow leading zero count to 8 bits. */
+    uint8x16_t row01_lz = vuzp1q_u8(vreinterpretq_u8_s16(row0_lz),
+                                    vreinterpretq_u8_s16(row1_lz));
+    uint8x16_t row23_lz = vuzp1q_u8(vreinterpretq_u8_s16(row2_lz),
+                                    vreinterpretq_u8_s16(row3_lz));
+    uint8x16_t row45_lz = vuzp1q_u8(vreinterpretq_u8_s16(row4_lz),
+                                    vreinterpretq_u8_s16(row5_lz));
+    uint8x16_t row67_lz = vuzp1q_u8(vreinterpretq_u8_s16(row6_lz),
+                                    vreinterpretq_u8_s16(row7_lz));
+    /* Compute nbits needed to specify magnitude of each coefficient. */
+    uint8x16_t row01_nbits = vsubq_u8(vdupq_n_u8(16), row01_lz);
+    uint8x16_t row23_nbits = vsubq_u8(vdupq_n_u8(16), row23_lz);
+    uint8x16_t row45_nbits = vsubq_u8(vdupq_n_u8(16), row45_lz);
+    uint8x16_t row67_nbits = vsubq_u8(vdupq_n_u8(16), row67_lz);
+    /* Store nbits. */
+    vst1q_u8(block_nbits + 0 * DCTSIZE, row01_nbits);
+    vst1q_u8(block_nbits + 2 * DCTSIZE, row23_nbits);
+    vst1q_u8(block_nbits + 4 * DCTSIZE, row45_nbits);
+    vst1q_u8(block_nbits + 6 * DCTSIZE, row67_nbits);
+    /* Mask bits not required to specify sign and amplitude of diff. */
+    uint16x8_t row1_mask = vshlq_u16(vcltzq_s16(row1), vnegq_s16(row1_lz));
+    uint16x8_t row2_mask = vshlq_u16(vcltzq_s16(row2), vnegq_s16(row2_lz));
+    uint16x8_t row3_mask = vshlq_u16(vcltzq_s16(row3), vnegq_s16(row3_lz));
+    uint16x8_t row4_mask = vshlq_u16(vcltzq_s16(row4), vnegq_s16(row4_lz));
+    uint16x8_t row5_mask = vshlq_u16(vcltzq_s16(row5), vnegq_s16(row5_lz));
+    uint16x8_t row6_mask = vshlq_u16(vcltzq_s16(row6), vnegq_s16(row6_lz));
+    uint16x8_t row7_mask = vshlq_u16(vcltzq_s16(row7), vnegq_s16(row7_lz));
+    /* diff = abs(coeff) ^ sign(coeff) [no-op for positive coefficients] */
+    uint16x8_t row1_diff = veorq_u16(vreinterpretq_u16_s16(abs_row1),
+                                     row1_mask);
+    uint16x8_t row2_diff = veorq_u16(vreinterpretq_u16_s16(abs_row2),
+                                     row2_mask);
+    uint16x8_t row3_diff = veorq_u16(vreinterpretq_u16_s16(abs_row3),
+                                     row3_mask);
+    uint16x8_t row4_diff = veorq_u16(vreinterpretq_u16_s16(abs_row4),
+                                     row4_mask);
+    uint16x8_t row5_diff = veorq_u16(vreinterpretq_u16_s16(abs_row5),
+                                     row5_mask);
+    uint16x8_t row6_diff = veorq_u16(vreinterpretq_u16_s16(abs_row6),
+                                     row6_mask);
+    uint16x8_t row7_diff = veorq_u16(vreinterpretq_u16_s16(abs_row7),
+                                     row7_mask);
+    /* Store diff bits. */
+    vst1q_u16(block_diff + 0 * DCTSIZE, row0_diff);
+    vst1q_u16(block_diff + 1 * DCTSIZE, row1_diff);
+    vst1q_u16(block_diff + 2 * DCTSIZE, row2_diff);
+    vst1q_u16(block_diff + 3 * DCTSIZE, row3_diff);
+    vst1q_u16(block_diff + 4 * DCTSIZE, row4_diff);
+    vst1q_u16(block_diff + 5 * DCTSIZE, row5_diff);
+    vst1q_u16(block_diff + 6 * DCTSIZE, row6_diff);
+    vst1q_u16(block_diff + 7 * DCTSIZE, row7_diff);
+
+    while (bitmap != 0) {
+      r = BUILTIN_CLZLL(bitmap);
+      i += r;
+      bitmap <<= r;
+      nbits = block_nbits[i];
+      diff = block_diff[i];
+      while (r > 15) {
+        /* If run length > 15, emit special run-length-16 codes. */
+        PUT_BITS(code_0xf0, size_0xf0)
+        r -= 16;
+      }
+      /* Emit Huffman symbol for run length / number of bits. (F.1.2.2.1) */
+      unsigned int rs = (r << 4) + nbits;
+      PUT_CODE(actbl->ehufco[rs], actbl->ehufsi[rs], diff)
+      i++;
+      bitmap <<= 1;
+    }
+  } else if (bitmap != 0) {
+    uint16_t block_abs[DCTSIZE2];
+    /* Compute and store absolute value of coefficients. */
+    int16x8_t abs_row1 = vabsq_s16(row1);
+    int16x8_t abs_row2 = vabsq_s16(row2);
+    int16x8_t abs_row3 = vabsq_s16(row3);
+    int16x8_t abs_row4 = vabsq_s16(row4);
+    int16x8_t abs_row5 = vabsq_s16(row5);
+    int16x8_t abs_row6 = vabsq_s16(row6);
+    int16x8_t abs_row7 = vabsq_s16(row7);
+    vst1q_u16(block_abs + 0 * DCTSIZE, vreinterpretq_u16_s16(abs_row0));
+    vst1q_u16(block_abs + 1 * DCTSIZE, vreinterpretq_u16_s16(abs_row1));
+    vst1q_u16(block_abs + 2 * DCTSIZE, vreinterpretq_u16_s16(abs_row2));
+    vst1q_u16(block_abs + 3 * DCTSIZE, vreinterpretq_u16_s16(abs_row3));
+    vst1q_u16(block_abs + 4 * DCTSIZE, vreinterpretq_u16_s16(abs_row4));
+    vst1q_u16(block_abs + 5 * DCTSIZE, vreinterpretq_u16_s16(abs_row5));
+    vst1q_u16(block_abs + 6 * DCTSIZE, vreinterpretq_u16_s16(abs_row6));
+    vst1q_u16(block_abs + 7 * DCTSIZE, vreinterpretq_u16_s16(abs_row7));
+    /* Compute diff bits (without nbits mask) and store. */
+    uint16x8_t row1_diff = veorq_u16(vreinterpretq_u16_s16(abs_row1),
+                                     vcltzq_s16(row1));
+    uint16x8_t row2_diff = veorq_u16(vreinterpretq_u16_s16(abs_row2),
+                                     vcltzq_s16(row2));
+    uint16x8_t row3_diff = veorq_u16(vreinterpretq_u16_s16(abs_row3),
+                                     vcltzq_s16(row3));
+    uint16x8_t row4_diff = veorq_u16(vreinterpretq_u16_s16(abs_row4),
+                                     vcltzq_s16(row4));
+    uint16x8_t row5_diff = veorq_u16(vreinterpretq_u16_s16(abs_row5),
+                                     vcltzq_s16(row5));
+    uint16x8_t row6_diff = veorq_u16(vreinterpretq_u16_s16(abs_row6),
+                                     vcltzq_s16(row6));
+    uint16x8_t row7_diff = veorq_u16(vreinterpretq_u16_s16(abs_row7),
+                                     vcltzq_s16(row7));
+    vst1q_u16(block_diff + 0 * DCTSIZE, row0_diff);
+    vst1q_u16(block_diff + 1 * DCTSIZE, row1_diff);
+    vst1q_u16(block_diff + 2 * DCTSIZE, row2_diff);
+    vst1q_u16(block_diff + 3 * DCTSIZE, row3_diff);
+    vst1q_u16(block_diff + 4 * DCTSIZE, row4_diff);
+    vst1q_u16(block_diff + 5 * DCTSIZE, row5_diff);
+    vst1q_u16(block_diff + 6 * DCTSIZE, row6_diff);
+    vst1q_u16(block_diff + 7 * DCTSIZE, row7_diff);
+
+    /* Same as above but must mask diff bits and compute nbits on demand. */
+    while (bitmap != 0) {
+      r = BUILTIN_CLZLL(bitmap);
+      i += r;
+      bitmap <<= r;
+      lz = BUILTIN_CLZ(block_abs[i]);
+      nbits = 32 - lz;
+      diff = ((unsigned int)block_diff[i] << lz) >> lz;
+      while (r > 15) {
+        /* If run length > 15, emit special run-length-16 codes. */
+        PUT_BITS(code_0xf0, size_0xf0)
+        r -= 16;
+      }
+      /* Emit Huffman symbol for run length / number of bits. (F.1.2.2.1) */
+      unsigned int rs = (r << 4) + nbits;
+      PUT_CODE(actbl->ehufco[rs], actbl->ehufsi[rs], diff)
+      i++;
+      bitmap <<= 1;
+    }
+  }
+
+  /* If the last coefficient(s) were zero, emit an end-of-block (EOB) code.
+   * The value of RS for the EOB code is 0.
+   */
+  if (i != 64) {
+    PUT_BITS(actbl->ehufco[0], actbl->ehufsi[0])
+  }
+
+  state_ptr->cur.put_buffer = put_buffer;
+  state_ptr->cur.free_bits = free_bits;
+
+  return buffer;
+}
@@ -0,0 +1,28 @@
+/*
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* How to obtain memory alignment for structures and variables */
+#if defined(_MSC_VER)
+#define ALIGN(alignment)  __declspec(align(alignment))
+#elif defined(__clang__) || defined(__GNUC__)
+#define ALIGN(alignment)  __attribute__((aligned(alignment)))
+#else
+#error "Unknown compiler"
+#endif
@@ -0,0 +1,160 @@
+/*
+ * jccolor-neon.c - colorspace conversion (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+#include "neon-compat.h"
+
+#include <arm_neon.h>
+
+
+/* RGB -> YCbCr conversion constants */
+
+#define F_0_298  19595
+#define F_0_587  38470
+#define F_0_113  7471
+#define F_0_168  11059
+#define F_0_331  21709
+#define F_0_500  32768
+#define F_0_418  27439
+#define F_0_081  5329
+
+ALIGN(16) static const uint16_t jsimd_rgb_ycc_neon_consts[] = {
+  F_0_298, F_0_587, F_0_113, F_0_168,
+  F_0_331, F_0_500, F_0_418, F_0_081
+};
+
+
+/* Include inline routines for colorspace extensions. */
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+
+#define RGB_RED  EXT_RGB_RED
+#define RGB_GREEN  EXT_RGB_GREEN
+#define RGB_BLUE  EXT_RGB_BLUE
+#define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+#define jsimd_rgb_ycc_convert_neon  jsimd_extrgb_ycc_convert_neon
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_ycc_convert_neon
+
+#define RGB_RED  EXT_RGBX_RED
+#define RGB_GREEN  EXT_RGBX_GREEN
+#define RGB_BLUE  EXT_RGBX_BLUE
+#define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+#define jsimd_rgb_ycc_convert_neon  jsimd_extrgbx_ycc_convert_neon
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_ycc_convert_neon
+
+#define RGB_RED  EXT_BGR_RED
+#define RGB_GREEN  EXT_BGR_GREEN
+#define RGB_BLUE  EXT_BGR_BLUE
+#define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+#define jsimd_rgb_ycc_convert_neon  jsimd_extbgr_ycc_convert_neon
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_ycc_convert_neon
+
+#define RGB_RED  EXT_BGRX_RED
+#define RGB_GREEN  EXT_BGRX_GREEN
+#define RGB_BLUE  EXT_BGRX_BLUE
+#define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+#define jsimd_rgb_ycc_convert_neon  jsimd_extbgrx_ycc_convert_neon
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_ycc_convert_neon
+
+#define RGB_RED  EXT_XBGR_RED
+#define RGB_GREEN  EXT_XBGR_GREEN
+#define RGB_BLUE  EXT_XBGR_BLUE
+#define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+#define jsimd_rgb_ycc_convert_neon  jsimd_extxbgr_ycc_convert_neon
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_ycc_convert_neon
+
+#define RGB_RED  EXT_XRGB_RED
+#define RGB_GREEN  EXT_XRGB_GREEN
+#define RGB_BLUE  EXT_XRGB_BLUE
+#define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+#define jsimd_rgb_ycc_convert_neon  jsimd_extxrgb_ycc_convert_neon
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
+#else
+#include "aarch32/jccolext-neon.c"
+#endif
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_ycc_convert_neon
@@ -0,0 +1,120 @@
+/*
+ * jcgray-neon.c - grayscale colorspace conversion (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+
+#include <arm_neon.h>
+
+
+/* RGB -> Grayscale conversion constants */
+
+#define F_0_298  19595
+#define F_0_587  38470
+#define F_0_113  7471
+
+
+/* Include inline routines for colorspace extensions. */
+
+#include "jcgryext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+
+#define RGB_RED  EXT_RGB_RED
+#define RGB_GREEN  EXT_RGB_GREEN
+#define RGB_BLUE  EXT_RGB_BLUE
+#define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+#define jsimd_rgb_gray_convert_neon  jsimd_extrgb_gray_convert_neon
+#include "jcgryext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_gray_convert_neon
+
+#define RGB_RED  EXT_RGBX_RED
+#define RGB_GREEN  EXT_RGBX_GREEN
+#define RGB_BLUE  EXT_RGBX_BLUE
+#define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+#define jsimd_rgb_gray_convert_neon  jsimd_extrgbx_gray_convert_neon
+#include "jcgryext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_gray_convert_neon
+
+#define RGB_RED  EXT_BGR_RED
+#define RGB_GREEN  EXT_BGR_GREEN
+#define RGB_BLUE  EXT_BGR_BLUE
+#define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+#define jsimd_rgb_gray_convert_neon  jsimd_extbgr_gray_convert_neon
+#include "jcgryext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_gray_convert_neon
+
+#define RGB_RED  EXT_BGRX_RED
+#define RGB_GREEN  EXT_BGRX_GREEN
+#define RGB_BLUE  EXT_BGRX_BLUE
+#define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+#define jsimd_rgb_gray_convert_neon  jsimd_extbgrx_gray_convert_neon
+#include "jcgryext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_gray_convert_neon
+
+#define RGB_RED  EXT_XBGR_RED
+#define RGB_GREEN  EXT_XBGR_GREEN
+#define RGB_BLUE  EXT_XBGR_BLUE
+#define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+#define jsimd_rgb_gray_convert_neon  jsimd_extxbgr_gray_convert_neon
+#include "jcgryext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_gray_convert_neon
+
+#define RGB_RED  EXT_XRGB_RED
+#define RGB_GREEN  EXT_XRGB_GREEN
+#define RGB_BLUE  EXT_XRGB_BLUE
+#define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+#define jsimd_rgb_gray_convert_neon  jsimd_extxrgb_gray_convert_neon
+#include "jcgryext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_rgb_gray_convert_neon
@@ -0,0 +1,106 @@
+/*
+ * jcgryext-neon.c - grayscale colorspace conversion (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* This file is included by jcgray-neon.c */
+
+
+/* RGB -> Grayscale conversion is defined by the following equation:
+ *    Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+ *
+ * Avoid floating point arithmetic by using shifted integer constants:
+ *    0.29899597 = 19595 * 2^-16
+ *    0.58700561 = 38470 * 2^-16
+ *    0.11399841 =  7471 * 2^-16
+ * These constants are defined in jcgray-neon.c
+ *
+ * This is the same computation as the RGB -> Y portion of RGB -> YCbCr.
+ */
+
+void jsimd_rgb_gray_convert_neon(JDIMENSION image_width, JSAMPARRAY input_buf,
+                                 JSAMPIMAGE output_buf, JDIMENSION output_row,
+                                 int num_rows)
+{
+  JSAMPROW inptr;
+  JSAMPROW outptr;
+  /* Allocate temporary buffer for final (image_width % 16) pixels in row. */
+  ALIGN(16) uint8_t tmp_buf[16 * RGB_PIXELSIZE];
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr = output_buf[0][output_row];
+    output_row++;
+
+    int cols_remaining = image_width;
+    for (; cols_remaining > 0; cols_remaining -= 16) {
+
+      /* To prevent buffer overread by the vector load instructions, the last
+       * (image_width % 16) columns of data are first memcopied to a temporary
+       * buffer large enough to accommodate the vector load.
+       */
+      if (cols_remaining < 16) {
+        memcpy(tmp_buf, inptr, cols_remaining * RGB_PIXELSIZE);
+        inptr = tmp_buf;
+      }
+
+#if RGB_PIXELSIZE == 4
+      uint8x16x4_t input_pixels = vld4q_u8(inptr);
+#else
+      uint8x16x3_t input_pixels = vld3q_u8(inptr);
+#endif
+      uint16x8_t r_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_RED]));
+      uint16x8_t r_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_RED]));
+      uint16x8_t g_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_GREEN]));
+      uint16x8_t g_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_GREEN]));
+      uint16x8_t b_l = vmovl_u8(vget_low_u8(input_pixels.val[RGB_BLUE]));
+      uint16x8_t b_h = vmovl_u8(vget_high_u8(input_pixels.val[RGB_BLUE]));
+
+      /* Compute Y = 0.29900 * R + 0.58700 * G + 0.11400 * B */
+      uint32x4_t y_ll = vmull_n_u16(vget_low_u16(r_l), F_0_298);
+      uint32x4_t y_lh = vmull_n_u16(vget_high_u16(r_l), F_0_298);
+      uint32x4_t y_hl = vmull_n_u16(vget_low_u16(r_h), F_0_298);
+      uint32x4_t y_hh = vmull_n_u16(vget_high_u16(r_h), F_0_298);
+      y_ll = vmlal_n_u16(y_ll, vget_low_u16(g_l), F_0_587);
+      y_lh = vmlal_n_u16(y_lh, vget_high_u16(g_l), F_0_587);
+      y_hl = vmlal_n_u16(y_hl, vget_low_u16(g_h), F_0_587);
+      y_hh = vmlal_n_u16(y_hh, vget_high_u16(g_h), F_0_587);
+      y_ll = vmlal_n_u16(y_ll, vget_low_u16(b_l), F_0_113);
+      y_lh = vmlal_n_u16(y_lh, vget_high_u16(b_l), F_0_113);
+      y_hl = vmlal_n_u16(y_hl, vget_low_u16(b_h), F_0_113);
+      y_hh = vmlal_n_u16(y_hh, vget_high_u16(b_h), F_0_113);
+
+      /* Descale Y values (rounding right shift) and narrow to 16-bit. */
+      uint16x8_t y_l = vcombine_u16(vrshrn_n_u32(y_ll, 16),
+                                    vrshrn_n_u32(y_lh, 16));
+      uint16x8_t y_h = vcombine_u16(vrshrn_n_u32(y_hl, 16),
+                                    vrshrn_n_u32(y_hh, 16));
+
+      /* Narrow Y values to 8-bit and store to memory.  Buffer overwrite is
+       * permitted up to the next multiple of ALIGN_SIZE bytes.
+       */
+      vst1q_u8(outptr, vcombine_u8(vmovn_u16(y_l), vmovn_u16(y_h)));
+
+      /* Increment pointers. */
+      inptr += (16 * RGB_PIXELSIZE);
+      outptr += 16;
+    }
+  }
+}
@@ -0,0 +1,131 @@
+/*
+ * jchuff.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2009, 2018, 2021, D. R. Commander.
+ * Copyright (C) 2018, Matthias Räncker.
+ * Copyright (C) 2020-2021, Arm Limited.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ */
+
+/* Expanded entropy encoder object for Huffman encoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+#define BIT_BUF_SIZE  64
+#else
+#define BIT_BUF_SIZE  32
+#endif
+
+typedef struct {
+  size_t put_buffer;                    /* current bit accumulation buffer */
+  int free_bits;                        /* # of bits available in it */
+  int last_dc_val[MAX_COMPS_IN_SCAN];   /* last DC coef for each component */
+} savable_state;
+
+typedef struct {
+  JOCTET *next_output_byte;     /* => next byte to write in buffer */
+  size_t free_in_buffer;        /* # of byte spaces remaining in buffer */
+  savable_state cur;            /* Current bit buffer & DC state */
+  j_compress_ptr cinfo;         /* dump_buffer needs access to this */
+  int simd;
+} working_state;
+
+/* Outputting bits to the file */
+
+/* Output byte b and, speculatively, an additional 0 byte. 0xFF must be encoded
+ * as 0xFF 0x00, so the output buffer pointer is advanced by 2 if the byte is
+ * 0xFF.  Otherwise, the output buffer pointer is advanced by 1, and the
+ * speculative 0 byte will be overwritten by the next byte.
+ */
+#define EMIT_BYTE(b) { \
+  buffer[0] = (JOCTET)(b); \
+  buffer[1] = 0; \
+  buffer -= -2 + ((JOCTET)(b) < 0xFF); \
+}
+
+/* Output the entire bit buffer.  If there are no 0xFF bytes in it, then write
+ * directly to the output buffer.  Otherwise, use the EMIT_BYTE() macro to
+ * encode 0xFF as 0xFF 0x00.
+ */
+#if defined(__aarch64__) || defined(_M_ARM64)
+
+#define FLUSH() { \
+  if (put_buffer & 0x8080808080808080 & ~(put_buffer + 0x0101010101010101)) { \
+    EMIT_BYTE(put_buffer >> 56) \
+    EMIT_BYTE(put_buffer >> 48) \
+    EMIT_BYTE(put_buffer >> 40) \
+    EMIT_BYTE(put_buffer >> 32) \
+    EMIT_BYTE(put_buffer >> 24) \
+    EMIT_BYTE(put_buffer >> 16) \
+    EMIT_BYTE(put_buffer >>  8) \
+    EMIT_BYTE(put_buffer      ) \
+  } else { \
+    *((uint64_t *)buffer) = BUILTIN_BSWAP64(put_buffer); \
+    buffer += 8; \
+  } \
+}
+
+#else
+
+#if defined(_MSC_VER) && !defined(__clang__)
+#define SPLAT() { \
+  buffer[0] = (JOCTET)(put_buffer >> 24); \
+  buffer[1] = (JOCTET)(put_buffer >> 16); \
+  buffer[2] = (JOCTET)(put_buffer >>  8); \
+  buffer[3] = (JOCTET)(put_buffer      ); \
+  buffer += 4; \
+}
+#else
+#define SPLAT() { \
+  put_buffer = __builtin_bswap32(put_buffer); \
+  __asm__("str %1, [%0], #4" : "+r" (buffer) : "r" (put_buffer)); \
+}
+#endif
+
+#define FLUSH() { \
+  if (put_buffer & 0x80808080 & ~(put_buffer + 0x01010101)) { \
+    EMIT_BYTE(put_buffer >> 24) \
+    EMIT_BYTE(put_buffer >> 16) \
+    EMIT_BYTE(put_buffer >>  8) \
+    EMIT_BYTE(put_buffer      ) \
+  } else { \
+    SPLAT(); \
+  } \
+}
+
+#endif
+
+/* Fill the bit buffer to capacity with the leading bits from code, then output
+ * the bit buffer and put the remaining bits from code into the bit buffer.
+ */
+#define PUT_AND_FLUSH(code, size) { \
+  put_buffer = (put_buffer << (size + free_bits)) | (code >> -free_bits); \
+  FLUSH() \
+  free_bits += BIT_BUF_SIZE; \
+  put_buffer = code; \
+}
+
+/* Insert code into the bit buffer and output the bit buffer if needed.
+ * NOTE: We can't flush with free_bits == 0, since the left shift in
+ * PUT_AND_FLUSH() would have undefined behavior.
+ */
+#define PUT_BITS(code, size) { \
+  free_bits -= size; \
+  if (free_bits < 0) \
+    PUT_AND_FLUSH(code, size) \
+  else \
+    put_buffer = (put_buffer << size) | code; \
+}
+
+#define PUT_CODE(code, size, diff) { \
+  diff |= code << nbits; \
+  nbits += size; \
+  PUT_BITS(diff, nbits) \
+}
@@ -0,0 +1,622 @@
+/*
+ * jcphuff-neon.c - prepare data for progressive Huffman encoding (Arm Neon)
+ *
+ * Copyright (C) 2020-2021, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "jconfigint.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "neon-compat.h"
+
+#include <arm_neon.h>
+
+
+/* Data preparation for encode_mcu_AC_first().
+ *
+ * The equivalent scalar C function (encode_mcu_AC_first_prepare()) can be
+ * found in jcphuff.c.
+ */
+
+void jsimd_encode_mcu_AC_first_prepare_neon
+  (const JCOEF *block, const int *jpeg_natural_order_start, int Sl, int Al,
+   JCOEF *values, size_t *zerobits)
+{
+  JCOEF *values_ptr = values;
+  JCOEF *diff_values_ptr = values + DCTSIZE2;
+
+  /* Rows of coefficients to zero (since they haven't been processed) */
+  int i, rows_to_zero = 8;
+
+  for (i = 0; i < Sl / 16; i++) {
+    int16x8_t coefs1 = vld1q_dup_s16(block + jpeg_natural_order_start[0]);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs1, 1);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs1, 2);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs1, 3);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs1, 4);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs1, 5);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs1, 6);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs1, 7);
+    int16x8_t coefs2 = vld1q_dup_s16(block + jpeg_natural_order_start[8]);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[9], coefs2, 1);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[10], coefs2, 2);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[11], coefs2, 3);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[12], coefs2, 4);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[13], coefs2, 5);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[14], coefs2, 6);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[15], coefs2, 7);
+
+    /* Isolate sign of coefficients. */
+    int16x8_t sign_coefs1 = vshrq_n_s16(coefs1, 15);
+    int16x8_t sign_coefs2 = vshrq_n_s16(coefs2, 15);
+    /* Compute absolute value of coefficients and apply point transform Al. */
+    int16x8_t abs_coefs1 = vabsq_s16(coefs1);
+    int16x8_t abs_coefs2 = vabsq_s16(coefs2);
+    coefs1 = vshlq_s16(abs_coefs1, vdupq_n_s16(-Al));
+    coefs2 = vshlq_s16(abs_coefs2, vdupq_n_s16(-Al));
+
+    /* Compute diff values. */
+    int16x8_t diff1 = veorq_s16(coefs1, sign_coefs1);
+    int16x8_t diff2 = veorq_s16(coefs2, sign_coefs2);
+
+    /* Store transformed coefficients and diff values. */
+    vst1q_s16(values_ptr, coefs1);
+    vst1q_s16(values_ptr + DCTSIZE, coefs2);
+    vst1q_s16(diff_values_ptr, diff1);
+    vst1q_s16(diff_values_ptr + DCTSIZE, diff2);
+    values_ptr += 16;
+    diff_values_ptr += 16;
+    jpeg_natural_order_start += 16;
+    rows_to_zero -= 2;
+  }
+
+  /* Same operation but for remaining partial vector */
+  int remaining_coefs = Sl % 16;
+  if (remaining_coefs > 8) {
+    int16x8_t coefs1 = vld1q_dup_s16(block + jpeg_natural_order_start[0]);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs1, 1);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs1, 2);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs1, 3);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs1, 4);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs1, 5);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs1, 6);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs1, 7);
+    int16x8_t coefs2 = vdupq_n_s16(0);
+    switch (remaining_coefs) {
+    case 15:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[14], coefs2, 6);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 14:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[13], coefs2, 5);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 13:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[12], coefs2, 4);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 12:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[11], coefs2, 3);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 11:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[10], coefs2, 2);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 10:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[9], coefs2, 1);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 9:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[8], coefs2, 0);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    default:
+      break;
+    }
+
+    /* Isolate sign of coefficients. */
+    int16x8_t sign_coefs1 = vshrq_n_s16(coefs1, 15);
+    int16x8_t sign_coefs2 = vshrq_n_s16(coefs2, 15);
+    /* Compute absolute value of coefficients and apply point transform Al. */
+    int16x8_t abs_coefs1 = vabsq_s16(coefs1);
+    int16x8_t abs_coefs2 = vabsq_s16(coefs2);
+    coefs1 = vshlq_s16(abs_coefs1, vdupq_n_s16(-Al));
+    coefs2 = vshlq_s16(abs_coefs2, vdupq_n_s16(-Al));
+
+    /* Compute diff values. */
+    int16x8_t diff1 = veorq_s16(coefs1, sign_coefs1);
+    int16x8_t diff2 = veorq_s16(coefs2, sign_coefs2);
+
+    /* Store transformed coefficients and diff values. */
+    vst1q_s16(values_ptr, coefs1);
+    vst1q_s16(values_ptr + DCTSIZE, coefs2);
+    vst1q_s16(diff_values_ptr, diff1);
+    vst1q_s16(diff_values_ptr + DCTSIZE, diff2);
+    values_ptr += 16;
+    diff_values_ptr += 16;
+    rows_to_zero -= 2;
+
+  } else if (remaining_coefs > 0) {
+    int16x8_t coefs = vdupq_n_s16(0);
+
+    switch (remaining_coefs) {
+    case 8:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs, 7);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 7:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs, 6);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 6:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs, 5);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 5:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs, 4);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 4:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs, 3);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 3:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs, 2);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 2:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs, 1);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 1:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[0], coefs, 0);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    default:
+      break;
+    }
+
+    /* Isolate sign of coefficients. */
+    int16x8_t sign_coefs = vshrq_n_s16(coefs, 15);
+    /* Compute absolute value of coefficients and apply point transform Al. */
+    int16x8_t abs_coefs = vabsq_s16(coefs);
+    coefs = vshlq_s16(abs_coefs, vdupq_n_s16(-Al));
+
+    /* Compute diff values. */
+    int16x8_t diff = veorq_s16(coefs, sign_coefs);
+
+    /* Store transformed coefficients and diff values. */
+    vst1q_s16(values_ptr, coefs);
+    vst1q_s16(diff_values_ptr, diff);
+    values_ptr += 8;
+    diff_values_ptr += 8;
+    rows_to_zero--;
+  }
+
+  /* Zero remaining memory in the values and diff_values blocks. */
+  for (i = 0; i < rows_to_zero; i++) {
+    vst1q_s16(values_ptr, vdupq_n_s16(0));
+    vst1q_s16(diff_values_ptr, vdupq_n_s16(0));
+    values_ptr += 8;
+    diff_values_ptr += 8;
+  }
+
+  /* Construct zerobits bitmap.  A set bit means that the corresponding
+   * coefficient != 0.
+   */
+  int16x8_t row0 = vld1q_s16(values + 0 * DCTSIZE);
+  int16x8_t row1 = vld1q_s16(values + 1 * DCTSIZE);
+  int16x8_t row2 = vld1q_s16(values + 2 * DCTSIZE);
+  int16x8_t row3 = vld1q_s16(values + 3 * DCTSIZE);
+  int16x8_t row4 = vld1q_s16(values + 4 * DCTSIZE);
+  int16x8_t row5 = vld1q_s16(values + 5 * DCTSIZE);
+  int16x8_t row6 = vld1q_s16(values + 6 * DCTSIZE);
+  int16x8_t row7 = vld1q_s16(values + 7 * DCTSIZE);
+
+  uint8x8_t row0_eq0 = vmovn_u16(vceqq_s16(row0, vdupq_n_s16(0)));
+  uint8x8_t row1_eq0 = vmovn_u16(vceqq_s16(row1, vdupq_n_s16(0)));
+  uint8x8_t row2_eq0 = vmovn_u16(vceqq_s16(row2, vdupq_n_s16(0)));
+  uint8x8_t row3_eq0 = vmovn_u16(vceqq_s16(row3, vdupq_n_s16(0)));
+  uint8x8_t row4_eq0 = vmovn_u16(vceqq_s16(row4, vdupq_n_s16(0)));
+  uint8x8_t row5_eq0 = vmovn_u16(vceqq_s16(row5, vdupq_n_s16(0)));
+  uint8x8_t row6_eq0 = vmovn_u16(vceqq_s16(row6, vdupq_n_s16(0)));
+  uint8x8_t row7_eq0 = vmovn_u16(vceqq_s16(row7, vdupq_n_s16(0)));
+
+  /* { 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80 } */
+  const uint8x8_t bitmap_mask =
+    vreinterpret_u8_u64(vmov_n_u64(0x8040201008040201));
+
+  row0_eq0 = vand_u8(row0_eq0, bitmap_mask);
+  row1_eq0 = vand_u8(row1_eq0, bitmap_mask);
+  row2_eq0 = vand_u8(row2_eq0, bitmap_mask);
+  row3_eq0 = vand_u8(row3_eq0, bitmap_mask);
+  row4_eq0 = vand_u8(row4_eq0, bitmap_mask);
+  row5_eq0 = vand_u8(row5_eq0, bitmap_mask);
+  row6_eq0 = vand_u8(row6_eq0, bitmap_mask);
+  row7_eq0 = vand_u8(row7_eq0, bitmap_mask);
+
+  uint8x8_t bitmap_rows_01 = vpadd_u8(row0_eq0, row1_eq0);
+  uint8x8_t bitmap_rows_23 = vpadd_u8(row2_eq0, row3_eq0);
+  uint8x8_t bitmap_rows_45 = vpadd_u8(row4_eq0, row5_eq0);
+  uint8x8_t bitmap_rows_67 = vpadd_u8(row6_eq0, row7_eq0);
+  uint8x8_t bitmap_rows_0123 = vpadd_u8(bitmap_rows_01, bitmap_rows_23);
+  uint8x8_t bitmap_rows_4567 = vpadd_u8(bitmap_rows_45, bitmap_rows_67);
+  uint8x8_t bitmap_all = vpadd_u8(bitmap_rows_0123, bitmap_rows_4567);
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+  /* Move bitmap to a 64-bit scalar register. */
+  uint64_t bitmap = vget_lane_u64(vreinterpret_u64_u8(bitmap_all), 0);
+  /* Store zerobits bitmap. */
+  *zerobits = ~bitmap;
+#else
+  /* Move bitmap to two 32-bit scalar registers. */
+  uint32_t bitmap0 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 0);
+  uint32_t bitmap1 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 1);
+  /* Store zerobits bitmap. */
+  zerobits[0] = ~bitmap0;
+  zerobits[1] = ~bitmap1;
+#endif
+}
+
+
+/* Data preparation for encode_mcu_AC_refine().
+ *
+ * The equivalent scalar C function (encode_mcu_AC_refine_prepare()) can be
+ * found in jcphuff.c.
+ */
+
+int jsimd_encode_mcu_AC_refine_prepare_neon
+  (const JCOEF *block, const int *jpeg_natural_order_start, int Sl, int Al,
+   JCOEF *absvalues, size_t *bits)
+{
+  /* Temporary storage buffers for data used to compute the signbits bitmap and
+   * the end-of-block (EOB) position
+   */
+  uint8_t coef_sign_bits[64];
+  uint8_t coef_eq1_bits[64];
+
+  JCOEF *absvalues_ptr = absvalues;
+  uint8_t *coef_sign_bits_ptr = coef_sign_bits;
+  uint8_t *eq1_bits_ptr = coef_eq1_bits;
+
+  /* Rows of coefficients to zero (since they haven't been processed) */
+  int i, rows_to_zero = 8;
+
+  for (i = 0; i < Sl / 16; i++) {
+    int16x8_t coefs1 = vld1q_dup_s16(block + jpeg_natural_order_start[0]);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs1, 1);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs1, 2);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs1, 3);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs1, 4);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs1, 5);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs1, 6);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs1, 7);
+    int16x8_t coefs2 = vld1q_dup_s16(block + jpeg_natural_order_start[8]);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[9], coefs2, 1);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[10], coefs2, 2);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[11], coefs2, 3);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[12], coefs2, 4);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[13], coefs2, 5);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[14], coefs2, 6);
+    coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[15], coefs2, 7);
+
+    /* Compute and store data for signbits bitmap. */
+    uint8x8_t sign_coefs1 =
+      vmovn_u16(vreinterpretq_u16_s16(vshrq_n_s16(coefs1, 15)));
+    uint8x8_t sign_coefs2 =
+      vmovn_u16(vreinterpretq_u16_s16(vshrq_n_s16(coefs2, 15)));
+    vst1_u8(coef_sign_bits_ptr, sign_coefs1);
+    vst1_u8(coef_sign_bits_ptr + DCTSIZE, sign_coefs2);
+
+    /* Compute absolute value of coefficients and apply point transform Al. */
+    int16x8_t abs_coefs1 = vabsq_s16(coefs1);
+    int16x8_t abs_coefs2 = vabsq_s16(coefs2);
+    coefs1 = vshlq_s16(abs_coefs1, vdupq_n_s16(-Al));
+    coefs2 = vshlq_s16(abs_coefs2, vdupq_n_s16(-Al));
+    vst1q_s16(absvalues_ptr, coefs1);
+    vst1q_s16(absvalues_ptr + DCTSIZE, coefs2);
+
+    /* Test whether transformed coefficient values == 1 (used to find EOB
+     * position.)
+     */
+    uint8x8_t coefs_eq11 = vmovn_u16(vceqq_s16(coefs1, vdupq_n_s16(1)));
+    uint8x8_t coefs_eq12 = vmovn_u16(vceqq_s16(coefs2, vdupq_n_s16(1)));
+    vst1_u8(eq1_bits_ptr, coefs_eq11);
+    vst1_u8(eq1_bits_ptr + DCTSIZE, coefs_eq12);
+
+    absvalues_ptr += 16;
+    coef_sign_bits_ptr += 16;
+    eq1_bits_ptr += 16;
+    jpeg_natural_order_start += 16;
+    rows_to_zero -= 2;
+  }
+
+  /* Same operation but for remaining partial vector */
+  int remaining_coefs = Sl % 16;
+  if (remaining_coefs > 8) {
+    int16x8_t coefs1 = vld1q_dup_s16(block + jpeg_natural_order_start[0]);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs1, 1);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs1, 2);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs1, 3);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs1, 4);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs1, 5);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs1, 6);
+    coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs1, 7);
+    int16x8_t coefs2 = vdupq_n_s16(0);
+    switch (remaining_coefs) {
+    case 15:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[14], coefs2, 6);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 14:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[13], coefs2, 5);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 13:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[12], coefs2, 4);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 12:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[11], coefs2, 3);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 11:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[10], coefs2, 2);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 10:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[9], coefs2, 1);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 9:
+      coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[8], coefs2, 0);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    default:
+      break;
+    }
+
+    /* Compute and store data for signbits bitmap. */
+    uint8x8_t sign_coefs1 =
+      vmovn_u16(vreinterpretq_u16_s16(vshrq_n_s16(coefs1, 15)));
+    uint8x8_t sign_coefs2 =
+      vmovn_u16(vreinterpretq_u16_s16(vshrq_n_s16(coefs2, 15)));
+    vst1_u8(coef_sign_bits_ptr, sign_coefs1);
+    vst1_u8(coef_sign_bits_ptr + DCTSIZE, sign_coefs2);
+
+    /* Compute absolute value of coefficients and apply point transform Al. */
+    int16x8_t abs_coefs1 = vabsq_s16(coefs1);
+    int16x8_t abs_coefs2 = vabsq_s16(coefs2);
+    coefs1 = vshlq_s16(abs_coefs1, vdupq_n_s16(-Al));
+    coefs2 = vshlq_s16(abs_coefs2, vdupq_n_s16(-Al));
+    vst1q_s16(absvalues_ptr, coefs1);
+    vst1q_s16(absvalues_ptr + DCTSIZE, coefs2);
+
+    /* Test whether transformed coefficient values == 1 (used to find EOB
+     * position.)
+     */
+    uint8x8_t coefs_eq11 = vmovn_u16(vceqq_s16(coefs1, vdupq_n_s16(1)));
+    uint8x8_t coefs_eq12 = vmovn_u16(vceqq_s16(coefs2, vdupq_n_s16(1)));
+    vst1_u8(eq1_bits_ptr, coefs_eq11);
+    vst1_u8(eq1_bits_ptr + DCTSIZE, coefs_eq12);
+
+    absvalues_ptr += 16;
+    coef_sign_bits_ptr += 16;
+    eq1_bits_ptr += 16;
+    jpeg_natural_order_start += 16;
+    rows_to_zero -= 2;
+
+  } else if (remaining_coefs > 0) {
+    int16x8_t coefs = vdupq_n_s16(0);
+
+    switch (remaining_coefs) {
+    case 8:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs, 7);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 7:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs, 6);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 6:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs, 5);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 5:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs, 4);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 4:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs, 3);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 3:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs, 2);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 2:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs, 1);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 1:
+      coefs = vld1q_lane_s16(block + jpeg_natural_order_start[0], coefs, 0);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    default:
+      break;
+    }
+
+    /* Compute and store data for signbits bitmap. */
+    uint8x8_t sign_coefs =
+      vmovn_u16(vreinterpretq_u16_s16(vshrq_n_s16(coefs, 15)));
+    vst1_u8(coef_sign_bits_ptr, sign_coefs);
+
+    /* Compute absolute value of coefficients and apply point transform Al. */
+    int16x8_t abs_coefs = vabsq_s16(coefs);
+    coefs = vshlq_s16(abs_coefs, vdupq_n_s16(-Al));
+    vst1q_s16(absvalues_ptr, coefs);
+
+    /* Test whether transformed coefficient values == 1 (used to find EOB
+     * position.)
+     */
+    uint8x8_t coefs_eq1 = vmovn_u16(vceqq_s16(coefs, vdupq_n_s16(1)));
+    vst1_u8(eq1_bits_ptr, coefs_eq1);
+
+    absvalues_ptr += 8;
+    coef_sign_bits_ptr += 8;
+    eq1_bits_ptr += 8;
+    rows_to_zero--;
+  }
+
+  /* Zero remaining memory in blocks. */
+  for (i = 0; i < rows_to_zero; i++) {
+    vst1q_s16(absvalues_ptr, vdupq_n_s16(0));
+    vst1_u8(coef_sign_bits_ptr, vdup_n_u8(0));
+    vst1_u8(eq1_bits_ptr, vdup_n_u8(0));
+    absvalues_ptr += 8;
+    coef_sign_bits_ptr += 8;
+    eq1_bits_ptr += 8;
+  }
+
+  /* Construct zerobits bitmap. */
+  int16x8_t abs_row0 = vld1q_s16(absvalues + 0 * DCTSIZE);
+  int16x8_t abs_row1 = vld1q_s16(absvalues + 1 * DCTSIZE);
+  int16x8_t abs_row2 = vld1q_s16(absvalues + 2 * DCTSIZE);
+  int16x8_t abs_row3 = vld1q_s16(absvalues + 3 * DCTSIZE);
+  int16x8_t abs_row4 = vld1q_s16(absvalues + 4 * DCTSIZE);
+  int16x8_t abs_row5 = vld1q_s16(absvalues + 5 * DCTSIZE);
+  int16x8_t abs_row6 = vld1q_s16(absvalues + 6 * DCTSIZE);
+  int16x8_t abs_row7 = vld1q_s16(absvalues + 7 * DCTSIZE);
+
+  uint8x8_t abs_row0_eq0 = vmovn_u16(vceqq_s16(abs_row0, vdupq_n_s16(0)));
+  uint8x8_t abs_row1_eq0 = vmovn_u16(vceqq_s16(abs_row1, vdupq_n_s16(0)));
+  uint8x8_t abs_row2_eq0 = vmovn_u16(vceqq_s16(abs_row2, vdupq_n_s16(0)));
+  uint8x8_t abs_row3_eq0 = vmovn_u16(vceqq_s16(abs_row3, vdupq_n_s16(0)));
+  uint8x8_t abs_row4_eq0 = vmovn_u16(vceqq_s16(abs_row4, vdupq_n_s16(0)));
+  uint8x8_t abs_row5_eq0 = vmovn_u16(vceqq_s16(abs_row5, vdupq_n_s16(0)));
+  uint8x8_t abs_row6_eq0 = vmovn_u16(vceqq_s16(abs_row6, vdupq_n_s16(0)));
+  uint8x8_t abs_row7_eq0 = vmovn_u16(vceqq_s16(abs_row7, vdupq_n_s16(0)));
+
+  /* { 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80 } */
+  const uint8x8_t bitmap_mask =
+    vreinterpret_u8_u64(vmov_n_u64(0x8040201008040201));
+
+  abs_row0_eq0 = vand_u8(abs_row0_eq0, bitmap_mask);
+  abs_row1_eq0 = vand_u8(abs_row1_eq0, bitmap_mask);
+  abs_row2_eq0 = vand_u8(abs_row2_eq0, bitmap_mask);
+  abs_row3_eq0 = vand_u8(abs_row3_eq0, bitmap_mask);
+  abs_row4_eq0 = vand_u8(abs_row4_eq0, bitmap_mask);
+  abs_row5_eq0 = vand_u8(abs_row5_eq0, bitmap_mask);
+  abs_row6_eq0 = vand_u8(abs_row6_eq0, bitmap_mask);
+  abs_row7_eq0 = vand_u8(abs_row7_eq0, bitmap_mask);
+
+  uint8x8_t bitmap_rows_01 = vpadd_u8(abs_row0_eq0, abs_row1_eq0);
+  uint8x8_t bitmap_rows_23 = vpadd_u8(abs_row2_eq0, abs_row3_eq0);
+  uint8x8_t bitmap_rows_45 = vpadd_u8(abs_row4_eq0, abs_row5_eq0);
+  uint8x8_t bitmap_rows_67 = vpadd_u8(abs_row6_eq0, abs_row7_eq0);
+  uint8x8_t bitmap_rows_0123 = vpadd_u8(bitmap_rows_01, bitmap_rows_23);
+  uint8x8_t bitmap_rows_4567 = vpadd_u8(bitmap_rows_45, bitmap_rows_67);
+  uint8x8_t bitmap_all = vpadd_u8(bitmap_rows_0123, bitmap_rows_4567);
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+  /* Move bitmap to a 64-bit scalar register. */
+  uint64_t bitmap = vget_lane_u64(vreinterpret_u64_u8(bitmap_all), 0);
+  /* Store zerobits bitmap. */
+  bits[0] = ~bitmap;
+#else
+  /* Move bitmap to two 32-bit scalar registers. */
+  uint32_t bitmap0 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 0);
+  uint32_t bitmap1 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 1);
+  /* Store zerobits bitmap. */
+  bits[0] = ~bitmap0;
+  bits[1] = ~bitmap1;
+#endif
+
+  /* Construct signbits bitmap. */
+  uint8x8_t signbits_row0 = vld1_u8(coef_sign_bits + 0 * DCTSIZE);
+  uint8x8_t signbits_row1 = vld1_u8(coef_sign_bits + 1 * DCTSIZE);
+  uint8x8_t signbits_row2 = vld1_u8(coef_sign_bits + 2 * DCTSIZE);
+  uint8x8_t signbits_row3 = vld1_u8(coef_sign_bits + 3 * DCTSIZE);
+  uint8x8_t signbits_row4 = vld1_u8(coef_sign_bits + 4 * DCTSIZE);
+  uint8x8_t signbits_row5 = vld1_u8(coef_sign_bits + 5 * DCTSIZE);
+  uint8x8_t signbits_row6 = vld1_u8(coef_sign_bits + 6 * DCTSIZE);
+  uint8x8_t signbits_row7 = vld1_u8(coef_sign_bits + 7 * DCTSIZE);
+
+  signbits_row0 = vand_u8(signbits_row0, bitmap_mask);
+  signbits_row1 = vand_u8(signbits_row1, bitmap_mask);
+  signbits_row2 = vand_u8(signbits_row2, bitmap_mask);
+  signbits_row3 = vand_u8(signbits_row3, bitmap_mask);
+  signbits_row4 = vand_u8(signbits_row4, bitmap_mask);
+  signbits_row5 = vand_u8(signbits_row5, bitmap_mask);
+  signbits_row6 = vand_u8(signbits_row6, bitmap_mask);
+  signbits_row7 = vand_u8(signbits_row7, bitmap_mask);
+
+  bitmap_rows_01 = vpadd_u8(signbits_row0, signbits_row1);
+  bitmap_rows_23 = vpadd_u8(signbits_row2, signbits_row3);
+  bitmap_rows_45 = vpadd_u8(signbits_row4, signbits_row5);
+  bitmap_rows_67 = vpadd_u8(signbits_row6, signbits_row7);
+  bitmap_rows_0123 = vpadd_u8(bitmap_rows_01, bitmap_rows_23);
+  bitmap_rows_4567 = vpadd_u8(bitmap_rows_45, bitmap_rows_67);
+  bitmap_all = vpadd_u8(bitmap_rows_0123, bitmap_rows_4567);
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+  /* Move bitmap to a 64-bit scalar register. */
+  bitmap = vget_lane_u64(vreinterpret_u64_u8(bitmap_all), 0);
+  /* Store signbits bitmap. */
+  bits[1] = ~bitmap;
+#else
+  /* Move bitmap to two 32-bit scalar registers. */
+  bitmap0 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 0);
+  bitmap1 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 1);
+  /* Store signbits bitmap. */
+  bits[2] = ~bitmap0;
+  bits[3] = ~bitmap1;
+#endif
+
+  /* Construct bitmap to find EOB position (the index of the last coefficient
+   * equal to 1.)
+   */
+  uint8x8_t row0_eq1 = vld1_u8(coef_eq1_bits + 0 * DCTSIZE);
+  uint8x8_t row1_eq1 = vld1_u8(coef_eq1_bits + 1 * DCTSIZE);
+  uint8x8_t row2_eq1 = vld1_u8(coef_eq1_bits + 2 * DCTSIZE);
+  uint8x8_t row3_eq1 = vld1_u8(coef_eq1_bits + 3 * DCTSIZE);
+  uint8x8_t row4_eq1 = vld1_u8(coef_eq1_bits + 4 * DCTSIZE);
+  uint8x8_t row5_eq1 = vld1_u8(coef_eq1_bits + 5 * DCTSIZE);
+  uint8x8_t row6_eq1 = vld1_u8(coef_eq1_bits + 6 * DCTSIZE);
+  uint8x8_t row7_eq1 = vld1_u8(coef_eq1_bits + 7 * DCTSIZE);
+
+  row0_eq1 = vand_u8(row0_eq1, bitmap_mask);
+  row1_eq1 = vand_u8(row1_eq1, bitmap_mask);
+  row2_eq1 = vand_u8(row2_eq1, bitmap_mask);
+  row3_eq1 = vand_u8(row3_eq1, bitmap_mask);
+  row4_eq1 = vand_u8(row4_eq1, bitmap_mask);
+  row5_eq1 = vand_u8(row5_eq1, bitmap_mask);
+  row6_eq1 = vand_u8(row6_eq1, bitmap_mask);
+  row7_eq1 = vand_u8(row7_eq1, bitmap_mask);
+
+  bitmap_rows_01 = vpadd_u8(row0_eq1, row1_eq1);
+  bitmap_rows_23 = vpadd_u8(row2_eq1, row3_eq1);
+  bitmap_rows_45 = vpadd_u8(row4_eq1, row5_eq1);
+  bitmap_rows_67 = vpadd_u8(row6_eq1, row7_eq1);
+  bitmap_rows_0123 = vpadd_u8(bitmap_rows_01, bitmap_rows_23);
+  bitmap_rows_4567 = vpadd_u8(bitmap_rows_45, bitmap_rows_67);
+  bitmap_all = vpadd_u8(bitmap_rows_0123, bitmap_rows_4567);
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+  /* Move bitmap to a 64-bit scalar register. */
+  bitmap = vget_lane_u64(vreinterpret_u64_u8(bitmap_all), 0);
+
+  /* Return EOB position. */
+  if (bitmap == 0) {
+    /* EOB position is defined to be 0 if all coefficients != 1. */
+    return 0;
+  } else {
+    return 63 - BUILTIN_CLZLL(bitmap);
+  }
+#else
+  /* Move bitmap to two 32-bit scalar registers. */
+  bitmap0 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 0);
+  bitmap1 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 1);
+
+  /* Return EOB position. */
+  if (bitmap0 == 0 && bitmap1 == 0) {
+    return 0;
+  } else if (bitmap1 != 0) {
+    return 63 - BUILTIN_CLZ(bitmap1);
+  } else {
+    return 31 - BUILTIN_CLZ(bitmap0);
+  }
+#endif
+}
@@ -0,0 +1,192 @@
+/*
+ * jcsample-neon.c - downsampling (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+
+#include <arm_neon.h>
+
+
+ALIGN(16) static const uint8_t jsimd_h2_downsample_consts[] = {
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 0 */
+  0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 1 */
+  0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0E,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 2 */
+  0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0D, 0x0D,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 3 */
+  0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0C, 0x0C, 0x0C,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 4 */
+  0x08, 0x09, 0x0A, 0x0B, 0x0B, 0x0B, 0x0B, 0x0B,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 5 */
+  0x08, 0x09, 0x0A, 0x0A, 0x0A, 0x0A, 0x0A, 0x0A,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 6 */
+  0x08, 0x09, 0x09, 0x09, 0x09, 0x09, 0x09, 0x09,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 7 */
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,   /* Pad 8 */
+  0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x06,   /* Pad 9 */
+  0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x05, 0x05,   /* Pad 10 */
+  0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x04, 0x04, 0x04,   /* Pad 11 */
+  0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04,
+  0x00, 0x01, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03,   /* Pad 12 */
+  0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03,
+  0x00, 0x01, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,   /* Pad 13 */
+  0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+  0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,   /* Pad 14 */
+  0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,   /* Pad 15 */
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
+};
+
+
+/* Downsample pixel values of a single component.
+ * This version handles the common case of 2:1 horizontal and 1:1 vertical,
+ * without smoothing.
+ */
+
+void jsimd_h2v1_downsample_neon(JDIMENSION image_width, int max_v_samp_factor,
+                                JDIMENSION v_samp_factor,
+                                JDIMENSION width_in_blocks,
+                                JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  JSAMPROW inptr, outptr;
+  /* Load expansion mask to pad remaining elements of last DCT block. */
+  const int mask_offset = 16 * ((width_in_blocks * 2 * DCTSIZE) - image_width);
+  const uint8x16_t expand_mask =
+    vld1q_u8(&jsimd_h2_downsample_consts[mask_offset]);
+  /* Load bias pattern (alternating every pixel.) */
+  /* { 0, 1, 0, 1, 0, 1, 0, 1 } */
+  const uint16x8_t bias = vreinterpretq_u16_u32(vdupq_n_u32(0x00010000));
+  unsigned i, outrow;
+
+  for (outrow = 0; outrow < v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr = input_data[outrow];
+
+    /* Downsample all but the last DCT block of pixels. */
+    for (i = 0; i < width_in_blocks - 1; i++) {
+      uint8x16_t pixels = vld1q_u8(inptr + i * 2 * DCTSIZE);
+      /* Add adjacent pixel values, widen to 16-bit, and add bias. */
+      uint16x8_t samples_u16 = vpadalq_u8(bias, pixels);
+      /* Divide total by 2 and narrow to 8-bit. */
+      uint8x8_t samples_u8 = vshrn_n_u16(samples_u16, 1);
+      /* Store samples to memory. */
+      vst1_u8(outptr + i * DCTSIZE, samples_u8);
+    }
+
+    /* Load pixels in last DCT block into a table. */
+    uint8x16_t pixels = vld1q_u8(inptr + (width_in_blocks - 1) * 2 * DCTSIZE);
+#if defined(__aarch64__) || defined(_M_ARM64)
+    /* Pad the empty elements with the value of the last pixel. */
+    pixels = vqtbl1q_u8(pixels, expand_mask);
+#else
+    uint8x8x2_t table = { { vget_low_u8(pixels), vget_high_u8(pixels) } };
+    pixels = vcombine_u8(vtbl2_u8(table, vget_low_u8(expand_mask)),
+                         vtbl2_u8(table, vget_high_u8(expand_mask)));
+#endif
+    /* Add adjacent pixel values, widen to 16-bit, and add bias. */
+    uint16x8_t samples_u16 = vpadalq_u8(bias, pixels);
+    /* Divide total by 2, narrow to 8-bit, and store. */
+    uint8x8_t samples_u8 = vshrn_n_u16(samples_u16, 1);
+    vst1_u8(outptr + (width_in_blocks - 1) * DCTSIZE, samples_u8);
+  }
+}
+
+
+/* Downsample pixel values of a single component.
+ * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+ * without smoothing.
+ */
+
+void jsimd_h2v2_downsample_neon(JDIMENSION image_width, int max_v_samp_factor,
+                                JDIMENSION v_samp_factor,
+                                JDIMENSION width_in_blocks,
+                                JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  JSAMPROW inptr0, inptr1, outptr;
+  /* Load expansion mask to pad remaining elements of last DCT block. */
+  const int mask_offset = 16 * ((width_in_blocks * 2 * DCTSIZE) - image_width);
+  const uint8x16_t expand_mask =
+    vld1q_u8(&jsimd_h2_downsample_consts[mask_offset]);
+  /* Load bias pattern (alternating every pixel.) */
+  /* { 1, 2, 1, 2, 1, 2, 1, 2 } */
+  const uint16x8_t bias = vreinterpretq_u16_u32(vdupq_n_u32(0x00020001));
+  unsigned i, outrow;
+
+  for (outrow = 0; outrow < v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr0 = input_data[outrow];
+    inptr1 = input_data[outrow + 1];
+
+    /* Downsample all but the last DCT block of pixels. */
+    for (i = 0; i < width_in_blocks - 1; i++) {
+      uint8x16_t pixels_r0 = vld1q_u8(inptr0 + i * 2 * DCTSIZE);
+      uint8x16_t pixels_r1 = vld1q_u8(inptr1 + i * 2 * DCTSIZE);
+      /* Add adjacent pixel values in row 0, widen to 16-bit, and add bias. */
+      uint16x8_t samples_u16 = vpadalq_u8(bias, pixels_r0);
+      /* Add adjacent pixel values in row 1, widen to 16-bit, and accumulate.
+       */
+      samples_u16 = vpadalq_u8(samples_u16, pixels_r1);
+      /* Divide total by 4 and narrow to 8-bit. */
+      uint8x8_t samples_u8 = vshrn_n_u16(samples_u16, 2);
+      /* Store samples to memory and increment pointers. */
+      vst1_u8(outptr + i * DCTSIZE, samples_u8);
+    }
+
+    /* Load pixels in last DCT block into a table. */
+    uint8x16_t pixels_r0 =
+      vld1q_u8(inptr0 + (width_in_blocks - 1) * 2 * DCTSIZE);
+    uint8x16_t pixels_r1 =
+      vld1q_u8(inptr1 + (width_in_blocks - 1) * 2 * DCTSIZE);
+#if defined(__aarch64__) || defined(_M_ARM64)
+    /* Pad the empty elements with the value of the last pixel. */
+    pixels_r0 = vqtbl1q_u8(pixels_r0, expand_mask);
+    pixels_r1 = vqtbl1q_u8(pixels_r1, expand_mask);
+#else
+    uint8x8x2_t table_r0 =
+      { { vget_low_u8(pixels_r0), vget_high_u8(pixels_r0) } };
+    uint8x8x2_t table_r1 =
+      { { vget_low_u8(pixels_r1), vget_high_u8(pixels_r1) } };
+    pixels_r0 = vcombine_u8(vtbl2_u8(table_r0, vget_low_u8(expand_mask)),
+                            vtbl2_u8(table_r0, vget_high_u8(expand_mask)));
+    pixels_r1 = vcombine_u8(vtbl2_u8(table_r1, vget_low_u8(expand_mask)),
+                            vtbl2_u8(table_r1, vget_high_u8(expand_mask)));
+#endif
+    /* Add adjacent pixel values in row 0, widen to 16-bit, and add bias. */
+    uint16x8_t samples_u16 = vpadalq_u8(bias, pixels_r0);
+    /* Add adjacent pixel values in row 1, widen to 16-bit, and accumulate. */
+    samples_u16 = vpadalq_u8(samples_u16, pixels_r1);
+    /* Divide total by 4, narrow to 8-bit, and store. */
+    uint8x8_t samples_u8 = vshrn_n_u16(samples_u16, 2);
+    vst1_u8(outptr + (width_in_blocks - 1) * DCTSIZE, samples_u8);
+  }
+}
@@ -0,0 +1,374 @@
+/*
+ * jdcolext-neon.c - colorspace conversion (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* This file is included by jdcolor-neon.c. */
+
+
+/* YCbCr -> RGB conversion is defined by the following equations:
+ *    R = Y                        + 1.40200 * (Cr - 128)
+ *    G = Y - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128)
+ *    B = Y + 1.77200 * (Cb - 128)
+ *
+ * Scaled integer constants are used to avoid floating-point arithmetic:
+ *    0.3441467 = 11277 * 2^-15
+ *    0.7141418 = 23401 * 2^-15
+ *    1.4020386 = 22971 * 2^-14
+ *    1.7720337 = 29033 * 2^-14
+ * These constants are defined in jdcolor-neon.c.
+ *
+ * To ensure correct results, rounding is used when descaling.
+ */
+
+/* Notes on safe memory access for YCbCr -> RGB conversion routines:
+ *
+ * Input memory buffers can be safely overread up to the next multiple of
+ * ALIGN_SIZE bytes, since they are always allocated by alloc_sarray() in
+ * jmemmgr.c.
+ *
+ * The output buffer cannot safely be written beyond output_width, since
+ * output_buf points to a possibly unpadded row in the decompressed image
+ * buffer allocated by the calling program.
+ */
+
+void jsimd_ycc_rgb_convert_neon(JDIMENSION output_width, JSAMPIMAGE input_buf,
+                                JDIMENSION input_row, JSAMPARRAY output_buf,
+                                int num_rows)
+{
+  JSAMPROW outptr;
+  /* Pointers to Y, Cb, and Cr data */
+  JSAMPROW inptr0, inptr1, inptr2;
+
+  const int16x4_t consts = vld1_s16(jsimd_ycc_rgb_convert_neon_consts);
+  const int16x8_t neg_128 = vdupq_n_s16(-128);
+
+  while (--num_rows >= 0) {
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    int cols_remaining = output_width;
+    for (; cols_remaining >= 16; cols_remaining -= 16) {
+      uint8x16_t y  = vld1q_u8(inptr0);
+      uint8x16_t cb = vld1q_u8(inptr1);
+      uint8x16_t cr = vld1q_u8(inptr2);
+      /* Subtract 128 from Cb and Cr. */
+      int16x8_t cr_128_l =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128),
+                                       vget_low_u8(cr)));
+      int16x8_t cr_128_h =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128),
+                                       vget_high_u8(cr)));
+      int16x8_t cb_128_l =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128),
+                                       vget_low_u8(cb)));
+      int16x8_t cb_128_h =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128),
+                                       vget_high_u8(cb)));
+      /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+      int32x4_t g_sub_y_ll = vmull_lane_s16(vget_low_s16(cb_128_l), consts, 0);
+      int32x4_t g_sub_y_lh = vmull_lane_s16(vget_high_s16(cb_128_l),
+                                            consts, 0);
+      int32x4_t g_sub_y_hl = vmull_lane_s16(vget_low_s16(cb_128_h), consts, 0);
+      int32x4_t g_sub_y_hh = vmull_lane_s16(vget_high_s16(cb_128_h),
+                                            consts, 0);
+      g_sub_y_ll = vmlsl_lane_s16(g_sub_y_ll, vget_low_s16(cr_128_l),
+                                  consts, 1);
+      g_sub_y_lh = vmlsl_lane_s16(g_sub_y_lh, vget_high_s16(cr_128_l),
+                                  consts, 1);
+      g_sub_y_hl = vmlsl_lane_s16(g_sub_y_hl, vget_low_s16(cr_128_h),
+                                  consts, 1);
+      g_sub_y_hh = vmlsl_lane_s16(g_sub_y_hh, vget_high_s16(cr_128_h),
+                                  consts, 1);
+      /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+      int16x8_t g_sub_y_l = vcombine_s16(vrshrn_n_s32(g_sub_y_ll, 15),
+                                         vrshrn_n_s32(g_sub_y_lh, 15));
+      int16x8_t g_sub_y_h = vcombine_s16(vrshrn_n_s32(g_sub_y_hl, 15),
+                                         vrshrn_n_s32(g_sub_y_hh, 15));
+      /* Compute R-Y: 1.40200 * (Cr - 128) */
+      int16x8_t r_sub_y_l = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128_l, 1),
+                                               consts, 2);
+      int16x8_t r_sub_y_h = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128_h, 1),
+                                               consts, 2);
+      /* Compute B-Y: 1.77200 * (Cb - 128) */
+      int16x8_t b_sub_y_l = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128_l, 1),
+                                               consts, 3);
+      int16x8_t b_sub_y_h = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128_h, 1),
+                                               consts, 3);
+      /* Add Y. */
+      int16x8_t r_l =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y_l),
+                                       vget_low_u8(y)));
+      int16x8_t r_h =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y_h),
+                                       vget_high_u8(y)));
+      int16x8_t b_l =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y_l),
+                                       vget_low_u8(y)));
+      int16x8_t b_h =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y_h),
+                                       vget_high_u8(y)));
+      int16x8_t g_l =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y_l),
+                                       vget_low_u8(y)));
+      int16x8_t g_h =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y_h),
+                                       vget_high_u8(y)));
+
+#if RGB_PIXELSIZE == 4
+      uint8x16x4_t rgba;
+      /* Convert each component to unsigned and narrow, clamping to [0-255]. */
+      rgba.val[RGB_RED] = vcombine_u8(vqmovun_s16(r_l), vqmovun_s16(r_h));
+      rgba.val[RGB_GREEN] = vcombine_u8(vqmovun_s16(g_l), vqmovun_s16(g_h));
+      rgba.val[RGB_BLUE] = vcombine_u8(vqmovun_s16(b_l), vqmovun_s16(b_h));
+      /* Set alpha channel to opaque (0xFF). */
+      rgba.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
+      /* Store RGBA pixel data to memory. */
+      vst4q_u8(outptr, rgba);
+#elif RGB_PIXELSIZE == 3
+      uint8x16x3_t rgb;
+      /* Convert each component to unsigned and narrow, clamping to [0-255]. */
+      rgb.val[RGB_RED] = vcombine_u8(vqmovun_s16(r_l), vqmovun_s16(r_h));
+      rgb.val[RGB_GREEN] = vcombine_u8(vqmovun_s16(g_l), vqmovun_s16(g_h));
+      rgb.val[RGB_BLUE] = vcombine_u8(vqmovun_s16(b_l), vqmovun_s16(b_h));
+      /* Store RGB pixel data to memory. */
+      vst3q_u8(outptr, rgb);
+#else
+      /* Pack R, G, and B values in ratio 5:6:5. */
+      uint16x8_t rgb565_l = vqshluq_n_s16(r_l, 8);
+      rgb565_l = vsriq_n_u16(rgb565_l, vqshluq_n_s16(g_l, 8), 5);
+      rgb565_l = vsriq_n_u16(rgb565_l, vqshluq_n_s16(b_l, 8), 11);
+      uint16x8_t rgb565_h = vqshluq_n_s16(r_h, 8);
+      rgb565_h = vsriq_n_u16(rgb565_h, vqshluq_n_s16(g_h, 8), 5);
+      rgb565_h = vsriq_n_u16(rgb565_h, vqshluq_n_s16(b_h, 8), 11);
+      /* Store RGB pixel data to memory. */
+      vst1q_u16((uint16_t *)outptr, rgb565_l);
+      vst1q_u16(((uint16_t *)outptr) + 8, rgb565_h);
+#endif
+
+      /* Increment pointers. */
+      inptr0 += 16;
+      inptr1 += 16;
+      inptr2 += 16;
+      outptr += (RGB_PIXELSIZE * 16);
+    }
+
+    if (cols_remaining >= 8) {
+      uint8x8_t y  = vld1_u8(inptr0);
+      uint8x8_t cb = vld1_u8(inptr1);
+      uint8x8_t cr = vld1_u8(inptr2);
+      /* Subtract 128 from Cb and Cr. */
+      int16x8_t cr_128 =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+      int16x8_t cb_128 =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+      /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+      int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+      int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+      g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+      g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+      /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+      int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+                                       vrshrn_n_s32(g_sub_y_h, 15));
+      /* Compute R-Y: 1.40200 * (Cr - 128) */
+      int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1),
+                                             consts, 2);
+      /* Compute B-Y: 1.77200 * (Cb - 128) */
+      int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1),
+                                             consts, 3);
+      /* Add Y. */
+      int16x8_t r =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y));
+      int16x8_t b =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y));
+      int16x8_t g =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y));
+
+#if RGB_PIXELSIZE == 4
+      uint8x8x4_t rgba;
+      /* Convert each component to unsigned and narrow, clamping to [0-255]. */
+      rgba.val[RGB_RED] = vqmovun_s16(r);
+      rgba.val[RGB_GREEN] = vqmovun_s16(g);
+      rgba.val[RGB_BLUE] = vqmovun_s16(b);
+      /* Set alpha channel to opaque (0xFF). */
+      rgba.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+      /* Store RGBA pixel data to memory. */
+      vst4_u8(outptr, rgba);
+#elif RGB_PIXELSIZE == 3
+      uint8x8x3_t rgb;
+      /* Convert each component to unsigned and narrow, clamping to [0-255]. */
+      rgb.val[RGB_RED] = vqmovun_s16(r);
+      rgb.val[RGB_GREEN] = vqmovun_s16(g);
+      rgb.val[RGB_BLUE] = vqmovun_s16(b);
+      /* Store RGB pixel data to memory. */
+      vst3_u8(outptr, rgb);
+#else
+      /* Pack R, G, and B values in ratio 5:6:5. */
+      uint16x8_t rgb565 = vqshluq_n_s16(r, 8);
+      rgb565 = vsriq_n_u16(rgb565, vqshluq_n_s16(g, 8), 5);
+      rgb565 = vsriq_n_u16(rgb565, vqshluq_n_s16(b, 8), 11);
+      /* Store RGB pixel data to memory. */
+      vst1q_u16((uint16_t *)outptr, rgb565);
+#endif
+
+      /* Increment pointers. */
+      inptr0 += 8;
+      inptr1 += 8;
+      inptr2 += 8;
+      outptr += (RGB_PIXELSIZE * 8);
+      cols_remaining -= 8;
+    }
+
+    /* Handle the tail elements. */
+    if (cols_remaining > 0) {
+      uint8x8_t y  = vld1_u8(inptr0);
+      uint8x8_t cb = vld1_u8(inptr1);
+      uint8x8_t cr = vld1_u8(inptr2);
+      /* Subtract 128 from Cb and Cr. */
+      int16x8_t cr_128 =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+      int16x8_t cb_128 =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+      /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+      int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+      int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+      g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+      g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+      /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+      int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+                                       vrshrn_n_s32(g_sub_y_h, 15));
+      /* Compute R-Y: 1.40200 * (Cr - 128) */
+      int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1),
+                                             consts, 2);
+      /* Compute B-Y: 1.77200 * (Cb - 128) */
+      int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1),
+                                             consts, 3);
+      /* Add Y. */
+      int16x8_t r =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y));
+      int16x8_t b =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y));
+      int16x8_t g =
+        vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y));
+
+#if RGB_PIXELSIZE == 4
+      uint8x8x4_t rgba;
+      /* Convert each component to unsigned and narrow, clamping to [0-255]. */
+      rgba.val[RGB_RED] = vqmovun_s16(r);
+      rgba.val[RGB_GREEN] = vqmovun_s16(g);
+      rgba.val[RGB_BLUE] = vqmovun_s16(b);
+      /* Set alpha channel to opaque (0xFF). */
+      rgba.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+      /* Store RGBA pixel data to memory. */
+      switch (cols_remaining) {
+      case 7:
+        vst4_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgba, 6);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 6:
+        vst4_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgba, 5);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 5:
+        vst4_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgba, 4);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 4:
+        vst4_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgba, 3);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 3:
+        vst4_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgba, 2);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 2:
+        vst4_lane_u8(outptr + RGB_PIXELSIZE, rgba, 1);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 1:
+        vst4_lane_u8(outptr, rgba, 0);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      default:
+        break;
+      }
+#elif RGB_PIXELSIZE == 3
+      uint8x8x3_t rgb;
+      /* Convert each component to unsigned and narrow, clamping to [0-255]. */
+      rgb.val[RGB_RED] = vqmovun_s16(r);
+      rgb.val[RGB_GREEN] = vqmovun_s16(g);
+      rgb.val[RGB_BLUE] = vqmovun_s16(b);
+      /* Store RGB pixel data to memory. */
+      switch (cols_remaining) {
+      case 7:
+        vst3_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgb, 6);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 6:
+        vst3_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgb, 5);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 5:
+        vst3_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgb, 4);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 4:
+        vst3_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgb, 3);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 3:
+        vst3_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgb, 2);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 2:
+        vst3_lane_u8(outptr + RGB_PIXELSIZE, rgb, 1);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 1:
+        vst3_lane_u8(outptr, rgb, 0);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      default:
+        break;
+      }
+#else
+      /* Pack R, G, and B values in ratio 5:6:5. */
+      uint16x8_t rgb565 = vqshluq_n_s16(r, 8);
+      rgb565 = vsriq_n_u16(rgb565, vqshluq_n_s16(g, 8), 5);
+      rgb565 = vsriq_n_u16(rgb565, vqshluq_n_s16(b, 8), 11);
+      /* Store RGB565 pixel data to memory. */
+      switch (cols_remaining) {
+      case 7:
+        vst1q_lane_u16((uint16_t *)(outptr + 6 * RGB_PIXELSIZE), rgb565, 6);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 6:
+        vst1q_lane_u16((uint16_t *)(outptr + 5 * RGB_PIXELSIZE), rgb565, 5);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 5:
+        vst1q_lane_u16((uint16_t *)(outptr + 4 * RGB_PIXELSIZE), rgb565, 4);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 4:
+        vst1q_lane_u16((uint16_t *)(outptr + 3 * RGB_PIXELSIZE), rgb565, 3);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 3:
+        vst1q_lane_u16((uint16_t *)(outptr + 2 * RGB_PIXELSIZE), rgb565, 2);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 2:
+        vst1q_lane_u16((uint16_t *)(outptr + RGB_PIXELSIZE), rgb565, 1);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      case 1:
+        vst1q_lane_u16((uint16_t *)outptr, rgb565, 0);
+        FALLTHROUGH             /*FALLTHROUGH*/
+      default:
+        break;
+      }
+#endif
+    }
+  }
+}
@@ -0,0 +1,142 @@
+/*
+ * jdcolor-neon.c - colorspace conversion (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "jconfigint.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+
+#include <arm_neon.h>
+
+
+/* YCbCr -> RGB conversion constants */
+
+#define F_0_344  11277  /* 0.3441467 = 11277 * 2^-15 */
+#define F_0_714  23401  /* 0.7141418 = 23401 * 2^-15 */
+#define F_1_402  22971  /* 1.4020386 = 22971 * 2^-14 */
+#define F_1_772  29033  /* 1.7720337 = 29033 * 2^-14 */
+
+ALIGN(16) static const int16_t jsimd_ycc_rgb_convert_neon_consts[] = {
+  -F_0_344, F_0_714, F_1_402, F_1_772
+};
+
+
+/* Include inline routines for colorspace extensions. */
+
+#include "jdcolext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+
+#define RGB_RED  EXT_RGB_RED
+#define RGB_GREEN  EXT_RGB_GREEN
+#define RGB_BLUE  EXT_RGB_BLUE
+#define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+#define jsimd_ycc_rgb_convert_neon  jsimd_ycc_extrgb_convert_neon
+#include "jdcolext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_ycc_rgb_convert_neon
+
+#define RGB_RED  EXT_RGBX_RED
+#define RGB_GREEN  EXT_RGBX_GREEN
+#define RGB_BLUE  EXT_RGBX_BLUE
+#define RGB_ALPHA  3
+#define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+#define jsimd_ycc_rgb_convert_neon  jsimd_ycc_extrgbx_convert_neon
+#include "jdcolext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef jsimd_ycc_rgb_convert_neon
+
+#define RGB_RED  EXT_BGR_RED
+#define RGB_GREEN  EXT_BGR_GREEN
+#define RGB_BLUE  EXT_BGR_BLUE
+#define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+#define jsimd_ycc_rgb_convert_neon  jsimd_ycc_extbgr_convert_neon
+#include "jdcolext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_ycc_rgb_convert_neon
+
+#define RGB_RED  EXT_BGRX_RED
+#define RGB_GREEN  EXT_BGRX_GREEN
+#define RGB_BLUE  EXT_BGRX_BLUE
+#define RGB_ALPHA  3
+#define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+#define jsimd_ycc_rgb_convert_neon  jsimd_ycc_extbgrx_convert_neon
+#include "jdcolext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef jsimd_ycc_rgb_convert_neon
+
+#define RGB_RED  EXT_XBGR_RED
+#define RGB_GREEN  EXT_XBGR_GREEN
+#define RGB_BLUE  EXT_XBGR_BLUE
+#define RGB_ALPHA  0
+#define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+#define jsimd_ycc_rgb_convert_neon  jsimd_ycc_extxbgr_convert_neon
+#include "jdcolext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef jsimd_ycc_rgb_convert_neon
+
+#define RGB_RED  EXT_XRGB_RED
+#define RGB_GREEN  EXT_XRGB_GREEN
+#define RGB_BLUE  EXT_XRGB_BLUE
+#define RGB_ALPHA  0
+#define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+#define jsimd_ycc_rgb_convert_neon  jsimd_ycc_extxrgb_convert_neon
+#include "jdcolext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef jsimd_ycc_rgb_convert_neon
+
+/* YCbCr -> RGB565 Conversion */
+
+#define RGB_PIXELSIZE  2
+#define jsimd_ycc_rgb_convert_neon  jsimd_ycc_rgb565_convert_neon
+#include "jdcolext-neon.c"
+#undef RGB_PIXELSIZE
+#undef jsimd_ycc_rgb_convert_neon
@@ -0,0 +1,145 @@
+/*
+ * jdmerge-neon.c - merged upsampling/color conversion (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "jconfigint.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+
+#include <arm_neon.h>
+
+
+/* YCbCr -> RGB conversion constants */
+
+#define F_0_344  11277  /* 0.3441467 = 11277 * 2^-15 */
+#define F_0_714  23401  /* 0.7141418 = 23401 * 2^-15 */
+#define F_1_402  22971  /* 1.4020386 = 22971 * 2^-14 */
+#define F_1_772  29033  /* 1.7720337 = 29033 * 2^-14 */
+
+ALIGN(16) static const int16_t jsimd_ycc_rgb_convert_neon_consts[] = {
+  -F_0_344, F_0_714, F_1_402, F_1_772
+};
+
+
+/* Include inline routines for colorspace extensions. */
+
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+
+#define RGB_RED  EXT_RGB_RED
+#define RGB_GREEN  EXT_RGB_GREEN
+#define RGB_BLUE  EXT_RGB_BLUE
+#define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+#define jsimd_h2v1_merged_upsample_neon  jsimd_h2v1_extrgb_merged_upsample_neon
+#define jsimd_h2v2_merged_upsample_neon  jsimd_h2v2_extrgb_merged_upsample_neon
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_h2v1_merged_upsample_neon
+#undef jsimd_h2v2_merged_upsample_neon
+
+#define RGB_RED  EXT_RGBX_RED
+#define RGB_GREEN  EXT_RGBX_GREEN
+#define RGB_BLUE  EXT_RGBX_BLUE
+#define RGB_ALPHA  3
+#define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+#define jsimd_h2v1_merged_upsample_neon  jsimd_h2v1_extrgbx_merged_upsample_neon
+#define jsimd_h2v2_merged_upsample_neon  jsimd_h2v2_extrgbx_merged_upsample_neon
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef jsimd_h2v1_merged_upsample_neon
+#undef jsimd_h2v2_merged_upsample_neon
+
+#define RGB_RED  EXT_BGR_RED
+#define RGB_GREEN  EXT_BGR_GREEN
+#define RGB_BLUE  EXT_BGR_BLUE
+#define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+#define jsimd_h2v1_merged_upsample_neon  jsimd_h2v1_extbgr_merged_upsample_neon
+#define jsimd_h2v2_merged_upsample_neon  jsimd_h2v2_extbgr_merged_upsample_neon
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_PIXELSIZE
+#undef jsimd_h2v1_merged_upsample_neon
+#undef jsimd_h2v2_merged_upsample_neon
+
+#define RGB_RED  EXT_BGRX_RED
+#define RGB_GREEN  EXT_BGRX_GREEN
+#define RGB_BLUE  EXT_BGRX_BLUE
+#define RGB_ALPHA  3
+#define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+#define jsimd_h2v1_merged_upsample_neon  jsimd_h2v1_extbgrx_merged_upsample_neon
+#define jsimd_h2v2_merged_upsample_neon  jsimd_h2v2_extbgrx_merged_upsample_neon
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef jsimd_h2v1_merged_upsample_neon
+#undef jsimd_h2v2_merged_upsample_neon
+
+#define RGB_RED  EXT_XBGR_RED
+#define RGB_GREEN  EXT_XBGR_GREEN
+#define RGB_BLUE  EXT_XBGR_BLUE
+#define RGB_ALPHA  0
+#define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+#define jsimd_h2v1_merged_upsample_neon  jsimd_h2v1_extxbgr_merged_upsample_neon
+#define jsimd_h2v2_merged_upsample_neon  jsimd_h2v2_extxbgr_merged_upsample_neon
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef jsimd_h2v1_merged_upsample_neon
+#undef jsimd_h2v2_merged_upsample_neon
+
+#define RGB_RED  EXT_XRGB_RED
+#define RGB_GREEN  EXT_XRGB_GREEN
+#define RGB_BLUE  EXT_XRGB_BLUE
+#define RGB_ALPHA  0
+#define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+#define jsimd_h2v1_merged_upsample_neon  jsimd_h2v1_extxrgb_merged_upsample_neon
+#define jsimd_h2v2_merged_upsample_neon  jsimd_h2v2_extxrgb_merged_upsample_neon
+#include "jdmrgext-neon.c"
+#undef RGB_RED
+#undef RGB_GREEN
+#undef RGB_BLUE
+#undef RGB_ALPHA
+#undef RGB_PIXELSIZE
+#undef jsimd_h2v1_merged_upsample_neon
@@ -0,0 +1,723 @@
+/*
+ * jdmrgext-neon.c - merged upsampling/color conversion (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* This file is included by jdmerge-neon.c. */
+
+
+/* These routines combine simple (non-fancy, i.e. non-smooth) h2v1 or h2v2
+ * chroma upsampling and YCbCr -> RGB color conversion into a single function.
