Version update to 4.8.1.

Use proper branch for CI pipelines.
CI update for 4.8.x branch.
2026-07-23 20:33:04 +04:00 · 2023-09-27 13:18:31 +03:00 · 2023-09-27 12:22:37 +03:00 · 2023-09-27 10:48:19 +03:00 · 2023-09-27 10:10:12 +03:00 · 2023-09-27 10:08:16 +03:00
934 changed files with 107998 additions and 24700 deletions
@@ -0,0 +1,3 @@
+# These are supported funding model platforms
+
+github: opencv
@@ -10,7 +10,7 @@ body:

 - type: textarea
  attributes:
-    label: Descripe the doc issue
+    label: Describe 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: |
@@ -10,7 +10,7 @@ body:

 - type: textarea
  attributes:
-    label: Descripe the feature and motivation
+    label: Describe the feature and motivation
    description: |
      Please provide a clear and concise proposal of the feature and outline the motivation.
  validations:
@@ -1,41 +1,47 @@
-name: PR:4.x
+name: PR:4.8.x

 on:
  pull_request:
    branches:
-      - 4.x
+      - 4.8.x

 jobs:
  Ubuntu2004-ARM64:
-    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-ARM64.yaml@main
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-ARM64.yaml@release_4.8.x

  Ubuntu2004-ARM64-Debug:
-    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-ARM64-Debug.yaml@main
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-ARM64-Debug.yaml@release_4.8.x

  Ubuntu2004-x64:
-    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-U20.yaml@main
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-U20.yaml@release_4.8.x

  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
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-U20-Cuda.yaml@release_4.8.x

  Windows10-x64:
-    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-W10.yaml@main
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-W10.yaml@release_4.8.x
+
+  Windows10-x64-Vulkan:
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-W10-Vulkan.yaml@release_4.8.x

  macOS-ARM64:
-    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-macOS-ARM64.yaml@main
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-macOS-ARM64.yaml@release_4.8.x

  macOS-x64:
-    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-macOS-x86_64.yaml@main
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-macOS-x86_64.yaml@release_4.8.x
+
+  macOS-ARM64-Vulkan:
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-macOS-ARM64-Vulkan.yaml@release_4.8.x

  iOS:
-    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-iOS.yaml@main
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-iOS.yaml@release_4.8.x

  Android:
-    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-Android.yaml@main
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-Android.yaml@release_4.8.x

  TIM-VX:
-    uses: opencv/ci-gha-workflow/.github/workflows/OCV-timvx-backend-tests-4.x.yml@main
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-timvx-backend-tests-4.x.yml@release_4.8.x

  docs:
-    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-docs.yaml@main
+    uses: opencv/ci-gha-workflow/.github/workflows/OCV-PR-4.x-docs.yaml@release_4.8.x
@@ -1,8 +1,8 @@
-# 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")
+# Binaries branch name: ffmpeg/4.x_20230622
+# Binaries were created for OpenCV: 61d48dd0f8d1cc1a115d26998705a61478f64a3c
+ocv_update(FFMPEG_BINARIES_COMMIT "7da61f0695eabf8972a2c302bf1632a3d99fb0d5")
+ocv_update(FFMPEG_FILE_HASH_BIN32 "4aaef1456e282e5ef665d65555f47f56")
+ocv_update(FFMPEG_FILE_HASH_BIN64 "38a638851e064c591ce812e27ed43f1f")
 ocv_update(FFMPEG_FILE_HASH_CMAKE "8862c87496e2e8c375965e1277dee1c7")

 function(download_win_ffmpeg script_var)
@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
@@ -0,0 +1 @@
+Origin: https://github.com/google/flatbuffers/tree/v23.5.9
@@ -0,0 +1,68 @@
+/*
+ * Copyright 2021 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_ALLOCATOR_H_
+#define FLATBUFFERS_ALLOCATOR_H_
+
+#include "flatbuffers/base.h"
+
+namespace flatbuffers {
+
+// Allocator interface. This is flatbuffers-specific and meant only for
+// `vector_downward` usage.
+class Allocator {
+ public:
+  virtual ~Allocator() {}
+
+  // Allocate `size` bytes of memory.
+  virtual uint8_t *allocate(size_t size) = 0;
+
+  // Deallocate `size` bytes of memory at `p` allocated by this allocator.
+  virtual void deallocate(uint8_t *p, size_t size) = 0;
+
+  // Reallocate `new_size` bytes of memory, replacing the old region of size
+  // `old_size` at `p`. In contrast to a normal realloc, this grows downwards,
+  // and is intended specifcally for `vector_downward` use.
+  // `in_use_back` and `in_use_front` indicate how much of `old_size` is
+  // actually in use at each end, and needs to be copied.
+  virtual uint8_t *reallocate_downward(uint8_t *old_p, size_t old_size,
+                                       size_t new_size, size_t in_use_back,
+                                       size_t in_use_front) {
+    FLATBUFFERS_ASSERT(new_size > old_size);  // vector_downward only grows
+    uint8_t *new_p = allocate(new_size);
+    memcpy_downward(old_p, old_size, new_p, new_size, in_use_back,
+                    in_use_front);
+    deallocate(old_p, old_size);
+    return new_p;
+  }
+
+ protected:
+  // Called by `reallocate_downward` to copy memory from `old_p` of `old_size`
+  // to `new_p` of `new_size`. Only memory of size `in_use_front` and
+  // `in_use_back` will be copied from the front and back of the old memory
+  // allocation.
+  void memcpy_downward(uint8_t *old_p, size_t old_size, uint8_t *new_p,
+                       size_t new_size, size_t in_use_back,
+                       size_t in_use_front) {
+    memcpy(new_p + new_size - in_use_back, old_p + old_size - in_use_back,
+           in_use_back);
+    memcpy(new_p, old_p, in_use_front);
+  }
+};
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_ALLOCATOR_H_
@@ -0,0 +1,256 @@
+/*
+ * Copyright 2021 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_ARRAY_H_
+#define FLATBUFFERS_ARRAY_H_
+
+#include <cstdint>
+#include <memory>
+
+#include "flatbuffers/base.h"
+#include "flatbuffers/stl_emulation.h"
+#include "flatbuffers/vector.h"
+
+namespace flatbuffers {
+
+// This is used as a helper type for accessing arrays.
+template<typename T, uint16_t length> class Array {
+  // Array<T> can carry only POD data types (scalars or structs).
+  typedef typename flatbuffers::bool_constant<flatbuffers::is_scalar<T>::value>
+      scalar_tag;
+  typedef
+      typename flatbuffers::conditional<scalar_tag::value, T, const T *>::type
+          IndirectHelperType;
+
+ public:
+  typedef uint16_t size_type;
+  typedef typename IndirectHelper<IndirectHelperType>::return_type return_type;
+  typedef VectorConstIterator<T, return_type, uoffset_t> const_iterator;
+  typedef VectorReverseIterator<const_iterator> const_reverse_iterator;
+
+  // If T is a LE-scalar or a struct (!scalar_tag::value).
+  static FLATBUFFERS_CONSTEXPR bool is_span_observable =
+      (scalar_tag::value && (FLATBUFFERS_LITTLEENDIAN || sizeof(T) == 1)) ||
+      !scalar_tag::value;
+
+  FLATBUFFERS_CONSTEXPR uint16_t size() const { return length; }
+
+  return_type Get(uoffset_t i) const {
+    FLATBUFFERS_ASSERT(i < size());
+    return IndirectHelper<IndirectHelperType>::Read(Data(), i);
+  }
+
+  return_type operator[](uoffset_t i) const { return Get(i); }
+
+  // If this is a Vector of enums, T will be its storage type, not the enum
+  // type. This function makes it convenient to retrieve value with enum
+  // type E.
+  template<typename E> E GetEnum(uoffset_t i) const {
+    return static_cast<E>(Get(i));
+  }
+
+  const_iterator begin() const { return const_iterator(Data(), 0); }
+  const_iterator end() const { return const_iterator(Data(), size()); }
+
+  const_reverse_iterator rbegin() const {
+    return const_reverse_iterator(end());
+  }
+  const_reverse_iterator rend() const {
+    return const_reverse_iterator(begin());
+  }
+
+  const_iterator cbegin() const { return begin(); }
+  const_iterator cend() const { return end(); }
+
+  const_reverse_iterator crbegin() const { return rbegin(); }
+  const_reverse_iterator crend() const { return rend(); }
+
+  // Get a mutable pointer to elements inside this array.
+  // This method used to mutate arrays of structs followed by a @p Mutate
+  // operation. For primitive types use @p Mutate directly.
+  // @warning Assignments and reads to/from the dereferenced pointer are not
+  //  automatically converted to the correct endianness.
+  typename flatbuffers::conditional<scalar_tag::value, void, T *>::type
+  GetMutablePointer(uoffset_t i) const {
+    FLATBUFFERS_ASSERT(i < size());
+    return const_cast<T *>(&data()[i]);
+  }
+
+  // Change elements if you have a non-const pointer to this object.
+  void Mutate(uoffset_t i, const T &val) { MutateImpl(scalar_tag(), i, val); }
+
+  // The raw data in little endian format. Use with care.
+  const uint8_t *Data() const { return data_; }
+
+  uint8_t *Data() { return data_; }
+
+  // Similarly, but typed, much like std::vector::data
+  const T *data() const { return reinterpret_cast<const T *>(Data()); }
+  T *data() { return reinterpret_cast<T *>(Data()); }
+
+  // Copy data from a span with endian conversion.
+  // If this Array and the span overlap, the behavior is undefined.
+  void CopyFromSpan(flatbuffers::span<const T, length> src) {
+    const auto p1 = reinterpret_cast<const uint8_t *>(src.data());
+    const auto p2 = Data();
+    FLATBUFFERS_ASSERT(!(p1 >= p2 && p1 < (p2 + length)) &&
+                       !(p2 >= p1 && p2 < (p1 + length)));
+    (void)p1;
+    (void)p2;
+    CopyFromSpanImpl(flatbuffers::bool_constant<is_span_observable>(), src);
+  }
+
+ protected:
+  void MutateImpl(flatbuffers::true_type, uoffset_t i, const T &val) {
+    FLATBUFFERS_ASSERT(i < size());
+    WriteScalar(data() + i, val);
+  }
+
+  void MutateImpl(flatbuffers::false_type, uoffset_t i, const T &val) {
+    *(GetMutablePointer(i)) = val;
+  }
+
+  void CopyFromSpanImpl(flatbuffers::true_type,
+                        flatbuffers::span<const T, length> src) {
+    // Use std::memcpy() instead of std::copy() to avoid performance degradation
+    // due to aliasing if T is char or unsigned char.
+    // The size is known at compile time, so memcpy would be inlined.
+    std::memcpy(data(), src.data(), length * sizeof(T));
+  }
+
+  // Copy data from flatbuffers::span with endian conversion.
+  void CopyFromSpanImpl(flatbuffers::false_type,
+                        flatbuffers::span<const T, length> src) {
+    for (size_type k = 0; k < length; k++) { Mutate(k, src[k]); }
+  }
+
+  // This class is only used to access pre-existing data. Don't ever
+  // try to construct these manually.
+  // 'constexpr' allows us to use 'size()' at compile time.
+  // @note Must not use 'FLATBUFFERS_CONSTEXPR' here, as const is not allowed on
+  //  a constructor.
+#if defined(__cpp_constexpr)
+  constexpr Array();
+#else
+  Array();
+#endif
+
+  uint8_t data_[length * sizeof(T)];
+
+ private:
+  // This class is a pointer. Copying will therefore create an invalid object.
+  // Private and unimplemented copy constructor.
+  Array(const Array &);
+  Array &operator=(const Array &);
+};
+
+// Specialization for Array[struct] with access using Offset<void> pointer.
+// This specialization used by idl_gen_text.cpp.
+template<typename T, uint16_t length, template<typename> class OffsetT>
+class Array<OffsetT<T>, length> {
+  static_assert(flatbuffers::is_same<T, void>::value, "unexpected type T");
+
+ public:
+  typedef const void *return_type;
+  typedef uint16_t size_type;
+
+  const uint8_t *Data() const { return data_; }
+
+  // Make idl_gen_text.cpp::PrintContainer happy.
+  return_type operator[](uoffset_t) const {
+    FLATBUFFERS_ASSERT(false);
+    return nullptr;
+  }
+
+ private:
+  // This class is only used to access pre-existing data.
+  Array();
+  Array(const Array &);
+  Array &operator=(const Array &);
+
+  uint8_t data_[1];
+};
+
+template<class U, uint16_t N>
+FLATBUFFERS_CONSTEXPR_CPP11 flatbuffers::span<U, N> make_span(Array<U, N> &arr)
+    FLATBUFFERS_NOEXCEPT {
+  static_assert(
+      Array<U, N>::is_span_observable,
+      "wrong type U, only plain struct, LE-scalar, or byte types are allowed");
+  return span<U, N>(arr.data(), N);
+}
+
+template<class U, uint16_t N>
+FLATBUFFERS_CONSTEXPR_CPP11 flatbuffers::span<const U, N> make_span(
+    const Array<U, N> &arr) FLATBUFFERS_NOEXCEPT {
+  static_assert(
+      Array<U, N>::is_span_observable,
+      "wrong type U, only plain struct, LE-scalar, or byte types are allowed");
+  return span<const U, N>(arr.data(), N);
+}
+
+template<class U, uint16_t N>
+FLATBUFFERS_CONSTEXPR_CPP11 flatbuffers::span<uint8_t, sizeof(U) * N>
+make_bytes_span(Array<U, N> &arr) FLATBUFFERS_NOEXCEPT {
+  static_assert(Array<U, N>::is_span_observable,
+                "internal error, Array<T> might hold only scalars or structs");
+  return span<uint8_t, sizeof(U) * N>(arr.Data(), sizeof(U) * N);
+}
+
+template<class U, uint16_t N>
+FLATBUFFERS_CONSTEXPR_CPP11 flatbuffers::span<const uint8_t, sizeof(U) * N>
+make_bytes_span(const Array<U, N> &arr) FLATBUFFERS_NOEXCEPT {
+  static_assert(Array<U, N>::is_span_observable,
+                "internal error, Array<T> might hold only scalars or structs");
+  return span<const uint8_t, sizeof(U) * N>(arr.Data(), sizeof(U) * N);
+}
+
+// Cast a raw T[length] to a raw flatbuffers::Array<T, length>
+// without endian conversion. Use with care.
+// TODO: move these Cast-methods to `internal` namespace.
+template<typename T, uint16_t length>
+Array<T, length> &CastToArray(T (&arr)[length]) {
+  return *reinterpret_cast<Array<T, length> *>(arr);
+}
+
+template<typename T, uint16_t length>
+const Array<T, length> &CastToArray(const T (&arr)[length]) {
+  return *reinterpret_cast<const Array<T, length> *>(arr);
+}
+
+template<typename E, typename T, uint16_t length>
+Array<E, length> &CastToArrayOfEnum(T (&arr)[length]) {
+  static_assert(sizeof(E) == sizeof(T), "invalid enum type E");
+  return *reinterpret_cast<Array<E, length> *>(arr);
+}
+
+template<typename E, typename T, uint16_t length>
+const Array<E, length> &CastToArrayOfEnum(const T (&arr)[length]) {
+  static_assert(sizeof(E) == sizeof(T), "invalid enum type E");
+  return *reinterpret_cast<const Array<E, length> *>(arr);
+}
+
+template<typename T, uint16_t length>
+bool operator==(const Array<T, length> &lhs,
+                const Array<T, length> &rhs) noexcept {
+  return std::addressof(lhs) == std::addressof(rhs) ||
+         (lhs.size() == rhs.size() &&
+          std::memcmp(lhs.Data(), rhs.Data(), rhs.size() * sizeof(T)) == 0);
+}
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_ARRAY_H_
@@ -0,0 +1,495 @@
+#ifndef FLATBUFFERS_BASE_H_
+#define FLATBUFFERS_BASE_H_
+
+// clang-format off
+
+// If activate should be declared and included first.
+#if defined(FLATBUFFERS_MEMORY_LEAK_TRACKING) && \
+    defined(_MSC_VER) && defined(_DEBUG)
+  // The _CRTDBG_MAP_ALLOC inside <crtdbg.h> will replace
+  // calloc/free (etc) to its debug version using #define directives.
+  #define _CRTDBG_MAP_ALLOC
+  #include <stdlib.h>
+  #include <crtdbg.h>
+  // Replace operator new by trace-enabled version.
+  #define DEBUG_NEW new(_NORMAL_BLOCK, __FILE__, __LINE__)
+  #define new DEBUG_NEW
+#endif
+
+#if !defined(FLATBUFFERS_ASSERT)
+#include <assert.h>
+#define FLATBUFFERS_ASSERT assert
+#elif defined(FLATBUFFERS_ASSERT_INCLUDE)
+// Include file with forward declaration
+#include FLATBUFFERS_ASSERT_INCLUDE
+#endif
+
+#ifndef ARDUINO
+#include <cstdint>
+#endif
+
+#include <cstddef>
+#include <cstdlib>
+#include <cstring>
+
+#if defined(ARDUINO) && !defined(ARDUINOSTL_M_H) && defined(__AVR__)
+  #include <utility.h>
+#else
+  #include <utility>
+#endif
+
+#include <string>
+#include <type_traits>
+#include <vector>
+#include <set>
+#include <algorithm>
+#include <limits>
+#include <iterator>
+#include <memory>
+
+#if defined(__unix__) && !defined(FLATBUFFERS_LOCALE_INDEPENDENT)
+  #include <unistd.h>
+#endif
+
+#ifdef __ANDROID__
+  #include <android/api-level.h>
+#endif
+
+#if defined(__ICCARM__)
+#include <intrinsics.h>
+#endif
+
+// Note the __clang__ check is needed, because clang presents itself
+// as an older GNUC compiler (4.2).
+// Clang 3.3 and later implement all of the ISO C++ 2011 standard.
+// Clang 3.4 and later implement all of the ISO C++ 2014 standard.
+// http://clang.llvm.org/cxx_status.html
+
+// Note the MSVC value '__cplusplus' may be incorrect:
+// The '__cplusplus' predefined macro in the MSVC stuck at the value 199711L,
+// indicating (erroneously!) that the compiler conformed to the C++98 Standard.
+// This value should be correct starting from MSVC2017-15.7-Preview-3.
+// The '__cplusplus' will be valid only if MSVC2017-15.7-P3 and the `/Zc:__cplusplus` switch is set.
+// Workaround (for details see MSDN):
+// Use the _MSC_VER and _MSVC_LANG definition instead of the __cplusplus  for compatibility.
+// The _MSVC_LANG macro reports the Standard version regardless of the '/Zc:__cplusplus' switch.
+
+#if defined(__GNUC__) && !defined(__clang__)
+  #define FLATBUFFERS_GCC (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__)
+#else
+  #define FLATBUFFERS_GCC 0
+#endif
+
+#if defined(__clang__)
+  #define FLATBUFFERS_CLANG (__clang_major__ * 10000 + __clang_minor__ * 100 + __clang_patchlevel__)
+#else
+  #define FLATBUFFERS_CLANG 0
+#endif
+
+/// @cond FLATBUFFERS_INTERNAL
+#if __cplusplus <= 199711L && \
+    (!defined(_MSC_VER) || _MSC_VER < 1600) && \
+    (!defined(__GNUC__) || \
+      (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__ < 40400))
+  #error A C++11 compatible compiler with support for the auto typing is \
+         required for FlatBuffers.
+  #error __cplusplus _MSC_VER __GNUC__  __GNUC_MINOR__  __GNUC_PATCHLEVEL__
+#endif
+
+#if !defined(__clang__) && \
+    defined(__GNUC__) && \
+    (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__ < 40600)
+  // Backwards compatibility for g++ 4.4, and 4.5 which don't have the nullptr
+  // and constexpr keywords. Note the __clang__ check is needed, because clang
+  // presents itself as an older GNUC compiler.
+  #ifndef nullptr_t
+    const class nullptr_t {
+    public:
+      template<class T> inline operator T*() const { return 0; }
+    private:
+      void operator&() const;
+    } nullptr = {};
+  #endif
+  #ifndef constexpr
+    #define constexpr const
+  #endif
+#endif
+
+// The wire format uses a little endian encoding (since that's efficient for
+// the common platforms).
+#if defined(__s390x__)
+  #define FLATBUFFERS_LITTLEENDIAN 0
+#endif // __s390x__
+#if !defined(FLATBUFFERS_LITTLEENDIAN)
+  #if defined(__GNUC__) || defined(__clang__) || defined(__ICCARM__)
+    #if (defined(__BIG_ENDIAN__) || \
+         (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__))
+      #define FLATBUFFERS_LITTLEENDIAN 0
+    #else
+      #define FLATBUFFERS_LITTLEENDIAN 1
+    #endif // __BIG_ENDIAN__
+  #elif defined(_MSC_VER)
+    #if defined(_M_PPC)
+      #define FLATBUFFERS_LITTLEENDIAN 0
+    #else
+      #define FLATBUFFERS_LITTLEENDIAN 1
+    #endif
+  #else
+    #error Unable to determine endianness, define FLATBUFFERS_LITTLEENDIAN.
+  #endif
+#endif // !defined(FLATBUFFERS_LITTLEENDIAN)
+
+#define FLATBUFFERS_VERSION_MAJOR 23
+#define FLATBUFFERS_VERSION_MINOR 5
+#define FLATBUFFERS_VERSION_REVISION 9
+#define FLATBUFFERS_STRING_EXPAND(X) #X
+#define FLATBUFFERS_STRING(X) FLATBUFFERS_STRING_EXPAND(X)
+namespace flatbuffers {
+  // Returns version as string  "MAJOR.MINOR.REVISION".
+  const char* FLATBUFFERS_VERSION();
+}
+
+#if (!defined(_MSC_VER) || _MSC_VER > 1600) && \
+    (!defined(__GNUC__) || (__GNUC__ * 100 + __GNUC_MINOR__ >= 407)) || \
+    defined(__clang__)
+  #define FLATBUFFERS_FINAL_CLASS final
+  #define FLATBUFFERS_OVERRIDE override
+  #define FLATBUFFERS_EXPLICIT_CPP11 explicit
+  #define FLATBUFFERS_VTABLE_UNDERLYING_TYPE : flatbuffers::voffset_t
+#else
+  #define FLATBUFFERS_FINAL_CLASS
+  #define FLATBUFFERS_OVERRIDE
+  #define FLATBUFFERS_EXPLICIT_CPP11
+  #define FLATBUFFERS_VTABLE_UNDERLYING_TYPE
+#endif
+
+#if (!defined(_MSC_VER) || _MSC_VER >= 1900) && \
+    (!defined(__GNUC__) || (__GNUC__ * 100 + __GNUC_MINOR__ >= 406)) || \
+    (defined(__cpp_constexpr) && __cpp_constexpr >= 200704)
+  #define FLATBUFFERS_CONSTEXPR constexpr
+  #define FLATBUFFERS_CONSTEXPR_CPP11 constexpr
+  #define FLATBUFFERS_CONSTEXPR_DEFINED
+#else
+  #define FLATBUFFERS_CONSTEXPR const
+  #define FLATBUFFERS_CONSTEXPR_CPP11
+#endif
+
+#if (defined(__cplusplus) && __cplusplus >= 201402L) || \
+    (defined(__cpp_constexpr) && __cpp_constexpr >= 201304)
+  #define FLATBUFFERS_CONSTEXPR_CPP14 FLATBUFFERS_CONSTEXPR_CPP11
+#else
+  #define FLATBUFFERS_CONSTEXPR_CPP14
+#endif
+
+#if (defined(__GXX_EXPERIMENTAL_CXX0X__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 406)) || \
+    (defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 190023026)) || \
+    defined(__clang__)
+  #define FLATBUFFERS_NOEXCEPT noexcept
+#else
+  #define FLATBUFFERS_NOEXCEPT
+#endif
+
+// NOTE: the FLATBUFFERS_DELETE_FUNC macro may change the access mode to
+// private, so be sure to put it at the end or reset access mode explicitly.
+#if (!defined(_MSC_VER) || _MSC_FULL_VER >= 180020827) && \
+    (!defined(__GNUC__) || (__GNUC__ * 100 + __GNUC_MINOR__ >= 404)) || \
+    defined(__clang__)
+  #define FLATBUFFERS_DELETE_FUNC(func) func = delete
+#else
+  #define FLATBUFFERS_DELETE_FUNC(func) private: func
+#endif
+
+#if (!defined(_MSC_VER) || _MSC_VER >= 1900) && \
+    (!defined(__GNUC__) || (__GNUC__ * 100 + __GNUC_MINOR__ >= 409)) || \
+    defined(__clang__)
+  #define FLATBUFFERS_DEFAULT_DECLARATION
+#endif
+
+// Check if we can use template aliases
+// Not possible if Microsoft Compiler before 2012
+// Possible is the language feature __cpp_alias_templates is defined well
+// Or possible if the C++ std is C+11 or newer
+#if (defined(_MSC_VER) && _MSC_VER > 1700 /* MSVC2012 */) \
+    || (defined(__cpp_alias_templates) && __cpp_alias_templates >= 200704) \
+    || (defined(__cplusplus) && __cplusplus >= 201103L)
+  #define FLATBUFFERS_TEMPLATES_ALIASES
+#endif
+
+#ifndef FLATBUFFERS_HAS_STRING_VIEW
+  // Only provide flatbuffers::string_view if __has_include can be used
+  // to detect a header that provides an implementation