+ *
+ * As with the standalone functions, YCbCr -> RGB conversion is defined by the
+ * following equations:
+ *    R = Y                        + 1.40200 * (Cr - 128)
+ *    G = Y - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128)
+ *    B = Y + 1.77200 * (Cb - 128)
+ *
+ * Scaled integer constants are used to avoid floating-point arithmetic:
+ *    0.3441467 = 11277 * 2^-15
+ *    0.7141418 = 23401 * 2^-15
+ *    1.4020386 = 22971 * 2^-14
+ *    1.7720337 = 29033 * 2^-14
+ * These constants are defined in jdmerge-neon.c.
+ *
+ * To ensure correct results, rounding is used when descaling.
+ */
+
+/* Notes on safe memory access for merged upsampling/YCbCr -> RGB conversion
+ * routines:
+ *
+ * Input memory buffers can be safely overread up to the next multiple of
+ * ALIGN_SIZE bytes, since they are always allocated by alloc_sarray() in
+ * jmemmgr.c.
+ *
+ * The output buffer cannot safely be written beyond output_width, since
+ * output_buf points to a possibly unpadded row in the decompressed image
+ * buffer allocated by the calling program.
+ */
+
+/* Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+ */
+
+void jsimd_h2v1_merged_upsample_neon(JDIMENSION output_width,
+                                     JSAMPIMAGE input_buf,
+                                     JDIMENSION in_row_group_ctr,
+                                     JSAMPARRAY output_buf)
+{
+  JSAMPROW outptr;
+  /* Pointers to Y, Cb, and Cr data */
+  JSAMPROW inptr0, inptr1, inptr2;
+
+  const int16x4_t consts = vld1_s16(jsimd_ycc_rgb_convert_neon_consts);
+  const int16x8_t neg_128 = vdupq_n_s16(-128);
+
+  inptr0 = input_buf[0][in_row_group_ctr];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr = output_buf[0];
+
+  int cols_remaining = output_width;
+  for (; cols_remaining >= 16; cols_remaining -= 16) {
+    /* De-interleave Y component values into two separate vectors, one
+     * containing the component values with even-numbered indices and one
+     * containing the component values with odd-numbered indices.
+     */
+    uint8x8x2_t y = vld2_u8(inptr0);
+    uint8x8_t cb = vld1_u8(inptr1);
+    uint8x8_t cr = vld1_u8(inptr2);
+    /* Subtract 128 from Cb and Cr. */
+    int16x8_t cr_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+    int16x8_t cb_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+    /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+    int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+    int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+    g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+    g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+    /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+    int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+                                     vrshrn_n_s32(g_sub_y_h, 15));
+    /* Compute R-Y: 1.40200 * (Cr - 128) */
+    int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
+    /* Compute B-Y: 1.77200 * (Cb - 128) */
+    int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
+    /* Add the chroma-derived values (G-Y, R-Y, and B-Y) to both the "even" and
+     * "odd" Y component values.  This effectively upsamples the chroma
+     * components horizontally.
+     */
+    int16x8_t g_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y.val[0]));
+    int16x8_t r_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y.val[0]));
+    int16x8_t b_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y.val[0]));
+    int16x8_t g_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y.val[1]));
+    int16x8_t r_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y.val[1]));
+    int16x8_t b_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y.val[1]));
+    /* Convert each component to unsigned and narrow, clamping to [0-255].
+     * Re-interleave the "even" and "odd" component values.
+     */
+    uint8x8x2_t r = vzip_u8(vqmovun_s16(r_even), vqmovun_s16(r_odd));
+    uint8x8x2_t g = vzip_u8(vqmovun_s16(g_even), vqmovun_s16(g_odd));
+    uint8x8x2_t b = vzip_u8(vqmovun_s16(b_even), vqmovun_s16(b_odd));
+
+#ifdef RGB_ALPHA
+    uint8x16x4_t rgba;
+    rgba.val[RGB_RED] = vcombine_u8(r.val[0], r.val[1]);
+    rgba.val[RGB_GREEN] = vcombine_u8(g.val[0], g.val[1]);
+    rgba.val[RGB_BLUE] = vcombine_u8(b.val[0], b.val[1]);
+    /* Set alpha channel to opaque (0xFF). */
+    rgba.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
+    /* Store RGBA pixel data to memory. */
+    vst4q_u8(outptr, rgba);
+#else
+    uint8x16x3_t rgb;
+    rgb.val[RGB_RED] = vcombine_u8(r.val[0], r.val[1]);
+    rgb.val[RGB_GREEN] = vcombine_u8(g.val[0], g.val[1]);
+    rgb.val[RGB_BLUE] = vcombine_u8(b.val[0], b.val[1]);
+    /* Store RGB pixel data to memory. */
+    vst3q_u8(outptr, rgb);
+#endif
+
+    /* Increment pointers. */
+    inptr0 += 16;
+    inptr1 += 8;
+    inptr2 += 8;
+    outptr += (RGB_PIXELSIZE * 16);
+  }
+
+  if (cols_remaining > 0) {
+    /* De-interleave Y component values into two separate vectors, one
+     * containing the component values with even-numbered indices and one
+     * containing the component values with odd-numbered indices.
+     */
+    uint8x8x2_t y = vld2_u8(inptr0);
+    uint8x8_t cb = vld1_u8(inptr1);
+    uint8x8_t cr = vld1_u8(inptr2);
+    /* Subtract 128 from Cb and Cr. */
+    int16x8_t cr_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+    int16x8_t cb_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+    /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+    int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+    int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+    g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+    g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+    /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+    int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+                                     vrshrn_n_s32(g_sub_y_h, 15));
+    /* Compute R-Y: 1.40200 * (Cr - 128) */
+    int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
+    /* Compute B-Y: 1.77200 * (Cb - 128) */
+    int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
+    /* Add the chroma-derived values (G-Y, R-Y, and B-Y) to both the "even" and
+     * "odd" Y component values.  This effectively upsamples the chroma
+     * components horizontally.
+     */
+    int16x8_t g_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y.val[0]));
+    int16x8_t r_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y.val[0]));
+    int16x8_t b_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y.val[0]));
+    int16x8_t g_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y.val[1]));
+    int16x8_t r_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y.val[1]));
+    int16x8_t b_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y.val[1]));
+    /* Convert each component to unsigned and narrow, clamping to [0-255].
+     * Re-interleave the "even" and "odd" component values.
+     */
+    uint8x8x2_t r = vzip_u8(vqmovun_s16(r_even), vqmovun_s16(r_odd));
+    uint8x8x2_t g = vzip_u8(vqmovun_s16(g_even), vqmovun_s16(g_odd));
+    uint8x8x2_t b = vzip_u8(vqmovun_s16(b_even), vqmovun_s16(b_odd));
+
+#ifdef RGB_ALPHA
+    uint8x8x4_t rgba_h;
+    rgba_h.val[RGB_RED] = r.val[1];
+    rgba_h.val[RGB_GREEN] = g.val[1];
+    rgba_h.val[RGB_BLUE] = b.val[1];
+    /* Set alpha channel to opaque (0xFF). */
+    rgba_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+    uint8x8x4_t rgba_l;
+    rgba_l.val[RGB_RED] = r.val[0];
+    rgba_l.val[RGB_GREEN] = g.val[0];
+    rgba_l.val[RGB_BLUE] = b.val[0];
+    /* Set alpha channel to opaque (0xFF). */
+    rgba_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+    /* Store RGBA pixel data to memory. */
+    switch (cols_remaining) {
+    case 15:
+      vst4_lane_u8(outptr + 14 * RGB_PIXELSIZE, rgba_h, 6);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 14:
+      vst4_lane_u8(outptr + 13 * RGB_PIXELSIZE, rgba_h, 5);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 13:
+      vst4_lane_u8(outptr + 12 * RGB_PIXELSIZE, rgba_h, 4);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 12:
+      vst4_lane_u8(outptr + 11 * RGB_PIXELSIZE, rgba_h, 3);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 11:
+      vst4_lane_u8(outptr + 10 * RGB_PIXELSIZE, rgba_h, 2);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 10:
+      vst4_lane_u8(outptr + 9 * RGB_PIXELSIZE, rgba_h, 1);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 9:
+      vst4_lane_u8(outptr + 8 * RGB_PIXELSIZE, rgba_h, 0);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 8:
+      vst4_u8(outptr, rgba_l);
+      break;
+    case 7:
+      vst4_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgba_l, 6);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 6:
+      vst4_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgba_l, 5);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 5:
+      vst4_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgba_l, 4);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 4:
+      vst4_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgba_l, 3);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 3:
+      vst4_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgba_l, 2);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 2:
+      vst4_lane_u8(outptr + RGB_PIXELSIZE, rgba_l, 1);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 1:
+      vst4_lane_u8(outptr, rgba_l, 0);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    default:
+      break;
+    }
+#else
+    uint8x8x3_t rgb_h;
+    rgb_h.val[RGB_RED] = r.val[1];
+    rgb_h.val[RGB_GREEN] = g.val[1];
+    rgb_h.val[RGB_BLUE] = b.val[1];
+    uint8x8x3_t rgb_l;
+    rgb_l.val[RGB_RED] = r.val[0];
+    rgb_l.val[RGB_GREEN] = g.val[0];
+    rgb_l.val[RGB_BLUE] = b.val[0];
+    /* Store RGB pixel data to memory. */
+    switch (cols_remaining) {
+    case 15:
+      vst3_lane_u8(outptr + 14 * RGB_PIXELSIZE, rgb_h, 6);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 14:
+      vst3_lane_u8(outptr + 13 * RGB_PIXELSIZE, rgb_h, 5);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 13:
+      vst3_lane_u8(outptr + 12 * RGB_PIXELSIZE, rgb_h, 4);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 12:
+      vst3_lane_u8(outptr + 11 * RGB_PIXELSIZE, rgb_h, 3);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 11:
+      vst3_lane_u8(outptr + 10 * RGB_PIXELSIZE, rgb_h, 2);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 10:
+      vst3_lane_u8(outptr + 9 * RGB_PIXELSIZE, rgb_h, 1);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 9:
+      vst3_lane_u8(outptr + 8 * RGB_PIXELSIZE, rgb_h, 0);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 8:
+      vst3_u8(outptr, rgb_l);
+      break;
+    case 7:
+      vst3_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgb_l, 6);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 6:
+      vst3_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgb_l, 5);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 5:
+      vst3_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgb_l, 4);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 4:
+      vst3_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgb_l, 3);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 3:
+      vst3_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgb_l, 2);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 2:
+      vst3_lane_u8(outptr + RGB_PIXELSIZE, rgb_l, 1);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 1:
+      vst3_lane_u8(outptr, rgb_l, 0);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    default:
+      break;
+    }
+#endif
+  }
+}
+
+
+/* Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+ *
+ * See comments above for details regarding color conversion and safe memory
+ * access.
+ */
+
+void jsimd_h2v2_merged_upsample_neon(JDIMENSION output_width,
+                                     JSAMPIMAGE input_buf,
+                                     JDIMENSION in_row_group_ctr,
+                                     JSAMPARRAY output_buf)
+{
+  JSAMPROW outptr0, outptr1;
+  /* Pointers to Y (both rows), Cb, and Cr data */
+  JSAMPROW inptr0_0, inptr0_1, inptr1, inptr2;
+
+  const int16x4_t consts = vld1_s16(jsimd_ycc_rgb_convert_neon_consts);
+  const int16x8_t neg_128 = vdupq_n_s16(-128);
+
+  inptr0_0 = input_buf[0][in_row_group_ctr * 2];
+  inptr0_1 = input_buf[0][in_row_group_ctr * 2 + 1];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr0 = output_buf[0];
+  outptr1 = output_buf[1];
+
+  int cols_remaining = output_width;
+  for (; cols_remaining >= 16; cols_remaining -= 16) {
+    /* For each row, de-interleave Y component values into two separate
+     * vectors, one containing the component values with even-numbered indices
+     * and one containing the component values with odd-numbered indices.
+     */
+    uint8x8x2_t y0 = vld2_u8(inptr0_0);
+    uint8x8x2_t y1 = vld2_u8(inptr0_1);
+    uint8x8_t cb = vld1_u8(inptr1);
+    uint8x8_t cr = vld1_u8(inptr2);
+    /* Subtract 128 from Cb and Cr. */
+    int16x8_t cr_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+    int16x8_t cb_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+    /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+    int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+    int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+    g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+    g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+    /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+    int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+                                     vrshrn_n_s32(g_sub_y_h, 15));
+    /* Compute R-Y: 1.40200 * (Cr - 128) */
+    int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
+    /* Compute B-Y: 1.77200 * (Cb - 128) */
+    int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
+    /* For each row, add the chroma-derived values (G-Y, R-Y, and B-Y) to both
+     * the "even" and "odd" Y component values.  This effectively upsamples the
+     * chroma components both horizontally and vertically.
+     */
+    int16x8_t g0_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y0.val[0]));
+    int16x8_t r0_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y0.val[0]));
+    int16x8_t b0_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y0.val[0]));
+    int16x8_t g0_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y0.val[1]));
+    int16x8_t r0_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y0.val[1]));
+    int16x8_t b0_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y0.val[1]));
+    int16x8_t g1_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y1.val[0]));
+    int16x8_t r1_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y1.val[0]));
+    int16x8_t b1_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y1.val[0]));
+    int16x8_t g1_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y1.val[1]));
+    int16x8_t r1_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y1.val[1]));
+    int16x8_t b1_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y1.val[1]));
+    /* Convert each component to unsigned and narrow, clamping to [0-255].
+     * Re-interleave the "even" and "odd" component values.
+     */
+    uint8x8x2_t r0 = vzip_u8(vqmovun_s16(r0_even), vqmovun_s16(r0_odd));
+    uint8x8x2_t r1 = vzip_u8(vqmovun_s16(r1_even), vqmovun_s16(r1_odd));
+    uint8x8x2_t g0 = vzip_u8(vqmovun_s16(g0_even), vqmovun_s16(g0_odd));
+    uint8x8x2_t g1 = vzip_u8(vqmovun_s16(g1_even), vqmovun_s16(g1_odd));
+    uint8x8x2_t b0 = vzip_u8(vqmovun_s16(b0_even), vqmovun_s16(b0_odd));
+    uint8x8x2_t b1 = vzip_u8(vqmovun_s16(b1_even), vqmovun_s16(b1_odd));
+
+#ifdef RGB_ALPHA
+    uint8x16x4_t rgba0, rgba1;
+    rgba0.val[RGB_RED] = vcombine_u8(r0.val[0], r0.val[1]);
+    rgba1.val[RGB_RED] = vcombine_u8(r1.val[0], r1.val[1]);
+    rgba0.val[RGB_GREEN] = vcombine_u8(g0.val[0], g0.val[1]);
+    rgba1.val[RGB_GREEN] = vcombine_u8(g1.val[0], g1.val[1]);
+    rgba0.val[RGB_BLUE] = vcombine_u8(b0.val[0], b0.val[1]);
+    rgba1.val[RGB_BLUE] = vcombine_u8(b1.val[0], b1.val[1]);
+    /* Set alpha channel to opaque (0xFF). */
+    rgba0.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
+    rgba1.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
+    /* Store RGBA pixel data to memory. */
+    vst4q_u8(outptr0, rgba0);
+    vst4q_u8(outptr1, rgba1);
+#else
+    uint8x16x3_t rgb0, rgb1;
+    rgb0.val[RGB_RED] = vcombine_u8(r0.val[0], r0.val[1]);
+    rgb1.val[RGB_RED] = vcombine_u8(r1.val[0], r1.val[1]);
+    rgb0.val[RGB_GREEN] = vcombine_u8(g0.val[0], g0.val[1]);
+    rgb1.val[RGB_GREEN] = vcombine_u8(g1.val[0], g1.val[1]);
+    rgb0.val[RGB_BLUE] = vcombine_u8(b0.val[0], b0.val[1]);
+    rgb1.val[RGB_BLUE] = vcombine_u8(b1.val[0], b1.val[1]);
+    /* Store RGB pixel data to memory. */
+    vst3q_u8(outptr0, rgb0);
+    vst3q_u8(outptr1, rgb1);
+#endif
+
+    /* Increment pointers. */
+    inptr0_0 += 16;
+    inptr0_1 += 16;
+    inptr1 += 8;
+    inptr2 += 8;
+    outptr0 += (RGB_PIXELSIZE * 16);
+    outptr1 += (RGB_PIXELSIZE * 16);
+  }
+
+  if (cols_remaining > 0) {
+    /* For each row, de-interleave Y component values into two separate
+     * vectors, one containing the component values with even-numbered indices
+     * and one containing the component values with odd-numbered indices.
+     */
+    uint8x8x2_t y0 = vld2_u8(inptr0_0);
+    uint8x8x2_t y1 = vld2_u8(inptr0_1);
+    uint8x8_t cb = vld1_u8(inptr1);
+    uint8x8_t cr = vld1_u8(inptr2);
+    /* Subtract 128 from Cb and Cr. */
+    int16x8_t cr_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+    int16x8_t cb_128 =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+    /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+    int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+    int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+    g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+    g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+    /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+    int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+                                     vrshrn_n_s32(g_sub_y_h, 15));
+    /* Compute R-Y: 1.40200 * (Cr - 128) */
+    int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
+    /* Compute B-Y: 1.77200 * (Cb - 128) */
+    int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
+    /* For each row, add the chroma-derived values (G-Y, R-Y, and B-Y) to both
+     * the "even" and "odd" Y component values.  This effectively upsamples the
+     * chroma components both horizontally and vertically.
+     */
+    int16x8_t g0_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y0.val[0]));
+    int16x8_t r0_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y0.val[0]));
+    int16x8_t b0_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y0.val[0]));
+    int16x8_t g0_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y0.val[1]));
+    int16x8_t r0_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y0.val[1]));
+    int16x8_t b0_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y0.val[1]));
+    int16x8_t g1_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y1.val[0]));
+    int16x8_t r1_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y1.val[0]));
+    int16x8_t b1_even =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y1.val[0]));
+    int16x8_t g1_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+                                     y1.val[1]));
+    int16x8_t r1_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+                                     y1.val[1]));
+    int16x8_t b1_odd =
+      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+                                     y1.val[1]));
+    /* Convert each component to unsigned and narrow, clamping to [0-255].
+     * Re-interleave the "even" and "odd" component values.
+     */
+    uint8x8x2_t r0 = vzip_u8(vqmovun_s16(r0_even), vqmovun_s16(r0_odd));
+    uint8x8x2_t r1 = vzip_u8(vqmovun_s16(r1_even), vqmovun_s16(r1_odd));
+    uint8x8x2_t g0 = vzip_u8(vqmovun_s16(g0_even), vqmovun_s16(g0_odd));
+    uint8x8x2_t g1 = vzip_u8(vqmovun_s16(g1_even), vqmovun_s16(g1_odd));
+    uint8x8x2_t b0 = vzip_u8(vqmovun_s16(b0_even), vqmovun_s16(b0_odd));
+    uint8x8x2_t b1 = vzip_u8(vqmovun_s16(b1_even), vqmovun_s16(b1_odd));
+
+#ifdef RGB_ALPHA
+    uint8x8x4_t rgba0_h, rgba1_h;
+    rgba0_h.val[RGB_RED] = r0.val[1];
+    rgba1_h.val[RGB_RED] = r1.val[1];
+    rgba0_h.val[RGB_GREEN] = g0.val[1];
+    rgba1_h.val[RGB_GREEN] = g1.val[1];
+    rgba0_h.val[RGB_BLUE] = b0.val[1];
+    rgba1_h.val[RGB_BLUE] = b1.val[1];
+    /* Set alpha channel to opaque (0xFF). */
+    rgba0_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+    rgba1_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+
+    uint8x8x4_t rgba0_l, rgba1_l;
+    rgba0_l.val[RGB_RED] = r0.val[0];
+    rgba1_l.val[RGB_RED] = r1.val[0];
+    rgba0_l.val[RGB_GREEN] = g0.val[0];
+    rgba1_l.val[RGB_GREEN] = g1.val[0];
+    rgba0_l.val[RGB_BLUE] = b0.val[0];
+    rgba1_l.val[RGB_BLUE] = b1.val[0];
+    /* Set alpha channel to opaque (0xFF). */
+    rgba0_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+    rgba1_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+    /* Store RGBA pixel data to memory. */
+    switch (cols_remaining) {
+    case 15:
+      vst4_lane_u8(outptr0 + 14 * RGB_PIXELSIZE, rgba0_h, 6);
+      vst4_lane_u8(outptr1 + 14 * RGB_PIXELSIZE, rgba1_h, 6);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 14:
+      vst4_lane_u8(outptr0 + 13 * RGB_PIXELSIZE, rgba0_h, 5);
+      vst4_lane_u8(outptr1 + 13 * RGB_PIXELSIZE, rgba1_h, 5);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 13:
+      vst4_lane_u8(outptr0 + 12 * RGB_PIXELSIZE, rgba0_h, 4);
+      vst4_lane_u8(outptr1 + 12 * RGB_PIXELSIZE, rgba1_h, 4);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 12:
+      vst4_lane_u8(outptr0 + 11 * RGB_PIXELSIZE, rgba0_h, 3);
+      vst4_lane_u8(outptr1 + 11 * RGB_PIXELSIZE, rgba1_h, 3);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 11:
+      vst4_lane_u8(outptr0 + 10 * RGB_PIXELSIZE, rgba0_h, 2);
+      vst4_lane_u8(outptr1 + 10 * RGB_PIXELSIZE, rgba1_h, 2);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 10:
+      vst4_lane_u8(outptr0 + 9 * RGB_PIXELSIZE, rgba0_h, 1);
+      vst4_lane_u8(outptr1 + 9 * RGB_PIXELSIZE, rgba1_h, 1);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 9:
+      vst4_lane_u8(outptr0 + 8 * RGB_PIXELSIZE, rgba0_h, 0);
+      vst4_lane_u8(outptr1 + 8 * RGB_PIXELSIZE, rgba1_h, 0);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 8:
+      vst4_u8(outptr0, rgba0_l);
+      vst4_u8(outptr1, rgba1_l);
+      break;
+    case 7:
+      vst4_lane_u8(outptr0 + 6 * RGB_PIXELSIZE, rgba0_l, 6);
+      vst4_lane_u8(outptr1 + 6 * RGB_PIXELSIZE, rgba1_l, 6);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 6:
+      vst4_lane_u8(outptr0 + 5 * RGB_PIXELSIZE, rgba0_l, 5);
+      vst4_lane_u8(outptr1 + 5 * RGB_PIXELSIZE, rgba1_l, 5);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 5:
+      vst4_lane_u8(outptr0 + 4 * RGB_PIXELSIZE, rgba0_l, 4);
+      vst4_lane_u8(outptr1 + 4 * RGB_PIXELSIZE, rgba1_l, 4);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 4:
+      vst4_lane_u8(outptr0 + 3 * RGB_PIXELSIZE, rgba0_l, 3);
+      vst4_lane_u8(outptr1 + 3 * RGB_PIXELSIZE, rgba1_l, 3);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 3:
+      vst4_lane_u8(outptr0 + 2 * RGB_PIXELSIZE, rgba0_l, 2);
+      vst4_lane_u8(outptr1 + 2 * RGB_PIXELSIZE, rgba1_l, 2);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 2:
+      vst4_lane_u8(outptr0 + 1 * RGB_PIXELSIZE, rgba0_l, 1);
+      vst4_lane_u8(outptr1 + 1 * RGB_PIXELSIZE, rgba1_l, 1);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 1:
+      vst4_lane_u8(outptr0, rgba0_l, 0);
+      vst4_lane_u8(outptr1, rgba1_l, 0);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    default:
+      break;
+    }
+#else
+    uint8x8x3_t rgb0_h, rgb1_h;
+    rgb0_h.val[RGB_RED] = r0.val[1];
+    rgb1_h.val[RGB_RED] = r1.val[1];
+    rgb0_h.val[RGB_GREEN] = g0.val[1];
+    rgb1_h.val[RGB_GREEN] = g1.val[1];
+    rgb0_h.val[RGB_BLUE] = b0.val[1];
+    rgb1_h.val[RGB_BLUE] = b1.val[1];
+
+    uint8x8x3_t rgb0_l, rgb1_l;
+    rgb0_l.val[RGB_RED] = r0.val[0];
+    rgb1_l.val[RGB_RED] = r1.val[0];
+    rgb0_l.val[RGB_GREEN] = g0.val[0];
+    rgb1_l.val[RGB_GREEN] = g1.val[0];
+    rgb0_l.val[RGB_BLUE] = b0.val[0];
+    rgb1_l.val[RGB_BLUE] = b1.val[0];
+    /* Store RGB pixel data to memory. */
+    switch (cols_remaining) {
+    case 15:
+      vst3_lane_u8(outptr0 + 14 * RGB_PIXELSIZE, rgb0_h, 6);
+      vst3_lane_u8(outptr1 + 14 * RGB_PIXELSIZE, rgb1_h, 6);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 14:
+      vst3_lane_u8(outptr0 + 13 * RGB_PIXELSIZE, rgb0_h, 5);
+      vst3_lane_u8(outptr1 + 13 * RGB_PIXELSIZE, rgb1_h, 5);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 13:
+      vst3_lane_u8(outptr0 + 12 * RGB_PIXELSIZE, rgb0_h, 4);
+      vst3_lane_u8(outptr1 + 12 * RGB_PIXELSIZE, rgb1_h, 4);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 12:
+      vst3_lane_u8(outptr0 + 11 * RGB_PIXELSIZE, rgb0_h, 3);
+      vst3_lane_u8(outptr1 + 11 * RGB_PIXELSIZE, rgb1_h, 3);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 11:
+      vst3_lane_u8(outptr0 + 10 * RGB_PIXELSIZE, rgb0_h, 2);
+      vst3_lane_u8(outptr1 + 10 * RGB_PIXELSIZE, rgb1_h, 2);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 10:
+      vst3_lane_u8(outptr0 + 9 * RGB_PIXELSIZE, rgb0_h, 1);
+      vst3_lane_u8(outptr1 + 9 * RGB_PIXELSIZE, rgb1_h, 1);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 9:
+      vst3_lane_u8(outptr0 + 8 * RGB_PIXELSIZE, rgb0_h, 0);
+      vst3_lane_u8(outptr1 + 8 * RGB_PIXELSIZE, rgb1_h, 0);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 8:
+      vst3_u8(outptr0, rgb0_l);
+      vst3_u8(outptr1, rgb1_l);
+      break;
+    case 7:
+      vst3_lane_u8(outptr0 + 6 * RGB_PIXELSIZE, rgb0_l, 6);
+      vst3_lane_u8(outptr1 + 6 * RGB_PIXELSIZE, rgb1_l, 6);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 6:
+      vst3_lane_u8(outptr0 + 5 * RGB_PIXELSIZE, rgb0_l, 5);
+      vst3_lane_u8(outptr1 + 5 * RGB_PIXELSIZE, rgb1_l, 5);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 5:
+      vst3_lane_u8(outptr0 + 4 * RGB_PIXELSIZE, rgb0_l, 4);
+      vst3_lane_u8(outptr1 + 4 * RGB_PIXELSIZE, rgb1_l, 4);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 4:
+      vst3_lane_u8(outptr0 + 3 * RGB_PIXELSIZE, rgb0_l, 3);
+      vst3_lane_u8(outptr1 + 3 * RGB_PIXELSIZE, rgb1_l, 3);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 3:
+      vst3_lane_u8(outptr0 + 2 * RGB_PIXELSIZE, rgb0_l, 2);
+      vst3_lane_u8(outptr1 + 2 * RGB_PIXELSIZE, rgb1_l, 2);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 2:
+      vst3_lane_u8(outptr0 + 1 * RGB_PIXELSIZE, rgb0_l, 1);
+      vst3_lane_u8(outptr1 + 1 * RGB_PIXELSIZE, rgb1_l, 1);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    case 1:
+      vst3_lane_u8(outptr0, rgb0_l, 0);
+      vst3_lane_u8(outptr1, rgb1_l, 0);
+      FALLTHROUGH               /*FALLTHROUGH*/
+    default:
+      break;
+    }
+#endif
+  }
+}
@@ -0,0 +1,569 @@
+/*
+ * jdsample-neon.c - upsampling (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+
+#include <arm_neon.h>
+
+
+/* The diagram below shows a row of samples produced by h2v1 downsampling.
+ *
+ *                s0        s1        s2
+ *            +---------+---------+---------+
+ *            |         |         |         |
+ *            | p0   p1 | p2   p3 | p4   p5 |
+ *            |         |         |         |
+ *            +---------+---------+---------+
+ *
+ * Samples s0-s2 were created by averaging the original pixel component values
+ * centered at positions p0-p5 above.  To approximate those original pixel
+ * component values, we proportionally blend the adjacent samples in each row.
+ *
+ * An upsampled pixel component value is computed by blending the sample
+ * containing the pixel center with the nearest neighboring sample, in the
+ * ratio 3:1.  For example:
+ *     p1(upsampled) = 3/4 * s0 + 1/4 * s1
+ *     p2(upsampled) = 3/4 * s1 + 1/4 * s0
+ * When computing the first and last pixel component values in the row, there
+ * is no adjacent sample to blend, so:
+ *     p0(upsampled) = s0
+ *     p5(upsampled) = s2
+ */
+
+void jsimd_h2v1_fancy_upsample_neon(int max_v_samp_factor,
+                                    JDIMENSION downsampled_width,
+                                    JSAMPARRAY input_data,
+                                    JSAMPARRAY *output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  JSAMPROW inptr, outptr;
+  int inrow;
+  unsigned colctr;
+  /* Set up constants. */
+  const uint16x8_t one_u16 = vdupq_n_u16(1);
+  const uint8x8_t three_u8 = vdup_n_u8(3);
+
+  for (inrow = 0; inrow < max_v_samp_factor; inrow++) {
+    inptr = input_data[inrow];
+    outptr = output_data[inrow];
+    /* First pixel component value in this row of the original image */
+    *outptr = (JSAMPLE)GETJSAMPLE(*inptr);
+
+    /*    3/4 * containing sample + 1/4 * nearest neighboring sample
+     * For p1: containing sample = s0, nearest neighboring sample = s1
+     * For p2: containing sample = s1, nearest neighboring sample = s0
+     */
+    uint8x16_t s0 = vld1q_u8(inptr);
+    uint8x16_t s1 = vld1q_u8(inptr + 1);
+    /* Multiplication makes vectors twice as wide.  '_l' and '_h' suffixes
+     * denote low half and high half respectively.
+     */
+    uint16x8_t s1_add_3s0_l =
+      vmlal_u8(vmovl_u8(vget_low_u8(s1)), vget_low_u8(s0), three_u8);
+    uint16x8_t s1_add_3s0_h =
+      vmlal_u8(vmovl_u8(vget_high_u8(s1)), vget_high_u8(s0), three_u8);
+    uint16x8_t s0_add_3s1_l =
+      vmlal_u8(vmovl_u8(vget_low_u8(s0)), vget_low_u8(s1), three_u8);
+    uint16x8_t s0_add_3s1_h =
+      vmlal_u8(vmovl_u8(vget_high_u8(s0)), vget_high_u8(s1), three_u8);
+    /* Add ordered dithering bias to odd pixel values. */
+    s0_add_3s1_l = vaddq_u16(s0_add_3s1_l, one_u16);
+    s0_add_3s1_h = vaddq_u16(s0_add_3s1_h, one_u16);
+
+    /* The offset is initially 1, because the first pixel component has already
+     * been stored.  However, in subsequent iterations of the SIMD loop, this
+     * offset is (2 * colctr - 1) to stay within the bounds of the sample
+     * buffers without having to resort to a slow scalar tail case for the last
+     * (downsampled_width % 16) samples.  See "Creation of 2-D sample arrays"
+     * in jmemmgr.c for more details.
+     */
+    unsigned outptr_offset = 1;
+    uint8x16x2_t output_pixels;
+
+    /* We use software pipelining to maximise performance.  The code indented
+     * an extra two spaces begins the next iteration of the loop.
+     */
+    for (colctr = 16; colctr < downsampled_width; colctr += 16) {
+
+        s0 = vld1q_u8(inptr + colctr - 1);
+        s1 = vld1q_u8(inptr + colctr);
+
+      /* Right-shift by 2 (divide by 4), narrow to 8-bit, and combine. */
+      output_pixels.val[0] = vcombine_u8(vrshrn_n_u16(s1_add_3s0_l, 2),
+                                         vrshrn_n_u16(s1_add_3s0_h, 2));
+      output_pixels.val[1] = vcombine_u8(vshrn_n_u16(s0_add_3s1_l, 2),
+                                         vshrn_n_u16(s0_add_3s1_h, 2));
+
+        /* Multiplication makes vectors twice as wide.  '_l' and '_h' suffixes
+         * denote low half and high half respectively.
+         */
+        s1_add_3s0_l =
+          vmlal_u8(vmovl_u8(vget_low_u8(s1)), vget_low_u8(s0), three_u8);
+        s1_add_3s0_h =
+          vmlal_u8(vmovl_u8(vget_high_u8(s1)), vget_high_u8(s0), three_u8);
+        s0_add_3s1_l =
+          vmlal_u8(vmovl_u8(vget_low_u8(s0)), vget_low_u8(s1), three_u8);
+        s0_add_3s1_h =
+          vmlal_u8(vmovl_u8(vget_high_u8(s0)), vget_high_u8(s1), three_u8);
+        /* Add ordered dithering bias to odd pixel values. */
+        s0_add_3s1_l = vaddq_u16(s0_add_3s1_l, one_u16);
+        s0_add_3s1_h = vaddq_u16(s0_add_3s1_h, one_u16);
+
+      /* Store pixel component values to memory. */
+      vst2q_u8(outptr + outptr_offset, output_pixels);
+      outptr_offset = 2 * colctr - 1;
+    }
+
+    /* Complete the last iteration of the loop. */
+
+    /* Right-shift by 2 (divide by 4), narrow to 8-bit, and combine. */
+    output_pixels.val[0] = vcombine_u8(vrshrn_n_u16(s1_add_3s0_l, 2),
+                                       vrshrn_n_u16(s1_add_3s0_h, 2));
+    output_pixels.val[1] = vcombine_u8(vshrn_n_u16(s0_add_3s1_l, 2),
+                                       vshrn_n_u16(s0_add_3s1_h, 2));
+    /* Store pixel component values to memory. */
+    vst2q_u8(outptr + outptr_offset, output_pixels);
+
+    /* Last pixel component value in this row of the original image */
+    outptr[2 * downsampled_width - 1] =
+      GETJSAMPLE(inptr[downsampled_width - 1]);
+  }
+}
+
+
+/* The diagram below shows an array of samples produced by h2v2 downsampling.
+ *
+ *                s0        s1        s2
+ *            +---------+---------+---------+
+ *            | p0   p1 | p2   p3 | p4   p5 |
+ *       sA   |         |         |         |
+ *            | p6   p7 | p8   p9 | p10  p11|
+ *            +---------+---------+---------+
+ *            | p12  p13| p14  p15| p16  p17|
+ *       sB   |         |         |         |
+ *            | p18  p19| p20  p21| p22  p23|
+ *            +---------+---------+---------+
+ *            | p24  p25| p26  p27| p28  p29|
+ *       sC   |         |         |         |
+ *            | p30  p31| p32  p33| p34  p35|
+ *            +---------+---------+---------+
+ *
+ * Samples s0A-s2C were created by averaging the original pixel component
+ * values centered at positions p0-p35 above.  To approximate one of those
+ * original pixel component values, we proportionally blend the sample
+ * containing the pixel center with the nearest neighboring samples in each
+ * row, column, and diagonal.
+ *
+ * An upsampled pixel component value is computed by first blending the sample
+ * containing the pixel center with the nearest neighboring samples in the
+ * same column, in the ratio 3:1, and then blending each column sum with the
+ * nearest neighboring column sum, in the ratio 3:1.  For example:
+ *     p14(upsampled) = 3/4 * (3/4 * s1B + 1/4 * s1A) +
+ *                      1/4 * (3/4 * s0B + 1/4 * s0A)
+ *                    = 9/16 * s1B + 3/16 * s1A + 3/16 * s0B + 1/16 * s0A
+ * When computing the first and last pixel component values in the row, there
+ * is no horizontally adjacent sample to blend, so:
+ *     p12(upsampled) = 3/4 * s0B + 1/4 * s0A
+ *     p23(upsampled) = 3/4 * s2B + 1/4 * s2C
+ * When computing the first and last pixel component values in the column,
+ * there is no vertically adjacent sample to blend, so:
+ *     p2(upsampled) = 3/4 * s1A + 1/4 * s0A
+ *     p33(upsampled) = 3/4 * s1C + 1/4 * s2C
+ * When computing the corner pixel component values, there is no adjacent
+ * sample to blend, so:
+ *     p0(upsampled) = s0A
+ *     p35(upsampled) = s2C
+ */
+
+void jsimd_h2v2_fancy_upsample_neon(int max_v_samp_factor,
+                                    JDIMENSION downsampled_width,
+                                    JSAMPARRAY input_data,
+                                    JSAMPARRAY *output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  JSAMPROW inptr0, inptr1, inptr2, outptr0, outptr1;
+  int inrow, outrow;
+  unsigned colctr;
+  /* Set up constants. */
+  const uint16x8_t seven_u16 = vdupq_n_u16(7);
+  const uint8x8_t three_u8 = vdup_n_u8(3);
+  const uint16x8_t three_u16 = vdupq_n_u16(3);
+
+  inrow = outrow = 0;
+  while (outrow < max_v_samp_factor) {
+    inptr0 = input_data[inrow - 1];
+    inptr1 = input_data[inrow];
+    inptr2 = input_data[inrow + 1];
+    /* Suffixes 0 and 1 denote the upper and lower rows of output pixels,
+     * respectively.
+     */
+    outptr0 = output_data[outrow++];
+    outptr1 = output_data[outrow++];
+
+    /* First pixel component value in this row of the original image */
+    int s0colsum0 = GETJSAMPLE(*inptr1) * 3 + GETJSAMPLE(*inptr0);
+    *outptr0 = (JSAMPLE)((s0colsum0 * 4 + 8) >> 4);
+    int s0colsum1 = GETJSAMPLE(*inptr1) * 3 + GETJSAMPLE(*inptr2);
+    *outptr1 = (JSAMPLE)((s0colsum1 * 4 + 8) >> 4);
+
+    /* Step 1: Blend samples vertically in columns s0 and s1.
+     * Leave the divide by 4 until the end, when it can be done for both
+     * dimensions at once, right-shifting by 4.
+     */
+
+    /* Load and compute s0colsum0 and s0colsum1. */
+    uint8x16_t s0A = vld1q_u8(inptr0);
+    uint8x16_t s0B = vld1q_u8(inptr1);
+    uint8x16_t s0C = vld1q_u8(inptr2);
+    /* Multiplication makes vectors twice as wide.  '_l' and '_h' suffixes
+     * denote low half and high half respectively.
+     */
+    uint16x8_t s0colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s0A)),
+                                      vget_low_u8(s0B), three_u8);
+    uint16x8_t s0colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s0A)),
+                                      vget_high_u8(s0B), three_u8);
+    uint16x8_t s0colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s0C)),
+                                      vget_low_u8(s0B), three_u8);
+    uint16x8_t s0colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s0C)),
+                                      vget_high_u8(s0B), three_u8);
+    /* Load and compute s1colsum0 and s1colsum1. */
+    uint8x16_t s1A = vld1q_u8(inptr0 + 1);
+    uint8x16_t s1B = vld1q_u8(inptr1 + 1);
+    uint8x16_t s1C = vld1q_u8(inptr2 + 1);
+    uint16x8_t s1colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s1A)),
+                                      vget_low_u8(s1B), three_u8);
+    uint16x8_t s1colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s1A)),
+                                      vget_high_u8(s1B), three_u8);
+    uint16x8_t s1colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s1C)),
+                                      vget_low_u8(s1B), three_u8);
+    uint16x8_t s1colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s1C)),
+                                      vget_high_u8(s1B), three_u8);
+
+    /* Step 2: Blend the already-blended columns. */
+
+    uint16x8_t output0_p1_l = vmlaq_u16(s1colsum0_l, s0colsum0_l, three_u16);
+    uint16x8_t output0_p1_h = vmlaq_u16(s1colsum0_h, s0colsum0_h, three_u16);
+    uint16x8_t output0_p2_l = vmlaq_u16(s0colsum0_l, s1colsum0_l, three_u16);
+    uint16x8_t output0_p2_h = vmlaq_u16(s0colsum0_h, s1colsum0_h, three_u16);
+    uint16x8_t output1_p1_l = vmlaq_u16(s1colsum1_l, s0colsum1_l, three_u16);
+    uint16x8_t output1_p1_h = vmlaq_u16(s1colsum1_h, s0colsum1_h, three_u16);
+    uint16x8_t output1_p2_l = vmlaq_u16(s0colsum1_l, s1colsum1_l, three_u16);
+    uint16x8_t output1_p2_h = vmlaq_u16(s0colsum1_h, s1colsum1_h, three_u16);
+    /* Add ordered dithering bias to odd pixel values. */
+    output0_p1_l = vaddq_u16(output0_p1_l, seven_u16);
+    output0_p1_h = vaddq_u16(output0_p1_h, seven_u16);
+    output1_p1_l = vaddq_u16(output1_p1_l, seven_u16);
+    output1_p1_h = vaddq_u16(output1_p1_h, seven_u16);
+    /* Right-shift by 4 (divide by 16), narrow to 8-bit, and combine. */
+    uint8x16x2_t output_pixels0 = { {
+      vcombine_u8(vshrn_n_u16(output0_p1_l, 4), vshrn_n_u16(output0_p1_h, 4)),
+      vcombine_u8(vrshrn_n_u16(output0_p2_l, 4), vrshrn_n_u16(output0_p2_h, 4))
+    } };
+    uint8x16x2_t output_pixels1 = { {
+      vcombine_u8(vshrn_n_u16(output1_p1_l, 4), vshrn_n_u16(output1_p1_h, 4)),
+      vcombine_u8(vrshrn_n_u16(output1_p2_l, 4), vrshrn_n_u16(output1_p2_h, 4))
+    } };
+
+    /* Store pixel component values to memory.
+     * The minimum size of the output buffer for each row is 64 bytes => no
+     * need to worry about buffer overflow here.  See "Creation of 2-D sample
+     * arrays" in jmemmgr.c for more details.
+     */
+    vst2q_u8(outptr0 + 1, output_pixels0);
+    vst2q_u8(outptr1 + 1, output_pixels1);
+
+    /* The first pixel of the image shifted our loads and stores by one byte.
+     * We have to re-align on a 32-byte boundary at some point before the end
+     * of the row (we do it now on the 32/33 pixel boundary) to stay within the
+     * bounds of the sample buffers without having to resort to a slow scalar
+     * tail case for the last (downsampled_width % 16) samples.  See "Creation
+     * of 2-D sample arrays" in jmemmgr.c for more details.
+     */
+    for (colctr = 16; colctr < downsampled_width; colctr += 16) {
+      /* Step 1: Blend samples vertically in columns s0 and s1. */
+
+      /* Load and compute s0colsum0 and s0colsum1. */
+      s0A = vld1q_u8(inptr0 + colctr - 1);
+      s0B = vld1q_u8(inptr1 + colctr - 1);
+      s0C = vld1q_u8(inptr2 + colctr - 1);
+      s0colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s0A)), vget_low_u8(s0B),
+                             three_u8);
+      s0colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s0A)), vget_high_u8(s0B),
+                             three_u8);
+      s0colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s0C)), vget_low_u8(s0B),
+                             three_u8);
+      s0colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s0C)), vget_high_u8(s0B),
+                             three_u8);
+      /* Load and compute s1colsum0 and s1colsum1. */
+      s1A = vld1q_u8(inptr0 + colctr);
+      s1B = vld1q_u8(inptr1 + colctr);
+      s1C = vld1q_u8(inptr2 + colctr);
+      s1colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s1A)), vget_low_u8(s1B),
+                             three_u8);
+      s1colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s1A)), vget_high_u8(s1B),
+                             three_u8);
+      s1colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s1C)), vget_low_u8(s1B),
+                             three_u8);
+      s1colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s1C)), vget_high_u8(s1B),
+                             three_u8);
+
+      /* Step 2: Blend the already-blended columns. */
+
+      output0_p1_l = vmlaq_u16(s1colsum0_l, s0colsum0_l, three_u16);
+      output0_p1_h = vmlaq_u16(s1colsum0_h, s0colsum0_h, three_u16);
+      output0_p2_l = vmlaq_u16(s0colsum0_l, s1colsum0_l, three_u16);
+      output0_p2_h = vmlaq_u16(s0colsum0_h, s1colsum0_h, three_u16);
+      output1_p1_l = vmlaq_u16(s1colsum1_l, s0colsum1_l, three_u16);
+      output1_p1_h = vmlaq_u16(s1colsum1_h, s0colsum1_h, three_u16);
+      output1_p2_l = vmlaq_u16(s0colsum1_l, s1colsum1_l, three_u16);
+      output1_p2_h = vmlaq_u16(s0colsum1_h, s1colsum1_h, three_u16);
+      /* Add ordered dithering bias to odd pixel values. */
+      output0_p1_l = vaddq_u16(output0_p1_l, seven_u16);
+      output0_p1_h = vaddq_u16(output0_p1_h, seven_u16);
+      output1_p1_l = vaddq_u16(output1_p1_l, seven_u16);
+      output1_p1_h = vaddq_u16(output1_p1_h, seven_u16);
+      /* Right-shift by 4 (divide by 16), narrow to 8-bit, and combine. */
+      output_pixels0.val[0] = vcombine_u8(vshrn_n_u16(output0_p1_l, 4),
+                                          vshrn_n_u16(output0_p1_h, 4));
+      output_pixels0.val[1] = vcombine_u8(vrshrn_n_u16(output0_p2_l, 4),
+                                          vrshrn_n_u16(output0_p2_h, 4));
+      output_pixels1.val[0] = vcombine_u8(vshrn_n_u16(output1_p1_l, 4),
+                                          vshrn_n_u16(output1_p1_h, 4));
+      output_pixels1.val[1] = vcombine_u8(vrshrn_n_u16(output1_p2_l, 4),
+                                          vrshrn_n_u16(output1_p2_h, 4));
+      /* Store pixel component values to memory. */
+      vst2q_u8(outptr0 + 2 * colctr - 1, output_pixels0);
+      vst2q_u8(outptr1 + 2 * colctr - 1, output_pixels1);
+    }
+
+    /* Last pixel component value in this row of the original image */
+    int s1colsum0 = GETJSAMPLE(inptr1[downsampled_width - 1]) * 3 +
+                    GETJSAMPLE(inptr0[downsampled_width - 1]);
+    outptr0[2 * downsampled_width - 1] = (JSAMPLE)((s1colsum0 * 4 + 7) >> 4);
+    int s1colsum1 = GETJSAMPLE(inptr1[downsampled_width - 1]) * 3 +
+                    GETJSAMPLE(inptr2[downsampled_width - 1]);
+    outptr1[2 * downsampled_width - 1] = (JSAMPLE)((s1colsum1 * 4 + 7) >> 4);
+    inrow++;
+  }
+}
+
+
+/* The diagram below shows a column of samples produced by h1v2 downsampling
+ * (or by losslessly rotating or transposing an h2v1-downsampled image.)
+ *
+ *            +---------+
+ *            |   p0    |
+ *     sA     |         |
+ *            |   p1    |
+ *            +---------+
+ *            |   p2    |
+ *     sB     |         |
+ *            |   p3    |
+ *            +---------+
+ *            |   p4    |
+ *     sC     |         |
+ *            |   p5    |
+ *            +---------+
+ *
+ * Samples sA-sC were created by averaging the original pixel component values
+ * centered at positions p0-p5 above.  To approximate those original pixel
+ * component values, we proportionally blend the adjacent samples in each
+ * column.
+ *
+ * An upsampled pixel component value is computed by blending the sample
+ * containing the pixel center with the nearest neighboring sample, in the
+ * ratio 3:1.  For example:
+ *     p1(upsampled) = 3/4 * sA + 1/4 * sB
+ *     p2(upsampled) = 3/4 * sB + 1/4 * sA
+ * When computing the first and last pixel component values in the column,
+ * there is no adjacent sample to blend, so:
+ *     p0(upsampled) = sA
+ *     p5(upsampled) = sC
+ */
+
+void jsimd_h1v2_fancy_upsample_neon(int max_v_samp_factor,
+                                    JDIMENSION downsampled_width,
+                                    JSAMPARRAY input_data,
+                                    JSAMPARRAY *output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  JSAMPROW inptr0, inptr1, inptr2, outptr0, outptr1;
+  int inrow, outrow;
+  unsigned colctr;
+  /* Set up constants. */
+  const uint16x8_t one_u16 = vdupq_n_u16(1);
+  const uint8x8_t three_u8 = vdup_n_u8(3);
+
+  inrow = outrow = 0;
+  while (outrow < max_v_samp_factor) {
+    inptr0 = input_data[inrow - 1];
+    inptr1 = input_data[inrow];
+    inptr2 = input_data[inrow + 1];
+    /* Suffixes 0 and 1 denote the upper and lower rows of output pixels,
+     * respectively.
+     */
+    outptr0 = output_data[outrow++];
+    outptr1 = output_data[outrow++];
+    inrow++;
+
+    /* The size of the input and output buffers is always a multiple of 32
+     * bytes => no need to worry about buffer overflow when reading/writing
+     * memory.  See "Creation of 2-D sample arrays" in jmemmgr.c for more
+     * details.
+     */
+    for (colctr = 0; colctr < downsampled_width; colctr += 16) {
+      /* Load samples. */
+      uint8x16_t sA = vld1q_u8(inptr0 + colctr);
+      uint8x16_t sB = vld1q_u8(inptr1 + colctr);
+      uint8x16_t sC = vld1q_u8(inptr2 + colctr);
+      /* Blend samples vertically. */
+      uint16x8_t colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(sA)),
+                                      vget_low_u8(sB), three_u8);
+      uint16x8_t colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(sA)),
+                                      vget_high_u8(sB), three_u8);
+      uint16x8_t colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(sC)),
+                                      vget_low_u8(sB), three_u8);
+      uint16x8_t colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(sC)),
+                                      vget_high_u8(sB), three_u8);
+      /* Add ordered dithering bias to pixel values in even output rows. */
+      colsum0_l = vaddq_u16(colsum0_l, one_u16);
+      colsum0_h = vaddq_u16(colsum0_h, one_u16);
+      /* Right-shift by 2 (divide by 4), narrow to 8-bit, and combine. */
+      uint8x16_t output_pixels0 = vcombine_u8(vshrn_n_u16(colsum0_l, 2),
+                                              vshrn_n_u16(colsum0_h, 2));
+      uint8x16_t output_pixels1 = vcombine_u8(vrshrn_n_u16(colsum1_l, 2),
+                                              vrshrn_n_u16(colsum1_h, 2));
+      /* Store pixel component values to memory. */
+      vst1q_u8(outptr0 + colctr, output_pixels0);
+      vst1q_u8(outptr1 + colctr, output_pixels1);
+    }
+  }
+}
+
+
+/* The diagram below shows a row of samples produced by h2v1 downsampling.
+ *
+ *                s0        s1
+ *            +---------+---------+
+ *            |         |         |
+ *            | p0   p1 | p2   p3 |
+ *            |         |         |
+ *            +---------+---------+
+ *
+ * Samples s0 and s1 were created by averaging the original pixel component
+ * values centered at positions p0-p3 above.  To approximate those original
+ * pixel component values, we duplicate the samples horizontally:
+ *     p0(upsampled) = p1(upsampled) = s0
+ *     p2(upsampled) = p3(upsampled) = s1
+ */
+
+void jsimd_h2v1_upsample_neon(int max_v_samp_factor, JDIMENSION output_width,
+                              JSAMPARRAY input_data,
+                              JSAMPARRAY *output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  JSAMPROW inptr, outptr;
+  int inrow;
+  unsigned colctr;
+
+  for (inrow = 0; inrow < max_v_samp_factor; inrow++) {
+    inptr = input_data[inrow];
+    outptr = output_data[inrow];
+    for (colctr = 0; 2 * colctr < output_width; colctr += 16) {
+      uint8x16_t samples = vld1q_u8(inptr + colctr);
+      /* Duplicate the samples.  The store operation below interleaves them so
+       * that adjacent pixel component values take on the same sample value,
+       * per above.
+       */
+      uint8x16x2_t output_pixels = { { samples, samples } };
+      /* Store pixel component values to memory.
+       * Due to the way sample buffers are allocated, we don't need to worry
+       * about tail cases when output_width is not a multiple of 32.  See
+       * "Creation of 2-D sample arrays" in jmemmgr.c for details.
+       */
+      vst2q_u8(outptr + 2 * colctr, output_pixels);
+    }
+  }
+}
+
+
+/* The diagram below shows an array of samples produced by h2v2 downsampling.
+ *
+ *                s0        s1
+ *            +---------+---------+
+ *            | p0   p1 | p2   p3 |
+ *       sA   |         |         |
+ *            | p4   p5 | p6   p7 |
+ *            +---------+---------+
+ *            | p8   p9 | p10  p11|
+ *       sB   |         |         |
+ *            | p12  p13| p14  p15|
+ *            +---------+---------+
+ *
+ * Samples s0A-s1B were created by averaging the original pixel component
+ * values centered at positions p0-p15 above.  To approximate those original
+ * pixel component values, we duplicate the samples both horizontally and
+ * vertically:
+ *     p0(upsampled) = p1(upsampled) = p4(upsampled) = p5(upsampled) = s0A
+ *     p2(upsampled) = p3(upsampled) = p6(upsampled) = p7(upsampled) = s1A
+ *     p8(upsampled) = p9(upsampled) = p12(upsampled) = p13(upsampled) = s0B
+ *     p10(upsampled) = p11(upsampled) = p14(upsampled) = p15(upsampled) = s1B
+ */
+
+void jsimd_h2v2_upsample_neon(int max_v_samp_factor, JDIMENSION output_width,
+                              JSAMPARRAY input_data,
+                              JSAMPARRAY *output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  JSAMPROW inptr, outptr0, outptr1;
+  int inrow, outrow;
+  unsigned colctr;
+
+  for (inrow = 0, outrow = 0; outrow < max_v_samp_factor; inrow++) {
+    inptr = input_data[inrow];
+    outptr0 = output_data[outrow++];
+    outptr1 = output_data[outrow++];
+
+    for (colctr = 0; 2 * colctr < output_width; colctr += 16) {
+      uint8x16_t samples = vld1q_u8(inptr + colctr);
+      /* Duplicate the samples.  The store operation below interleaves them so
+       * that adjacent pixel component values take on the same sample value,
+       * per above.
+       */
+      uint8x16x2_t output_pixels = { { samples, samples } };
+      /* Store pixel component values for both output rows to memory.
+       * Due to the way sample buffers are allocated, we don't need to worry
+       * about tail cases when output_width is not a multiple of 32.  See
+       * "Creation of 2-D sample arrays" in jmemmgr.c for details.
+       */
+      vst2q_u8(outptr0 + 2 * colctr, output_pixels);
+      vst2q_u8(outptr1 + 2 * colctr, output_pixels);
+    }
+  }
+}
@@ -0,0 +1,214 @@
+/*
+ * jfdctfst-neon.c - fast integer FDCT (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+
+#include <arm_neon.h>
+
+
+/* jsimd_fdct_ifast_neon() performs a fast, not so accurate forward DCT
+ * (Discrete Cosine Transform) on one block of samples.  It uses the same
+ * calculations and produces exactly the same output as IJG's original
+ * jpeg_fdct_ifast() function, which can be found in jfdctfst.c.
+ *
+ * Scaled integer constants are used to avoid floating-point arithmetic:
+ *    0.382683433 = 12544 * 2^-15
+ *    0.541196100 = 17795 * 2^-15
+ *    0.707106781 = 23168 * 2^-15
+ *    0.306562965 =  9984 * 2^-15
+ *
+ * See jfdctfst.c for further details of the DCT algorithm.  Where possible,
+ * the variable names and comments here in jsimd_fdct_ifast_neon() match up
+ * with those in jpeg_fdct_ifast().
+ */
+
+#define F_0_382  12544
+#define F_0_541  17792
+#define F_0_707  23168
+#define F_0_306  9984
+
+
+ALIGN(16) static const int16_t jsimd_fdct_ifast_neon_consts[] = {
+  F_0_382, F_0_541, F_0_707, F_0_306
+};
+
+void jsimd_fdct_ifast_neon(DCTELEM *data)
+{
+  /* Load an 8x8 block of samples into Neon registers.  De-interleaving loads
+   * are used, followed by vuzp to transpose the block such that we have a
+   * column of samples per vector - allowing all rows to be processed at once.
+   */
+  int16x8x4_t data1 = vld4q_s16(data);
+  int16x8x4_t data2 = vld4q_s16(data + 4 * DCTSIZE);
+
+  int16x8x2_t cols_04 = vuzpq_s16(data1.val[0], data2.val[0]);
+  int16x8x2_t cols_15 = vuzpq_s16(data1.val[1], data2.val[1]);
+  int16x8x2_t cols_26 = vuzpq_s16(data1.val[2], data2.val[2]);
+  int16x8x2_t cols_37 = vuzpq_s16(data1.val[3], data2.val[3]);
+
+  int16x8_t col0 = cols_04.val[0];
+  int16x8_t col1 = cols_15.val[0];
+  int16x8_t col2 = cols_26.val[0];
+  int16x8_t col3 = cols_37.val[0];
+  int16x8_t col4 = cols_04.val[1];
+  int16x8_t col5 = cols_15.val[1];
+  int16x8_t col6 = cols_26.val[1];
+  int16x8_t col7 = cols_37.val[1];
+
+  /* Pass 1: process rows. */
+
+  /* Load DCT conversion constants. */
+  const int16x4_t consts = vld1_s16(jsimd_fdct_ifast_neon_consts);
+
+  int16x8_t tmp0 = vaddq_s16(col0, col7);
+  int16x8_t tmp7 = vsubq_s16(col0, col7);
+  int16x8_t tmp1 = vaddq_s16(col1, col6);
+  int16x8_t tmp6 = vsubq_s16(col1, col6);
+  int16x8_t tmp2 = vaddq_s16(col2, col5);
+  int16x8_t tmp5 = vsubq_s16(col2, col5);
+  int16x8_t tmp3 = vaddq_s16(col3, col4);
+  int16x8_t tmp4 = vsubq_s16(col3, col4);
+
+  /* Even part */
+  int16x8_t tmp10 = vaddq_s16(tmp0, tmp3);    /* phase 2 */
+  int16x8_t tmp13 = vsubq_s16(tmp0, tmp3);
+  int16x8_t tmp11 = vaddq_s16(tmp1, tmp2);
+  int16x8_t tmp12 = vsubq_s16(tmp1, tmp2);
+
+  col0 = vaddq_s16(tmp10, tmp11);             /* phase 3 */
+  col4 = vsubq_s16(tmp10, tmp11);
+
+  int16x8_t z1 = vqdmulhq_lane_s16(vaddq_s16(tmp12, tmp13), consts, 2);
+  col2 = vaddq_s16(tmp13, z1);                /* phase 5 */
+  col6 = vsubq_s16(tmp13, z1);
+
+  /* Odd part */
+  tmp10 = vaddq_s16(tmp4, tmp5);              /* phase 2 */
+  tmp11 = vaddq_s16(tmp5, tmp6);
+  tmp12 = vaddq_s16(tmp6, tmp7);
+
+  int16x8_t z5 = vqdmulhq_lane_s16(vsubq_s16(tmp10, tmp12), consts, 0);
+  int16x8_t z2 = vqdmulhq_lane_s16(tmp10, consts, 1);
+  z2 = vaddq_s16(z2, z5);
+  int16x8_t z4 = vqdmulhq_lane_s16(tmp12, consts, 3);
+  z5 = vaddq_s16(tmp12, z5);
+  z4 = vaddq_s16(z4, z5);
+  int16x8_t z3 = vqdmulhq_lane_s16(tmp11, consts, 2);
+
+  int16x8_t z11 = vaddq_s16(tmp7, z3);        /* phase 5 */
+  int16x8_t z13 = vsubq_s16(tmp7, z3);
+
+  col5 = vaddq_s16(z13, z2);                  /* phase 6 */
+  col3 = vsubq_s16(z13, z2);
+  col1 = vaddq_s16(z11, z4);
+  col7 = vsubq_s16(z11, z4);
+
+  /* Transpose to work on columns in pass 2. */
+  int16x8x2_t cols_01 = vtrnq_s16(col0, col1);
+  int16x8x2_t cols_23 = vtrnq_s16(col2, col3);
+  int16x8x2_t cols_45 = vtrnq_s16(col4, col5);
+  int16x8x2_t cols_67 = vtrnq_s16(col6, col7);
+
+  int32x4x2_t cols_0145_l = vtrnq_s32(vreinterpretq_s32_s16(cols_01.val[0]),
+                                      vreinterpretq_s32_s16(cols_45.val[0]));
+  int32x4x2_t cols_0145_h = vtrnq_s32(vreinterpretq_s32_s16(cols_01.val[1]),
+                                      vreinterpretq_s32_s16(cols_45.val[1]));
+  int32x4x2_t cols_2367_l = vtrnq_s32(vreinterpretq_s32_s16(cols_23.val[0]),
+                                      vreinterpretq_s32_s16(cols_67.val[0]));
+  int32x4x2_t cols_2367_h = vtrnq_s32(vreinterpretq_s32_s16(cols_23.val[1]),
+                                      vreinterpretq_s32_s16(cols_67.val[1]));
+
+  int32x4x2_t rows_04 = vzipq_s32(cols_0145_l.val[0], cols_2367_l.val[0]);
+  int32x4x2_t rows_15 = vzipq_s32(cols_0145_h.val[0], cols_2367_h.val[0]);
+  int32x4x2_t rows_26 = vzipq_s32(cols_0145_l.val[1], cols_2367_l.val[1]);
+  int32x4x2_t rows_37 = vzipq_s32(cols_0145_h.val[1], cols_2367_h.val[1]);
+
+  int16x8_t row0 = vreinterpretq_s16_s32(rows_04.val[0]);
+  int16x8_t row1 = vreinterpretq_s16_s32(rows_15.val[0]);
+  int16x8_t row2 = vreinterpretq_s16_s32(rows_26.val[0]);
+  int16x8_t row3 = vreinterpretq_s16_s32(rows_37.val[0]);
+  int16x8_t row4 = vreinterpretq_s16_s32(rows_04.val[1]);
+  int16x8_t row5 = vreinterpretq_s16_s32(rows_15.val[1]);
+  int16x8_t row6 = vreinterpretq_s16_s32(rows_26.val[1]);
+  int16x8_t row7 = vreinterpretq_s16_s32(rows_37.val[1]);
+
+  /* Pass 2: process columns. */
+
+  tmp0 = vaddq_s16(row0, row7);
+  tmp7 = vsubq_s16(row0, row7);
+  tmp1 = vaddq_s16(row1, row6);
+  tmp6 = vsubq_s16(row1, row6);
+  tmp2 = vaddq_s16(row2, row5);
+  tmp5 = vsubq_s16(row2, row5);
+  tmp3 = vaddq_s16(row3, row4);
+  tmp4 = vsubq_s16(row3, row4);
+
+  /* Even part */
+  tmp10 = vaddq_s16(tmp0, tmp3);              /* phase 2 */
+  tmp13 = vsubq_s16(tmp0, tmp3);
+  tmp11 = vaddq_s16(tmp1, tmp2);
+  tmp12 = vsubq_s16(tmp1, tmp2);
+
+  row0 = vaddq_s16(tmp10, tmp11);             /* phase 3 */
+  row4 = vsubq_s16(tmp10, tmp11);
+
+  z1 = vqdmulhq_lane_s16(vaddq_s16(tmp12, tmp13), consts, 2);
+  row2 = vaddq_s16(tmp13, z1);                /* phase 5 */
+  row6 = vsubq_s16(tmp13, z1);
+
+  /* Odd part */
+  tmp10 = vaddq_s16(tmp4, tmp5);              /* phase 2 */
+  tmp11 = vaddq_s16(tmp5, tmp6);
+  tmp12 = vaddq_s16(tmp6, tmp7);
+
+  z5 = vqdmulhq_lane_s16(vsubq_s16(tmp10, tmp12), consts, 0);
+  z2 = vqdmulhq_lane_s16(tmp10, consts, 1);
+  z2 = vaddq_s16(z2, z5);
+  z4 = vqdmulhq_lane_s16(tmp12, consts, 3);
+  z5 = vaddq_s16(tmp12, z5);
+  z4 = vaddq_s16(z4, z5);
+  z3 = vqdmulhq_lane_s16(tmp11, consts, 2);
+
+  z11 = vaddq_s16(tmp7, z3);                  /* phase 5 */
+  z13 = vsubq_s16(tmp7, z3);
+
+  row5 = vaddq_s16(z13, z2);                  /* phase 6 */
+  row3 = vsubq_s16(z13, z2);
+  row1 = vaddq_s16(z11, z4);
+  row7 = vsubq_s16(z11, z4);
+
+  vst1q_s16(data + 0 * DCTSIZE, row0);
+  vst1q_s16(data + 1 * DCTSIZE, row1);
+  vst1q_s16(data + 2 * DCTSIZE, row2);
+  vst1q_s16(data + 3 * DCTSIZE, row3);
+  vst1q_s16(data + 4 * DCTSIZE, row4);
+  vst1q_s16(data + 5 * DCTSIZE, row5);
+  vst1q_s16(data + 6 * DCTSIZE, row6);
+  vst1q_s16(data + 7 * DCTSIZE, row7);
+}
@@ -0,0 +1,376 @@
+/*
+ * jfdctint-neon.c - accurate integer FDCT (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+#include "neon-compat.h"
+
+#include <arm_neon.h>
+
+
+/* jsimd_fdct_islow_neon() performs a slower but more accurate forward DCT
+ * (Discrete Cosine Transform) on one block of samples.  It uses the same
+ * calculations and produces exactly the same output as IJG's original
+ * jpeg_fdct_islow() function, which can be found in jfdctint.c.
+ *
+ * Scaled integer constants are used to avoid floating-point arithmetic:
+ *    0.298631336 =  2446 * 2^-13
+ *    0.390180644 =  3196 * 2^-13
+ *    0.541196100 =  4433 * 2^-13
+ *    0.765366865 =  6270 * 2^-13
+ *    0.899976223 =  7373 * 2^-13
+ *    1.175875602 =  9633 * 2^-13
+ *    1.501321110 = 12299 * 2^-13
+ *    1.847759065 = 15137 * 2^-13
+ *    1.961570560 = 16069 * 2^-13
+ *    2.053119869 = 16819 * 2^-13
+ *    2.562915447 = 20995 * 2^-13
+ *    3.072711026 = 25172 * 2^-13
+ *
+ * See jfdctint.c for further details of the DCT algorithm.  Where possible,
+ * the variable names and comments here in jsimd_fdct_islow_neon() match up
+ * with those in jpeg_fdct_islow().
+ */
+
+#define CONST_BITS  13
+#define PASS1_BITS  2
+
+#define DESCALE_P1  (CONST_BITS - PASS1_BITS)
+#define DESCALE_P2  (CONST_BITS + PASS1_BITS)
+
+#define F_0_298  2446
+#define F_0_390  3196
+#define F_0_541  4433
+#define F_0_765  6270
+#define F_0_899  7373
+#define F_1_175  9633
+#define F_1_501  12299
+#define F_1_847  15137
+#define F_1_961  16069
+#define F_2_053  16819
+#define F_2_562  20995
+#define F_3_072  25172
+
+
+ALIGN(16) static const int16_t jsimd_fdct_islow_neon_consts[] = {
+  F_0_298, -F_0_390,  F_0_541,  F_0_765,
+ -F_0_899,  F_1_175,  F_1_501, -F_1_847,
+ -F_1_961,  F_2_053, -F_2_562,  F_3_072
+};
+
+void jsimd_fdct_islow_neon(DCTELEM *data)
+{
+  /* Load DCT constants. */
+#ifdef HAVE_VLD1_S16_X3
+  const int16x4x3_t consts = vld1_s16_x3(jsimd_fdct_islow_neon_consts);
+#else
+  /* GCC does not currently support the intrinsic vld1_<type>_x3(). */
+  const int16x4_t consts1 = vld1_s16(jsimd_fdct_islow_neon_consts);
+  const int16x4_t consts2 = vld1_s16(jsimd_fdct_islow_neon_consts + 4);
+  const int16x4_t consts3 = vld1_s16(jsimd_fdct_islow_neon_consts + 8);
+  const int16x4x3_t consts = { { consts1, consts2, consts3 } };
+#endif
+
+  /* Load an 8x8 block of samples into Neon registers.  De-interleaving loads
+   * are used, followed by vuzp to transpose the block such that we have a
+   * column of samples per vector - allowing all rows to be processed at once.
+   */
+  int16x8x4_t s_rows_0123 = vld4q_s16(data);
+  int16x8x4_t s_rows_4567 = vld4q_s16(data + 4 * DCTSIZE);
+
+  int16x8x2_t cols_04 = vuzpq_s16(s_rows_0123.val[0], s_rows_4567.val[0]);
+  int16x8x2_t cols_15 = vuzpq_s16(s_rows_0123.val[1], s_rows_4567.val[1]);
+  int16x8x2_t cols_26 = vuzpq_s16(s_rows_0123.val[2], s_rows_4567.val[2]);
+  int16x8x2_t cols_37 = vuzpq_s16(s_rows_0123.val[3], s_rows_4567.val[3]);
+
+  int16x8_t col0 = cols_04.val[0];
+  int16x8_t col1 = cols_15.val[0];
+  int16x8_t col2 = cols_26.val[0];
+  int16x8_t col3 = cols_37.val[0];
+  int16x8_t col4 = cols_04.val[1];
+  int16x8_t col5 = cols_15.val[1];
+  int16x8_t col6 = cols_26.val[1];
+  int16x8_t col7 = cols_37.val[1];
+
+  /* Pass 1: process rows. */
+
+  int16x8_t tmp0 = vaddq_s16(col0, col7);
+  int16x8_t tmp7 = vsubq_s16(col0, col7);
+  int16x8_t tmp1 = vaddq_s16(col1, col6);
+  int16x8_t tmp6 = vsubq_s16(col1, col6);
+  int16x8_t tmp2 = vaddq_s16(col2, col5);
+  int16x8_t tmp5 = vsubq_s16(col2, col5);
+  int16x8_t tmp3 = vaddq_s16(col3, col4);
+  int16x8_t tmp4 = vsubq_s16(col3, col4);
+
+  /* Even part */
+  int16x8_t tmp10 = vaddq_s16(tmp0, tmp3);
+  int16x8_t tmp13 = vsubq_s16(tmp0, tmp3);
+  int16x8_t tmp11 = vaddq_s16(tmp1, tmp2);
+  int16x8_t tmp12 = vsubq_s16(tmp1, tmp2);
+
+  col0 = vshlq_n_s16(vaddq_s16(tmp10, tmp11), PASS1_BITS);
+  col4 = vshlq_n_s16(vsubq_s16(tmp10, tmp11), PASS1_BITS);
+
+  int16x8_t tmp12_add_tmp13 = vaddq_s16(tmp12, tmp13);
+  int32x4_t z1_l =
+    vmull_lane_s16(vget_low_s16(tmp12_add_tmp13), consts.val[0], 2);
+  int32x4_t z1_h =
+    vmull_lane_s16(vget_high_s16(tmp12_add_tmp13), consts.val[0], 2);
+
+  int32x4_t col2_scaled_l =
+    vmlal_lane_s16(z1_l, vget_low_s16(tmp13), consts.val[0], 3);
+  int32x4_t col2_scaled_h =
+    vmlal_lane_s16(z1_h, vget_high_s16(tmp13), consts.val[0], 3);
+  col2 = vcombine_s16(vrshrn_n_s32(col2_scaled_l, DESCALE_P1),
+                      vrshrn_n_s32(col2_scaled_h, DESCALE_P1));
+
+  int32x4_t col6_scaled_l =
+    vmlal_lane_s16(z1_l, vget_low_s16(tmp12), consts.val[1], 3);
+  int32x4_t col6_scaled_h =
+    vmlal_lane_s16(z1_h, vget_high_s16(tmp12), consts.val[1], 3);
+  col6 = vcombine_s16(vrshrn_n_s32(col6_scaled_l, DESCALE_P1),
+                      vrshrn_n_s32(col6_scaled_h, DESCALE_P1));
+
+  /* Odd part */
+  int16x8_t z1 = vaddq_s16(tmp4, tmp7);
+  int16x8_t z2 = vaddq_s16(tmp5, tmp6);
+  int16x8_t z3 = vaddq_s16(tmp4, tmp6);
+  int16x8_t z4 = vaddq_s16(tmp5, tmp7);
+  /* sqrt(2) * c3 */
+  int32x4_t z5_l = vmull_lane_s16(vget_low_s16(z3), consts.val[1], 1);
+  int32x4_t z5_h = vmull_lane_s16(vget_high_s16(z3), consts.val[1], 1);
+  z5_l = vmlal_lane_s16(z5_l, vget_low_s16(z4), consts.val[1], 1);
+  z5_h = vmlal_lane_s16(z5_h, vget_high_s16(z4), consts.val[1], 1);
+
+  /* sqrt(2) * (-c1+c3+c5-c7) */
+  int32x4_t tmp4_l = vmull_lane_s16(vget_low_s16(tmp4), consts.val[0], 0);
+  int32x4_t tmp4_h = vmull_lane_s16(vget_high_s16(tmp4), consts.val[0], 0);
+  /* sqrt(2) * ( c1+c3-c5+c7) */
+  int32x4_t tmp5_l = vmull_lane_s16(vget_low_s16(tmp5), consts.val[2], 1);
+  int32x4_t tmp5_h = vmull_lane_s16(vget_high_s16(tmp5), consts.val[2], 1);
+  /* sqrt(2) * ( c1+c3+c5-c7) */
+  int32x4_t tmp6_l = vmull_lane_s16(vget_low_s16(tmp6), consts.val[2], 3);
+  int32x4_t tmp6_h = vmull_lane_s16(vget_high_s16(tmp6), consts.val[2], 3);
+  /* sqrt(2) * ( c1+c3-c5-c7) */
+  int32x4_t tmp7_l = vmull_lane_s16(vget_low_s16(tmp7), consts.val[1], 2);
+  int32x4_t tmp7_h = vmull_lane_s16(vget_high_s16(tmp7), consts.val[1], 2);
+
+  /* sqrt(2) * (c7-c3) */
+  z1_l = vmull_lane_s16(vget_low_s16(z1), consts.val[1], 0);
+  z1_h = vmull_lane_s16(vget_high_s16(z1), consts.val[1], 0);
+  /* sqrt(2) * (-c1-c3) */
+  int32x4_t z2_l = vmull_lane_s16(vget_low_s16(z2), consts.val[2], 2);
+  int32x4_t z2_h = vmull_lane_s16(vget_high_s16(z2), consts.val[2], 2);
+  /* sqrt(2) * (-c3-c5) */
+  int32x4_t z3_l = vmull_lane_s16(vget_low_s16(z3), consts.val[2], 0);
+  int32x4_t z3_h = vmull_lane_s16(vget_high_s16(z3), consts.val[2], 0);
+  /* sqrt(2) * (c5-c3) */
+  int32x4_t z4_l = vmull_lane_s16(vget_low_s16(z4), consts.val[0], 1);
+  int32x4_t z4_h = vmull_lane_s16(vget_high_s16(z4), consts.val[0], 1);
+
+  z3_l = vaddq_s32(z3_l, z5_l);
+  z3_h = vaddq_s32(z3_h, z5_h);
+  z4_l = vaddq_s32(z4_l, z5_l);
+  z4_h = vaddq_s32(z4_h, z5_h);
+
+  tmp4_l = vaddq_s32(tmp4_l, z1_l);
+  tmp4_h = vaddq_s32(tmp4_h, z1_h);
+  tmp4_l = vaddq_s32(tmp4_l, z3_l);
+  tmp4_h = vaddq_s32(tmp4_h, z3_h);
+  col7 = vcombine_s16(vrshrn_n_s32(tmp4_l, DESCALE_P1),
+                      vrshrn_n_s32(tmp4_h, DESCALE_P1));
+
+  tmp5_l = vaddq_s32(tmp5_l, z2_l);
+  tmp5_h = vaddq_s32(tmp5_h, z2_h);
+  tmp5_l = vaddq_s32(tmp5_l, z4_l);
+  tmp5_h = vaddq_s32(tmp5_h, z4_h);
+  col5 = vcombine_s16(vrshrn_n_s32(tmp5_l, DESCALE_P1),
+                      vrshrn_n_s32(tmp5_h, DESCALE_P1));
+
+  tmp6_l = vaddq_s32(tmp6_l, z2_l);
+  tmp6_h = vaddq_s32(tmp6_h, z2_h);
+  tmp6_l = vaddq_s32(tmp6_l, z3_l);
+  tmp6_h = vaddq_s32(tmp6_h, z3_h);
+  col3 = vcombine_s16(vrshrn_n_s32(tmp6_l, DESCALE_P1),
+                      vrshrn_n_s32(tmp6_h, DESCALE_P1));
+
+  tmp7_l = vaddq_s32(tmp7_l, z1_l);
+  tmp7_h = vaddq_s32(tmp7_h, z1_h);
+  tmp7_l = vaddq_s32(tmp7_l, z4_l);
+  tmp7_h = vaddq_s32(tmp7_h, z4_h);
+  col1 = vcombine_s16(vrshrn_n_s32(tmp7_l, DESCALE_P1),
+                      vrshrn_n_s32(tmp7_h, DESCALE_P1));
+
+  /* Transpose to work on columns in pass 2. */
+  int16x8x2_t cols_01 = vtrnq_s16(col0, col1);
+  int16x8x2_t cols_23 = vtrnq_s16(col2, col3);
+  int16x8x2_t cols_45 = vtrnq_s16(col4, col5);
+  int16x8x2_t cols_67 = vtrnq_s16(col6, col7);
+
+  int32x4x2_t cols_0145_l = vtrnq_s32(vreinterpretq_s32_s16(cols_01.val[0]),
+                                      vreinterpretq_s32_s16(cols_45.val[0]));
+  int32x4x2_t cols_0145_h = vtrnq_s32(vreinterpretq_s32_s16(cols_01.val[1]),
+                                      vreinterpretq_s32_s16(cols_45.val[1]));
+  int32x4x2_t cols_2367_l = vtrnq_s32(vreinterpretq_s32_s16(cols_23.val[0]),
+                                      vreinterpretq_s32_s16(cols_67.val[0]));
+  int32x4x2_t cols_2367_h = vtrnq_s32(vreinterpretq_s32_s16(cols_23.val[1]),
+                                      vreinterpretq_s32_s16(cols_67.val[1]));
+
+  int32x4x2_t rows_04 = vzipq_s32(cols_0145_l.val[0], cols_2367_l.val[0]);
+  int32x4x2_t rows_15 = vzipq_s32(cols_0145_h.val[0], cols_2367_h.val[0]);
+  int32x4x2_t rows_26 = vzipq_s32(cols_0145_l.val[1], cols_2367_l.val[1]);
+  int32x4x2_t rows_37 = vzipq_s32(cols_0145_h.val[1], cols_2367_h.val[1]);
+
+  int16x8_t row0 = vreinterpretq_s16_s32(rows_04.val[0]);
+  int16x8_t row1 = vreinterpretq_s16_s32(rows_15.val[0]);
+  int16x8_t row2 = vreinterpretq_s16_s32(rows_26.val[0]);
+  int16x8_t row3 = vreinterpretq_s16_s32(rows_37.val[0]);
+  int16x8_t row4 = vreinterpretq_s16_s32(rows_04.val[1]);
+  int16x8_t row5 = vreinterpretq_s16_s32(rows_15.val[1]);
+  int16x8_t row6 = vreinterpretq_s16_s32(rows_26.val[1]);
+  int16x8_t row7 = vreinterpretq_s16_s32(rows_37.val[1]);
+
+  /* Pass 2: process columns. */
+
+  tmp0 = vaddq_s16(row0, row7);
+  tmp7 = vsubq_s16(row0, row7);
+  tmp1 = vaddq_s16(row1, row6);
+  tmp6 = vsubq_s16(row1, row6);
+  tmp2 = vaddq_s16(row2, row5);
+  tmp5 = vsubq_s16(row2, row5);
+  tmp3 = vaddq_s16(row3, row4);
+  tmp4 = vsubq_s16(row3, row4);
+
+  /* Even part */
+  tmp10 = vaddq_s16(tmp0, tmp3);
+  tmp13 = vsubq_s16(tmp0, tmp3);
+  tmp11 = vaddq_s16(tmp1, tmp2);
+  tmp12 = vsubq_s16(tmp1, tmp2);
+
+  row0 = vrshrq_n_s16(vaddq_s16(tmp10, tmp11), PASS1_BITS);
+  row4 = vrshrq_n_s16(vsubq_s16(tmp10, tmp11), PASS1_BITS);
+
+  tmp12_add_tmp13 = vaddq_s16(tmp12, tmp13);
+  z1_l = vmull_lane_s16(vget_low_s16(tmp12_add_tmp13), consts.val[0], 2);
+  z1_h = vmull_lane_s16(vget_high_s16(tmp12_add_tmp13), consts.val[0], 2);
+
+  int32x4_t row2_scaled_l =
+    vmlal_lane_s16(z1_l, vget_low_s16(tmp13), consts.val[0], 3);
+  int32x4_t row2_scaled_h =
+    vmlal_lane_s16(z1_h, vget_high_s16(tmp13), consts.val[0], 3);
+  row2 = vcombine_s16(vrshrn_n_s32(row2_scaled_l, DESCALE_P2),
+                      vrshrn_n_s32(row2_scaled_h, DESCALE_P2));
+
+  int32x4_t row6_scaled_l =
+    vmlal_lane_s16(z1_l, vget_low_s16(tmp12), consts.val[1], 3);
+  int32x4_t row6_scaled_h =
+    vmlal_lane_s16(z1_h, vget_high_s16(tmp12), consts.val[1], 3);
+  row6 = vcombine_s16(vrshrn_n_s32(row6_scaled_l, DESCALE_P2),
+                      vrshrn_n_s32(row6_scaled_h, DESCALE_P2));
+
+  /* Odd part */
+  z1 = vaddq_s16(tmp4, tmp7);
+  z2 = vaddq_s16(tmp5, tmp6);
+  z3 = vaddq_s16(tmp4, tmp6);
+  z4 = vaddq_s16(tmp5, tmp7);
+  /* sqrt(2) * c3 */
+  z5_l = vmull_lane_s16(vget_low_s16(z3), consts.val[1], 1);
+  z5_h = vmull_lane_s16(vget_high_s16(z3), consts.val[1], 1);
+  z5_l = vmlal_lane_s16(z5_l, vget_low_s16(z4), consts.val[1], 1);
+  z5_h = vmlal_lane_s16(z5_h, vget_high_s16(z4), consts.val[1], 1);
+
+  /* sqrt(2) * (-c1+c3+c5-c7) */
+  tmp4_l = vmull_lane_s16(vget_low_s16(tmp4), consts.val[0], 0);
+  tmp4_h = vmull_lane_s16(vget_high_s16(tmp4), consts.val[0], 0);
+  /* sqrt(2) * ( c1+c3-c5+c7) */
+  tmp5_l = vmull_lane_s16(vget_low_s16(tmp5), consts.val[2], 1);
+  tmp5_h = vmull_lane_s16(vget_high_s16(tmp5), consts.val[2], 1);
+  /* sqrt(2) * ( c1+c3+c5-c7) */
+  tmp6_l = vmull_lane_s16(vget_low_s16(tmp6), consts.val[2], 3);
+  tmp6_h = vmull_lane_s16(vget_high_s16(tmp6), consts.val[2], 3);
+  /* sqrt(2) * ( c1+c3-c5-c7) */
+  tmp7_l = vmull_lane_s16(vget_low_s16(tmp7), consts.val[1], 2);
+  tmp7_h = vmull_lane_s16(vget_high_s16(tmp7), consts.val[1], 2);
+
+  /* sqrt(2) * (c7-c3) */
+  z1_l = vmull_lane_s16(vget_low_s16(z1), consts.val[1], 0);
+  z1_h = vmull_lane_s16(vget_high_s16(z1), consts.val[1], 0);
+  /* sqrt(2) * (-c1-c3) */
+  z2_l = vmull_lane_s16(vget_low_s16(z2), consts.val[2], 2);
+  z2_h = vmull_lane_s16(vget_high_s16(z2), consts.val[2], 2);
+  /* sqrt(2) * (-c3-c5) */
+  z3_l = vmull_lane_s16(vget_low_s16(z3), consts.val[2], 0);
+  z3_h = vmull_lane_s16(vget_high_s16(z3), consts.val[2], 0);
+  /* sqrt(2) * (c5-c3) */
+  z4_l = vmull_lane_s16(vget_low_s16(z4), consts.val[0], 1);
+  z4_h = vmull_lane_s16(vget_high_s16(z4), consts.val[0], 1);
+
+  z3_l = vaddq_s32(z3_l, z5_l);
+  z3_h = vaddq_s32(z3_h, z5_h);
+  z4_l = vaddq_s32(z4_l, z5_l);
+  z4_h = vaddq_s32(z4_h, z5_h);
+
+  tmp4_l = vaddq_s32(tmp4_l, z1_l);
+  tmp4_h = vaddq_s32(tmp4_h, z1_h);
+  tmp4_l = vaddq_s32(tmp4_l, z3_l);
+  tmp4_h = vaddq_s32(tmp4_h, z3_h);
+  row7 = vcombine_s16(vrshrn_n_s32(tmp4_l, DESCALE_P2),
+                      vrshrn_n_s32(tmp4_h, DESCALE_P2));
+
+  tmp5_l = vaddq_s32(tmp5_l, z2_l);
+  tmp5_h = vaddq_s32(tmp5_h, z2_h);
+  tmp5_l = vaddq_s32(tmp5_l, z4_l);
+  tmp5_h = vaddq_s32(tmp5_h, z4_h);
+  row5 = vcombine_s16(vrshrn_n_s32(tmp5_l, DESCALE_P2),
+                      vrshrn_n_s32(tmp5_h, DESCALE_P2));
+
+  tmp6_l = vaddq_s32(tmp6_l, z2_l);
+  tmp6_h = vaddq_s32(tmp6_h, z2_h);
+  tmp6_l = vaddq_s32(tmp6_l, z3_l);
+  tmp6_h = vaddq_s32(tmp6_h, z3_h);
+  row3 = vcombine_s16(vrshrn_n_s32(tmp6_l, DESCALE_P2),
+                      vrshrn_n_s32(tmp6_h, DESCALE_P2));
+
+  tmp7_l = vaddq_s32(tmp7_l, z1_l);
+  tmp7_h = vaddq_s32(tmp7_h, z1_h);
+  tmp7_l = vaddq_s32(tmp7_l, z4_l);
+  tmp7_h = vaddq_s32(tmp7_h, z4_h);
+  row1 = vcombine_s16(vrshrn_n_s32(tmp7_l, DESCALE_P2),
+                      vrshrn_n_s32(tmp7_h, DESCALE_P2));
+
+  vst1q_s16(data + 0 * DCTSIZE, row0);
+  vst1q_s16(data + 1 * DCTSIZE, row1);
+  vst1q_s16(data + 2 * DCTSIZE, row2);
+  vst1q_s16(data + 3 * DCTSIZE, row3);
+  vst1q_s16(data + 4 * DCTSIZE, row4);
+  vst1q_s16(data + 5 * DCTSIZE, row5);
+  vst1q_s16(data + 6 * DCTSIZE, row6);
+  vst1q_s16(data + 7 * DCTSIZE, row7);
+}
@@ -0,0 +1,472 @@
+/*
+ * jidctfst-neon.c - fast integer IDCT (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+
+#include <arm_neon.h>
+
+
+/* jsimd_idct_ifast_neon() performs dequantization and a fast, not so accurate
+ * inverse DCT (Discrete Cosine Transform) on one block of coefficients.  It
+ * uses the same calculations and produces exactly the same output as IJG's
+ * original jpeg_idct_ifast() function, which can be found in jidctfst.c.
+ *
+ * Scaled integer constants are used to avoid floating-point arithmetic:
+ *    0.082392200 =  2688 * 2^-15
+ *    0.414213562 = 13568 * 2^-15
+ *    0.847759065 = 27776 * 2^-15
+ *    0.613125930 = 20096 * 2^-15
+ *
+ * See jidctfst.c for further details of the IDCT algorithm.  Where possible,
+ * the variable names and comments here in jsimd_idct_ifast_neon() match up
+ * with those in jpeg_idct_ifast().
+ */
+
+#define PASS1_BITS  2
+
+#define F_0_082  2688
+#define F_0_414  13568
+#define F_0_847  27776
+#define F_0_613  20096
+
+
+ALIGN(16) static const int16_t jsimd_idct_ifast_neon_consts[] = {
+  F_0_082, F_0_414, F_0_847, F_0_613
+};
+
+void jsimd_idct_ifast_neon(void *dct_table, JCOEFPTR coef_block,
+                           JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  IFAST_MULT_TYPE *quantptr = dct_table;
+
+  /* Load DCT coefficients. */
+  int16x8_t row0 = vld1q_s16(coef_block + 0 * DCTSIZE);
+  int16x8_t row1 = vld1q_s16(coef_block + 1 * DCTSIZE);
+  int16x8_t row2 = vld1q_s16(coef_block + 2 * DCTSIZE);
+  int16x8_t row3 = vld1q_s16(coef_block + 3 * DCTSIZE);
+  int16x8_t row4 = vld1q_s16(coef_block + 4 * DCTSIZE);
+  int16x8_t row5 = vld1q_s16(coef_block + 5 * DCTSIZE);
+  int16x8_t row6 = vld1q_s16(coef_block + 6 * DCTSIZE);
+  int16x8_t row7 = vld1q_s16(coef_block + 7 * DCTSIZE);
+
+  /* Load quantization table values for DC coefficients. */
+  int16x8_t quant_row0 = vld1q_s16(quantptr + 0 * DCTSIZE);
+  /* Dequantize DC coefficients. */
+  row0 = vmulq_s16(row0, quant_row0);
+
+  /* Construct bitmap to test if all AC coefficients are 0. */
+  int16x8_t bitmap = vorrq_s16(row1, row2);
+  bitmap = vorrq_s16(bitmap, row3);
+  bitmap = vorrq_s16(bitmap, row4);
+  bitmap = vorrq_s16(bitmap, row5);
+  bitmap = vorrq_s16(bitmap, row6);
+  bitmap = vorrq_s16(bitmap, row7);
+
+  int64_t left_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 0);
+  int64_t right_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 1);
+
+  /* Load IDCT conversion constants. */
+  const int16x4_t consts = vld1_s16(jsimd_idct_ifast_neon_consts);
+
+  if (left_ac_bitmap == 0 && right_ac_bitmap == 0) {
+    /* All AC coefficients are zero.
+     * Compute DC values and duplicate into vectors.
+     */
+    int16x8_t dcval = row0;
+    row1 = dcval;
+    row2 = dcval;
+    row3 = dcval;
+    row4 = dcval;
+    row5 = dcval;
+    row6 = dcval;
+    row7 = dcval;
+  } else if (left_ac_bitmap == 0) {
+    /* AC coefficients are zero for columns 0, 1, 2, and 3.
+     * Use DC values for these columns.
+     */
+    int16x4_t dcval = vget_low_s16(row0);
+
+    /* Commence regular fast IDCT computation for columns 4, 5, 6, and 7. */
+
+    /* Load quantization table. */
+    int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE + 4);
+    int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE + 4);
+    int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE + 4);
+    int16x4_t quant_row4 = vld1_s16(quantptr + 4 * DCTSIZE + 4);
+    int16x4_t quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE + 4);
+    int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE + 4);
+    int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE + 4);
+
+    /* Even part: dequantize DCT coefficients. */
+    int16x4_t tmp0 = vget_high_s16(row0);
+    int16x4_t tmp1 = vmul_s16(vget_high_s16(row2), quant_row2);
+    int16x4_t tmp2 = vmul_s16(vget_high_s16(row4), quant_row4);
+    int16x4_t tmp3 = vmul_s16(vget_high_s16(row6), quant_row6);
+
+    int16x4_t tmp10 = vadd_s16(tmp0, tmp2);   /* phase 3 */
+    int16x4_t tmp11 = vsub_s16(tmp0, tmp2);
+
+    int16x4_t tmp13 = vadd_s16(tmp1, tmp3);   /* phases 5-3 */
+    int16x4_t tmp1_sub_tmp3 = vsub_s16(tmp1, tmp3);
+    int16x4_t tmp12 = vqdmulh_lane_s16(tmp1_sub_tmp3, consts, 1);
+    tmp12 = vadd_s16(tmp12, tmp1_sub_tmp3);
+    tmp12 = vsub_s16(tmp12, tmp13);
+
+    tmp0 = vadd_s16(tmp10, tmp13);            /* phase 2 */
+    tmp3 = vsub_s16(tmp10, tmp13);
+    tmp1 = vadd_s16(tmp11, tmp12);
+    tmp2 = vsub_s16(tmp11, tmp12);
+
+    /* Odd part: dequantize DCT coefficients. */
+    int16x4_t tmp4 = vmul_s16(vget_high_s16(row1), quant_row1);
+    int16x4_t tmp5 = vmul_s16(vget_high_s16(row3), quant_row3);
+    int16x4_t tmp6 = vmul_s16(vget_high_s16(row5), quant_row5);
+    int16x4_t tmp7 = vmul_s16(vget_high_s16(row7), quant_row7);
+
+    int16x4_t z13 = vadd_s16(tmp6, tmp5);     /* phase 6 */
+    int16x4_t neg_z10 = vsub_s16(tmp5, tmp6);
+    int16x4_t z11 = vadd_s16(tmp4, tmp7);
+    int16x4_t z12 = vsub_s16(tmp4, tmp7);
+
+    tmp7 = vadd_s16(z11, z13);                /* phase 5 */
+    int16x4_t z11_sub_z13 = vsub_s16(z11, z13);
+    tmp11 = vqdmulh_lane_s16(z11_sub_z13, consts, 1);
+    tmp11 = vadd_s16(tmp11, z11_sub_z13);
+
+    int16x4_t z10_add_z12 = vsub_s16(z12, neg_z10);
+    int16x4_t z5 = vqdmulh_lane_s16(z10_add_z12, consts, 2);
+    z5 = vadd_s16(z5, z10_add_z12);
+    tmp10 = vqdmulh_lane_s16(z12, consts, 0);
+    tmp10 = vadd_s16(tmp10, z12);
+    tmp10 = vsub_s16(tmp10, z5);
+    tmp12 = vqdmulh_lane_s16(neg_z10, consts, 3);
+    tmp12 = vadd_s16(tmp12, vadd_s16(neg_z10, neg_z10));
+    tmp12 = vadd_s16(tmp12, z5);
+
+    tmp6 = vsub_s16(tmp12, tmp7);             /* phase 2 */
+    tmp5 = vsub_s16(tmp11, tmp6);
+    tmp4 = vadd_s16(tmp10, tmp5);
+
+    row0 = vcombine_s16(dcval, vadd_s16(tmp0, tmp7));
+    row7 = vcombine_s16(dcval, vsub_s16(tmp0, tmp7));
+    row1 = vcombine_s16(dcval, vadd_s16(tmp1, tmp6));
+    row6 = vcombine_s16(dcval, vsub_s16(tmp1, tmp6));
+    row2 = vcombine_s16(dcval, vadd_s16(tmp2, tmp5));
+    row5 = vcombine_s16(dcval, vsub_s16(tmp2, tmp5));
+    row4 = vcombine_s16(dcval, vadd_s16(tmp3, tmp4));
+    row3 = vcombine_s16(dcval, vsub_s16(tmp3, tmp4));
+  } else if (right_ac_bitmap == 0) {
+    /* AC coefficients are zero for columns 4, 5, 6, and 7.
+     * Use DC values for these columns.
+     */
+    int16x4_t dcval = vget_high_s16(row0);
+
+    /* Commence regular fast IDCT computation for columns 0, 1, 2, and 3. */
+
+    /* Load quantization table. */
+    int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE);
+    int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE);
+    int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE);
+    int16x4_t quant_row4 = vld1_s16(quantptr + 4 * DCTSIZE);
+    int16x4_t quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE);
+    int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE);
+    int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE);
+
+    /* Even part: dequantize DCT coefficients. */
+    int16x4_t tmp0 = vget_low_s16(row0);
+    int16x4_t tmp1 = vmul_s16(vget_low_s16(row2), quant_row2);
+    int16x4_t tmp2 = vmul_s16(vget_low_s16(row4), quant_row4);
+    int16x4_t tmp3 = vmul_s16(vget_low_s16(row6), quant_row6);
+
+    int16x4_t tmp10 = vadd_s16(tmp0, tmp2);   /* phase 3 */
+    int16x4_t tmp11 = vsub_s16(tmp0, tmp2);
+
+    int16x4_t tmp13 = vadd_s16(tmp1, tmp3);   /* phases 5-3 */
+    int16x4_t tmp1_sub_tmp3 = vsub_s16(tmp1, tmp3);
+    int16x4_t tmp12 = vqdmulh_lane_s16(tmp1_sub_tmp3, consts, 1);
+    tmp12 = vadd_s16(tmp12, tmp1_sub_tmp3);
+    tmp12 = vsub_s16(tmp12, tmp13);
+
+    tmp0 = vadd_s16(tmp10, tmp13);            /* phase 2 */
+    tmp3 = vsub_s16(tmp10, tmp13);
+    tmp1 = vadd_s16(tmp11, tmp12);
+    tmp2 = vsub_s16(tmp11, tmp12);
+
+    /* Odd part: dequantize DCT coefficients. */
+    int16x4_t tmp4 = vmul_s16(vget_low_s16(row1), quant_row1);
+    int16x4_t tmp5 = vmul_s16(vget_low_s16(row3), quant_row3);
+    int16x4_t tmp6 = vmul_s16(vget_low_s16(row5), quant_row5);
+    int16x4_t tmp7 = vmul_s16(vget_low_s16(row7), quant_row7);
+
+    int16x4_t z13 = vadd_s16(tmp6, tmp5);     /* phase 6 */
+    int16x4_t neg_z10 = vsub_s16(tmp5, tmp6);
+    int16x4_t z11 = vadd_s16(tmp4, tmp7);
+    int16x4_t z12 = vsub_s16(tmp4, tmp7);
+
+    tmp7 = vadd_s16(z11, z13);                /* phase 5 */
+    int16x4_t z11_sub_z13 = vsub_s16(z11, z13);
+    tmp11 = vqdmulh_lane_s16(z11_sub_z13, consts, 1);
+    tmp11 = vadd_s16(tmp11, z11_sub_z13);
+
+    int16x4_t z10_add_z12 = vsub_s16(z12, neg_z10);
+    int16x4_t z5 = vqdmulh_lane_s16(z10_add_z12, consts, 2);
+    z5 = vadd_s16(z5, z10_add_z12);
+    tmp10 = vqdmulh_lane_s16(z12, consts, 0);
+    tmp10 = vadd_s16(tmp10, z12);
+    tmp10 = vsub_s16(tmp10, z5);
+    tmp12 = vqdmulh_lane_s16(neg_z10, consts, 3);
+    tmp12 = vadd_s16(tmp12, vadd_s16(neg_z10, neg_z10));
+    tmp12 = vadd_s16(tmp12, z5);
+
+    tmp6 = vsub_s16(tmp12, tmp7);             /* phase 2 */
+    tmp5 = vsub_s16(tmp11, tmp6);
+    tmp4 = vadd_s16(tmp10, tmp5);
+
+    row0 = vcombine_s16(vadd_s16(tmp0, tmp7), dcval);
+    row7 = vcombine_s16(vsub_s16(tmp0, tmp7), dcval);
+    row1 = vcombine_s16(vadd_s16(tmp1, tmp6), dcval);
+    row6 = vcombine_s16(vsub_s16(tmp1, tmp6), dcval);
+    row2 = vcombine_s16(vadd_s16(tmp2, tmp5), dcval);
+    row5 = vcombine_s16(vsub_s16(tmp2, tmp5), dcval);
+    row4 = vcombine_s16(vadd_s16(tmp3, tmp4), dcval);
+    row3 = vcombine_s16(vsub_s16(tmp3, tmp4), dcval);
+  } else {
+    /* Some AC coefficients are non-zero; full IDCT calculation required. */
+
+    /* Load quantization table. */
+    int16x8_t quant_row1 = vld1q_s16(quantptr + 1 * DCTSIZE);
+    int16x8_t quant_row2 = vld1q_s16(quantptr + 2 * DCTSIZE);
+    int16x8_t quant_row3 = vld1q_s16(quantptr + 3 * DCTSIZE);
+    int16x8_t quant_row4 = vld1q_s16(quantptr + 4 * DCTSIZE);
+    int16x8_t quant_row5 = vld1q_s16(quantptr + 5 * DCTSIZE);
+    int16x8_t quant_row6 = vld1q_s16(quantptr + 6 * DCTSIZE);
+    int16x8_t quant_row7 = vld1q_s16(quantptr + 7 * DCTSIZE);
+
+    /* Even part: dequantize DCT coefficients. */
+    int16x8_t tmp0 = row0;
+    int16x8_t tmp1 = vmulq_s16(row2, quant_row2);
+    int16x8_t tmp2 = vmulq_s16(row4, quant_row4);
+    int16x8_t tmp3 = vmulq_s16(row6, quant_row6);
+
+    int16x8_t tmp10 = vaddq_s16(tmp0, tmp2);   /* phase 3 */
+    int16x8_t tmp11 = vsubq_s16(tmp0, tmp2);
+
+    int16x8_t tmp13 = vaddq_s16(tmp1, tmp3);   /* phases 5-3 */
+    int16x8_t tmp1_sub_tmp3 = vsubq_s16(tmp1, tmp3);
+    int16x8_t tmp12 = vqdmulhq_lane_s16(tmp1_sub_tmp3, consts, 1);
+    tmp12 = vaddq_s16(tmp12, tmp1_sub_tmp3);
+    tmp12 = vsubq_s16(tmp12, tmp13);
+
+    tmp0 = vaddq_s16(tmp10, tmp13);            /* phase 2 */
+    tmp3 = vsubq_s16(tmp10, tmp13);
+    tmp1 = vaddq_s16(tmp11, tmp12);
+    tmp2 = vsubq_s16(tmp11, tmp12);
+
+    /* Odd part: dequantize DCT coefficients. */
+    int16x8_t tmp4 = vmulq_s16(row1, quant_row1);
+    int16x8_t tmp5 = vmulq_s16(row3, quant_row3);
+    int16x8_t tmp6 = vmulq_s16(row5, quant_row5);
+    int16x8_t tmp7 = vmulq_s16(row7, quant_row7);
+
+    int16x8_t z13 = vaddq_s16(tmp6, tmp5);     /* phase 6 */
+    int16x8_t neg_z10 = vsubq_s16(tmp5, tmp6);
+    int16x8_t z11 = vaddq_s16(tmp4, tmp7);
+    int16x8_t z12 = vsubq_s16(tmp4, tmp7);
+
+    tmp7 = vaddq_s16(z11, z13);                /* phase 5 */
+    int16x8_t z11_sub_z13 = vsubq_s16(z11, z13);
+    tmp11 = vqdmulhq_lane_s16(z11_sub_z13, consts, 1);
+    tmp11 = vaddq_s16(tmp11, z11_sub_z13);
+
+    int16x8_t z10_add_z12 = vsubq_s16(z12, neg_z10);
+    int16x8_t z5 = vqdmulhq_lane_s16(z10_add_z12, consts, 2);
+    z5 = vaddq_s16(z5, z10_add_z12);
+    tmp10 = vqdmulhq_lane_s16(z12, consts, 0);
+    tmp10 = vaddq_s16(tmp10, z12);
+    tmp10 = vsubq_s16(tmp10, z5);
+    tmp12 = vqdmulhq_lane_s16(neg_z10, consts, 3);
+    tmp12 = vaddq_s16(tmp12, vaddq_s16(neg_z10, neg_z10));
+    tmp12 = vaddq_s16(tmp12, z5);
+
+    tmp6 = vsubq_s16(tmp12, tmp7);             /* phase 2 */
+    tmp5 = vsubq_s16(tmp11, tmp6);
+    tmp4 = vaddq_s16(tmp10, tmp5);
+
+    row0 = vaddq_s16(tmp0, tmp7);
+    row7 = vsubq_s16(tmp0, tmp7);
+    row1 = vaddq_s16(tmp1, tmp6);
+    row6 = vsubq_s16(tmp1, tmp6);
+    row2 = vaddq_s16(tmp2, tmp5);
+    row5 = vsubq_s16(tmp2, tmp5);
+    row4 = vaddq_s16(tmp3, tmp4);
+    row3 = vsubq_s16(tmp3, tmp4);
+  }
+
+  /* Transpose rows to work on columns in pass 2. */
+  int16x8x2_t rows_01 = vtrnq_s16(row0, row1);
+  int16x8x2_t rows_23 = vtrnq_s16(row2, row3);
+  int16x8x2_t rows_45 = vtrnq_s16(row4, row5);
+  int16x8x2_t rows_67 = vtrnq_s16(row6, row7);
+
+  int32x4x2_t rows_0145_l = vtrnq_s32(vreinterpretq_s32_s16(rows_01.val[0]),
+                                      vreinterpretq_s32_s16(rows_45.val[0]));
+  int32x4x2_t rows_0145_h = vtrnq_s32(vreinterpretq_s32_s16(rows_01.val[1]),
+                                      vreinterpretq_s32_s16(rows_45.val[1]));
+  int32x4x2_t rows_2367_l = vtrnq_s32(vreinterpretq_s32_s16(rows_23.val[0]),
+                                      vreinterpretq_s32_s16(rows_67.val[0]));
+  int32x4x2_t rows_2367_h = vtrnq_s32(vreinterpretq_s32_s16(rows_23.val[1]),
+                                      vreinterpretq_s32_s16(rows_67.val[1]));
+
+  int32x4x2_t cols_04 = vzipq_s32(rows_0145_l.val[0], rows_2367_l.val[0]);
+  int32x4x2_t cols_15 = vzipq_s32(rows_0145_h.val[0], rows_2367_h.val[0]);
+  int32x4x2_t cols_26 = vzipq_s32(rows_0145_l.val[1], rows_2367_l.val[1]);
+  int32x4x2_t cols_37 = vzipq_s32(rows_0145_h.val[1], rows_2367_h.val[1]);
+
+  int16x8_t col0 = vreinterpretq_s16_s32(cols_04.val[0]);
+  int16x8_t col1 = vreinterpretq_s16_s32(cols_15.val[0]);
+  int16x8_t col2 = vreinterpretq_s16_s32(cols_26.val[0]);
+  int16x8_t col3 = vreinterpretq_s16_s32(cols_37.val[0]);
+  int16x8_t col4 = vreinterpretq_s16_s32(cols_04.val[1]);
+  int16x8_t col5 = vreinterpretq_s16_s32(cols_15.val[1]);
+  int16x8_t col6 = vreinterpretq_s16_s32(cols_26.val[1]);
+  int16x8_t col7 = vreinterpretq_s16_s32(cols_37.val[1]);
+
+  /* 1-D IDCT, pass 2 */
+
+  /* Even part */
+  int16x8_t tmp10 = vaddq_s16(col0, col4);
+  int16x8_t tmp11 = vsubq_s16(col0, col4);
+
+  int16x8_t tmp13 = vaddq_s16(col2, col6);
+  int16x8_t col2_sub_col6 = vsubq_s16(col2, col6);
+  int16x8_t tmp12 = vqdmulhq_lane_s16(col2_sub_col6, consts, 1);
+  tmp12 = vaddq_s16(tmp12, col2_sub_col6);
+  tmp12 = vsubq_s16(tmp12, tmp13);
+
+  int16x8_t tmp0 = vaddq_s16(tmp10, tmp13);
+  int16x8_t tmp3 = vsubq_s16(tmp10, tmp13);
+  int16x8_t tmp1 = vaddq_s16(tmp11, tmp12);
+  int16x8_t tmp2 = vsubq_s16(tmp11, tmp12);
+
+  /* Odd part */
+  int16x8_t z13 = vaddq_s16(col5, col3);
+  int16x8_t neg_z10 = vsubq_s16(col3, col5);
+  int16x8_t z11 = vaddq_s16(col1, col7);
+  int16x8_t z12 = vsubq_s16(col1, col7);
+
+  int16x8_t tmp7 = vaddq_s16(z11, z13);      /* phase 5 */
+  int16x8_t z11_sub_z13 = vsubq_s16(z11, z13);
+  tmp11 = vqdmulhq_lane_s16(z11_sub_z13, consts, 1);
+  tmp11 = vaddq_s16(tmp11, z11_sub_z13);
+
+  int16x8_t z10_add_z12 = vsubq_s16(z12, neg_z10);
+  int16x8_t z5 = vqdmulhq_lane_s16(z10_add_z12, consts, 2);
+  z5 = vaddq_s16(z5, z10_add_z12);
+  tmp10 = vqdmulhq_lane_s16(z12, consts, 0);
+  tmp10 = vaddq_s16(tmp10, z12);
+  tmp10 = vsubq_s16(tmp10, z5);
+  tmp12 = vqdmulhq_lane_s16(neg_z10, consts, 3);
+  tmp12 = vaddq_s16(tmp12, vaddq_s16(neg_z10, neg_z10));
+  tmp12 = vaddq_s16(tmp12, z5);
+
+  int16x8_t tmp6 = vsubq_s16(tmp12, tmp7);   /* phase 2 */
+  int16x8_t tmp5 = vsubq_s16(tmp11, tmp6);
+  int16x8_t tmp4 = vaddq_s16(tmp10, tmp5);
+
+  col0 = vaddq_s16(tmp0, tmp7);
+  col7 = vsubq_s16(tmp0, tmp7);
+  col1 = vaddq_s16(tmp1, tmp6);
+  col6 = vsubq_s16(tmp1, tmp6);
+  col2 = vaddq_s16(tmp2, tmp5);
+  col5 = vsubq_s16(tmp2, tmp5);
+  col4 = vaddq_s16(tmp3, tmp4);
+  col3 = vsubq_s16(tmp3, tmp4);
+
+  /* Scale down by a factor of 8, narrowing to 8-bit. */
+  int8x16_t cols_01_s8 = vcombine_s8(vqshrn_n_s16(col0, PASS1_BITS + 3),
+                                     vqshrn_n_s16(col1, PASS1_BITS + 3));
+  int8x16_t cols_45_s8 = vcombine_s8(vqshrn_n_s16(col4, PASS1_BITS + 3),
+                                     vqshrn_n_s16(col5, PASS1_BITS + 3));
+  int8x16_t cols_23_s8 = vcombine_s8(vqshrn_n_s16(col2, PASS1_BITS + 3),
+                                     vqshrn_n_s16(col3, PASS1_BITS + 3));
+  int8x16_t cols_67_s8 = vcombine_s8(vqshrn_n_s16(col6, PASS1_BITS + 3),
+                                     vqshrn_n_s16(col7, PASS1_BITS + 3));
+  /* Clamp to range [0-255]. */
+  uint8x16_t cols_01 =
+    vreinterpretq_u8_s8
+      (vaddq_s8(cols_01_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
+  uint8x16_t cols_45 =
+    vreinterpretq_u8_s8
+      (vaddq_s8(cols_45_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
+  uint8x16_t cols_23 =
+    vreinterpretq_u8_s8
+      (vaddq_s8(cols_23_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
+  uint8x16_t cols_67 =
+    vreinterpretq_u8_s8
+      (vaddq_s8(cols_67_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
+
+  /* Transpose block to prepare for store. */
+  uint32x4x2_t cols_0415 = vzipq_u32(vreinterpretq_u32_u8(cols_01),
+                                     vreinterpretq_u32_u8(cols_45));
+  uint32x4x2_t cols_2637 = vzipq_u32(vreinterpretq_u32_u8(cols_23),
+                                     vreinterpretq_u32_u8(cols_67));
+
+  uint8x16x2_t cols_0145 = vtrnq_u8(vreinterpretq_u8_u32(cols_0415.val[0]),
+                                    vreinterpretq_u8_u32(cols_0415.val[1]));
+  uint8x16x2_t cols_2367 = vtrnq_u8(vreinterpretq_u8_u32(cols_2637.val[0]),
+                                    vreinterpretq_u8_u32(cols_2637.val[1]));
+  uint16x8x2_t rows_0426 = vtrnq_u16(vreinterpretq_u16_u8(cols_0145.val[0]),
+                                     vreinterpretq_u16_u8(cols_2367.val[0]));
+  uint16x8x2_t rows_1537 = vtrnq_u16(vreinterpretq_u16_u8(cols_0145.val[1]),
+                                     vreinterpretq_u16_u8(cols_2367.val[1]));
+
+  uint8x16_t rows_04 = vreinterpretq_u8_u16(rows_0426.val[0]);
+  uint8x16_t rows_15 = vreinterpretq_u8_u16(rows_1537.val[0]);
+  uint8x16_t rows_26 = vreinterpretq_u8_u16(rows_0426.val[1]);
+  uint8x16_t rows_37 = vreinterpretq_u8_u16(rows_1537.val[1]);
+
+  JSAMPROW outptr0 = output_buf[0] + output_col;
+  JSAMPROW outptr1 = output_buf[1] + output_col;
+  JSAMPROW outptr2 = output_buf[2] + output_col;
+  JSAMPROW outptr3 = output_buf[3] + output_col;
+  JSAMPROW outptr4 = output_buf[4] + output_col;
+  JSAMPROW outptr5 = output_buf[5] + output_col;
+  JSAMPROW outptr6 = output_buf[6] + output_col;
+  JSAMPROW outptr7 = output_buf[7] + output_col;
+
+  /* Store DCT block to memory. */
+  vst1q_lane_u64((uint64_t *)outptr0, vreinterpretq_u64_u8(rows_04), 0);
+  vst1q_lane_u64((uint64_t *)outptr1, vreinterpretq_u64_u8(rows_15), 0);
+  vst1q_lane_u64((uint64_t *)outptr2, vreinterpretq_u64_u8(rows_26), 0);
+  vst1q_lane_u64((uint64_t *)outptr3, vreinterpretq_u64_u8(rows_37), 0);
+  vst1q_lane_u64((uint64_t *)outptr4, vreinterpretq_u64_u8(rows_04), 1);
+  vst1q_lane_u64((uint64_t *)outptr5, vreinterpretq_u64_u8(rows_15), 1);
+  vst1q_lane_u64((uint64_t *)outptr6, vreinterpretq_u64_u8(rows_26), 1);
+  vst1q_lane_u64((uint64_t *)outptr7, vreinterpretq_u64_u8(rows_37), 1);
+}
@@ -0,0 +1,802 @@
+/*
+ * jidctint-neon.c - accurate integer IDCT (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "jconfigint.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+#include "neon-compat.h"
+
+#include <arm_neon.h>
+
+
+#define CONST_BITS  13
+#define PASS1_BITS  2
+
+#define DESCALE_P1  (CONST_BITS - PASS1_BITS)
+#define DESCALE_P2  (CONST_BITS + PASS1_BITS + 3)
+
+/* The computation of the inverse DCT requires the use of constants known at
+ * compile time.  Scaled integer constants are used to avoid floating-point
+ * arithmetic:
+ *    0.298631336 =  2446 * 2^-13
+ *    0.390180644 =  3196 * 2^-13
+ *    0.541196100 =  4433 * 2^-13
+ *    0.765366865 =  6270 * 2^-13
+ *    0.899976223 =  7373 * 2^-13
+ *    1.175875602 =  9633 * 2^-13
+ *    1.501321110 = 12299 * 2^-13
+ *    1.847759065 = 15137 * 2^-13
+ *    1.961570560 = 16069 * 2^-13
+ *    2.053119869 = 16819 * 2^-13
+ *    2.562915447 = 20995 * 2^-13
+ *    3.072711026 = 25172 * 2^-13
+ */
+
+#define F_0_298  2446
+#define F_0_390  3196
+#define F_0_541  4433
+#define F_0_765  6270
+#define F_0_899  7373
+#define F_1_175  9633
+#define F_1_501  12299
+#define F_1_847  15137
+#define F_1_961  16069
+#define F_2_053  16819
+#define F_2_562  20995
+#define F_3_072  25172
+
+#define F_1_175_MINUS_1_961  (F_1_175 - F_1_961)
+#define F_1_175_MINUS_0_390  (F_1_175 - F_0_390)
+#define F_0_541_MINUS_1_847  (F_0_541 - F_1_847)
+#define F_3_072_MINUS_2_562  (F_3_072 - F_2_562)
+#define F_0_298_MINUS_0_899  (F_0_298 - F_0_899)
+#define F_1_501_MINUS_0_899  (F_1_501 - F_0_899)
+#define F_2_053_MINUS_2_562  (F_2_053 - F_2_562)
+#define F_0_541_PLUS_0_765   (F_0_541 + F_0_765)
+
+
+ALIGN(16) static const int16_t jsimd_idct_islow_neon_consts[] = {
+  F_0_899,             F_0_541,
+  F_2_562,             F_0_298_MINUS_0_899,
+  F_1_501_MINUS_0_899, F_2_053_MINUS_2_562,
+  F_0_541_PLUS_0_765,  F_1_175,
+  F_1_175_MINUS_0_390, F_0_541_MINUS_1_847,
+  F_3_072_MINUS_2_562, F_1_175_MINUS_1_961,
+  0, 0, 0, 0
+};
+
+
+/* Forward declaration of regular and sparse IDCT helper functions */
+
+static INLINE void jsimd_idct_islow_pass1_regular(int16x4_t row0,
+                                                  int16x4_t row1,
+                                                  int16x4_t row2,
+                                                  int16x4_t row3,
+                                                  int16x4_t row4,
+                                                  int16x4_t row5,
+                                                  int16x4_t row6,
+                                                  int16x4_t row7,
+                                                  int16x4_t quant_row0,
+                                                  int16x4_t quant_row1,
+                                                  int16x4_t quant_row2,
+                                                  int16x4_t quant_row3,
+                                                  int16x4_t quant_row4,
+                                                  int16x4_t quant_row5,
+                                                  int16x4_t quant_row6,
+                                                  int16x4_t quant_row7,
+                                                  int16_t *workspace_1,
+                                                  int16_t *workspace_2);
+
+static INLINE void jsimd_idct_islow_pass1_sparse(int16x4_t row0,
+                                                 int16x4_t row1,
+                                                 int16x4_t row2,
+                                                 int16x4_t row3,
+                                                 int16x4_t quant_row0,
+                                                 int16x4_t quant_row1,
+                                                 int16x4_t quant_row2,
+                                                 int16x4_t quant_row3,
+                                                 int16_t *workspace_1,
+                                                 int16_t *workspace_2);
+
+static INLINE void jsimd_idct_islow_pass2_regular(int16_t *workspace,
+                                                  JSAMPARRAY output_buf,
+                                                  JDIMENSION output_col,
+                                                  unsigned buf_offset);
+
+static INLINE void jsimd_idct_islow_pass2_sparse(int16_t *workspace,
+                                                 JSAMPARRAY output_buf,
+                                                 JDIMENSION output_col,
+                                                 unsigned buf_offset);
+
+
+/* Perform dequantization and inverse DCT on one block of coefficients.  For
+ * reference, the C implementation (jpeg_idct_slow()) can be found in
+ * jidctint.c.
+ *
+ * Optimization techniques used for fast data access:
+ *
+ * In each pass, the inverse DCT is computed for the left and right 4x8 halves
+ * of the DCT block.  This avoids spilling due to register pressure, and the
+ * increased granularity allows for an optimized calculation depending on the
+ * values of the DCT coefficients.  Between passes, intermediate data is stored
+ * in 4x8 workspace buffers.
+ *
+ * Transposing the 8x8 DCT block after each pass can be achieved by transposing
+ * each of the four 4x4 quadrants and swapping quadrants 1 and 2 (refer to the
+ * diagram below.)  Swapping quadrants is cheap, since the second pass can just
+ * swap the workspace buffer pointers.
+ *
+ *      +-------+-------+                   +-------+-------+
+ *      |       |       |                   |       |       |
+ *      |   0   |   1   |                   |   0   |   2   |
+ *      |       |       |    transpose      |       |       |
+ *      +-------+-------+     ------>       +-------+-------+
+ *      |       |       |                   |       |       |
+ *      |   2   |   3   |                   |   1   |   3   |
+ *      |       |       |                   |       |       |
+ *      +-------+-------+                   +-------+-------+
+ *
+ * Optimization techniques used to accelerate the inverse DCT calculation:
+ *
+ * In a DCT coefficient block, the coefficients are increasingly likely to be 0
+ * as you move diagonally from top left to bottom right.  If whole rows of
+ * coefficients are 0, then the inverse DCT calculation can be simplified.  On
+ * the first pass of the inverse DCT, we test for three special cases before
+ * defaulting to a full "regular" inverse DCT:
+ *
+ * 1) Coefficients in rows 4-7 are all zero.  In this case, we perform a
+ *    "sparse" simplified inverse DCT on rows 0-3.
+ * 2) AC coefficients (rows 1-7) are all zero.  In this case, the inverse DCT
+ *    result is equal to the dequantized DC coefficients.
+ * 3) AC and DC coefficients are all zero.  In this case, the inverse DCT
+ *    result is all zero.  For the left 4x8 half, this is handled identically
+ *    to Case 2 above.  For the right 4x8 half, we do no work and signal that
+ *    the "sparse" algorithm is required for the second pass.
+ *
+ * In the second pass, only a single special case is tested: whether the AC and
+ * DC coefficients were all zero in the right 4x8 block during the first pass
+ * (refer to Case 3 above.)  If this is the case, then a "sparse" variant of
+ * the second pass is performed for both the left and right halves of the DCT
+ * block.  (The transposition after the first pass means that the right 4x8
+ * block during the first pass becomes rows 4-7 during the second pass.)
+ */
+
+void jsimd_idct_islow_neon(void *dct_table, JCOEFPTR coef_block,
+                           JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  ISLOW_MULT_TYPE *quantptr = dct_table;
+
+  int16_t workspace_l[8 * DCTSIZE / 2];
+  int16_t workspace_r[8 * DCTSIZE / 2];
+
+  /* Compute IDCT first pass on left 4x8 coefficient block. */
+
+  /* Load DCT coefficients in left 4x8 block. */
+  int16x4_t row0 = vld1_s16(coef_block + 0 * DCTSIZE);
+  int16x4_t row1 = vld1_s16(coef_block + 1 * DCTSIZE);
+  int16x4_t row2 = vld1_s16(coef_block + 2 * DCTSIZE);
+  int16x4_t row3 = vld1_s16(coef_block + 3 * DCTSIZE);
+  int16x4_t row4 = vld1_s16(coef_block + 4 * DCTSIZE);
+  int16x4_t row5 = vld1_s16(coef_block + 5 * DCTSIZE);
+  int16x4_t row6 = vld1_s16(coef_block + 6 * DCTSIZE);
+  int16x4_t row7 = vld1_s16(coef_block + 7 * DCTSIZE);
+
+  /* Load quantization table for left 4x8 block. */
+  int16x4_t quant_row0 = vld1_s16(quantptr + 0 * DCTSIZE);
+  int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE);
+  int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE);
+  int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE);
+  int16x4_t quant_row4 = vld1_s16(quantptr + 4 * DCTSIZE);
+  int16x4_t quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE);
+  int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE);
+  int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE);
+
+  /* Construct bitmap to test if DCT coefficients in left 4x8 block are 0. */
+  int16x4_t bitmap = vorr_s16(row7, row6);
+  bitmap = vorr_s16(bitmap, row5);
+  bitmap = vorr_s16(bitmap, row4);
+  int64_t bitmap_rows_4567 = vget_lane_s64(vreinterpret_s64_s16(bitmap), 0);
+
+  if (bitmap_rows_4567 == 0) {
+    bitmap = vorr_s16(bitmap, row3);
+    bitmap = vorr_s16(bitmap, row2);
+    bitmap = vorr_s16(bitmap, row1);
+    int64_t left_ac_bitmap = vget_lane_s64(vreinterpret_s64_s16(bitmap), 0);
+
+    if (left_ac_bitmap == 0) {
+      int16x4_t dcval = vshl_n_s16(vmul_s16(row0, quant_row0), PASS1_BITS);
+      int16x4x4_t quadrant = { { dcval, dcval, dcval, dcval } };
+      /* Store 4x4 blocks to workspace, transposing in the process. */
+      vst4_s16(workspace_l, quadrant);
+      vst4_s16(workspace_r, quadrant);
+    } else {
+      jsimd_idct_islow_pass1_sparse(row0, row1, row2, row3, quant_row0,
+                                    quant_row1, quant_row2, quant_row3,
+                                    workspace_l, workspace_r);
+    }
+  } else {
+    jsimd_idct_islow_pass1_regular(row0, row1, row2, row3, row4, row5,
+                                   row6, row7, quant_row0, quant_row1,
+                                   quant_row2, quant_row3, quant_row4,
+                                   quant_row5, quant_row6, quant_row7,
+                                   workspace_l, workspace_r);
+  }
+
+  /* Compute IDCT first pass on right 4x8 coefficient block. */
+
+  /* Load DCT coefficients in right 4x8 block. */
+  row0 = vld1_s16(coef_block + 0 * DCTSIZE + 4);
+  row1 = vld1_s16(coef_block + 1 * DCTSIZE + 4);
+  row2 = vld1_s16(coef_block + 2 * DCTSIZE + 4);
+  row3 = vld1_s16(coef_block + 3 * DCTSIZE + 4);
+  row4 = vld1_s16(coef_block + 4 * DCTSIZE + 4);
+  row5 = vld1_s16(coef_block + 5 * DCTSIZE + 4);
+  row6 = vld1_s16(coef_block + 6 * DCTSIZE + 4);
+  row7 = vld1_s16(coef_block + 7 * DCTSIZE + 4);
+
+  /* Load quantization table for right 4x8 block. */
+  quant_row0 = vld1_s16(quantptr + 0 * DCTSIZE + 4);
+  quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE + 4);
+  quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE + 4);
+  quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE + 4);
+  quant_row4 = vld1_s16(quantptr + 4 * DCTSIZE + 4);
+  quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE + 4);
+  quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE + 4);
+  quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE + 4);
+
+  /* Construct bitmap to test if DCT coefficients in right 4x8 block are 0. */
+  bitmap = vorr_s16(row7, row6);
+  bitmap = vorr_s16(bitmap, row5);
+  bitmap = vorr_s16(bitmap, row4);
+  bitmap_rows_4567 = vget_lane_s64(vreinterpret_s64_s16(bitmap), 0);
+  bitmap = vorr_s16(bitmap, row3);
+  bitmap = vorr_s16(bitmap, row2);
+  bitmap = vorr_s16(bitmap, row1);
+  int64_t right_ac_bitmap = vget_lane_s64(vreinterpret_s64_s16(bitmap), 0);
+
+  /* If this remains non-zero, a "regular" second pass will be performed. */
+  int64_t right_ac_dc_bitmap = 1;
+
+  if (right_ac_bitmap == 0) {
+    bitmap = vorr_s16(bitmap, row0);
+    right_ac_dc_bitmap = vget_lane_s64(vreinterpret_s64_s16(bitmap), 0);
+
+    if (right_ac_dc_bitmap != 0) {
+      int16x4_t dcval = vshl_n_s16(vmul_s16(row0, quant_row0), PASS1_BITS);
+      int16x4x4_t quadrant = { { dcval, dcval, dcval, dcval } };
+      /* Store 4x4 blocks to workspace, transposing in the process. */
+      vst4_s16(workspace_l + 4 * DCTSIZE / 2, quadrant);
+      vst4_s16(workspace_r + 4 * DCTSIZE / 2, quadrant);
+    }
+  } else {
+    if (bitmap_rows_4567 == 0) {
+      jsimd_idct_islow_pass1_sparse(row0, row1, row2, row3, quant_row0,
+                                    quant_row1, quant_row2, quant_row3,
+                                    workspace_l + 4 * DCTSIZE / 2,
+                                    workspace_r + 4 * DCTSIZE / 2);
+    } else {
+      jsimd_idct_islow_pass1_regular(row0, row1, row2, row3, row4, row5,
+                                     row6, row7, quant_row0, quant_row1,
+                                     quant_row2, quant_row3, quant_row4,
+                                     quant_row5, quant_row6, quant_row7,
+                                     workspace_l + 4 * DCTSIZE / 2,
+                                     workspace_r + 4 * DCTSIZE / 2);
+    }
+  }
+
+  /* Second pass: compute IDCT on rows in workspace. */
+
+  /* If all coefficients in right 4x8 block are 0, use "sparse" second pass. */
+  if (right_ac_dc_bitmap == 0) {
+    jsimd_idct_islow_pass2_sparse(workspace_l, output_buf, output_col, 0);
+    jsimd_idct_islow_pass2_sparse(workspace_r, output_buf, output_col, 4);
+  } else {
+    jsimd_idct_islow_pass2_regular(workspace_l, output_buf, output_col, 0);
+    jsimd_idct_islow_pass2_regular(workspace_r, output_buf, output_col, 4);
+  }
+}
+
+
+/* Perform dequantization and the first pass of the accurate inverse DCT on a
+ * 4x8 block of coefficients.  (To process the full 8x8 DCT block, this
+ * function-- or some other optimized variant-- needs to be called for both the
+ * left and right 4x8 blocks.)
+ *
+ * This "regular" version assumes that no optimization can be made to the IDCT
+ * calculation, since no useful set of AC coefficients is all 0.
+ *
+ * The original C implementation of the accurate IDCT (jpeg_idct_slow()) can be
+ * found in jidctint.c.  Algorithmic changes made here are documented inline.
+ */
+
+static INLINE void jsimd_idct_islow_pass1_regular(int16x4_t row0,
+                                                  int16x4_t row1,
+                                                  int16x4_t row2,
+                                                  int16x4_t row3,
+                                                  int16x4_t row4,
+                                                  int16x4_t row5,
+                                                  int16x4_t row6,
+                                                  int16x4_t row7,
+                                                  int16x4_t quant_row0,
+                                                  int16x4_t quant_row1,
+                                                  int16x4_t quant_row2,
+                                                  int16x4_t quant_row3,
+                                                  int16x4_t quant_row4,
+                                                  int16x4_t quant_row5,
+                                                  int16x4_t quant_row6,
+                                                  int16x4_t quant_row7,
+                                                  int16_t *workspace_1,
+                                                  int16_t *workspace_2)
+{
+  /* Load constants for IDCT computation. */
+#ifdef HAVE_VLD1_S16_X3
+  const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_islow_neon_consts);
+#else
+  const int16x4_t consts1 = vld1_s16(jsimd_idct_islow_neon_consts);
+  const int16x4_t consts2 = vld1_s16(jsimd_idct_islow_neon_consts + 4);
+  const int16x4_t consts3 = vld1_s16(jsimd_idct_islow_neon_consts + 8);
+  const int16x4x3_t consts = { { consts1, consts2, consts3 } };
+#endif
+
+  /* Even part */
+  int16x4_t z2_s16 = vmul_s16(row2, quant_row2);
+  int16x4_t z3_s16 = vmul_s16(row6, quant_row6);
+
+  int32x4_t tmp2 = vmull_lane_s16(z2_s16, consts.val[0], 1);
+  int32x4_t tmp3 = vmull_lane_s16(z2_s16, consts.val[1], 2);
+  tmp2 = vmlal_lane_s16(tmp2, z3_s16, consts.val[2], 1);
+  tmp3 = vmlal_lane_s16(tmp3, z3_s16, consts.val[0], 1);
+
+  z2_s16 = vmul_s16(row0, quant_row0);
+  z3_s16 = vmul_s16(row4, quant_row4);
+
+  int32x4_t tmp0 = vshll_n_s16(vadd_s16(z2_s16, z3_s16), CONST_BITS);
+  int32x4_t tmp1 = vshll_n_s16(vsub_s16(z2_s16, z3_s16), CONST_BITS);
+
+  int32x4_t tmp10 = vaddq_s32(tmp0, tmp3);
+  int32x4_t tmp13 = vsubq_s32(tmp0, tmp3);
+  int32x4_t tmp11 = vaddq_s32(tmp1, tmp2);
+  int32x4_t tmp12 = vsubq_s32(tmp1, tmp2);
+
+  /* Odd part */
+  int16x4_t tmp0_s16 = vmul_s16(row7, quant_row7);
+  int16x4_t tmp1_s16 = vmul_s16(row5, quant_row5);
+  int16x4_t tmp2_s16 = vmul_s16(row3, quant_row3);
+  int16x4_t tmp3_s16 = vmul_s16(row1, quant_row1);
+
+  z3_s16 = vadd_s16(tmp0_s16, tmp2_s16);
+  int16x4_t z4_s16 = vadd_s16(tmp1_s16, tmp3_s16);
+
+  /* Implementation as per jpeg_idct_islow() in jidctint.c:
+   *   z5 = (z3 + z4) * 1.175875602;
+   *   z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+   *   z3 += z5;  z4 += z5;
+   *
+   * This implementation:
+   *   z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+   *   z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+   */
+
+  int32x4_t z3 = vmull_lane_s16(z3_s16, consts.val[2], 3);
+  int32x4_t z4 = vmull_lane_s16(z3_s16, consts.val[1], 3);
+  z3 = vmlal_lane_s16(z3, z4_s16, consts.val[1], 3);
+  z4 = vmlal_lane_s16(z4, z4_s16, consts.val[2], 0);
+
+  /* Implementation as per jpeg_idct_islow() in jidctint.c:
+   *   z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
+   *   tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
+   *   tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
+   *   z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+   *   tmp0 += z1 + z3;  tmp1 += z2 + z4;
+   *   tmp2 += z2 + z3;  tmp3 += z1 + z4;
+   *
+   * This implementation:
+   *   tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
+   *   tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
+   *   tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
+   *   tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
+   *   tmp0 += z3;  tmp1 += z4;
+   *   tmp2 += z3;  tmp3 += z4;
+   */
+
+  tmp0 = vmull_lane_s16(tmp0_s16, consts.val[0], 3);
+  tmp1 = vmull_lane_s16(tmp1_s16, consts.val[1], 1);
+  tmp2 = vmull_lane_s16(tmp2_s16, consts.val[2], 2);
+  tmp3 = vmull_lane_s16(tmp3_s16, consts.val[1], 0);
+
+  tmp0 = vmlsl_lane_s16(tmp0, tmp3_s16, consts.val[0], 0);
+  tmp1 = vmlsl_lane_s16(tmp1, tmp2_s16, consts.val[0], 2);
+  tmp2 = vmlsl_lane_s16(tmp2, tmp1_s16, consts.val[0], 2);
+  tmp3 = vmlsl_lane_s16(tmp3, tmp0_s16, consts.val[0], 0);
+
+  tmp0 = vaddq_s32(tmp0, z3);
+  tmp1 = vaddq_s32(tmp1, z4);
+  tmp2 = vaddq_s32(tmp2, z3);
+  tmp3 = vaddq_s32(tmp3, z4);
+
+  /* Final output stage: descale and narrow to 16-bit. */
+  int16x4x4_t rows_0123 = { {
+    vrshrn_n_s32(vaddq_s32(tmp10, tmp3), DESCALE_P1),
+    vrshrn_n_s32(vaddq_s32(tmp11, tmp2), DESCALE_P1),
+    vrshrn_n_s32(vaddq_s32(tmp12, tmp1), DESCALE_P1),
+    vrshrn_n_s32(vaddq_s32(tmp13, tmp0), DESCALE_P1)
+  } };
+  int16x4x4_t rows_4567 = { {
+    vrshrn_n_s32(vsubq_s32(tmp13, tmp0), DESCALE_P1),
+    vrshrn_n_s32(vsubq_s32(tmp12, tmp1), DESCALE_P1),
+    vrshrn_n_s32(vsubq_s32(tmp11, tmp2), DESCALE_P1),
+    vrshrn_n_s32(vsubq_s32(tmp10, tmp3), DESCALE_P1)
+  } };
+
+  /* Store 4x4 blocks to the intermediate workspace, ready for the second pass.