+  #if defined(__has_include)
+    // Check for std::string_view (in c++17)
+    #if __has_include(<string_view>) && (__cplusplus >= 201606 || (defined(_HAS_CXX17) && _HAS_CXX17))
+      #include <string_view>
+      namespace flatbuffers {
+        typedef std::string_view string_view;
+      }
+      #define FLATBUFFERS_HAS_STRING_VIEW 1
+    // Check for std::experimental::string_view (in c++14, compiler-dependent)
+    #elif __has_include(<experimental/string_view>) && (__cplusplus >= 201411)
+      #include <experimental/string_view>
+      namespace flatbuffers {
+        typedef std::experimental::string_view string_view;
+      }
+      #define FLATBUFFERS_HAS_STRING_VIEW 1
+    // Check for absl::string_view
+    #elif __has_include("absl/strings/string_view.h") && \
+          __has_include("absl/base/config.h") && \
+          (__cplusplus >= 201411)
+      #include "absl/base/config.h"
+      #if !defined(ABSL_USES_STD_STRING_VIEW)
+        #include "absl/strings/string_view.h"
+        namespace flatbuffers {
+          typedef absl::string_view string_view;
+        }
+        #define FLATBUFFERS_HAS_STRING_VIEW 1
+      #endif
+    #endif
+  #endif // __has_include
+#endif // !FLATBUFFERS_HAS_STRING_VIEW
+
+#ifndef FLATBUFFERS_GENERAL_HEAP_ALLOC_OK
+  // Allow heap allocations to be used
+  #define FLATBUFFERS_GENERAL_HEAP_ALLOC_OK 1
+#endif // !FLATBUFFERS_GENERAL_HEAP_ALLOC_OK
+
+#ifndef FLATBUFFERS_HAS_NEW_STRTOD
+  // Modern (C++11) strtod and strtof functions are available for use.
+  // 1) nan/inf strings as argument of strtod;
+  // 2) hex-float  as argument of  strtod/strtof.
+  #if (defined(_MSC_VER) && _MSC_VER >= 1900) || \
+      (defined(__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 409)) || \
+      (defined(__clang__))
+    #define FLATBUFFERS_HAS_NEW_STRTOD 1
+  #endif
+#endif // !FLATBUFFERS_HAS_NEW_STRTOD
+
+#ifndef FLATBUFFERS_LOCALE_INDEPENDENT
+  // Enable locale independent functions {strtof_l, strtod_l,strtoll_l,
+  // strtoull_l}.
+  #if (defined(_MSC_VER) && _MSC_VER >= 1800) || \
+      (defined(__ANDROID_API__) && __ANDROID_API__>= 21) || \
+      (defined(_XOPEN_VERSION) && (_XOPEN_VERSION >= 700)) && \
+        (!defined(__Fuchsia__) && !defined(__ANDROID_API__))
+    #define FLATBUFFERS_LOCALE_INDEPENDENT 1
+  #else
+    #define FLATBUFFERS_LOCALE_INDEPENDENT 0
+  #endif
+#endif  // !FLATBUFFERS_LOCALE_INDEPENDENT
+
+// Suppress Undefined Behavior Sanitizer (recoverable only). Usage:
+// - __suppress_ubsan__("undefined")
+// - __suppress_ubsan__("signed-integer-overflow")
+#if defined(__clang__) && (__clang_major__ > 3 || (__clang_major__ == 3 && __clang_minor__ >=7))
+  #define __suppress_ubsan__(type) __attribute__((no_sanitize(type)))
+#elif defined(__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 409)
+  #define __suppress_ubsan__(type) __attribute__((no_sanitize_undefined))
+#else
+  #define __suppress_ubsan__(type)
+#endif
+
+// This is constexpr function used for checking compile-time constants.
+// Avoid `#pragma warning(disable: 4127) // C4127: expression is constant`.
+template<typename T> FLATBUFFERS_CONSTEXPR inline bool IsConstTrue(T t) {
+  return !!t;
+}
+
+// Enable C++ attribute [[]] if std:c++17 or higher.
+#if ((__cplusplus >= 201703L) \
+    || (defined(_MSVC_LANG) &&  (_MSVC_LANG >= 201703L)))
+  // All attributes unknown to an implementation are ignored without causing an error.
+  #define FLATBUFFERS_ATTRIBUTE(attr) attr
+
+  #define FLATBUFFERS_FALLTHROUGH() [[fallthrough]]
+#else
+  #define FLATBUFFERS_ATTRIBUTE(attr)
+
+  #if FLATBUFFERS_CLANG >= 30800
+    #define FLATBUFFERS_FALLTHROUGH() [[clang::fallthrough]]
+  #elif FLATBUFFERS_GCC >= 70300
+    #define FLATBUFFERS_FALLTHROUGH() [[gnu::fallthrough]]
+  #else
+    #define FLATBUFFERS_FALLTHROUGH()
+  #endif
+#endif
+
+/// @endcond
+
+/// @file
+namespace flatbuffers {
+
+/// @cond FLATBUFFERS_INTERNAL
+// Our default offset / size type, 32bit on purpose on 64bit systems.
+// Also, using a consistent offset type maintains compatibility of serialized
+// offset values between 32bit and 64bit systems.
+typedef uint32_t uoffset_t;
+typedef uint64_t uoffset64_t;
+
+// Signed offsets for references that can go in both directions.
+typedef int32_t soffset_t;
+typedef int64_t soffset64_t;
+
+// Offset/index used in v-tables, can be changed to uint8_t in
+// format forks to save a bit of space if desired.
+typedef uint16_t voffset_t;
+
+typedef uintmax_t largest_scalar_t;
+
+// In 32bits, this evaluates to 2GB - 1
+#define FLATBUFFERS_MAX_BUFFER_SIZE std::numeric_limits<::flatbuffers::soffset_t>::max()
+#define FLATBUFFERS_MAX_64_BUFFER_SIZE std::numeric_limits<::flatbuffers::soffset64_t>::max()
+
+// The minimum size buffer that can be a valid flatbuffer.
+// Includes the offset to the root table (uoffset_t), the offset to the vtable
+// of the root table (soffset_t), the size of the vtable (uint16_t), and the
+// size of the referring table (uint16_t).
+#define FLATBUFFERS_MIN_BUFFER_SIZE sizeof(uoffset_t) + sizeof(soffset_t) + \
+   sizeof(uint16_t) + sizeof(uint16_t)
+
+// We support aligning the contents of buffers up to this size.
+#ifndef FLATBUFFERS_MAX_ALIGNMENT
+  #define FLATBUFFERS_MAX_ALIGNMENT 32
+#endif
+
+/// @brief The length of a FlatBuffer file header.
+static const size_t kFileIdentifierLength = 4;
+
+inline bool VerifyAlignmentRequirements(size_t align, size_t min_align = 1) {
+  return (min_align <= align) && (align <= (FLATBUFFERS_MAX_ALIGNMENT)) &&
+         (align & (align - 1)) == 0;  // must be power of 2
+}
+
+#if defined(_MSC_VER)
+  #pragma warning(disable: 4351) // C4351: new behavior: elements of array ... will be default initialized
+  #pragma warning(push)
+  #pragma warning(disable: 4127) // C4127: conditional expression is constant
+#endif
+
+template<typename T> T EndianSwap(T t) {
+  #if defined(_MSC_VER)
+    #define FLATBUFFERS_BYTESWAP16 _byteswap_ushort
+    #define FLATBUFFERS_BYTESWAP32 _byteswap_ulong
+    #define FLATBUFFERS_BYTESWAP64 _byteswap_uint64
+  #elif defined(__ICCARM__)
+    #define FLATBUFFERS_BYTESWAP16 __REV16
+    #define FLATBUFFERS_BYTESWAP32 __REV
+    #define FLATBUFFERS_BYTESWAP64(x) \
+       ((__REV(static_cast<uint32_t>(x >> 32U))) | (static_cast<uint64_t>(__REV(static_cast<uint32_t>(x)))) << 32U)
+  #else
+    #if defined(__GNUC__) && __GNUC__ * 100 + __GNUC_MINOR__ < 408 && !defined(__clang__)
+      // __builtin_bswap16 was missing prior to GCC 4.8.
+      #define FLATBUFFERS_BYTESWAP16(x) \
+        static_cast<uint16_t>(__builtin_bswap32(static_cast<uint32_t>(x) << 16))
+    #else
+      #define FLATBUFFERS_BYTESWAP16 __builtin_bswap16
+    #endif
+    #define FLATBUFFERS_BYTESWAP32 __builtin_bswap32
+    #define FLATBUFFERS_BYTESWAP64 __builtin_bswap64
+  #endif
+  if (sizeof(T) == 1) {   // Compile-time if-then's.
+    return t;
+  } else if (sizeof(T) == 2) {
+    union { T t; uint16_t i; } u = { t };
+    u.i = FLATBUFFERS_BYTESWAP16(u.i);
+    return u.t;
+  } else if (sizeof(T) == 4) {
+    union { T t; uint32_t i; } u = { t };
+    u.i = FLATBUFFERS_BYTESWAP32(u.i);
+    return u.t;
+  } else if (sizeof(T) == 8) {
+    union { T t; uint64_t i; } u = { t };
+    u.i = FLATBUFFERS_BYTESWAP64(u.i);
+    return u.t;
+  } else {
+    FLATBUFFERS_ASSERT(0);
+    return t;
+  }
+}
+
+#if defined(_MSC_VER)
+  #pragma warning(pop)
+#endif
+
+
+template<typename T> T EndianScalar(T t) {
+  #if FLATBUFFERS_LITTLEENDIAN
+    return t;
+  #else
+    return EndianSwap(t);
+  #endif
+}
+
+template<typename T>
+// UBSAN: C++ aliasing type rules, see std::bit_cast<> for details.
+__suppress_ubsan__("alignment")
+T ReadScalar(const void *p) {
+  return EndianScalar(*reinterpret_cast<const T *>(p));
+}
+
+// See https://github.com/google/flatbuffers/issues/5950
+
+#if (FLATBUFFERS_GCC >= 100000) && (FLATBUFFERS_GCC < 110000)
+  #pragma GCC diagnostic push
+  #pragma GCC diagnostic ignored "-Wstringop-overflow"
+#endif
+
+template<typename T>
+// UBSAN: C++ aliasing type rules, see std::bit_cast<> for details.
+__suppress_ubsan__("alignment")
+void WriteScalar(void *p, T t) {
+  *reinterpret_cast<T *>(p) = EndianScalar(t);
+}
+
+template<typename T> struct Offset;
+template<typename T> __suppress_ubsan__("alignment") void WriteScalar(void *p, Offset<T> t) {
+  *reinterpret_cast<uoffset_t *>(p) = EndianScalar(t.o);
+}
+
+#if (FLATBUFFERS_GCC >= 100000) && (FLATBUFFERS_GCC < 110000)
+  #pragma GCC diagnostic pop
+#endif
+
+// Computes how many bytes you'd have to pad to be able to write an
+// "scalar_size" scalar if the buffer had grown to "buf_size" (downwards in
+// memory).
+__suppress_ubsan__("unsigned-integer-overflow")
+inline size_t PaddingBytes(size_t buf_size, size_t scalar_size) {
+  return ((~buf_size) + 1) & (scalar_size - 1);
+}
+
+// Generic 'operator==' with conditional specialisations.
+// T e - new value of a scalar field.
+// T def - default of scalar (is known at compile-time).
+template<typename T> inline bool IsTheSameAs(T e, T def) { return e == def; }
+
+#if defined(FLATBUFFERS_NAN_DEFAULTS) && \
+    defined(FLATBUFFERS_HAS_NEW_STRTOD) && (FLATBUFFERS_HAS_NEW_STRTOD > 0)
+// Like `operator==(e, def)` with weak NaN if T=(float|double).
+template<typename T> inline bool IsFloatTheSameAs(T e, T def) {
+  return (e == def) || ((def != def) && (e != e));
+}
+template<> inline bool IsTheSameAs<float>(float e, float def) {
+  return IsFloatTheSameAs(e, def);
+}
+template<> inline bool IsTheSameAs<double>(double e, double def) {
+  return IsFloatTheSameAs(e, def);
+}
+#endif
+
+// Check 'v' is out of closed range [low; high].
+// Workaround for GCC warning [-Werror=type-limits]:
+// comparison is always true due to limited range of data type.
+template<typename T>
+inline bool IsOutRange(const T &v, const T &low, const T &high) {
+  return (v < low) || (high < v);
+}
+
+// Check 'v' is in closed range [low; high].
+template<typename T>
+inline bool IsInRange(const T &v, const T &low, const T &high) {
+  return !IsOutRange(v, low, high);
+}
+
+}  // namespace flatbuffers
+#endif  // FLATBUFFERS_BASE_H_
@@ -0,0 +1,199 @@
+/*
+ * Copyright 2021 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_BUFFER_H_
+#define FLATBUFFERS_BUFFER_H_
+
+#include <algorithm>
+
+#include "flatbuffers/base.h"
+
+namespace flatbuffers {
+
+// Wrapper for uoffset_t to allow safe template specialization.
+// Value is allowed to be 0 to indicate a null object (see e.g. AddOffset).
+template<typename T = void> struct Offset {
+  // The type of offset to use.
+  typedef uoffset_t offset_type;
+
+  offset_type o;
+  Offset() : o(0) {}
+  Offset(const offset_type _o) : o(_o) {}
+  Offset<> Union() const { return o; }
+  bool IsNull() const { return !o; }
+};
+
+// Wrapper for uoffset64_t Offsets.
+template<typename T = void> struct Offset64 {
+  // The type of offset to use.
+  typedef uoffset64_t offset_type;
+
+  offset_type o;
+  Offset64() : o(0) {}
+  Offset64(const offset_type offset) : o(offset) {}
+  Offset64<> Union() const { return o; }
+  bool IsNull() const { return !o; }
+};
+
+// Litmus check for ensuring the Offsets are the expected size.
+static_assert(sizeof(Offset<>) == 4, "Offset has wrong size");
+static_assert(sizeof(Offset64<>) == 8, "Offset64 has wrong size");
+
+inline void EndianCheck() {
+  int endiantest = 1;
+  // If this fails, see FLATBUFFERS_LITTLEENDIAN above.
+  FLATBUFFERS_ASSERT(*reinterpret_cast<char *>(&endiantest) ==
+                     FLATBUFFERS_LITTLEENDIAN);
+  (void)endiantest;
+}
+
+template<typename T> FLATBUFFERS_CONSTEXPR size_t AlignOf() {
+  // clang-format off
+  #ifdef _MSC_VER
+    return __alignof(T);
+  #else
+    #ifndef alignof
+      return __alignof__(T);
+    #else
+      return alignof(T);
+    #endif
+  #endif
+  // clang-format on
+}
+
+// Lexicographically compare two strings (possibly containing nulls), and
+// return true if the first is less than the second.
+static inline bool StringLessThan(const char *a_data, uoffset_t a_size,
+                                  const char *b_data, uoffset_t b_size) {
+  const auto cmp = memcmp(a_data, b_data, (std::min)(a_size, b_size));
+  return cmp == 0 ? a_size < b_size : cmp < 0;
+}
+
+// When we read serialized data from memory, in the case of most scalars,
+// we want to just read T, but in the case of Offset, we want to actually
+// perform the indirection and return a pointer.
+// The template specialization below does just that.
+// It is wrapped in a struct since function templates can't overload on the
+// return type like this.
+// The typedef is for the convenience of callers of this function
+// (avoiding the need for a trailing return decltype)
+template<typename T> struct IndirectHelper {
+  typedef T return_type;
+  typedef T mutable_return_type;
+  static const size_t element_stride = sizeof(T);
+
+  static return_type Read(const uint8_t *p, const size_t i) {
+    return EndianScalar((reinterpret_cast<const T *>(p))[i]);
+  }
+  static mutable_return_type Read(uint8_t *p, const size_t i) {
+    return reinterpret_cast<mutable_return_type>(
+        Read(const_cast<const uint8_t *>(p), i));
+  }
+};
+
+// For vector of Offsets.
+template<typename T, template<typename> class OffsetT>
+struct IndirectHelper<OffsetT<T>> {
+  typedef const T *return_type;
+  typedef T *mutable_return_type;
+  typedef typename OffsetT<T>::offset_type offset_type;
+  static const offset_type element_stride = sizeof(offset_type);
+
+  static return_type Read(const uint8_t *const p, const offset_type i) {
+    // Offsets are relative to themselves, so first update the pointer to
+    // point to the offset location.
+    const uint8_t *const offset_location = p + i * element_stride;
+
+    // Then read the scalar value of the offset (which may be 32 or 64-bits) and
+    // then determine the relative location from the offset location.
+    return reinterpret_cast<return_type>(
+        offset_location + ReadScalar<offset_type>(offset_location));
+  }
+  static mutable_return_type Read(uint8_t *const p, const offset_type i) {
+    // Offsets are relative to themselves, so first update the pointer to
+    // point to the offset location.
+    uint8_t *const offset_location = p + i * element_stride;
+
+    // Then read the scalar value of the offset (which may be 32 or 64-bits) and
+    // then determine the relative location from the offset location.
+    return reinterpret_cast<mutable_return_type>(
+        offset_location + ReadScalar<offset_type>(offset_location));
+  }
+};
+
+// For vector of structs.
+template<typename T> struct IndirectHelper<const T *> {
+  typedef const T *return_type;
+  typedef T *mutable_return_type;
+  static const size_t element_stride = sizeof(T);
+
+  static return_type Read(const uint8_t *const p, const size_t i) {
+    // Structs are stored inline, relative to the first struct pointer.
+    return reinterpret_cast<return_type>(p + i * element_stride);
+  }
+  static mutable_return_type Read(uint8_t *const p, const size_t i) {
+    // Structs are stored inline, relative to the first struct pointer.
+    return reinterpret_cast<mutable_return_type>(p + i * element_stride);
+  }
+};
+
+/// @brief Get a pointer to the file_identifier section of the buffer.
+/// @return Returns a const char pointer to the start of the file_identifier
+/// characters in the buffer.  The returned char * has length
+/// 'flatbuffers::FlatBufferBuilder::kFileIdentifierLength'.
+/// This function is UNDEFINED for FlatBuffers whose schema does not include
+/// a file_identifier (likely points at padding or the start of a the root
+/// vtable).
+inline const char *GetBufferIdentifier(const void *buf,
+                                       bool size_prefixed = false) {
+  return reinterpret_cast<const char *>(buf) +
+         ((size_prefixed) ? 2 * sizeof(uoffset_t) : sizeof(uoffset_t));
+}
+
+// Helper to see if the identifier in a buffer has the expected value.
+inline bool BufferHasIdentifier(const void *buf, const char *identifier,
+                                bool size_prefixed = false) {
+  return strncmp(GetBufferIdentifier(buf, size_prefixed), identifier,
+                 flatbuffers::kFileIdentifierLength) == 0;
+}
+
+/// @cond FLATBUFFERS_INTERNAL
+// Helpers to get a typed pointer to the root object contained in the buffer.
+template<typename T> T *GetMutableRoot(void *buf) {
+  if (!buf) return nullptr;
+  EndianCheck();
+  return reinterpret_cast<T *>(
+      reinterpret_cast<uint8_t *>(buf) +
+      EndianScalar(*reinterpret_cast<uoffset_t *>(buf)));
+}
+
+template<typename T, typename SizeT = uoffset_t>
+T *GetMutableSizePrefixedRoot(void *buf) {
+  return GetMutableRoot<T>(reinterpret_cast<uint8_t *>(buf) + sizeof(SizeT));
+}
+
+template<typename T> const T *GetRoot(const void *buf) {
+  return GetMutableRoot<T>(const_cast<void *>(buf));
+}
+
+template<typename T, typename SizeT = uoffset_t>
+const T *GetSizePrefixedRoot(const void *buf) {
+  return GetRoot<T>(reinterpret_cast<const uint8_t *>(buf) + sizeof(SizeT));
+}
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_BUFFER_H_
@@ -0,0 +1,53 @@
+/*
+ * Copyright 2021 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_BUFFER_REF_H_
+#define FLATBUFFERS_BUFFER_REF_H_
+
+#include "flatbuffers/base.h"
+#include "flatbuffers/verifier.h"
+
+namespace flatbuffers {
+
+// Convenient way to bundle a buffer and its length, to pass it around
+// typed by its root.
+// A BufferRef does not own its buffer.
+struct BufferRefBase {};  // for std::is_base_of
+
+template<typename T> struct BufferRef : BufferRefBase {
+  BufferRef() : buf(nullptr), len(0), must_free(false) {}
+  BufferRef(uint8_t *_buf, uoffset_t _len)
+      : buf(_buf), len(_len), must_free(false) {}
+
+  ~BufferRef() {
+    if (must_free) free(buf);
+  }
+
+  const T *GetRoot() const { return flatbuffers::GetRoot<T>(buf); }
+
+  bool Verify() {
+    Verifier verifier(buf, len);
+    return verifier.VerifyBuffer<T>(nullptr);
+  }
+
+  uint8_t *buf;
+  uoffset_t len;
+  bool must_free;
+};
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_BUFFER_REF_H_
@@ -0,0 +1,64 @@
+/*
+ * Copyright 2021 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_DEFAULT_ALLOCATOR_H_
+#define FLATBUFFERS_DEFAULT_ALLOCATOR_H_
+
+#include "flatbuffers/allocator.h"
+#include "flatbuffers/base.h"
+
+namespace flatbuffers {
+
+// DefaultAllocator uses new/delete to allocate memory regions
+class DefaultAllocator : public Allocator {
+ public:
+  uint8_t *allocate(size_t size) FLATBUFFERS_OVERRIDE {
+    return new uint8_t[size];
+  }
+
+  void deallocate(uint8_t *p, size_t) FLATBUFFERS_OVERRIDE { delete[] p; }
+
+  static void dealloc(void *p, size_t) { delete[] static_cast<uint8_t *>(p); }
+};
+
+// These functions allow for a null allocator to mean use the default allocator,
+// as used by DetachedBuffer and vector_downward below.
+// This is to avoid having a statically or dynamically allocated default
+// allocator, or having to move it between the classes that may own it.
+inline uint8_t *Allocate(Allocator *allocator, size_t size) {
+  return allocator ? allocator->allocate(size)
+                   : DefaultAllocator().allocate(size);
+}
+
+inline void Deallocate(Allocator *allocator, uint8_t *p, size_t size) {
+  if (allocator)
+    allocator->deallocate(p, size);
+  else
+    DefaultAllocator().deallocate(p, size);
+}
+
+inline uint8_t *ReallocateDownward(Allocator *allocator, uint8_t *old_p,
+                                   size_t old_size, size_t new_size,
+                                   size_t in_use_back, size_t in_use_front) {
+  return allocator ? allocator->reallocate_downward(old_p, old_size, new_size,
+                                                    in_use_back, in_use_front)
+                   : DefaultAllocator().reallocate_downward(
+                         old_p, old_size, new_size, in_use_back, in_use_front);
+}
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_DEFAULT_ALLOCATOR_H_
@@ -0,0 +1,114 @@
+/*
+ * Copyright 2021 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_DETACHED_BUFFER_H_
+#define FLATBUFFERS_DETACHED_BUFFER_H_
+
+#include "flatbuffers/allocator.h"
+#include "flatbuffers/base.h"
+#include "flatbuffers/default_allocator.h"
+
+namespace flatbuffers {
+
+// DetachedBuffer is a finished flatbuffer memory region, detached from its
+// builder. The original memory region and allocator are also stored so that
+// the DetachedBuffer can manage the memory lifetime.
+class DetachedBuffer {
+ public:
+  DetachedBuffer()
+      : allocator_(nullptr),
+        own_allocator_(false),
+        buf_(nullptr),
+        reserved_(0),
+        cur_(nullptr),
+        size_(0) {}
+
+  DetachedBuffer(Allocator *allocator, bool own_allocator, uint8_t *buf,
+                 size_t reserved, uint8_t *cur, size_t sz)
+      : allocator_(allocator),
+        own_allocator_(own_allocator),
+        buf_(buf),
+        reserved_(reserved),
+        cur_(cur),
+        size_(sz) {}
+
+  DetachedBuffer(DetachedBuffer &&other) noexcept
+      : allocator_(other.allocator_),
+        own_allocator_(other.own_allocator_),
+        buf_(other.buf_),
+        reserved_(other.reserved_),
+        cur_(other.cur_),
+        size_(other.size_) {
+    other.reset();
+  }
+
+  DetachedBuffer &operator=(DetachedBuffer &&other) noexcept {
+    if (this == &other) return *this;
+
+    destroy();
+
+    allocator_ = other.allocator_;
+    own_allocator_ = other.own_allocator_;
+    buf_ = other.buf_;
+    reserved_ = other.reserved_;
+    cur_ = other.cur_;
+    size_ = other.size_;
+
+    other.reset();
+
+    return *this;
+  }
+
+  ~DetachedBuffer() { destroy(); }
+
+  const uint8_t *data() const { return cur_; }
+
+  uint8_t *data() { return cur_; }
+
+  size_t size() const { return size_; }
+
+  // These may change access mode, leave these at end of public section
+  FLATBUFFERS_DELETE_FUNC(DetachedBuffer(const DetachedBuffer &other));
+  FLATBUFFERS_DELETE_FUNC(
+      DetachedBuffer &operator=(const DetachedBuffer &other));
+
+ protected:
+  Allocator *allocator_;
+  bool own_allocator_;
+  uint8_t *buf_;
+  size_t reserved_;
+  uint8_t *cur_;
+  size_t size_;
+
+  inline void destroy() {
+    if (buf_) Deallocate(allocator_, buf_, reserved_);
+    if (own_allocator_ && allocator_) { delete allocator_; }
+    reset();
+  }
+
+  inline void reset() {
+    allocator_ = nullptr;
+    own_allocator_ = false;
+    buf_ = nullptr;
+    reserved_ = 0;
+    cur_ = nullptr;
+    size_ = 0;
+  }
+};
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_DETACHED_BUFFER_H_
@@ -0,0 +1,273 @@
+/*
+ * Copyright 2014 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_H_
+#define FLATBUFFERS_H_
+
+#include <algorithm>
+
+// TODO: These includes are for mitigating the pains of users editing their
+// source because they relied on flatbuffers.h to include everything for them.
+#include "flatbuffers/array.h"
+#include "flatbuffers/base.h"
+#include "flatbuffers/buffer.h"
+#include "flatbuffers/buffer_ref.h"
+#include "flatbuffers/detached_buffer.h"
+#include "flatbuffers/flatbuffer_builder.h"
+#include "flatbuffers/stl_emulation.h"
+#include "flatbuffers/string.h"
+#include "flatbuffers/struct.h"
+#include "flatbuffers/table.h"
+#include "flatbuffers/vector.h"
+#include "flatbuffers/vector_downward.h"
+#include "flatbuffers/verifier.h"
+
+namespace flatbuffers {
+
+/// @brief This can compute the start of a FlatBuffer from a root pointer, i.e.
+/// it is the opposite transformation of GetRoot().
+/// This may be useful if you want to pass on a root and have the recipient
+/// delete the buffer afterwards.
+inline const uint8_t *GetBufferStartFromRootPointer(const void *root) {
+  auto table = reinterpret_cast<const Table *>(root);
+  auto vtable = table->GetVTable();
+  // Either the vtable is before the root or after the root.
+  auto start = (std::min)(vtable, reinterpret_cast<const uint8_t *>(root));
+  // Align to at least sizeof(uoffset_t).
+  start = reinterpret_cast<const uint8_t *>(reinterpret_cast<uintptr_t>(start) &
+                                            ~(sizeof(uoffset_t) - 1));
+  // Additionally, there may be a file_identifier in the buffer, and the root
+  // offset. The buffer may have been aligned to any size between
+  // sizeof(uoffset_t) and FLATBUFFERS_MAX_ALIGNMENT (see "force_align").
+  // Sadly, the exact alignment is only known when constructing the buffer,
+  // since it depends on the presence of values with said alignment properties.
+  // So instead, we simply look at the next uoffset_t values (root,
+  // file_identifier, and alignment padding) to see which points to the root.
+  // None of the other values can "impersonate" the root since they will either
+  // be 0 or four ASCII characters.
+  static_assert(flatbuffers::kFileIdentifierLength == sizeof(uoffset_t),
+                "file_identifier is assumed to be the same size as uoffset_t");
+  for (auto possible_roots = FLATBUFFERS_MAX_ALIGNMENT / sizeof(uoffset_t) + 1;
+       possible_roots; possible_roots--) {
+    start -= sizeof(uoffset_t);
+    if (ReadScalar<uoffset_t>(start) + start ==
+        reinterpret_cast<const uint8_t *>(root))
+      return start;
+  }
+  // We didn't find the root, either the "root" passed isn't really a root,
+  // or the buffer is corrupt.
+  // Assert, because calling this function with bad data may cause reads
+  // outside of buffer boundaries.
+  FLATBUFFERS_ASSERT(false);
+  return nullptr;
+}
+
+/// @brief This return the prefixed size of a FlatBuffer.
+template<typename SizeT = uoffset_t>
+inline SizeT GetPrefixedSize(const uint8_t *buf) {
+  return ReadScalar<SizeT>(buf);
+}
+
+// Base class for native objects (FlatBuffer data de-serialized into native
+// C++ data structures).
+// Contains no functionality, purely documentative.
+struct NativeTable {};
+
+/// @brief Function types to be used with resolving hashes into objects and
+/// back again. The resolver gets a pointer to a field inside an object API
+/// object that is of the type specified in the schema using the attribute
+/// `cpp_type` (it is thus important whatever you write to this address
+/// matches that type). The value of this field is initially null, so you
+/// may choose to implement a delayed binding lookup using this function
+/// if you wish. The resolver does the opposite lookup, for when the object
+/// is being serialized again.
+typedef uint64_t hash_value_t;
+typedef std::function<void(void **pointer_adr, hash_value_t hash)>
+    resolver_function_t;
+typedef std::function<hash_value_t(void *pointer)> rehasher_function_t;
+
+// Helper function to test if a field is present, using any of the field
+// enums in the generated code.
+// `table` must be a generated table type. Since this is a template parameter,
+// this is not typechecked to be a subclass of Table, so beware!
+// Note: this function will return false for fields equal to the default
+// value, since they're not stored in the buffer (unless force_defaults was
+// used).
+template<typename T>
+bool IsFieldPresent(const T *table, typename T::FlatBuffersVTableOffset field) {
+  // Cast, since Table is a private baseclass of any table types.
+  return reinterpret_cast<const Table *>(table)->CheckField(
+      static_cast<voffset_t>(field));
+}
+
+// Utility function for reverse lookups on the EnumNames*() functions
+// (in the generated C++ code)
+// names must be NULL terminated.
+inline int LookupEnum(const char **names, const char *name) {
+  for (const char **p = names; *p; p++)
+    if (!strcmp(*p, name)) return static_cast<int>(p - names);
+  return -1;
+}
+
+// These macros allow us to layout a struct with a guarantee that they'll end
+// up looking the same on different compilers and platforms.
+// It does this by disallowing the compiler to do any padding, and then
+// does padding itself by inserting extra padding fields that make every
+// element aligned to its own size.
+// Additionally, it manually sets the alignment of the struct as a whole,
+// which is typically its largest element, or a custom size set in the schema
+// by the force_align attribute.
+// These are used in the generated code only.
+
+// clang-format off
+#if defined(_MSC_VER)
+  #define FLATBUFFERS_MANUALLY_ALIGNED_STRUCT(alignment) \
+    __pragma(pack(1)) \
+    struct __declspec(align(alignment))
+  #define FLATBUFFERS_STRUCT_END(name, size) \
+    __pragma(pack()) \
+    static_assert(sizeof(name) == size, "compiler breaks packing rules")
+#elif defined(__GNUC__) || defined(__clang__) || defined(__ICCARM__)
+  #define FLATBUFFERS_MANUALLY_ALIGNED_STRUCT(alignment) \
+    _Pragma("pack(1)") \
+    struct __attribute__((aligned(alignment)))
+  #define FLATBUFFERS_STRUCT_END(name, size) \
+    _Pragma("pack()") \
+    static_assert(sizeof(name) == size, "compiler breaks packing rules")
+#else
+  #error Unknown compiler, please define structure alignment macros
+#endif
+// clang-format on
+
+// Minimal reflection via code generation.
+// Besides full-fat reflection (see reflection.h) and parsing/printing by
+// loading schemas (see idl.h), we can also have code generation for minimal
+// reflection data which allows pretty-printing and other uses without needing
+// a schema or a parser.
+// Generate code with --reflect-types (types only) or --reflect-names (names
+// also) to enable.
+// See minireflect.h for utilities using this functionality.
+
+// These types are organized slightly differently as the ones in idl.h.
+enum SequenceType { ST_TABLE, ST_STRUCT, ST_UNION, ST_ENUM };
+
+// Scalars have the same order as in idl.h
+// clang-format off
+#define FLATBUFFERS_GEN_ELEMENTARY_TYPES(ET) \
+  ET(ET_UTYPE) \
+  ET(ET_BOOL) \
+  ET(ET_CHAR) \
+  ET(ET_UCHAR) \
+  ET(ET_SHORT) \
+  ET(ET_USHORT) \
+  ET(ET_INT) \
+  ET(ET_UINT) \
+  ET(ET_LONG) \
+  ET(ET_ULONG) \
+  ET(ET_FLOAT) \
+  ET(ET_DOUBLE) \
+  ET(ET_STRING) \
+  ET(ET_SEQUENCE)  // See SequenceType.
+
+enum ElementaryType {
+  #define FLATBUFFERS_ET(E) E,
+    FLATBUFFERS_GEN_ELEMENTARY_TYPES(FLATBUFFERS_ET)
+  #undef FLATBUFFERS_ET
+};
+
+inline const char * const *ElementaryTypeNames() {
+  static const char * const names[] = {
+    #define FLATBUFFERS_ET(E) #E,
+      FLATBUFFERS_GEN_ELEMENTARY_TYPES(FLATBUFFERS_ET)
+    #undef FLATBUFFERS_ET
+  };
+  return names;
+}
+// clang-format on
+
+// Basic type info cost just 16bits per field!
+// We're explicitly defining the signedness since the signedness of integer
+// bitfields is otherwise implementation-defined and causes warnings on older
+// GCC compilers.
+struct TypeCode {
+  // ElementaryType
+  unsigned short base_type : 4;
+  // Either vector (in table) or array (in struct)
+  unsigned short is_repeating : 1;
+  // Index into type_refs below, or -1 for none.
+  signed short sequence_ref : 11;
+};
+
+static_assert(sizeof(TypeCode) == 2, "TypeCode");
+
+struct TypeTable;
+
+// Signature of the static method present in each type.
+typedef const TypeTable *(*TypeFunction)();
+
+struct TypeTable {
+  SequenceType st;
+  size_t num_elems;  // of type_codes, values, names (but not type_refs).
+  const TypeCode *type_codes;     // num_elems count
+  const TypeFunction *type_refs;  // less than num_elems entries (see TypeCode).
+  const int16_t *array_sizes;     // less than num_elems entries (see TypeCode).
+  const int64_t *values;  // Only set for non-consecutive enum/union or structs.
+  const char *const *names;  // Only set if compiled with --reflect-names.
+};
+
+// String which identifies the current version of FlatBuffers.
+inline const char *flatbuffers_version_string() {
+  return "FlatBuffers " FLATBUFFERS_STRING(FLATBUFFERS_VERSION_MAJOR) "."
+      FLATBUFFERS_STRING(FLATBUFFERS_VERSION_MINOR) "."
+      FLATBUFFERS_STRING(FLATBUFFERS_VERSION_REVISION);
+}
+
+// clang-format off
+#define FLATBUFFERS_DEFINE_BITMASK_OPERATORS(E, T)\
+    inline E operator | (E lhs, E rhs){\
+        return E(T(lhs) | T(rhs));\
+    }\
+    inline E operator & (E lhs, E rhs){\
+        return E(T(lhs) & T(rhs));\
+    }\
+    inline E operator ^ (E lhs, E rhs){\
+        return E(T(lhs) ^ T(rhs));\
+    }\
+    inline E operator ~ (E lhs){\
+        return E(~T(lhs));\
+    }\
+    inline E operator |= (E &lhs, E rhs){\
+        lhs = lhs | rhs;\
+        return lhs;\
+    }\
+    inline E operator &= (E &lhs, E rhs){\
+        lhs = lhs & rhs;\
+        return lhs;\
+    }\
+    inline E operator ^= (E &lhs, E rhs){\
+        lhs = lhs ^ rhs;\
+        return lhs;\
+    }\
+    inline bool operator !(E rhs) \
+    {\
+        return !bool(T(rhs)); \
+    }
+/// @endcond
+}  // namespace flatbuffers
+
+// clang-format on
+
+#endif  // FLATBUFFERS_H_
@@ -0,0 +1,513 @@
+/*
+ * Copyright 2017 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_STL_EMULATION_H_
+#define FLATBUFFERS_STL_EMULATION_H_
+
+// clang-format off
+#include "flatbuffers/base.h"
+
+#include <string>
+#include <type_traits>
+#include <vector>
+#include <memory>
+#include <limits>
+
+#ifndef FLATBUFFERS_USE_STD_OPTIONAL
+  // Detect C++17 compatible compiler.
+  // __cplusplus >= 201703L - a compiler has support of 'static inline' variables.
+  #if (defined(__cplusplus) && __cplusplus >= 201703L) \
+      || (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L)
+    #define FLATBUFFERS_USE_STD_OPTIONAL 1
+  #else
+    #define FLATBUFFERS_USE_STD_OPTIONAL 0
+  #endif // (defined(__cplusplus) && __cplusplus >= 201703L) ...
+#endif // FLATBUFFERS_USE_STD_OPTIONAL
+
+#if FLATBUFFERS_USE_STD_OPTIONAL
+  #include <optional>
+#endif
+
+#ifndef FLATBUFFERS_USE_STD_SPAN
+  // Testing __cpp_lib_span requires including either <version> or <span>,
+  // both of which were added in C++20.
+  // See: https://en.cppreference.com/w/cpp/utility/feature_test
+  #if defined(__cplusplus) && __cplusplus >= 202002L
+    #define FLATBUFFERS_USE_STD_SPAN 1
+  #endif
+#endif // FLATBUFFERS_USE_STD_SPAN
+
+#if defined(FLATBUFFERS_USE_STD_SPAN)
+  #include <array>
+  #include <span>
+#else
+  // Disable non-trivial ctors if FLATBUFFERS_SPAN_MINIMAL defined.
+  #if !defined(FLATBUFFERS_TEMPLATES_ALIASES)
+    #define FLATBUFFERS_SPAN_MINIMAL
+  #else
+    // Enable implicit construction of a span<T,N> from a std::array<T,N>.
+    #include <array>
+  #endif
+#endif // defined(FLATBUFFERS_USE_STD_SPAN)
+
+// This header provides backwards compatibility for older versions of the STL.
+namespace flatbuffers {
+
+#if defined(FLATBUFFERS_TEMPLATES_ALIASES)
+  template <typename T>
+  using numeric_limits = std::numeric_limits<T>;
+#else
+  template <typename T> class numeric_limits :
+    public std::numeric_limits<T> {};
+#endif  // defined(FLATBUFFERS_TEMPLATES_ALIASES)
+
+#if defined(FLATBUFFERS_TEMPLATES_ALIASES)
+  template <typename T> using is_scalar = std::is_scalar<T>;
+  template <typename T, typename U> using is_same = std::is_same<T,U>;
+  template <typename T> using is_floating_point = std::is_floating_point<T>;
+  template <typename T> using is_unsigned = std::is_unsigned<T>;
+  template <typename T> using is_enum = std::is_enum<T>;
+  template <typename T> using make_unsigned = std::make_unsigned<T>;
+  template<bool B, class T, class F>
+  using conditional = std::conditional<B, T, F>;
+  template<class T, T v>
+  using integral_constant = std::integral_constant<T, v>;
+  template <bool B>
+  using bool_constant = integral_constant<bool, B>;
+  using true_type  = std::true_type;
+  using false_type = std::false_type;
+#else
+  // MSVC 2010 doesn't support C++11 aliases.
+  template <typename T> struct is_scalar : public std::is_scalar<T> {};
+  template <typename T, typename U> struct is_same : public std::is_same<T,U> {};
+  template <typename T> struct is_floating_point :
+        public std::is_floating_point<T> {};
+  template <typename T> struct is_unsigned : public std::is_unsigned<T> {};
+  template <typename T> struct is_enum : public std::is_enum<T> {};
+  template <typename T> struct make_unsigned : public std::make_unsigned<T> {};
+  template<bool B, class T, class F>
+  struct conditional : public std::conditional<B, T, F> {};
+  template<class T, T v>
+  struct integral_constant : public std::integral_constant<T, v> {};
+  template <bool B>
+  struct bool_constant : public integral_constant<bool, B> {};
+  typedef bool_constant<true>  true_type;
+  typedef bool_constant<false> false_type;
+#endif  // defined(FLATBUFFERS_TEMPLATES_ALIASES)
+
+#if defined(FLATBUFFERS_TEMPLATES_ALIASES)
+  template <class T> using unique_ptr = std::unique_ptr<T>;
+#else
+  // MSVC 2010 doesn't support C++11 aliases.
+  // We're manually "aliasing" the class here as we want to bring unique_ptr
+  // into the flatbuffers namespace.  We have unique_ptr in the flatbuffers
+  // namespace we have a completely independent implementation (see below)
+  // for C++98 STL implementations.
+  template <class T> class unique_ptr : public std::unique_ptr<T> {
+    public:
+    unique_ptr() {}
+    explicit unique_ptr(T* p) : std::unique_ptr<T>(p) {}
+    unique_ptr(std::unique_ptr<T>&& u) { *this = std::move(u); }
+    unique_ptr(unique_ptr&& u) { *this = std::move(u); }
+    unique_ptr& operator=(std::unique_ptr<T>&& u) {
+      std::unique_ptr<T>::reset(u.release());
+      return *this;
+    }
+    unique_ptr& operator=(unique_ptr&& u) {
+      std::unique_ptr<T>::reset(u.release());
+      return *this;
+    }
+    unique_ptr& operator=(T* p) {
+      return std::unique_ptr<T>::operator=(p);
+    }
+  };
+#endif  // defined(FLATBUFFERS_TEMPLATES_ALIASES)
+
+#if FLATBUFFERS_USE_STD_OPTIONAL
+template<class T>
+using Optional = std::optional<T>;