+   * (VST4 transposes the blocks.  We need to operate on rows in the next
+   * pass.)
+   */
+  vst4_s16(workspace_1, rows_0123);
+  vst4_s16(workspace_2, rows_4567);
+}
+
+
+/* Perform dequantization and the first pass of the accurate inverse DCT on a
+ * 4x8 block of coefficients.
+ *
+ * This "sparse" version assumes that the AC coefficients in rows 4-7 are all
+ * 0.  This simplifies the IDCT calculation, accelerating overall performance.
+ */
+
+static INLINE void jsimd_idct_islow_pass1_sparse(int16x4_t row0,
+                                                 int16x4_t row1,
+                                                 int16x4_t row2,
+                                                 int16x4_t row3,
+                                                 int16x4_t quant_row0,
+                                                 int16x4_t quant_row1,
+                                                 int16x4_t quant_row2,
+                                                 int16x4_t quant_row3,
+                                                 int16_t *workspace_1,
+                                                 int16_t *workspace_2)
+{
+  /* Load constants for IDCT computation. */
+#ifdef HAVE_VLD1_S16_X3
+  const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_islow_neon_consts);
+#else
+  const int16x4_t consts1 = vld1_s16(jsimd_idct_islow_neon_consts);
+  const int16x4_t consts2 = vld1_s16(jsimd_idct_islow_neon_consts + 4);
+  const int16x4_t consts3 = vld1_s16(jsimd_idct_islow_neon_consts + 8);
+  const int16x4x3_t consts = { { consts1, consts2, consts3 } };
+#endif
+
+  /* Even part (z3 is all 0) */
+  int16x4_t z2_s16 = vmul_s16(row2, quant_row2);
+
+  int32x4_t tmp2 = vmull_lane_s16(z2_s16, consts.val[0], 1);
+  int32x4_t tmp3 = vmull_lane_s16(z2_s16, consts.val[1], 2);
+
+  z2_s16 = vmul_s16(row0, quant_row0);
+  int32x4_t tmp0 = vshll_n_s16(z2_s16, CONST_BITS);
+  int32x4_t tmp1 = vshll_n_s16(z2_s16, CONST_BITS);
+
+  int32x4_t tmp10 = vaddq_s32(tmp0, tmp3);
+  int32x4_t tmp13 = vsubq_s32(tmp0, tmp3);
+  int32x4_t tmp11 = vaddq_s32(tmp1, tmp2);
+  int32x4_t tmp12 = vsubq_s32(tmp1, tmp2);
+
+  /* Odd part (tmp0 and tmp1 are both all 0) */
+  int16x4_t tmp2_s16 = vmul_s16(row3, quant_row3);
+  int16x4_t tmp3_s16 = vmul_s16(row1, quant_row1);
+
+  int16x4_t z3_s16 = tmp2_s16;
+  int16x4_t z4_s16 = tmp3_s16;
+
+  int32x4_t z3 = vmull_lane_s16(z3_s16, consts.val[2], 3);
+  int32x4_t z4 = vmull_lane_s16(z3_s16, consts.val[1], 3);
+  z3 = vmlal_lane_s16(z3, z4_s16, consts.val[1], 3);
+  z4 = vmlal_lane_s16(z4, z4_s16, consts.val[2], 0);
+
+  tmp0 = vmlsl_lane_s16(z3, tmp3_s16, consts.val[0], 0);
+  tmp1 = vmlsl_lane_s16(z4, tmp2_s16, consts.val[0], 2);
+  tmp2 = vmlal_lane_s16(z3, tmp2_s16, consts.val[2], 2);
+  tmp3 = vmlal_lane_s16(z4, tmp3_s16, consts.val[1], 0);
+
+  /* Final output stage: descale and narrow to 16-bit. */
+  int16x4x4_t rows_0123 = { {
+    vrshrn_n_s32(vaddq_s32(tmp10, tmp3), DESCALE_P1),
+    vrshrn_n_s32(vaddq_s32(tmp11, tmp2), DESCALE_P1),
+    vrshrn_n_s32(vaddq_s32(tmp12, tmp1), DESCALE_P1),
+    vrshrn_n_s32(vaddq_s32(tmp13, tmp0), DESCALE_P1)
+  } };
+  int16x4x4_t rows_4567 = { {
+    vrshrn_n_s32(vsubq_s32(tmp13, tmp0), DESCALE_P1),
+    vrshrn_n_s32(vsubq_s32(tmp12, tmp1), DESCALE_P1),
+    vrshrn_n_s32(vsubq_s32(tmp11, tmp2), DESCALE_P1),
+    vrshrn_n_s32(vsubq_s32(tmp10, tmp3), DESCALE_P1)
+  } };
+
+  /* Store 4x4 blocks to the intermediate workspace, ready for the second pass.
+   * (VST4 transposes the blocks.  We need to operate on rows in the next
+   * pass.)
+   */
+  vst4_s16(workspace_1, rows_0123);
+  vst4_s16(workspace_2, rows_4567);
+}
+
+
+/* Perform the second pass of the accurate inverse DCT on a 4x8 block of
+ * coefficients.  (To process the full 8x8 DCT block, this function-- or some
+ * other optimized variant-- needs to be called for both the right and left 4x8
+ * blocks.)
+ *
+ * This "regular" version assumes that no optimization can be made to the IDCT
+ * calculation, since no useful set of coefficient values are all 0 after the
+ * first pass.
+ *
+ * Again, the original C implementation of the accurate IDCT (jpeg_idct_slow())
+ * can be found in jidctint.c.  Algorithmic changes made here are documented
+ * inline.
+ */
+
+static INLINE void jsimd_idct_islow_pass2_regular(int16_t *workspace,
+                                                  JSAMPARRAY output_buf,
+                                                  JDIMENSION output_col,
+                                                  unsigned buf_offset)
+{
+  /* Load constants for IDCT computation. */
+#ifdef HAVE_VLD1_S16_X3
+  const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_islow_neon_consts);
+#else
+  const int16x4_t consts1 = vld1_s16(jsimd_idct_islow_neon_consts);
+  const int16x4_t consts2 = vld1_s16(jsimd_idct_islow_neon_consts + 4);
+  const int16x4_t consts3 = vld1_s16(jsimd_idct_islow_neon_consts + 8);
+  const int16x4x3_t consts = { { consts1, consts2, consts3 } };
+#endif
+
+  /* Even part */
+  int16x4_t z2_s16 = vld1_s16(workspace + 2 * DCTSIZE / 2);
+  int16x4_t z3_s16 = vld1_s16(workspace + 6 * DCTSIZE / 2);
+
+  int32x4_t tmp2 = vmull_lane_s16(z2_s16, consts.val[0], 1);
+  int32x4_t tmp3 = vmull_lane_s16(z2_s16, consts.val[1], 2);
+  tmp2 = vmlal_lane_s16(tmp2, z3_s16, consts.val[2], 1);
+  tmp3 = vmlal_lane_s16(tmp3, z3_s16, consts.val[0], 1);
+
+  z2_s16 = vld1_s16(workspace + 0 * DCTSIZE / 2);
+  z3_s16 = vld1_s16(workspace + 4 * DCTSIZE / 2);
+
+  int32x4_t tmp0 = vshll_n_s16(vadd_s16(z2_s16, z3_s16), CONST_BITS);
+  int32x4_t tmp1 = vshll_n_s16(vsub_s16(z2_s16, z3_s16), CONST_BITS);
+
+  int32x4_t tmp10 = vaddq_s32(tmp0, tmp3);
+  int32x4_t tmp13 = vsubq_s32(tmp0, tmp3);
+  int32x4_t tmp11 = vaddq_s32(tmp1, tmp2);
+  int32x4_t tmp12 = vsubq_s32(tmp1, tmp2);
+
+  /* Odd part */
+  int16x4_t tmp0_s16 = vld1_s16(workspace + 7 * DCTSIZE / 2);
+  int16x4_t tmp1_s16 = vld1_s16(workspace + 5 * DCTSIZE / 2);
+  int16x4_t tmp2_s16 = vld1_s16(workspace + 3 * DCTSIZE / 2);
+  int16x4_t tmp3_s16 = vld1_s16(workspace + 1 * DCTSIZE / 2);
+
+  z3_s16 = vadd_s16(tmp0_s16, tmp2_s16);
+  int16x4_t z4_s16 = vadd_s16(tmp1_s16, tmp3_s16);
+
+  /* Implementation as per jpeg_idct_islow() in jidctint.c:
+   *   z5 = (z3 + z4) * 1.175875602;
+   *   z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
+   *   z3 += z5;  z4 += z5;
+   *
+   * This implementation:
+   *   z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
+   *   z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
+   */
+
+  int32x4_t z3 = vmull_lane_s16(z3_s16, consts.val[2], 3);
+  int32x4_t z4 = vmull_lane_s16(z3_s16, consts.val[1], 3);
+  z3 = vmlal_lane_s16(z3, z4_s16, consts.val[1], 3);
+  z4 = vmlal_lane_s16(z4, z4_s16, consts.val[2], 0);
+
+  /* Implementation as per jpeg_idct_islow() in jidctint.c:
+   *   z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
+   *   tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
+   *   tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
+   *   z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
+   *   tmp0 += z1 + z3;  tmp1 += z2 + z4;
+   *   tmp2 += z2 + z3;  tmp3 += z1 + z4;
+   *
+   * This implementation:
+   *   tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
+   *   tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
+   *   tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
+   *   tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
+   *   tmp0 += z3;  tmp1 += z4;
+   *   tmp2 += z3;  tmp3 += z4;
+   */
+
+  tmp0 = vmull_lane_s16(tmp0_s16, consts.val[0], 3);
+  tmp1 = vmull_lane_s16(tmp1_s16, consts.val[1], 1);
+  tmp2 = vmull_lane_s16(tmp2_s16, consts.val[2], 2);
+  tmp3 = vmull_lane_s16(tmp3_s16, consts.val[1], 0);
+
+  tmp0 = vmlsl_lane_s16(tmp0, tmp3_s16, consts.val[0], 0);
+  tmp1 = vmlsl_lane_s16(tmp1, tmp2_s16, consts.val[0], 2);
+  tmp2 = vmlsl_lane_s16(tmp2, tmp1_s16, consts.val[0], 2);
+  tmp3 = vmlsl_lane_s16(tmp3, tmp0_s16, consts.val[0], 0);
+
+  tmp0 = vaddq_s32(tmp0, z3);
+  tmp1 = vaddq_s32(tmp1, z4);
+  tmp2 = vaddq_s32(tmp2, z3);
+  tmp3 = vaddq_s32(tmp3, z4);
+
+  /* Final output stage: descale and narrow to 16-bit. */
+  int16x8_t cols_02_s16 = vcombine_s16(vaddhn_s32(tmp10, tmp3),
+                                       vaddhn_s32(tmp12, tmp1));
+  int16x8_t cols_13_s16 = vcombine_s16(vaddhn_s32(tmp11, tmp2),
+                                       vaddhn_s32(tmp13, tmp0));
+  int16x8_t cols_46_s16 = vcombine_s16(vsubhn_s32(tmp13, tmp0),
+                                       vsubhn_s32(tmp11, tmp2));
+  int16x8_t cols_57_s16 = vcombine_s16(vsubhn_s32(tmp12, tmp1),
+                                       vsubhn_s32(tmp10, tmp3));
+  /* Descale and narrow to 8-bit. */
+  int8x8_t cols_02_s8 = vqrshrn_n_s16(cols_02_s16, DESCALE_P2 - 16);
+  int8x8_t cols_13_s8 = vqrshrn_n_s16(cols_13_s16, DESCALE_P2 - 16);
+  int8x8_t cols_46_s8 = vqrshrn_n_s16(cols_46_s16, DESCALE_P2 - 16);
+  int8x8_t cols_57_s8 = vqrshrn_n_s16(cols_57_s16, DESCALE_P2 - 16);
+  /* Clamp to range [0-255]. */
+  uint8x8_t cols_02_u8 = vadd_u8(vreinterpret_u8_s8(cols_02_s8),
+                                 vdup_n_u8(CENTERJSAMPLE));
+  uint8x8_t cols_13_u8 = vadd_u8(vreinterpret_u8_s8(cols_13_s8),
+                                 vdup_n_u8(CENTERJSAMPLE));
+  uint8x8_t cols_46_u8 = vadd_u8(vreinterpret_u8_s8(cols_46_s8),
+                                 vdup_n_u8(CENTERJSAMPLE));
+  uint8x8_t cols_57_u8 = vadd_u8(vreinterpret_u8_s8(cols_57_s8),
+                                 vdup_n_u8(CENTERJSAMPLE));
+
+  /* Transpose 4x8 block and store to memory.  (Zipping adjacent columns
+   * together allows us to store 16-bit elements.)
+   */
+  uint8x8x2_t cols_01_23 = vzip_u8(cols_02_u8, cols_13_u8);
+  uint8x8x2_t cols_45_67 = vzip_u8(cols_46_u8, cols_57_u8);
+  uint16x4x4_t cols_01_23_45_67 = { {
+    vreinterpret_u16_u8(cols_01_23.val[0]),
+    vreinterpret_u16_u8(cols_01_23.val[1]),
+    vreinterpret_u16_u8(cols_45_67.val[0]),
+    vreinterpret_u16_u8(cols_45_67.val[1])
+  } };
+
+  JSAMPROW outptr0 = output_buf[buf_offset + 0] + output_col;
+  JSAMPROW outptr1 = output_buf[buf_offset + 1] + output_col;
+  JSAMPROW outptr2 = output_buf[buf_offset + 2] + output_col;
+  JSAMPROW outptr3 = output_buf[buf_offset + 3] + output_col;
+  /* VST4 of 16-bit elements completes the transpose. */
+  vst4_lane_u16((uint16_t *)outptr0, cols_01_23_45_67, 0);
+  vst4_lane_u16((uint16_t *)outptr1, cols_01_23_45_67, 1);
+  vst4_lane_u16((uint16_t *)outptr2, cols_01_23_45_67, 2);
+  vst4_lane_u16((uint16_t *)outptr3, cols_01_23_45_67, 3);
+}
+
+
+/* Performs the second pass of the accurate inverse DCT on a 4x8 block
+ * of coefficients.
+ *
+ * This "sparse" version assumes that the coefficient values (after the first
+ * pass) in rows 4-7 are all 0.  This simplifies the IDCT calculation,
+ * accelerating overall performance.
+ */
+
+static INLINE void jsimd_idct_islow_pass2_sparse(int16_t *workspace,
+                                                 JSAMPARRAY output_buf,
+                                                 JDIMENSION output_col,
+                                                 unsigned buf_offset)
+{
+  /* Load constants for IDCT computation. */
+#ifdef HAVE_VLD1_S16_X3
+  const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_islow_neon_consts);
+#else
+  const int16x4_t consts1 = vld1_s16(jsimd_idct_islow_neon_consts);
+  const int16x4_t consts2 = vld1_s16(jsimd_idct_islow_neon_consts + 4);
+  const int16x4_t consts3 = vld1_s16(jsimd_idct_islow_neon_consts + 8);
+  const int16x4x3_t consts = { { consts1, consts2, consts3 } };
+#endif
+
+  /* Even part (z3 is all 0) */
+  int16x4_t z2_s16 = vld1_s16(workspace + 2 * DCTSIZE / 2);
+
+  int32x4_t tmp2 = vmull_lane_s16(z2_s16, consts.val[0], 1);
+  int32x4_t tmp3 = vmull_lane_s16(z2_s16, consts.val[1], 2);
+
+  z2_s16 = vld1_s16(workspace + 0 * DCTSIZE / 2);
+  int32x4_t tmp0 = vshll_n_s16(z2_s16, CONST_BITS);
+  int32x4_t tmp1 = vshll_n_s16(z2_s16, CONST_BITS);
+
+  int32x4_t tmp10 = vaddq_s32(tmp0, tmp3);
+  int32x4_t tmp13 = vsubq_s32(tmp0, tmp3);
+  int32x4_t tmp11 = vaddq_s32(tmp1, tmp2);
+  int32x4_t tmp12 = vsubq_s32(tmp1, tmp2);
+
+  /* Odd part (tmp0 and tmp1 are both all 0) */
+  int16x4_t tmp2_s16 = vld1_s16(workspace + 3 * DCTSIZE / 2);
+  int16x4_t tmp3_s16 = vld1_s16(workspace + 1 * DCTSIZE / 2);
+
+  int16x4_t z3_s16 = tmp2_s16;
+  int16x4_t z4_s16 = tmp3_s16;
+
+  int32x4_t z3 = vmull_lane_s16(z3_s16, consts.val[2], 3);
+  z3 = vmlal_lane_s16(z3, z4_s16, consts.val[1], 3);
+  int32x4_t z4 = vmull_lane_s16(z3_s16, consts.val[1], 3);
+  z4 = vmlal_lane_s16(z4, z4_s16, consts.val[2], 0);
+
+  tmp0 = vmlsl_lane_s16(z3, tmp3_s16, consts.val[0], 0);
+  tmp1 = vmlsl_lane_s16(z4, tmp2_s16, consts.val[0], 2);
+  tmp2 = vmlal_lane_s16(z3, tmp2_s16, consts.val[2], 2);
+  tmp3 = vmlal_lane_s16(z4, tmp3_s16, consts.val[1], 0);
+
+  /* Final output stage: descale and narrow to 16-bit. */
+  int16x8_t cols_02_s16 = vcombine_s16(vaddhn_s32(tmp10, tmp3),
+                                       vaddhn_s32(tmp12, tmp1));
+  int16x8_t cols_13_s16 = vcombine_s16(vaddhn_s32(tmp11, tmp2),
+                                       vaddhn_s32(tmp13, tmp0));
+  int16x8_t cols_46_s16 = vcombine_s16(vsubhn_s32(tmp13, tmp0),
+                                       vsubhn_s32(tmp11, tmp2));
+  int16x8_t cols_57_s16 = vcombine_s16(vsubhn_s32(tmp12, tmp1),
+                                       vsubhn_s32(tmp10, tmp3));
+  /* Descale and narrow to 8-bit. */
+  int8x8_t cols_02_s8 = vqrshrn_n_s16(cols_02_s16, DESCALE_P2 - 16);
+  int8x8_t cols_13_s8 = vqrshrn_n_s16(cols_13_s16, DESCALE_P2 - 16);
+  int8x8_t cols_46_s8 = vqrshrn_n_s16(cols_46_s16, DESCALE_P2 - 16);
+  int8x8_t cols_57_s8 = vqrshrn_n_s16(cols_57_s16, DESCALE_P2 - 16);
+  /* Clamp to range [0-255]. */
+  uint8x8_t cols_02_u8 = vadd_u8(vreinterpret_u8_s8(cols_02_s8),
+                                 vdup_n_u8(CENTERJSAMPLE));
+  uint8x8_t cols_13_u8 = vadd_u8(vreinterpret_u8_s8(cols_13_s8),
+                                 vdup_n_u8(CENTERJSAMPLE));
+  uint8x8_t cols_46_u8 = vadd_u8(vreinterpret_u8_s8(cols_46_s8),
+                                 vdup_n_u8(CENTERJSAMPLE));
+  uint8x8_t cols_57_u8 = vadd_u8(vreinterpret_u8_s8(cols_57_s8),
+                                 vdup_n_u8(CENTERJSAMPLE));
+
+  /* Transpose 4x8 block and store to memory.  (Zipping adjacent columns
+   * together allows us to store 16-bit elements.)
+   */
+  uint8x8x2_t cols_01_23 = vzip_u8(cols_02_u8, cols_13_u8);
+  uint8x8x2_t cols_45_67 = vzip_u8(cols_46_u8, cols_57_u8);
+  uint16x4x4_t cols_01_23_45_67 = { {
+    vreinterpret_u16_u8(cols_01_23.val[0]),
+    vreinterpret_u16_u8(cols_01_23.val[1]),
+    vreinterpret_u16_u8(cols_45_67.val[0]),
+    vreinterpret_u16_u8(cols_45_67.val[1])
+  } };
+
+  JSAMPROW outptr0 = output_buf[buf_offset + 0] + output_col;
+  JSAMPROW outptr1 = output_buf[buf_offset + 1] + output_col;
+  JSAMPROW outptr2 = output_buf[buf_offset + 2] + output_col;
+  JSAMPROW outptr3 = output_buf[buf_offset + 3] + output_col;
+  /* VST4 of 16-bit elements completes the transpose. */
+  vst4_lane_u16((uint16_t *)outptr0, cols_01_23_45_67, 0);
+  vst4_lane_u16((uint16_t *)outptr1, cols_01_23_45_67, 1);
+  vst4_lane_u16((uint16_t *)outptr2, cols_01_23_45_67, 2);
+  vst4_lane_u16((uint16_t *)outptr3, cols_01_23_45_67, 3);
+}
@@ -0,0 +1,486 @@
+/*
+ * jidctred-neon.c - reduced-size IDCT (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+#include "neon-compat.h"
+
+#include <arm_neon.h>
+
+
+#define CONST_BITS  13
+#define PASS1_BITS  2
+
+#define F_0_211  1730
+#define F_0_509  4176
+#define F_0_601  4926
+#define F_0_720  5906
+#define F_0_765  6270
+#define F_0_850  6967
+#define F_0_899  7373
+#define F_1_061  8697
+#define F_1_272  10426
+#define F_1_451  11893
+#define F_1_847  15137
+#define F_2_172  17799
+#define F_2_562  20995
+#define F_3_624  29692
+
+
+/* jsimd_idct_2x2_neon() is an inverse DCT function that produces reduced-size
+ * 2x2 output from an 8x8 DCT block.  It uses the same calculations and
+ * produces exactly the same output as IJG's original jpeg_idct_2x2() function
+ * from jpeg-6b, which can be found in jidctred.c.
+ *
+ * Scaled integer constants are used to avoid floating-point arithmetic:
+ *    0.720959822 =  5906 * 2^-13
+ *    0.850430095 =  6967 * 2^-13
+ *    1.272758580 = 10426 * 2^-13
+ *    3.624509785 = 29692 * 2^-13
+ *
+ * See jidctred.c for further details of the 2x2 IDCT algorithm.  Where
+ * possible, the variable names and comments here in jsimd_idct_2x2_neon()
+ * match up with those in jpeg_idct_2x2().
+ */
+
+ALIGN(16) static const int16_t jsimd_idct_2x2_neon_consts[] = {
+  -F_0_720, F_0_850, -F_1_272, F_3_624
+};
+
+void jsimd_idct_2x2_neon(void *dct_table, JCOEFPTR coef_block,
+                         JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  ISLOW_MULT_TYPE *quantptr = dct_table;
+
+  /* Load DCT coefficients. */
+  int16x8_t row0 = vld1q_s16(coef_block + 0 * DCTSIZE);
+  int16x8_t row1 = vld1q_s16(coef_block + 1 * DCTSIZE);
+  int16x8_t row3 = vld1q_s16(coef_block + 3 * DCTSIZE);
+  int16x8_t row5 = vld1q_s16(coef_block + 5 * DCTSIZE);
+  int16x8_t row7 = vld1q_s16(coef_block + 7 * DCTSIZE);
+
+  /* Load quantization table values. */
+  int16x8_t quant_row0 = vld1q_s16(quantptr + 0 * DCTSIZE);
+  int16x8_t quant_row1 = vld1q_s16(quantptr + 1 * DCTSIZE);
+  int16x8_t quant_row3 = vld1q_s16(quantptr + 3 * DCTSIZE);
+  int16x8_t quant_row5 = vld1q_s16(quantptr + 5 * DCTSIZE);
+  int16x8_t quant_row7 = vld1q_s16(quantptr + 7 * DCTSIZE);
+
+  /* Dequantize DCT coefficients. */
+  row0 = vmulq_s16(row0, quant_row0);
+  row1 = vmulq_s16(row1, quant_row1);
+  row3 = vmulq_s16(row3, quant_row3);
+  row5 = vmulq_s16(row5, quant_row5);
+  row7 = vmulq_s16(row7, quant_row7);
+
+  /* Load IDCT conversion constants. */
+  const int16x4_t consts = vld1_s16(jsimd_idct_2x2_neon_consts);
+
+  /* Pass 1: process columns from input, put results in vectors row0 and
+   * row1.
+   */
+
+  /* Even part */
+  int32x4_t tmp10_l = vshll_n_s16(vget_low_s16(row0), CONST_BITS + 2);
+  int32x4_t tmp10_h = vshll_n_s16(vget_high_s16(row0), CONST_BITS + 2);
+
+  /* Odd part */
+  int32x4_t tmp0_l = vmull_lane_s16(vget_low_s16(row1), consts, 3);
+  tmp0_l = vmlal_lane_s16(tmp0_l, vget_low_s16(row3), consts, 2);
+  tmp0_l = vmlal_lane_s16(tmp0_l, vget_low_s16(row5), consts, 1);
+  tmp0_l = vmlal_lane_s16(tmp0_l, vget_low_s16(row7), consts, 0);
+  int32x4_t tmp0_h = vmull_lane_s16(vget_high_s16(row1), consts, 3);
+  tmp0_h = vmlal_lane_s16(tmp0_h, vget_high_s16(row3), consts, 2);
+  tmp0_h = vmlal_lane_s16(tmp0_h, vget_high_s16(row5), consts, 1);
+  tmp0_h = vmlal_lane_s16(tmp0_h, vget_high_s16(row7), consts, 0);
+
+  /* Final output stage: descale and narrow to 16-bit. */
+  row0 = vcombine_s16(vrshrn_n_s32(vaddq_s32(tmp10_l, tmp0_l), CONST_BITS),
+                      vrshrn_n_s32(vaddq_s32(tmp10_h, tmp0_h), CONST_BITS));
+  row1 = vcombine_s16(vrshrn_n_s32(vsubq_s32(tmp10_l, tmp0_l), CONST_BITS),
+                      vrshrn_n_s32(vsubq_s32(tmp10_h, tmp0_h), CONST_BITS));
+
+  /* Transpose two rows, ready for second pass. */
+  int16x8x2_t cols_0246_1357 = vtrnq_s16(row0, row1);
+  int16x8_t cols_0246 = cols_0246_1357.val[0];
+  int16x8_t cols_1357 = cols_0246_1357.val[1];
+  /* Duplicate columns such that each is accessible in its own vector. */
+  int32x4x2_t cols_1155_3377 = vtrnq_s32(vreinterpretq_s32_s16(cols_1357),
+                                         vreinterpretq_s32_s16(cols_1357));
+  int16x8_t cols_1155 = vreinterpretq_s16_s32(cols_1155_3377.val[0]);
+  int16x8_t cols_3377 = vreinterpretq_s16_s32(cols_1155_3377.val[1]);
+
+  /* Pass 2: process two rows, store to output array. */
+
+  /* Even part: we're only interested in col0; the top half of tmp10 is "don't
+   * care."
+   */
+  int32x4_t tmp10 = vshll_n_s16(vget_low_s16(cols_0246), CONST_BITS + 2);
+
+  /* Odd part: we're only interested in the bottom half of tmp0. */
+  int32x4_t tmp0 = vmull_lane_s16(vget_low_s16(cols_1155), consts, 3);
+  tmp0 = vmlal_lane_s16(tmp0, vget_low_s16(cols_3377), consts, 2);
+  tmp0 = vmlal_lane_s16(tmp0, vget_high_s16(cols_1155), consts, 1);
+  tmp0 = vmlal_lane_s16(tmp0, vget_high_s16(cols_3377), consts, 0);
+
+  /* Final output stage: descale and clamp to range [0-255]. */
+  int16x8_t output_s16 = vcombine_s16(vaddhn_s32(tmp10, tmp0),
+                                      vsubhn_s32(tmp10, tmp0));
+  output_s16 = vrsraq_n_s16(vdupq_n_s16(CENTERJSAMPLE), output_s16,
+                            CONST_BITS + PASS1_BITS + 3 + 2 - 16);
+  /* Narrow to 8-bit and convert to unsigned. */
+  uint8x8_t output_u8 = vqmovun_s16(output_s16);
+
+  /* Store 2x2 block to memory. */
+  vst1_lane_u8(output_buf[0] + output_col, output_u8, 0);
+  vst1_lane_u8(output_buf[1] + output_col, output_u8, 1);
+  vst1_lane_u8(output_buf[0] + output_col + 1, output_u8, 4);
+  vst1_lane_u8(output_buf[1] + output_col + 1, output_u8, 5);
+}
+
+
+/* jsimd_idct_4x4_neon() is an inverse DCT function that produces reduced-size
+ * 4x4 output from an 8x8 DCT block.  It uses the same calculations and
+ * produces exactly the same output as IJG's original jpeg_idct_4x4() function
+ * from jpeg-6b, which can be found in jidctred.c.
+ *
+ * Scaled integer constants are used to avoid floating-point arithmetic:
+ *    0.211164243 =  1730 * 2^-13
+ *    0.509795579 =  4176 * 2^-13
+ *    0.601344887 =  4926 * 2^-13
+ *    0.765366865 =  6270 * 2^-13
+ *    0.899976223 =  7373 * 2^-13
+ *    1.061594337 =  8697 * 2^-13
+ *    1.451774981 = 11893 * 2^-13
+ *    1.847759065 = 15137 * 2^-13
+ *    2.172734803 = 17799 * 2^-13
+ *    2.562915447 = 20995 * 2^-13
+ *
+ * See jidctred.c for further details of the 4x4 IDCT algorithm.  Where
+ * possible, the variable names and comments here in jsimd_idct_4x4_neon()
+ * match up with those in jpeg_idct_4x4().
+ */
+
+ALIGN(16) static const int16_t jsimd_idct_4x4_neon_consts[] = {
+  F_1_847, -F_0_765, -F_0_211,  F_1_451,
+ -F_2_172,  F_1_061, -F_0_509, -F_0_601,
+  F_0_899,  F_2_562,        0,        0
+};
+
+void jsimd_idct_4x4_neon(void *dct_table, JCOEFPTR coef_block,
+                         JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  ISLOW_MULT_TYPE *quantptr = dct_table;
+
+  /* Load DCT coefficients. */
+  int16x8_t row0  = vld1q_s16(coef_block + 0 * DCTSIZE);
+  int16x8_t row1  = vld1q_s16(coef_block + 1 * DCTSIZE);
+  int16x8_t row2  = vld1q_s16(coef_block + 2 * DCTSIZE);
+  int16x8_t row3  = vld1q_s16(coef_block + 3 * DCTSIZE);
+  int16x8_t row5  = vld1q_s16(coef_block + 5 * DCTSIZE);
+  int16x8_t row6  = vld1q_s16(coef_block + 6 * DCTSIZE);
+  int16x8_t row7  = vld1q_s16(coef_block + 7 * DCTSIZE);
+
+  /* Load quantization table values for DC coefficients. */
+  int16x8_t quant_row0 = vld1q_s16(quantptr + 0 * DCTSIZE);
+  /* Dequantize DC coefficients. */
+  row0 = vmulq_s16(row0, quant_row0);
+
+  /* Construct bitmap to test if all AC coefficients are 0. */
+  int16x8_t bitmap = vorrq_s16(row1, row2);
+  bitmap = vorrq_s16(bitmap, row3);
+  bitmap = vorrq_s16(bitmap, row5);
+  bitmap = vorrq_s16(bitmap, row6);
+  bitmap = vorrq_s16(bitmap, row7);
+
+  int64_t left_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 0);
+  int64_t right_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 1);
+
+  /* Load constants for IDCT computation. */
+#ifdef HAVE_VLD1_S16_X3
+  const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_4x4_neon_consts);
+#else
+  /* GCC does not currently support the intrinsic vld1_<type>_x3(). */
+  const int16x4_t consts1 = vld1_s16(jsimd_idct_4x4_neon_consts);
+  const int16x4_t consts2 = vld1_s16(jsimd_idct_4x4_neon_consts + 4);
+  const int16x4_t consts3 = vld1_s16(jsimd_idct_4x4_neon_consts + 8);
+  const int16x4x3_t consts = { { consts1, consts2, consts3 } };
+#endif
+
+  if (left_ac_bitmap == 0 && right_ac_bitmap == 0) {
+    /* All AC coefficients are zero.
+     * Compute DC values and duplicate into row vectors 0, 1, 2, and 3.
+     */
+    int16x8_t dcval = vshlq_n_s16(row0, PASS1_BITS);
+    row0 = dcval;
+    row1 = dcval;
+    row2 = dcval;
+    row3 = dcval;
+  } else if (left_ac_bitmap == 0) {
+    /* AC coefficients are zero for columns 0, 1, 2, and 3.
+     * Compute DC values for these columns.
+     */
+    int16x4_t dcval = vshl_n_s16(vget_low_s16(row0), PASS1_BITS);
+
+    /* Commence regular IDCT computation for columns 4, 5, 6, and 7. */
+
+    /* Load quantization table. */
+    int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE + 4);
+    int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE + 4);
+    int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE + 4);
+    int16x4_t quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE + 4);
+    int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE + 4);
+    int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE + 4);
+
+    /* Even part */
+    int32x4_t tmp0 = vshll_n_s16(vget_high_s16(row0), CONST_BITS + 1);
+
+    int16x4_t z2 = vmul_s16(vget_high_s16(row2), quant_row2);
+    int16x4_t z3 = vmul_s16(vget_high_s16(row6), quant_row6);
+
+    int32x4_t tmp2 = vmull_lane_s16(z2, consts.val[0], 0);
+    tmp2 = vmlal_lane_s16(tmp2, z3, consts.val[0], 1);
+
+    int32x4_t tmp10 = vaddq_s32(tmp0, tmp2);
+    int32x4_t tmp12 = vsubq_s32(tmp0, tmp2);
+
+    /* Odd part */
+    int16x4_t z1 = vmul_s16(vget_high_s16(row7), quant_row7);
+    z2 = vmul_s16(vget_high_s16(row5), quant_row5);
+    z3 = vmul_s16(vget_high_s16(row3), quant_row3);
+    int16x4_t z4 = vmul_s16(vget_high_s16(row1), quant_row1);
+
+    tmp0 = vmull_lane_s16(z1, consts.val[0], 2);
+    tmp0 = vmlal_lane_s16(tmp0, z2, consts.val[0], 3);
+    tmp0 = vmlal_lane_s16(tmp0, z3, consts.val[1], 0);
+    tmp0 = vmlal_lane_s16(tmp0, z4, consts.val[1], 1);
+
+    tmp2 = vmull_lane_s16(z1, consts.val[1], 2);
+    tmp2 = vmlal_lane_s16(tmp2, z2, consts.val[1], 3);
+    tmp2 = vmlal_lane_s16(tmp2, z3, consts.val[2], 0);
+    tmp2 = vmlal_lane_s16(tmp2, z4, consts.val[2], 1);
+
+    /* Final output stage: descale and narrow to 16-bit. */
+    row0 = vcombine_s16(dcval, vrshrn_n_s32(vaddq_s32(tmp10, tmp2),
+                                            CONST_BITS - PASS1_BITS + 1));
+    row3 = vcombine_s16(dcval, vrshrn_n_s32(vsubq_s32(tmp10, tmp2),
+                                            CONST_BITS - PASS1_BITS + 1));
+    row1 = vcombine_s16(dcval, vrshrn_n_s32(vaddq_s32(tmp12, tmp0),
+                                            CONST_BITS - PASS1_BITS + 1));
+    row2 = vcombine_s16(dcval, vrshrn_n_s32(vsubq_s32(tmp12, tmp0),
+                                            CONST_BITS - PASS1_BITS + 1));
+  } else if (right_ac_bitmap == 0) {
+    /* AC coefficients are zero for columns 4, 5, 6, and 7.
+     * Compute DC values for these columns.
+     */
+    int16x4_t dcval = vshl_n_s16(vget_high_s16(row0), PASS1_BITS);
+
+    /* Commence regular IDCT computation for columns 0, 1, 2, and 3. */
+
+    /* Load quantization table. */
+    int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE);
+    int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE);
+    int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE);
+    int16x4_t quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE);
+    int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE);
+    int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE);
+
+    /* Even part */
+    int32x4_t tmp0 = vshll_n_s16(vget_low_s16(row0), CONST_BITS + 1);
+
+    int16x4_t z2 = vmul_s16(vget_low_s16(row2), quant_row2);
+    int16x4_t z3 = vmul_s16(vget_low_s16(row6), quant_row6);
+
+    int32x4_t tmp2 = vmull_lane_s16(z2, consts.val[0], 0);
+    tmp2 = vmlal_lane_s16(tmp2, z3, consts.val[0], 1);
+
+    int32x4_t tmp10 = vaddq_s32(tmp0, tmp2);
+    int32x4_t tmp12 = vsubq_s32(tmp0, tmp2);
+
+    /* Odd part */
+    int16x4_t z1 = vmul_s16(vget_low_s16(row7), quant_row7);
+    z2 = vmul_s16(vget_low_s16(row5), quant_row5);
+    z3 = vmul_s16(vget_low_s16(row3), quant_row3);
+    int16x4_t z4 = vmul_s16(vget_low_s16(row1), quant_row1);
+
+    tmp0 = vmull_lane_s16(z1, consts.val[0], 2);
+    tmp0 = vmlal_lane_s16(tmp0, z2, consts.val[0], 3);
+    tmp0 = vmlal_lane_s16(tmp0, z3, consts.val[1], 0);
+    tmp0 = vmlal_lane_s16(tmp0, z4, consts.val[1], 1);
+
+    tmp2 = vmull_lane_s16(z1, consts.val[1], 2);
+    tmp2 = vmlal_lane_s16(tmp2, z2, consts.val[1], 3);
+    tmp2 = vmlal_lane_s16(tmp2, z3, consts.val[2], 0);
+    tmp2 = vmlal_lane_s16(tmp2, z4, consts.val[2], 1);
+
+    /* Final output stage: descale and narrow to 16-bit. */
+    row0 = vcombine_s16(vrshrn_n_s32(vaddq_s32(tmp10, tmp2),
+                                     CONST_BITS - PASS1_BITS + 1), dcval);
+    row3 = vcombine_s16(vrshrn_n_s32(vsubq_s32(tmp10, tmp2),
+                                     CONST_BITS - PASS1_BITS + 1), dcval);
+    row1 = vcombine_s16(vrshrn_n_s32(vaddq_s32(tmp12, tmp0),
+                                     CONST_BITS - PASS1_BITS + 1), dcval);
+    row2 = vcombine_s16(vrshrn_n_s32(vsubq_s32(tmp12, tmp0),
+                                     CONST_BITS - PASS1_BITS + 1), dcval);
+  } else {
+    /* All AC coefficients are non-zero; full IDCT calculation required. */
+    int16x8_t quant_row1 = vld1q_s16(quantptr + 1 * DCTSIZE);
+    int16x8_t quant_row2 = vld1q_s16(quantptr + 2 * DCTSIZE);
+    int16x8_t quant_row3 = vld1q_s16(quantptr + 3 * DCTSIZE);
+    int16x8_t quant_row5 = vld1q_s16(quantptr + 5 * DCTSIZE);
+    int16x8_t quant_row6 = vld1q_s16(quantptr + 6 * DCTSIZE);
+    int16x8_t quant_row7 = vld1q_s16(quantptr + 7 * DCTSIZE);
+
+    /* Even part */
+    int32x4_t tmp0_l = vshll_n_s16(vget_low_s16(row0), CONST_BITS + 1);
+    int32x4_t tmp0_h = vshll_n_s16(vget_high_s16(row0), CONST_BITS + 1);
+
+    int16x8_t z2 = vmulq_s16(row2, quant_row2);
+    int16x8_t z3 = vmulq_s16(row6, quant_row6);
+
+    int32x4_t tmp2_l = vmull_lane_s16(vget_low_s16(z2), consts.val[0], 0);
+    int32x4_t tmp2_h = vmull_lane_s16(vget_high_s16(z2), consts.val[0], 0);
+    tmp2_l = vmlal_lane_s16(tmp2_l, vget_low_s16(z3), consts.val[0], 1);
+    tmp2_h = vmlal_lane_s16(tmp2_h, vget_high_s16(z3), consts.val[0], 1);
+
+    int32x4_t tmp10_l = vaddq_s32(tmp0_l, tmp2_l);
+    int32x4_t tmp10_h = vaddq_s32(tmp0_h, tmp2_h);
+    int32x4_t tmp12_l = vsubq_s32(tmp0_l, tmp2_l);
+    int32x4_t tmp12_h = vsubq_s32(tmp0_h, tmp2_h);
+
+    /* Odd part */
+    int16x8_t z1 = vmulq_s16(row7, quant_row7);
+    z2 = vmulq_s16(row5, quant_row5);
+    z3 = vmulq_s16(row3, quant_row3);
+    int16x8_t z4 = vmulq_s16(row1, quant_row1);
+
+    tmp0_l = vmull_lane_s16(vget_low_s16(z1), consts.val[0], 2);
+    tmp0_l = vmlal_lane_s16(tmp0_l, vget_low_s16(z2), consts.val[0], 3);
+    tmp0_l = vmlal_lane_s16(tmp0_l, vget_low_s16(z3), consts.val[1], 0);
+    tmp0_l = vmlal_lane_s16(tmp0_l, vget_low_s16(z4), consts.val[1], 1);
+    tmp0_h = vmull_lane_s16(vget_high_s16(z1), consts.val[0], 2);
+    tmp0_h = vmlal_lane_s16(tmp0_h, vget_high_s16(z2), consts.val[0], 3);
+    tmp0_h = vmlal_lane_s16(tmp0_h, vget_high_s16(z3), consts.val[1], 0);
+    tmp0_h = vmlal_lane_s16(tmp0_h, vget_high_s16(z4), consts.val[1], 1);
+
+    tmp2_l = vmull_lane_s16(vget_low_s16(z1), consts.val[1], 2);
+    tmp2_l = vmlal_lane_s16(tmp2_l, vget_low_s16(z2), consts.val[1], 3);
+    tmp2_l = vmlal_lane_s16(tmp2_l, vget_low_s16(z3), consts.val[2], 0);
+    tmp2_l = vmlal_lane_s16(tmp2_l, vget_low_s16(z4), consts.val[2], 1);
+    tmp2_h = vmull_lane_s16(vget_high_s16(z1), consts.val[1], 2);
+    tmp2_h = vmlal_lane_s16(tmp2_h, vget_high_s16(z2), consts.val[1], 3);
+    tmp2_h = vmlal_lane_s16(tmp2_h, vget_high_s16(z3), consts.val[2], 0);
+    tmp2_h = vmlal_lane_s16(tmp2_h, vget_high_s16(z4), consts.val[2], 1);
+
+    /* Final output stage: descale and narrow to 16-bit. */
+    row0 = vcombine_s16(vrshrn_n_s32(vaddq_s32(tmp10_l, tmp2_l),
+                                     CONST_BITS - PASS1_BITS + 1),
+                        vrshrn_n_s32(vaddq_s32(tmp10_h, tmp2_h),
+                                     CONST_BITS - PASS1_BITS + 1));
+    row3 = vcombine_s16(vrshrn_n_s32(vsubq_s32(tmp10_l, tmp2_l),
+                                     CONST_BITS - PASS1_BITS + 1),
+                        vrshrn_n_s32(vsubq_s32(tmp10_h, tmp2_h),
+                                     CONST_BITS - PASS1_BITS + 1));
+    row1 = vcombine_s16(vrshrn_n_s32(vaddq_s32(tmp12_l, tmp0_l),
+                                     CONST_BITS - PASS1_BITS + 1),
+                        vrshrn_n_s32(vaddq_s32(tmp12_h, tmp0_h),
+                                     CONST_BITS - PASS1_BITS + 1));
+    row2 = vcombine_s16(vrshrn_n_s32(vsubq_s32(tmp12_l, tmp0_l),
+                                     CONST_BITS - PASS1_BITS + 1),
+                        vrshrn_n_s32(vsubq_s32(tmp12_h, tmp0_h),
+                                     CONST_BITS - PASS1_BITS + 1));
+  }
+
+  /* Transpose 8x4 block to perform IDCT on rows in second pass. */
+  int16x8x2_t row_01 = vtrnq_s16(row0, row1);
+  int16x8x2_t row_23 = vtrnq_s16(row2, row3);
+
+  int32x4x2_t cols_0426 = vtrnq_s32(vreinterpretq_s32_s16(row_01.val[0]),
+                                    vreinterpretq_s32_s16(row_23.val[0]));
+  int32x4x2_t cols_1537 = vtrnq_s32(vreinterpretq_s32_s16(row_01.val[1]),
+                                    vreinterpretq_s32_s16(row_23.val[1]));
+
+  int16x4_t col0 = vreinterpret_s16_s32(vget_low_s32(cols_0426.val[0]));
+  int16x4_t col1 = vreinterpret_s16_s32(vget_low_s32(cols_1537.val[0]));
+  int16x4_t col2 = vreinterpret_s16_s32(vget_low_s32(cols_0426.val[1]));
+  int16x4_t col3 = vreinterpret_s16_s32(vget_low_s32(cols_1537.val[1]));
+  int16x4_t col5 = vreinterpret_s16_s32(vget_high_s32(cols_1537.val[0]));
+  int16x4_t col6 = vreinterpret_s16_s32(vget_high_s32(cols_0426.val[1]));
+  int16x4_t col7 = vreinterpret_s16_s32(vget_high_s32(cols_1537.val[1]));
+
+  /* Commence second pass of IDCT. */
+
+  /* Even part */
+  int32x4_t tmp0 = vshll_n_s16(col0, CONST_BITS + 1);
+  int32x4_t tmp2 = vmull_lane_s16(col2, consts.val[0], 0);
+  tmp2 = vmlal_lane_s16(tmp2, col6, consts.val[0], 1);
+
+  int32x4_t tmp10 = vaddq_s32(tmp0, tmp2);
+  int32x4_t tmp12 = vsubq_s32(tmp0, tmp2);
+
+  /* Odd part */
+  tmp0 = vmull_lane_s16(col7, consts.val[0], 2);
+  tmp0 = vmlal_lane_s16(tmp0, col5, consts.val[0], 3);
+  tmp0 = vmlal_lane_s16(tmp0, col3, consts.val[1], 0);
+  tmp0 = vmlal_lane_s16(tmp0, col1, consts.val[1], 1);
+
+  tmp2 = vmull_lane_s16(col7, consts.val[1], 2);
+  tmp2 = vmlal_lane_s16(tmp2, col5, consts.val[1], 3);
+  tmp2 = vmlal_lane_s16(tmp2, col3, consts.val[2], 0);
+  tmp2 = vmlal_lane_s16(tmp2, col1, consts.val[2], 1);
+
+  /* Final output stage: descale and clamp to range [0-255]. */
+  int16x8_t output_cols_02 = vcombine_s16(vaddhn_s32(tmp10, tmp2),
+                                          vsubhn_s32(tmp12, tmp0));
+  int16x8_t output_cols_13 = vcombine_s16(vaddhn_s32(tmp12, tmp0),
+                                          vsubhn_s32(tmp10, tmp2));
+  output_cols_02 = vrsraq_n_s16(vdupq_n_s16(CENTERJSAMPLE), output_cols_02,
+                                CONST_BITS + PASS1_BITS + 3 + 1 - 16);
+  output_cols_13 = vrsraq_n_s16(vdupq_n_s16(CENTERJSAMPLE), output_cols_13,
+                                CONST_BITS + PASS1_BITS + 3 + 1 - 16);
+  /* Narrow to 8-bit and convert to unsigned while zipping 8-bit elements.
+   * An interleaving store completes the transpose.
+   */
+  uint8x8x2_t output_0123 = vzip_u8(vqmovun_s16(output_cols_02),
+                                    vqmovun_s16(output_cols_13));
+  uint16x4x2_t output_01_23 = { {
+    vreinterpret_u16_u8(output_0123.val[0]),
+    vreinterpret_u16_u8(output_0123.val[1])
+  } };
+
+  /* Store 4x4 block to memory. */
+  JSAMPROW outptr0 = output_buf[0] + output_col;
+  JSAMPROW outptr1 = output_buf[1] + output_col;
+  JSAMPROW outptr2 = output_buf[2] + output_col;
+  JSAMPROW outptr3 = output_buf[3] + output_col;
+  vst2_lane_u16((uint16_t *)outptr0, output_01_23, 0);
+  vst2_lane_u16((uint16_t *)outptr1, output_01_23, 1);
+  vst2_lane_u16((uint16_t *)outptr2, output_01_23, 2);
+  vst2_lane_u16((uint16_t *)outptr3, output_01_23, 3);
+}
@@ -0,0 +1,193 @@
+/*
+ * jquanti-neon.c - sample data conversion and quantization (Arm Neon)
+ *
+ * Copyright (C) 2020-2021, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+
+#include <arm_neon.h>
+
+
+/* After downsampling, the resulting sample values are in the range [0, 255],
+ * but the Discrete Cosine Transform (DCT) operates on values centered around
+ * 0.
+ *
+ * To prepare sample values for the DCT, load samples into a DCT workspace,
+ * subtracting CENTERJSAMPLE (128).  The samples, now in the range [-128, 127],
+ * are also widened from 8- to 16-bit.
+ *
+ * The equivalent scalar C function convsamp() can be found in jcdctmgr.c.
+ */
+
+void jsimd_convsamp_neon(JSAMPARRAY sample_data, JDIMENSION start_col,
+                         DCTELEM *workspace)
+{
+  uint8x8_t samp_row0 = vld1_u8(sample_data[0] + start_col);
+  uint8x8_t samp_row1 = vld1_u8(sample_data[1] + start_col);
+  uint8x8_t samp_row2 = vld1_u8(sample_data[2] + start_col);
+  uint8x8_t samp_row3 = vld1_u8(sample_data[3] + start_col);
+  uint8x8_t samp_row4 = vld1_u8(sample_data[4] + start_col);
+  uint8x8_t samp_row5 = vld1_u8(sample_data[5] + start_col);
+  uint8x8_t samp_row6 = vld1_u8(sample_data[6] + start_col);
+  uint8x8_t samp_row7 = vld1_u8(sample_data[7] + start_col);
+
+  int16x8_t row0 =
+    vreinterpretq_s16_u16(vsubl_u8(samp_row0, vdup_n_u8(CENTERJSAMPLE)));
+  int16x8_t row1 =
+    vreinterpretq_s16_u16(vsubl_u8(samp_row1, vdup_n_u8(CENTERJSAMPLE)));
+  int16x8_t row2 =
+    vreinterpretq_s16_u16(vsubl_u8(samp_row2, vdup_n_u8(CENTERJSAMPLE)));
+  int16x8_t row3 =
+    vreinterpretq_s16_u16(vsubl_u8(samp_row3, vdup_n_u8(CENTERJSAMPLE)));
+  int16x8_t row4 =
+    vreinterpretq_s16_u16(vsubl_u8(samp_row4, vdup_n_u8(CENTERJSAMPLE)));
+  int16x8_t row5 =
+    vreinterpretq_s16_u16(vsubl_u8(samp_row5, vdup_n_u8(CENTERJSAMPLE)));
+  int16x8_t row6 =
+    vreinterpretq_s16_u16(vsubl_u8(samp_row6, vdup_n_u8(CENTERJSAMPLE)));
+  int16x8_t row7 =
+    vreinterpretq_s16_u16(vsubl_u8(samp_row7, vdup_n_u8(CENTERJSAMPLE)));
+
+  vst1q_s16(workspace + 0 * DCTSIZE, row0);
+  vst1q_s16(workspace + 1 * DCTSIZE, row1);
+  vst1q_s16(workspace + 2 * DCTSIZE, row2);
+  vst1q_s16(workspace + 3 * DCTSIZE, row3);
+  vst1q_s16(workspace + 4 * DCTSIZE, row4);
+  vst1q_s16(workspace + 5 * DCTSIZE, row5);
+  vst1q_s16(workspace + 6 * DCTSIZE, row6);
+  vst1q_s16(workspace + 7 * DCTSIZE, row7);
+}
+
+
+/* After the DCT, the resulting array of coefficient values needs to be divided
+ * by an array of quantization values.
+ *
+ * To avoid a slow division operation, the DCT coefficients are multiplied by
+ * the (scaled) reciprocals of the quantization values and then right-shifted.
+ *
+ * The equivalent scalar C function quantize() can be found in jcdctmgr.c.
+ */
+
+void jsimd_quantize_neon(JCOEFPTR coef_block, DCTELEM *divisors,
+                         DCTELEM *workspace)
+{
+  JCOEFPTR out_ptr = coef_block;
+  UDCTELEM *recip_ptr = (UDCTELEM *)divisors;
+  UDCTELEM *corr_ptr = (UDCTELEM *)divisors + DCTSIZE2;
+  DCTELEM *shift_ptr = divisors + 3 * DCTSIZE2;
+  int i;
+
+#if defined(__clang__) && (defined(__aarch64__) || defined(_M_ARM64))
+#pragma unroll
+#endif
+  for (i = 0; i < DCTSIZE; i += DCTSIZE / 2) {
+    /* Load DCT coefficients. */
+    int16x8_t row0 = vld1q_s16(workspace + (i + 0) * DCTSIZE);
+    int16x8_t row1 = vld1q_s16(workspace + (i + 1) * DCTSIZE);
+    int16x8_t row2 = vld1q_s16(workspace + (i + 2) * DCTSIZE);
+    int16x8_t row3 = vld1q_s16(workspace + (i + 3) * DCTSIZE);
+    /* Load reciprocals of quantization values. */
+    uint16x8_t recip0 = vld1q_u16(recip_ptr + (i + 0) * DCTSIZE);
+    uint16x8_t recip1 = vld1q_u16(recip_ptr + (i + 1) * DCTSIZE);
+    uint16x8_t recip2 = vld1q_u16(recip_ptr + (i + 2) * DCTSIZE);
+    uint16x8_t recip3 = vld1q_u16(recip_ptr + (i + 3) * DCTSIZE);
+    uint16x8_t corr0 = vld1q_u16(corr_ptr + (i + 0) * DCTSIZE);
+    uint16x8_t corr1 = vld1q_u16(corr_ptr + (i + 1) * DCTSIZE);
+    uint16x8_t corr2 = vld1q_u16(corr_ptr + (i + 2) * DCTSIZE);
+    uint16x8_t corr3 = vld1q_u16(corr_ptr + (i + 3) * DCTSIZE);
+    int16x8_t shift0 = vld1q_s16(shift_ptr + (i + 0) * DCTSIZE);
+    int16x8_t shift1 = vld1q_s16(shift_ptr + (i + 1) * DCTSIZE);
+    int16x8_t shift2 = vld1q_s16(shift_ptr + (i + 2) * DCTSIZE);
+    int16x8_t shift3 = vld1q_s16(shift_ptr + (i + 3) * DCTSIZE);
+
+    /* Extract sign from coefficients. */
+    int16x8_t sign_row0 = vshrq_n_s16(row0, 15);
+    int16x8_t sign_row1 = vshrq_n_s16(row1, 15);
+    int16x8_t sign_row2 = vshrq_n_s16(row2, 15);
+    int16x8_t sign_row3 = vshrq_n_s16(row3, 15);
+    /* Get absolute value of DCT coefficients. */
+    uint16x8_t abs_row0 = vreinterpretq_u16_s16(vabsq_s16(row0));
+    uint16x8_t abs_row1 = vreinterpretq_u16_s16(vabsq_s16(row1));
+    uint16x8_t abs_row2 = vreinterpretq_u16_s16(vabsq_s16(row2));
+    uint16x8_t abs_row3 = vreinterpretq_u16_s16(vabsq_s16(row3));
+    /* Add correction. */
+    abs_row0 = vaddq_u16(abs_row0, corr0);
+    abs_row1 = vaddq_u16(abs_row1, corr1);
+    abs_row2 = vaddq_u16(abs_row2, corr2);
+    abs_row3 = vaddq_u16(abs_row3, corr3);
+
+    /* Multiply DCT coefficients by quantization reciprocals. */
+    int32x4_t row0_l = vreinterpretq_s32_u32(vmull_u16(vget_low_u16(abs_row0),
+                                                       vget_low_u16(recip0)));
+    int32x4_t row0_h = vreinterpretq_s32_u32(vmull_u16(vget_high_u16(abs_row0),
+                                                       vget_high_u16(recip0)));
+    int32x4_t row1_l = vreinterpretq_s32_u32(vmull_u16(vget_low_u16(abs_row1),
+                                                       vget_low_u16(recip1)));
+    int32x4_t row1_h = vreinterpretq_s32_u32(vmull_u16(vget_high_u16(abs_row1),
+                                                       vget_high_u16(recip1)));
+    int32x4_t row2_l = vreinterpretq_s32_u32(vmull_u16(vget_low_u16(abs_row2),
+                                                       vget_low_u16(recip2)));
+    int32x4_t row2_h = vreinterpretq_s32_u32(vmull_u16(vget_high_u16(abs_row2),
+                                                       vget_high_u16(recip2)));
+    int32x4_t row3_l = vreinterpretq_s32_u32(vmull_u16(vget_low_u16(abs_row3),
+                                                       vget_low_u16(recip3)));
+    int32x4_t row3_h = vreinterpretq_s32_u32(vmull_u16(vget_high_u16(abs_row3),
+                                                       vget_high_u16(recip3)));
+    /* Narrow back to 16-bit. */
+    row0 = vcombine_s16(vshrn_n_s32(row0_l, 16), vshrn_n_s32(row0_h, 16));
+    row1 = vcombine_s16(vshrn_n_s32(row1_l, 16), vshrn_n_s32(row1_h, 16));
+    row2 = vcombine_s16(vshrn_n_s32(row2_l, 16), vshrn_n_s32(row2_h, 16));
+    row3 = vcombine_s16(vshrn_n_s32(row3_l, 16), vshrn_n_s32(row3_h, 16));
+
+    /* Since VSHR only supports an immediate as its second argument, negate the
+     * shift value and shift left.
+     */
+    row0 = vreinterpretq_s16_u16(vshlq_u16(vreinterpretq_u16_s16(row0),
+                                           vnegq_s16(shift0)));
+    row1 = vreinterpretq_s16_u16(vshlq_u16(vreinterpretq_u16_s16(row1),
+                                           vnegq_s16(shift1)));
+    row2 = vreinterpretq_s16_u16(vshlq_u16(vreinterpretq_u16_s16(row2),
+                                           vnegq_s16(shift2)));
+    row3 = vreinterpretq_s16_u16(vshlq_u16(vreinterpretq_u16_s16(row3),
+                                           vnegq_s16(shift3)));
+
+    /* Restore sign to original product. */
+    row0 = veorq_s16(row0, sign_row0);
+    row0 = vsubq_s16(row0, sign_row0);
+    row1 = veorq_s16(row1, sign_row1);
+    row1 = vsubq_s16(row1, sign_row1);
+    row2 = veorq_s16(row2, sign_row2);
+    row2 = vsubq_s16(row2, sign_row2);
+    row3 = veorq_s16(row3, sign_row3);
+    row3 = vsubq_s16(row3, sign_row3);
+
+    /* Store quantized coefficients to memory. */
+    vst1q_s16(out_ptr + (i + 0) * DCTSIZE, row0);
+    vst1q_s16(out_ptr + (i + 1) * DCTSIZE, row1);
+    vst1q_s16(out_ptr + (i + 2) * DCTSIZE, row2);
+    vst1q_s16(out_ptr + (i + 3) * DCTSIZE, row3);
+  }
+}
@@ -0,0 +1,37 @@
+/*
+ * Copyright (C) 2020, D. R. Commander.  All Rights Reserved.