+using nullopt_t = std::nullopt_t;
+inline constexpr nullopt_t nullopt = std::nullopt;
+
+#else
+// Limited implementation of Optional<T> type for a scalar T.
+// This implementation limited by trivial types compatible with
+// std::is_arithmetic<T> or std::is_enum<T> type traits.
+
+// A tag to indicate an empty flatbuffers::optional<T>.
+struct nullopt_t {
+  explicit FLATBUFFERS_CONSTEXPR_CPP11 nullopt_t(int) {}
+};
+
+#if defined(FLATBUFFERS_CONSTEXPR_DEFINED)
+  namespace internal {
+    template <class> struct nullopt_holder {
+      static constexpr nullopt_t instance_ = nullopt_t(0);
+    };
+    template<class Dummy>
+    constexpr nullopt_t nullopt_holder<Dummy>::instance_;
+  }
+  static constexpr const nullopt_t &nullopt = internal::nullopt_holder<void>::instance_;
+
+#else
+  namespace internal {
+    template <class> struct nullopt_holder {
+      static const nullopt_t instance_;
+    };
+    template<class Dummy>
+    const nullopt_t nullopt_holder<Dummy>::instance_  = nullopt_t(0);
+  }
+  static const nullopt_t &nullopt = internal::nullopt_holder<void>::instance_;
+
+#endif
+
+template<class T>
+class Optional FLATBUFFERS_FINAL_CLASS {
+  // Non-scalar 'T' would extremely complicated Optional<T>.
+  // Use is_scalar<T> checking because flatbuffers flatbuffers::is_arithmetic<T>
+  // isn't implemented.
+  static_assert(flatbuffers::is_scalar<T>::value, "unexpected type T");
+
+ public:
+  ~Optional() {}
+
+  FLATBUFFERS_CONSTEXPR_CPP11 Optional() FLATBUFFERS_NOEXCEPT
+    : value_(), has_value_(false) {}
+
+  FLATBUFFERS_CONSTEXPR_CPP11 Optional(nullopt_t) FLATBUFFERS_NOEXCEPT
+    : value_(), has_value_(false) {}
+
+  FLATBUFFERS_CONSTEXPR_CPP11 Optional(T val) FLATBUFFERS_NOEXCEPT
+    : value_(val), has_value_(true) {}
+
+  FLATBUFFERS_CONSTEXPR_CPP11 Optional(const Optional &other) FLATBUFFERS_NOEXCEPT
+    : value_(other.value_), has_value_(other.has_value_) {}
+
+  FLATBUFFERS_CONSTEXPR_CPP14 Optional &operator=(const Optional &other) FLATBUFFERS_NOEXCEPT {
+    value_ = other.value_;
+    has_value_ = other.has_value_;
+    return *this;
+  }
+
+  FLATBUFFERS_CONSTEXPR_CPP14 Optional &operator=(nullopt_t) FLATBUFFERS_NOEXCEPT {
+    value_ = T();
+    has_value_ = false;
+    return *this;
+  }
+
+  FLATBUFFERS_CONSTEXPR_CPP14 Optional &operator=(T val) FLATBUFFERS_NOEXCEPT {
+    value_ = val;
+    has_value_ = true;
+    return *this;
+  }
+
+  void reset() FLATBUFFERS_NOEXCEPT {
+    *this = nullopt;
+  }
+
+  void swap(Optional &other) FLATBUFFERS_NOEXCEPT {
+    std::swap(value_, other.value_);
+    std::swap(has_value_, other.has_value_);
+  }
+
+  FLATBUFFERS_CONSTEXPR_CPP11 FLATBUFFERS_EXPLICIT_CPP11 operator bool() const FLATBUFFERS_NOEXCEPT {
+    return has_value_;
+  }
+
+  FLATBUFFERS_CONSTEXPR_CPP11 bool has_value() const FLATBUFFERS_NOEXCEPT {
+    return has_value_;
+  }
+
+  FLATBUFFERS_CONSTEXPR_CPP11 const T& operator*() const FLATBUFFERS_NOEXCEPT {
+    return value_;
+  }
+
+  const T& value() const {
+    FLATBUFFERS_ASSERT(has_value());
+    return value_;
+  }
+
+  T value_or(T default_value) const FLATBUFFERS_NOEXCEPT {
+    return has_value() ? value_ : default_value;
+  }
+
+ private:
+  T value_;
+  bool has_value_;
+};
+
+template<class T>
+FLATBUFFERS_CONSTEXPR_CPP11 bool operator==(const Optional<T>& opt, nullopt_t) FLATBUFFERS_NOEXCEPT {
+  return !opt;
+}
+template<class T>
+FLATBUFFERS_CONSTEXPR_CPP11 bool operator==(nullopt_t, const Optional<T>& opt) FLATBUFFERS_NOEXCEPT {
+  return !opt;
+}
+
+template<class T, class U>
+FLATBUFFERS_CONSTEXPR_CPP11 bool operator==(const Optional<T>& lhs, const U& rhs) FLATBUFFERS_NOEXCEPT {
+  return static_cast<bool>(lhs) && (*lhs == rhs);
+}
+
+template<class T, class U>
+FLATBUFFERS_CONSTEXPR_CPP11 bool operator==(const T& lhs, const Optional<U>& rhs) FLATBUFFERS_NOEXCEPT {
+  return static_cast<bool>(rhs) && (lhs == *rhs);
+}
+
+template<class T, class U>
+FLATBUFFERS_CONSTEXPR_CPP11 bool operator==(const Optional<T>& lhs, const Optional<U>& rhs) FLATBUFFERS_NOEXCEPT {
+  return static_cast<bool>(lhs) != static_cast<bool>(rhs)
+              ? false
+              : !static_cast<bool>(lhs) ? false : (*lhs == *rhs);
+}
+#endif // FLATBUFFERS_USE_STD_OPTIONAL
+
+
+// Very limited and naive partial implementation of C++20 std::span<T,Extent>.
+#if defined(FLATBUFFERS_USE_STD_SPAN)
+  inline constexpr std::size_t dynamic_extent = std::dynamic_extent;
+  template<class T, std::size_t Extent = std::dynamic_extent>
+  using span = std::span<T, Extent>;
+
+#else // !defined(FLATBUFFERS_USE_STD_SPAN)
+FLATBUFFERS_CONSTEXPR std::size_t dynamic_extent = static_cast<std::size_t>(-1);
+
+// Exclude this code if MSVC2010 or non-STL Android is active.
+// The non-STL Android doesn't have `std::is_convertible` required for SFINAE.
+#if !defined(FLATBUFFERS_SPAN_MINIMAL)
+namespace internal {
+  // This is SFINAE helper class for checking of a common condition:
+  // > This overload only participates in overload resolution
+  // > Check whether a pointer to an array of From can be converted
+  // > to a pointer to an array of To.
+  // This helper is used for checking of 'From -> const From'.
+  template<class To, std::size_t Extent, class From, std::size_t N>
+  struct is_span_convertible {
+    using type =
+      typename std::conditional<std::is_convertible<From (*)[], To (*)[]>::value
+                                && (Extent == dynamic_extent || N == Extent),
+                                int, void>::type;
+  };
+
+  template<typename T>
+  struct SpanIterator {
+    // TODO: upgrade to std::random_access_iterator_tag.
+    using iterator_category = std::forward_iterator_tag;
+    using difference_type  = std::ptrdiff_t;
+    using value_type = typename std::remove_cv<T>::type;
+    using reference = T&;
+    using pointer   = T*;
+
+    // Convince MSVC compiler that this iterator is trusted (it is verified).
+    #ifdef _MSC_VER
+      using _Unchecked_type = pointer;
+    #endif // _MSC_VER
+
+    SpanIterator(pointer ptr) : ptr_(ptr) {}
+    reference operator*() const { return *ptr_; }
+    pointer operator->() { return ptr_; }
+    SpanIterator& operator++() { ptr_++; return *this; }  
+    SpanIterator  operator++(int) { auto tmp = *this; ++(*this); return tmp; }
+
+    friend bool operator== (const SpanIterator& lhs, const SpanIterator& rhs) { return lhs.ptr_ == rhs.ptr_; }
+    friend bool operator!= (const SpanIterator& lhs, const SpanIterator& rhs) { return lhs.ptr_ != rhs.ptr_; }
+
+   private:
+    pointer ptr_;
+  };
+}  // namespace internal
+#endif  // !defined(FLATBUFFERS_SPAN_MINIMAL)
+
+// T - element type; must be a complete type that is not an abstract
+// class type.
+// Extent - the number of elements in the sequence, or dynamic.
+template<class T, std::size_t Extent = dynamic_extent>
+class span FLATBUFFERS_FINAL_CLASS {
+ public:
+  typedef T element_type;
+  typedef T& reference;
+  typedef const T& const_reference;
+  typedef T* pointer;
+  typedef const T* const_pointer;
+  typedef std::size_t size_type;
+
+  static FLATBUFFERS_CONSTEXPR size_type extent = Extent;
+
+  // Returns the number of elements in the span.
+  FLATBUFFERS_CONSTEXPR_CPP11 size_type size() const FLATBUFFERS_NOEXCEPT {
+    return count_;
+  }
+
+  // Returns the size of the sequence in bytes.
+  FLATBUFFERS_CONSTEXPR_CPP11
+  size_type size_bytes() const FLATBUFFERS_NOEXCEPT {
+    return size() * sizeof(element_type);
+  }
+
+  // Checks if the span is empty.
+  FLATBUFFERS_CONSTEXPR_CPP11 bool empty() const FLATBUFFERS_NOEXCEPT {
+    return size() == 0;
+  }
+
+  // Returns a pointer to the beginning of the sequence.
+  FLATBUFFERS_CONSTEXPR_CPP11 pointer data() const FLATBUFFERS_NOEXCEPT {
+    return data_;
+  }
+
+  #if !defined(FLATBUFFERS_SPAN_MINIMAL)
+    using Iterator = internal::SpanIterator<T>;
+
+    Iterator begin() const { return Iterator(data()); }
+    Iterator end() const   { return Iterator(data() + size()); }
+  #endif
+
+  // Returns a reference to the idx-th element of the sequence.
+  // The behavior is undefined if the idx is greater than or equal to size().
+  FLATBUFFERS_CONSTEXPR_CPP11 reference operator[](size_type idx) const {
+    return data()[idx];
+  }
+
+  FLATBUFFERS_CONSTEXPR_CPP11 span(const span &other) FLATBUFFERS_NOEXCEPT
+      : data_(other.data_), count_(other.count_) {}
+
+  FLATBUFFERS_CONSTEXPR_CPP14 span &operator=(const span &other)
+      FLATBUFFERS_NOEXCEPT {
+    data_ = other.data_;
+    count_ = other.count_;
+  }
+
+  // Limited implementation of
+  // `template <class It> constexpr std::span(It first, size_type count);`.
+  //
+  // Constructs a span that is a view over the range [first, first + count);
+  // the resulting span has: data() == first and size() == count.
+  // The behavior is undefined if [first, first + count) is not a valid range,
+  // or if (extent != flatbuffers::dynamic_extent && count != extent).
+  FLATBUFFERS_CONSTEXPR_CPP11
+  explicit span(pointer first, size_type count) FLATBUFFERS_NOEXCEPT
+    : data_ (Extent == dynamic_extent ? first : (Extent == count ? first : nullptr)),
+      count_(Extent == dynamic_extent ? count : (Extent == count ? Extent : 0)) {
+      // Make span empty if the count argument is incompatible with span<T,N>.
+  }
+
+  // Exclude this code if MSVC2010 is active. The MSVC2010 isn't C++11
+  // compliant, it doesn't support default template arguments for functions.
+  #if defined(FLATBUFFERS_SPAN_MINIMAL)
+  FLATBUFFERS_CONSTEXPR_CPP11 span() FLATBUFFERS_NOEXCEPT : data_(nullptr),
+                                                            count_(0) {
+    static_assert(extent == 0 || extent == dynamic_extent, "invalid span");
+  }
+
+  #else
+  // Constructs an empty span whose data() == nullptr and size() == 0.
+  // This overload only participates in overload resolution if
+  // extent == 0 || extent == flatbuffers::dynamic_extent.
+  // A dummy template argument N is need dependency for SFINAE.
+  template<std::size_t N = 0,
+    typename internal::is_span_convertible<element_type, Extent, element_type, (N - N)>::type = 0>
+  FLATBUFFERS_CONSTEXPR_CPP11 span() FLATBUFFERS_NOEXCEPT : data_(nullptr),
+                                                            count_(0) {
+    static_assert(extent == 0 || extent == dynamic_extent, "invalid span");
+  }
+
+  // Constructs a span that is a view over the array arr; the resulting span
+  // has size() == N and data() == std::data(arr). These overloads only
+  // participate in overload resolution if
+  // extent == std::dynamic_extent || N == extent is true and
+  // std::remove_pointer_t<decltype(std::data(arr))>(*)[]
+  // is convertible to element_type (*)[].
+  template<std::size_t N,
+    typename internal::is_span_convertible<element_type, Extent, element_type, N>::type = 0>
+  FLATBUFFERS_CONSTEXPR_CPP11 span(element_type (&arr)[N]) FLATBUFFERS_NOEXCEPT
+      : data_(arr), count_(N) {}
+
+  template<class U, std::size_t N,
+    typename internal::is_span_convertible<element_type, Extent, U, N>::type = 0>
+  FLATBUFFERS_CONSTEXPR_CPP11 span(std::array<U, N> &arr) FLATBUFFERS_NOEXCEPT
+     : data_(arr.data()), count_(N) {}
+
+  //template<class U, std::size_t N,
+  //  int = 0>
+  //FLATBUFFERS_CONSTEXPR_CPP11 span(std::array<U, N> &arr) FLATBUFFERS_NOEXCEPT
+  //   : data_(arr.data()), count_(N) {}
+
+  template<class U, std::size_t N,
+    typename internal::is_span_convertible<element_type, Extent, U, N>::type = 0>
+  FLATBUFFERS_CONSTEXPR_CPP11 span(const std::array<U, N> &arr) FLATBUFFERS_NOEXCEPT
+    : data_(arr.data()), count_(N) {}
+
+  // Converting constructor from another span s;
+  // the resulting span has size() == s.size() and data() == s.data().
+  // This overload only participates in overload resolution
+  // if extent == std::dynamic_extent || N == extent is true and U (*)[]
+  // is convertible to element_type (*)[].
+  template<class U, std::size_t N,
+    typename internal::is_span_convertible<element_type, Extent, U, N>::type = 0>
+  FLATBUFFERS_CONSTEXPR_CPP11 span(const flatbuffers::span<U, N> &s) FLATBUFFERS_NOEXCEPT
+      : span(s.data(), s.size()) {
+  }
+
+  #endif  // !defined(FLATBUFFERS_SPAN_MINIMAL)
+
+ private:
+  // This is a naive implementation with 'count_' member even if (Extent != dynamic_extent).
+  pointer const data_;
+  size_type count_;
+};
+#endif  // defined(FLATBUFFERS_USE_STD_SPAN)
+
+#if !defined(FLATBUFFERS_SPAN_MINIMAL)
+template<class ElementType, std::size_t Extent>
+FLATBUFFERS_CONSTEXPR_CPP11
+flatbuffers::span<ElementType, Extent> make_span(ElementType(&arr)[Extent]) FLATBUFFERS_NOEXCEPT {
+  return span<ElementType, Extent>(arr);
+}
+
+template<class ElementType, std::size_t Extent>
+FLATBUFFERS_CONSTEXPR_CPP11
+flatbuffers::span<const ElementType, Extent> make_span(const ElementType(&arr)[Extent]) FLATBUFFERS_NOEXCEPT {
+  return span<const ElementType, Extent>(arr);
+}
+
+template<class ElementType, std::size_t Extent>
+FLATBUFFERS_CONSTEXPR_CPP11
+flatbuffers::span<ElementType, Extent> make_span(std::array<ElementType, Extent> &arr) FLATBUFFERS_NOEXCEPT {
+  return span<ElementType, Extent>(arr);
+}
+
+template<class ElementType, std::size_t Extent>
+FLATBUFFERS_CONSTEXPR_CPP11
+flatbuffers::span<const ElementType, Extent> make_span(const std::array<ElementType, Extent> &arr) FLATBUFFERS_NOEXCEPT {
+  return span<const ElementType, Extent>(arr);
+}
+
+template<class ElementType, std::size_t Extent>
+FLATBUFFERS_CONSTEXPR_CPP11
+flatbuffers::span<ElementType, dynamic_extent> make_span(ElementType *first, std::size_t count) FLATBUFFERS_NOEXCEPT {
+  return span<ElementType, dynamic_extent>(first, count);
+}
+
+template<class ElementType, std::size_t Extent>
+FLATBUFFERS_CONSTEXPR_CPP11
+flatbuffers::span<const ElementType, dynamic_extent> make_span(const ElementType *first, std::size_t count) FLATBUFFERS_NOEXCEPT {
+  return span<const ElementType, dynamic_extent>(first, count);
+}
+#endif // !defined(FLATBUFFERS_SPAN_MINIMAL)
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_STL_EMULATION_H_
@@ -0,0 +1,64 @@
+/*
+ * Copyright 2021 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_STRING_H_
+#define FLATBUFFERS_STRING_H_
+
+#include "flatbuffers/base.h"
+#include "flatbuffers/vector.h"
+
+namespace flatbuffers {
+
+struct String : public Vector<char> {
+  const char *c_str() const { return reinterpret_cast<const char *>(Data()); }
+  std::string str() const { return std::string(c_str(), size()); }
+
+  // clang-format off
+  #ifdef FLATBUFFERS_HAS_STRING_VIEW
+  flatbuffers::string_view string_view() const {
+    return flatbuffers::string_view(c_str(), size());
+  }
+  #endif // FLATBUFFERS_HAS_STRING_VIEW
+  // clang-format on
+
+  bool operator<(const String &o) const {
+    return StringLessThan(this->data(), this->size(), o.data(), o.size());
+  }
+};
+
+// Convenience function to get std::string from a String returning an empty
+// string on null pointer.
+static inline std::string GetString(const String *str) {
+  return str ? str->str() : "";
+}
+
+// Convenience function to get char* from a String returning an empty string on
+// null pointer.
+static inline const char *GetCstring(const String *str) {
+  return str ? str->c_str() : "";
+}
+
+#ifdef FLATBUFFERS_HAS_STRING_VIEW
+// Convenience function to get string_view from a String returning an empty
+// string_view on null pointer.
+static inline flatbuffers::string_view GetStringView(const String *str) {
+  return str ? str->string_view() : flatbuffers::string_view();
+}
+#endif  // FLATBUFFERS_HAS_STRING_VIEW
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_STRING_H_
@@ -0,0 +1,53 @@
+/*
+ * Copyright 2021 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_STRUCT_H_
+#define FLATBUFFERS_STRUCT_H_
+
+#include "flatbuffers/base.h"
+
+namespace flatbuffers {
+
+// "structs" are flat structures that do not have an offset table, thus
+// always have all members present and do not support forwards/backwards
+// compatible extensions.
+
+class Struct FLATBUFFERS_FINAL_CLASS {
+ public:
+  template<typename T> T GetField(uoffset_t o) const {
+    return ReadScalar<T>(&data_[o]);
+  }
+
+  template<typename T> T GetStruct(uoffset_t o) const {
+    return reinterpret_cast<T>(&data_[o]);
+  }
+
+  const uint8_t *GetAddressOf(uoffset_t o) const { return &data_[o]; }
+  uint8_t *GetAddressOf(uoffset_t o) { return &data_[o]; }
+
+ private:
+  // private constructor & copy constructor: you obtain instances of this
+  // class by pointing to existing data only
+  Struct();
+  Struct(const Struct &);
+  Struct &operator=(const Struct &);
+
+  uint8_t data_[1];
+};
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_STRUCT_H_
@@ -0,0 +1,188 @@
+/*
+ * Copyright 2021 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_TABLE_H_
+#define FLATBUFFERS_TABLE_H_
+
+#include "flatbuffers/base.h"
+#include "flatbuffers/verifier.h"
+
+namespace flatbuffers {
+
+// "tables" use an offset table (possibly shared) that allows fields to be
+// omitted and added at will, but uses an extra indirection to read.
+class Table {
+ public:
+  const uint8_t *GetVTable() const {
+    return data_ - ReadScalar<soffset_t>(data_);
+  }
+
+  // This gets the field offset for any of the functions below it, or 0
+  // if the field was not present.
+  voffset_t GetOptionalFieldOffset(voffset_t field) const {
+    // The vtable offset is always at the start.
+    auto vtable = GetVTable();
+    // The first element is the size of the vtable (fields + type id + itself).
+    auto vtsize = ReadScalar<voffset_t>(vtable);
+    // If the field we're accessing is outside the vtable, we're reading older
+    // data, so it's the same as if the offset was 0 (not present).
+    return field < vtsize ? ReadScalar<voffset_t>(vtable + field) : 0;
+  }
+
+  template<typename T> T GetField(voffset_t field, T defaultval) const {
+    auto field_offset = GetOptionalFieldOffset(field);
+    return field_offset ? ReadScalar<T>(data_ + field_offset) : defaultval;
+  }
+
+  template<typename P, typename OffsetSize = uoffset_t>
+  P GetPointer(voffset_t field) {
+    auto field_offset = GetOptionalFieldOffset(field);
+    auto p = data_ + field_offset;
+    return field_offset ? reinterpret_cast<P>(p + ReadScalar<OffsetSize>(p))
+                        : nullptr;
+  }
+  template<typename P, typename OffsetSize = uoffset_t>
+  P GetPointer(voffset_t field) const {
+    return const_cast<Table *>(this)->GetPointer<P, OffsetSize>(field);
+  }
+
+  template<typename P> P GetPointer64(voffset_t field) {
+    return GetPointer<P, uoffset64_t>(field);
+  }
+
+  template<typename P> P GetPointer64(voffset_t field) const {
+    return GetPointer<P, uoffset64_t>(field);
+  }
+
+  template<typename P> P GetStruct(voffset_t field) const {