+ * Copyright (C) 2020-2021, Arm Limited.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#cmakedefine HAVE_VLD1_S16_X3
+#cmakedefine HAVE_VLD1_U16_X2
+#cmakedefine HAVE_VLD1Q_U8_X4
+
+/* Define compiler-independent count-leading-zeros and byte-swap macros */
+#if defined(_MSC_VER) && !defined(__clang__)
+#define BUILTIN_CLZ(x)  _CountLeadingZeros(x)
+#define BUILTIN_CLZLL(x)  _CountLeadingZeros64(x)
+#define BUILTIN_BSWAP64(x)  _byteswap_uint64(x)
+#elif defined(__clang__) || defined(__GNUC__)
+#define BUILTIN_CLZ(x)  __builtin_clz(x)
+#define BUILTIN_CLZLL(x)  __builtin_clzll(x)
+#define BUILTIN_BSWAP64(x)  __builtin_bswap64(x)
+#else
+#error "Unknown compiler"
+#endif
@@ -0,0 +1,578 @@
+;
+; jccolext.asm - colorspace conversion (AVX2)
+;
+; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_ycc_convert_avx2(JDIMENSION img_width, JSAMPARRAY input_buf,
+;                            JSAMPIMAGE output_buf, JDIMENSION output_row,
+;                            int num_rows);
+;
+
+%define img_width(b)   (b) + 8          ; JDIMENSION img_width
+%define input_buf(b)   (b) + 12         ; JSAMPARRAY input_buf
+%define output_buf(b)  (b) + 16         ; JSAMPIMAGE output_buf
+%define output_row(b)  (b) + 20         ; JDIMENSION output_row
+%define num_rows(b)    (b) + 24         ; int num_rows
+
+%define original_ebp   ebp + 0
+%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_YMMWORD
+                                        ; ymmword wk[WK_NUM]
+%define WK_NUM         8
+%define gotptr         wk(0) - SIZEOF_POINTER  ; void * gotptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_rgb_ycc_convert_avx2)
+
+EXTN(jsimd_rgb_ycc_convert_avx2):
+    push        ebp
+    mov         eax, esp                     ; eax = original ebp
+    sub         esp, byte 4
+    and         esp, byte (-SIZEOF_YMMWORD)  ; align to 256 bits
+    mov         [esp], eax
+    mov         ebp, esp                     ; ebp = aligned ebp
+    lea         esp, [wk(0)]
+    pushpic     eax                     ; make a room for GOT address
+    push        ebx
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    get_GOT     ebx                     ; get GOT address
+    movpic      POINTER [gotptr], ebx   ; save GOT address
+
+    mov         ecx, JDIMENSION [img_width(eax)]
+    test        ecx, ecx
+    jz          near .return
+
+    push        ecx
+
+    mov         esi, JSAMPIMAGE [output_buf(eax)]
+    mov         ecx, JDIMENSION [output_row(eax)]
+    mov         edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
+    mov         ebx, JSAMPARRAY [esi+1*SIZEOF_JSAMPARRAY]
+    mov         edx, JSAMPARRAY [esi+2*SIZEOF_JSAMPARRAY]
+    lea         edi, [edi+ecx*SIZEOF_JSAMPROW]
+    lea         ebx, [ebx+ecx*SIZEOF_JSAMPROW]
+    lea         edx, [edx+ecx*SIZEOF_JSAMPROW]
+
+    pop         ecx
+
+    mov         esi, JSAMPARRAY [input_buf(eax)]
+    mov         eax, INT [num_rows(eax)]
+    test        eax, eax
+    jle         near .return
+    alignx      16, 7
+.rowloop:
+    pushpic     eax
+    push        edx
+    push        ebx
+    push        edi
+    push        esi
+    push        ecx                     ; col
+
+    mov         esi, JSAMPROW [esi]     ; inptr
+    mov         edi, JSAMPROW [edi]     ; outptr0
+    mov         ebx, JSAMPROW [ebx]     ; outptr1
+    mov         edx, JSAMPROW [edx]     ; outptr2
+    movpic      eax, POINTER [gotptr]   ; load GOT address (eax)
+
+    cmp         ecx, byte SIZEOF_YMMWORD
+    jae         near .columnloop
+    alignx      16, 7
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+.column_ld1:
+    push        eax
+    push        edx
+    lea         ecx, [ecx+ecx*2]        ; imul ecx,RGB_PIXELSIZE
+    test        cl, SIZEOF_BYTE
+    jz          short .column_ld2
+    sub         ecx, byte SIZEOF_BYTE
+    movzx       eax, byte [esi+ecx]
+.column_ld2:
+    test        cl, SIZEOF_WORD
+    jz          short .column_ld4
+    sub         ecx, byte SIZEOF_WORD
+    movzx       edx, word [esi+ecx]
+    shl         eax, WORD_BIT
+    or          eax, edx
+.column_ld4:
+    vmovd       xmmA, eax
+    pop         edx
+    pop         eax
+    test        cl, SIZEOF_DWORD
+    jz          short .column_ld8
+    sub         ecx, byte SIZEOF_DWORD
+    vmovd       xmmF, XMM_DWORD [esi+ecx]
+    vpslldq     xmmA, xmmA, SIZEOF_DWORD
+    vpor        xmmA, xmmA, xmmF
+.column_ld8:
+    test        cl, SIZEOF_MMWORD
+    jz          short .column_ld16
+    sub         ecx, byte SIZEOF_MMWORD
+    vmovq       xmmB, XMM_MMWORD [esi+ecx]
+    vpslldq     xmmA, xmmA, SIZEOF_MMWORD
+    vpor        xmmA, xmmA, xmmB
+.column_ld16:
+    test        cl, SIZEOF_XMMWORD
+    jz          short .column_ld32
+    sub         ecx, byte SIZEOF_XMMWORD
+    vmovdqu     xmmB, XMM_MMWORD [esi+ecx]
+    vperm2i128  ymmA, ymmA, ymmA, 1
+    vpor        ymmA, ymmB
+.column_ld32:
+    test        cl, SIZEOF_YMMWORD
+    jz          short .column_ld64
+    sub         ecx, byte SIZEOF_YMMWORD
+    vmovdqa     ymmF, ymmA
+    vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
+.column_ld64:
+    test        cl, 2*SIZEOF_YMMWORD
+    mov         ecx, SIZEOF_YMMWORD
+    jz          short .rgb_ycc_cnv
+    vmovdqa     ymmB, ymmA
+    vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD]
+    jmp         short .rgb_ycc_cnv
+    alignx      16, 7
+
+.columnloop:
+    vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD]
+    vmovdqu     ymmB, YMMWORD [esi+2*SIZEOF_YMMWORD]
+
+.rgb_ycc_cnv:
+    ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+    ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+    ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    ; ymmB=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
+    ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+
+    vmovdqu     ymmC, ymmA
+    vinserti128 ymmA, ymmF, xmmA, 0  ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+                                     ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    vinserti128 ymmC, ymmC, xmmB, 0  ; ymmC=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
+                                     ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    vinserti128 ymmB, ymmB, xmmF, 0  ; ymmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+                                     ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+    vperm2i128  ymmF, ymmC, ymmC, 1  ; ymmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A
+                                     ;       1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q)
+
+    vmovdqa     ymmG, ymmA
+    vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12
+                                  ;       22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I)
+    vpsrldq     ymmG, ymmG, 8     ; ymmG=(22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I
+                                  ;       2I 0J 1J 2J 0K 1K 2K 0L -- -- -- -- -- -- -- --)
+
+    vpunpckhbw  ymmA, ymmA, ymmF  ; ymmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A
+                                  ;       0G 0O 1G 1O 2G 2O 0H 0P 1H 1P 2H 2P 0I 0Q 1I 1Q)
+    vpslldq     ymmF, ymmF, 8     ; ymmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27
+                                  ;       08 18 28 09 19 29 0A 1A 1L 2L 0M 1M 2M 0N 1N 2N)
+
+    vpunpcklbw  ymmG, ymmG, ymmB  ; ymmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D
+                                  ;       2I 2Q 0J 0R 1J 1R 2J 2R 0K 0S 1K 1S 2K 2S 0L 0T)
+    vpunpckhbw  ymmF, ymmF, ymmB  ; ymmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F
+                                  ;       1L 1T 2L 2T 0M 0U 1M 1U 2M 2U 0N 0V 1N 1V 2N 2V)
+
+    vmovdqa     ymmD, ymmA
+    vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09
+                                  ;       11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P)
+    vpsrldq     ymmD, ymmD, 8     ; ymmD=(11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P
+                                  ;       1H 1P 2H 2P 0I 0Q 1I 1Q -- -- -- -- -- -- -- --)
+
+    vpunpckhbw  ymmA, ymmA, ymmG  ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D
+                                  ;       0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 0H 0L 0P 0T)
+    vpslldq     ymmG, ymmG, 8     ; ymmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B
+                                  ;       04 0C 14 1C 24 2C 05 0D 2I 2Q 0J 0R 1J 1R 2J 2R)
+
+    vpunpcklbw  ymmD, ymmD, ymmF  ; ymmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E
+                                  ;       1H 1L 1P 1T 2H 2L 2P 2T 0I 0M 0Q 0U 1I 1M 1Q 1U)
+    vpunpckhbw  ymmG, ymmG, ymmF  ; ymmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F
+                                  ;       2I 2M 2Q 2U 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V)
+
+    vmovdqa     ymmE, ymmA
+    vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C
+                                  ;       20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S)
+    vpsrldq     ymmE, ymmE, 8     ; ymmE=(20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S
+                                  ;       2G 2K 2O 2S 0H 0L 0P 0T -- -- -- -- -- -- -- --)
+
+    vpunpckhbw  ymmA, ymmA, ymmD  ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E
+                                  ;       0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U)
+    vpslldq     ymmD, ymmD, 8     ; ymmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D
+                                  ;       02 06 0A 0E 12 16 1A 1E 1H 1L 1P 1T 2H 2L 2P 2T)
+
+    vpunpcklbw  ymmE, ymmE, ymmG  ; ymmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F
+                                  ;       2G 2I 2K 2M 2O 2Q 2S 2U 0H 0J 0L 0N 0P 0R 0T 0V)
+    vpunpckhbw  ymmD, ymmD, ymmG  ; ymmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F
+                                  ;       1H 1J 1L 1N 1P 1R 1T 1V 2H 2J 2L 2N 2P 2R 2T 2V)
+
+    vpxor       ymmH, ymmH, ymmH
+
+    vmovdqa     ymmC, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmH  ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U)
+    vpunpckhbw  ymmC, ymmC, ymmH  ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U)
+
+    vmovdqa     ymmB, ymmE
+    vpunpcklbw  ymmE, ymmE, ymmH  ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U)
+    vpunpckhbw  ymmB, ymmB, ymmH  ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V)
+
+    vmovdqa     ymmF, ymmD
+    vpunpcklbw  ymmD, ymmD, ymmH  ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V)
+    vpunpckhbw  ymmF, ymmF, ymmH  ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V)
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+    test        cl, SIZEOF_XMMWORD/16
+    jz          short .column_ld2
+    sub         ecx, byte SIZEOF_XMMWORD/16
+    vmovd       xmmA, XMM_DWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld2:
+    test        cl, SIZEOF_XMMWORD/8
+    jz          short .column_ld4
+    sub         ecx, byte SIZEOF_XMMWORD/8
+    vmovq       xmmF, XMM_MMWORD [esi+ecx*RGB_PIXELSIZE]
+    vpslldq     xmmA, xmmA, SIZEOF_MMWORD
+    vpor        xmmA, xmmA, xmmF
+.column_ld4:
+    test        cl, SIZEOF_XMMWORD/4
+    jz          short .column_ld8
+    sub         ecx, byte SIZEOF_XMMWORD/4
+    vmovdqa     xmmF, xmmA
+    vperm2i128  ymmF, ymmF, ymmF, 1
+    vmovdqu     xmmA, XMMWORD [esi+ecx*RGB_PIXELSIZE]
+    vpor        ymmA, ymmA, ymmF
+.column_ld8:
+    test        cl, SIZEOF_XMMWORD/2
+    jz          short .column_ld16
+    sub         ecx, byte SIZEOF_XMMWORD/2
+    vmovdqa     ymmF, ymmA
+    vmovdqu     ymmA, YMMWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld16:
+    test        cl, SIZEOF_XMMWORD
+    mov         ecx, SIZEOF_YMMWORD
+    jz          short .rgb_ycc_cnv
+    vmovdqa     ymmE, ymmA
+    vmovdqa     ymmH, ymmF
+    vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD]
+    jmp         short .rgb_ycc_cnv
+    alignx      16, 7
+
+.columnloop:
+    vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD]
+    vmovdqu     ymmE, YMMWORD [esi+2*SIZEOF_YMMWORD]
+    vmovdqu     ymmH, YMMWORD [esi+3*SIZEOF_YMMWORD]
+
+.rgb_ycc_cnv:
+    ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+    ;       04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+    ;       0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+    ; ymmE=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J
+    ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+    ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R
+    ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+
+    vmovdqa     ymmB, ymmA
+    vinserti128 ymmA, ymmA, xmmE, 1     ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+                                        ;       0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J)
+    vperm2i128  ymmE, ymmB, ymmE, 0x31  ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+                                        ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+
+    vmovdqa     ymmB, ymmF
+    vinserti128 ymmF, ymmF, xmmH, 1     ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+                                        ;       0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R)
+    vperm2i128  ymmH, ymmB, ymmH, 0x31  ; ymmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F
+                                        ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+
+    vmovdqa     ymmD, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmE      ; ymmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35
+                                      ;       0G 0K 1G 1K 2G 2K 3G 3K 0H 0L 1H 1L 2H 2L 3H 3L)
+    vpunpckhbw  ymmD, ymmD, ymmE      ; ymmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37
+                                      ;       0I 0M 1I 1M 2I 2M 3I 3M 0J 0N 1J 1N 2J 2N 3J 3N)
+
+    vmovdqa     ymmC, ymmF
+    vpunpcklbw  ymmF, ymmF, ymmH      ; ymmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D
+                                      ;       0O 0S 1O 1S 2O 2S 3O 3S 0P 0T 1P 1T 2P 2T 3P 3T)
+    vpunpckhbw  ymmC, ymmC, ymmH      ; ymmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F
+                                      ;       0Q 0U 1Q 1U 2Q 2U 3Q 3U 0R 0V 1R 1V 2R 2V 3R 3V)
+
+    vmovdqa     ymmB, ymmA
+    vpunpcklwd  ymmA, ymmA, ymmF      ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C
+                                      ;       0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 3G 3K 3O 3S)
+    vpunpckhwd  ymmB, ymmB, ymmF      ; ymmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D
+                                      ;       0H 0L 0P 0T 1H 1L 1P 1T 2H 2L 2P 2T 3H 3L 3P 3T)
+
+    vmovdqa     ymmG, ymmD
+    vpunpcklwd  ymmD, ymmD, ymmC      ; ymmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E
+                                      ;       0I 0M 0Q 0U 1I 1M 1Q 1U 2I 2M 2Q 2U 3I 3M 3Q 3U)
+    vpunpckhwd  ymmG, ymmG, ymmC      ; ymmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F
+                                      ;       0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V 3J 3N 3R 3V)
+
+    vmovdqa     ymmE, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmD      ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E
+                                      ;       0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U)
+    vpunpckhbw  ymmE, ymmE, ymmD      ; ymmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E
+                                      ;       2G 2I 2K 2M 2O 2Q 2S 2U 3G 3I 3K 3M 3O 3Q 3S 3U)
+
+    vmovdqa     ymmH, ymmB
+    vpunpcklbw  ymmB, ymmB, ymmG      ; ymmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F
+                                      ;       0H 0J 0L 0N 0P 0R 0T 0V 1H 1J 1L 1N 1P 1R 1T 1V)
+    vpunpckhbw  ymmH, ymmH, ymmG      ; ymmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F
+                                      ;       2H 2J 2L 2N 2P 2R 2T 2V 3H 3J 3L 3N 3P 3R 3T 3V)
+
+    vpxor       ymmF, ymmF, ymmF
+
+    vmovdqa     ymmC, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmF      ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U)
+    vpunpckhbw  ymmC, ymmC, ymmF      ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U)
+
+    vmovdqa     ymmD, ymmB
+    vpunpcklbw  ymmB, ymmB, ymmF      ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V)
+    vpunpckhbw  ymmD, ymmD, ymmF      ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V)
+
+    vmovdqa     ymmG, ymmE
+    vpunpcklbw  ymmE, ymmE, ymmF      ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U)
+    vpunpckhbw  ymmG, ymmG, ymmF      ; ymmG=(30 32 34 36 38 3A 3C 3E 3G 3I 3K 3M 3O 3Q 3S 3U)
+
+    vpunpcklbw  ymmF, ymmF, ymmH
+    vpunpckhbw  ymmH, ymmH, ymmH
+    vpsrlw      ymmF, ymmF, BYTE_BIT  ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V)
+    vpsrlw      ymmH, ymmH, BYTE_BIT  ; ymmH=(31 33 35 37 39 3B 3D 3F 3H 3J 3L 3N 3P 3R 3T 3V)
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    ; ymm0=R(02468ACEGIKMOQSU)=RE, ymm2=G(02468ACEGIKMOQSU)=GE, ymm4=B(02468ACEGIKMOQSU)=BE
+    ; ymm1=R(13579BDFHJLNPRTV)=RO, ymm3=G(13579BDFHJLNPRTV)=GO, ymm5=B(13579BDFHJLNPRTV)=BO
+
+    ; (Original)
+    ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+    ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+    ; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+    ;
+    ; (This implementation)
+    ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+    ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+    ; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+
+    vmovdqa     YMMWORD [wk(0)], ymm0   ; wk(0)=RE
+    vmovdqa     YMMWORD [wk(1)], ymm1   ; wk(1)=RO
+    vmovdqa     YMMWORD [wk(2)], ymm4   ; wk(2)=BE
+    vmovdqa     YMMWORD [wk(3)], ymm5   ; wk(3)=BO
+
+    vmovdqa     ymm6, ymm1
+    vpunpcklwd  ymm1, ymm1, ymm3
+    vpunpckhwd  ymm6, ymm6, ymm3
+    vmovdqa     ymm7, ymm1
+    vmovdqa     ymm4, ymm6
+    vpmaddwd    ymm1, ymm1, [GOTOFF(eax,PW_F0299_F0337)]  ; ymm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+    vpmaddwd    ymm6, ymm6, [GOTOFF(eax,PW_F0299_F0337)]  ; ymm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+    vpmaddwd    ymm7, ymm7, [GOTOFF(eax,PW_MF016_MF033)]  ; ymm7=ROL*-FIX(0.168)+GOL*-FIX(0.331)
+    vpmaddwd    ymm4, ymm4, [GOTOFF(eax,PW_MF016_MF033)]  ; ymm4=ROH*-FIX(0.168)+GOH*-FIX(0.331)
+
+    vmovdqa     YMMWORD [wk(4)], ymm1   ; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337)
+    vmovdqa     YMMWORD [wk(5)], ymm6   ; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    vpxor       ymm1, ymm1, ymm1
+    vpxor       ymm6, ymm6, ymm6
+    vpunpcklwd  ymm1, ymm1, ymm5        ; ymm1=BOL
+    vpunpckhwd  ymm6, ymm6, ymm5        ; ymm6=BOH
+    vpsrld      ymm1, ymm1, 1           ; ymm1=BOL*FIX(0.500)
+    vpsrld      ymm6, ymm6, 1           ; ymm6=BOH*FIX(0.500)
+
+    vmovdqa     ymm5, [GOTOFF(eax,PD_ONEHALFM1_CJ)]  ; ymm5=[PD_ONEHALFM1_CJ]
+
+    vpaddd      ymm7, ymm7, ymm1
+    vpaddd      ymm4, ymm4, ymm6
+    vpaddd      ymm7, ymm7, ymm5
+    vpaddd      ymm4, ymm4, ymm5
+    vpsrld      ymm7, ymm7, SCALEBITS   ; ymm7=CbOL
+    vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=CbOH
+    vpackssdw   ymm7, ymm7, ymm4        ; ymm7=CbO
+
+    vmovdqa     ymm1, YMMWORD [wk(2)]   ; ymm1=BE
+
+    vmovdqa     ymm6, ymm0
+    vpunpcklwd  ymm0, ymm0, ymm2
+    vpunpckhwd  ymm6, ymm6, ymm2
+    vmovdqa     ymm5, ymm0
+    vmovdqa     ymm4, ymm6
+    vpmaddwd    ymm0, ymm0, [GOTOFF(eax,PW_F0299_F0337)]  ; ymm0=REL*FIX(0.299)+GEL*FIX(0.337)
+    vpmaddwd    ymm6, ymm6, [GOTOFF(eax,PW_F0299_F0337)]  ; ymm6=REH*FIX(0.299)+GEH*FIX(0.337)
+    vpmaddwd    ymm5, ymm5, [GOTOFF(eax,PW_MF016_MF033)]  ; ymm5=REL*-FIX(0.168)+GEL*-FIX(0.331)
+    vpmaddwd    ymm4, ymm4, [GOTOFF(eax,PW_MF016_MF033)]  ; ymm4=REH*-FIX(0.168)+GEH*-FIX(0.331)
+
+    vmovdqa     YMMWORD [wk(6)], ymm0   ; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337)
+    vmovdqa     YMMWORD [wk(7)], ymm6   ; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    vpxor       ymm0, ymm0, ymm0
+    vpxor       ymm6, ymm6, ymm6
+    vpunpcklwd  ymm0, ymm0, ymm1        ; ymm0=BEL
+    vpunpckhwd  ymm6, ymm6, ymm1        ; ymm6=BEH
+    vpsrld      ymm0, ymm0, 1           ; ymm0=BEL*FIX(0.500)
+    vpsrld      ymm6, ymm6, 1           ; ymm6=BEH*FIX(0.500)
+
+    vmovdqa     ymm1, [GOTOFF(eax,PD_ONEHALFM1_CJ)]  ; ymm1=[PD_ONEHALFM1_CJ]
+
+    vpaddd      ymm5, ymm5, ymm0
+    vpaddd      ymm4, ymm4, ymm6
+    vpaddd      ymm5, ymm5, ymm1
+    vpaddd      ymm4, ymm4, ymm1
+    vpsrld      ymm5, ymm5, SCALEBITS   ; ymm5=CbEL
+    vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=CbEH
+    vpackssdw   ymm5, ymm5, ymm4        ; ymm5=CbE
+
+    vpsllw      ymm7, ymm7, BYTE_BIT
+    vpor        ymm5, ymm5, ymm7        ; ymm5=Cb
+    vmovdqu     YMMWORD [ebx], ymm5     ; Save Cb
+
+    vmovdqa     ymm0, YMMWORD [wk(3)]   ; ymm0=BO
+    vmovdqa     ymm6, YMMWORD [wk(2)]   ; ymm6=BE
+    vmovdqa     ymm1, YMMWORD [wk(1)]   ; ymm1=RO
+
+    vmovdqa     ymm4, ymm0
+    vpunpcklwd  ymm0, ymm0, ymm3
+    vpunpckhwd  ymm4, ymm4, ymm3
+    vmovdqa     ymm7, ymm0
+    vmovdqa     ymm5, ymm4
+    vpmaddwd    ymm0, ymm0, [GOTOFF(eax,PW_F0114_F0250)]  ; ymm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+    vpmaddwd    ymm4, ymm4, [GOTOFF(eax,PW_F0114_F0250)]  ; ymm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+    vpmaddwd    ymm7, ymm7, [GOTOFF(eax,PW_MF008_MF041)]  ; ymm7=BOL*-FIX(0.081)+GOL*-FIX(0.418)
+    vpmaddwd    ymm5, ymm5, [GOTOFF(eax,PW_MF008_MF041)]  ; ymm5=BOH*-FIX(0.081)+GOH*-FIX(0.418)
+
+    vmovdqa     ymm3, [GOTOFF(eax,PD_ONEHALF)]            ; ymm3=[PD_ONEHALF]
+
+    vpaddd      ymm0, ymm0, YMMWORD [wk(4)]
+    vpaddd      ymm4, ymm4, YMMWORD [wk(5)]
+    vpaddd      ymm0, ymm0, ymm3
+    vpaddd      ymm4, ymm4, ymm3
+    vpsrld      ymm0, ymm0, SCALEBITS   ; ymm0=YOL
+    vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=YOH
+    vpackssdw   ymm0, ymm0, ymm4        ; ymm0=YO
+
+    vpxor       ymm3, ymm3, ymm3
+    vpxor       ymm4, ymm4, ymm4
+    vpunpcklwd  ymm3, ymm3, ymm1        ; ymm3=ROL
+    vpunpckhwd  ymm4, ymm4, ymm1        ; ymm4=ROH
+    vpsrld      ymm3, ymm3, 1           ; ymm3=ROL*FIX(0.500)
+    vpsrld      ymm4, ymm4, 1           ; ymm4=ROH*FIX(0.500)
+
+    vmovdqa     ymm1, [GOTOFF(eax,PD_ONEHALFM1_CJ)]  ; ymm1=[PD_ONEHALFM1_CJ]
+
+    vpaddd      ymm7, ymm7, ymm3
+    vpaddd      ymm5, ymm5, ymm4
+    vpaddd      ymm7, ymm7, ymm1
+    vpaddd      ymm5, ymm5, ymm1
+    vpsrld      ymm7, ymm7, SCALEBITS   ; ymm7=CrOL
+    vpsrld      ymm5, ymm5, SCALEBITS   ; ymm5=CrOH
+    vpackssdw   ymm7, ymm7, ymm5        ; ymm7=CrO
+
+    vmovdqa     ymm3, YMMWORD [wk(0)]   ; ymm3=RE
+
+    vmovdqa     ymm4, ymm6
+    vpunpcklwd  ymm6, ymm6, ymm2
+    vpunpckhwd  ymm4, ymm4, ymm2
+    vmovdqa     ymm1, ymm6
+    vmovdqa     ymm5, ymm4
+    vpmaddwd    ymm6, ymm6, [GOTOFF(eax,PW_F0114_F0250)]  ; ymm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+    vpmaddwd    ymm4, ymm4, [GOTOFF(eax,PW_F0114_F0250)]  ; ymm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+    vpmaddwd    ymm1, ymm1, [GOTOFF(eax,PW_MF008_MF041)]  ; ymm1=BEL*-FIX(0.081)+GEL*-FIX(0.418)
+    vpmaddwd    ymm5, ymm5, [GOTOFF(eax,PW_MF008_MF041)]  ; ymm5=BEH*-FIX(0.081)+GEH*-FIX(0.418)
+
+    vmovdqa     ymm2, [GOTOFF(eax,PD_ONEHALF)]            ; ymm2=[PD_ONEHALF]
+
+    vpaddd      ymm6, ymm6, YMMWORD [wk(6)]
+    vpaddd      ymm4, ymm4, YMMWORD [wk(7)]
+    vpaddd      ymm6, ymm6, ymm2
+    vpaddd      ymm4, ymm4, ymm2
+    vpsrld      ymm6, ymm6, SCALEBITS   ; ymm6=YEL
+    vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=YEH
+    vpackssdw   ymm6, ymm6, ymm4        ; ymm6=YE
+
+    vpsllw      ymm0, ymm0, BYTE_BIT
+    vpor        ymm6, ymm6, ymm0        ; ymm6=Y
+    vmovdqu     YMMWORD [edi], ymm6     ; Save Y
+
+    vpxor       ymm2, ymm2, ymm2
+    vpxor       ymm4, ymm4, ymm4
+    vpunpcklwd  ymm2, ymm2, ymm3        ; ymm2=REL
+    vpunpckhwd  ymm4, ymm4, ymm3        ; ymm4=REH
+    vpsrld      ymm2, ymm2, 1           ; ymm2=REL*FIX(0.500)
+    vpsrld      ymm4, ymm4, 1           ; ymm4=REH*FIX(0.500)
+
+    vmovdqa     ymm0, [GOTOFF(eax,PD_ONEHALFM1_CJ)]  ; ymm0=[PD_ONEHALFM1_CJ]
+
+    vpaddd      ymm1, ymm1, ymm2
+    vpaddd      ymm5, ymm5, ymm4
+    vpaddd      ymm1, ymm1, ymm0
+    vpaddd      ymm5, ymm5, ymm0
+    vpsrld      ymm1, ymm1, SCALEBITS   ; ymm1=CrEL
+    vpsrld      ymm5, ymm5, SCALEBITS   ; ymm5=CrEH
+    vpackssdw   ymm1, ymm1, ymm5        ; ymm1=CrE
+
+    vpsllw      ymm7, ymm7, BYTE_BIT
+    vpor        ymm1, ymm1, ymm7        ; ymm1=Cr
+    vmovdqu     YMMWORD [edx], ymm1     ; Save Cr
+
+    sub         ecx, byte SIZEOF_YMMWORD
+    add         esi, RGB_PIXELSIZE*SIZEOF_YMMWORD  ; inptr
+    add         edi, byte SIZEOF_YMMWORD           ; outptr0
+    add         ebx, byte SIZEOF_YMMWORD           ; outptr1
+    add         edx, byte SIZEOF_YMMWORD           ; outptr2
+    cmp         ecx, byte SIZEOF_YMMWORD
+    jae         near .columnloop
+    test        ecx, ecx
+    jnz         near .column_ld1
+
+    pop         ecx                     ; col
+    pop         esi
+    pop         edi
+    pop         ebx
+    pop         edx
+    poppic      eax
+
+    add         esi, byte SIZEOF_JSAMPROW  ; input_buf
+    add         edi, byte SIZEOF_JSAMPROW
+    add         ebx, byte SIZEOF_JSAMPROW
+    add         edx, byte SIZEOF_JSAMPROW
+    dec         eax                        ; num_rows
+    jg          near .rowloop
+
+.return:
+    vzeroupper
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+    pop         ebx
+    mov         esp, ebp                ; esp <- aligned ebp
+    pop         esp                     ; esp <- original ebp
+    pop         ebp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,476 @@
+;
+; jccolext.asm - colorspace conversion (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_ycc_convert_mmx(JDIMENSION img_width, JSAMPARRAY input_buf,
+;                           JSAMPIMAGE output_buf, JDIMENSION output_row,
+;                           int num_rows);
+;
+
+%define img_width(b)   (b) + 8          ; JDIMENSION img_width
+%define input_buf(b)   (b) + 12         ; JSAMPARRAY input_buf
+%define output_buf(b)  (b) + 16         ; JSAMPIMAGE output_buf
+%define output_row(b)  (b) + 20         ; JDIMENSION output_row
+%define num_rows(b)    (b) + 24         ; int num_rows
+
+%define original_ebp   ebp + 0
+%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_MMWORD
+                                        ; mmword wk[WK_NUM]
+%define WK_NUM         8
+%define gotptr         wk(0) - SIZEOF_POINTER  ; void * gotptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_rgb_ycc_convert_mmx)
+
+EXTN(jsimd_rgb_ycc_convert_mmx):
+    push        ebp
+    mov         eax, esp                    ; eax = original ebp
+    sub         esp, byte 4
+    and         esp, byte (-SIZEOF_MMWORD)  ; align to 64 bits
+    mov         [esp], eax
+    mov         ebp, esp                    ; ebp = aligned ebp
+    lea         esp, [wk(0)]
+    pushpic     eax                     ; make a room for GOT address
+    push        ebx
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    get_GOT     ebx                     ; get GOT address
+    movpic      POINTER [gotptr], ebx   ; save GOT address
+
+    mov         ecx, JDIMENSION [img_width(eax)]  ; num_cols
+    test        ecx, ecx
+    jz          near .return
+
+    push        ecx
+
+    mov         esi, JSAMPIMAGE [output_buf(eax)]
+    mov         ecx, JDIMENSION [output_row(eax)]
+    mov         edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
+    mov         ebx, JSAMPARRAY [esi+1*SIZEOF_JSAMPARRAY]
+    mov         edx, JSAMPARRAY [esi+2*SIZEOF_JSAMPARRAY]
+    lea         edi, [edi+ecx*SIZEOF_JSAMPROW]
+    lea         ebx, [ebx+ecx*SIZEOF_JSAMPROW]
+    lea         edx, [edx+ecx*SIZEOF_JSAMPROW]
+
+    pop         ecx
+
+    mov         esi, JSAMPARRAY [input_buf(eax)]
+    mov         eax, INT [num_rows(eax)]
+    test        eax, eax
+    jle         near .return
+    alignx      16, 7
+.rowloop:
+    pushpic     eax
+    push        edx
+    push        ebx
+    push        edi
+    push        esi
+    push        ecx                     ; col
+
+    mov         esi, JSAMPROW [esi]     ; inptr
+    mov         edi, JSAMPROW [edi]     ; outptr0
+    mov         ebx, JSAMPROW [ebx]     ; outptr1
+    mov         edx, JSAMPROW [edx]     ; outptr2
+    movpic      eax, POINTER [gotptr]   ; load GOT address (eax)
+
+    cmp         ecx, byte SIZEOF_MMWORD
+    jae         short .columnloop
+    alignx      16, 7
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+.column_ld1:
+    push        eax
+    push        edx
+    lea         ecx, [ecx+ecx*2]        ; imul ecx,RGB_PIXELSIZE
+    test        cl, SIZEOF_BYTE
+    jz          short .column_ld2
+    sub         ecx, byte SIZEOF_BYTE
+    xor         eax, eax
+    mov         al, byte [esi+ecx]
+.column_ld2:
+    test        cl, SIZEOF_WORD
+    jz          short .column_ld4
+    sub         ecx, byte SIZEOF_WORD
+    xor         edx, edx
+    mov         dx, word [esi+ecx]
+    shl         eax, WORD_BIT
+    or          eax, edx
+.column_ld4:
+    movd        mmA, eax
+    pop         edx
+    pop         eax
+    test        cl, SIZEOF_DWORD
+    jz          short .column_ld8
+    sub         ecx, byte SIZEOF_DWORD
+    movd        mmG, dword [esi+ecx]
+    psllq       mmA, DWORD_BIT
+    por         mmA, mmG
+.column_ld8:
+    test        cl, SIZEOF_MMWORD
+    jz          short .column_ld16
+    movq        mmG, mmA
+    movq        mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+    mov         ecx, SIZEOF_MMWORD
+    jmp         short .rgb_ycc_cnv
+.column_ld16:
+    test        cl, 2*SIZEOF_MMWORD
+    mov         ecx, SIZEOF_MMWORD
+    jz          short .rgb_ycc_cnv
+    movq        mmF, mmA
+    movq        mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+    movq        mmG, MMWORD [esi+1*SIZEOF_MMWORD]
+    jmp         short .rgb_ycc_cnv
+    alignx      16, 7
+
+.columnloop:
+    movq        mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+    movq        mmG, MMWORD [esi+1*SIZEOF_MMWORD]
+    movq        mmF, MMWORD [esi+2*SIZEOF_MMWORD]
+
+.rgb_ycc_cnv:
+    ; mmA=(00 10 20 01 11 21 02 12)
+    ; mmG=(22 03 13 23 04 14 24 05)
+    ; mmF=(15 25 06 16 26 07 17 27)
+
+    movq        mmD, mmA
+    psllq       mmA, 4*BYTE_BIT         ; mmA=(-- -- -- -- 00 10 20 01)
+    psrlq       mmD, 4*BYTE_BIT         ; mmD=(11 21 02 12 -- -- -- --)
+
+    punpckhbw   mmA, mmG                ; mmA=(00 04 10 14 20 24 01 05)
+    psllq       mmG, 4*BYTE_BIT         ; mmG=(-- -- -- -- 22 03 13 23)
+
+    punpcklbw   mmD, mmF                ; mmD=(11 15 21 25 02 06 12 16)
+    punpckhbw   mmG, mmF                ; mmG=(22 26 03 07 13 17 23 27)
+
+    movq        mmE, mmA
+    psllq       mmA, 4*BYTE_BIT         ; mmA=(-- -- -- -- 00 04 10 14)
+    psrlq       mmE, 4*BYTE_BIT         ; mmE=(20 24 01 05 -- -- -- --)
+
+    punpckhbw   mmA, mmD                ; mmA=(00 02 04 06 10 12 14 16)
+    psllq       mmD, 4*BYTE_BIT         ; mmD=(-- -- -- -- 11 15 21 25)
+
+    punpcklbw   mmE, mmG                ; mmE=(20 22 24 26 01 03 05 07)
+    punpckhbw   mmD, mmG                ; mmD=(11 13 15 17 21 23 25 27)
+
+    pxor        mmH, mmH
+
+    movq        mmC, mmA
+    punpcklbw   mmA, mmH                ; mmA=(00 02 04 06)
+    punpckhbw   mmC, mmH                ; mmC=(10 12 14 16)
+
+    movq        mmB, mmE
+    punpcklbw   mmE, mmH                ; mmE=(20 22 24 26)
+    punpckhbw   mmB, mmH                ; mmB=(01 03 05 07)
+
+    movq        mmF, mmD
+    punpcklbw   mmD, mmH                ; mmD=(11 13 15 17)
+    punpckhbw   mmF, mmH                ; mmF=(21 23 25 27)
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+    test        cl, SIZEOF_MMWORD/8
+    jz          short .column_ld2
+    sub         ecx, byte SIZEOF_MMWORD/8
+    movd        mmA, dword [esi+ecx*RGB_PIXELSIZE]
+.column_ld2:
+    test        cl, SIZEOF_MMWORD/4
+    jz          short .column_ld4
+    sub         ecx, byte SIZEOF_MMWORD/4
+    movq        mmF, mmA
+    movq        mmA, MMWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld4:
+    test        cl, SIZEOF_MMWORD/2
+    mov         ecx, SIZEOF_MMWORD
+    jz          short .rgb_ycc_cnv
+    movq        mmD, mmA
+    movq        mmC, mmF
+    movq        mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+    movq        mmF, MMWORD [esi+1*SIZEOF_MMWORD]
+    jmp         short .rgb_ycc_cnv
+    alignx      16, 7
+
+.columnloop:
+    movq        mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+    movq        mmF, MMWORD [esi+1*SIZEOF_MMWORD]
+    movq        mmD, MMWORD [esi+2*SIZEOF_MMWORD]
+    movq        mmC, MMWORD [esi+3*SIZEOF_MMWORD]
+
+.rgb_ycc_cnv:
+    ; mmA=(00 10 20 30 01 11 21 31)
+    ; mmF=(02 12 22 32 03 13 23 33)
+    ; mmD=(04 14 24 34 05 15 25 35)
+    ; mmC=(06 16 26 36 07 17 27 37)
+
+    movq        mmB, mmA
+    punpcklbw   mmA, mmF                ; mmA=(00 02 10 12 20 22 30 32)
+    punpckhbw   mmB, mmF                ; mmB=(01 03 11 13 21 23 31 33)
+
+    movq        mmG, mmD
+    punpcklbw   mmD, mmC                ; mmD=(04 06 14 16 24 26 34 36)
+    punpckhbw   mmG, mmC                ; mmG=(05 07 15 17 25 27 35 37)
+
+    movq        mmE, mmA
+    punpcklwd   mmA, mmD                ; mmA=(00 02 04 06 10 12 14 16)
+    punpckhwd   mmE, mmD                ; mmE=(20 22 24 26 30 32 34 36)
+
+    movq        mmH, mmB
+    punpcklwd   mmB, mmG                ; mmB=(01 03 05 07 11 13 15 17)
+    punpckhwd   mmH, mmG                ; mmH=(21 23 25 27 31 33 35 37)
+
+    pxor        mmF, mmF
+
+    movq        mmC, mmA
+    punpcklbw   mmA, mmF                ; mmA=(00 02 04 06)
+    punpckhbw   mmC, mmF                ; mmC=(10 12 14 16)
+
+    movq        mmD, mmB
+    punpcklbw   mmB, mmF                ; mmB=(01 03 05 07)
+    punpckhbw   mmD, mmF                ; mmD=(11 13 15 17)
+
+    movq        mmG, mmE
+    punpcklbw   mmE, mmF                ; mmE=(20 22 24 26)
+    punpckhbw   mmG, mmF                ; mmG=(30 32 34 36)
+
+    punpcklbw   mmF, mmH
+    punpckhbw   mmH, mmH
+    psrlw       mmF, BYTE_BIT           ; mmF=(21 23 25 27)
+    psrlw       mmH, BYTE_BIT           ; mmH=(31 33 35 37)
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    ; mm0=(R0 R2 R4 R6)=RE, mm2=(G0 G2 G4 G6)=GE, mm4=(B0 B2 B4 B6)=BE
+    ; mm1=(R1 R3 R5 R7)=RO, mm3=(G1 G3 G5 G7)=GO, mm5=(B1 B3 B5 B7)=BO
+
+    ; (Original)
+    ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+    ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+    ; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+    ;
+    ; (This implementation)
+    ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+    ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+    ; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+
+    movq        MMWORD [wk(0)], mm0     ; wk(0)=RE
+    movq        MMWORD [wk(1)], mm1     ; wk(1)=RO
+    movq        MMWORD [wk(2)], mm4     ; wk(2)=BE
+    movq        MMWORD [wk(3)], mm5     ; wk(3)=BO
+
+    movq        mm6, mm1
+    punpcklwd   mm1, mm3
+    punpckhwd   mm6, mm3
+    movq        mm7, mm1
+    movq        mm4, mm6
+    pmaddwd     mm1, [GOTOFF(eax,PW_F0299_F0337)]  ; mm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+    pmaddwd     mm6, [GOTOFF(eax,PW_F0299_F0337)]  ; mm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+    pmaddwd     mm7, [GOTOFF(eax,PW_MF016_MF033)]  ; mm7=ROL*-FIX(0.168)+GOL*-FIX(0.331)
+    pmaddwd     mm4, [GOTOFF(eax,PW_MF016_MF033)]  ; mm4=ROH*-FIX(0.168)+GOH*-FIX(0.331)
+
+    movq        MMWORD [wk(4)], mm1     ; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337)
+    movq        MMWORD [wk(5)], mm6     ; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    pxor        mm1, mm1
+    pxor        mm6, mm6
+    punpcklwd   mm1, mm5                ; mm1=BOL
+    punpckhwd   mm6, mm5                ; mm6=BOH
+    psrld       mm1, 1                  ; mm1=BOL*FIX(0.500)
+    psrld       mm6, 1                  ; mm6=BOH*FIX(0.500)
+
+    movq        mm5, [GOTOFF(eax,PD_ONEHALFM1_CJ)]  ; mm5=[PD_ONEHALFM1_CJ]
+
+    paddd       mm7, mm1
+    paddd       mm4, mm6
+    paddd       mm7, mm5
+    paddd       mm4, mm5
+    psrld       mm7, SCALEBITS          ; mm7=CbOL
+    psrld       mm4, SCALEBITS          ; mm4=CbOH
+    packssdw    mm7, mm4                ; mm7=CbO
+
+    movq        mm1, MMWORD [wk(2)]     ; mm1=BE
+
+    movq        mm6, mm0
+    punpcklwd   mm0, mm2
+    punpckhwd   mm6, mm2
+    movq        mm5, mm0
+    movq        mm4, mm6
+    pmaddwd     mm0, [GOTOFF(eax,PW_F0299_F0337)]  ; mm0=REL*FIX(0.299)+GEL*FIX(0.337)
+    pmaddwd     mm6, [GOTOFF(eax,PW_F0299_F0337)]  ; mm6=REH*FIX(0.299)+GEH*FIX(0.337)
+    pmaddwd     mm5, [GOTOFF(eax,PW_MF016_MF033)]  ; mm5=REL*-FIX(0.168)+GEL*-FIX(0.331)
+    pmaddwd     mm4, [GOTOFF(eax,PW_MF016_MF033)]  ; mm4=REH*-FIX(0.168)+GEH*-FIX(0.331)
+
+    movq        MMWORD [wk(6)], mm0     ; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337)
+    movq        MMWORD [wk(7)], mm6     ; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    pxor        mm0, mm0
+    pxor        mm6, mm6
+    punpcklwd   mm0, mm1                ; mm0=BEL
+    punpckhwd   mm6, mm1                ; mm6=BEH
+    psrld       mm0, 1                  ; mm0=BEL*FIX(0.500)
+    psrld       mm6, 1                  ; mm6=BEH*FIX(0.500)
+
+    movq        mm1, [GOTOFF(eax,PD_ONEHALFM1_CJ)]  ; mm1=[PD_ONEHALFM1_CJ]
+
+    paddd       mm5, mm0
+    paddd       mm4, mm6
+    paddd       mm5, mm1
+    paddd       mm4, mm1
+    psrld       mm5, SCALEBITS          ; mm5=CbEL
+    psrld       mm4, SCALEBITS          ; mm4=CbEH
+    packssdw    mm5, mm4                ; mm5=CbE
+
+    psllw       mm7, BYTE_BIT
+    por         mm5, mm7                ; mm5=Cb
+    movq        MMWORD [ebx], mm5       ; Save Cb
+
+    movq        mm0, MMWORD [wk(3)]     ; mm0=BO
+    movq        mm6, MMWORD [wk(2)]     ; mm6=BE
+    movq        mm1, MMWORD [wk(1)]     ; mm1=RO
+
+    movq        mm4, mm0
+    punpcklwd   mm0, mm3
+    punpckhwd   mm4, mm3
+    movq        mm7, mm0
+    movq        mm5, mm4
+    pmaddwd     mm0, [GOTOFF(eax,PW_F0114_F0250)]  ; mm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+    pmaddwd     mm4, [GOTOFF(eax,PW_F0114_F0250)]  ; mm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+    pmaddwd     mm7, [GOTOFF(eax,PW_MF008_MF041)]  ; mm7=BOL*-FIX(0.081)+GOL*-FIX(0.418)
+    pmaddwd     mm5, [GOTOFF(eax,PW_MF008_MF041)]  ; mm5=BOH*-FIX(0.081)+GOH*-FIX(0.418)
+
+    movq        mm3, [GOTOFF(eax,PD_ONEHALF)]  ; mm3=[PD_ONEHALF]
+
+    paddd       mm0, MMWORD [wk(4)]
+    paddd       mm4, MMWORD [wk(5)]
+    paddd       mm0, mm3
+    paddd       mm4, mm3
+    psrld       mm0, SCALEBITS          ; mm0=YOL
+    psrld       mm4, SCALEBITS          ; mm4=YOH
+    packssdw    mm0, mm4                ; mm0=YO
+
+    pxor        mm3, mm3
+    pxor        mm4, mm4
+    punpcklwd   mm3, mm1                ; mm3=ROL
+    punpckhwd   mm4, mm1                ; mm4=ROH
+    psrld       mm3, 1                  ; mm3=ROL*FIX(0.500)
+    psrld       mm4, 1                  ; mm4=ROH*FIX(0.500)
+
+    movq        mm1, [GOTOFF(eax,PD_ONEHALFM1_CJ)]  ; mm1=[PD_ONEHALFM1_CJ]
+
+    paddd       mm7, mm3
+    paddd       mm5, mm4
+    paddd       mm7, mm1
+    paddd       mm5, mm1
+    psrld       mm7, SCALEBITS          ; mm7=CrOL
+    psrld       mm5, SCALEBITS          ; mm5=CrOH
+    packssdw    mm7, mm5                ; mm7=CrO
+
+    movq        mm3, MMWORD [wk(0)]     ; mm3=RE
+
+    movq        mm4, mm6
+    punpcklwd   mm6, mm2
+    punpckhwd   mm4, mm2
+    movq        mm1, mm6
+    movq        mm5, mm4
+    pmaddwd     mm6, [GOTOFF(eax,PW_F0114_F0250)]  ; mm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+    pmaddwd     mm4, [GOTOFF(eax,PW_F0114_F0250)]  ; mm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+    pmaddwd     mm1, [GOTOFF(eax,PW_MF008_MF041)]  ; mm1=BEL*-FIX(0.081)+GEL*-FIX(0.418)
+    pmaddwd     mm5, [GOTOFF(eax,PW_MF008_MF041)]  ; mm5=BEH*-FIX(0.081)+GEH*-FIX(0.418)
+
+    movq        mm2, [GOTOFF(eax,PD_ONEHALF)]      ; mm2=[PD_ONEHALF]
+
+    paddd       mm6, MMWORD [wk(6)]
+    paddd       mm4, MMWORD [wk(7)]
+    paddd       mm6, mm2
+    paddd       mm4, mm2
+    psrld       mm6, SCALEBITS          ; mm6=YEL
+    psrld       mm4, SCALEBITS          ; mm4=YEH
+    packssdw    mm6, mm4                ; mm6=YE
+
+    psllw       mm0, BYTE_BIT
+    por         mm6, mm0                ; mm6=Y
+    movq        MMWORD [edi], mm6       ; Save Y
+
+    pxor        mm2, mm2
+    pxor        mm4, mm4
+    punpcklwd   mm2, mm3                ; mm2=REL
+    punpckhwd   mm4, mm3                ; mm4=REH
+    psrld       mm2, 1                  ; mm2=REL*FIX(0.500)
+    psrld       mm4, 1                  ; mm4=REH*FIX(0.500)
+
+    movq        mm0, [GOTOFF(eax,PD_ONEHALFM1_CJ)]  ; mm0=[PD_ONEHALFM1_CJ]
+
+    paddd       mm1, mm2
+    paddd       mm5, mm4
+    paddd       mm1, mm0
+    paddd       mm5, mm0
+    psrld       mm1, SCALEBITS          ; mm1=CrEL
+    psrld       mm5, SCALEBITS          ; mm5=CrEH
+    packssdw    mm1, mm5                ; mm1=CrE
+
+    psllw       mm7, BYTE_BIT
+    por         mm1, mm7                ; mm1=Cr
+    movq        MMWORD [edx], mm1       ; Save Cr
+
+    sub         ecx, byte SIZEOF_MMWORD
+    add         esi, byte RGB_PIXELSIZE*SIZEOF_MMWORD  ; inptr
+    add         edi, byte SIZEOF_MMWORD                ; outptr0
+    add         ebx, byte SIZEOF_MMWORD                ; outptr1
+    add         edx, byte SIZEOF_MMWORD                ; outptr2
+    cmp         ecx, byte SIZEOF_MMWORD
+    jae         near .columnloop
+    test        ecx, ecx
+    jnz         near .column_ld1
+
+    pop         ecx                     ; col
+    pop         esi
+    pop         edi
+    pop         ebx
+    pop         edx
+    poppic      eax
+
+    add         esi, byte SIZEOF_JSAMPROW  ; input_buf
+    add         edi, byte SIZEOF_JSAMPROW
+    add         ebx, byte SIZEOF_JSAMPROW
+    add         edx, byte SIZEOF_JSAMPROW
+    dec         eax                        ; num_rows
+    jg          near .rowloop
+
+    emms                                ; empty MMX state
+
+.return:
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+    pop         ebx
+    mov         esp, ebp                ; esp <- aligned ebp
+    pop         esp                     ; esp <- original ebp
+    pop         ebp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,503 @@
+;
+; jccolext.asm - colorspace conversion (SSE2)
+;
+; Copyright (C) 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_ycc_convert_sse2(JDIMENSION img_width, JSAMPARRAY input_buf,
+;                            JSAMPIMAGE output_buf, JDIMENSION output_row,
+;                            int num_rows);
+;
+
+%define img_width(b)   (b) + 8          ; JDIMENSION img_width
+%define input_buf(b)   (b) + 12         ; JSAMPARRAY input_buf
+%define output_buf(b)  (b) + 16         ; JSAMPIMAGE output_buf
+%define output_row(b)  (b) + 20         ; JDIMENSION output_row
+%define num_rows(b)    (b) + 24         ; int num_rows
+
+%define original_ebp   ebp + 0
+%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_XMMWORD
+                                        ; xmmword wk[WK_NUM]
+%define WK_NUM         8
+%define gotptr         wk(0) - SIZEOF_POINTER  ; void * gotptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_rgb_ycc_convert_sse2)
+
+EXTN(jsimd_rgb_ycc_convert_sse2):
+    push        ebp
+    mov         eax, esp                     ; eax = original ebp
+    sub         esp, byte 4
+    and         esp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [esp], eax
+    mov         ebp, esp                     ; ebp = aligned ebp
+    lea         esp, [wk(0)]
+    pushpic     eax                     ; make a room for GOT address
+    push        ebx
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    get_GOT     ebx                     ; get GOT address
+    movpic      POINTER [gotptr], ebx   ; save GOT address
+
+    mov         ecx, JDIMENSION [img_width(eax)]
+    test        ecx, ecx
+    jz          near .return
+
+    push        ecx
+
+    mov         esi, JSAMPIMAGE [output_buf(eax)]
+    mov         ecx, JDIMENSION [output_row(eax)]
+    mov         edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
+    mov         ebx, JSAMPARRAY [esi+1*SIZEOF_JSAMPARRAY]
+    mov         edx, JSAMPARRAY [esi+2*SIZEOF_JSAMPARRAY]
+    lea         edi, [edi+ecx*SIZEOF_JSAMPROW]
+    lea         ebx, [ebx+ecx*SIZEOF_JSAMPROW]
+    lea         edx, [edx+ecx*SIZEOF_JSAMPROW]
+
+    pop         ecx
+
+    mov         esi, JSAMPARRAY [input_buf(eax)]
+    mov         eax, INT [num_rows(eax)]
+    test        eax, eax
+    jle         near .return
+    alignx      16, 7
+.rowloop:
+    pushpic     eax
+    push        edx
+    push        ebx
+    push        edi
+    push        esi
+    push        ecx                     ; col
+
+    mov         esi, JSAMPROW [esi]     ; inptr
+    mov         edi, JSAMPROW [edi]     ; outptr0
+    mov         ebx, JSAMPROW [ebx]     ; outptr1
+    mov         edx, JSAMPROW [edx]     ; outptr2
+    movpic      eax, POINTER [gotptr]   ; load GOT address (eax)
+
+    cmp         ecx, byte SIZEOF_XMMWORD
+    jae         near .columnloop
+    alignx      16, 7
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+.column_ld1:
+    push        eax
+    push        edx
+    lea         ecx, [ecx+ecx*2]        ; imul ecx,RGB_PIXELSIZE
+    test        cl, SIZEOF_BYTE
+    jz          short .column_ld2
+    sub         ecx, byte SIZEOF_BYTE
+    movzx       eax, byte [esi+ecx]
+.column_ld2:
+    test        cl, SIZEOF_WORD
+    jz          short .column_ld4
+    sub         ecx, byte SIZEOF_WORD
+    movzx       edx, word [esi+ecx]
+    shl         eax, WORD_BIT
+    or          eax, edx
+.column_ld4:
+    movd        xmmA, eax
+    pop         edx
+    pop         eax
+    test        cl, SIZEOF_DWORD
+    jz          short .column_ld8
+    sub         ecx, byte SIZEOF_DWORD
+    movd        xmmF, XMM_DWORD [esi+ecx]
+    pslldq      xmmA, SIZEOF_DWORD
+    por         xmmA, xmmF
+.column_ld8:
+    test        cl, SIZEOF_MMWORD
+    jz          short .column_ld16
+    sub         ecx, byte SIZEOF_MMWORD
+    movq        xmmB, XMM_MMWORD [esi+ecx]
+    pslldq      xmmA, SIZEOF_MMWORD
+    por         xmmA, xmmB
+.column_ld16:
+    test        cl, SIZEOF_XMMWORD
+    jz          short .column_ld32
+    movdqa      xmmF, xmmA
+    movdqu      xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    mov         ecx, SIZEOF_XMMWORD
+    jmp         short .rgb_ycc_cnv
+.column_ld32:
+    test        cl, 2*SIZEOF_XMMWORD
+    mov         ecx, SIZEOF_XMMWORD
+    jz          short .rgb_ycc_cnv
+    movdqa      xmmB, xmmA
+    movdqu      xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    movdqu      xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD]
+    jmp         short .rgb_ycc_cnv
+    alignx      16, 7
+
+.columnloop:
+    movdqu      xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    movdqu      xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD]
+    movdqu      xmmB, XMMWORD [esi+2*SIZEOF_XMMWORD]
+
+.rgb_ycc_cnv:
+    ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+    ; xmmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    ; xmmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+    movdqa      xmmG, xmmA
+    pslldq      xmmA, 8     ; xmmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12)
+    psrldq      xmmG, 8     ; xmmG=(22 03 13 23 04 14 24 05 -- -- -- -- -- -- -- --)
+
+    punpckhbw   xmmA, xmmF  ; xmmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A)
+    pslldq      xmmF, 8     ; xmmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27)
+
+    punpcklbw   xmmG, xmmB  ; xmmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D)
+    punpckhbw   xmmF, xmmB  ; xmmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F)
+
+    movdqa      xmmD, xmmA
+    pslldq      xmmA, 8     ; xmmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09)
+    psrldq      xmmD, 8     ; xmmD=(11 19 21 29 02 0A 12 1A -- -- -- -- -- -- -- --)
+
+    punpckhbw   xmmA, xmmG  ; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D)
+    pslldq      xmmG, 8     ; xmmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B)
+
+    punpcklbw   xmmD, xmmF  ; xmmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E)
+    punpckhbw   xmmG, xmmF  ; xmmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F)
+
+    movdqa      xmmE, xmmA
+    pslldq      xmmA, 8     ; xmmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C)
+    psrldq      xmmE, 8     ; xmmE=(20 24 28 2C 01 05 09 0D -- -- -- -- -- -- -- --)
+
+    punpckhbw   xmmA, xmmD  ; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
+    pslldq      xmmD, 8     ; xmmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D)
+
+    punpcklbw   xmmE, xmmG  ; xmmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F)
+    punpckhbw   xmmD, xmmG  ; xmmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F)
+
+    pxor        xmmH, xmmH
+
+    movdqa      xmmC, xmmA
+    punpcklbw   xmmA, xmmH  ; xmmA=(00 02 04 06 08 0A 0C 0E)
+    punpckhbw   xmmC, xmmH  ; xmmC=(10 12 14 16 18 1A 1C 1E)
+
+    movdqa      xmmB, xmmE
+    punpcklbw   xmmE, xmmH  ; xmmE=(20 22 24 26 28 2A 2C 2E)
+    punpckhbw   xmmB, xmmH  ; xmmB=(01 03 05 07 09 0B 0D 0F)
+
+    movdqa      xmmF, xmmD
+    punpcklbw   xmmD, xmmH  ; xmmD=(11 13 15 17 19 1B 1D 1F)
+    punpckhbw   xmmF, xmmH  ; xmmF=(21 23 25 27 29 2B 2D 2F)
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+    test        cl, SIZEOF_XMMWORD/16
+    jz          short .column_ld2
+    sub         ecx, byte SIZEOF_XMMWORD/16
+    movd        xmmA, XMM_DWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld2:
+    test        cl, SIZEOF_XMMWORD/8
+    jz          short .column_ld4
+    sub         ecx, byte SIZEOF_XMMWORD/8
+    movq        xmmE, XMM_MMWORD [esi+ecx*RGB_PIXELSIZE]
+    pslldq      xmmA, SIZEOF_MMWORD
+    por         xmmA, xmmE
+.column_ld4:
+    test        cl, SIZEOF_XMMWORD/4
+    jz          short .column_ld8
+    sub         ecx, byte SIZEOF_XMMWORD/4
+    movdqa      xmmE, xmmA
+    movdqu      xmmA, XMMWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld8:
+    test        cl, SIZEOF_XMMWORD/2
+    mov         ecx, SIZEOF_XMMWORD
+    jz          short .rgb_ycc_cnv
+    movdqa      xmmF, xmmA
+    movdqa      xmmH, xmmE
+    movdqu      xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    movdqu      xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD]
+    jmp         short .rgb_ycc_cnv
+    alignx      16, 7
+
+.columnloop:
+    movdqu      xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    movdqu      xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD]
+    movdqu      xmmF, XMMWORD [esi+2*SIZEOF_XMMWORD]
+    movdqu      xmmH, XMMWORD [esi+3*SIZEOF_XMMWORD]
+
+.rgb_ycc_cnv:
+    ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+    ; xmmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    ; xmmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+    ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+    movdqa      xmmD, xmmA
+    punpcklbw   xmmA, xmmE      ; xmmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35)
+    punpckhbw   xmmD, xmmE      ; xmmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37)
+
+    movdqa      xmmC, xmmF
+    punpcklbw   xmmF, xmmH      ; xmmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D)
+    punpckhbw   xmmC, xmmH      ; xmmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F)
+
+    movdqa      xmmB, xmmA
+    punpcklwd   xmmA, xmmF      ; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C)
+    punpckhwd   xmmB, xmmF      ; xmmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D)
+
+    movdqa      xmmG, xmmD
+    punpcklwd   xmmD, xmmC      ; xmmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E)
+    punpckhwd   xmmG, xmmC      ; xmmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F)
+
+    movdqa      xmmE, xmmA
+    punpcklbw   xmmA, xmmD      ; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
+    punpckhbw   xmmE, xmmD      ; xmmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E)
+
+    movdqa      xmmH, xmmB
+    punpcklbw   xmmB, xmmG      ; xmmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F)
+    punpckhbw   xmmH, xmmG      ; xmmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F)
+
+    pxor        xmmF, xmmF
+
+    movdqa      xmmC, xmmA
+    punpcklbw   xmmA, xmmF      ; xmmA=(00 02 04 06 08 0A 0C 0E)
+    punpckhbw   xmmC, xmmF      ; xmmC=(10 12 14 16 18 1A 1C 1E)
+
+    movdqa      xmmD, xmmB
+    punpcklbw   xmmB, xmmF      ; xmmB=(01 03 05 07 09 0B 0D 0F)
+    punpckhbw   xmmD, xmmF      ; xmmD=(11 13 15 17 19 1B 1D 1F)
+
+    movdqa      xmmG, xmmE
+    punpcklbw   xmmE, xmmF      ; xmmE=(20 22 24 26 28 2A 2C 2E)
+    punpckhbw   xmmG, xmmF      ; xmmG=(30 32 34 36 38 3A 3C 3E)
+
+    punpcklbw   xmmF, xmmH
+    punpckhbw   xmmH, xmmH
+    psrlw       xmmF, BYTE_BIT  ; xmmF=(21 23 25 27 29 2B 2D 2F)
+    psrlw       xmmH, BYTE_BIT  ; xmmH=(31 33 35 37 39 3B 3D 3F)
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    ; xmm0=R(02468ACE)=RE, xmm2=G(02468ACE)=GE, xmm4=B(02468ACE)=BE
+    ; xmm1=R(13579BDF)=RO, xmm3=G(13579BDF)=GO, xmm5=B(13579BDF)=BO
+
+    ; (Original)
+    ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+    ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+    ; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+    ;
+    ; (This implementation)
+    ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+    ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
+    ; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
+
+    movdqa      XMMWORD [wk(0)], xmm0   ; wk(0)=RE
+    movdqa      XMMWORD [wk(1)], xmm1   ; wk(1)=RO
+    movdqa      XMMWORD [wk(2)], xmm4   ; wk(2)=BE
+    movdqa      XMMWORD [wk(3)], xmm5   ; wk(3)=BO
+
+    movdqa      xmm6, xmm1
+    punpcklwd   xmm1, xmm3
+    punpckhwd   xmm6, xmm3
+    movdqa      xmm7, xmm1
+    movdqa      xmm4, xmm6
+    pmaddwd     xmm1, [GOTOFF(eax,PW_F0299_F0337)]  ; xmm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+    pmaddwd     xmm6, [GOTOFF(eax,PW_F0299_F0337)]  ; xmm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+    pmaddwd     xmm7, [GOTOFF(eax,PW_MF016_MF033)]  ; xmm7=ROL*-FIX(0.168)+GOL*-FIX(0.331)
+    pmaddwd     xmm4, [GOTOFF(eax,PW_MF016_MF033)]  ; xmm4=ROH*-FIX(0.168)+GOH*-FIX(0.331)
+
+    movdqa      XMMWORD [wk(4)], xmm1   ; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337)
+    movdqa      XMMWORD [wk(5)], xmm6   ; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    pxor        xmm1, xmm1
+    pxor        xmm6, xmm6
+    punpcklwd   xmm1, xmm5              ; xmm1=BOL
+    punpckhwd   xmm6, xmm5              ; xmm6=BOH
+    psrld       xmm1, 1                 ; xmm1=BOL*FIX(0.500)
+    psrld       xmm6, 1                 ; xmm6=BOH*FIX(0.500)
+
+    movdqa      xmm5, [GOTOFF(eax,PD_ONEHALFM1_CJ)]  ; xmm5=[PD_ONEHALFM1_CJ]
+
+    paddd       xmm7, xmm1
+    paddd       xmm4, xmm6
+    paddd       xmm7, xmm5
+    paddd       xmm4, xmm5
+    psrld       xmm7, SCALEBITS         ; xmm7=CbOL
+    psrld       xmm4, SCALEBITS         ; xmm4=CbOH
+    packssdw    xmm7, xmm4              ; xmm7=CbO
+
+    movdqa      xmm1, XMMWORD [wk(2)]   ; xmm1=BE
+
+    movdqa      xmm6, xmm0
+    punpcklwd   xmm0, xmm2
+    punpckhwd   xmm6, xmm2
+    movdqa      xmm5, xmm0
+    movdqa      xmm4, xmm6
+    pmaddwd     xmm0, [GOTOFF(eax,PW_F0299_F0337)]  ; xmm0=REL*FIX(0.299)+GEL*FIX(0.337)
+    pmaddwd     xmm6, [GOTOFF(eax,PW_F0299_F0337)]  ; xmm6=REH*FIX(0.299)+GEH*FIX(0.337)
+    pmaddwd     xmm5, [GOTOFF(eax,PW_MF016_MF033)]  ; xmm5=REL*-FIX(0.168)+GEL*-FIX(0.331)
+    pmaddwd     xmm4, [GOTOFF(eax,PW_MF016_MF033)]  ; xmm4=REH*-FIX(0.168)+GEH*-FIX(0.331)
+
+    movdqa      XMMWORD [wk(6)], xmm0   ; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337)
+    movdqa      XMMWORD [wk(7)], xmm6   ; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    pxor        xmm0, xmm0
+    pxor        xmm6, xmm6
+    punpcklwd   xmm0, xmm1              ; xmm0=BEL
+    punpckhwd   xmm6, xmm1              ; xmm6=BEH
+    psrld       xmm0, 1                 ; xmm0=BEL*FIX(0.500)
+    psrld       xmm6, 1                 ; xmm6=BEH*FIX(0.500)
+
+    movdqa      xmm1, [GOTOFF(eax,PD_ONEHALFM1_CJ)]  ; xmm1=[PD_ONEHALFM1_CJ]
+
+    paddd       xmm5, xmm0
+    paddd       xmm4, xmm6
+    paddd       xmm5, xmm1
+    paddd       xmm4, xmm1
+    psrld       xmm5, SCALEBITS         ; xmm5=CbEL
+    psrld       xmm4, SCALEBITS         ; xmm4=CbEH
+    packssdw    xmm5, xmm4              ; xmm5=CbE
+
+    psllw       xmm7, BYTE_BIT
+    por         xmm5, xmm7              ; xmm5=Cb
+    movdqa      XMMWORD [ebx], xmm5     ; Save Cb
+
+    movdqa      xmm0, XMMWORD [wk(3)]   ; xmm0=BO
+    movdqa      xmm6, XMMWORD [wk(2)]   ; xmm6=BE
+    movdqa      xmm1, XMMWORD [wk(1)]   ; xmm1=RO
+
+    movdqa      xmm4, xmm0
+    punpcklwd   xmm0, xmm3
+    punpckhwd   xmm4, xmm3
+    movdqa      xmm7, xmm0
+    movdqa      xmm5, xmm4
+    pmaddwd     xmm0, [GOTOFF(eax,PW_F0114_F0250)]  ; xmm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+    pmaddwd     xmm4, [GOTOFF(eax,PW_F0114_F0250)]  ; xmm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+    pmaddwd     xmm7, [GOTOFF(eax,PW_MF008_MF041)]  ; xmm7=BOL*-FIX(0.081)+GOL*-FIX(0.418)
+    pmaddwd     xmm5, [GOTOFF(eax,PW_MF008_MF041)]  ; xmm5=BOH*-FIX(0.081)+GOH*-FIX(0.418)
+
+    movdqa      xmm3, [GOTOFF(eax,PD_ONEHALF)]  ; xmm3=[PD_ONEHALF]
+
+    paddd       xmm0, XMMWORD [wk(4)]
+    paddd       xmm4, XMMWORD [wk(5)]
+    paddd       xmm0, xmm3
+    paddd       xmm4, xmm3
+    psrld       xmm0, SCALEBITS         ; xmm0=YOL
+    psrld       xmm4, SCALEBITS         ; xmm4=YOH
+    packssdw    xmm0, xmm4              ; xmm0=YO
+
+    pxor        xmm3, xmm3
+    pxor        xmm4, xmm4
+    punpcklwd   xmm3, xmm1              ; xmm3=ROL
+    punpckhwd   xmm4, xmm1              ; xmm4=ROH
+    psrld       xmm3, 1                 ; xmm3=ROL*FIX(0.500)
+    psrld       xmm4, 1                 ; xmm4=ROH*FIX(0.500)
+
+    movdqa      xmm1, [GOTOFF(eax,PD_ONEHALFM1_CJ)]  ; xmm1=[PD_ONEHALFM1_CJ]
+
+    paddd       xmm7, xmm3
+    paddd       xmm5, xmm4
+    paddd       xmm7, xmm1
+    paddd       xmm5, xmm1
+    psrld       xmm7, SCALEBITS         ; xmm7=CrOL
+    psrld       xmm5, SCALEBITS         ; xmm5=CrOH
+    packssdw    xmm7, xmm5              ; xmm7=CrO
+
+    movdqa      xmm3, XMMWORD [wk(0)]   ; xmm3=RE
+
+    movdqa      xmm4, xmm6
+    punpcklwd   xmm6, xmm2
+    punpckhwd   xmm4, xmm2
+    movdqa      xmm1, xmm6
+    movdqa      xmm5, xmm4
+    pmaddwd     xmm6, [GOTOFF(eax,PW_F0114_F0250)]  ; xmm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+    pmaddwd     xmm4, [GOTOFF(eax,PW_F0114_F0250)]  ; xmm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+    pmaddwd     xmm1, [GOTOFF(eax,PW_MF008_MF041)]  ; xmm1=BEL*-FIX(0.081)+GEL*-FIX(0.418)
+    pmaddwd     xmm5, [GOTOFF(eax,PW_MF008_MF041)]  ; xmm5=BEH*-FIX(0.081)+GEH*-FIX(0.418)
+
+    movdqa      xmm2, [GOTOFF(eax,PD_ONEHALF)]      ; xmm2=[PD_ONEHALF]
+
+    paddd       xmm6, XMMWORD [wk(6)]
+    paddd       xmm4, XMMWORD [wk(7)]
+    paddd       xmm6, xmm2
+    paddd       xmm4, xmm2
+    psrld       xmm6, SCALEBITS         ; xmm6=YEL
+    psrld       xmm4, SCALEBITS         ; xmm4=YEH
+    packssdw    xmm6, xmm4              ; xmm6=YE
+
+    psllw       xmm0, BYTE_BIT
+    por         xmm6, xmm0              ; xmm6=Y
+    movdqa      XMMWORD [edi], xmm6     ; Save Y
+
+    pxor        xmm2, xmm2
+    pxor        xmm4, xmm4
+    punpcklwd   xmm2, xmm3              ; xmm2=REL
+    punpckhwd   xmm4, xmm3              ; xmm4=REH
+    psrld       xmm2, 1                 ; xmm2=REL*FIX(0.500)
+    psrld       xmm4, 1                 ; xmm4=REH*FIX(0.500)
+
+    movdqa      xmm0, [GOTOFF(eax,PD_ONEHALFM1_CJ)]  ; xmm0=[PD_ONEHALFM1_CJ]
+
+    paddd       xmm1, xmm2
+    paddd       xmm5, xmm4
+    paddd       xmm1, xmm0
+    paddd       xmm5, xmm0
+    psrld       xmm1, SCALEBITS         ; xmm1=CrEL
+    psrld       xmm5, SCALEBITS         ; xmm5=CrEH
+    packssdw    xmm1, xmm5              ; xmm1=CrE
+
+    psllw       xmm7, BYTE_BIT
+    por         xmm1, xmm7              ; xmm1=Cr
+    movdqa      XMMWORD [edx], xmm1     ; Save Cr
+
+    sub         ecx, byte SIZEOF_XMMWORD
+    add         esi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; inptr
+    add         edi, byte SIZEOF_XMMWORD                ; outptr0
+    add         ebx, byte SIZEOF_XMMWORD                ; outptr1
+    add         edx, byte SIZEOF_XMMWORD                ; outptr2
+    cmp         ecx, byte SIZEOF_XMMWORD
+    jae         near .columnloop
+    test        ecx, ecx
+    jnz         near .column_ld1
+
+    pop         ecx                     ; col
+    pop         esi
+    pop         edi
+    pop         ebx
+    pop         edx
+    poppic      eax
+
+    add         esi, byte SIZEOF_JSAMPROW  ; input_buf
+    add         edi, byte SIZEOF_JSAMPROW
+    add         ebx, byte SIZEOF_JSAMPROW
+    add         edx, byte SIZEOF_JSAMPROW
+    dec         eax                        ; num_rows
+    jg          near .rowloop
+
+.return:
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+    pop         ebx
+    mov         esp, ebp                ; esp <- aligned ebp
+    pop         esp                     ; esp <- original ebp
+    pop         ebp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,121 @@
+;
+; jccolor.asm - colorspace conversion (AVX2)
+;
+; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_081 equ  5329                ; FIX(0.08131)
+F_0_114 equ  7471                ; FIX(0.11400)
+F_0_168 equ 11059                ; FIX(0.16874)
+F_0_250 equ 16384                ; FIX(0.25000)
+F_0_299 equ 19595                ; FIX(0.29900)
+F_0_331 equ 21709                ; FIX(0.33126)
+F_0_418 equ 27439                ; FIX(0.41869)
+F_0_587 equ 38470                ; FIX(0.58700)
+F_0_337 equ (F_0_587 - F_0_250)  ; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_rgb_ycc_convert_avx2)
+
+EXTN(jconst_rgb_ycc_convert_avx2):
+
+PW_F0299_F0337  times 8 dw  F_0_299,  F_0_337
+PW_F0114_F0250  times 8 dw  F_0_114,  F_0_250
+PW_MF016_MF033  times 8 dw -F_0_168, -F_0_331
+PW_MF008_MF041  times 8 dw -F_0_081, -F_0_418
+PD_ONEHALFM1_CJ times 8 dd  (1 << (SCALEBITS - 1)) - 1 + \
+                            (CENTERJSAMPLE << SCALEBITS)
+PD_ONEHALF      times 8 dd  (1 << (SCALEBITS - 1))
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+
+%include "jccolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_rgb_ycc_convert_avx2  jsimd_extrgb_ycc_convert_avx2
+%include "jccolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_rgb_ycc_convert_avx2  jsimd_extrgbx_ycc_convert_avx2
+%include "jccolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_rgb_ycc_convert_avx2  jsimd_extbgr_ycc_convert_avx2
+%include "jccolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_rgb_ycc_convert_avx2  jsimd_extbgrx_ycc_convert_avx2
+%include "jccolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_rgb_ycc_convert_avx2  jsimd_extxbgr_ycc_convert_avx2
+%include "jccolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_rgb_ycc_convert_avx2  jsimd_extxrgb_ycc_convert_avx2
+%include "jccolext-avx2.asm"
@@ -0,0 +1,121 @@
+;
+; jccolor.asm - colorspace conversion (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_081 equ  5329                ; FIX(0.08131)
+F_0_114 equ  7471                ; FIX(0.11400)
+F_0_168 equ 11059                ; FIX(0.16874)
+F_0_250 equ 16384                ; FIX(0.25000)
+F_0_299 equ 19595                ; FIX(0.29900)
+F_0_331 equ 21709                ; FIX(0.33126)
+F_0_418 equ 27439                ; FIX(0.41869)
+F_0_587 equ 38470                ; FIX(0.58700)
+F_0_337 equ (F_0_587 - F_0_250)  ; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_rgb_ycc_convert_mmx)
+
+EXTN(jconst_rgb_ycc_convert_mmx):
+
+PW_F0299_F0337  times 2 dw  F_0_299,  F_0_337
+PW_F0114_F0250  times 2 dw  F_0_114,  F_0_250
+PW_MF016_MF033  times 2 dw -F_0_168, -F_0_331
+PW_MF008_MF041  times 2 dw -F_0_081, -F_0_418
+PD_ONEHALFM1_CJ times 2 dd  (1 << (SCALEBITS - 1)) - 1 + \
+                            (CENTERJSAMPLE << SCALEBITS)
+PD_ONEHALF      times 2 dd  (1 << (SCALEBITS - 1))
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+
+%include "jccolext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_rgb_ycc_convert_mmx  jsimd_extrgb_ycc_convert_mmx
+%include "jccolext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_rgb_ycc_convert_mmx  jsimd_extrgbx_ycc_convert_mmx
+%include "jccolext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_rgb_ycc_convert_mmx  jsimd_extbgr_ycc_convert_mmx
+%include "jccolext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_rgb_ycc_convert_mmx  jsimd_extbgrx_ycc_convert_mmx
+%include "jccolext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_rgb_ycc_convert_mmx  jsimd_extxbgr_ycc_convert_mmx
+%include "jccolext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_rgb_ycc_convert_mmx  jsimd_extxrgb_ycc_convert_mmx
+%include "jccolext-mmx.asm"
@@ -0,0 +1,120 @@
+;
+; jccolor.asm - colorspace conversion (SSE2)
+;
+; Copyright (C) 2009, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_081 equ  5329                ; FIX(0.08131)
+F_0_114 equ  7471                ; FIX(0.11400)
+F_0_168 equ 11059                ; FIX(0.16874)
+F_0_250 equ 16384                ; FIX(0.25000)
+F_0_299 equ 19595                ; FIX(0.29900)
+F_0_331 equ 21709                ; FIX(0.33126)
+F_0_418 equ 27439                ; FIX(0.41869)
+F_0_587 equ 38470                ; FIX(0.58700)
+F_0_337 equ (F_0_587 - F_0_250)  ; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_rgb_ycc_convert_sse2)
+
+EXTN(jconst_rgb_ycc_convert_sse2):
+
+PW_F0299_F0337  times 4 dw  F_0_299,  F_0_337
+PW_F0114_F0250  times 4 dw  F_0_114,  F_0_250
+PW_MF016_MF033  times 4 dw -F_0_168, -F_0_331
+PW_MF008_MF041  times 4 dw -F_0_081, -F_0_418
+PD_ONEHALFM1_CJ times 4 dd  (1 << (SCALEBITS - 1)) - 1 + \
+                            (CENTERJSAMPLE << SCALEBITS)
+PD_ONEHALF      times 4 dd  (1 << (SCALEBITS - 1))
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+
+%include "jccolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_rgb_ycc_convert_sse2  jsimd_extrgb_ycc_convert_sse2
+%include "jccolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_rgb_ycc_convert_sse2  jsimd_extrgbx_ycc_convert_sse2
+%include "jccolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_rgb_ycc_convert_sse2  jsimd_extbgr_ycc_convert_sse2
+%include "jccolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_rgb_ycc_convert_sse2  jsimd_extbgrx_ycc_convert_sse2
+%include "jccolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_rgb_ycc_convert_sse2  jsimd_extxbgr_ycc_convert_sse2
+%include "jccolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_rgb_ycc_convert_sse2  jsimd_extxrgb_ycc_convert_sse2
+%include "jccolext-sse2.asm"
@@ -0,0 +1,113 @@
+;
+; jcgray.asm - grayscale colorspace conversion (AVX2)
+;
+; Copyright (C) 2011, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_114 equ  7471                ; FIX(0.11400)
+F_0_250 equ 16384                ; FIX(0.25000)
+F_0_299 equ 19595                ; FIX(0.29900)
+F_0_587 equ 38470                ; FIX(0.58700)
+F_0_337 equ (F_0_587 - F_0_250)  ; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_rgb_gray_convert_avx2)
+
+EXTN(jconst_rgb_gray_convert_avx2):
+
+PW_F0299_F0337 times 8 dw F_0_299, F_0_337
+PW_F0114_F0250 times 8 dw F_0_114, F_0_250
+PD_ONEHALF     times 8 dd (1 << (SCALEBITS - 1))
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+
+%include "jcgryext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_rgb_gray_convert_avx2  jsimd_extrgb_gray_convert_avx2
+%include "jcgryext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_rgb_gray_convert_avx2  jsimd_extrgbx_gray_convert_avx2
+%include "jcgryext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_rgb_gray_convert_avx2  jsimd_extbgr_gray_convert_avx2
+%include "jcgryext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_rgb_gray_convert_avx2  jsimd_extbgrx_gray_convert_avx2
+%include "jcgryext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_rgb_gray_convert_avx2  jsimd_extxbgr_gray_convert_avx2
+%include "jcgryext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_rgb_gray_convert_avx2  jsimd_extxrgb_gray_convert_avx2
+%include "jcgryext-avx2.asm"
@@ -0,0 +1,113 @@
+;
+; jcgray.asm - grayscale colorspace conversion (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2011, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_114 equ  7471                ; FIX(0.11400)
+F_0_250 equ 16384                ; FIX(0.25000)
+F_0_299 equ 19595                ; FIX(0.29900)
+F_0_587 equ 38470                ; FIX(0.58700)
+F_0_337 equ (F_0_587 - F_0_250)  ; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_rgb_gray_convert_mmx)
+
+EXTN(jconst_rgb_gray_convert_mmx):
+
+PW_F0299_F0337 times 2 dw F_0_299, F_0_337
+PW_F0114_F0250 times 2 dw F_0_114, F_0_250
+PD_ONEHALF     times 2 dd (1 << (SCALEBITS - 1))
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+
+%include "jcgryext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_rgb_gray_convert_mmx  jsimd_extrgb_gray_convert_mmx
+%include "jcgryext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_rgb_gray_convert_mmx  jsimd_extrgbx_gray_convert_mmx
+%include "jcgryext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_rgb_gray_convert_mmx  jsimd_extbgr_gray_convert_mmx
+%include "jcgryext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_rgb_gray_convert_mmx  jsimd_extbgrx_gray_convert_mmx
+%include "jcgryext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_rgb_gray_convert_mmx  jsimd_extxbgr_gray_convert_mmx
+%include "jcgryext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_rgb_gray_convert_mmx  jsimd_extxrgb_gray_convert_mmx
+%include "jcgryext-mmx.asm"
@@ -0,0 +1,112 @@
+;
+; jcgray.asm - grayscale colorspace conversion (SSE2)
+;
+; Copyright (C) 2011, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_114 equ  7471                ; FIX(0.11400)
+F_0_250 equ 16384                ; FIX(0.25000)
+F_0_299 equ 19595                ; FIX(0.29900)
+F_0_587 equ 38470                ; FIX(0.58700)
+F_0_337 equ (F_0_587 - F_0_250)  ; FIX(0.58700) - FIX(0.25000)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_rgb_gray_convert_sse2)
+
+EXTN(jconst_rgb_gray_convert_sse2):
+
+PW_F0299_F0337 times 4 dw F_0_299, F_0_337
+PW_F0114_F0250 times 4 dw F_0_114, F_0_250
+PD_ONEHALF     times 4 dd (1 << (SCALEBITS - 1))
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+
+%include "jcgryext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_rgb_gray_convert_sse2  jsimd_extrgb_gray_convert_sse2
+%include "jcgryext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_rgb_gray_convert_sse2  jsimd_extrgbx_gray_convert_sse2
+%include "jcgryext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_rgb_gray_convert_sse2  jsimd_extbgr_gray_convert_sse2
+%include "jcgryext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_rgb_gray_convert_sse2  jsimd_extbgrx_gray_convert_sse2
+%include "jcgryext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_rgb_gray_convert_sse2  jsimd_extxbgr_gray_convert_sse2
+%include "jcgryext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_rgb_gray_convert_sse2  jsimd_extxrgb_gray_convert_sse2
+%include "jcgryext-sse2.asm"
@@ -0,0 +1,457 @@
+;
+; jcgryext.asm - grayscale colorspace conversion (AVX2)
+;
+; Copyright (C) 2011, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_gray_convert_avx2(JDIMENSION img_width, JSAMPARRAY input_buf,
+;                             JSAMPIMAGE output_buf, JDIMENSION output_row,
+;                             int num_rows);
+;
+
+%define img_width(b)   (b) + 8          ; JDIMENSION img_width
+%define input_buf(b)   (b) + 12         ; JSAMPARRAY input_buf
+%define output_buf(b)  (b) + 16         ; JSAMPIMAGE output_buf
+%define output_row(b)  (b) + 20         ; JDIMENSION output_row
+%define num_rows(b)    (b) + 24         ; int num_rows
+
+%define original_ebp   ebp + 0
+%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_YMMWORD
+                                        ; ymmword wk[WK_NUM]
+%define WK_NUM         2
+%define gotptr         wk(0) - SIZEOF_POINTER  ; void * gotptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_rgb_gray_convert_avx2)
+
+EXTN(jsimd_rgb_gray_convert_avx2):
+    push        ebp
+    mov         eax, esp                     ; eax = original ebp
+    sub         esp, byte 4
+    and         esp, byte (-SIZEOF_YMMWORD)  ; align to 256 bits
+    mov         [esp], eax
+    mov         ebp, esp                     ; ebp = aligned ebp
+    lea         esp, [wk(0)]
+    pushpic     eax                     ; make a room for GOT address
+    push        ebx
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    get_GOT     ebx                     ; get GOT address
+    movpic      POINTER [gotptr], ebx   ; save GOT address
+
+    mov         ecx, JDIMENSION [img_width(eax)]
+    test        ecx, ecx
+    jz          near .return
+
+    push        ecx
+
+    mov         esi, JSAMPIMAGE [output_buf(eax)]
+    mov         ecx, JDIMENSION [output_row(eax)]
+    mov         edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
+    lea         edi, [edi+ecx*SIZEOF_JSAMPROW]
+
+    pop         ecx
+
+    mov         esi, JSAMPARRAY [input_buf(eax)]
+    mov         eax, INT [num_rows(eax)]
+    test        eax, eax
+    jle         near .return
+    alignx      16, 7
+.rowloop:
+    pushpic     eax
+    push        edi
+    push        esi
+    push        ecx                     ; col
+
+    mov         esi, JSAMPROW [esi]     ; inptr
+    mov         edi, JSAMPROW [edi]     ; outptr0
+    movpic      eax, POINTER [gotptr]   ; load GOT address (eax)
+
+    cmp         ecx, byte SIZEOF_YMMWORD
+    jae         near .columnloop
+    alignx      16, 7
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+.column_ld1:
+    push        eax
+    push        edx
+    lea         ecx, [ecx+ecx*2]        ; imul ecx,RGB_PIXELSIZE
+    test        cl, SIZEOF_BYTE
+    jz          short .column_ld2
+    sub         ecx, byte SIZEOF_BYTE
+    movzx       eax, byte [esi+ecx]
+.column_ld2:
+    test        cl, SIZEOF_WORD
+    jz          short .column_ld4
+    sub         ecx, byte SIZEOF_WORD
+    movzx       edx, word [esi+ecx]
+    shl         eax, WORD_BIT
+    or          eax, edx
+.column_ld4:
+    vmovd       xmmA, eax
+    pop         edx
+    pop         eax
+    test        cl, SIZEOF_DWORD
+    jz          short .column_ld8
+    sub         ecx, byte SIZEOF_DWORD
+    vmovd       xmmF, XMM_DWORD [esi+ecx]
+    vpslldq     xmmA, xmmA, SIZEOF_DWORD
+    vpor        xmmA, xmmA, xmmF
+.column_ld8:
+    test        cl, SIZEOF_MMWORD
+    jz          short .column_ld16
+    sub         ecx, byte SIZEOF_MMWORD
+    vmovq       xmmB, XMM_MMWORD [esi+ecx]
+    vpslldq     xmmA, xmmA, SIZEOF_MMWORD
+    vpor        xmmA, xmmA, xmmB
+.column_ld16:
+    test        cl, SIZEOF_XMMWORD
+    jz          short .column_ld32
+    sub         ecx, byte SIZEOF_XMMWORD
+    vmovdqu     xmmB, XMM_MMWORD [esi+ecx]
+    vperm2i128  ymmA, ymmA, ymmA, 1
+    vpor        ymmA, ymmB
+.column_ld32:
+    test        cl, SIZEOF_YMMWORD
+    jz          short .column_ld64
+    sub         ecx, byte SIZEOF_YMMWORD
+    vmovdqa     ymmF, ymmA
+    vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
+.column_ld64:
+    test        cl, 2*SIZEOF_YMMWORD
+    mov         ecx, SIZEOF_YMMWORD
+    jz          short .rgb_gray_cnv
+    vmovdqa     ymmB, ymmA
+    vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD]
+    jmp         short .rgb_gray_cnv
+    alignx      16, 7
+
+.columnloop:
+    vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD]
+    vmovdqu     ymmB, YMMWORD [esi+2*SIZEOF_YMMWORD]
+
+.rgb_gray_cnv:
+    ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+    ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+    ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    ; ymmB=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
+    ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+
+    vmovdqu     ymmC, ymmA
+    vinserti128 ymmA, ymmF, xmmA, 0  ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+                                     ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    vinserti128 ymmC, ymmC, xmmB, 0  ; ymmC=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
+                                     ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    vinserti128 ymmB, ymmB, xmmF, 0  ; ymmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+                                     ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+    vperm2i128  ymmF, ymmC, ymmC, 1  ; ymmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A
+                                     ;       1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q)
+
+    vmovdqa     ymmG, ymmA
+    vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12
+                                  ;       22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I)
+    vpsrldq     ymmG, ymmG, 8     ; ymmG=(22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I
+                                  ;       2I 0J 1J 2J 0K 1K 2K 0L -- -- -- -- -- -- -- --)
+
+    vpunpckhbw  ymmA, ymmA, ymmF  ; ymmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A
+                                  ;       0G 0O 1G 1O 2G 2O 0H 0P 1H 1P 2H 2P 0I 0Q 1I 1Q)
+    vpslldq     ymmF, ymmF, 8     ; ymmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27
+                                  ;       08 18 28 09 19 29 0A 1A 1L 2L 0M 1M 2M 0N 1N 2N)
+
+    vpunpcklbw  ymmG, ymmG, ymmB  ; ymmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D
+                                  ;       2I 2Q 0J 0R 1J 1R 2J 2R 0K 0S 1K 1S 2K 2S 0L 0T)
+    vpunpckhbw  ymmF, ymmF, ymmB  ; ymmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F
+                                  ;       1L 1T 2L 2T 0M 0U 1M 1U 2M 2U 0N 0V 1N 1V 2N 2V)
+
+    vmovdqa     ymmD, ymmA
+    vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09
+                                  ;       11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P)
+    vpsrldq     ymmD, ymmD, 8     ; ymmD=(11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P
+                                  ;       1H 1P 2H 2P 0I 0Q 1I 1Q -- -- -- -- -- -- -- --)
+
+    vpunpckhbw  ymmA, ymmA, ymmG  ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D
+                                  ;       0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 0H 0L 0P 0T)
+    vpslldq     ymmG, ymmG, 8     ; ymmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B
+                                  ;       04 0C 14 1C 24 2C 05 0D 2I 2Q 0J 0R 1J 1R 2J 2R)
+
+    vpunpcklbw  ymmD, ymmD, ymmF  ; ymmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E
+                                  ;       1H 1L 1P 1T 2H 2L 2P 2T 0I 0M 0Q 0U 1I 1M 1Q 1U)
+    vpunpckhbw  ymmG, ymmG, ymmF  ; ymmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F
+                                  ;       2I 2M 2Q 2U 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V)
+
+    vmovdqa     ymmE, ymmA
+    vpslldq     ymmA, ymmA, 8     ; ymmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C
+                                  ;       20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S)
+    vpsrldq     ymmE, ymmE, 8     ; ymmE=(20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S
+                                  ;       2G 2K 2O 2S 0H 0L 0P 0T -- -- -- -- -- -- -- --)
+
+    vpunpckhbw  ymmA, ymmA, ymmD  ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E
+                                  ;       0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U)
+    vpslldq     ymmD, ymmD, 8     ; ymmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D
+                                  ;       02 06 0A 0E 12 16 1A 1E 1H 1L 1P 1T 2H 2L 2P 2T)
+
+    vpunpcklbw  ymmE, ymmE, ymmG  ; ymmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F
+                                  ;       2G 2I 2K 2M 2O 2Q 2S 2U 0H 0J 0L 0N 0P 0R 0T 0V)
+    vpunpckhbw  ymmD, ymmD, ymmG  ; ymmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F
+                                  ;       1H 1J 1L 1N 1P 1R 1T 1V 2H 2J 2L 2N 2P 2R 2T 2V)
+
+    vpxor       ymmH, ymmH, ymmH
+
+    vmovdqa     ymmC, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmH  ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U)
+    vpunpckhbw  ymmC, ymmC, ymmH  ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U)
+
+    vmovdqa     ymmB, ymmE
+    vpunpcklbw  ymmE, ymmE, ymmH  ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U)
+    vpunpckhbw  ymmB, ymmB, ymmH  ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V)
+
+    vmovdqa     ymmF, ymmD
+    vpunpcklbw  ymmD, ymmD, ymmH  ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V)
+    vpunpckhbw  ymmF, ymmF, ymmH  ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V)
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+    test        cl, SIZEOF_XMMWORD/16
+    jz          short .column_ld2
+    sub         ecx, byte SIZEOF_XMMWORD/16
+    vmovd       xmmA, XMM_DWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld2:
+    test        cl, SIZEOF_XMMWORD/8
+    jz          short .column_ld4
+    sub         ecx, byte SIZEOF_XMMWORD/8
+    vmovq       xmmF, XMM_MMWORD [esi+ecx*RGB_PIXELSIZE]
+    vpslldq     xmmA, xmmA, SIZEOF_MMWORD
+    vpor        xmmA, xmmA, xmmF
+.column_ld4:
+    test        cl, SIZEOF_XMMWORD/4
+    jz          short .column_ld8
+    sub         ecx, byte SIZEOF_XMMWORD/4
+    vmovdqa     xmmF, xmmA
+    vperm2i128  ymmF, ymmF, ymmF, 1
+    vmovdqu     xmmA, XMMWORD [esi+ecx*RGB_PIXELSIZE]
+    vpor        ymmA, ymmA, ymmF
+.column_ld8:
+    test        cl, SIZEOF_XMMWORD/2
+    jz          short .column_ld16
+    sub         ecx, byte SIZEOF_XMMWORD/2
+    vmovdqa     ymmF, ymmA
+    vmovdqu     ymmA, YMMWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld16:
+    test        cl, SIZEOF_XMMWORD
+    mov         ecx, SIZEOF_YMMWORD
+    jz          short .rgb_gray_cnv
+    vmovdqa     ymmE, ymmA
+    vmovdqa     ymmH, ymmF
+    vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD]
+    jmp         short .rgb_gray_cnv
+    alignx      16, 7
+
+.columnloop:
+    vmovdqu     ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD]
+    vmovdqu     ymmE, YMMWORD [esi+2*SIZEOF_YMMWORD]
+    vmovdqu     ymmH, YMMWORD [esi+3*SIZEOF_YMMWORD]
+
+.rgb_gray_cnv:
+    ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+    ;       04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+    ;       0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+    ; ymmE=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J
+    ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+    ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R
+    ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+
+    vmovdqa     ymmB, ymmA
+    vinserti128 ymmA, ymmA, xmmE, 1     ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+                                        ;       0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J)
+    vperm2i128  ymmE, ymmB, ymmE, 0x31  ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+                                        ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+
+    vmovdqa     ymmB, ymmF
+    vinserti128 ymmF, ymmF, xmmH, 1     ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+                                        ;       0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R)
+    vperm2i128  ymmH, ymmB, ymmH, 0x31  ; ymmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F
+                                        ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+
+    vmovdqa     ymmD, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmE      ; ymmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35
+                                      ;       0G 0K 1G 1K 2G 2K 3G 3K 0H 0L 1H 1L 2H 2L 3H 3L)
+    vpunpckhbw  ymmD, ymmD, ymmE      ; ymmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37
+                                      ;       0I 0M 1I 1M 2I 2M 3I 3M 0J 0N 1J 1N 2J 2N 3J 3N)
+
+    vmovdqa     ymmC, ymmF
+    vpunpcklbw  ymmF, ymmF, ymmH      ; ymmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D
+                                      ;       0O 0S 1O 1S 2O 2S 3O 3S 0P 0T 1P 1T 2P 2T 3P 3T)
+    vpunpckhbw  ymmC, ymmC, ymmH      ; ymmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F
+                                      ;       0Q 0U 1Q 1U 2Q 2U 3Q 3U 0R 0V 1R 1V 2R 2V 3R 3V)
+
+    vmovdqa     ymmB, ymmA
+    vpunpcklwd  ymmA, ymmA, ymmF      ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C
+                                      ;       0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 3G 3K 3O 3S)
+    vpunpckhwd  ymmB, ymmB, ymmF      ; ymmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D
+                                      ;       0H 0L 0P 0T 1H 1L 1P 1T 2H 2L 2P 2T 3H 3L 3P 3T)
+
+    vmovdqa     ymmG, ymmD
+    vpunpcklwd  ymmD, ymmD, ymmC      ; ymmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E
+                                      ;       0I 0M 0Q 0U 1I 1M 1Q 1U 2I 2M 2Q 2U 3I 3M 3Q 3U)
+    vpunpckhwd  ymmG, ymmG, ymmC      ; ymmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F
+                                      ;       0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V 3J 3N 3R 3V)
+
+    vmovdqa     ymmE, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmD      ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E
+                                      ;       0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U)
+    vpunpckhbw  ymmE, ymmE, ymmD      ; ymmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E
+                                      ;       2G 2I 2K 2M 2O 2Q 2S 2U 3G 3I 3K 3M 3O 3Q 3S 3U)
+
+    vmovdqa     ymmH, ymmB
+    vpunpcklbw  ymmB, ymmB, ymmG      ; ymmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F
+                                      ;       0H 0J 0L 0N 0P 0R 0T 0V 1H 1J 1L 1N 1P 1R 1T 1V)
+    vpunpckhbw  ymmH, ymmH, ymmG      ; ymmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F
+                                      ;       2H 2J 2L 2N 2P 2R 2T 2V 3H 3J 3L 3N 3P 3R 3T 3V)
+
+    vpxor       ymmF, ymmF, ymmF
+
+    vmovdqa     ymmC, ymmA
+    vpunpcklbw  ymmA, ymmA, ymmF      ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U)
+    vpunpckhbw  ymmC, ymmC, ymmF      ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U)
+
+    vmovdqa     ymmD, ymmB
+    vpunpcklbw  ymmB, ymmB, ymmF      ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V)
+    vpunpckhbw  ymmD, ymmD, ymmF      ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V)
+
+    vmovdqa     ymmG, ymmE
+    vpunpcklbw  ymmE, ymmE, ymmF      ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U)
+    vpunpckhbw  ymmG, ymmG, ymmF      ; ymmG=(30 32 34 36 38 3A 3C 3E 3G 3I 3K 3M 3O 3Q 3S 3U)
+
+    vpunpcklbw  ymmF, ymmF, ymmH
+    vpunpckhbw  ymmH, ymmH, ymmH
+    vpsrlw      ymmF, ymmF, BYTE_BIT  ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V)
+    vpsrlw      ymmH, ymmH, BYTE_BIT  ; ymmH=(31 33 35 37 39 3B 3D 3F 3H 3J 3L 3N 3P 3R 3T 3V)
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    ; ymm0=R(02468ACEGIKMOQSU)=RE, ymm2=G(02468ACEGIKMOQSU)=GE, ymm4=B(02468ACEGIKMOQSU)=BE
+    ; ymm1=R(13579BDFHJLNPRTV)=RO, ymm3=G(13579BDFHJLNPRTV)=GO, ymm5=B(13579BDFHJLNPRTV)=BO
+
+    ; (Original)
+    ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+    ;
+    ; (This implementation)
+    ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+
+    vmovdqa     ymm6, ymm1
+    vpunpcklwd  ymm1, ymm1, ymm3
+    vpunpckhwd  ymm6, ymm6, ymm3
+    vpmaddwd    ymm1, ymm1, [GOTOFF(eax,PW_F0299_F0337)]  ; ymm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+    vpmaddwd    ymm6, ymm6, [GOTOFF(eax,PW_F0299_F0337)]  ; ymm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    vmovdqa     ymm7, ymm6              ; ymm7=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    vmovdqa     ymm6, ymm0
+    vpunpcklwd  ymm0, ymm0, ymm2
+    vpunpckhwd  ymm6, ymm6, ymm2
+    vpmaddwd    ymm0, ymm0, [GOTOFF(eax,PW_F0299_F0337)]  ; ymm0=REL*FIX(0.299)+GEL*FIX(0.337)
+    vpmaddwd    ymm6, ymm6, [GOTOFF(eax,PW_F0299_F0337)]  ; ymm6=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    vmovdqa     YMMWORD [wk(0)], ymm0   ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337)
+    vmovdqa     YMMWORD [wk(1)], ymm6   ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    vmovdqa     ymm0, ymm5              ; ymm0=BO
+    vmovdqa     ymm6, ymm4              ; ymm6=BE
+
+    vmovdqa     ymm4, ymm0
+    vpunpcklwd  ymm0, ymm0, ymm3
+    vpunpckhwd  ymm4, ymm4, ymm3
+    vpmaddwd    ymm0, ymm0, [GOTOFF(eax,PW_F0114_F0250)]  ; ymm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+    vpmaddwd    ymm4, ymm4, [GOTOFF(eax,PW_F0114_F0250)]  ; ymm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+
+    vmovdqa     ymm3, [GOTOFF(eax,PD_ONEHALF)]            ; ymm3=[PD_ONEHALF]
+
+    vpaddd      ymm0, ymm0, ymm1
+    vpaddd      ymm4, ymm4, ymm7
+    vpaddd      ymm0, ymm0, ymm3
+    vpaddd      ymm4, ymm4, ymm3
+    vpsrld      ymm0, ymm0, SCALEBITS   ; ymm0=YOL
+    vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=YOH
+    vpackssdw   ymm0, ymm0, ymm4        ; ymm0=YO
+
+    vmovdqa     ymm4, ymm6
+    vpunpcklwd  ymm6, ymm6, ymm2
+    vpunpckhwd  ymm4, ymm4, ymm2
+    vpmaddwd    ymm6, ymm6, [GOTOFF(eax,PW_F0114_F0250)]  ; ymm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+    vpmaddwd    ymm4, ymm4, [GOTOFF(eax,PW_F0114_F0250)]  ; ymm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+
+    vmovdqa     ymm2, [GOTOFF(eax,PD_ONEHALF)]            ; ymm2=[PD_ONEHALF]
+
+    vpaddd      ymm6, ymm6, YMMWORD [wk(0)]
+    vpaddd      ymm4, ymm4, YMMWORD [wk(1)]
+    vpaddd      ymm6, ymm6, ymm2
+    vpaddd      ymm4, ymm4, ymm2
+    vpsrld      ymm6, ymm6, SCALEBITS   ; ymm6=YEL
+    vpsrld      ymm4, ymm4, SCALEBITS   ; ymm4=YEH
+    vpackssdw   ymm6, ymm6, ymm4        ; ymm6=YE
+
+    vpsllw      ymm0, ymm0, BYTE_BIT
+    vpor        ymm6, ymm6, ymm0        ; ymm6=Y
+    vmovdqu     YMMWORD [edi], ymm6     ; Save Y
+
+    sub         ecx, byte SIZEOF_YMMWORD
+    add         esi, RGB_PIXELSIZE*SIZEOF_YMMWORD  ; inptr
+    add         edi, byte SIZEOF_YMMWORD           ; outptr0
+    cmp         ecx, byte SIZEOF_YMMWORD
+    jae         near .columnloop
+    test        ecx, ecx
+    jnz         near .column_ld1
+
+    pop         ecx                     ; col
+    pop         esi
+    pop         edi
+    poppic      eax
+
+    add         esi, byte SIZEOF_JSAMPROW  ; input_buf
+    add         edi, byte SIZEOF_JSAMPROW
+    dec         eax                        ; num_rows
+    jg          near .rowloop
+
+.return:
+    vzeroupper
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+    pop         ebx
+    mov         esp, ebp                ; esp <- aligned ebp
+    pop         esp                     ; esp <- original ebp
+    pop         ebp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,355 @@
+;
+; jcgryext.asm - grayscale colorspace conversion (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2011, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_gray_convert_mmx(JDIMENSION img_width, JSAMPARRAY input_buf,
+;                            JSAMPIMAGE output_buf, JDIMENSION output_row,
+;                            int num_rows);
+;
+
+%define img_width(b)   (b) + 8          ; JDIMENSION img_width
+%define input_buf(b)   (b) + 12         ; JSAMPARRAY input_buf
+%define output_buf(b)  (b) + 16         ; JSAMPIMAGE output_buf
+%define output_row(b)  (b) + 20         ; JDIMENSION output_row
+%define num_rows(b)    (b) + 24         ; int num_rows
+
+%define original_ebp   ebp + 0
+%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_MMWORD
+                                        ; mmword wk[WK_NUM]
+%define WK_NUM         2
+%define gotptr         wk(0) - SIZEOF_POINTER  ; void * gotptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_rgb_gray_convert_mmx)
+
+EXTN(jsimd_rgb_gray_convert_mmx):
+    push        ebp
+    mov         eax, esp                    ; eax = original ebp
+    sub         esp, byte 4
+    and         esp, byte (-SIZEOF_MMWORD)  ; align to 64 bits
+    mov         [esp], eax
+    mov         ebp, esp                    ; ebp = aligned ebp
+    lea         esp, [wk(0)]
+    pushpic     eax                     ; make a room for GOT address
+    push        ebx
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    get_GOT     ebx                     ; get GOT address
+    movpic      POINTER [gotptr], ebx   ; save GOT address
+
+    mov         ecx, JDIMENSION [img_width(eax)]  ; num_cols
+    test        ecx, ecx
+    jz          near .return
+
+    push        ecx
+
+    mov         esi, JSAMPIMAGE [output_buf(eax)]
+    mov         ecx, JDIMENSION [output_row(eax)]
+    mov         edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
+    lea         edi, [edi+ecx*SIZEOF_JSAMPROW]
+
+    pop         ecx
+
+    mov         esi, JSAMPARRAY [input_buf(eax)]
+    mov         eax, INT [num_rows(eax)]
+    test        eax, eax
+    jle         near .return
+    alignx      16, 7
+.rowloop:
+    pushpic     eax
+    push        edi
+    push        esi
+    push        ecx                     ; col
+
+    mov         esi, JSAMPROW [esi]     ; inptr
+    mov         edi, JSAMPROW [edi]     ; outptr0
+    movpic      eax, POINTER [gotptr]   ; load GOT address (eax)
+
+    cmp         ecx, byte SIZEOF_MMWORD
+    jae         short .columnloop
+    alignx      16, 7
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+.column_ld1:
+    push        eax
+    push        edx
+    lea         ecx, [ecx+ecx*2]        ; imul ecx,RGB_PIXELSIZE
+    test        cl, SIZEOF_BYTE
+    jz          short .column_ld2
+    sub         ecx, byte SIZEOF_BYTE
+    xor         eax, eax
+    mov         al, byte [esi+ecx]
+.column_ld2:
+    test        cl, SIZEOF_WORD
+    jz          short .column_ld4
+    sub         ecx, byte SIZEOF_WORD
+    xor         edx, edx
+    mov         dx, word [esi+ecx]
+    shl         eax, WORD_BIT
+    or          eax, edx
+.column_ld4:
+    movd        mmA, eax
+    pop         edx
+    pop         eax
+    test        cl, SIZEOF_DWORD
+    jz          short .column_ld8
+    sub         ecx, byte SIZEOF_DWORD
+    movd        mmG, dword [esi+ecx]
+    psllq       mmA, DWORD_BIT
+    por         mmA, mmG
+.column_ld8:
+    test        cl, SIZEOF_MMWORD
+    jz          short .column_ld16
+    movq        mmG, mmA
+    movq        mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+    mov         ecx, SIZEOF_MMWORD
+    jmp         short .rgb_gray_cnv
+.column_ld16:
+    test        cl, 2*SIZEOF_MMWORD
+    mov         ecx, SIZEOF_MMWORD
+    jz          short .rgb_gray_cnv
+    movq        mmF, mmA
+    movq        mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+    movq        mmG, MMWORD [esi+1*SIZEOF_MMWORD]
+    jmp         short .rgb_gray_cnv
+    alignx      16, 7
+
+.columnloop:
+    movq        mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+    movq        mmG, MMWORD [esi+1*SIZEOF_MMWORD]
+    movq        mmF, MMWORD [esi+2*SIZEOF_MMWORD]
+
+.rgb_gray_cnv:
+    ; mmA=(00 10 20 01 11 21 02 12)
+    ; mmG=(22 03 13 23 04 14 24 05)
+    ; mmF=(15 25 06 16 26 07 17 27)
+
+    movq        mmD, mmA
+    psllq       mmA, 4*BYTE_BIT         ; mmA=(-- -- -- -- 00 10 20 01)
+    psrlq       mmD, 4*BYTE_BIT         ; mmD=(11 21 02 12 -- -- -- --)
+
+    punpckhbw   mmA, mmG                ; mmA=(00 04 10 14 20 24 01 05)
+    psllq       mmG, 4*BYTE_BIT         ; mmG=(-- -- -- -- 22 03 13 23)
+
+    punpcklbw   mmD, mmF                ; mmD=(11 15 21 25 02 06 12 16)
+    punpckhbw   mmG, mmF                ; mmG=(22 26 03 07 13 17 23 27)
+
+    movq        mmE, mmA
+    psllq       mmA, 4*BYTE_BIT         ; mmA=(-- -- -- -- 00 04 10 14)
+    psrlq       mmE, 4*BYTE_BIT         ; mmE=(20 24 01 05 -- -- -- --)
+
+    punpckhbw   mmA, mmD                ; mmA=(00 02 04 06 10 12 14 16)
+    psllq       mmD, 4*BYTE_BIT         ; mmD=(-- -- -- -- 11 15 21 25)
+
+    punpcklbw   mmE, mmG                ; mmE=(20 22 24 26 01 03 05 07)
+    punpckhbw   mmD, mmG                ; mmD=(11 13 15 17 21 23 25 27)
+
+    pxor        mmH, mmH
+
+    movq        mmC, mmA
+    punpcklbw   mmA, mmH                ; mmA=(00 02 04 06)
+    punpckhbw   mmC, mmH                ; mmC=(10 12 14 16)
+
+    movq        mmB, mmE
+    punpcklbw   mmE, mmH                ; mmE=(20 22 24 26)
+    punpckhbw   mmB, mmH                ; mmB=(01 03 05 07)
+
+    movq        mmF, mmD
+    punpcklbw   mmD, mmH                ; mmD=(11 13 15 17)
+    punpckhbw   mmF, mmH                ; mmF=(21 23 25 27)
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+    test        cl, SIZEOF_MMWORD/8
+    jz          short .column_ld2
+    sub         ecx, byte SIZEOF_MMWORD/8
+    movd        mmA, dword [esi+ecx*RGB_PIXELSIZE]
+.column_ld2:
+    test        cl, SIZEOF_MMWORD/4
+    jz          short .column_ld4
+    sub         ecx, byte SIZEOF_MMWORD/4
+    movq        mmF, mmA
+    movq        mmA, MMWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld4:
+    test        cl, SIZEOF_MMWORD/2
+    mov         ecx, SIZEOF_MMWORD
+    jz          short .rgb_gray_cnv
+    movq        mmD, mmA
+    movq        mmC, mmF
+    movq        mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+    movq        mmF, MMWORD [esi+1*SIZEOF_MMWORD]
+    jmp         short .rgb_gray_cnv
+    alignx      16, 7
+
+.columnloop:
+    movq        mmA, MMWORD [esi+0*SIZEOF_MMWORD]
+    movq        mmF, MMWORD [esi+1*SIZEOF_MMWORD]
+    movq        mmD, MMWORD [esi+2*SIZEOF_MMWORD]
+    movq        mmC, MMWORD [esi+3*SIZEOF_MMWORD]
+
+.rgb_gray_cnv:
+    ; mmA=(00 10 20 30 01 11 21 31)
+    ; mmF=(02 12 22 32 03 13 23 33)
+    ; mmD=(04 14 24 34 05 15 25 35)
+    ; mmC=(06 16 26 36 07 17 27 37)
+
+    movq        mmB, mmA
+    punpcklbw   mmA, mmF                ; mmA=(00 02 10 12 20 22 30 32)
+    punpckhbw   mmB, mmF                ; mmB=(01 03 11 13 21 23 31 33)
+
+    movq        mmG, mmD
+    punpcklbw   mmD, mmC                ; mmD=(04 06 14 16 24 26 34 36)
+    punpckhbw   mmG, mmC                ; mmG=(05 07 15 17 25 27 35 37)
+
+    movq        mmE, mmA
+    punpcklwd   mmA, mmD                ; mmA=(00 02 04 06 10 12 14 16)
+    punpckhwd   mmE, mmD                ; mmE=(20 22 24 26 30 32 34 36)
+
+    movq        mmH, mmB
+    punpcklwd   mmB, mmG                ; mmB=(01 03 05 07 11 13 15 17)
+    punpckhwd   mmH, mmG                ; mmH=(21 23 25 27 31 33 35 37)
+
+    pxor        mmF, mmF
+
+    movq        mmC, mmA
+    punpcklbw   mmA, mmF                ; mmA=(00 02 04 06)
+    punpckhbw   mmC, mmF                ; mmC=(10 12 14 16)
+
+    movq        mmD, mmB
+    punpcklbw   mmB, mmF                ; mmB=(01 03 05 07)
+    punpckhbw   mmD, mmF                ; mmD=(11 13 15 17)
+
+    movq        mmG, mmE
+    punpcklbw   mmE, mmF                ; mmE=(20 22 24 26)
+    punpckhbw   mmG, mmF                ; mmG=(30 32 34 36)
+
+    punpcklbw   mmF, mmH
+    punpckhbw   mmH, mmH
+    psrlw       mmF, BYTE_BIT           ; mmF=(21 23 25 27)
+    psrlw       mmH, BYTE_BIT           ; mmH=(31 33 35 37)
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    ; mm0=(R0 R2 R4 R6)=RE, mm2=(G0 G2 G4 G6)=GE, mm4=(B0 B2 B4 B6)=BE
+    ; mm1=(R1 R3 R5 R7)=RO, mm3=(G1 G3 G5 G7)=GO, mm5=(B1 B3 B5 B7)=BO
+
+    ; (Original)
+    ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+    ;
+    ; (This implementation)
+    ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+
+    movq        mm6, mm1
+    punpcklwd   mm1, mm3
+    punpckhwd   mm6, mm3
+    pmaddwd     mm1, [GOTOFF(eax,PW_F0299_F0337)]  ; mm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+    pmaddwd     mm6, [GOTOFF(eax,PW_F0299_F0337)]  ; mm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    movq        mm7,  mm6               ; mm7=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    movq        mm6, mm0
+    punpcklwd   mm0, mm2
+    punpckhwd   mm6, mm2
+    pmaddwd     mm0, [GOTOFF(eax,PW_F0299_F0337)]  ; mm0=REL*FIX(0.299)+GEL*FIX(0.337)
+    pmaddwd     mm6, [GOTOFF(eax,PW_F0299_F0337)]  ; mm6=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    movq        MMWORD [wk(0)], mm0     ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337)
+    movq        MMWORD [wk(1)], mm6     ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    movq        mm0, mm5                ; mm0=BO
+    movq        mm6, mm4                ; mm6=BE
+
+    movq        mm4, mm0
+    punpcklwd   mm0, mm3
+    punpckhwd   mm4, mm3
+    pmaddwd     mm0, [GOTOFF(eax,PW_F0114_F0250)]  ; mm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+    pmaddwd     mm4, [GOTOFF(eax,PW_F0114_F0250)]  ; mm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+
+    movq        mm3, [GOTOFF(eax,PD_ONEHALF)]  ; mm3=[PD_ONEHALF]
+
+    paddd       mm0, mm1
+    paddd       mm4, mm7
+    paddd       mm0, mm3
+    paddd       mm4, mm3
+    psrld       mm0, SCALEBITS          ; mm0=YOL
+    psrld       mm4, SCALEBITS          ; mm4=YOH
+    packssdw    mm0, mm4                ; mm0=YO
+
+    movq        mm4, mm6
+    punpcklwd   mm6, mm2
+    punpckhwd   mm4, mm2
+    pmaddwd     mm6, [GOTOFF(eax,PW_F0114_F0250)]  ; mm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+    pmaddwd     mm4, [GOTOFF(eax,PW_F0114_F0250)]  ; mm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+
+    movq        mm2, [GOTOFF(eax,PD_ONEHALF)]      ; mm2=[PD_ONEHALF]
+
+    paddd       mm6, MMWORD [wk(0)]
+    paddd       mm4, MMWORD [wk(1)]
+    paddd       mm6, mm2
+    paddd       mm4, mm2
+    psrld       mm6, SCALEBITS          ; mm6=YEL
+    psrld       mm4, SCALEBITS          ; mm4=YEH
+    packssdw    mm6, mm4                ; mm6=YE
+
+    psllw       mm0, BYTE_BIT
+    por         mm6, mm0                ; mm6=Y
+    movq        MMWORD [edi], mm6       ; Save Y
+
+    sub         ecx, byte SIZEOF_MMWORD
+    add         esi, byte RGB_PIXELSIZE*SIZEOF_MMWORD  ; inptr
+    add         edi, byte SIZEOF_MMWORD                ; outptr0
+    cmp         ecx, byte SIZEOF_MMWORD
+    jae         near .columnloop
+    test        ecx, ecx
+    jnz         near .column_ld1
+
+    pop         ecx                     ; col
+    pop         esi
+    pop         edi
+    poppic      eax
+
+    add         esi, byte SIZEOF_JSAMPROW  ; input_buf
+    add         edi, byte SIZEOF_JSAMPROW
+    dec         eax                        ; num_rows
+    jg          near .rowloop
+
+    emms                                ; empty MMX state
+
+.return:
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+    pop         ebx
+    mov         esp, ebp                ; esp <- aligned ebp
+    pop         esp                     ; esp <- original ebp
+    pop         ebp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,382 @@
+;
+; jcgryext.asm - grayscale colorspace conversion (SSE2)
+;
+; Copyright (C) 2011, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_rgb_gray_convert_sse2(JDIMENSION img_width, JSAMPARRAY input_buf,
+;                             JSAMPIMAGE output_buf, JDIMENSION output_row,
+;                             int num_rows);
+;
+
+%define img_width(b)   (b) + 8          ; JDIMENSION img_width
+%define input_buf(b)   (b) + 12         ; JSAMPARRAY input_buf
+%define output_buf(b)  (b) + 16         ; JSAMPIMAGE output_buf
+%define output_row(b)  (b) + 20         ; JDIMENSION output_row
+%define num_rows(b)    (b) + 24         ; int num_rows
+
+%define original_ebp   ebp + 0
+%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_XMMWORD
+                                        ; xmmword wk[WK_NUM]
+%define WK_NUM         2
+%define gotptr         wk(0) - SIZEOF_POINTER  ; void * gotptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_rgb_gray_convert_sse2)
+
+EXTN(jsimd_rgb_gray_convert_sse2):
+    push        ebp
+    mov         eax, esp                     ; eax = original ebp
+    sub         esp, byte 4
+    and         esp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [esp], eax
+    mov         ebp, esp                     ; ebp = aligned ebp
+    lea         esp, [wk(0)]
+    pushpic     eax                     ; make a room for GOT address
+    push        ebx
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    get_GOT     ebx                     ; get GOT address
+    movpic      POINTER [gotptr], ebx   ; save GOT address
+
+    mov         ecx, JDIMENSION [img_width(eax)]
+    test        ecx, ecx
+    jz          near .return
+
+    push        ecx
+
+    mov         esi, JSAMPIMAGE [output_buf(eax)]
+    mov         ecx, JDIMENSION [output_row(eax)]
+    mov         edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
+    lea         edi, [edi+ecx*SIZEOF_JSAMPROW]
+
+    pop         ecx
+
+    mov         esi, JSAMPARRAY [input_buf(eax)]
+    mov         eax, INT [num_rows(eax)]
+    test        eax, eax
+    jle         near .return
+    alignx      16, 7
+.rowloop:
+    pushpic     eax
+    push        edi
+    push        esi
+    push        ecx                     ; col
+
+    mov         esi, JSAMPROW [esi]     ; inptr
+    mov         edi, JSAMPROW [edi]     ; outptr0
+    movpic      eax, POINTER [gotptr]   ; load GOT address (eax)
+
+    cmp         ecx, byte SIZEOF_XMMWORD
+    jae         near .columnloop
+    alignx      16, 7
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+.column_ld1:
+    push        eax
+    push        edx
+    lea         ecx, [ecx+ecx*2]        ; imul ecx,RGB_PIXELSIZE
+    test        cl, SIZEOF_BYTE
+    jz          short .column_ld2
+    sub         ecx, byte SIZEOF_BYTE
+    movzx       eax, byte [esi+ecx]
+.column_ld2:
+    test        cl, SIZEOF_WORD
+    jz          short .column_ld4
+    sub         ecx, byte SIZEOF_WORD
+    movzx       edx, word [esi+ecx]
+    shl         eax, WORD_BIT
+    or          eax, edx
+.column_ld4:
+    movd        xmmA, eax
+    pop         edx
+    pop         eax
+    test        cl, SIZEOF_DWORD
+    jz          short .column_ld8
+    sub         ecx, byte SIZEOF_DWORD
+    movd        xmmF, XMM_DWORD [esi+ecx]
+    pslldq      xmmA, SIZEOF_DWORD
+    por         xmmA, xmmF
+.column_ld8:
+    test        cl, SIZEOF_MMWORD
+    jz          short .column_ld16
+    sub         ecx, byte SIZEOF_MMWORD
+    movq        xmmB, XMM_MMWORD [esi+ecx]
+    pslldq      xmmA, SIZEOF_MMWORD
+    por         xmmA, xmmB
+.column_ld16:
+    test        cl, SIZEOF_XMMWORD
+    jz          short .column_ld32
+    movdqa      xmmF, xmmA
+    movdqu      xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    mov         ecx, SIZEOF_XMMWORD
+    jmp         short .rgb_gray_cnv
+.column_ld32:
+    test        cl, 2*SIZEOF_XMMWORD
+    mov         ecx, SIZEOF_XMMWORD
+    jz          short .rgb_gray_cnv
+    movdqa      xmmB, xmmA
+    movdqu      xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    movdqu      xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD]
+    jmp         short .rgb_gray_cnv
+    alignx      16, 7
+
+.columnloop:
+    movdqu      xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    movdqu      xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD]
+    movdqu      xmmB, XMMWORD [esi+2*SIZEOF_XMMWORD]
+
+.rgb_gray_cnv:
+    ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+    ; xmmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    ; xmmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+    movdqa      xmmG, xmmA
+    pslldq      xmmA, 8     ; xmmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12)
+    psrldq      xmmG, 8     ; xmmG=(22 03 13 23 04 14 24 05 -- -- -- -- -- -- -- --)
+
+    punpckhbw   xmmA, xmmF  ; xmmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A)
+    pslldq      xmmF, 8     ; xmmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27)
+
+    punpcklbw   xmmG, xmmB  ; xmmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D)
+    punpckhbw   xmmF, xmmB  ; xmmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F)
+
+    movdqa      xmmD, xmmA
+    pslldq      xmmA, 8     ; xmmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09)
+    psrldq      xmmD, 8     ; xmmD=(11 19 21 29 02 0A 12 1A -- -- -- -- -- -- -- --)
+
+    punpckhbw   xmmA, xmmG  ; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D)
+    pslldq      xmmG, 8     ; xmmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B)
+
+    punpcklbw   xmmD, xmmF  ; xmmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E)
+    punpckhbw   xmmG, xmmF  ; xmmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F)
+
+    movdqa      xmmE, xmmA
+    pslldq      xmmA, 8     ; xmmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C)
+    psrldq      xmmE, 8     ; xmmE=(20 24 28 2C 01 05 09 0D -- -- -- -- -- -- -- --)
+
+    punpckhbw   xmmA, xmmD  ; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
+    pslldq      xmmD, 8     ; xmmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D)
+
+    punpcklbw   xmmE, xmmG  ; xmmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F)
+    punpckhbw   xmmD, xmmG  ; xmmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F)
+
+    pxor        xmmH, xmmH
+
+    movdqa      xmmC, xmmA
+    punpcklbw   xmmA, xmmH  ; xmmA=(00 02 04 06 08 0A 0C 0E)
+    punpckhbw   xmmC, xmmH  ; xmmC=(10 12 14 16 18 1A 1C 1E)
+
+    movdqa      xmmB, xmmE
+    punpcklbw   xmmE, xmmH  ; xmmE=(20 22 24 26 28 2A 2C 2E)
+    punpckhbw   xmmB, xmmH  ; xmmB=(01 03 05 07 09 0B 0D 0F)
+
+    movdqa      xmmF, xmmD
+    punpcklbw   xmmD, xmmH  ; xmmD=(11 13 15 17 19 1B 1D 1F)
+    punpckhbw   xmmF, xmmH  ; xmmF=(21 23 25 27 29 2B 2D 2F)
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+.column_ld1:
+    test        cl, SIZEOF_XMMWORD/16
+    jz          short .column_ld2
+    sub         ecx, byte SIZEOF_XMMWORD/16
+    movd        xmmA, XMM_DWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld2:
+    test        cl, SIZEOF_XMMWORD/8
+    jz          short .column_ld4
+    sub         ecx, byte SIZEOF_XMMWORD/8
+    movq        xmmE, XMM_MMWORD [esi+ecx*RGB_PIXELSIZE]
+    pslldq      xmmA, SIZEOF_MMWORD
+    por         xmmA, xmmE
+.column_ld4:
+    test        cl, SIZEOF_XMMWORD/4
+    jz          short .column_ld8
+    sub         ecx, byte SIZEOF_XMMWORD/4
+    movdqa      xmmE, xmmA
+    movdqu      xmmA, XMMWORD [esi+ecx*RGB_PIXELSIZE]
+.column_ld8:
+    test        cl, SIZEOF_XMMWORD/2
+    mov         ecx, SIZEOF_XMMWORD
+    jz          short .rgb_gray_cnv
+    movdqa      xmmF, xmmA
+    movdqa      xmmH, xmmE
+    movdqu      xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    movdqu      xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD]
+    jmp         short .rgb_gray_cnv
+    alignx      16, 7
+
+.columnloop:
+    movdqu      xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    movdqu      xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD]
+    movdqu      xmmF, XMMWORD [esi+2*SIZEOF_XMMWORD]
+    movdqu      xmmH, XMMWORD [esi+3*SIZEOF_XMMWORD]
+
+.rgb_gray_cnv:
+    ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+    ; xmmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    ; xmmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+    ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+    movdqa      xmmD, xmmA
+    punpcklbw   xmmA, xmmE      ; xmmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35)
+    punpckhbw   xmmD, xmmE      ; xmmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37)
+
+    movdqa      xmmC, xmmF
+    punpcklbw   xmmF, xmmH      ; xmmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D)
+    punpckhbw   xmmC, xmmH      ; xmmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F)
+
+    movdqa      xmmB, xmmA
+    punpcklwd   xmmA, xmmF      ; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C)
+    punpckhwd   xmmB, xmmF      ; xmmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D)
+
+    movdqa      xmmG, xmmD
+    punpcklwd   xmmD, xmmC      ; xmmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E)
+    punpckhwd   xmmG, xmmC      ; xmmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F)
+
+    movdqa      xmmE, xmmA
+    punpcklbw   xmmA, xmmD      ; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
+    punpckhbw   xmmE, xmmD      ; xmmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E)
+
+    movdqa      xmmH, xmmB
+    punpcklbw   xmmB, xmmG      ; xmmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F)
+    punpckhbw   xmmH, xmmG      ; xmmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F)
+
+    pxor        xmmF, xmmF
+
+    movdqa      xmmC, xmmA
+    punpcklbw   xmmA, xmmF      ; xmmA=(00 02 04 06 08 0A 0C 0E)
+    punpckhbw   xmmC, xmmF      ; xmmC=(10 12 14 16 18 1A 1C 1E)
+
+    movdqa      xmmD, xmmB
+    punpcklbw   xmmB, xmmF      ; xmmB=(01 03 05 07 09 0B 0D 0F)
+    punpckhbw   xmmD, xmmF      ; xmmD=(11 13 15 17 19 1B 1D 1F)
+
+    movdqa      xmmG, xmmE
+    punpcklbw   xmmE, xmmF      ; xmmE=(20 22 24 26 28 2A 2C 2E)
+    punpckhbw   xmmG, xmmF      ; xmmG=(30 32 34 36 38 3A 3C 3E)
+
+    punpcklbw   xmmF, xmmH
+    punpckhbw   xmmH, xmmH
+    psrlw       xmmF, BYTE_BIT  ; xmmF=(21 23 25 27 29 2B 2D 2F)
+    psrlw       xmmH, BYTE_BIT  ; xmmH=(31 33 35 37 39 3B 3D 3F)
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    ; xmm0=R(02468ACE)=RE, xmm2=G(02468ACE)=GE, xmm4=B(02468ACE)=BE
+    ; xmm1=R(13579BDF)=RO, xmm3=G(13579BDF)=GO, xmm5=B(13579BDF)=BO
+
+    ; (Original)
+    ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+    ;
+    ; (This implementation)
+    ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+
+    movdqa      xmm6, xmm1
+    punpcklwd   xmm1, xmm3
+    punpckhwd   xmm6, xmm3
+    pmaddwd     xmm1, [GOTOFF(eax,PW_F0299_F0337)]  ; xmm1=ROL*FIX(0.299)+GOL*FIX(0.337)
+    pmaddwd     xmm6, [GOTOFF(eax,PW_F0299_F0337)]  ; xmm6=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    movdqa      xmm7, xmm6              ; xmm7=ROH*FIX(0.299)+GOH*FIX(0.337)
+
+    movdqa      xmm6, xmm0
+    punpcklwd   xmm0, xmm2
+    punpckhwd   xmm6, xmm2
+    pmaddwd     xmm0, [GOTOFF(eax,PW_F0299_F0337)]  ; xmm0=REL*FIX(0.299)+GEL*FIX(0.337)
+    pmaddwd     xmm6, [GOTOFF(eax,PW_F0299_F0337)]  ; xmm6=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    movdqa      XMMWORD [wk(0)], xmm0   ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337)
+    movdqa      XMMWORD [wk(1)], xmm6   ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337)
+
+    movdqa      xmm0, xmm5              ; xmm0=BO
+    movdqa      xmm6, xmm4              ; xmm6=BE
+
+    movdqa      xmm4, xmm0
+    punpcklwd   xmm0, xmm3
+    punpckhwd   xmm4, xmm3
+    pmaddwd     xmm0, [GOTOFF(eax,PW_F0114_F0250)]  ; xmm0=BOL*FIX(0.114)+GOL*FIX(0.250)
+    pmaddwd     xmm4, [GOTOFF(eax,PW_F0114_F0250)]  ; xmm4=BOH*FIX(0.114)+GOH*FIX(0.250)
+
+    movdqa      xmm3, [GOTOFF(eax,PD_ONEHALF)]      ; xmm3=[PD_ONEHALF]
+
+    paddd       xmm0, xmm1
+    paddd       xmm4, xmm7
+    paddd       xmm0, xmm3
+    paddd       xmm4, xmm3
+    psrld       xmm0, SCALEBITS         ; xmm0=YOL
+    psrld       xmm4, SCALEBITS         ; xmm4=YOH
+    packssdw    xmm0, xmm4              ; xmm0=YO
+
+    movdqa      xmm4, xmm6
+    punpcklwd   xmm6, xmm2
+    punpckhwd   xmm4, xmm2
+    pmaddwd     xmm6, [GOTOFF(eax,PW_F0114_F0250)]  ; xmm6=BEL*FIX(0.114)+GEL*FIX(0.250)
+    pmaddwd     xmm4, [GOTOFF(eax,PW_F0114_F0250)]  ; xmm4=BEH*FIX(0.114)+GEH*FIX(0.250)
+
+    movdqa      xmm2, [GOTOFF(eax,PD_ONEHALF)]      ; xmm2=[PD_ONEHALF]
+
+    paddd       xmm6, XMMWORD [wk(0)]
+    paddd       xmm4, XMMWORD [wk(1)]
+    paddd       xmm6, xmm2
+    paddd       xmm4, xmm2
+    psrld       xmm6, SCALEBITS         ; xmm6=YEL
+    psrld       xmm4, SCALEBITS         ; xmm4=YEH
+    packssdw    xmm6, xmm4              ; xmm6=YE
+
+    psllw       xmm0, BYTE_BIT
+    por         xmm6, xmm0              ; xmm6=Y
+    movdqa      XMMWORD [edi], xmm6     ; Save Y
+
+    sub         ecx, byte SIZEOF_XMMWORD
+    add         esi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; inptr
+    add         edi, byte SIZEOF_XMMWORD                ; outptr0
+    cmp         ecx, byte SIZEOF_XMMWORD
+    jae         near .columnloop
+    test        ecx, ecx
+    jnz         near .column_ld1
+
+    pop         ecx                     ; col
+    pop         esi
+    pop         edi
+    poppic      eax
+
+    add         esi, byte SIZEOF_JSAMPROW  ; input_buf
+    add         edi, byte SIZEOF_JSAMPROW
+    dec         eax                        ; num_rows
+    jg          near .rowloop
+
+.return:
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+    pop         ebx
+    mov         esp, ebp                ; esp <- aligned ebp
+    pop         esp                     ; esp <- original ebp
+    pop         ebp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,761 @@
+;
+; jchuff-sse2.asm - Huffman entropy encoding (SSE2)
+;
+; Copyright (C) 2009-2011, 2014-2017, 2019, D. R. Commander.
+; Copyright (C) 2015, Matthieu Darbois.
+; Copyright (C) 2018, Matthias Räncker.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains an SSE2 implementation for Huffman coding of one block.
+; The following code is based on jchuff.c; see jchuff.c for more details.
+
+%include "jsimdext.inc"
+
+struc working_state
+.next_output_byte:   resp 1     ; => next byte to write in buffer
+.free_in_buffer:     resp 1     ; # of byte spaces remaining in buffer
+.cur.put_buffer.simd resq 1     ; current bit accumulation buffer
+.cur.free_bits       resd 1     ; # of bits available in it
+.cur.last_dc_val     resd 4     ; last DC coef for each component
+.cinfo:              resp 1     ; dump_buffer needs access to this
+endstruc
+
+struc c_derived_tbl
+.ehufco:             resd 256   ; code for each symbol
+.ehufsi:             resb 256   ; length of code for each symbol
+; If no code has been allocated for a symbol S, ehufsi[S] contains 0
+endstruc
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    GLOBAL_DATA(jconst_huff_encode_one_block)
+
+EXTN(jconst_huff_encode_one_block):
+
+    alignz      32
+
+jpeg_mask_bits dq 0x0000, 0x0001, 0x0003, 0x0007
+               dq 0x000f, 0x001f, 0x003f, 0x007f
+               dq 0x00ff, 0x01ff, 0x03ff, 0x07ff
+               dq 0x0fff, 0x1fff, 0x3fff, 0x7fff
+
+times 1 << 14 db 15
+times 1 << 13 db 14
+times 1 << 12 db 13
+times 1 << 11 db 12
+times 1 << 10 db 11
+times 1 <<  9 db 10
+times 1 <<  8 db  9
+times 1 <<  7 db  8
+times 1 <<  6 db  7
+times 1 <<  5 db  6
+times 1 <<  4 db  5
+times 1 <<  3 db  4
+times 1 <<  2 db  3
+times 1 <<  1 db  2
+times 1 <<  0 db  1
+times 1       db  0
+jpeg_nbits_table:
+times 1       db  0
+times 1 <<  0 db  1
+times 1 <<  1 db  2
+times 1 <<  2 db  3
+times 1 <<  3 db  4
+times 1 <<  4 db  5
+times 1 <<  5 db  6
+times 1 <<  6 db  7
+times 1 <<  7 db  8
+times 1 <<  8 db  9
+times 1 <<  9 db 10
+times 1 << 10 db 11
+times 1 << 11 db 12
+times 1 << 12 db 13
+times 1 << 13 db 14
+times 1 << 14 db 15
+
+    alignz      32
+
+%ifdef PIC
+%define NBITS(x)      nbits_base + x
+%else
+%define NBITS(x)      jpeg_nbits_table + x
+%endif
+%define MASK_BITS(x)  NBITS((x) * 8) + (jpeg_mask_bits - jpeg_nbits_table)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+
+%define mm_put_buffer     mm0
+%define mm_all_0xff       mm1
+%define mm_temp           mm2
+%define mm_nbits          mm3
+%define mm_code_bits      mm3
+%define mm_code           mm4
+%define mm_overflow_bits  mm5
+%define mm_save_nbits     mm6
+
+; Shorthand used to describe SIMD operations:
+; wN:  xmmN treated as eight signed 16-bit values
+; wN[i]:  perform the same operation on all eight signed 16-bit values, i=0..7
+; bN:  xmmN treated as 16 unsigned 8-bit values, or
+;      mmN treated as eight unsigned 8-bit values
+; bN[i]:  perform the same operation on all unsigned 8-bit values,
+;         i=0..15 (SSE register) or i=0..7 (MMX register)
+; Contents of SIMD registers are shown in memory order.
+
+; Fill the bit buffer to capacity with the leading bits from code, then output
+; the bit buffer and put the remaining bits from code into the bit buffer.
+;
+; Usage:
+; code - contains the bits to shift into the bit buffer (LSB-aligned)
+; %1 - temp register
+; %2 - low byte of temp register
+; %3 - second byte of temp register
+; %4-%8 (optional) - extra instructions to execute before the macro completes
+; %9 - the label to which to jump when the macro completes
+;
+; Upon completion, free_bits will be set to the number of remaining bits from
+; code, and put_buffer will contain those remaining bits.  temp and code will
+; be clobbered.
+;
+; This macro encodes any 0xFF bytes as 0xFF 0x00, as does the EMIT_BYTE()
+; macro in jchuff.c.
+
+%macro EMIT_QWORD 9
+%define %%temp   %1
+%define %%tempb  %2
+%define %%temph  %3
+    add         nbits, free_bits             ; nbits += free_bits;
+    neg         free_bits                    ; free_bits = -free_bits;
+    movq        mm_temp, mm_code             ; temp = code;
+    movd        mm_nbits, nbits              ; nbits --> MMX register
+    movd        mm_overflow_bits, free_bits  ; overflow_bits (temp register) = free_bits;
+    neg         free_bits                    ; free_bits = -free_bits;
+    psllq       mm_put_buffer, mm_nbits      ; put_buffer <<= nbits;
+    psrlq       mm_temp, mm_overflow_bits    ; temp >>= overflow_bits;
+    add         free_bits, 64                ; free_bits += 64;
+    por         mm_temp, mm_put_buffer       ; temp |= put_buffer;
+%ifidn %%temp, nbits_base
+    movd        mm_save_nbits, nbits_base    ; save nbits_base
+%endif
+    movq        mm_code_bits, mm_temp        ; code_bits (temp register) = temp;
+    movq        mm_put_buffer, mm_code       ; put_buffer = code;
+    pcmpeqb     mm_temp, mm_all_0xff         ; b_temp[i] = (b_temp[i] == 0xFF ? 0xFF : 0);
+    movq        mm_code, mm_code_bits        ; code = code_bits;
+    psrlq       mm_code_bits, 32             ; code_bits >>= 32;
+    pmovmskb    nbits, mm_temp               ; nbits = 0;  nbits |= ((b_temp[i] >> 7) << i);
+    movd        %%temp, mm_code_bits         ; temp = code_bits;
+    bswap       %%temp                       ; temp = htonl(temp);
+    test        nbits, nbits                 ; if (nbits != 0)  /* Some 0xFF bytes */
+    jnz         %%.SLOW                      ;   goto %%.SLOW
+    mov         dword [buffer], %%temp       ; *(uint32_t)buffer = temp;
+%ifidn %%temp, nbits_base
+    movd        nbits_base, mm_save_nbits    ; restore nbits_base
+%endif
+    %4
+    movd        nbits, mm_code               ; nbits = (uint32_t)(code);
+    %5
+    bswap       nbits                        ; nbits = htonl(nbits);
+    mov         dword [buffer + 4], nbits    ; *(uint32_t)(buffer + 4) = nbits;
+    lea         buffer, [buffer + 8]         ; buffer += 8;
+    %6
+    %7
+    %8
+    jmp %9                                   ; return
+%%.SLOW:
+    ; Execute the equivalent of the EMIT_BYTE() macro in jchuff.c for all 8
+    ; bytes in the qword.
+    mov         byte [buffer], %%tempb     ; buffer[0] = temp[0];
+    cmp         %%tempb, 0xFF              ; Set CF if temp[0] < 0xFF
+    mov         byte [buffer+1], 0         ; buffer[1] = 0;
+    sbb         buffer, -2                 ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0));
+    mov         byte [buffer], %%temph     ; buffer[0] = temp[1];
+    cmp         %%temph, 0xFF              ; Set CF if temp[1] < 0xFF
+    mov         byte [buffer+1], 0         ; buffer[1] = 0;
+    sbb         buffer, -2                 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
+    shr         %%temp, 16                 ; temp >>= 16;
+    mov         byte [buffer], %%tempb     ; buffer[0] = temp[0];
+    cmp         %%tempb, 0xFF              ; Set CF if temp[0] < 0xFF
+    mov         byte [buffer+1], 0         ; buffer[1] = 0;
+    sbb         buffer, -2                 ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0));
+    mov         byte [buffer], %%temph     ; buffer[0] = temp[1];
+    cmp         %%temph, 0xFF              ; Set CF if temp[1] < 0xFF
+    mov         byte [buffer+1], 0         ; buffer[1] = 0;
+    sbb         buffer, -2                 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
+    movd        nbits, mm_code             ; nbits (temp register) = (uint32_t)(code)
+%ifidn %%temp, nbits_base
+    movd        nbits_base, mm_save_nbits  ; restore nbits_base
+%endif
+    bswap       nbits                      ; nbits = htonl(nbits)
+    mov         byte [buffer], nbitsb      ; buffer[0] = nbits[0];
+    cmp         nbitsb, 0xFF               ; Set CF if nbits[0] < 0xFF
+    mov         byte [buffer+1], 0         ; buffer[1] = 0;
+    sbb         buffer, -2                 ; buffer -= (-2 + (nbits[0] < 0xFF ? 1 : 0));
+    mov         byte [buffer], nbitsh      ; buffer[0] = nbits[1];
+    cmp         nbitsh, 0xFF               ; Set CF if nbits[1] < 0xFF
+    mov         byte [buffer+1], 0         ; buffer[1] = 0;
+    sbb         buffer, -2                 ; buffer -= (-2 + (nbits[1] < 0xFF ? 1 : 0));
+    shr         nbits, 16                  ; nbits >>= 16;
+    mov         byte [buffer], nbitsb      ; buffer[0] = nbits[0];
+    cmp         nbitsb, 0xFF               ; Set CF if nbits[0] < 0xFF
+    mov         byte [buffer+1], 0         ; buffer[1] = 0;
+    sbb         buffer, -2                 ; buffer -= (-2 + (nbits[0] < 0xFF ? 1 : 0));
+    mov         byte [buffer], nbitsh      ; buffer[0] = nbits[1];
+    %4
+    cmp         nbitsh, 0xFF               ; Set CF if nbits[1] < 0xFF
+    mov         byte [buffer+1], 0         ; buffer[1] = 0;
+    sbb         buffer, -2                 ; buffer -= (-2 + (nbits[1] < 0xFF ? 1 : 0));
+    %5
+    %6
+    %7
+    %8
+    jmp %9                                 ; return;
+%endmacro
+
+%macro PUSH 1
+    push        %1
+%assign stack_offset  stack_offset + 4
+%endmacro
+
+%macro POP 1
+    pop         %1
+%assign stack_offset  stack_offset - 4
+%endmacro
+
+; If PIC is defined, load the address of a symbol defined in this file into a
+; register.  Equivalent to
+;   get_GOT     %1
+;   lea         %1, [GOTOFF(%1, %2)]
+; without using the GOT.
+;
+; Usage:
+; %1 - register into which to load the address of the symbol
+; %2 - symbol whose address should be loaded
+; %3 - optional multi-line macro to execute before the symbol address is loaded
+; %4 - optional multi-line macro to execute after the symbol address is loaded
+;
+; If PIC is not defined, then %3 and %4 are executed in order.
+
+%macro GET_SYM 2-4
+%ifdef PIC
+    call        %%.geteip
+%%.ref:
+    %4
+    add         %1, %2 - %%.ref
+    jmp         short %%.done
+    align       32
+%%.geteip:
+    %3          4               ; must adjust stack pointer because of call
+    mov         %1, POINTER [esp]
+    ret
+    align       32
+%%.done:
+%else
+    %3          0
+    %4
+%endif
+%endmacro
+
+;
+; Encode a single block's worth of coefficients.
+;
+; GLOBAL(JOCTET *)
+; jsimd_huff_encode_one_block_sse2(working_state *state, JOCTET *buffer,
+;                                  JCOEFPTR block, int last_dc_val,
+;                                  c_derived_tbl *dctbl, c_derived_tbl *actbl)
+;
+; Stack layout:
+; Function args
+; Return address
+; Saved ebx
+; Saved ebp
+; Saved esi
+; Saved edi <-- esp_save
+; ...
+; esp_save
+; t_ 64*2 bytes (aligned to 128 bytes)
+;
+; esp is used (as t) to point into t_ (data in lower indices is not used once
+; esp passes over them, so this is signal-safe.)  Aligning to 128 bytes allows
+; us to find the rest of the data again.
+;
+; NOTES:
+; When shuffling data, we try to avoid pinsrw as much as possible, since it is
+; slow on many CPUs.  Its reciprocal throughput (issue latency) is 1 even on
+; modern CPUs, so chains of pinsrw instructions (even with different outputs)
+; can limit performance.  pinsrw is a VectorPath instruction on AMD K8 and
+; requires 2 µops (with memory operand) on Intel.  In either case, only one
+; pinsrw instruction can be decoded per cycle (and nothing else if they are
+; back-to-back), so out-of-order execution cannot be used to work around long
+; pinsrw chains (though for Sandy Bridge and later, this may be less of a
+; problem if the code runs from the µop cache.)
+;
+; We use tzcnt instead of bsf without checking for support.  The instruction is
+; executed as bsf on CPUs that don't support tzcnt (encoding is equivalent to
+; rep bsf.)  The destination (first) operand of bsf (and tzcnt on some CPUs) is
+; an input dependency (although the behavior is not formally defined, Intel
+; CPUs usually leave the destination unmodified if the source is zero.)  This
+; can prevent out-of-order execution, so we clear the destination before
+; invoking tzcnt.
+;
+; Initial register allocation
+; eax - frame --> buffer
+; ebx - nbits_base (PIC) / emit_temp
+; ecx - dctbl --> size --> state
+; edx - block --> nbits
+; esi - code_temp --> state --> actbl
+; edi - index_temp --> free_bits
+; esp - t
+; ebp - index
+
+%define frame       eax
+%ifdef PIC
+%define nbits_base  ebx
+%endif
+%define emit_temp   ebx
+%define emit_tempb  bl
+%define emit_temph  bh
+%define dctbl       ecx
+%define block       edx
+%define code_temp   esi
+%define index_temp  edi
+%define t           esp
+%define index       ebp
+
+%assign save_frame  DCTSIZE2 * SIZEOF_WORD
+
+; Step 1: Re-arrange input data according to jpeg_natural_order
+; xx 01 02 03 04 05 06 07      xx 01 08 16 09 02 03 10
+; 08 09 10 11 12 13 14 15      17 24 32 25 18 11 04 05
+; 16 17 18 19 20 21 22 23      12 19 26 33 40 48 41 34
+; 24 25 26 27 28 29 30 31 ==>  27 20 13 06 07 14 21 28
+; 32 33 34 35 36 37 38 39      35 42 49 56 57 50 43 36
+; 40 41 42 43 44 45 46 47      29 22 15 23 30 37 44 51
+; 48 49 50 51 52 53 54 55      58 59 52 45 38 31 39 46
+; 56 57 58 59 60 61 62 63      53 60 61 54 47 55 62 63
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_huff_encode_one_block_sse2)
+
+EXTN(jsimd_huff_encode_one_block_sse2):
+
+%assign stack_offset      0
+%define arg_state         4 + stack_offset
+%define arg_buffer        8 + stack_offset
+%define arg_block        12 + stack_offset
+%define arg_last_dc_val  16 + stack_offset
+%define arg_dctbl        20 + stack_offset
+%define arg_actbl        24 + stack_offset
+
+                                                          ;X: X = code stream
+    mov         block, [esp + arg_block]
+    PUSH        ebx
+    PUSH        ebp
+    movups      xmm3, XMMWORD [block + 0 * SIZEOF_WORD]   ;D: w3 = xx 01 02 03 04 05 06 07
+    PUSH        esi
+    PUSH        edi
+    movdqa      xmm0, xmm3                                ;A: w0 = xx 01 02 03 04 05 06 07
+    mov         frame, esp
+    lea         t, [frame - (save_frame + 4)]
+    movups      xmm1, XMMWORD [block + 8 * SIZEOF_WORD]   ;B: w1 = 08 09 10 11 12 13 14 15
+    and         t, -DCTSIZE2 * SIZEOF_WORD                                             ; t = &t_[0]
+    mov         [t + save_frame], frame
+    pxor        xmm4, xmm4                                ;A: w4[i] = 0;
+    punpckldq   xmm0, xmm1                                ;A: w0 = xx 01 08 09 02 03 10 11
+    pshuflw     xmm0, xmm0, 11001001b                     ;A: w0 = 01 08 xx 09 02 03 10 11
+    pinsrw      xmm0, word [block + 16 * SIZEOF_WORD], 2  ;A: w0 = 01 08 16 09 02 03 10 11
+    punpckhdq   xmm3, xmm1                                ;D: w3 = 04 05 12 13 06 07 14 15
+    punpcklqdq  xmm1, xmm3                                ;B: w1 = 08 09 10 11 04 05 12 13
+    pinsrw      xmm0, word [block + 17 * SIZEOF_WORD], 7  ;A: w0 = 01 08 16 09 02 03 10 17
+                                                          ;A:      (Row 0, offset 1)
+    pcmpgtw     xmm4, xmm0                                ;A: w4[i] = (w0[i] < 0 ? -1 : 0);
+    paddw       xmm0, xmm4                                ;A: w0[i] += w4[i];
+    movaps      XMMWORD [t + 0 * SIZEOF_WORD], xmm0       ;A: t[i] = w0[i];
+
+    movq        xmm2, qword [block + 24 * SIZEOF_WORD]    ;B: w2 = 24 25 26 27 -- -- -- --
+    pshuflw     xmm2, xmm2, 11011000b                     ;B: w2 = 24 26 25 27 -- -- -- --
+    pslldq      xmm1, 1 * SIZEOF_WORD                     ;B: w1 = -- 08 09 10 11 04 05 12
+    movups      xmm5, XMMWORD [block + 48 * SIZEOF_WORD]  ;H: w5 = 48 49 50 51 52 53 54 55
+    movsd       xmm1, xmm2                                ;B: w1 = 24 26 25 27 11 04 05 12
+    punpcklqdq  xmm2, xmm5                                ;C: w2 = 24 26 25 27 48 49 50 51
+    pinsrw      xmm1, word [block + 32 * SIZEOF_WORD], 1  ;B: w1 = 24 32 25 27 11 04 05 12
+    pxor        xmm4, xmm4                                ;A: w4[i] = 0;
+    psrldq      xmm3, 2 * SIZEOF_WORD                     ;D: w3 = 12 13 06 07 14 15 -- --
+    pcmpeqw     xmm0, xmm4                                ;A: w0[i] = (w0[i] == 0 ? -1 : 0);
+    pinsrw      xmm1, word [block + 18 * SIZEOF_WORD], 3  ;B: w1 = 24 32 25 18 11 04 05 12
+                                                          ;        (Row 1, offset 1)
+    pcmpgtw     xmm4, xmm1                                ;B: w4[i] = (w1[i] < 0 ? -1 : 0);
+    paddw       xmm1, xmm4                                ;B: w1[i] += w4[i];
+    movaps      XMMWORD [t + 8 * SIZEOF_WORD], xmm1       ;B: t[i+8] = w1[i];
+    pxor        xmm4, xmm4                                ;B: w4[i] = 0;
+    pcmpeqw     xmm1, xmm4                                ;B: w1[i] = (w1[i] == 0 ? -1 : 0);
+
+    packsswb    xmm0, xmm1                                ;AB: b0[i] = w0[i], b0[i+8] = w1[i]
+                                                          ;    w/ signed saturation
+
+    pinsrw      xmm3, word [block + 20 * SIZEOF_WORD], 0  ;D: w3 = 20 13 06 07 14 15 -- --
+    pinsrw      xmm3, word [block + 21 * SIZEOF_WORD], 5  ;D: w3 = 20 13 06 07 14 21 -- --
+    pinsrw      xmm3, word [block + 28 * SIZEOF_WORD], 6  ;D: w3 = 20 13 06 07 14 21 28 --
+    pinsrw      xmm3, word [block + 35 * SIZEOF_WORD], 7  ;D: w3 = 20 13 06 07 14 21 28 35
+                                                          ;        (Row 3, offset 1)
+    pcmpgtw     xmm4, xmm3                                ;D: w4[i] = (w3[i] < 0 ? -1 : 0);
+    paddw       xmm3, xmm4                                ;D: w3[i] += w4[i];
+    movaps      XMMWORD [t + 24 * SIZEOF_WORD], xmm3      ;D: t[i+24] = w3[i];
+    pxor        xmm4, xmm4                                ;D: w4[i] = 0;
+    pcmpeqw     xmm3, xmm4                                ;D: w3[i] = (w3[i] == 0 ? -1 : 0);
+
+    pinsrw      xmm2, word [block + 19 * SIZEOF_WORD], 0  ;C: w2 = 19 26 25 27 48 49 50 51
+    pinsrw      xmm2, word [block + 33 * SIZEOF_WORD], 2  ;C: w2 = 19 26 33 27 48 49 50 51
+    pinsrw      xmm2, word [block + 40 * SIZEOF_WORD], 3  ;C: w2 = 19 26 33 40 48 49 50 51
+    pinsrw      xmm2, word [block + 41 * SIZEOF_WORD], 5  ;C: w2 = 19 26 33 40 48 41 50 51
+    pinsrw      xmm2, word [block + 34 * SIZEOF_WORD], 6  ;C: w2 = 19 26 33 40 48 41 34 51
+    pinsrw      xmm2, word [block + 27 * SIZEOF_WORD], 7  ;C: w2 = 19 26 33 40 48 41 34 27
+                                                          ;        (Row 2, offset 1)
+    pcmpgtw     xmm4, xmm2                                ;C: w4[i] = (w2[i] < 0 ? -1 : 0);
+    paddw       xmm2, xmm4                                ;C: w2[i] += w4[i];
+    movsx       code_temp, word [block]                   ;Z:     code_temp = block[0];
+
+; %1 - stack pointer adjustment
+%macro GET_SYM_BEFORE 1
+    movaps      XMMWORD [t + 16 * SIZEOF_WORD + %1], xmm2
+                                                          ;C: t[i+16] = w2[i];
+    pxor        xmm4, xmm4                                ;C: w4[i] = 0;
+    pcmpeqw     xmm2, xmm4                                ;C: w2[i] = (w2[i] == 0 ? -1 : 0);
+    sub         code_temp, [frame + arg_last_dc_val]      ;Z:     code_temp -= last_dc_val;
+
+    packsswb    xmm2, xmm3                                ;CD: b2[i] = w2[i], b2[i+8] = w3[i]
+                                                          ;    w/ signed saturation
+
+    movdqa      xmm3, xmm5                                ;H: w3 = 48 49 50 51 52 53 54 55
+    pmovmskb    index_temp, xmm2                          ;Z:     index_temp = 0;  index_temp |= ((b2[i] >> 7) << i);
+    pmovmskb    index, xmm0                               ;Z:     index = 0;  index |= ((b0[i] >> 7) << i);
+    movups      xmm0, XMMWORD [block + 56 * SIZEOF_WORD]  ;H: w0 = 56 57 58 59 60 61 62 63
+    punpckhdq   xmm3, xmm0                                ;H: w3 = 52 53 60 61 54 55 62 63
+    shl         index_temp, 16                            ;Z:     index_temp <<= 16;
+    psrldq      xmm3, 1 * SIZEOF_WORD                     ;H: w3 = 53 60 61 54 55 62 63 --
+    pxor        xmm2, xmm2                                ;H: w2[i] = 0;
+    pshuflw     xmm3, xmm3, 00111001b                     ;H: w3 = 60 61 54 53 55 62 63 --
+    or          index, index_temp                         ;Z:     index |= index_temp;
+%undef index_temp
+%define free_bits  edi
+%endmacro
+
+%macro GET_SYM_AFTER 0
+    movq        xmm1, qword [block + 44 * SIZEOF_WORD]    ;G: w1 = 44 45 46 47 -- -- -- --
+    unpcklps    xmm5, xmm0                                ;E: w5 = 48 49 56 57 50 51 58 59
+    pxor        xmm0, xmm0                                ;H: w0[i] = 0;
+    not         index                                     ;Z:     index = ~index;
+    pinsrw      xmm3, word [block + 47 * SIZEOF_WORD], 3  ;H: w3 = 60 61 54 47 55 62 63 --
+                                                          ;        (Row 7, offset 1)
+    pcmpgtw     xmm2, xmm3                                ;H: w2[i] = (w3[i] < 0 ? -1 : 0);
+    mov         dctbl, [frame + arg_dctbl]
+    paddw       xmm3, xmm2                                ;H: w3[i] += w2[i];
+    movaps      XMMWORD [t + 56 * SIZEOF_WORD], xmm3      ;H: t[i+56] = w3[i];
+    movq        xmm4, qword [block + 36 * SIZEOF_WORD]    ;G: w4 = 36 37 38 39 -- -- -- --
+    pcmpeqw     xmm3, xmm0                                ;H: w3[i] = (w3[i] == 0 ? -1 : 0);
+    punpckldq   xmm4, xmm1                                ;G: w4 = 36 37 44 45 38 39 46 47
+    movdqa      xmm1, xmm4                                ;F: w1 = 36 37 44 45 38 39 46 47
+    pcmpeqw     mm_all_0xff, mm_all_0xff                  ;Z:     all_0xff[i] = 0xFF;
+%endmacro
+
+    GET_SYM     nbits_base, jpeg_nbits_table, GET_SYM_BEFORE, GET_SYM_AFTER
+
+    psrldq      xmm4, 1 * SIZEOF_WORD                     ;G: w4 = 37 44 45 38 39 46 47 --
+    shufpd      xmm1, xmm5, 10b                           ;F: w1 = 36 37 44 45 50 51 58 59
+    pshufhw     xmm4, xmm4, 11010011b                     ;G: w4 = 37 44 45 38 -- 39 46 --
+    pslldq      xmm1, 1 * SIZEOF_WORD                     ;F: w1 = -- 36 37 44 45 50 51 58
+    pinsrw      xmm4, word [block + 59 * SIZEOF_WORD], 0  ;G: w4 = 59 44 45 38 -- 39 46 --
+    pshufd      xmm1, xmm1, 11011000b                     ;F: w1 = -- 36 45 50 37 44 51 58
+    cmp         code_temp, 1 << 31                        ;Z:     Set CF if code_temp < 0x80000000,
+                                                          ;Z:     i.e. if code_temp is positive
+    pinsrw      xmm4, word [block + 52 * SIZEOF_WORD], 1  ;G: w4 = 59 52 45 38 -- 39 46 --
+    movlps      xmm1, qword [block + 20 * SIZEOF_WORD]    ;F: w1 = 20 21 22 23 37 44 51 58
+    pinsrw      xmm4, word [block + 31 * SIZEOF_WORD], 4  ;G: w4 = 59 52 45 38 31 39 46 --
+    pshuflw     xmm1, xmm1, 01110010b                     ;F: w1 = 22 20 23 21 37 44 51 58
+    pinsrw      xmm4, word [block + 53 * SIZEOF_WORD], 7  ;G: w4 = 59 52 45 38 31 39 46 53
+                                                          ;        (Row 6, offset 1)
+    adc         code_temp, -1                             ;Z:     code_temp += -1 + (code_temp >= 0 ? 1 : 0);
+    pxor        xmm2, xmm2                                ;G: w2[i] = 0;
+    pcmpgtw     xmm0, xmm4                                ;G: w0[i] = (w4[i] < 0 ? -1 : 0);
+    pinsrw      xmm1, word [block + 15 * SIZEOF_WORD], 1  ;F: w1 = 22 15 23 21 37 44 51 58
+    paddw       xmm4, xmm0                                ;G: w4[i] += w0[i];
+    movaps      XMMWORD [t + 48 * SIZEOF_WORD], xmm4      ;G: t[48+i] = w4[i];
+    movd        mm_temp, code_temp                        ;Z:     temp = code_temp
+    pinsrw      xmm1, word [block + 30 * SIZEOF_WORD], 3  ;F: w1 = 22 15 23 30 37 44 51 58
+                                                          ;        (Row 5, offset 1)
+    pcmpeqw     xmm4, xmm2                                ;G: w4[i] = (w4[i] == 0 ? -1 : 0);
+
+    packsswb    xmm4, xmm3                                ;GH: b4[i] = w4[i], b4[i+8] = w3[i]
+                                                          ;    w/ signed saturation
+
+    lea         t, [t - SIZEOF_WORD]                      ;Z:     t = &t[-1]
+    pxor        xmm0, xmm0                                ;F: w0[i] = 0;
+    pcmpgtw     xmm2, xmm1                                ;F: w2[i] = (w1[i] < 0 ? -1 : 0);
+    paddw       xmm1, xmm2                                ;F: w1[i] += w2[i];
+    movaps      XMMWORD [t + (40+1) * SIZEOF_WORD], xmm1  ;F: t[40+i] = w1[i];
+    pcmpeqw     xmm1, xmm0                                ;F: w1[i] = (w1[i] == 0 ? -1 : 0);
+    pinsrw      xmm5, word [block + 42 * SIZEOF_WORD], 0  ;E: w5 = 42 49 56 57 50 51 58 59
+    pinsrw      xmm5, word [block + 43 * SIZEOF_WORD], 5  ;E: w5 = 42 49 56 57 50 43 58 59
+    pinsrw      xmm5, word [block + 36 * SIZEOF_WORD], 6  ;E: w5 = 42 49 56 57 50 43 36 59
+    pinsrw      xmm5, word [block + 29 * SIZEOF_WORD], 7  ;E: w5 = 42 49 56 57 50 43 36 29
+                                                          ;        (Row 4, offset 1)
+%undef block
+%define nbits  edx
+%define nbitsb  dl
+%define nbitsh  dh
+    movzx       nbits, byte [NBITS(code_temp)]            ;Z:     nbits = JPEG_NBITS(code_temp);
+%undef code_temp
+%define state  esi
+    pxor        xmm2, xmm2                                ;E: w2[i] = 0;
+    mov         state, [frame + arg_state]
+    movd        mm_nbits, nbits                           ;Z:     nbits --> MMX register
+    pcmpgtw     xmm0, xmm5                                ;E: w0[i] = (w5[i] < 0 ? -1 : 0);
+    movd        mm_code, dword [dctbl + c_derived_tbl.ehufco + nbits * 4]
+                                                          ;Z:     code = dctbl->ehufco[nbits];
+%define size  ecx
+%define sizeb  cl
+%define sizeh  ch
+    paddw       xmm5, xmm0                                ;E: w5[i] += w0[i];
+    movaps      XMMWORD [t + (32+1) * SIZEOF_WORD], xmm5  ;E: t[32+i] = w5[i];
+    movzx       size, byte [dctbl + c_derived_tbl.ehufsi + nbits]
+                                                          ;Z:     size = dctbl->ehufsi[nbits];
+%undef dctbl
+    pcmpeqw     xmm5, xmm2                                ;E: w5[i] = (w5[i] == 0 ? -1 : 0);
+
+    packsswb    xmm5, xmm1                                ;EF: b5[i] = w5[i], b5[i+8] = w1[i]
+                                                          ;    w/ signed saturation
+
+    movq        mm_put_buffer, [state + working_state.cur.put_buffer.simd]
+                                                          ;Z:     put_buffer = state->cur.put_buffer.simd;
+    mov         free_bits, [state + working_state.cur.free_bits]
+                                                          ;Z:     free_bits = state->cur.free_bits;
+%undef state
+%define actbl  esi
+    mov         actbl, [frame + arg_actbl]
+%define buffer  eax
+    mov         buffer, [frame + arg_buffer]
+%undef frame
+    jmp        .BEGIN
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    align       16
+; size <= 32, so this is not really a loop
+.BRLOOP1:                                                 ; .BRLOOP1:
+    movzx       nbits, byte [actbl + c_derived_tbl.ehufsi + 0xf0]
+                                                          ; nbits = actbl->ehufsi[0xf0];
+    movd        mm_code, dword [actbl + c_derived_tbl.ehufco + 0xf0 * 4]
+                                                          ; code = actbl->ehufco[0xf0];
+    and         index, 0x7ffffff                          ; clear index if size == 32
+    sub         size, 16                                  ; size -= 16;
+    sub         free_bits, nbits                          ; if ((free_bits -= nbits) <= 0)
+    jle         .EMIT_BRLOOP1                             ;   goto .EMIT_BRLOOP1;
+    movd        mm_nbits, nbits                           ; nbits --> MMX register
+    psllq       mm_put_buffer, mm_nbits                   ; put_buffer <<= nbits;
+    por         mm_put_buffer, mm_code                    ; put_buffer |= code;
+    jmp         .ERLOOP1                                  ; goto .ERLOOP1;
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    align       16
+%ifdef PIC
+    times 6     nop
+%else
+    times 2     nop
+%endif
+.BLOOP1:                                                  ; do {  /* size = # of zero bits/elements to skip */
+; if size == 32, index remains unchanged.  Correct in .BRLOOP.