+    auto field_offset = GetOptionalFieldOffset(field);
+    auto p = const_cast<uint8_t *>(data_ + field_offset);
+    return field_offset ? reinterpret_cast<P>(p) : nullptr;
+  }
+
+  template<typename Raw, typename Face>
+  flatbuffers::Optional<Face> GetOptional(voffset_t field) const {
+    auto field_offset = GetOptionalFieldOffset(field);
+    auto p = data_ + field_offset;
+    return field_offset ? Optional<Face>(static_cast<Face>(ReadScalar<Raw>(p)))
+                        : Optional<Face>();
+  }
+
+  template<typename T> bool SetField(voffset_t field, T val, T def) {
+    auto field_offset = GetOptionalFieldOffset(field);
+    if (!field_offset) return IsTheSameAs(val, def);
+    WriteScalar(data_ + field_offset, val);
+    return true;
+  }
+  template<typename T> bool SetField(voffset_t field, T val) {
+    auto field_offset = GetOptionalFieldOffset(field);
+    if (!field_offset) return false;
+    WriteScalar(data_ + field_offset, val);
+    return true;
+  }
+
+  bool SetPointer(voffset_t field, const uint8_t *val) {
+    auto field_offset = GetOptionalFieldOffset(field);
+    if (!field_offset) return false;
+    WriteScalar(data_ + field_offset,
+                static_cast<uoffset_t>(val - (data_ + field_offset)));
+    return true;
+  }
+
+  uint8_t *GetAddressOf(voffset_t field) {
+    auto field_offset = GetOptionalFieldOffset(field);
+    return field_offset ? data_ + field_offset : nullptr;
+  }
+  const uint8_t *GetAddressOf(voffset_t field) const {
+    return const_cast<Table *>(this)->GetAddressOf(field);
+  }
+
+  bool CheckField(voffset_t field) const {
+    return GetOptionalFieldOffset(field) != 0;
+  }
+
+  // Verify the vtable of this table.
+  // Call this once per table, followed by VerifyField once per field.
+  bool VerifyTableStart(Verifier &verifier) const {
+    return verifier.VerifyTableStart(data_);
+  }
+
+  // Verify a particular field.
+  template<typename T>
+  bool VerifyField(const Verifier &verifier, voffset_t field,
+                   size_t align) const {
+    // Calling GetOptionalFieldOffset should be safe now thanks to
+    // VerifyTable().
+    auto field_offset = GetOptionalFieldOffset(field);
+    // Check the actual field.
+    return !field_offset || verifier.VerifyField<T>(data_, field_offset, align);
+  }
+
+  // VerifyField for required fields.
+  template<typename T>
+  bool VerifyFieldRequired(const Verifier &verifier, voffset_t field,
+                           size_t align) const {
+    auto field_offset = GetOptionalFieldOffset(field);
+    return verifier.Check(field_offset != 0) &&
+           verifier.VerifyField<T>(data_, field_offset, align);
+  }
+
+  // Versions for offsets.
+  template<typename OffsetT = uoffset_t>
+  bool VerifyOffset(const Verifier &verifier, voffset_t field) const {
+    auto field_offset = GetOptionalFieldOffset(field);
+    return !field_offset || verifier.VerifyOffset<OffsetT>(data_, field_offset);
+  }
+
+  template<typename OffsetT = uoffset_t>
+  bool VerifyOffsetRequired(const Verifier &verifier, voffset_t field) const {
+    auto field_offset = GetOptionalFieldOffset(field);
+    return verifier.Check(field_offset != 0) &&
+           verifier.VerifyOffset<OffsetT>(data_, field_offset);
+  }
+ 
+  bool VerifyOffset64(const Verifier &verifier, voffset_t field) const {
+    return VerifyOffset<uoffset64_t>(verifier, field);
+  }
+
+  bool VerifyOffset64Required(const Verifier &verifier, voffset_t field) const {
+    return VerifyOffsetRequired<uoffset64_t>(verifier, field);
+  }
+
+ private:
+  // private constructor & copy constructor: you obtain instances of this
+  // class by pointing to existing data only
+  Table();
+  Table(const Table &other);
+  Table &operator=(const Table &);
+
+  uint8_t data_[1];
+};
+
+// This specialization allows avoiding warnings like:
+// MSVC C4800: type: forcing value to bool 'true' or 'false'.
+template<>
+inline flatbuffers::Optional<bool> Table::GetOptional<uint8_t, bool>(
+    voffset_t field) const {
+  auto field_offset = GetOptionalFieldOffset(field);
+  auto p = data_ + field_offset;
+  return field_offset ? Optional<bool>(ReadScalar<uint8_t>(p) != 0)
+                      : Optional<bool>();
+}
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_TABLE_H_
@@ -0,0 +1,400 @@
+/*
+ * Copyright 2021 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_VECTOR_H_
+#define FLATBUFFERS_VECTOR_H_
+
+#include "flatbuffers/base.h"
+#include "flatbuffers/buffer.h"
+#include "flatbuffers/stl_emulation.h"
+
+namespace flatbuffers {
+
+struct String;
+
+// An STL compatible iterator implementation for Vector below, effectively
+// calling Get() for every element.
+template<typename T, typename IT, typename Data = uint8_t *,
+         typename SizeT = uoffset_t>
+struct VectorIterator {
+  typedef std::random_access_iterator_tag iterator_category;
+  typedef IT value_type;
+  typedef ptrdiff_t difference_type;
+  typedef IT *pointer;
+  typedef IT &reference;
+
+  static const SizeT element_stride = IndirectHelper<T>::element_stride;
+
+  VectorIterator(Data data, SizeT i) : data_(data + element_stride * i) {}
+  VectorIterator(const VectorIterator &other) : data_(other.data_) {}
+  VectorIterator() : data_(nullptr) {}
+
+  VectorIterator &operator=(const VectorIterator &other) {
+    data_ = other.data_;
+    return *this;
+  }
+
+  VectorIterator &operator=(VectorIterator &&other) {
+    data_ = other.data_;
+    return *this;
+  }
+
+  bool operator==(const VectorIterator &other) const {
+    return data_ == other.data_;
+  }
+
+  bool operator<(const VectorIterator &other) const {
+    return data_ < other.data_;
+  }
+
+  bool operator!=(const VectorIterator &other) const {
+    return data_ != other.data_;
+  }
+
+  difference_type operator-(const VectorIterator &other) const {
+    return (data_ - other.data_) / element_stride;
+  }
+
+  // Note: return type is incompatible with the standard
+  // `reference operator*()`.
+  IT operator*() const { return IndirectHelper<T>::Read(data_, 0); }
+
+  // Note: return type is incompatible with the standard
+  // `pointer operator->()`.
+  IT operator->() const { return IndirectHelper<T>::Read(data_, 0); }
+
+  VectorIterator &operator++() {
+    data_ += element_stride;
+    return *this;
+  }
+
+  VectorIterator operator++(int) {
+    VectorIterator temp(data_, 0);
+    data_ += element_stride;
+    return temp;
+  }
+
+  VectorIterator operator+(const SizeT &offset) const {
+    return VectorIterator(data_ + offset * element_stride, 0);
+  }
+
+  VectorIterator &operator+=(const SizeT &offset) {
+    data_ += offset * element_stride;
+    return *this;
+  }
+
+  VectorIterator &operator--() {
+    data_ -= element_stride;
+    return *this;
+  }
+
+  VectorIterator operator--(int) {
+    VectorIterator temp(data_, 0);
+    data_ -= element_stride;
+    return temp;
+  }
+
+  VectorIterator operator-(const SizeT &offset) const {
+    return VectorIterator(data_ - offset * element_stride, 0);
+  }
+
+  VectorIterator &operator-=(const SizeT &offset) {
+    data_ -= offset * element_stride;
+    return *this;
+  }
+
+ private:
+  Data data_;
+};
+
+template<typename T, typename IT, typename SizeT = uoffset_t>
+using VectorConstIterator = VectorIterator<T, IT, const uint8_t *, SizeT>;
+
+template<typename Iterator>
+struct VectorReverseIterator : public std::reverse_iterator<Iterator> {
+  explicit VectorReverseIterator(Iterator iter)
+      : std::reverse_iterator<Iterator>(iter) {}
+
+  // Note: return type is incompatible with the standard
+  // `reference operator*()`.
+  typename Iterator::value_type operator*() const {
+    auto tmp = std::reverse_iterator<Iterator>::current;
+    return *--tmp;
+  }
+
+  // Note: return type is incompatible with the standard
+  // `pointer operator->()`.
+  typename Iterator::value_type operator->() const {
+    auto tmp = std::reverse_iterator<Iterator>::current;
+    return *--tmp;
+  }
+};
+
+// This is used as a helper type for accessing vectors.
+// Vector::data() assumes the vector elements start after the length field.
+template<typename T, typename SizeT = uoffset_t> class Vector {
+ public:
+  typedef VectorIterator<T,
+                         typename IndirectHelper<T>::mutable_return_type,
+                         uint8_t *, SizeT>
+      iterator;
+  typedef VectorConstIterator<T, typename IndirectHelper<T>::return_type,
+                              SizeT>
+      const_iterator;
+  typedef VectorReverseIterator<iterator> reverse_iterator;
+  typedef VectorReverseIterator<const_iterator> const_reverse_iterator;
+
+  typedef typename flatbuffers::bool_constant<flatbuffers::is_scalar<T>::value>
+      scalar_tag;
+
+  static FLATBUFFERS_CONSTEXPR bool is_span_observable =
+      scalar_tag::value && (FLATBUFFERS_LITTLEENDIAN || sizeof(T) == 1);
+
+  SizeT size() const { return EndianScalar(length_); }
+
+  // Deprecated: use size(). Here for backwards compatibility.
+  FLATBUFFERS_ATTRIBUTE([[deprecated("use size() instead")]])
+  SizeT Length() const { return size(); }
+
+  typedef SizeT size_type;
+  typedef typename IndirectHelper<T>::return_type return_type;
+  typedef typename IndirectHelper<T>::mutable_return_type
+      mutable_return_type;
+  typedef return_type value_type;
+
+  return_type Get(SizeT i) const {
+    FLATBUFFERS_ASSERT(i < size());
+    return IndirectHelper<T>::Read(Data(), i);
+  }
+
+  return_type operator[](SizeT i) const { return Get(i); }
+
+  // If this is a Vector of enums, T will be its storage type, not the enum
+  // type. This function makes it convenient to retrieve value with enum
+  // type E.
+  template<typename E> E GetEnum(SizeT i) const {
+    return static_cast<E>(Get(i));
+  }
+
+  // If this a vector of unions, this does the cast for you. There's no check
+  // to make sure this is the right type!
+  template<typename U> const U *GetAs(SizeT i) const {
+    return reinterpret_cast<const U *>(Get(i));
+  }
+
+  // If this a vector of unions, this does the cast for you. There's no check
+  // to make sure this is actually a string!
+  const String *GetAsString(SizeT i) const {
+    return reinterpret_cast<const String *>(Get(i));
+  }
+
+  const void *GetStructFromOffset(size_t o) const {
+    return reinterpret_cast<const void *>(Data() + o);
+  }
+
+  iterator begin() { return iterator(Data(), 0); }
+  const_iterator begin() const { return const_iterator(Data(), 0); }
+
+  iterator end() { return iterator(Data(), size()); }
+  const_iterator end() const { return const_iterator(Data(), size()); }
+
+  reverse_iterator rbegin() { return reverse_iterator(end()); }
+  const_reverse_iterator rbegin() const {
+    return const_reverse_iterator(end());
+  }
+
+  reverse_iterator rend() { return reverse_iterator(begin()); }
+  const_reverse_iterator rend() const {
+    return const_reverse_iterator(begin());
+  }
+
+  const_iterator cbegin() const { return begin(); }
+
+  const_iterator cend() const { return end(); }
+
+  const_reverse_iterator crbegin() const { return rbegin(); }
+
+  const_reverse_iterator crend() const { return rend(); }
+
+  // Change elements if you have a non-const pointer to this object.
+  // Scalars only. See reflection.h, and the documentation.
+  void Mutate(SizeT i, const T &val) {
+    FLATBUFFERS_ASSERT(i < size());
+    WriteScalar(data() + i, val);
+  }
+
+  // Change an element of a vector of tables (or strings).
+  // "val" points to the new table/string, as you can obtain from
+  // e.g. reflection::AddFlatBuffer().
+  void MutateOffset(SizeT i, const uint8_t *val) {
+    FLATBUFFERS_ASSERT(i < size());
+    static_assert(sizeof(T) == sizeof(SizeT), "Unrelated types");
+    WriteScalar(data() + i,
+                static_cast<SizeT>(val - (Data() + i * sizeof(SizeT))));
+  }
+
+  // Get a mutable pointer to tables/strings inside this vector.
+  mutable_return_type GetMutableObject(SizeT i) const {
+    FLATBUFFERS_ASSERT(i < size());
+    return const_cast<mutable_return_type>(IndirectHelper<T>::Read(Data(), i));
+  }
+
+  // The raw data in little endian format. Use with care.
+  const uint8_t *Data() const {
+    return reinterpret_cast<const uint8_t *>(&length_ + 1);
+  }
+
+  uint8_t *Data() { return reinterpret_cast<uint8_t *>(&length_ + 1); }
+
+  // Similarly, but typed, much like std::vector::data
+  const T *data() const { return reinterpret_cast<const T *>(Data()); }
+  T *data() { return reinterpret_cast<T *>(Data()); }
+
+  template<typename K> return_type LookupByKey(K key) const {
+    void *search_result = std::bsearch(
+        &key, Data(), size(), IndirectHelper<T>::element_stride, KeyCompare<K>);
+
+    if (!search_result) {
+      return nullptr;  // Key not found.
+    }
+
+    const uint8_t *element = reinterpret_cast<const uint8_t *>(search_result);
+
+    return IndirectHelper<T>::Read(element, 0);
+  }
+
+  template<typename K> mutable_return_type MutableLookupByKey(K key) {
+    return const_cast<mutable_return_type>(LookupByKey(key));
+  }
+
+ protected:
+  // This class is only used to access pre-existing data. Don't ever
+  // try to construct these manually.
+  Vector();
+
+  SizeT length_;
+
+ private:
+  // This class is a pointer. Copying will therefore create an invalid object.
+  // Private and unimplemented copy constructor.
+  Vector(const Vector &);
+  Vector &operator=(const Vector &);
+
+  template<typename K> static int KeyCompare(const void *ap, const void *bp) {
+    const K *key = reinterpret_cast<const K *>(ap);
+    const uint8_t *data = reinterpret_cast<const uint8_t *>(bp);
+    auto table = IndirectHelper<T>::Read(data, 0);
+
+    // std::bsearch compares with the operands transposed, so we negate the
+    // result here.
+    return -table->KeyCompareWithValue(*key);
+  }
+};
+
+template<typename T> using Vector64 = Vector<T, uoffset64_t>;
+
+template<class U>
+FLATBUFFERS_CONSTEXPR_CPP11 flatbuffers::span<U> make_span(Vector<U> &vec)
+    FLATBUFFERS_NOEXCEPT {
+  static_assert(Vector<U>::is_span_observable,
+                "wrong type U, only LE-scalar, or byte types are allowed");
+  return span<U>(vec.data(), vec.size());
+}
+
+template<class U>
+FLATBUFFERS_CONSTEXPR_CPP11 flatbuffers::span<const U> make_span(
+    const Vector<U> &vec) FLATBUFFERS_NOEXCEPT {
+  static_assert(Vector<U>::is_span_observable,
+                "wrong type U, only LE-scalar, or byte types are allowed");
+  return span<const U>(vec.data(), vec.size());
+}
+
+template<class U>
+FLATBUFFERS_CONSTEXPR_CPP11 flatbuffers::span<uint8_t> make_bytes_span(
+    Vector<U> &vec) FLATBUFFERS_NOEXCEPT {
+  static_assert(Vector<U>::scalar_tag::value,
+                "wrong type U, only LE-scalar, or byte types are allowed");
+  return span<uint8_t>(vec.Data(), vec.size() * sizeof(U));
+}
+
+template<class U>
+FLATBUFFERS_CONSTEXPR_CPP11 flatbuffers::span<const uint8_t> make_bytes_span(
+    const Vector<U> &vec) FLATBUFFERS_NOEXCEPT {
+  static_assert(Vector<U>::scalar_tag::value,
+                "wrong type U, only LE-scalar, or byte types are allowed");
+  return span<const uint8_t>(vec.Data(), vec.size() * sizeof(U));
+}
+
+// Convenient helper functions to get a span of any vector, regardless
+// of whether it is null or not (the field is not set).
+template<class U>
+FLATBUFFERS_CONSTEXPR_CPP11 flatbuffers::span<U> make_span(Vector<U> *ptr)
+    FLATBUFFERS_NOEXCEPT {
+  static_assert(Vector<U>::is_span_observable,
+                "wrong type U, only LE-scalar, or byte types are allowed");
+  return ptr ? make_span(*ptr) : span<U>();
+}
+
+template<class U>
+FLATBUFFERS_CONSTEXPR_CPP11 flatbuffers::span<const U> make_span(
+    const Vector<U> *ptr) FLATBUFFERS_NOEXCEPT {
+  static_assert(Vector<U>::is_span_observable,
+                "wrong type U, only LE-scalar, or byte types are allowed");
+  return ptr ? make_span(*ptr) : span<const U>();
+}
+
+// Represent a vector much like the template above, but in this case we
+// don't know what the element types are (used with reflection.h).
+class VectorOfAny {
+ public:
+  uoffset_t size() const { return EndianScalar(length_); }
+
+  const uint8_t *Data() const {
+    return reinterpret_cast<const uint8_t *>(&length_ + 1);
+  }
+  uint8_t *Data() { return reinterpret_cast<uint8_t *>(&length_ + 1); }
+
+ protected:
+  VectorOfAny();
+
+  uoffset_t length_;
+
+ private:
+  VectorOfAny(const VectorOfAny &);
+  VectorOfAny &operator=(const VectorOfAny &);
+};
+
+template<typename T, typename U>
+Vector<Offset<T>> *VectorCast(Vector<Offset<U>> *ptr) {
+  static_assert(std::is_base_of<T, U>::value, "Unrelated types");
+  return reinterpret_cast<Vector<Offset<T>> *>(ptr);
+}
+
+template<typename T, typename U>
+const Vector<Offset<T>> *VectorCast(const Vector<Offset<U>> *ptr) {
+  static_assert(std::is_base_of<T, U>::value, "Unrelated types");
+  return reinterpret_cast<const Vector<Offset<T>> *>(ptr);
+}
+
+// Convenient helper function to get the length of any vector, regardless
+// of whether it is null or not (the field is not set).
+template<typename T> static inline size_t VectorLength(const Vector<T> *v) {
+  return v ? v->size() : 0;
+}
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_VERIFIER_H_
@@ -0,0 +1,288 @@
+/*
+ * Copyright 2021 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_VECTOR_DOWNWARD_H_
+#define FLATBUFFERS_VECTOR_DOWNWARD_H_
+
+#include <cstdint>
+
+#include <algorithm>
+
+#include "flatbuffers/base.h"
+#include "flatbuffers/default_allocator.h"
+#include "flatbuffers/detached_buffer.h"
+
+namespace flatbuffers {
+
+// This is a minimal replication of std::vector<uint8_t> functionality,
+// except growing from higher to lower addresses. i.e. push_back() inserts data
+// in the lowest address in the vector.
+// Since this vector leaves the lower part unused, we support a "scratch-pad"
+// that can be stored there for temporary data, to share the allocated space.
+// Essentially, this supports 2 std::vectors in a single buffer.
+template<typename SizeT = uoffset_t> class vector_downward {
+ public:
+  explicit vector_downward(size_t initial_size, Allocator *allocator,
+                           bool own_allocator, size_t buffer_minalign,
+                           const SizeT max_size = FLATBUFFERS_MAX_BUFFER_SIZE)
+      : allocator_(allocator),
+        own_allocator_(own_allocator),
+        initial_size_(initial_size),
+        max_size_(max_size),
+        buffer_minalign_(buffer_minalign),
+        reserved_(0),
+        size_(0),
+        buf_(nullptr),
+        cur_(nullptr),
+        scratch_(nullptr) {}
+
+  vector_downward(vector_downward &&other) noexcept
+      // clang-format on
+      : allocator_(other.allocator_),
+        own_allocator_(other.own_allocator_),
+        initial_size_(other.initial_size_),
+        max_size_(other.max_size_),