+    shr         index, sizeb                              ;   index >>= size;
+    lea         t, [t + size * SIZEOF_WORD]               ;   t += size;
+    cmp         size, 16                                  ;   if (size > 16)
+    jg          .BRLOOP1                                  ;     goto .BRLOOP1;
+.ERLOOP1:                                                 ; .ERLOOP1:
+    movsx       nbits, word [t]                           ;   nbits = *t;
+%ifdef PIC
+    add         size, size                                ;   size += size;
+%else
+    lea         size, [size * 2]                          ;   size += size;
+%endif
+    movd        mm_temp, nbits                            ;   temp = nbits;
+    movzx       nbits, byte [NBITS(nbits)]                ;   nbits = JPEG_NBITS(nbits);
+    lea         size, [size * 8 + nbits]                  ;   size = size * 8 + nbits;
+    movd        mm_nbits, nbits                           ;   nbits --> MMX register
+    movd        mm_code, dword [actbl + c_derived_tbl.ehufco + (size - 16) * 4]
+                                                          ;   code = actbl->ehufco[size-16];
+    movzx       size, byte [actbl + c_derived_tbl.ehufsi + (size - 16)]
+                                                          ;   size = actbl->ehufsi[size-16];
+.BEGIN:                                                   ; .BEGIN:
+    pand        mm_temp, [MASK_BITS(nbits)]               ;   temp &= (1 << nbits) - 1;
+    psllq       mm_code, mm_nbits                         ;   code <<= nbits;
+    add         nbits, size                               ;   nbits += size;
+    por         mm_code, mm_temp                          ;   code |= temp;
+    sub         free_bits, nbits                          ;   if ((free_bits -= nbits) <= 0)
+    jle         .EMIT_ERLOOP1                             ;     insert code, flush buffer, init size, goto .BLOOP1
+    xor         size, size                                ;   size = 0;  /* kill tzcnt input dependency */
+    tzcnt       size, index                               ;   size = # of trailing 0 bits in index
+    movd        mm_nbits, nbits                           ;   nbits --> MMX register
+    psllq       mm_put_buffer, mm_nbits                   ;   put_buffer <<= nbits;
+    inc         size                                      ;   ++size;
+    por         mm_put_buffer, mm_code                    ;   put_buffer |= code;
+    test        index, index
+    jnz         .BLOOP1                                   ; } while (index != 0);
+; Round 2
+; t points to the last used word, possibly below t_ if the previous index had 32 zero bits.
+.ELOOP1:                                                  ; .ELOOP1:
+    pmovmskb    size, xmm4                                ; size = 0;  size |= ((b4[i] >> 7) << i);
+    pmovmskb    index, xmm5                               ; index = 0;  index |= ((b5[i] >> 7) << i);
+    shl         size, 16                                  ; size <<= 16;
+    or          index, size                               ; index |= size;
+    not         index                                     ; index = ~index;
+    lea         nbits, [t + (1 + DCTSIZE2) * SIZEOF_WORD]
+                                                          ; nbits = t + 1 + 64;
+    and         nbits, -DCTSIZE2 * SIZEOF_WORD            ; nbits &= -128;  /* now points to &t_[64] */
+    sub         nbits, t                                  ; nbits -= t;
+    shr         nbits, 1                                  ; nbits >>= 1;  /* # of leading 0 bits in old index + 33 */
+    tzcnt       size, index                               ; size = # of trailing 0 bits in index
+    inc         size                                      ; ++size;
+    test        index, index                              ; if (index == 0)
+    jz          .ELOOP2                                   ;   goto .ELOOP2;
+; NOTE: size == 32 cannot happen, since the last element is always 0.
+    shr         index, sizeb                              ; index >>= size;
+    lea         size, [size + nbits - 33]                 ; size = size + nbits - 33;
+    lea         t, [t + size * SIZEOF_WORD]               ; t += size;
+    cmp         size, 16                                  ; if (size <= 16)
+    jle         .ERLOOP2                                  ;   goto .ERLOOP2;
+.BRLOOP2:                                                 ; do {
+    movzx       nbits, byte [actbl + c_derived_tbl.ehufsi + 0xf0]
+                                                          ;   nbits = actbl->ehufsi[0xf0];
+    sub         size, 16                                  ;   size -= 16;
+    movd        mm_code, dword [actbl + c_derived_tbl.ehufco + 0xf0 * 4]
+                                                          ;   code = actbl->ehufco[0xf0];
+    sub         free_bits, nbits                          ;   if ((free_bits -= nbits) <= 0)
+    jle         .EMIT_BRLOOP2                             ;     insert code and flush put_buffer
+    movd        mm_nbits, nbits                           ;   else { nbits --> MMX register
+    psllq       mm_put_buffer, mm_nbits                   ;     put_buffer <<= nbits;
+    por         mm_put_buffer, mm_code                    ;     put_buffer |= code;
+    cmp         size, 16                                  ;     if (size <= 16)
+    jle        .ERLOOP2                                   ;       goto .ERLOOP2;
+    jmp        .BRLOOP2                                   ; } while (1);
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    align      16
+.BLOOP2:                                                  ; do {  /* size = # of zero bits/elements to skip */
+    shr         index, sizeb                              ;   index >>= size;
+    lea         t, [t + size * SIZEOF_WORD]               ;   t += size;
+    cmp         size, 16                                  ;   if (size > 16)
+    jg          .BRLOOP2                                  ;     goto .BRLOOP2;
+.ERLOOP2:                                                 ; .ERLOOP2:
+    movsx       nbits, word [t]                           ;   nbits = *t;
+    add         size, size                                ;   size += size;
+    movd        mm_temp, nbits                            ;   temp = nbits;
+    movzx       nbits, byte [NBITS(nbits)]                ;   nbits = JPEG_NBITS(nbits);
+    movd        mm_nbits, nbits                           ;   nbits --> MMX register
+    lea         size, [size * 8 + nbits]                  ;   size = size * 8 + nbits;
+    movd        mm_code, dword [actbl + c_derived_tbl.ehufco + (size - 16) * 4]
+                                                          ;   code = actbl->ehufco[size-16];
+    movzx       size, byte [actbl + c_derived_tbl.ehufsi + (size - 16)]
+                                                          ;   size = actbl->ehufsi[size-16];
+    psllq       mm_code, mm_nbits                         ;   code <<= nbits;
+    pand        mm_temp, [MASK_BITS(nbits)]               ;   temp &= (1 << nbits) - 1;
+    lea         nbits, [nbits + size]                     ;   nbits += size;
+    por         mm_code, mm_temp                          ;   code |= temp;
+    xor         size, size                                ;   size = 0;  /* kill tzcnt input dependency */
+    sub         free_bits, nbits                          ;   if ((free_bits -= nbits) <= 0)
+    jle         .EMIT_ERLOOP2                             ;     insert code, flush buffer, init size, goto .BLOOP2
+    tzcnt       size, index                               ;   size = # of trailing 0 bits in index
+    movd        mm_nbits, nbits                           ;   nbits --> MMX register
+    psllq       mm_put_buffer, mm_nbits                   ;   put_buffer <<= nbits;
+    inc         size                                      ;   ++size;
+    por         mm_put_buffer, mm_code                    ;   put_buffer |= code;
+    test        index, index
+    jnz         .BLOOP2                                   ; } while (index != 0);
+.ELOOP2:                                                  ; .ELOOP2:
+    mov         nbits, t                                  ; nbits = t;
+    lea         t, [t + SIZEOF_WORD]                      ; t = &t[1];
+    and         nbits, DCTSIZE2 * SIZEOF_WORD - 1         ; nbits &= 127;
+    and         t, -DCTSIZE2 * SIZEOF_WORD                ; t &= -128;  /* t = &t_[0]; */
+    cmp         nbits, (DCTSIZE2 - 2) * SIZEOF_WORD       ; if (nbits != 62 * 2)
+    je          .EFN                                      ; {
+    movd        mm_code, dword [actbl + c_derived_tbl.ehufco + 0]
+                                                          ;   code = actbl->ehufco[0];
+    movzx       nbits, byte [actbl + c_derived_tbl.ehufsi + 0]
+                                                          ;   nbits = actbl->ehufsi[0];
+    sub         free_bits, nbits                          ;   if ((free_bits -= nbits) <= 0)
+    jg          .EFN_SKIP_EMIT_CODE                       ;   {
+    EMIT_QWORD  size, sizeb, sizeh, , , , , , .EFN        ;     insert code, flush put_buffer
+    align       16
+.EFN_SKIP_EMIT_CODE:                                      ;   } else {
+    movd        mm_nbits, nbits                           ;     nbits --> MMX register
+    psllq       mm_put_buffer, mm_nbits                   ;     put_buffer <<= nbits;
+    por         mm_put_buffer, mm_code                    ;     put_buffer |= code;
+.EFN:                                                     ; } }
+%define frame  esp
+    mov         frame, [t + save_frame]
+%define state  ecx
+    mov         state, [frame + arg_state]
+    movq        [state + working_state.cur.put_buffer.simd], mm_put_buffer
+                                                          ; state->cur.put_buffer.simd = put_buffer;
+    emms
+    mov         [state + working_state.cur.free_bits], free_bits
+                                                          ; state->cur.free_bits = free_bits;
+    POP         edi
+    POP         esi
+    POP         ebp
+    POP         ebx
+    ret
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    align       16
+.EMIT_BRLOOP1:
+    EMIT_QWORD  emit_temp, emit_tempb, emit_temph, , , , , , \
+      .ERLOOP1
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    align       16
+.EMIT_ERLOOP1:
+    EMIT_QWORD  size, sizeb, sizeh, \
+      { xor     size, size }, \
+      { tzcnt   size, index }, \
+      { inc     size }, \
+      { test    index, index }, \
+      { jnz     .BLOOP1 }, \
+      .ELOOP1
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    align       16
+.EMIT_BRLOOP2:
+    EMIT_QWORD  emit_temp, emit_tempb, emit_temph, , , , \
+      { cmp     size, 16 }, \
+      { jle     .ERLOOP2 }, \
+      .BRLOOP2
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    align       16
+.EMIT_ERLOOP2:
+    EMIT_QWORD  size, sizeb, sizeh, \
+      { xor     size, size }, \
+      { tzcnt   size, index }, \
+      { inc     size }, \
+      { test    index, index }, \
+      { jnz     .BLOOP2 }, \
+      .ELOOP2
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,662 @@
+;
+; jcphuff-sse2.asm - prepare data for progressive Huffman encoding (SSE2)
+;
+; Copyright (C) 2016, 2018, Matthieu Darbois
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains an SSE2 implementation of data preparation for progressive
+; Huffman encoding.  See jcphuff.c for more details.
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+
+; --------------------------------------------------------------------------
+; Macros to load data for jsimd_encode_mcu_AC_first_prepare_sse2() and
+; jsimd_encode_mcu_AC_refine_prepare_sse2()
+
+%macro LOAD16 0
+    pxor        N0, N0
+    pxor        N1, N1
+
+    mov         T0, INT [LUT +  0*SIZEOF_INT]
+    mov         T1, INT [LUT +  8*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 0
+    pinsrw      X1, word [BLOCK + T1 * 2], 0
+
+    mov         T0, INT [LUT +  1*SIZEOF_INT]
+    mov         T1, INT [LUT +  9*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 1
+    pinsrw      X1, word [BLOCK + T1 * 2], 1
+
+    mov         T0, INT [LUT +  2*SIZEOF_INT]
+    mov         T1, INT [LUT + 10*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 2
+    pinsrw      X1, word [BLOCK + T1 * 2], 2
+
+    mov         T0, INT [LUT +  3*SIZEOF_INT]
+    mov         T1, INT [LUT + 11*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 3
+    pinsrw      X1, word [BLOCK + T1 * 2], 3
+
+    mov         T0, INT [LUT +  4*SIZEOF_INT]
+    mov         T1, INT [LUT + 12*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 4
+    pinsrw      X1, word [BLOCK + T1 * 2], 4
+
+    mov         T0, INT [LUT +  5*SIZEOF_INT]
+    mov         T1, INT [LUT + 13*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 5
+    pinsrw      X1, word [BLOCK + T1 * 2], 5
+
+    mov         T0, INT [LUT +  6*SIZEOF_INT]
+    mov         T1, INT [LUT + 14*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 6
+    pinsrw      X1, word [BLOCK + T1 * 2], 6
+
+    mov         T0, INT [LUT +  7*SIZEOF_INT]
+    mov         T1, INT [LUT + 15*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 7
+    pinsrw      X1, word [BLOCK + T1 * 2], 7
+%endmacro
+
+%macro LOAD15 0
+    pxor        N0, N0
+    pxor        N1, N1
+    pxor        X1, X1
+
+    mov         T0, INT [LUT +  0*SIZEOF_INT]
+    mov         T1, INT [LUT +  8*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 0
+    pinsrw      X1, word [BLOCK + T1 * 2], 0
+
+    mov         T0, INT [LUT +  1*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 1
+
+    mov         T0, INT [LUT +  2*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 2
+
+    mov         T0, INT [LUT +  3*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 3
+
+    mov         T0, INT [LUT +  4*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 4
+
+    mov         T0, INT [LUT +  5*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 5
+
+    mov         T0, INT [LUT +  6*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 6
+
+    mov         T0, INT [LUT +  7*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 7
+
+    cmp         LENEND, 2
+    jl          %%.ELOAD15
+    mov         T1, INT [LUT +  9*SIZEOF_INT]
+    pinsrw      X1, word [BLOCK + T1 * 2], 1
+
+    cmp         LENEND, 3
+    jl          %%.ELOAD15
+    mov         T1, INT [LUT + 10*SIZEOF_INT]
+    pinsrw      X1, word [BLOCK + T1 * 2], 2
+
+    cmp         LENEND, 4
+    jl          %%.ELOAD15
+    mov         T1, INT [LUT + 11*SIZEOF_INT]
+    pinsrw      X1, word [BLOCK + T1 * 2], 3
+
+    cmp         LENEND, 5
+    jl          %%.ELOAD15
+    mov         T1, INT [LUT + 12*SIZEOF_INT]
+    pinsrw      X1, word [BLOCK + T1 * 2], 4
+
+    cmp         LENEND, 6
+    jl          %%.ELOAD15
+    mov         T1, INT [LUT + 13*SIZEOF_INT]
+    pinsrw      X1, word [BLOCK + T1 * 2], 5
+
+    cmp         LENEND, 7
+    jl          %%.ELOAD15
+    mov         T1, INT [LUT + 14*SIZEOF_INT]
+    pinsrw      X1, word [BLOCK + T1 * 2], 6
+%%.ELOAD15:
+%endmacro
+
+%macro LOAD8 0
+    pxor        N0, N0
+
+    mov         T0, INT [LUT +  0*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 0
+
+    mov         T0, INT [LUT +  1*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 1
+
+    mov         T0, INT [LUT +  2*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 2
+
+    mov         T0, INT [LUT +  3*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 3
+
+    mov         T0, INT [LUT +  4*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 4
+
+    mov         T0, INT [LUT +  5*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 5
+
+    mov         T0, INT [LUT +  6*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 6
+
+    mov         T0, INT [LUT +  7*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T0 * 2], 7
+%endmacro
+
+%macro LOAD7 0
+    pxor        N0, N0
+    pxor        X0, X0
+
+    mov         T1, INT [LUT +  0*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 0
+
+    cmp         LENEND, 2
+    jl          %%.ELOAD7
+    mov         T1, INT [LUT +  1*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 1
+
+    cmp         LENEND, 3
+    jl          %%.ELOAD7
+    mov         T1, INT [LUT +  2*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 2
+
+    cmp         LENEND, 4
+    jl          %%.ELOAD7
+    mov         T1, INT [LUT +  3*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 3
+
+    cmp         LENEND, 5
+    jl          %%.ELOAD7
+    mov         T1, INT [LUT +  4*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 4
+
+    cmp         LENEND, 6
+    jl          %%.ELOAD7
+    mov         T1, INT [LUT +  5*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 5
+
+    cmp         LENEND, 7
+    jl          %%.ELOAD7
+    mov         T1, INT [LUT +  6*SIZEOF_INT]
+    pinsrw      X0, word [BLOCK + T1 * 2], 6
+%%.ELOAD7:
+%endmacro
+
+%macro REDUCE0 0
+    movdqa      xmm0, XMMWORD [VALUES + ( 0*2)]
+    movdqa      xmm1, XMMWORD [VALUES + ( 8*2)]
+    movdqa      xmm2, XMMWORD [VALUES + (16*2)]
+    movdqa      xmm3, XMMWORD [VALUES + (24*2)]
+    movdqa      xmm4, XMMWORD [VALUES + (32*2)]
+    movdqa      xmm5, XMMWORD [VALUES + (40*2)]
+    movdqa      xmm6, XMMWORD [VALUES + (48*2)]
+
+    pcmpeqw     xmm0, ZERO
+    pcmpeqw     xmm1, ZERO
+    pcmpeqw     xmm2, ZERO
+    pcmpeqw     xmm3, ZERO
+    pcmpeqw     xmm4, ZERO
+    pcmpeqw     xmm5, ZERO
+    pcmpeqw     xmm6, ZERO
+    pcmpeqw     xmm7, XMMWORD [VALUES + (56*2)]
+
+    packsswb    xmm0, xmm1
+    packsswb    xmm2, xmm3
+    packsswb    xmm4, xmm5
+    packsswb    xmm6, xmm7
+
+    pmovmskb    eax, xmm0
+    pmovmskb    ecx, xmm2
+    pmovmskb    edx, xmm4
+    pmovmskb    esi, xmm6
+
+    shl         ecx, 16
+    shl         esi, 16
+
+    or          eax, ecx
+    or          edx, esi
+
+    not         eax
+    not         edx
+
+    mov         edi, ZEROBITS
+
+    mov         INT [edi], eax
+    mov         INT [edi+SIZEOF_INT], edx
+%endmacro
+
+;
+; Prepare data for jsimd_encode_mcu_AC_first().
+;
+; GLOBAL(void)
+; jsimd_encode_mcu_AC_first_prepare_sse2(const JCOEF *block,
+;                                        const int *jpeg_natural_order_start,
+;                                        int Sl, int Al, JCOEF *values,
+;                                        size_t *zerobits)
+;
+; eax + 8 = const JCOEF *block
+; eax + 12 = const int *jpeg_natural_order_start
+; eax + 16 = int Sl
+; eax + 20 = int Al
+; eax + 24 = JCOEF *values
+; eax + 28 = size_t *zerobits
+
+%define ZERO    xmm7
+%define X0      xmm0
+%define X1      xmm1
+%define N0      xmm2
+%define N1      xmm3
+%define AL      xmm4
+%define K       eax
+%define LENEND  eax
+%define LUT     ebx
+%define T0      ecx
+%define T1      edx
+%define BLOCK   esi
+%define VALUES  edi
+%define LEN     ebp
+
+%define ZEROBITS  INT [esp + 5 * 4]
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_encode_mcu_AC_first_prepare_sse2)
+
+EXTN(jsimd_encode_mcu_AC_first_prepare_sse2):
+    push        ebp
+    mov         eax, esp                     ; eax = original ebp
+    sub         esp, byte 4
+    and         esp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [esp], eax
+    mov         ebp, esp                     ; ebp = aligned ebp
+    sub         esp, 4
+    push        ebx
+    push        ecx
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+    push        ebp
+
+    mov         BLOCK, INT [eax + 8]
+    mov         LUT, INT [eax + 12]
+    mov         VALUES, INT [eax + 24]
+    movd        AL, INT [eax + 20]
+    mov         T0, INT [eax + 28]
+    mov         ZEROBITS, T0
+    mov         LEN, INT [eax + 16]
+    pxor        ZERO, ZERO
+    mov         K, LEN
+    and         K, -16
+    shr         K, 4
+    jz          .ELOOP16
+.BLOOP16:
+    LOAD16
+    pcmpgtw     N0, X0
+    pcmpgtw     N1, X1
+    paddw       X0, N0
+    paddw       X1, N1
+    pxor        X0, N0
+    pxor        X1, N1
+    psrlw       X0, AL
+    psrlw       X1, AL
+    pxor        N0, X0
+    pxor        N1, X1
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    movdqa      XMMWORD [VALUES + (8) * 2], X1
+    movdqa      XMMWORD [VALUES + (0 + DCTSIZE2) * 2], N0
+    movdqa      XMMWORD [VALUES + (8 + DCTSIZE2) * 2], N1
+    add         VALUES, 16*2
+    add         LUT, 16*SIZEOF_INT
+    dec         K
+    jnz         .BLOOP16
+    test        LEN, 15
+    je          .PADDING
+.ELOOP16:
+    mov         LENEND, LEN
+    and         LENEND, 7
+
+    test        LEN, 8
+    jz          .TRY7
+    test        LEN, 7
+    jz          .TRY8
+
+    LOAD15
+    pcmpgtw     N0, X0
+    pcmpgtw     N1, X1
+    paddw       X0, N0
+    paddw       X1, N1
+    pxor        X0, N0
+    pxor        X1, N1
+    psrlw       X0, AL
+    psrlw       X1, AL
+    pxor        N0, X0
+    pxor        N1, X1
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    movdqa      XMMWORD [VALUES + (8) * 2], X1
+    movdqa      XMMWORD [VALUES + (0 + DCTSIZE2) * 2], N0
+    movdqa      XMMWORD [VALUES + (8 + DCTSIZE2) * 2], N1
+    add         VALUES, 16*2
+    jmp         .PADDING
+.TRY8:
+    LOAD8
+    pcmpgtw     N0, X0
+    paddw       X0, N0
+    pxor        X0, N0
+    psrlw       X0, AL
+    pxor        N0, X0
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    movdqa      XMMWORD [VALUES + (0 + DCTSIZE2) * 2], N0
+    add         VALUES, 8*2
+    jmp         .PADDING
+.TRY7:
+    LOAD7
+    pcmpgtw     N0, X0
+    paddw       X0, N0
+    pxor        X0, N0
+    psrlw       X0, AL
+    pxor        N0, X0
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    movdqa      XMMWORD [VALUES + (0 + DCTSIZE2) * 2], N0
+    add         VALUES, 8*2
+.PADDING:
+    mov         K, LEN
+    add         K, 7
+    and         K, -8
+    shr         K, 3
+    sub         K, DCTSIZE2/8
+    jz          .EPADDING
+    align       16
+.ZEROLOOP:
+    movdqa      XMMWORD [VALUES + 0], ZERO
+    add         VALUES, 8*2
+    inc         K
+    jnz         .ZEROLOOP
+.EPADDING:
+    sub         VALUES, DCTSIZE2*2
+
+    REDUCE0
+
+    pop         ebp
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+    pop         ecx
+    pop         ebx
+    mov         esp, ebp                ; esp <- aligned ebp
+    pop         esp                     ; esp <- original ebp
+    pop         ebp
+    ret
+
+%undef ZERO
+%undef X0
+%undef X1
+%undef N0
+%undef N1
+%undef AL
+%undef K
+%undef LUT
+%undef T0
+%undef T1
+%undef BLOCK
+%undef VALUES
+%undef LEN
+
+;
+; Prepare data for jsimd_encode_mcu_AC_refine().
+;
+; GLOBAL(int)
+; jsimd_encode_mcu_AC_refine_prepare_sse2(const JCOEF *block,
+;                                         const int *jpeg_natural_order_start,
+;                                         int Sl, int Al, JCOEF *absvalues,
+;                                         size_t *bits)
+;
+; eax + 8 = const JCOEF *block
+; eax + 12 = const int *jpeg_natural_order_start
+; eax + 16 = int Sl
+; eax + 20 = int Al
+; eax + 24 = JCOEF *values
+; eax + 28 = size_t *bits
+
+%define ZERO    xmm7
+%define ONE     xmm5
+%define X0      xmm0
+%define X1      xmm1
+%define N0      xmm2
+%define N1      xmm3
+%define AL      xmm4
+%define K       eax
+%define LENEND  eax
+%define LUT     ebx
+%define T0      ecx
+%define T0w      cx
+%define T1      edx
+%define BLOCK   esi
+%define VALUES  edi
+%define KK      ebp
+
+%define ZEROBITS  INT [esp + 5 * 4]
+%define EOB       INT [esp + 5 * 4 + 4]
+%define LEN       INT [esp + 5 * 4 + 8]
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_encode_mcu_AC_refine_prepare_sse2)
+
+EXTN(jsimd_encode_mcu_AC_refine_prepare_sse2):
+    push        ebp
+    mov         eax, esp                     ; eax = original ebp
+    sub         esp, byte 4
+    and         esp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [esp], eax
+    mov         ebp, esp                     ; ebp = aligned ebp
+    sub         esp, 16
+    push        ebx
+    push        ecx
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+    push        ebp
+
+    pcmpeqw     ONE, ONE
+    psrlw       ONE, 15
+    mov         BLOCK, INT [eax + 8]
+    mov         LUT, INT [eax + 12]
+    mov         VALUES, INT [eax + 24]
+    movd        AL, INT [eax + 20]
+    mov         T0, INT [eax + 28]
+    mov         K,  INT [eax + 16]
+    mov         INT [T0 + 2 * SIZEOF_INT], -1
+    mov         INT [T0 + 3 * SIZEOF_INT], -1
+    mov         ZEROBITS, T0
+    mov         LEN, K
+    pxor        ZERO, ZERO
+    and         K, -16
+    mov         EOB, 0
+    xor         KK, KK
+    shr         K, 4
+    jz          .ELOOPR16
+.BLOOPR16:
+    LOAD16
+    pcmpgtw     N0, X0
+    pcmpgtw     N1, X1
+    paddw       X0, N0
+    paddw       X1, N1
+    pxor        X0, N0
+    pxor        X1, N1
+    psrlw       X0, AL
+    psrlw       X1, AL
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    movdqa      XMMWORD [VALUES + (8) * 2], X1
+    pcmpeqw     X0, ONE
+    pcmpeqw     X1, ONE
+    packsswb    N0, N1
+    packsswb    X0, X1
+    pmovmskb    T0, N0                  ; lsignbits.val16u[k>>4] = _mm_movemask_epi8(neg);
+    mov         T1, ZEROBITS
+    not         T0
+    mov         word [T1 + 2 * SIZEOF_INT + KK], T0w
+    pmovmskb    T1, X0                  ; idx = _mm_movemask_epi8(x1);
+    bsr         T1, T1                  ;  idx = 16 - (__builtin_clz(idx)>>1);
+    jz          .CONTINUER16            ; if (idx) {
+    lea         T1, [T1+KK*8]
+    mov         EOB, T1                 ; EOB = k + idx;
+.CONTINUER16:
+    add         VALUES, 16*2
+    add         LUT, 16*SIZEOF_INT
+    add         KK, 2
+    dec         K
+    jnz         .BLOOPR16
+    test        LEN, 15
+    je          .PADDINGR
+.ELOOPR16:
+    mov         LENEND, LEN
+
+    test        LENEND, 8
+    jz          .TRYR7
+    test        LENEND, 7
+    jz          .TRYR8
+
+    and         LENEND, 7
+    LOAD15
+    pcmpgtw     N0, X0
+    pcmpgtw     N1, X1
+    paddw       X0, N0
+    paddw       X1, N1
+    pxor        X0, N0
+    pxor        X1, N1
+    psrlw       X0, AL
+    psrlw       X1, AL
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    movdqa      XMMWORD [VALUES + (8) * 2], X1
+    pcmpeqw     X0, ONE
+    pcmpeqw     X1, ONE
+    packsswb    N0, N1
+    packsswb    X0, X1
+    pmovmskb    T0, N0                  ; lsignbits.val16u[k>>4] = _mm_movemask_epi8(neg);
+    mov         T1, ZEROBITS
+    not         T0
+    mov         word [T1 + 2 * SIZEOF_INT + KK], T0w
+    pmovmskb    T1, X0                  ; idx = _mm_movemask_epi8(x1);
+    bsr         T1, T1                  ;  idx = 16 - (__builtin_clz(idx)>>1);
+    jz          .CONTINUER15            ; if (idx) {
+    lea         T1, [T1+KK*8]
+    mov         EOB, T1                 ; EOB = k + idx;
+.CONTINUER15:
+    add         VALUES, 16*2
+    jmp         .PADDINGR
+.TRYR8:
+    LOAD8
+
+    pcmpgtw     N0, X0
+    paddw       X0, N0
+    pxor        X0, N0
+    psrlw       X0, AL
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    pcmpeqw     X0, ONE
+    packsswb    N0, ZERO
+    packsswb    X0, ZERO
+    pmovmskb    T0, N0                  ; lsignbits.val16u[k>>4] = _mm_movemask_epi8(neg);
+    mov         T1, ZEROBITS
+    not         T0
+    mov         word [T1 + 2 * SIZEOF_INT + KK], T0w
+    pmovmskb    T1, X0                  ; idx = _mm_movemask_epi8(x1);
+    bsr         T1, T1                  ;  idx = 16 - (__builtin_clz(idx)>>1);
+    jz          .CONTINUER8             ; if (idx) {
+    lea         T1, [T1+KK*8]
+    mov         EOB, T1                 ; EOB = k + idx;
+.CONTINUER8:
+    add         VALUES, 8*2
+    jmp         .PADDINGR
+.TRYR7:
+    and         LENEND, 7
+    LOAD7
+
+    pcmpgtw     N0, X0
+    paddw       X0, N0
+    pxor        X0, N0
+    psrlw       X0, AL
+    movdqa      XMMWORD [VALUES + (0) * 2], X0
+    pcmpeqw     X0, ONE
+    packsswb    N0, ZERO
+    packsswb    X0, ZERO
+    pmovmskb    T0, N0                  ; lsignbits.val16u[k>>4] = _mm_movemask_epi8(neg);
+    mov         T1, ZEROBITS
+    not         T0
+    mov         word [T1 + 2 * SIZEOF_INT + KK], T0w
+    pmovmskb    T1, X0                  ; idx = _mm_movemask_epi8(x1);
+    bsr         T1, T1                  ;  idx = 16 - (__builtin_clz(idx)>>1);
+    jz          .CONTINUER7             ; if (idx) {
+    lea         T1, [T1+KK*8]
+    mov         EOB, T1                 ; EOB = k + idx;
+.CONTINUER7:
+    add         VALUES, 8*2
+.PADDINGR:
+    mov         K, LEN
+    add         K, 7
+    and         K, -8
+    shr         K, 3
+    sub         K, DCTSIZE2/8
+    jz          .EPADDINGR
+    align       16
+.ZEROLOOPR:
+    movdqa      XMMWORD [VALUES + 0], ZERO
+    add         VALUES, 8*2
+    inc         K
+    jnz         .ZEROLOOPR
+.EPADDINGR:
+    sub         VALUES, DCTSIZE2*2
+
+    REDUCE0
+
+    mov         eax, EOB
+
+    pop         ebp
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+    pop         ecx
+    pop         ebx
+    mov         esp, ebp                ; esp <- aligned ebp
+    pop         esp                     ; esp <- original ebp
+    pop         ebp
+    ret
+
+%undef ZERO
+%undef ONE
+%undef X0
+%undef X1
+%undef N0
+%undef N1
+%undef AL
+%undef K
+%undef KK
+%undef EOB
+%undef SIGN
+%undef LUT
+%undef T0
+%undef T1
+%undef BLOCK
+%undef VALUES
+%undef LEN
+%undef LENEND
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,388 @@
+;
+; jcsample.asm - downsampling (AVX2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+;
+; Downsample pixel values of a single component.
+; This version handles the common case of 2:1 horizontal and 1:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v1_downsample_avx2(JDIMENSION image_width, int max_v_samp_factor,
+;                            JDIMENSION v_samp_factor,
+;                            JDIMENSION width_in_blocks, JSAMPARRAY input_data,
+;                            JSAMPARRAY output_data);
+;
+
+%define img_width(b)    (b) + 8         ; JDIMENSION image_width
+%define max_v_samp(b)   (b) + 12        ; int max_v_samp_factor
+%define v_samp(b)       (b) + 16        ; JDIMENSION v_samp_factor
+%define width_blks(b)   (b) + 20        ; JDIMENSION width_in_blocks
+%define input_data(b)   (b) + 24        ; JSAMPARRAY input_data
+%define output_data(b)  (b) + 28        ; JSAMPARRAY output_data
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v1_downsample_avx2)
+
+EXTN(jsimd_h2v1_downsample_avx2):
+    push        ebp
+    mov         ebp, esp
+;   push        ebx                     ; unused
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    mov         ecx, JDIMENSION [width_blks(ebp)]
+    shl         ecx, 3                  ; imul ecx,DCTSIZE (ecx = output_cols)
+    jz          near .return
+
+    mov         edx, JDIMENSION [img_width(ebp)]
+
+    ; -- expand_right_edge
+
+    push        ecx
+    shl         ecx, 1                  ; output_cols * 2
+    sub         ecx, edx
+    jle         short .expand_end
+
+    mov         eax, INT [max_v_samp(ebp)]
+    test        eax, eax
+    jle         short .expand_end
+
+    cld
+    mov         esi, JSAMPARRAY [input_data(ebp)]  ; input_data
+    alignx      16, 7
+.expandloop:
+    push        eax
+    push        ecx
+
+    mov         edi, JSAMPROW [esi]
+    add         edi, edx
+    mov         al, JSAMPLE [edi-1]
+
+    rep stosb
+
+    pop         ecx
+    pop         eax
+
+    add         esi, byte SIZEOF_JSAMPROW
+    dec         eax
+    jg          short .expandloop
+
+.expand_end:
+    pop         ecx                     ; output_cols
+
+    ; -- h2v1_downsample
+
+    mov         eax, JDIMENSION [v_samp(ebp)]  ; rowctr
+    test        eax, eax
+    jle         near .return
+
+    mov         edx, 0x00010000         ; bias pattern
+    vmovd       xmm7, edx
+    vpshufd     xmm7, xmm7, 0x00        ; xmm7={0, 1, 0, 1, 0, 1, 0, 1}
+    vperm2i128  ymm7, ymm7, ymm7, 0     ; ymm7={xmm7, xmm7}
+    vpcmpeqw    ymm6, ymm6, ymm6
+    vpsrlw      ymm6, ymm6, BYTE_BIT    ; ymm6={0xFF 0x00 0xFF 0x00 ..}
+
+    mov         esi, JSAMPARRAY [input_data(ebp)]   ; input_data
+    mov         edi, JSAMPARRAY [output_data(ebp)]  ; output_data
+    alignx      16, 7
+.rowloop:
+    push        ecx
+    push        edi
+    push        esi
+
+    mov         esi, JSAMPROW [esi]     ; inptr
+    mov         edi, JSAMPROW [edi]     ; outptr
+
+    cmp         ecx, byte SIZEOF_YMMWORD
+    jae         short .columnloop
+    alignx      16, 7
+
+.columnloop_r24:
+    ; ecx can possibly be 8, 16, 24
+    cmp         ecx, 24
+    jne         .columnloop_r16
+    vmovdqu     ymm0, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vmovdqu     xmm1, XMMWORD [esi+1*SIZEOF_YMMWORD]
+    mov         ecx, SIZEOF_YMMWORD
+    jmp         short .downsample
+
+.columnloop_r16:
+    cmp         ecx, 16
+    jne         .columnloop_r8
+    vmovdqu     ymm0, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vpxor       ymm1, ymm1, ymm1
+    mov         ecx, SIZEOF_YMMWORD
+    jmp         short .downsample
+
+.columnloop_r8:
+    vmovdqu     xmm0, XMMWORD[esi+0*SIZEOF_YMMWORD]
+    vpxor       ymm1, ymm1, ymm1
+    mov         ecx, SIZEOF_YMMWORD
+    jmp         short .downsample
+    alignx      16, 7
+
+.columnloop:
+    vmovdqu     ymm0, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymm1, YMMWORD [esi+1*SIZEOF_YMMWORD]
+
+.downsample:
+    vpsrlw      ymm2, ymm0, BYTE_BIT
+    vpand       ymm0, ymm0, ymm6
+    vpsrlw      ymm3, ymm1, BYTE_BIT
+    vpand       ymm1, ymm1, ymm6
+
+    vpaddw      ymm0, ymm0, ymm2
+    vpaddw      ymm1, ymm1, ymm3
+    vpaddw      ymm0, ymm0, ymm7
+    vpaddw      ymm1, ymm1, ymm7
+    vpsrlw      ymm0, ymm0, 1
+    vpsrlw      ymm1, ymm1, 1
+
+    vpackuswb   ymm0, ymm0, ymm1
+    vpermq      ymm0, ymm0, 0xd8
+
+    vmovdqu     YMMWORD [edi+0*SIZEOF_YMMWORD], ymm0
+
+    sub         ecx, byte SIZEOF_YMMWORD    ; outcol
+    add         esi, byte 2*SIZEOF_YMMWORD  ; inptr
+    add         edi, byte 1*SIZEOF_YMMWORD  ; outptr
+    cmp         ecx, byte SIZEOF_YMMWORD
+    jae         short .columnloop
+    test        ecx, ecx
+    jnz         near .columnloop_r24
+
+    pop         esi
+    pop         edi
+    pop         ecx
+
+    add         esi, byte SIZEOF_JSAMPROW  ; input_data
+    add         edi, byte SIZEOF_JSAMPROW  ; output_data
+    dec         eax                        ; rowctr
+    jg          near .rowloop
+
+.return:
+    vzeroupper
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+;   pop         ebx                     ; unused
+    pop         ebp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Downsample pixel values of a single component.
+; This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v2_downsample_avx2(JDIMENSION image_width, int max_v_samp_factor,
+;                            JDIMENSION v_samp_factor,
+;                            JDIMENSION width_in_blocks, JSAMPARRAY input_data,
+;                            JSAMPARRAY output_data);
+;
+
+%define img_width(b)    (b) + 8         ; JDIMENSION image_width
+%define max_v_samp(b)   (b) + 12        ; int max_v_samp_factor
+%define v_samp(b)       (b) + 16        ; JDIMENSION v_samp_factor
+%define width_blks(b)   (b) + 20        ; JDIMENSION width_in_blocks
+%define input_data(b)   (b) + 24        ; JSAMPARRAY input_data
+%define output_data(b)  (b) + 28        ; JSAMPARRAY output_data
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v2_downsample_avx2)
+
+EXTN(jsimd_h2v2_downsample_avx2):
+    push        ebp
+    mov         ebp, esp
+;   push        ebx                     ; unused
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    mov         ecx, JDIMENSION [width_blks(ebp)]
+    shl         ecx, 3                  ; imul ecx,DCTSIZE (ecx = output_cols)
+    jz          near .return
+
+    mov         edx, JDIMENSION [img_width(ebp)]
+
+    ; -- expand_right_edge
+
+    push        ecx
+    shl         ecx, 1                  ; output_cols * 2
+    sub         ecx, edx
+    jle         short .expand_end
+
+    mov         eax, INT [max_v_samp(ebp)]
+    test        eax, eax
+    jle         short .expand_end
+
+    cld
+    mov         esi, JSAMPARRAY [input_data(ebp)]  ; input_data
+    alignx      16, 7
+.expandloop:
+    push        eax
+    push        ecx
+
+    mov         edi, JSAMPROW [esi]
+    add         edi, edx
+    mov         al, JSAMPLE [edi-1]
+
+    rep stosb
+
+    pop         ecx
+    pop         eax
+
+    add         esi, byte SIZEOF_JSAMPROW
+    dec         eax
+    jg          short .expandloop
+
+.expand_end:
+    pop         ecx                     ; output_cols
+
+    ; -- h2v2_downsample
+
+    mov         eax, JDIMENSION [v_samp(ebp)]  ; rowctr
+    test        eax, eax
+    jle         near .return
+
+    mov         edx, 0x00020001         ; bias pattern
+    vmovd       xmm7, edx
+    vpcmpeqw    ymm6, ymm6, ymm6
+    vpshufd     xmm7, xmm7, 0x00        ; ymm7={1, 2, 1, 2, 1, 2, 1, 2}
+    vperm2i128  ymm7, ymm7, ymm7, 0
+    vpsrlw      ymm6, ymm6, BYTE_BIT    ; ymm6={0xFF 0x00 0xFF 0x00 ..}
+
+    mov         esi, JSAMPARRAY [input_data(ebp)]   ; input_data
+    mov         edi, JSAMPARRAY [output_data(ebp)]  ; output_data
+    alignx      16, 7
+.rowloop:
+    push        ecx
+    push        edi
+    push        esi
+
+    mov         edx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]  ; inptr0
+    mov         esi, JSAMPROW [esi+1*SIZEOF_JSAMPROW]  ; inptr1
+    mov         edi, JSAMPROW [edi]                    ; outptr
+
+    cmp         ecx, byte SIZEOF_YMMWORD
+    jae         short .columnloop
+    alignx      16, 7
+
+.columnloop_r24:
+    cmp         ecx, 24
+    jne         .columnloop_r16
+    vmovdqu     ymm0, YMMWORD [edx+0*SIZEOF_YMMWORD]
+    vmovdqu     ymm1, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vmovdqu     xmm2, XMMWORD [edx+1*SIZEOF_YMMWORD]
+    vmovdqu     xmm3, XMMWORD [esi+1*SIZEOF_YMMWORD]
+    mov         ecx, SIZEOF_YMMWORD
+    jmp         short .downsample
+
+.columnloop_r16:
+    cmp         ecx, 16
+    jne         .columnloop_r8
+    vmovdqu     ymm0, YMMWORD [edx+0*SIZEOF_YMMWORD]
+    vmovdqu     ymm1, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vpxor       ymm2, ymm2, ymm2
+    vpxor       ymm3, ymm3, ymm3
+    mov         ecx, SIZEOF_YMMWORD
+    jmp         short .downsample
+
+.columnloop_r8:
+    vmovdqu     xmm0, XMMWORD [edx+0*SIZEOF_XMMWORD]
+    vmovdqu     xmm1, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    vpxor       ymm2, ymm2, ymm2
+    vpxor       ymm3, ymm3, ymm3
+    mov         ecx, SIZEOF_YMMWORD
+    jmp         short .downsample
+    alignx      16, 7
+
+.columnloop:
+    vmovdqu     ymm0, YMMWORD [edx+0*SIZEOF_YMMWORD]
+    vmovdqu     ymm1, YMMWORD [esi+0*SIZEOF_YMMWORD]
+    vmovdqu     ymm2, YMMWORD [edx+1*SIZEOF_YMMWORD]
+    vmovdqu     ymm3, YMMWORD [esi+1*SIZEOF_YMMWORD]
+
+.downsample:
+    vpand       ymm4, ymm0, ymm6
+    vpsrlw      ymm0, ymm0, BYTE_BIT
+    vpand       ymm5, ymm1, ymm6
+    vpsrlw      ymm1, ymm1, BYTE_BIT
+    vpaddw      ymm0, ymm0, ymm4
+    vpaddw      ymm1, ymm1, ymm5
+
+    vpand       ymm4, ymm2, ymm6
+    vpsrlw      ymm2, ymm2, BYTE_BIT
+    vpand       ymm5, ymm3, ymm6
+    vpsrlw      ymm3, ymm3, BYTE_BIT
+    vpaddw      ymm2, ymm2, ymm4
+    vpaddw      ymm3, ymm3, ymm5
+
+    vpaddw      ymm0, ymm0, ymm1
+    vpaddw      ymm2, ymm2, ymm3
+    vpaddw      ymm0, ymm0, ymm7
+    vpaddw      ymm2, ymm2, ymm7
+    vpsrlw      ymm0, ymm0, 2
+    vpsrlw      ymm2, ymm2, 2
+
+    vpackuswb   ymm0, ymm0, ymm2
+    vpermq      ymm0, ymm0, 0xd8
+
+    vmovdqu     YMMWORD [edi+0*SIZEOF_YMMWORD], ymm0
+
+    sub         ecx, byte SIZEOF_YMMWORD    ; outcol
+    add         edx, byte 2*SIZEOF_YMMWORD  ; inptr0
+    add         esi, byte 2*SIZEOF_YMMWORD  ; inptr1
+    add         edi, byte 1*SIZEOF_YMMWORD  ; outptr
+    cmp         ecx, byte SIZEOF_YMMWORD
+    jae         near .columnloop
+    test        ecx, ecx
+    jnz         near .columnloop_r24
+
+    pop         esi
+    pop         edi
+    pop         ecx
+
+    add         esi, byte 2*SIZEOF_JSAMPROW  ; input_data
+    add         edi, byte 1*SIZEOF_JSAMPROW  ; output_data
+    dec         eax                          ; rowctr
+    jg          near .rowloop
+
+.return:
+    vzeroupper
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+;   pop         ebx                     ; unused
+    pop         ebp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,324 @@
+;
+; jcsample.asm - downsampling (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+;
+; Downsample pixel values of a single component.
+; This version handles the common case of 2:1 horizontal and 1:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v1_downsample_mmx(JDIMENSION image_width, int max_v_samp_factor,
+;                           JDIMENSION v_samp_factor,
+;                           JDIMENSION width_in_blocks, JSAMPARRAY input_data,
+;                           JSAMPARRAY output_data);
+;
+
+%define img_width(b)    (b) + 8         ; JDIMENSION image_width
+%define max_v_samp(b)   (b) + 12        ; int max_v_samp_factor
+%define v_samp(b)       (b) + 16        ; JDIMENSION v_samp_factor
+%define width_blks(b)   (b) + 20        ; JDIMENSION width_in_blocks
+%define input_data(b)   (b) + 24        ; JSAMPARRAY input_data
+%define output_data(b)  (b) + 28        ; JSAMPARRAY output_data
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v1_downsample_mmx)
+
+EXTN(jsimd_h2v1_downsample_mmx):
+    push        ebp
+    mov         ebp, esp
+;   push        ebx                     ; unused
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    mov         ecx, JDIMENSION [width_blks(ebp)]
+    shl         ecx, 3                  ; imul ecx,DCTSIZE (ecx = output_cols)
+    jz          near .return
+
+    mov         edx, JDIMENSION [img_width(ebp)]
+
+    ; -- expand_right_edge
+
+    push        ecx
+    shl         ecx, 1                  ; output_cols * 2
+    sub         ecx, edx
+    jle         short .expand_end
+
+    mov         eax, INT [max_v_samp(ebp)]
+    test        eax, eax
+    jle         short .expand_end
+
+    cld
+    mov         esi, JSAMPARRAY [input_data(ebp)]  ; input_data
+    alignx      16, 7
+.expandloop:
+    push        eax
+    push        ecx
+
+    mov         edi, JSAMPROW [esi]
+    add         edi, edx
+    mov         al, JSAMPLE [edi-1]
+
+    rep stosb
+
+    pop         ecx
+    pop         eax
+
+    add         esi, byte SIZEOF_JSAMPROW
+    dec         eax
+    jg          short .expandloop
+
+.expand_end:
+    pop         ecx                     ; output_cols
+
+    ; -- h2v1_downsample
+
+    mov         eax, JDIMENSION [v_samp(ebp)]  ; rowctr
+    test        eax, eax
+    jle         near .return
+
+    mov         edx, 0x00010000         ; bias pattern
+    movd        mm7, edx
+    pcmpeqw     mm6, mm6
+    punpckldq   mm7, mm7                ; mm7={0, 1, 0, 1}
+    psrlw       mm6, BYTE_BIT           ; mm6={0xFF 0x00 0xFF 0x00 ..}
+
+    mov         esi, JSAMPARRAY [input_data(ebp)]   ; input_data
+    mov         edi, JSAMPARRAY [output_data(ebp)]  ; output_data
+    alignx      16, 7
+.rowloop:
+    push        ecx
+    push        edi
+    push        esi
+
+    mov         esi, JSAMPROW [esi]     ; inptr
+    mov         edi, JSAMPROW [edi]     ; outptr
+    alignx      16, 7
+.columnloop:
+
+    movq        mm0, MMWORD [esi+0*SIZEOF_MMWORD]
+    movq        mm1, MMWORD [esi+1*SIZEOF_MMWORD]
+    movq        mm2, mm0
+    movq        mm3, mm1
+
+    pand        mm0, mm6
+    psrlw       mm2, BYTE_BIT
+    pand        mm1, mm6
+    psrlw       mm3, BYTE_BIT
+
+    paddw       mm0, mm2
+    paddw       mm1, mm3
+    paddw       mm0, mm7
+    paddw       mm1, mm7
+    psrlw       mm0, 1
+    psrlw       mm1, 1
+
+    packuswb    mm0, mm1
+
+    movq        MMWORD [edi+0*SIZEOF_MMWORD], mm0
+
+    add         esi, byte 2*SIZEOF_MMWORD  ; inptr
+    add         edi, byte 1*SIZEOF_MMWORD  ; outptr
+    sub         ecx, byte SIZEOF_MMWORD    ; outcol
+    jnz         short .columnloop
+
+    pop         esi
+    pop         edi
+    pop         ecx
+
+    add         esi, byte SIZEOF_JSAMPROW  ; input_data
+    add         edi, byte SIZEOF_JSAMPROW  ; output_data
+    dec         eax                        ; rowctr
+    jg          short .rowloop
+
+    emms                                ; empty MMX state
+
+.return:
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+;   pop         ebx                     ; unused
+    pop         ebp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Downsample pixel values of a single component.
+; This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v2_downsample_mmx(JDIMENSION image_width, int max_v_samp_factor,
+;                           JDIMENSION v_samp_factor,
+;                           JDIMENSION width_in_blocks, JSAMPARRAY input_data,
+;                           JSAMPARRAY output_data);
+;
+
+%define img_width(b)    (b) + 8         ; JDIMENSION image_width
+%define max_v_samp(b)   (b) + 12        ; int max_v_samp_factor
+%define v_samp(b)       (b) + 16        ; JDIMENSION v_samp_factor
+%define width_blks(b)   (b) + 20        ; JDIMENSION width_in_blocks
+%define input_data(b)   (b) + 24        ; JSAMPARRAY input_data
+%define output_data(b)  (b) + 28        ; JSAMPARRAY output_data
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v2_downsample_mmx)
+
+EXTN(jsimd_h2v2_downsample_mmx):
+    push        ebp
+    mov         ebp, esp
+;   push        ebx                     ; unused
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    mov         ecx, JDIMENSION [width_blks(ebp)]
+    shl         ecx, 3                  ; imul ecx,DCTSIZE (ecx = output_cols)
+    jz          near .return
+
+    mov         edx, JDIMENSION [img_width(ebp)]
+
+    ; -- expand_right_edge
+
+    push        ecx
+    shl         ecx, 1                  ; output_cols * 2
+    sub         ecx, edx
+    jle         short .expand_end
+
+    mov         eax, INT [max_v_samp(ebp)]
+    test        eax, eax
+    jle         short .expand_end
+
+    cld
+    mov         esi, JSAMPARRAY [input_data(ebp)]  ; input_data
+    alignx      16, 7
+.expandloop:
+    push        eax
+    push        ecx
+
+    mov         edi, JSAMPROW [esi]
+    add         edi, edx
+    mov         al, JSAMPLE [edi-1]
+
+    rep stosb
+
+    pop         ecx
+    pop         eax
+
+    add         esi, byte SIZEOF_JSAMPROW
+    dec         eax
+    jg          short .expandloop
+
+.expand_end:
+    pop         ecx                     ; output_cols
+
+    ; -- h2v2_downsample
+
+    mov         eax, JDIMENSION [v_samp(ebp)]  ; rowctr
+    test        eax, eax
+    jle         near .return
+
+    mov         edx, 0x00020001         ; bias pattern
+    movd        mm7, edx
+    pcmpeqw     mm6, mm6
+    punpckldq   mm7, mm7                ; mm7={1, 2, 1, 2}
+    psrlw       mm6, BYTE_BIT           ; mm6={0xFF 0x00 0xFF 0x00 ..}
+
+    mov         esi, JSAMPARRAY [input_data(ebp)]   ; input_data
+    mov         edi, JSAMPARRAY [output_data(ebp)]  ; output_data
+    alignx      16, 7
+.rowloop:
+    push        ecx
+    push        edi
+    push        esi
+
+    mov         edx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]  ; inptr0
+    mov         esi, JSAMPROW [esi+1*SIZEOF_JSAMPROW]  ; inptr1
+    mov         edi, JSAMPROW [edi]                    ; outptr
+    alignx      16, 7
+.columnloop:
+
+    movq        mm0, MMWORD [edx+0*SIZEOF_MMWORD]
+    movq        mm1, MMWORD [esi+0*SIZEOF_MMWORD]
+    movq        mm2, MMWORD [edx+1*SIZEOF_MMWORD]
+    movq        mm3, MMWORD [esi+1*SIZEOF_MMWORD]
+
+    movq        mm4, mm0
+    movq        mm5, mm1
+    pand        mm0, mm6
+    psrlw       mm4, BYTE_BIT
+    pand        mm1, mm6
+    psrlw       mm5, BYTE_BIT
+    paddw       mm0, mm4
+    paddw       mm1, mm5
+
+    movq        mm4, mm2
+    movq        mm5, mm3
+    pand        mm2, mm6
+    psrlw       mm4, BYTE_BIT
+    pand        mm3, mm6
+    psrlw       mm5, BYTE_BIT
+    paddw       mm2, mm4
+    paddw       mm3, mm5
+
+    paddw       mm0, mm1
+    paddw       mm2, mm3
+    paddw       mm0, mm7
+    paddw       mm2, mm7
+    psrlw       mm0, 2
+    psrlw       mm2, 2
+
+    packuswb    mm0, mm2
+
+    movq        MMWORD [edi+0*SIZEOF_MMWORD], mm0
+
+    add         edx, byte 2*SIZEOF_MMWORD  ; inptr0
+    add         esi, byte 2*SIZEOF_MMWORD  ; inptr1
+    add         edi, byte 1*SIZEOF_MMWORD  ; outptr
+    sub         ecx, byte SIZEOF_MMWORD    ; outcol
+    jnz         near .columnloop
+
+    pop         esi
+    pop         edi
+    pop         ecx
+
+    add         esi, byte 2*SIZEOF_JSAMPROW  ; input_data
+    add         edi, byte 1*SIZEOF_JSAMPROW  ; output_data
+    dec         eax                          ; rowctr
+    jg          near .rowloop
+
+    emms                                ; empty MMX state
+
+.return:
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+;   pop         ebx                     ; unused
+    pop         ebp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,351 @@
+;
+; jcsample.asm - downsampling (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+;
+; Downsample pixel values of a single component.
+; This version handles the common case of 2:1 horizontal and 1:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v1_downsample_sse2(JDIMENSION image_width, int max_v_samp_factor,
+;                            JDIMENSION v_samp_factor,
+;                            JDIMENSION width_in_blocks, JSAMPARRAY input_data,
+;                            JSAMPARRAY output_data);
+;
+
+%define img_width(b)    (b) + 8         ; JDIMENSION image_width
+%define max_v_samp(b)   (b) + 12        ; int max_v_samp_factor
+%define v_samp(b)       (b) + 16        ; JDIMENSION v_samp_factor
+%define width_blks(b)   (b) + 20        ; JDIMENSION width_in_blocks
+%define input_data(b)   (b) + 24        ; JSAMPARRAY input_data
+%define output_data(b)  (b) + 28        ; JSAMPARRAY output_data
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v1_downsample_sse2)
+
+EXTN(jsimd_h2v1_downsample_sse2):
+    push        ebp
+    mov         ebp, esp
+;   push        ebx                     ; unused
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    mov         ecx, JDIMENSION [width_blks(ebp)]
+    shl         ecx, 3                  ; imul ecx,DCTSIZE (ecx = output_cols)
+    jz          near .return
+
+    mov         edx, JDIMENSION [img_width(ebp)]
+
+    ; -- expand_right_edge
+
+    push        ecx
+    shl         ecx, 1                  ; output_cols * 2
+    sub         ecx, edx
+    jle         short .expand_end
+
+    mov         eax, INT [max_v_samp(ebp)]
+    test        eax, eax
+    jle         short .expand_end
+
+    cld
+    mov         esi, JSAMPARRAY [input_data(ebp)]  ; input_data
+    alignx      16, 7
+.expandloop:
+    push        eax
+    push        ecx
+
+    mov         edi, JSAMPROW [esi]
+    add         edi, edx
+    mov         al, JSAMPLE [edi-1]
+
+    rep stosb
+
+    pop         ecx
+    pop         eax
+
+    add         esi, byte SIZEOF_JSAMPROW
+    dec         eax
+    jg          short .expandloop
+
+.expand_end:
+    pop         ecx                     ; output_cols
+
+    ; -- h2v1_downsample
+
+    mov         eax, JDIMENSION [v_samp(ebp)]  ; rowctr
+    test        eax, eax
+    jle         near .return
+
+    mov         edx, 0x00010000         ; bias pattern
+    movd        xmm7, edx
+    pcmpeqw     xmm6, xmm6
+    pshufd      xmm7, xmm7, 0x00        ; xmm7={0, 1, 0, 1, 0, 1, 0, 1}
+    psrlw       xmm6, BYTE_BIT          ; xmm6={0xFF 0x00 0xFF 0x00 ..}
+
+    mov         esi, JSAMPARRAY [input_data(ebp)]   ; input_data
+    mov         edi, JSAMPARRAY [output_data(ebp)]  ; output_data
+    alignx      16, 7
+.rowloop:
+    push        ecx
+    push        edi
+    push        esi
+
+    mov         esi, JSAMPROW [esi]     ; inptr
+    mov         edi, JSAMPROW [edi]     ; outptr
+
+    cmp         ecx, byte SIZEOF_XMMWORD
+    jae         short .columnloop
+    alignx      16, 7
+
+.columnloop_r8:
+    movdqa      xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    pxor        xmm1, xmm1
+    mov         ecx, SIZEOF_XMMWORD
+    jmp         short .downsample
+    alignx      16, 7
+
+.columnloop:
+    movdqa      xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    movdqa      xmm1, XMMWORD [esi+1*SIZEOF_XMMWORD]
+
+.downsample:
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm1
+
+    pand        xmm0, xmm6
+    psrlw       xmm2, BYTE_BIT
+    pand        xmm1, xmm6
+    psrlw       xmm3, BYTE_BIT
+
+    paddw       xmm0, xmm2
+    paddw       xmm1, xmm3
+    paddw       xmm0, xmm7
+    paddw       xmm1, xmm7
+    psrlw       xmm0, 1
+    psrlw       xmm1, 1
+
+    packuswb    xmm0, xmm1
+
+    movdqa      XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0
+
+    sub         ecx, byte SIZEOF_XMMWORD    ; outcol
+    add         esi, byte 2*SIZEOF_XMMWORD  ; inptr
+    add         edi, byte 1*SIZEOF_XMMWORD  ; outptr
+    cmp         ecx, byte SIZEOF_XMMWORD
+    jae         short .columnloop
+    test        ecx, ecx
+    jnz         short .columnloop_r8
+
+    pop         esi
+    pop         edi
+    pop         ecx
+
+    add         esi, byte SIZEOF_JSAMPROW  ; input_data
+    add         edi, byte SIZEOF_JSAMPROW  ; output_data
+    dec         eax                        ; rowctr
+    jg          near .rowloop
+
+.return:
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+;   pop         ebx                     ; unused
+    pop         ebp
+    ret
+
+; --------------------------------------------------------------------------
+;
+; Downsample pixel values of a single component.
+; This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+; without smoothing.
+;
+; GLOBAL(void)
+; jsimd_h2v2_downsample_sse2(JDIMENSION image_width, int max_v_samp_factor,
+;                            JDIMENSION v_samp_factor,
+;                            JDIMENSION width_in_blocks, JSAMPARRAY input_data,
+;                            JSAMPARRAY output_data);
+;
+
+%define img_width(b)    (b) + 8         ; JDIMENSION image_width
+%define max_v_samp(b)   (b) + 12        ; int max_v_samp_factor
+%define v_samp(b)       (b) + 16        ; JDIMENSION v_samp_factor
+%define width_blks(b)   (b) + 20        ; JDIMENSION width_in_blocks
+%define input_data(b)   (b) + 24        ; JSAMPARRAY input_data
+%define output_data(b)  (b) + 28        ; JSAMPARRAY output_data
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_h2v2_downsample_sse2)
+
+EXTN(jsimd_h2v2_downsample_sse2):
+    push        ebp
+    mov         ebp, esp
+;   push        ebx                     ; unused
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    mov         ecx, JDIMENSION [width_blks(ebp)]
+    shl         ecx, 3                  ; imul ecx,DCTSIZE (ecx = output_cols)
+    jz          near .return
+
+    mov         edx, JDIMENSION [img_width(ebp)]
+
+    ; -- expand_right_edge
+
+    push        ecx
+    shl         ecx, 1                  ; output_cols * 2
+    sub         ecx, edx
+    jle         short .expand_end
+
+    mov         eax, INT [max_v_samp(ebp)]
+    test        eax, eax
+    jle         short .expand_end
+
+    cld
+    mov         esi, JSAMPARRAY [input_data(ebp)]  ; input_data
+    alignx      16, 7
+.expandloop:
+    push        eax
+    push        ecx
+
+    mov         edi, JSAMPROW [esi]
+    add         edi, edx
+    mov         al, JSAMPLE [edi-1]
+
+    rep stosb
+
+    pop         ecx
+    pop         eax
+
+    add         esi, byte SIZEOF_JSAMPROW
+    dec         eax
+    jg          short .expandloop
+
+.expand_end:
+    pop         ecx                     ; output_cols
+
+    ; -- h2v2_downsample
+
+    mov         eax, JDIMENSION [v_samp(ebp)]  ; rowctr
+    test        eax, eax
+    jle         near .return
+
+    mov         edx, 0x00020001         ; bias pattern
+    movd        xmm7, edx
+    pcmpeqw     xmm6, xmm6
+    pshufd      xmm7, xmm7, 0x00        ; xmm7={1, 2, 1, 2, 1, 2, 1, 2}
+    psrlw       xmm6, BYTE_BIT          ; xmm6={0xFF 0x00 0xFF 0x00 ..}
+
+    mov         esi, JSAMPARRAY [input_data(ebp)]   ; input_data
+    mov         edi, JSAMPARRAY [output_data(ebp)]  ; output_data
+    alignx      16, 7
+.rowloop:
+    push        ecx
+    push        edi
+    push        esi
+
+    mov         edx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]  ; inptr0
+    mov         esi, JSAMPROW [esi+1*SIZEOF_JSAMPROW]  ; inptr1
+    mov         edi, JSAMPROW [edi]                    ; outptr
+
+    cmp         ecx, byte SIZEOF_XMMWORD
+    jae         short .columnloop
+    alignx      16, 7
+
+.columnloop_r8:
+    movdqa      xmm0, XMMWORD [edx+0*SIZEOF_XMMWORD]
+    movdqa      xmm1, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    pxor        xmm2, xmm2
+    pxor        xmm3, xmm3
+    mov         ecx, SIZEOF_XMMWORD
+    jmp         short .downsample
+    alignx      16, 7
+
+.columnloop:
+    movdqa      xmm0, XMMWORD [edx+0*SIZEOF_XMMWORD]
+    movdqa      xmm1, XMMWORD [esi+0*SIZEOF_XMMWORD]
+    movdqa      xmm2, XMMWORD [edx+1*SIZEOF_XMMWORD]
+    movdqa      xmm3, XMMWORD [esi+1*SIZEOF_XMMWORD]
+
+.downsample:
+    movdqa      xmm4, xmm0
+    movdqa      xmm5, xmm1
+    pand        xmm0, xmm6
+    psrlw       xmm4, BYTE_BIT
+    pand        xmm1, xmm6
+    psrlw       xmm5, BYTE_BIT
+    paddw       xmm0, xmm4
+    paddw       xmm1, xmm5
+
+    movdqa      xmm4, xmm2
+    movdqa      xmm5, xmm3
+    pand        xmm2, xmm6
+    psrlw       xmm4, BYTE_BIT
+    pand        xmm3, xmm6
+    psrlw       xmm5, BYTE_BIT
+    paddw       xmm2, xmm4
+    paddw       xmm3, xmm5
+
+    paddw       xmm0, xmm1
+    paddw       xmm2, xmm3
+    paddw       xmm0, xmm7
+    paddw       xmm2, xmm7
+    psrlw       xmm0, 2
+    psrlw       xmm2, 2
+
+    packuswb    xmm0, xmm2
+
+    movdqa      XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0
+
+    sub         ecx, byte SIZEOF_XMMWORD    ; outcol
+    add         edx, byte 2*SIZEOF_XMMWORD  ; inptr0
+    add         esi, byte 2*SIZEOF_XMMWORD  ; inptr1
+    add         edi, byte 1*SIZEOF_XMMWORD  ; outptr
+    cmp         ecx, byte SIZEOF_XMMWORD
+    jae         near .columnloop
+    test        ecx, ecx
+    jnz         near .columnloop_r8
+
+    pop         esi
+    pop         edi
+    pop         ecx
+
+    add         esi, byte 2*SIZEOF_JSAMPROW  ; input_data
+    add         edi, byte 1*SIZEOF_JSAMPROW  ; output_data
+    dec         eax                          ; rowctr
+    jg          near .rowloop
+
+.return:
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+;   pop         ebx                     ; unused
+    pop         ebp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,515 @@
+;
+; jdcolext.asm - colorspace conversion (AVX2)
+;
+; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2012, 2016, D. R. Commander.
+; Copyright (C) 2015, Intel Corporation.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_ycc_rgb_convert_avx2(JDIMENSION out_width, JSAMPIMAGE input_buf,
+;                            JDIMENSION input_row, JSAMPARRAY output_buf,
+;                            int num_rows)
+;
+
+%define out_width(b)   (b) + 8          ; JDIMENSION out_width
+%define input_buf(b)   (b) + 12         ; JSAMPIMAGE input_buf
+%define input_row(b)   (b) + 16         ; JDIMENSION input_row
+%define output_buf(b)  (b) + 20         ; JSAMPARRAY output_buf
+%define num_rows(b)    (b) + 24         ; int num_rows
+
+%define original_ebp   ebp + 0
+%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_YMMWORD
+                                        ; ymmword wk[WK_NUM]
+%define WK_NUM         2
+%define gotptr         wk(0) - SIZEOF_POINTER  ; void * gotptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_avx2)
+
+EXTN(jsimd_ycc_rgb_convert_avx2):
+    push        ebp
+    mov         eax, esp                     ; eax = original ebp
+    sub         esp, byte 4
+    and         esp, byte (-SIZEOF_YMMWORD)  ; align to 256 bits
+    mov         [esp], eax
+    mov         ebp, esp                     ; ebp = aligned ebp
+    lea         esp, [wk(0)]
+    pushpic     eax                     ; make a room for GOT address
+    push        ebx
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    get_GOT     ebx                     ; get GOT address
+    movpic      POINTER [gotptr], ebx   ; save GOT address
+
+    mov         ecx, JDIMENSION [out_width(eax)]  ; num_cols
+    test        ecx, ecx
+    jz          near .return
+
+    push        ecx
+
+    mov         edi, JSAMPIMAGE [input_buf(eax)]
+    mov         ecx, JDIMENSION [input_row(eax)]
+    mov         esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
+    mov         ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
+    mov         edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
+    lea         esi, [esi+ecx*SIZEOF_JSAMPROW]
+    lea         ebx, [ebx+ecx*SIZEOF_JSAMPROW]
+    lea         edx, [edx+ecx*SIZEOF_JSAMPROW]
+
+    pop         ecx
+
+    mov         edi, JSAMPARRAY [output_buf(eax)]
+    mov         eax, INT [num_rows(eax)]
+    test        eax, eax
+    jle         near .return
+    alignx      16, 7
+.rowloop:
+    push        eax
+    push        edi
+    push        edx
+    push        ebx
+    push        esi
+    push        ecx                     ; col
+
+    mov         esi, JSAMPROW [esi]     ; inptr0
+    mov         ebx, JSAMPROW [ebx]     ; inptr1
+    mov         edx, JSAMPROW [edx]     ; inptr2
+    mov         edi, JSAMPROW [edi]     ; outptr
+    movpic      eax, POINTER [gotptr]   ; load GOT address (eax)
+    alignx      16, 7
+.columnloop:
+
+    vmovdqu     ymm5, YMMWORD [ebx]     ; ymm5=Cb(0123456789ABCDEFGHIJKLMNOPQRSTUV)
+    vmovdqu     ymm1, YMMWORD [edx]     ; ymm1=Cr(0123456789ABCDEFGHIJKLMNOPQRSTUV)
+
+    vpcmpeqw    ymm0, ymm0, ymm0
+    vpcmpeqw    ymm7, ymm7, ymm7
+    vpsrlw      ymm0, ymm0, BYTE_BIT    ; ymm0={0xFF 0x00 0xFF 0x00 ..}
+    vpsllw      ymm7, ymm7, 7           ; ymm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+
+    vpand       ymm4, ymm0, ymm5        ; ymm4=Cb(02468ACEGIKMOQSU)=CbE
+    vpsrlw      ymm5, ymm5, BYTE_BIT    ; ymm5=Cb(13579BDFHJLNPRTV)=CbO
+    vpand       ymm0, ymm0, ymm1        ; ymm0=Cr(02468ACEGIKMOQSU)=CrE
+    vpsrlw      ymm1, ymm1, BYTE_BIT    ; ymm1=Cr(13579BDFHJLNPRTV)=CrO
+
+    vpaddw      ymm2, ymm4, ymm7
+    vpaddw      ymm3, ymm5, ymm7
+    vpaddw      ymm6, ymm0, ymm7
+    vpaddw      ymm7, ymm1, ymm7
+
+    ; (Original)
+    ; R = Y                + 1.40200 * Cr
+    ; G = Y - 0.34414 * Cb - 0.71414 * Cr
+    ; B = Y + 1.77200 * Cb
+    ;
+    ; (This implementation)
+    ; R = Y                + 0.40200 * Cr + Cr
+    ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+    ; B = Y - 0.22800 * Cb + Cb + Cb
+
+    vpaddw      ymm4, ymm2, ymm2                     ; ymm4=2*CbE
+    vpaddw      ymm5, ymm3, ymm3                     ; ymm5=2*CbO
+    vpaddw      ymm0, ymm6, ymm6                     ; ymm0=2*CrE
+    vpaddw      ymm1, ymm7, ymm7                     ; ymm1=2*CrO
+
+    vpmulhw     ymm4, ymm4, [GOTOFF(eax,PW_MF0228)]  ; ymm4=(2*CbE * -FIX(0.22800))
+    vpmulhw     ymm5, ymm5, [GOTOFF(eax,PW_MF0228)]  ; ymm5=(2*CbO * -FIX(0.22800))
+    vpmulhw     ymm0, ymm0, [GOTOFF(eax,PW_F0402)]   ; ymm0=(2*CrE * FIX(0.40200))
+    vpmulhw     ymm1, ymm1, [GOTOFF(eax,PW_F0402)]   ; ymm1=(2*CrO * FIX(0.40200))
+
+    vpaddw      ymm4, ymm4, [GOTOFF(eax,PW_ONE)]
+    vpaddw      ymm5, ymm5, [GOTOFF(eax,PW_ONE)]
+    vpsraw      ymm4, ymm4, 1                        ; ymm4=(CbE * -FIX(0.22800))
+    vpsraw      ymm5, ymm5, 1                        ; ymm5=(CbO * -FIX(0.22800))
+    vpaddw      ymm0, ymm0, [GOTOFF(eax,PW_ONE)]
+    vpaddw      ymm1, ymm1, [GOTOFF(eax,PW_ONE)]
+    vpsraw      ymm0, ymm0, 1                        ; ymm0=(CrE * FIX(0.40200))
+    vpsraw      ymm1, ymm1, 1                        ; ymm1=(CrO * FIX(0.40200))
+
+    vpaddw      ymm4, ymm4, ymm2
+    vpaddw      ymm5, ymm5, ymm3
+    vpaddw      ymm4, ymm4, ymm2                     ; ymm4=(CbE * FIX(1.77200))=(B-Y)E
+    vpaddw      ymm5, ymm5, ymm3                     ; ymm5=(CbO * FIX(1.77200))=(B-Y)O
+    vpaddw      ymm0, ymm0, ymm6                     ; ymm0=(CrE * FIX(1.40200))=(R-Y)E
+    vpaddw      ymm1, ymm1, ymm7                     ; ymm1=(CrO * FIX(1.40200))=(R-Y)O
+
+    vmovdqa     YMMWORD [wk(0)], ymm4                ; wk(0)=(B-Y)E
+    vmovdqa     YMMWORD [wk(1)], ymm5                ; wk(1)=(B-Y)O
+
+    vpunpckhwd  ymm4, ymm2, ymm6
+    vpunpcklwd  ymm2, ymm2, ymm6
+    vpmaddwd    ymm2, ymm2, [GOTOFF(eax,PW_MF0344_F0285)]
+    vpmaddwd    ymm4, ymm4, [GOTOFF(eax,PW_MF0344_F0285)]
+    vpunpckhwd  ymm5, ymm3, ymm7
+    vpunpcklwd  ymm3, ymm3, ymm7
+    vpmaddwd    ymm3, ymm3, [GOTOFF(eax,PW_MF0344_F0285)]
+    vpmaddwd    ymm5, ymm5, [GOTOFF(eax,PW_MF0344_F0285)]
+
+    vpaddd      ymm2, ymm2, [GOTOFF(eax,PD_ONEHALF)]
+    vpaddd      ymm4, ymm4, [GOTOFF(eax,PD_ONEHALF)]
+    vpsrad      ymm2, ymm2, SCALEBITS
+    vpsrad      ymm4, ymm4, SCALEBITS
+    vpaddd      ymm3, ymm3, [GOTOFF(eax,PD_ONEHALF)]
+    vpaddd      ymm5, ymm5, [GOTOFF(eax,PD_ONEHALF)]
+    vpsrad      ymm3, ymm3, SCALEBITS
+    vpsrad      ymm5, ymm5, SCALEBITS
+
+    vpackssdw   ymm2, ymm2, ymm4             ; ymm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
+    vpackssdw   ymm3, ymm3, ymm5             ; ymm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
+    vpsubw      ymm2, ymm2, ymm6             ; ymm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
+    vpsubw      ymm3, ymm3, ymm7             ; ymm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
+
+    vmovdqu     ymm5, YMMWORD [esi]          ; ymm5=Y(0123456789ABCDEFGHIJKLMNOPQRSTUV)
+
+    vpcmpeqw    ymm4, ymm4, ymm4
+    vpsrlw      ymm4, ymm4, BYTE_BIT         ; ymm4={0xFF 0x00 0xFF 0x00 ..}
+    vpand       ymm4, ymm4, ymm5             ; ymm4=Y(02468ACEGIKMOQSU)=YE
+    vpsrlw      ymm5, ymm5, BYTE_BIT         ; ymm5=Y(13579BDFHJLNPRTV)=YO
+
+    vpaddw      ymm0, ymm0, ymm4             ; ymm0=((R-Y)E+YE)=RE=R(02468ACEGIKMOQSU)
+    vpaddw      ymm1, ymm1, ymm5             ; ymm1=((R-Y)O+YO)=RO=R(13579BDFHJLNPRTV)
+    vpackuswb   ymm0, ymm0, ymm0             ; ymm0=R(02468ACE********GIKMOQSU********)
+    vpackuswb   ymm1, ymm1, ymm1             ; ymm1=R(13579BDF********HJLNPRTV********)
+
+    vpaddw      ymm2, ymm2, ymm4             ; ymm2=((G-Y)E+YE)=GE=G(02468ACEGIKMOQSU)
+    vpaddw      ymm3, ymm3, ymm5             ; ymm3=((G-Y)O+YO)=GO=G(13579BDFHJLNPRTV)
+    vpackuswb   ymm2, ymm2, ymm2             ; ymm2=G(02468ACE********GIKMOQSU********)
+    vpackuswb   ymm3, ymm3, ymm3             ; ymm3=G(13579BDF********HJLNPRTV********)
+
+    vpaddw      ymm4, ymm4, YMMWORD [wk(0)]  ; ymm4=(YE+(B-Y)E)=BE=B(02468ACEGIKMOQSU)
+    vpaddw      ymm5, ymm5, YMMWORD [wk(1)]  ; ymm5=(YO+(B-Y)O)=BO=B(13579BDFHJLNPRTV)
+    vpackuswb   ymm4, ymm4, ymm4             ; ymm4=B(02468ACE********GIKMOQSU********)
+    vpackuswb   ymm5, ymm5, ymm5             ; ymm5=B(13579BDF********HJLNPRTV********)
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+    ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **)
+    ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **)
+    ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **)
+    ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **)
+    ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **)
+    ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **)
+    ; ymmG=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **)
+    ; ymmH=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **)
+
+    vpunpcklbw  ymmA, ymmA, ymmC        ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E
+                                        ;       0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U)
+    vpunpcklbw  ymmE, ymmE, ymmB        ; ymmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F
+                                        ;       2G 0H 2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V)
+    vpunpcklbw  ymmD, ymmD, ymmF        ; ymmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F
+                                        ;       1H 2H 1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V)
+
+    vpsrldq     ymmH, ymmA, 2           ; ymmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E 0G 1G
+                                        ;       0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U -- --)
+    vpunpckhwd  ymmG, ymmA, ymmE        ; ymmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F
+                                        ;       0O 1O 2O 0P 0Q 1Q 2Q 0R 0S 1S 2S 0T 0U 1U 2U 0V)
+    vpunpcklwd  ymmA, ymmA, ymmE        ; ymmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07
+                                        ;       0G 1G 2G 0H 0I 1I 2I 0J 0K 1K 2K 0L 0M 1M 2M 0N)
+
+    vpsrldq     ymmE, ymmE, 2           ; ymmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F 2G 0H
+                                        ;       2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V -- --)
+
+    vpsrldq     ymmB, ymmD, 2           ; ymmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F 1H 2H
+                                        ;       1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V -- --)
+    vpunpckhwd  ymmC, ymmD, ymmH        ; ymmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F 0G 1G
+                                        ;       1P 2P 0Q 1Q 1R 2R 0S 1S 1T 2T 0U 1U 1V 2V -- --)
+    vpunpcklwd  ymmD, ymmD, ymmH        ; ymmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18
+                                        ;       1H 2H 0I 1I 1J 2J 0K 1K 1L 2L 0M 1M 1N 2N 0O 1O)
+
+    vpunpckhwd  ymmF, ymmE, ymmB        ; ymmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F 2G 0H 1H 2H
+                                        ;       2Q 0R 1R 2R 2S 0T 1T 2T 2U 0V 1V 2V -- -- -- --)
+    vpunpcklwd  ymmE, ymmE, ymmB        ; ymmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29
+                                        ;       2I 0J 1J 2J 2K 0L 1L 2L 2M 0N 1N 2N 2O 0P 1P 2P)
+
+    vpshufd     ymmH, ymmA, 0x4E        ; ymmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03
+                                        ;       0K 1K 2K 0L 0M 1M 2M 0N 0G 1G 2G 0H 0I 1I 2I 0J)
+    vpunpckldq  ymmA, ymmA, ymmD        ; ymmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14
+                                        ;       0G 1G 2G 0H 1H 2H 0I 1I 0I 1I 2I 0J 1J 2J 0K 1K)
+    vpunpckhdq  ymmD, ymmD, ymmE        ; ymmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29
+                                        ;       1L 2L 0M 1M 2M 0N 1N 2N 1N 2N 0O 1O 2O 0P 1P 2P)
+    vpunpckldq  ymmE, ymmE, ymmH        ; ymmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07
+                                        ;       2I 0J 1J 2J 0K 1K 2K 0L 2K 0L 1L 2L 0M 1M 2M 0N)
+
+    vpshufd     ymmH, ymmG, 0x4E        ; ymmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B
+                                        ;       0S 1S 2S 0T 0U 1U 2U 0V 0O 1O 2O 0P 0Q 1Q 2Q 0R)
+    vpunpckldq  ymmG, ymmG, ymmC        ; ymmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C
+                                        ;       0O 1O 2O 0P 1P 2P 0Q 1Q 0Q 1Q 2Q 0R 1R 2R 0S 1S)
+    vpunpckhdq  ymmC, ymmC, ymmF        ; ymmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F 0G 1G 2G 0H 1H 2H
+                                        ;       1T 2T 0U 1U 2U 0V 1V 2V 1V 2V -- -- -- -- -- --)
+    vpunpckldq  ymmF, ymmF, ymmH        ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F
+                                        ;       2Q 0R 1R 2R 0S 1S 2S 0T 2S 0T 1T 2T 0U 1U 2U 0V)
+
+    vpunpcklqdq ymmH, ymmA, ymmE        ; ymmH=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+                                        ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    vpunpcklqdq ymmG, ymmD, ymmG        ; ymmG=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A
+                                        ;       1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q)
+    vpunpcklqdq ymmC, ymmF, ymmC        ; ymmC=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+                                        ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+
+    vperm2i128  ymmA, ymmH, ymmG, 0x20  ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
+                                        ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    vperm2i128  ymmD, ymmC, ymmH, 0x30  ; ymmD=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
+                                        ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
+    vperm2i128  ymmF, ymmG, ymmC, 0x31  ; ymmF=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
+                                        ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
+
+    cmp         ecx, byte SIZEOF_YMMWORD
+    jb          short .column_st64
+
+    test        edi, SIZEOF_YMMWORD-1
+    jnz         short .out1
+    ; --(aligned)-------------------
+    vmovntdq    YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA
+    vmovntdq    YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD
+    vmovntdq    YMMWORD [edi+2*SIZEOF_YMMWORD], ymmF
+    jmp         short .out0
+.out1:  ; --(unaligned)-----------------
+    vmovdqu     YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA
+    vmovdqu     YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD
+    vmovdqu     YMMWORD [edi+2*SIZEOF_YMMWORD], ymmF
+.out0:
+    add         edi, byte RGB_PIXELSIZE*SIZEOF_YMMWORD  ; outptr
+    sub         ecx, byte SIZEOF_YMMWORD
+    jz          near .nextrow
+
+    add         esi, byte SIZEOF_YMMWORD  ; inptr0
+    add         ebx, byte SIZEOF_YMMWORD  ; inptr1
+    add         edx, byte SIZEOF_YMMWORD  ; inptr2
+    jmp         near .columnloop
+    alignx      16, 7
+
+.column_st64:
+    lea         ecx, [ecx+ecx*2]            ; imul ecx, RGB_PIXELSIZE
+    cmp         ecx, byte 2*SIZEOF_YMMWORD
+    jb          short .column_st32
+    vmovdqu     YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA
+    vmovdqu     YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD
+    add         edi, byte 2*SIZEOF_YMMWORD  ; outptr
+    vmovdqa     ymmA, ymmF
+    sub         ecx, byte 2*SIZEOF_YMMWORD
+    jmp         short .column_st31
+.column_st32:
+    cmp         ecx, byte SIZEOF_YMMWORD
+    jb          short .column_st31
+    vmovdqu     YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA
+    add         edi, byte SIZEOF_YMMWORD    ; outptr
+    vmovdqa     ymmA, ymmD
+    sub         ecx, byte SIZEOF_YMMWORD
+    jmp         short .column_st31
+.column_st31:
+    cmp         ecx, byte SIZEOF_XMMWORD
+    jb          short .column_st15
+    vmovdqu     XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+    add         edi, byte SIZEOF_XMMWORD    ; outptr
+    vperm2i128  ymmA, ymmA, ymmA, 1
+    sub         ecx, byte SIZEOF_XMMWORD
+.column_st15:
+    ; Store the lower 8 bytes of xmmA to the output when it has enough
+    ; space.
+    cmp         ecx, byte SIZEOF_MMWORD
+    jb          short .column_st7
+    vmovq       XMM_MMWORD [edi], xmmA
+    add         edi, byte SIZEOF_MMWORD
+    sub         ecx, byte SIZEOF_MMWORD
+    vpsrldq     xmmA, xmmA, SIZEOF_MMWORD
+.column_st7:
+    ; Store the lower 4 bytes of xmmA to the output when it has enough
+    ; space.
+    cmp         ecx, byte SIZEOF_DWORD
+    jb          short .column_st3
+    vmovd       XMM_DWORD [edi], xmmA
+    add         edi, byte SIZEOF_DWORD
+    sub         ecx, byte SIZEOF_DWORD
+    vpsrldq     xmmA, xmmA, SIZEOF_DWORD
+.column_st3:
+    ; Store the lower 2 bytes of eax to the output when it has enough
+    ; space.
+    vmovd       eax, xmmA
+    cmp         ecx, byte SIZEOF_WORD
+    jb          short .column_st1
+    mov         word [edi], ax
+    add         edi, byte SIZEOF_WORD
+    sub         ecx, byte SIZEOF_WORD
+    shr         eax, 16
+.column_st1:
+    ; Store the lower 1 byte of eax to the output when it has enough
+    ; space.
+    test        ecx, ecx
+    jz          short .nextrow
+    mov         byte [edi], al
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+    vpcmpeqb    ymm6, ymm6, ymm6        ; ymm6=XE=X(02468ACE********GIKMOQSU********)
+    vpcmpeqb    ymm7, ymm7, ymm7        ; ymm7=XO=X(13579BDF********HJLNPRTV********)
+%else
+    vpxor       ymm6, ymm6, ymm6        ; ymm6=XE=X(02468ACE********GIKMOQSU********)
+    vpxor       ymm7, ymm7, ymm7        ; ymm7=XO=X(13579BDF********HJLNPRTV********)
+%endif
+    ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **)
+    ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **)
+    ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **)
+    ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **)
+    ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **)
+    ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **)
+    ; ymmG=(30 32 34 36 38 3A 3C 3E ** 3G 3I 3K 3M 3O 3Q 3S 3U **)
+    ; ymmH=(31 33 35 37 39 3B 3D 3F ** 3H 3J 3L 3N 3P 3R 3T 3V **)
+
+    vpunpcklbw  ymmA, ymmA, ymmC        ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E
+                                        ;       0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U)
+    vpunpcklbw  ymmE, ymmE, ymmG        ; ymmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E
+                                        ;       2G 3G 2I 3I 2K 3K 2M 3M 2O 3O 2Q 3Q 2S 3S 2U 3U)
+    vpunpcklbw  ymmB, ymmB, ymmD        ; ymmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F
+                                        ;       0H 1H 0J 1J 0L 1L 0N 1N 0P 1P 0R 1R 0T 1T 0V 1V)
+    vpunpcklbw  ymmF, ymmF, ymmH        ; ymmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F
+                                        ;       2H 3H 2J 3J 2L 3L 2N 3N 2P 3P 2R 3R 2T 3T 2V 3V)
+
+    vpunpckhwd  ymmC, ymmA, ymmE        ; ymmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E
+                                        ;       0O 1O 2O 3O 0Q 1Q 2Q 3Q 0S 1S 2S 3S 0U 1U 2U 3U)
+    vpunpcklwd  ymmA, ymmA, ymmE        ; ymmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36
+                                        ;       0G 1G 2G 3G 0I 1I 2I 3I 0K 1K 2K 3K 0M 1M 2M 3M)
+    vpunpckhwd  ymmG, ymmB, ymmF        ; ymmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F
+                                        ;       0P 1P 2P 3P 0R 1R 2R 3R 0T 1T 2T 3T 0V 1V 2V 3V)
+    vpunpcklwd  ymmB, ymmB, ymmF        ; ymmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37
+                                        ;       0H 1H 2H 3H 0J 1J 2J 3J 0L 1L 2L 3L 0N 1N 2N 3N)
+
+    vpunpckhdq  ymmE, ymmA, ymmB        ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+                                        ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+    vpunpckldq  ymmB, ymmA, ymmB        ; ymmB=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+                                        ;       0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J)
+    vpunpckhdq  ymmF, ymmC, ymmG        ; ymmF=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F
+                                        ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+    vpunpckldq  ymmG, ymmC, ymmG        ; ymmG=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+                                        ;       0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R)
+
+    vperm2i128  ymmA, ymmB, ymmE, 0x20  ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+                                        ;       04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    vperm2i128  ymmD, ymmG, ymmF, 0x20  ; ymmD=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
+                                        ;       0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+    vperm2i128  ymmC, ymmB, ymmE, 0x31  ; ymmC=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J
+                                        ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
+    vperm2i128  ymmH, ymmG, ymmF, 0x31  ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R
+                                        ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
+
+    cmp         ecx, byte SIZEOF_YMMWORD
+    jb          short .column_st64
+
+    test        edi, SIZEOF_YMMWORD-1
+    jnz         short .out1
+    ; --(aligned)-------------------
+    vmovntdq    YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA
+    vmovntdq    YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD
+    vmovntdq    YMMWORD [edi+2*SIZEOF_YMMWORD], ymmC
+    vmovntdq    YMMWORD [edi+3*SIZEOF_YMMWORD], ymmH
+    jmp         short .out0
+.out1:  ; --(unaligned)-----------------
+    vmovdqu     YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA
+    vmovdqu     YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD
+    vmovdqu     YMMWORD [edi+2*SIZEOF_YMMWORD], ymmC
+    vmovdqu     YMMWORD [edi+3*SIZEOF_YMMWORD], ymmH
+.out0:
+    add         edi, RGB_PIXELSIZE*SIZEOF_YMMWORD  ; outptr
+    sub         ecx, byte SIZEOF_YMMWORD
+    jz          near .nextrow
+
+    add         esi, byte SIZEOF_YMMWORD  ; inptr0
+    add         ebx, byte SIZEOF_YMMWORD  ; inptr1
+    add         edx, byte SIZEOF_YMMWORD  ; inptr2
+    jmp         near .columnloop
+    alignx      16, 7
+
+.column_st64:
+    cmp         ecx, byte SIZEOF_YMMWORD/2
+    jb          short .column_st32
+    vmovdqu     YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA
+    vmovdqu     YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD
+    add         edi, byte 2*SIZEOF_YMMWORD  ; outptr
+    vmovdqa     ymmA, ymmC
+    vmovdqa     ymmD, ymmH
+    sub         ecx, byte SIZEOF_YMMWORD/2
+.column_st32:
+    cmp         ecx, byte SIZEOF_YMMWORD/4
+    jb          short .column_st16
+    vmovdqu     YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA
+    add         edi, byte SIZEOF_YMMWORD    ; outptr
+    vmovdqa     ymmA, ymmD
+    sub         ecx, byte SIZEOF_YMMWORD/4
+.column_st16:
+    cmp         ecx, byte SIZEOF_YMMWORD/8
+    jb          short .column_st15
+    vmovdqu     XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+    vperm2i128  ymmA, ymmA, ymmA, 1
+    add         edi, byte SIZEOF_XMMWORD    ; outptr
+    sub         ecx, byte SIZEOF_YMMWORD/8
+.column_st15:
+    ; Store two pixels (8 bytes) of ymmA to the output when it has enough
+    ; space.
+    cmp         ecx, byte SIZEOF_YMMWORD/16
+    jb          short .column_st7
+    vmovq       MMWORD [edi], xmmA
+    add         edi, byte SIZEOF_YMMWORD/16*4
+    sub         ecx, byte SIZEOF_YMMWORD/16
+    vpsrldq     xmmA, SIZEOF_YMMWORD/16*4
+.column_st7:
+    ; Store one pixel (4 bytes) of ymmA to the output when it has enough
+    ; space.
+    test        ecx, ecx
+    jz          short .nextrow
+    vmovd       XMM_DWORD [edi], xmmA
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    alignx      16, 7
+
+.nextrow:
+    pop         ecx
+    pop         esi
+    pop         ebx
+    pop         edx
+    pop         edi
+    pop         eax
+
+    add         esi, byte SIZEOF_JSAMPROW
+    add         ebx, byte SIZEOF_JSAMPROW
+    add         edx, byte SIZEOF_JSAMPROW
+    add         edi, byte SIZEOF_JSAMPROW  ; output_buf
+    dec         eax                        ; num_rows
+    jg          near .rowloop
+
+    sfence                              ; flush the write buffer
+
+.return:
+    vzeroupper
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+    pop         ebx
+    mov         esp, ebp                ; esp <- aligned ebp
+    pop         esp                     ; esp <- original ebp
+    pop         ebp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,404 @@
+;
+; jdcolext.asm - colorspace conversion (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_ycc_rgb_convert_mmx(JDIMENSION out_width, JSAMPIMAGE input_buf,
+;                           JDIMENSION input_row, JSAMPARRAY output_buf,
+;                           int num_rows)
+;
+
+%define out_width(b)   (b) + 8          ; JDIMENSION out_width
+%define input_buf(b)   (b) + 12         ; JSAMPIMAGE input_buf
+%define input_row(b)   (b) + 16         ; JDIMENSION input_row
+%define output_buf(b)  (b) + 20         ; JSAMPARRAY output_buf
+%define num_rows(b)    (b) + 24         ; int num_rows
+
+%define original_ebp   ebp + 0
+%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_MMWORD
+                                        ; mmword wk[WK_NUM]
+%define WK_NUM         2
+%define gotptr         wk(0) - SIZEOF_POINTER  ; void * gotptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_mmx)
+
+EXTN(jsimd_ycc_rgb_convert_mmx):
+    push        ebp
+    mov         eax, esp                    ; eax = original ebp
+    sub         esp, byte 4
+    and         esp, byte (-SIZEOF_MMWORD)  ; align to 64 bits
+    mov         [esp], eax
+    mov         ebp, esp                    ; ebp = aligned ebp
+    lea         esp, [wk(0)]
+    pushpic     eax                     ; make a room for GOT address
+    push        ebx
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    get_GOT     ebx                     ; get GOT address
+    movpic      POINTER [gotptr], ebx   ; save GOT address
+
+    mov         ecx, JDIMENSION [out_width(eax)]  ; num_cols
+    test        ecx, ecx
+    jz          near .return
+
+    push        ecx
+
+    mov         edi, JSAMPIMAGE [input_buf(eax)]
+    mov         ecx, JDIMENSION [input_row(eax)]
+    mov         esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
+    mov         ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
+    mov         edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
+    lea         esi, [esi+ecx*SIZEOF_JSAMPROW]
+    lea         ebx, [ebx+ecx*SIZEOF_JSAMPROW]
+    lea         edx, [edx+ecx*SIZEOF_JSAMPROW]
+
+    pop         ecx
+
+    mov         edi, JSAMPARRAY [output_buf(eax)]
+    mov         eax, INT [num_rows(eax)]
+    test        eax, eax
+    jle         near .return
+    alignx      16, 7
+.rowloop:
+    push        eax
+    push        edi
+    push        edx
+    push        ebx
+    push        esi
+    push        ecx                     ; col
+
+    mov         esi, JSAMPROW [esi]     ; inptr0
+    mov         ebx, JSAMPROW [ebx]     ; inptr1
+    mov         edx, JSAMPROW [edx]     ; inptr2
+    mov         edi, JSAMPROW [edi]     ; outptr
+    movpic      eax, POINTER [gotptr]   ; load GOT address (eax)
+    alignx      16, 7
+.columnloop:
+
+    movq        mm5, MMWORD [ebx]       ; mm5=Cb(01234567)
+    movq        mm1, MMWORD [edx]       ; mm1=Cr(01234567)
+
+    pcmpeqw     mm4, mm4
+    pcmpeqw     mm7, mm7
+    psrlw       mm4, BYTE_BIT
+    psllw       mm7, 7                  ; mm7={0xFF80 0xFF80 0xFF80 0xFF80}
+    movq        mm0, mm4                ; mm0=mm4={0xFF 0x00 0xFF 0x00 ..}
+
+    pand        mm4, mm5                ; mm4=Cb(0246)=CbE
+    psrlw       mm5, BYTE_BIT           ; mm5=Cb(1357)=CbO
+    pand        mm0, mm1                ; mm0=Cr(0246)=CrE
+    psrlw       mm1, BYTE_BIT           ; mm1=Cr(1357)=CrO
+
+    paddw       mm4, mm7
+    paddw       mm5, mm7
+    paddw       mm0, mm7
+    paddw       mm1, mm7
+
+    ; (Original)
+    ; R = Y                + 1.40200 * Cr
+    ; G = Y - 0.34414 * Cb - 0.71414 * Cr
+    ; B = Y + 1.77200 * Cb
+    ;
+    ; (This implementation)
+    ; R = Y                + 0.40200 * Cr + Cr
+    ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+    ; B = Y - 0.22800 * Cb + Cb + Cb
+
+    movq        mm2, mm4                ; mm2=CbE
+    movq        mm3, mm5                ; mm3=CbO
+    paddw       mm4, mm4                ; mm4=2*CbE
+    paddw       mm5, mm5                ; mm5=2*CbO
+    movq        mm6, mm0                ; mm6=CrE
+    movq        mm7, mm1                ; mm7=CrO
+    paddw       mm0, mm0                ; mm0=2*CrE
+    paddw       mm1, mm1                ; mm1=2*CrO
+
+    pmulhw      mm4, [GOTOFF(eax,PW_MF0228)]  ; mm4=(2*CbE * -FIX(0.22800))
+    pmulhw      mm5, [GOTOFF(eax,PW_MF0228)]  ; mm5=(2*CbO * -FIX(0.22800))
+    pmulhw      mm0, [GOTOFF(eax,PW_F0402)]   ; mm0=(2*CrE * FIX(0.40200))
+    pmulhw      mm1, [GOTOFF(eax,PW_F0402)]   ; mm1=(2*CrO * FIX(0.40200))
+
+    paddw       mm4, [GOTOFF(eax,PW_ONE)]
+    paddw       mm5, [GOTOFF(eax,PW_ONE)]
+    psraw       mm4, 1                  ; mm4=(CbE * -FIX(0.22800))
+    psraw       mm5, 1                  ; mm5=(CbO * -FIX(0.22800))
+    paddw       mm0, [GOTOFF(eax,PW_ONE)]
+    paddw       mm1, [GOTOFF(eax,PW_ONE)]
+    psraw       mm0, 1                  ; mm0=(CrE * FIX(0.40200))
+    psraw       mm1, 1                  ; mm1=(CrO * FIX(0.40200))
+
+    paddw       mm4, mm2
+    paddw       mm5, mm3
+    paddw       mm4, mm2                ; mm4=(CbE * FIX(1.77200))=(B-Y)E
+    paddw       mm5, mm3                ; mm5=(CbO * FIX(1.77200))=(B-Y)O
+    paddw       mm0, mm6                ; mm0=(CrE * FIX(1.40200))=(R-Y)E
+    paddw       mm1, mm7                ; mm1=(CrO * FIX(1.40200))=(R-Y)O
+
+    movq        MMWORD [wk(0)], mm4     ; wk(0)=(B-Y)E
+    movq        MMWORD [wk(1)], mm5     ; wk(1)=(B-Y)O
+
+    movq        mm4, mm2
+    movq        mm5, mm3
+    punpcklwd   mm2, mm6
+    punpckhwd   mm4, mm6
+    pmaddwd     mm2, [GOTOFF(eax,PW_MF0344_F0285)]
+    pmaddwd     mm4, [GOTOFF(eax,PW_MF0344_F0285)]
+    punpcklwd   mm3, mm7
+    punpckhwd   mm5, mm7
+    pmaddwd     mm3, [GOTOFF(eax,PW_MF0344_F0285)]
+    pmaddwd     mm5, [GOTOFF(eax,PW_MF0344_F0285)]
+
+    paddd       mm2, [GOTOFF(eax,PD_ONEHALF)]
+    paddd       mm4, [GOTOFF(eax,PD_ONEHALF)]
+    psrad       mm2, SCALEBITS
+    psrad       mm4, SCALEBITS
+    paddd       mm3, [GOTOFF(eax,PD_ONEHALF)]
+    paddd       mm5, [GOTOFF(eax,PD_ONEHALF)]
+    psrad       mm3, SCALEBITS
+    psrad       mm5, SCALEBITS
+
+    packssdw    mm2, mm4                ; mm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
+    packssdw    mm3, mm5                ; mm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
+    psubw       mm2, mm6                ; mm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
+    psubw       mm3, mm7                ; mm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
+
+    movq        mm5, MMWORD [esi]       ; mm5=Y(01234567)
+
+    pcmpeqw     mm4, mm4
+    psrlw       mm4, BYTE_BIT           ; mm4={0xFF 0x00 0xFF 0x00 ..}
+    pand        mm4, mm5                ; mm4=Y(0246)=YE
+    psrlw       mm5, BYTE_BIT           ; mm5=Y(1357)=YO
+
+    paddw       mm0, mm4                ; mm0=((R-Y)E+YE)=RE=(R0 R2 R4 R6)
+    paddw       mm1, mm5                ; mm1=((R-Y)O+YO)=RO=(R1 R3 R5 R7)
+    packuswb    mm0, mm0                ; mm0=(R0 R2 R4 R6 ** ** ** **)
+    packuswb    mm1, mm1                ; mm1=(R1 R3 R5 R7 ** ** ** **)
+
+    paddw       mm2, mm4                ; mm2=((G-Y)E+YE)=GE=(G0 G2 G4 G6)
+    paddw       mm3, mm5                ; mm3=((G-Y)O+YO)=GO=(G1 G3 G5 G7)
+    packuswb    mm2, mm2                ; mm2=(G0 G2 G4 G6 ** ** ** **)
+    packuswb    mm3, mm3                ; mm3=(G1 G3 G5 G7 ** ** ** **)
+
+    paddw       mm4,  MMWORD [wk(0)]    ; mm4=(YE+(B-Y)E)=BE=(B0 B2 B4 B6)
+    paddw       mm5,  MMWORD [wk(1)]    ; mm5=(YO+(B-Y)O)=BO=(B1 B3 B5 B7)
+    packuswb    mm4, mm4                ; mm4=(B0 B2 B4 B6 ** ** ** **)
+    packuswb    mm5, mm5                ; mm5=(B1 B3 B5 B7 ** ** ** **)
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+    ; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **)
+    ; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **)
+    ; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **)
+    ; mmG=(** ** ** ** ** ** ** **), mmH=(** ** ** ** ** ** ** **)
+
+    punpcklbw   mmA, mmC                ; mmA=(00 10 02 12 04 14 06 16)
+    punpcklbw   mmE, mmB                ; mmE=(20 01 22 03 24 05 26 07)
+    punpcklbw   mmD, mmF                ; mmD=(11 21 13 23 15 25 17 27)
+
+    movq        mmG, mmA
+    movq        mmH, mmA
+    punpcklwd   mmA, mmE                ; mmA=(00 10 20 01 02 12 22 03)
+    punpckhwd   mmG, mmE                ; mmG=(04 14 24 05 06 16 26 07)
+
+    psrlq       mmH, 2*BYTE_BIT         ; mmH=(02 12 04 14 06 16 -- --)
+    psrlq       mmE, 2*BYTE_BIT         ; mmE=(22 03 24 05 26 07 -- --)
+
+    movq        mmC, mmD
+    movq        mmB, mmD
+    punpcklwd   mmD, mmH                ; mmD=(11 21 02 12 13 23 04 14)
+    punpckhwd   mmC, mmH                ; mmC=(15 25 06 16 17 27 -- --)
+
+    psrlq       mmB, 2*BYTE_BIT         ; mmB=(13 23 15 25 17 27 -- --)
+
+    movq        mmF, mmE
+    punpcklwd   mmE, mmB                ; mmE=(22 03 13 23 24 05 15 25)
+    punpckhwd   mmF, mmB                ; mmF=(26 07 17 27 -- -- -- --)
+
+    punpckldq   mmA, mmD                ; mmA=(00 10 20 01 11 21 02 12)
+    punpckldq   mmE, mmG                ; mmE=(22 03 13 23 04 14 24 05)
+    punpckldq   mmC, mmF                ; mmC=(15 25 06 16 26 07 17 27)
+
+    cmp         ecx, byte SIZEOF_MMWORD
+    jb          short .column_st16
+
+    movq        MMWORD [edi+0*SIZEOF_MMWORD], mmA
+    movq        MMWORD [edi+1*SIZEOF_MMWORD], mmE
+    movq        MMWORD [edi+2*SIZEOF_MMWORD], mmC
+
+    sub         ecx, byte SIZEOF_MMWORD
+    jz          short .nextrow
+
+    add         esi, byte SIZEOF_MMWORD                ; inptr0
+    add         ebx, byte SIZEOF_MMWORD                ; inptr1
+    add         edx, byte SIZEOF_MMWORD                ; inptr2
+    add         edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD  ; outptr
+    jmp         near .columnloop
+    alignx      16, 7
+
+.column_st16:
+    lea         ecx, [ecx+ecx*2]        ; imul ecx, RGB_PIXELSIZE
+    cmp         ecx, byte 2*SIZEOF_MMWORD
+    jb          short .column_st8
+    movq        MMWORD [edi+0*SIZEOF_MMWORD], mmA
+    movq        MMWORD [edi+1*SIZEOF_MMWORD], mmE
+    movq        mmA, mmC
+    sub         ecx, byte 2*SIZEOF_MMWORD
+    add         edi, byte 2*SIZEOF_MMWORD
+    jmp         short .column_st4
+.column_st8:
+    cmp         ecx, byte SIZEOF_MMWORD
+    jb          short .column_st4
+    movq        MMWORD [edi+0*SIZEOF_MMWORD], mmA
+    movq        mmA, mmE
+    sub         ecx, byte SIZEOF_MMWORD
+    add         edi, byte SIZEOF_MMWORD
+.column_st4:
+    movd        eax, mmA
+    cmp         ecx, byte SIZEOF_DWORD
+    jb          short .column_st2
+    mov         dword [edi+0*SIZEOF_DWORD], eax
+    psrlq       mmA, DWORD_BIT
+    movd        eax, mmA
+    sub         ecx, byte SIZEOF_DWORD
+    add         edi, byte SIZEOF_DWORD
+.column_st2:
+    cmp         ecx, byte SIZEOF_WORD
+    jb          short .column_st1
+    mov         word [edi+0*SIZEOF_WORD], ax
+    shr         eax, WORD_BIT
+    sub         ecx, byte SIZEOF_WORD
+    add         edi, byte SIZEOF_WORD
+.column_st1:
+    cmp         ecx, byte SIZEOF_BYTE
+    jb          short .nextrow
+    mov         byte [edi+0*SIZEOF_BYTE], al
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+    pcmpeqb     mm6, mm6                ; mm6=(X0 X2 X4 X6 ** ** ** **)
+    pcmpeqb     mm7, mm7                ; mm7=(X1 X3 X5 X7 ** ** ** **)
+%else
+    pxor        mm6, mm6                ; mm6=(X0 X2 X4 X6 ** ** ** **)
+    pxor        mm7, mm7                ; mm7=(X1 X3 X5 X7 ** ** ** **)
+%endif
+    ; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **)
+    ; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **)
+    ; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **)
+    ; mmG=(30 32 34 36 ** ** ** **), mmH=(31 33 35 37 ** ** ** **)
+
+    punpcklbw   mmA, mmC                ; mmA=(00 10 02 12 04 14 06 16)
+    punpcklbw   mmE, mmG                ; mmE=(20 30 22 32 24 34 26 36)
+    punpcklbw   mmB, mmD                ; mmB=(01 11 03 13 05 15 07 17)
+    punpcklbw   mmF, mmH                ; mmF=(21 31 23 33 25 35 27 37)
+
+    movq        mmC, mmA
+    punpcklwd   mmA, mmE                ; mmA=(00 10 20 30 02 12 22 32)
+    punpckhwd   mmC, mmE                ; mmC=(04 14 24 34 06 16 26 36)
+    movq        mmG, mmB
+    punpcklwd   mmB, mmF                ; mmB=(01 11 21 31 03 13 23 33)
+    punpckhwd   mmG, mmF                ; mmG=(05 15 25 35 07 17 27 37)
+
+    movq        mmD, mmA
+    punpckldq   mmA, mmB                ; mmA=(00 10 20 30 01 11 21 31)
+    punpckhdq   mmD, mmB                ; mmD=(02 12 22 32 03 13 23 33)
+    movq        mmH, mmC
+    punpckldq   mmC, mmG                ; mmC=(04 14 24 34 05 15 25 35)
+    punpckhdq   mmH, mmG                ; mmH=(06 16 26 36 07 17 27 37)
+
+    cmp         ecx, byte SIZEOF_MMWORD
+    jb          short .column_st16
+
+    movq        MMWORD [edi+0*SIZEOF_MMWORD], mmA
+    movq        MMWORD [edi+1*SIZEOF_MMWORD], mmD
+    movq        MMWORD [edi+2*SIZEOF_MMWORD], mmC
+    movq        MMWORD [edi+3*SIZEOF_MMWORD], mmH
+
+    sub         ecx, byte SIZEOF_MMWORD
+    jz          short .nextrow
+
+    add         esi, byte SIZEOF_MMWORD                ; inptr0
+    add         ebx, byte SIZEOF_MMWORD                ; inptr1
+    add         edx, byte SIZEOF_MMWORD                ; inptr2
+    add         edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD  ; outptr
+    jmp         near .columnloop
+    alignx      16, 7
+
+.column_st16:
+    cmp         ecx, byte SIZEOF_MMWORD/2
+    jb          short .column_st8
+    movq        MMWORD [edi+0*SIZEOF_MMWORD], mmA
+    movq        MMWORD [edi+1*SIZEOF_MMWORD], mmD
+    movq        mmA, mmC
+    movq        mmD, mmH
+    sub         ecx, byte SIZEOF_MMWORD/2
+    add         edi, byte 2*SIZEOF_MMWORD
+.column_st8:
+    cmp         ecx, byte SIZEOF_MMWORD/4
+    jb          short .column_st4
+    movq        MMWORD [edi+0*SIZEOF_MMWORD], mmA
+    movq        mmA, mmD
+    sub         ecx, byte SIZEOF_MMWORD/4
+    add         edi, byte 1*SIZEOF_MMWORD
+.column_st4:
+    cmp         ecx, byte SIZEOF_MMWORD/8
+    jb          short .nextrow
+    movd        dword [edi+0*SIZEOF_DWORD], mmA
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    alignx      16, 7
+
+.nextrow:
+    pop         ecx
+    pop         esi
+    pop         ebx
+    pop         edx
+    pop         edi
+    pop         eax
+
+    add         esi, byte SIZEOF_JSAMPROW
+    add         ebx, byte SIZEOF_JSAMPROW
+    add         edx, byte SIZEOF_JSAMPROW
+    add         edi, byte SIZEOF_JSAMPROW  ; output_buf
+    dec         eax                        ; num_rows
+    jg          near .rowloop
+
+    emms                                ; empty MMX state
+
+.return:
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+    pop         ebx
+    mov         esp, ebp                ; esp <- aligned ebp
+    pop         esp                     ; esp <- original ebp
+    pop         ebp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,458 @@
+;
+; jdcolext.asm - colorspace conversion (SSE2)
+;
+; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2012, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jcolsamp.inc"
+
+; --------------------------------------------------------------------------
+;
+; Convert some rows of samples to the output colorspace.
+;
+; GLOBAL(void)
+; jsimd_ycc_rgb_convert_sse2(JDIMENSION out_width, JSAMPIMAGE input_buf,
+;                            JDIMENSION input_row, JSAMPARRAY output_buf,
+;                            int num_rows)
+;
+
+%define out_width(b)   (b) + 8          ; JDIMENSION out_width
+%define input_buf(b)   (b) + 12         ; JSAMPIMAGE input_buf
+%define input_row(b)   (b) + 16         ; JDIMENSION input_row
+%define output_buf(b)  (b) + 20         ; JSAMPARRAY output_buf
+%define num_rows(b)    (b) + 24         ; int num_rows
+
+%define original_ebp   ebp + 0
+%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_XMMWORD
+                                        ; xmmword wk[WK_NUM]
+%define WK_NUM         2
+%define gotptr         wk(0) - SIZEOF_POINTER  ; void * gotptr
+
+    align       32
+    GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_sse2)
+
+EXTN(jsimd_ycc_rgb_convert_sse2):
+    push        ebp
+    mov         eax, esp                     ; eax = original ebp
+    sub         esp, byte 4
+    and         esp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+    mov         [esp], eax
+    mov         ebp, esp                     ; ebp = aligned ebp
+    lea         esp, [wk(0)]
+    pushpic     eax                     ; make a room for GOT address
+    push        ebx
+;   push        ecx                     ; need not be preserved
+;   push        edx                     ; need not be preserved
+    push        esi
+    push        edi
+
+    get_GOT     ebx                     ; get GOT address
+    movpic      POINTER [gotptr], ebx   ; save GOT address
+
+    mov         ecx, JDIMENSION [out_width(eax)]  ; num_cols
+    test        ecx, ecx
+    jz          near .return
+
+    push        ecx
+
+    mov         edi, JSAMPIMAGE [input_buf(eax)]
+    mov         ecx, JDIMENSION [input_row(eax)]
+    mov         esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
+    mov         ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
+    mov         edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
+    lea         esi, [esi+ecx*SIZEOF_JSAMPROW]
+    lea         ebx, [ebx+ecx*SIZEOF_JSAMPROW]
+    lea         edx, [edx+ecx*SIZEOF_JSAMPROW]
+
+    pop         ecx
+
+    mov         edi, JSAMPARRAY [output_buf(eax)]
+    mov         eax, INT [num_rows(eax)]
+    test        eax, eax
+    jle         near .return
+    alignx      16, 7
+.rowloop:
+    push        eax
+    push        edi
+    push        edx
+    push        ebx
+    push        esi
+    push        ecx                     ; col
+
+    mov         esi, JSAMPROW [esi]     ; inptr0
+    mov         ebx, JSAMPROW [ebx]     ; inptr1
+    mov         edx, JSAMPROW [edx]     ; inptr2
+    mov         edi, JSAMPROW [edi]     ; outptr
+    movpic      eax, POINTER [gotptr]   ; load GOT address (eax)
+    alignx      16, 7
+.columnloop:
+
+    movdqa      xmm5, XMMWORD [ebx]     ; xmm5=Cb(0123456789ABCDEF)
+    movdqa      xmm1, XMMWORD [edx]     ; xmm1=Cr(0123456789ABCDEF)
+
+    pcmpeqw     xmm4, xmm4
+    pcmpeqw     xmm7, xmm7
+    psrlw       xmm4, BYTE_BIT
+    psllw       xmm7, 7                 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
+    movdqa      xmm0, xmm4              ; xmm0=xmm4={0xFF 0x00 0xFF 0x00 ..}
+
+    pand        xmm4, xmm5              ; xmm4=Cb(02468ACE)=CbE
+    psrlw       xmm5, BYTE_BIT          ; xmm5=Cb(13579BDF)=CbO
+    pand        xmm0, xmm1              ; xmm0=Cr(02468ACE)=CrE
+    psrlw       xmm1, BYTE_BIT          ; xmm1=Cr(13579BDF)=CrO
+
+    paddw       xmm4, xmm7
+    paddw       xmm5, xmm7
+    paddw       xmm0, xmm7
+    paddw       xmm1, xmm7
+
+    ; (Original)
+    ; R = Y                + 1.40200 * Cr
+    ; G = Y - 0.34414 * Cb - 0.71414 * Cr
+    ; B = Y + 1.77200 * Cb
+    ;
+    ; (This implementation)
+    ; R = Y                + 0.40200 * Cr + Cr
+    ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+    ; B = Y - 0.22800 * Cb + Cb + Cb
+
+    movdqa      xmm2, xmm4              ; xmm2=CbE
+    movdqa      xmm3, xmm5              ; xmm3=CbO
+    paddw       xmm4, xmm4              ; xmm4=2*CbE
+    paddw       xmm5, xmm5              ; xmm5=2*CbO
+    movdqa      xmm6, xmm0              ; xmm6=CrE
+    movdqa      xmm7, xmm1              ; xmm7=CrO
+    paddw       xmm0, xmm0              ; xmm0=2*CrE
+    paddw       xmm1, xmm1              ; xmm1=2*CrO
+
+    pmulhw      xmm4, [GOTOFF(eax,PW_MF0228)]  ; xmm4=(2*CbE * -FIX(0.22800))
+    pmulhw      xmm5, [GOTOFF(eax,PW_MF0228)]  ; xmm5=(2*CbO * -FIX(0.22800))
+    pmulhw      xmm0, [GOTOFF(eax,PW_F0402)]   ; xmm0=(2*CrE * FIX(0.40200))
+    pmulhw      xmm1, [GOTOFF(eax,PW_F0402)]   ; xmm1=(2*CrO * FIX(0.40200))
+
+    paddw       xmm4, [GOTOFF(eax,PW_ONE)]
+    paddw       xmm5, [GOTOFF(eax,PW_ONE)]
+    psraw       xmm4, 1                 ; xmm4=(CbE * -FIX(0.22800))
+    psraw       xmm5, 1                 ; xmm5=(CbO * -FIX(0.22800))
+    paddw       xmm0, [GOTOFF(eax,PW_ONE)]
+    paddw       xmm1, [GOTOFF(eax,PW_ONE)]
+    psraw       xmm0, 1                 ; xmm0=(CrE * FIX(0.40200))
+    psraw       xmm1, 1                 ; xmm1=(CrO * FIX(0.40200))
+
+    paddw       xmm4, xmm2
+    paddw       xmm5, xmm3
+    paddw       xmm4, xmm2              ; xmm4=(CbE * FIX(1.77200))=(B-Y)E
+    paddw       xmm5, xmm3              ; xmm5=(CbO * FIX(1.77200))=(B-Y)O
+    paddw       xmm0, xmm6              ; xmm0=(CrE * FIX(1.40200))=(R-Y)E
+    paddw       xmm1, xmm7              ; xmm1=(CrO * FIX(1.40200))=(R-Y)O
+
+    movdqa      XMMWORD [wk(0)], xmm4   ; wk(0)=(B-Y)E
+    movdqa      XMMWORD [wk(1)], xmm5   ; wk(1)=(B-Y)O
+
+    movdqa      xmm4, xmm2
+    movdqa      xmm5, xmm3
+    punpcklwd   xmm2, xmm6
+    punpckhwd   xmm4, xmm6
+    pmaddwd     xmm2, [GOTOFF(eax,PW_MF0344_F0285)]
+    pmaddwd     xmm4, [GOTOFF(eax,PW_MF0344_F0285)]
+    punpcklwd   xmm3, xmm7
+    punpckhwd   xmm5, xmm7
+    pmaddwd     xmm3, [GOTOFF(eax,PW_MF0344_F0285)]
+    pmaddwd     xmm5, [GOTOFF(eax,PW_MF0344_F0285)]
+
+    paddd       xmm2, [GOTOFF(eax,PD_ONEHALF)]
+    paddd       xmm4, [GOTOFF(eax,PD_ONEHALF)]
+    psrad       xmm2, SCALEBITS
+    psrad       xmm4, SCALEBITS
+    paddd       xmm3, [GOTOFF(eax,PD_ONEHALF)]
+    paddd       xmm5, [GOTOFF(eax,PD_ONEHALF)]
+    psrad       xmm3, SCALEBITS
+    psrad       xmm5, SCALEBITS
+
+    packssdw    xmm2, xmm4              ; xmm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
+    packssdw    xmm3, xmm5              ; xmm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
+    psubw       xmm2, xmm6              ; xmm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
+    psubw       xmm3, xmm7              ; xmm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
+
+    movdqa      xmm5, XMMWORD [esi]     ; xmm5=Y(0123456789ABCDEF)
+
+    pcmpeqw     xmm4, xmm4
+    psrlw       xmm4, BYTE_BIT          ; xmm4={0xFF 0x00 0xFF 0x00 ..}
+    pand        xmm4, xmm5              ; xmm4=Y(02468ACE)=YE
+    psrlw       xmm5, BYTE_BIT          ; xmm5=Y(13579BDF)=YO
+
+    paddw       xmm0, xmm4              ; xmm0=((R-Y)E+YE)=RE=R(02468ACE)
+    paddw       xmm1, xmm5              ; xmm1=((R-Y)O+YO)=RO=R(13579BDF)
+    packuswb    xmm0, xmm0              ; xmm0=R(02468ACE********)
+    packuswb    xmm1, xmm1              ; xmm1=R(13579BDF********)
+
+    paddw       xmm2, xmm4              ; xmm2=((G-Y)E+YE)=GE=G(02468ACE)
+    paddw       xmm3, xmm5              ; xmm3=((G-Y)O+YO)=GO=G(13579BDF)
+    packuswb    xmm2, xmm2              ; xmm2=G(02468ACE********)
+    packuswb    xmm3, xmm3              ; xmm3=G(13579BDF********)
+
+    paddw       xmm4, XMMWORD [wk(0)]   ; xmm4=(YE+(B-Y)E)=BE=B(02468ACE)
+    paddw       xmm5, XMMWORD [wk(1)]   ; xmm5=(YO+(B-Y)O)=BO=B(13579BDF)
+    packuswb    xmm4, xmm4              ; xmm4=B(02468ACE********)
+    packuswb    xmm5, xmm5              ; xmm5=B(13579BDF********)
+
+%if RGB_PIXELSIZE == 3  ; ---------------
+
+    ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+    ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+    ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+    ; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
+
+    punpcklbw   xmmA, xmmC        ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+    punpcklbw   xmmE, xmmB        ; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
+    punpcklbw   xmmD, xmmF        ; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
+
+    movdqa      xmmG, xmmA
+    movdqa      xmmH, xmmA
+    punpcklwd   xmmA, xmmE        ; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
+    punpckhwd   xmmG, xmmE        ; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
+
+    psrldq      xmmH, 2           ; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
+    psrldq      xmmE, 2           ; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
+
+    movdqa      xmmC, xmmD
+    movdqa      xmmB, xmmD
+    punpcklwd   xmmD, xmmH        ; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
+    punpckhwd   xmmC, xmmH        ; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
+
+    psrldq      xmmB, 2           ; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
+
+    movdqa      xmmF, xmmE
+    punpcklwd   xmmE, xmmB        ; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
+    punpckhwd   xmmF, xmmB        ; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
+
+    pshufd      xmmH, xmmA, 0x4E  ; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
+    movdqa      xmmB, xmmE
+    punpckldq   xmmA, xmmD        ; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
+    punpckldq   xmmE, xmmH        ; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
+    punpckhdq   xmmD, xmmB        ; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
+
+    pshufd      xmmH, xmmG, 0x4E  ; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
+    movdqa      xmmB, xmmF
+    punpckldq   xmmG, xmmC        ; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
+    punpckldq   xmmF, xmmH        ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
+    punpckhdq   xmmC, xmmB        ; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
+
+    punpcklqdq  xmmA, xmmE        ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
+    punpcklqdq  xmmD, xmmG        ; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
+    punpcklqdq  xmmF, xmmC        ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
+
+    cmp         ecx, byte SIZEOF_XMMWORD
+    jb          short .column_st32
+
+    test        edi, SIZEOF_XMMWORD-1
+    jnz         short .out1
+    ; --(aligned)-------------------
+    movntdq     XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+    movntdq     XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
+    movntdq     XMMWORD [edi+2*SIZEOF_XMMWORD], xmmF
+    jmp         short .out0
+.out1:  ; --(unaligned)-----------------
+    movdqu      XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+    movdqu      XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
+    movdqu      XMMWORD [edi+2*SIZEOF_XMMWORD], xmmF
+.out0:
+    add         edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
+    sub         ecx, byte SIZEOF_XMMWORD
+    jz          near .nextrow
+
+    add         esi, byte SIZEOF_XMMWORD  ; inptr0
+    add         ebx, byte SIZEOF_XMMWORD  ; inptr1
+    add         edx, byte SIZEOF_XMMWORD  ; inptr2
+    jmp         near .columnloop
+    alignx      16, 7
+
+.column_st32:
+    lea         ecx, [ecx+ecx*2]        ; imul ecx, RGB_PIXELSIZE
+    cmp         ecx, byte 2*SIZEOF_XMMWORD
+    jb          short .column_st16
+    movdqu      XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+    movdqu      XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
+    add         edi, byte 2*SIZEOF_XMMWORD  ; outptr
+    movdqa      xmmA, xmmF
+    sub         ecx, byte 2*SIZEOF_XMMWORD
+    jmp         short .column_st15
+.column_st16:
+    cmp         ecx, byte SIZEOF_XMMWORD
+    jb          short .column_st15
+    movdqu      XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+    add         edi, byte SIZEOF_XMMWORD    ; outptr
+    movdqa      xmmA, xmmD
+    sub         ecx, byte SIZEOF_XMMWORD
+.column_st15:
+    ; Store the lower 8 bytes of xmmA to the output when it has enough
+    ; space.
+    cmp         ecx, byte SIZEOF_MMWORD
+    jb          short .column_st7
+    movq        XMM_MMWORD [edi], xmmA
+    add         edi, byte SIZEOF_MMWORD
+    sub         ecx, byte SIZEOF_MMWORD
+    psrldq      xmmA, SIZEOF_MMWORD
+.column_st7:
+    ; Store the lower 4 bytes of xmmA to the output when it has enough
+    ; space.
+    cmp         ecx, byte SIZEOF_DWORD
+    jb          short .column_st3
+    movd        XMM_DWORD [edi], xmmA
+    add         edi, byte SIZEOF_DWORD
+    sub         ecx, byte SIZEOF_DWORD
+    psrldq      xmmA, SIZEOF_DWORD
+.column_st3:
+    ; Store the lower 2 bytes of eax to the output when it has enough
+    ; space.
+    movd        eax, xmmA
+    cmp         ecx, byte SIZEOF_WORD
+    jb          short .column_st1
+    mov         word [edi], ax
+    add         edi, byte SIZEOF_WORD
+    sub         ecx, byte SIZEOF_WORD
+    shr         eax, 16
+.column_st1:
+    ; Store the lower 1 byte of eax to the output when it has enough
+    ; space.
+    test        ecx, ecx
+    jz          short .nextrow
+    mov         byte [edi], al
+
+%else  ; RGB_PIXELSIZE == 4 ; -----------
+
+%ifdef RGBX_FILLER_0XFF
+    pcmpeqb     xmm6, xmm6              ; xmm6=XE=X(02468ACE********)
+    pcmpeqb     xmm7, xmm7              ; xmm7=XO=X(13579BDF********)
+%else
+    pxor        xmm6, xmm6              ; xmm6=XE=X(02468ACE********)
+    pxor        xmm7, xmm7              ; xmm7=XO=X(13579BDF********)
+%endif
+    ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
+    ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
+    ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
+    ; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
+
+    punpcklbw   xmmA, xmmC  ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
+    punpcklbw   xmmE, xmmG  ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
+    punpcklbw   xmmB, xmmD  ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
+    punpcklbw   xmmF, xmmH  ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
+
+    movdqa      xmmC, xmmA
+    punpcklwd   xmmA, xmmE  ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
+    punpckhwd   xmmC, xmmE  ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
+    movdqa      xmmG, xmmB
+    punpcklwd   xmmB, xmmF  ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
+    punpckhwd   xmmG, xmmF  ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
+
+    movdqa      xmmD, xmmA
+    punpckldq   xmmA, xmmB  ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
+    punpckhdq   xmmD, xmmB  ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
+    movdqa      xmmH, xmmC
+    punpckldq   xmmC, xmmG  ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
+    punpckhdq   xmmH, xmmG  ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
+
+    cmp         ecx, byte SIZEOF_XMMWORD
+    jb          short .column_st32
+
+    test        edi, SIZEOF_XMMWORD-1
+    jnz         short .out1
+    ; --(aligned)-------------------
+    movntdq     XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+    movntdq     XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
+    movntdq     XMMWORD [edi+2*SIZEOF_XMMWORD], xmmC
+    movntdq     XMMWORD [edi+3*SIZEOF_XMMWORD], xmmH
+    jmp         short .out0
+.out1:  ; --(unaligned)-----------------
+    movdqu      XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+    movdqu      XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
+    movdqu      XMMWORD [edi+2*SIZEOF_XMMWORD], xmmC
+    movdqu      XMMWORD [edi+3*SIZEOF_XMMWORD], xmmH
+.out0:
+    add         edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
+    sub         ecx, byte SIZEOF_XMMWORD
+    jz          near .nextrow
+
+    add         esi, byte SIZEOF_XMMWORD  ; inptr0
+    add         ebx, byte SIZEOF_XMMWORD  ; inptr1
+    add         edx, byte SIZEOF_XMMWORD  ; inptr2
+    jmp         near .columnloop
+    alignx      16, 7
+
+.column_st32:
+    cmp         ecx, byte SIZEOF_XMMWORD/2
+    jb          short .column_st16
+    movdqu      XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+    movdqu      XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
+    add         edi, byte 2*SIZEOF_XMMWORD  ; outptr
+    movdqa      xmmA, xmmC
+    movdqa      xmmD, xmmH
+    sub         ecx, byte SIZEOF_XMMWORD/2
+.column_st16:
+    cmp         ecx, byte SIZEOF_XMMWORD/4
+    jb          short .column_st15
+    movdqu      XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
+    add         edi, byte SIZEOF_XMMWORD    ; outptr
+    movdqa      xmmA, xmmD
+    sub         ecx, byte SIZEOF_XMMWORD/4
+.column_st15:
+    ; Store two pixels (8 bytes) of xmmA to the output when it has enough
+    ; space.
+    cmp         ecx, byte SIZEOF_XMMWORD/8
+    jb          short .column_st7
+    movq        XMM_MMWORD [edi], xmmA
+    add         edi, byte SIZEOF_XMMWORD/8*4
+    sub         ecx, byte SIZEOF_XMMWORD/8
+    psrldq      xmmA, SIZEOF_XMMWORD/8*4
+.column_st7:
+    ; Store one pixel (4 bytes) of xmmA to the output when it has enough
+    ; space.
+    test        ecx, ecx
+    jz          short .nextrow
+    movd        XMM_DWORD [edi], xmmA
+
+%endif  ; RGB_PIXELSIZE ; ---------------
+
+    alignx      16, 7
+
+.nextrow:
+    pop         ecx
+    pop         esi
+    pop         ebx
+    pop         edx
+    pop         edi
+    pop         eax
+
+    add         esi, byte SIZEOF_JSAMPROW
+    add         ebx, byte SIZEOF_JSAMPROW
+    add         edx, byte SIZEOF_JSAMPROW
+    add         edi, byte SIZEOF_JSAMPROW  ; output_buf
+    dec         eax                        ; num_rows
+    jg          near .rowloop
+
+    sfence                              ; flush the write buffer
+
+.return:
+    pop         edi
+    pop         esi
+;   pop         edx                     ; need not be preserved
+;   pop         ecx                     ; need not be preserved
+    pop         ebx
+    mov         esp, ebp                ; esp <- aligned ebp
+    pop         esp                     ; esp <- original ebp
+    pop         ebp
+    ret
+
+; For some reason, the OS X linker does not honor the request to align the
+; segment unless we do this.
+    align       32
@@ -0,0 +1,118 @@
+;
+; jdcolor.asm - colorspace conversion (AVX2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_344 equ  22554              ; FIX(0.34414)
+F_0_714 equ  46802              ; FIX(0.71414)
+F_1_402 equ  91881              ; FIX(1.40200)
+F_1_772 equ 116130              ; FIX(1.77200)
+F_0_402 equ (F_1_402 - 65536)   ; FIX(1.40200) - FIX(1)
+F_0_285 equ ( 65536 - F_0_714)  ; FIX(1) - FIX(0.71414)
+F_0_228 equ (131072 - F_1_772)  ; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_ycc_rgb_convert_avx2)
+
+EXTN(jconst_ycc_rgb_convert_avx2):
+
+PW_F0402        times 16 dw  F_0_402
+PW_MF0228       times 16 dw -F_0_228
+PW_MF0344_F0285 times 8  dw -F_0_344, F_0_285
+PW_ONE          times 16 dw  1
+PD_ONEHALF      times 8  dd  1 << (SCALEBITS - 1)
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+
+%include "jdcolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_ycc_rgb_convert_avx2  jsimd_ycc_extrgb_convert_avx2
+%include "jdcolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_ycc_rgb_convert_avx2  jsimd_ycc_extrgbx_convert_avx2
+%include "jdcolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_ycc_rgb_convert_avx2  jsimd_ycc_extbgr_convert_avx2
+%include "jdcolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_ycc_rgb_convert_avx2  jsimd_ycc_extbgrx_convert_avx2
+%include "jdcolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_ycc_rgb_convert_avx2  jsimd_ycc_extxbgr_convert_avx2
+%include "jdcolext-avx2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_ycc_rgb_convert_avx2  jsimd_ycc_extxrgb_convert_avx2
+%include "jdcolext-avx2.asm"
@@ -0,0 +1,117 @@
+;
+; jdcolor.asm - colorspace conversion (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_344 equ  22554              ; FIX(0.34414)
+F_0_714 equ  46802              ; FIX(0.71414)
+F_1_402 equ  91881              ; FIX(1.40200)
+F_1_772 equ 116130              ; FIX(1.77200)
+F_0_402 equ (F_1_402 - 65536)   ; FIX(1.40200) - FIX(1)
+F_0_285 equ ( 65536 - F_0_714)  ; FIX(1) - FIX(0.71414)
+F_0_228 equ (131072 - F_1_772)  ; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_ycc_rgb_convert_mmx)
+
+EXTN(jconst_ycc_rgb_convert_mmx):
+
+PW_F0402        times 4 dw  F_0_402
+PW_MF0228       times 4 dw -F_0_228
+PW_MF0344_F0285 times 2 dw -F_0_344, F_0_285
+PW_ONE          times 4 dw  1
+PD_ONEHALF      times 2 dd  1 << (SCALEBITS - 1)
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+
+%include "jdcolext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_ycc_rgb_convert_mmx  jsimd_ycc_extrgb_convert_mmx
+%include "jdcolext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_ycc_rgb_convert_mmx  jsimd_ycc_extrgbx_convert_mmx
+%include "jdcolext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_ycc_rgb_convert_mmx  jsimd_ycc_extbgr_convert_mmx
+%include "jdcolext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_ycc_rgb_convert_mmx  jsimd_ycc_extbgrx_convert_mmx
+%include "jdcolext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_ycc_rgb_convert_mmx  jsimd_ycc_extxbgr_convert_mmx
+%include "jdcolext-mmx.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_ycc_rgb_convert_mmx  jsimd_ycc_extxrgb_convert_mmx
+%include "jdcolext-mmx.asm"
@@ -0,0 +1,117 @@
+;
+; jdcolor.asm - colorspace conversion (SSE2)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+; Copyright (C) 2009, 2016, D. R. Commander.
+;
+; Based on the x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+
+%include "jsimdext.inc"
+
+; --------------------------------------------------------------------------
+
+%define SCALEBITS  16
+
+F_0_344 equ  22554              ; FIX(0.34414)
+F_0_714 equ  46802              ; FIX(0.71414)
+F_1_402 equ  91881              ; FIX(1.40200)
+F_1_772 equ 116130              ; FIX(1.77200)
+F_0_402 equ (F_1_402 - 65536)   ; FIX(1.40200) - FIX(1)
+F_0_285 equ ( 65536 - F_0_714)  ; FIX(1) - FIX(0.71414)
+F_0_228 equ (131072 - F_1_772)  ; FIX(2) - FIX(1.77200)
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_CONST
+
+    alignz      32
+    GLOBAL_DATA(jconst_ycc_rgb_convert_sse2)
+
+EXTN(jconst_ycc_rgb_convert_sse2):
+
+PW_F0402        times 8 dw  F_0_402
+PW_MF0228       times 8 dw -F_0_228
+PW_MF0344_F0285 times 4 dw -F_0_344, F_0_285
+PW_ONE          times 8 dw  1
+PD_ONEHALF      times 4 dd  1 << (SCALEBITS - 1)
+
+    alignz      32
+
+; --------------------------------------------------------------------------
+    SECTION     SEG_TEXT
+    BITS        32
+
+%include "jdcolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGB_RED
+%define RGB_GREEN  EXT_RGB_GREEN
+%define RGB_BLUE  EXT_RGB_BLUE
+%define RGB_PIXELSIZE  EXT_RGB_PIXELSIZE
+%define jsimd_ycc_rgb_convert_sse2  jsimd_ycc_extrgb_convert_sse2
+%include "jdcolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_RGBX_RED
+%define RGB_GREEN  EXT_RGBX_GREEN
+%define RGB_BLUE  EXT_RGBX_BLUE
+%define RGB_PIXELSIZE  EXT_RGBX_PIXELSIZE
+%define jsimd_ycc_rgb_convert_sse2  jsimd_ycc_extrgbx_convert_sse2
+%include "jdcolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGR_RED
+%define RGB_GREEN  EXT_BGR_GREEN
+%define RGB_BLUE  EXT_BGR_BLUE
+%define RGB_PIXELSIZE  EXT_BGR_PIXELSIZE
+%define jsimd_ycc_rgb_convert_sse2  jsimd_ycc_extbgr_convert_sse2
+%include "jdcolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_BGRX_RED
+%define RGB_GREEN  EXT_BGRX_GREEN
+%define RGB_BLUE  EXT_BGRX_BLUE
+%define RGB_PIXELSIZE  EXT_BGRX_PIXELSIZE
+%define jsimd_ycc_rgb_convert_sse2  jsimd_ycc_extbgrx_convert_sse2
+%include "jdcolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XBGR_RED
+%define RGB_GREEN  EXT_XBGR_GREEN
+%define RGB_BLUE  EXT_XBGR_BLUE
+%define RGB_PIXELSIZE  EXT_XBGR_PIXELSIZE
+%define jsimd_ycc_rgb_convert_sse2  jsimd_ycc_extxbgr_convert_sse2
+%include "jdcolext-sse2.asm"
+
+%undef RGB_RED
+%undef RGB_GREEN
+%undef RGB_BLUE
+%undef RGB_PIXELSIZE
+%define RGB_RED  EXT_XRGB_RED
+%define RGB_GREEN  EXT_XRGB_GREEN
+%define RGB_BLUE  EXT_XRGB_BLUE
+%define RGB_PIXELSIZE  EXT_XRGB_PIXELSIZE
+%define jsimd_ycc_rgb_convert_sse2  jsimd_ycc_extxrgb_convert_sse2
+%include "jdcolext-sse2.asm"
--- a/Show More
+++ b/Show More