+        buffer_minalign_(other.buffer_minalign_),
+        reserved_(other.reserved_),
+        size_(other.size_),
+        buf_(other.buf_),
+        cur_(other.cur_),
+        scratch_(other.scratch_) {
+    // No change in other.allocator_
+    // No change in other.initial_size_
+    // No change in other.buffer_minalign_
+    other.own_allocator_ = false;
+    other.reserved_ = 0;
+    other.buf_ = nullptr;
+    other.cur_ = nullptr;
+    other.scratch_ = nullptr;
+  }
+
+  vector_downward &operator=(vector_downward &&other) noexcept {
+    // Move construct a temporary and swap idiom
+    vector_downward temp(std::move(other));
+    swap(temp);
+    return *this;
+  }
+
+  ~vector_downward() {
+    clear_buffer();
+    clear_allocator();
+  }
+
+  void reset() {
+    clear_buffer();
+    clear();
+  }
+
+  void clear() {
+    if (buf_) {
+      cur_ = buf_ + reserved_;
+    } else {
+      reserved_ = 0;
+      cur_ = nullptr;
+    }
+    size_ = 0;
+    clear_scratch();
+  }
+
+  void clear_scratch() { scratch_ = buf_; }
+
+  void clear_allocator() {
+    if (own_allocator_ && allocator_) { delete allocator_; }
+    allocator_ = nullptr;
+    own_allocator_ = false;
+  }
+
+  void clear_buffer() {
+    if (buf_) Deallocate(allocator_, buf_, reserved_);
+    buf_ = nullptr;
+  }
+
+  // Relinquish the pointer to the caller.
+  uint8_t *release_raw(size_t &allocated_bytes, size_t &offset) {
+    auto *buf = buf_;
+    allocated_bytes = reserved_;
+    offset = vector_downward::offset();
+
+    // release_raw only relinquishes the buffer ownership.
+    // Does not deallocate or reset the allocator. Destructor will do that.
+    buf_ = nullptr;
+    clear();
+    return buf;
+  }
+
+  // Relinquish the pointer to the caller.
+  DetachedBuffer release() {
+    // allocator ownership (if any) is transferred to DetachedBuffer.
+    DetachedBuffer fb(allocator_, own_allocator_, buf_, reserved_, cur_,
+                      size());
+    if (own_allocator_) {
+      allocator_ = nullptr;
+      own_allocator_ = false;
+    }
+    buf_ = nullptr;
+    clear();
+    return fb;
+  }
+
+  size_t ensure_space(size_t len) {
+    FLATBUFFERS_ASSERT(cur_ >= scratch_ && scratch_ >= buf_);
+    // If the length is larger than the unused part of the buffer, we need to
+    // grow.
+    if (len > unused_buffer_size()) { reallocate(len); }
+    FLATBUFFERS_ASSERT(size() < max_size_);
+    return len;
+  }
+
+  inline uint8_t *make_space(size_t len) {
+    if (len) {
+      ensure_space(len);
+      cur_ -= len;
+      size_ += static_cast<SizeT>(len);
+    }
+    return cur_;
+  }
+
+  // Returns nullptr if using the DefaultAllocator.
+  Allocator *get_custom_allocator() { return allocator_; }
+
+  // The current offset into the buffer.
+  size_t offset() const { return cur_ - buf_; }
+
+  // The total size of the vector (both the buffer and scratch parts).
+  inline SizeT size() const { return size_; }
+
+  // The size of the buffer part of the vector that is currently unused.
+  SizeT unused_buffer_size() const { return static_cast<SizeT>(cur_ - scratch_); }
+
+  // The size of the scratch part of the vector.
+  SizeT scratch_size() const { return static_cast<SizeT>(scratch_ - buf_); }
+
+  size_t capacity() const { return reserved_; }
+
+  uint8_t *data() const {
+    FLATBUFFERS_ASSERT(cur_);
+    return cur_;
+  }
+
+  uint8_t *scratch_data() const {
+    FLATBUFFERS_ASSERT(buf_);
+    return buf_;
+  }
+
+  uint8_t *scratch_end() const {
+    FLATBUFFERS_ASSERT(scratch_);
+    return scratch_;
+  }
+
+  uint8_t *data_at(size_t offset) const { return buf_ + reserved_ - offset; }
+
+  void push(const uint8_t *bytes, size_t num) {
+    if (num > 0) { memcpy(make_space(num), bytes, num); }
+  }
+
+  // Specialized version of push() that avoids memcpy call for small data.
+  template<typename T> void push_small(const T &little_endian_t) {
+    make_space(sizeof(T));
+    *reinterpret_cast<T *>(cur_) = little_endian_t;
+  }
+
+  template<typename T> void scratch_push_small(const T &t) {
+    ensure_space(sizeof(T));
+    *reinterpret_cast<T *>(scratch_) = t;
+    scratch_ += sizeof(T);
+  }
+
+  // fill() is most frequently called with small byte counts (<= 4),
+  // which is why we're using loops rather than calling memset.
+  void fill(size_t zero_pad_bytes) {
+    make_space(zero_pad_bytes);
+    for (size_t i = 0; i < zero_pad_bytes; i++) cur_[i] = 0;
+  }
+
+  // Version for when we know the size is larger.
+  // Precondition: zero_pad_bytes > 0
+  void fill_big(size_t zero_pad_bytes) {
+    memset(make_space(zero_pad_bytes), 0, zero_pad_bytes);
+  }
+
+  void pop(size_t bytes_to_remove) {
+    cur_ += bytes_to_remove;
+    size_ -= static_cast<SizeT>(bytes_to_remove);
+  }
+
+  void scratch_pop(size_t bytes_to_remove) { scratch_ -= bytes_to_remove; }
+
+  void swap(vector_downward &other) {
+    using std::swap;
+    swap(allocator_, other.allocator_);
+    swap(own_allocator_, other.own_allocator_);
+    swap(initial_size_, other.initial_size_);
+    swap(buffer_minalign_, other.buffer_minalign_);
+    swap(reserved_, other.reserved_);
+    swap(size_, other.size_);
+    swap(max_size_, other.max_size_);
+    swap(buf_, other.buf_);
+    swap(cur_, other.cur_);
+    swap(scratch_, other.scratch_);
+  }
+
+  void swap_allocator(vector_downward &other) {
+    using std::swap;
+    swap(allocator_, other.allocator_);
+    swap(own_allocator_, other.own_allocator_);
+  }
+
+ private:
+  // You shouldn't really be copying instances of this class.
+  FLATBUFFERS_DELETE_FUNC(vector_downward(const vector_downward &));
+  FLATBUFFERS_DELETE_FUNC(vector_downward &operator=(const vector_downward &));
+
+  Allocator *allocator_;
+  bool own_allocator_;
+  size_t initial_size_;
+
+  // The maximum size the vector can be.
+  SizeT max_size_;
+  size_t buffer_minalign_;
+  size_t reserved_;
+  SizeT size_;
+  uint8_t *buf_;
+  uint8_t *cur_;  // Points at location between empty (below) and used (above).
+  uint8_t *scratch_;  // Points to the end of the scratchpad in use.
+
+  void reallocate(size_t len) {
+    auto old_reserved = reserved_;
+    auto old_size = size();
+    auto old_scratch_size = scratch_size();
+    reserved_ +=
+        (std::max)(len, old_reserved ? old_reserved / 2 : initial_size_);
+    reserved_ = (reserved_ + buffer_minalign_ - 1) & ~(buffer_minalign_ - 1);
+    if (buf_) {
+      buf_ = ReallocateDownward(allocator_, buf_, old_reserved, reserved_,
+                                old_size, old_scratch_size);
+    } else {
+      buf_ = Allocate(allocator_, reserved_);
+    }
+    cur_ = buf_ + reserved_ - old_size;
+    scratch_ = buf_ + old_scratch_size;
+  }
+};
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_VECTOR_DOWNWARD_H_
@@ -0,0 +1,330 @@
+/*
+ * Copyright 2021 Google Inc. All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FLATBUFFERS_VERIFIER_H_
+#define FLATBUFFERS_VERIFIER_H_
+
+#include "flatbuffers/base.h"
+#include "flatbuffers/vector.h"
+
+namespace flatbuffers {
+
+// Helper class to verify the integrity of a FlatBuffer
+class Verifier FLATBUFFERS_FINAL_CLASS {
+ public:
+  struct Options {
+    // The maximum nesting of tables and vectors before we call it invalid.
+    uoffset_t max_depth = 64;
+    // The maximum number of tables we will verify before we call it invalid.
+    uoffset_t max_tables = 1000000;
+    // If true, verify all data is aligned.
+    bool check_alignment = true;
+    // If true, run verifier on nested flatbuffers
+    bool check_nested_flatbuffers = true;
+    // The maximum size of a buffer.
+    size_t max_size = FLATBUFFERS_MAX_BUFFER_SIZE;
+    // Use assertions to check for errors.
+    bool assert = false;
+  };
+
+  explicit Verifier(const uint8_t *const buf, const size_t buf_len,
+                    const Options &opts)
+      : buf_(buf), size_(buf_len), opts_(opts) {
+    FLATBUFFERS_ASSERT(size_ < opts.max_size);
+  }
+
+  // Deprecated API, please construct with Verifier::Options.
+  Verifier(const uint8_t *const buf, const size_t buf_len,
+           const uoffset_t max_depth = 64, const uoffset_t max_tables = 1000000,
+           const bool check_alignment = true)
+      : Verifier(buf, buf_len, [&] {
+          Options opts;
+          opts.max_depth = max_depth;
+          opts.max_tables = max_tables;
+          opts.check_alignment = check_alignment;
+          return opts;
+        }()) {}
+
+  // Central location where any verification failures register.
+  bool Check(const bool ok) const {
+    // clang-format off
+    #ifdef FLATBUFFERS_DEBUG_VERIFICATION_FAILURE
+       if (opts_.assert) { FLATBUFFERS_ASSERT(ok); }
+    #endif
+    #ifdef FLATBUFFERS_TRACK_VERIFIER_BUFFER_SIZE
+      if (!ok)
+        upper_bound_ = 0;
+    #endif
+    // clang-format on
+    return ok;
+  }
+
+  // Verify any range within the buffer.
+  bool Verify(const size_t elem, const size_t elem_len) const {
+    // clang-format off
+    #ifdef FLATBUFFERS_TRACK_VERIFIER_BUFFER_SIZE
+      auto upper_bound = elem + elem_len;
+      if (upper_bound_ < upper_bound)
+        upper_bound_ =  upper_bound;
+    #endif
+    // clang-format on
+    return Check(elem_len < size_ && elem <= size_ - elem_len);
+  }
+
+  bool VerifyAlignment(const size_t elem, const size_t align) const {
+    return Check((elem & (align - 1)) == 0 || !opts_.check_alignment);
+  }
+
+  // Verify a range indicated by sizeof(T).
+  template<typename T> bool Verify(const size_t elem) const {
+    return VerifyAlignment(elem, sizeof(T)) && Verify(elem, sizeof(T));
+  }
+
+  bool VerifyFromPointer(const uint8_t *const p, const size_t len) {
+    return Verify(static_cast<size_t>(p - buf_), len);
+  }
+
+  // Verify relative to a known-good base pointer.
+  bool VerifyFieldStruct(const uint8_t *const base, const voffset_t elem_off,
+                         const size_t elem_len, const size_t align) const {
+    const auto f = static_cast<size_t>(base - buf_) + elem_off;
+    return VerifyAlignment(f, align) && Verify(f, elem_len);
+  }
+
+  template<typename T>
+  bool VerifyField(const uint8_t *const base, const voffset_t elem_off,
+                   const size_t align) const {
+    const auto f = static_cast<size_t>(base - buf_) + elem_off;
+    return VerifyAlignment(f, align) && Verify(f, sizeof(T));
+  }
+
+  // Verify a pointer (may be NULL) of a table type.
+  template<typename T> bool VerifyTable(const T *const table) {
+    return !table || table->Verify(*this);
+  }
+
+  // Verify a pointer (may be NULL) of any vector type.
+  template<int &..., typename T, typename LenT>
+  bool VerifyVector(const Vector<T, LenT> *const vec) const {
+    return !vec || VerifyVectorOrString<LenT>(
+                       reinterpret_cast<const uint8_t *>(vec), sizeof(T));
+  }
+
+  // Verify a pointer (may be NULL) of a vector to struct.
+  template<int &..., typename T, typename LenT>
+  bool VerifyVector(const Vector<const T *, LenT> *const vec) const {
+    return VerifyVector(reinterpret_cast<const Vector<T, LenT> *>(vec));
+  }
+
+  // Verify a pointer (may be NULL) to string.
+  bool VerifyString(const String *const str) const {
+    size_t end;
+    return !str || (VerifyVectorOrString<uoffset_t>(
+                        reinterpret_cast<const uint8_t *>(str), 1, &end) &&
+                    Verify(end, 1) &&           // Must have terminator
+                    Check(buf_[end] == '\0'));  // Terminating byte must be 0.
+  }
+
+  // Common code between vectors and strings.
+  template<typename LenT = uoffset_t>
+  bool VerifyVectorOrString(const uint8_t *const vec, const size_t elem_size,
+                            size_t *const end = nullptr) const {
+    const auto vec_offset = static_cast<size_t>(vec - buf_);
+    // Check we can read the size field.
+    if (!Verify<LenT>(vec_offset)) return false;
+    // Check the whole array. If this is a string, the byte past the array must
+    // be 0.
+    const LenT size = ReadScalar<LenT>(vec);
+    const auto max_elems = opts_.max_size / elem_size;
+    if (!Check(size < max_elems))
+      return false;  // Protect against byte_size overflowing.
+    const auto byte_size = sizeof(LenT) + elem_size * size;
+    if (end) *end = vec_offset + byte_size;
+    return Verify(vec_offset, byte_size);
+  }
+
+  // Special case for string contents, after the above has been called.
+  bool VerifyVectorOfStrings(const Vector<Offset<String>> *const vec) const {
+    if (vec) {
+      for (uoffset_t i = 0; i < vec->size(); i++) {
+        if (!VerifyString(vec->Get(i))) return false;
+      }
+    }
+    return true;
+  }
+
+  // Special case for table contents, after the above has been called.
+  template<typename T>
+  bool VerifyVectorOfTables(const Vector<Offset<T>> *const vec) {
+    if (vec) {
+      for (uoffset_t i = 0; i < vec->size(); i++) {
+        if (!vec->Get(i)->Verify(*this)) return false;
+      }
+    }
+    return true;
+  }
+
+  __suppress_ubsan__("unsigned-integer-overflow") bool VerifyTableStart(
+      const uint8_t *const table) {
+    // Check the vtable offset.
+    const auto tableo = static_cast<size_t>(table - buf_);
+    if (!Verify<soffset_t>(tableo)) return false;
+    // This offset may be signed, but doing the subtraction unsigned always
+    // gives the result we want.
+    const auto vtableo =
+        tableo - static_cast<size_t>(ReadScalar<soffset_t>(table));
+    // Check the vtable size field, then check vtable fits in its entirety.
+    if (!(VerifyComplexity() && Verify<voffset_t>(vtableo) &&
+          VerifyAlignment(ReadScalar<voffset_t>(buf_ + vtableo),
+                          sizeof(voffset_t))))
+      return false;
+    const auto vsize = ReadScalar<voffset_t>(buf_ + vtableo);
+    return Check((vsize & 1) == 0) && Verify(vtableo, vsize);
+  }
+
+  template<typename T>
+  bool VerifyBufferFromStart(const char *const identifier, const size_t start) {
+    // Buffers have to be of some size to be valid. The reason it is a runtime
+    // check instead of static_assert, is that nested flatbuffers go through
+    // this call and their size is determined at runtime.
+    if (!Check(size_ >= FLATBUFFERS_MIN_BUFFER_SIZE)) return false;
+
+    // If an identifier is provided, check that we have a buffer
+    if (identifier && !Check((size_ >= 2 * sizeof(flatbuffers::uoffset_t) &&
+                              BufferHasIdentifier(buf_ + start, identifier)))) {
+      return false;
+    }
+
+    // Call T::Verify, which must be in the generated code for this type.
+    const auto o = VerifyOffset<uoffset_t>(start);
+    return Check(o != 0) &&
+           reinterpret_cast<const T *>(buf_ + start + o)->Verify(*this)
+    // clang-format off
+    #ifdef FLATBUFFERS_TRACK_VERIFIER_BUFFER_SIZE
+           && GetComputedSize()
+    #endif
+        ;
+    // clang-format on
+  }
+
+  template<typename T, int &..., typename SizeT>
+  bool VerifyNestedFlatBuffer(const Vector<uint8_t, SizeT> *const buf,
+                              const char *const identifier) {
+    // Caller opted out of this.
+    if (!opts_.check_nested_flatbuffers) return true;
+
+    // An empty buffer is OK as it indicates not present.
+    if (!buf) return true;
+
+    // If there is a nested buffer, it must be greater than the min size.
+    if (!Check(buf->size() >= FLATBUFFERS_MIN_BUFFER_SIZE)) return false;
+
+    Verifier nested_verifier(buf->data(), buf->size(), opts_);
+    return nested_verifier.VerifyBuffer<T>(identifier);
+  }
+
+  // Verify this whole buffer, starting with root type T.
+  template<typename T> bool VerifyBuffer() { return VerifyBuffer<T>(nullptr); }
+
+  template<typename T> bool VerifyBuffer(const char *const identifier) {
+    return VerifyBufferFromStart<T>(identifier, 0);
+  }
+
+  template<typename T, typename SizeT = uoffset_t>
+  bool VerifySizePrefixedBuffer(const char *const identifier) {
+    return Verify<SizeT>(0U) &&
+           Check(ReadScalar<SizeT>(buf_) == size_ - sizeof(SizeT)) &&
+           VerifyBufferFromStart<T>(identifier, sizeof(SizeT));
+  }
+
+  template<typename OffsetT = uoffset_t, typename SOffsetT = soffset_t>
+  size_t VerifyOffset(const size_t start) const {
+    if (!Verify<OffsetT>(start)) return 0;
+    const auto o = ReadScalar<OffsetT>(buf_ + start);
+    // May not point to itself.
+    if (!Check(o != 0)) return 0;
+    // Can't wrap around larger than the max size.
+    if (!Check(static_cast<SOffsetT>(o) >= 0)) return 0;
+    // Must be inside the buffer to create a pointer from it (pointer outside
+    // buffer is UB).
+    if (!Verify(start + o, 1)) return 0;
+    return o;
+  }
+
+  template<typename OffsetT = uoffset_t>
+  size_t VerifyOffset(const uint8_t *const base, const voffset_t start) const {
+    return VerifyOffset<OffsetT>(static_cast<size_t>(base - buf_) + start);
+  }
+
+  // Called at the start of a table to increase counters measuring data
+  // structure depth and amount, and possibly bails out with false if limits set
+  // by the constructor have been hit. Needs to be balanced with EndTable().
+  bool VerifyComplexity() {
+    depth_++;
+    num_tables_++;
+    return Check(depth_ <= opts_.max_depth && num_tables_ <= opts_.max_tables);
+  }
+
+  // Called at the end of a table to pop the depth count.
+  bool EndTable() {
+    depth_--;
+    return true;
+  }
+
+  // Returns the message size in bytes
+  size_t GetComputedSize() const {
+    // clang-format off
+    #ifdef FLATBUFFERS_TRACK_VERIFIER_BUFFER_SIZE
+      uintptr_t size = upper_bound_;
+      // Align the size to uoffset_t
+      size = (size - 1 + sizeof(uoffset_t)) & ~(sizeof(uoffset_t) - 1);
+      return (size > size_) ?  0 : size;
+    #else
+      // Must turn on FLATBUFFERS_TRACK_VERIFIER_BUFFER_SIZE for this to work.
+      (void)upper_bound_;
+      FLATBUFFERS_ASSERT(false);
+      return 0;
+    #endif
+    // clang-format on
+  }
+
+  std::vector<uint8_t> *GetFlexReuseTracker() { return flex_reuse_tracker_; }
+
+  void SetFlexReuseTracker(std::vector<uint8_t> *const rt) {
+    flex_reuse_tracker_ = rt;
+  }
+
+ private:
+  const uint8_t *buf_;
+  const size_t size_;
+  const Options opts_;
+
+  mutable size_t upper_bound_ = 0;
+
+  uoffset_t depth_ = 0;
+  uoffset_t num_tables_ = 0;
+  std::vector<uint8_t> *flex_reuse_tracker_ = nullptr;
+};
+
+// Specialization for 64-bit offsets.
+template<>
+inline size_t Verifier::VerifyOffset<uoffset64_t>(const size_t start) const {
+  return VerifyOffset<uoffset64_t, soffset64_t>(start);
+}
+
+}  // namespace flatbuffers
+
+#endif  // FLATBUFFERS_VERIFIER_H_
@@ -2,19 +2,9 @@
 // File: vk_platform.h
 //
 /*
-** Copyright (c) 2014-2017 The Khronos Group Inc.
+** Copyright 2014-2023 The Khronos Group Inc.
 **
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
+** SPDX-License-Identifier: Apache-2.0
 */


@@ -52,7 +42,7 @@ extern "C"
    #define VKAPI_CALL __stdcall
    #define VKAPI_PTR  VKAPI_CALL
 #elif defined(__ANDROID__) && defined(__ARM_ARCH) && __ARM_ARCH < 7
-    #error "Vulkan isn't supported for the 'armeabi' NDK ABI"
+    #error "Vulkan is not supported for the 'armeabi' NDK ABI"
 #elif defined(__ANDROID__) && defined(__ARM_ARCH) && __ARM_ARCH >= 7 && defined(__ARM_32BIT_STATE)
    // On Android 32-bit ARM targets, Vulkan functions use the "hardfloat"
    // calling convention, i.e. float parameters are passed in registers. This
@@ -68,7 +58,9 @@ extern "C"
    #define VKAPI_PTR
 #endif

-#include <stddef.h>
+#if !defined(VK_NO_STDDEF_H)
+    #include <stddef.h>
+#endif // !defined(VK_NO_STDDEF_H)

 #if !defined(VK_NO_STDINT_H)
    #if defined(_MSC_VER) && (_MSC_VER < 1600)
@@ -2,19 +2,9 @@
 #define VULKAN_H_ 1

 /*
-** Copyright (c) 2015-2018 The Khronos Group Inc.
+** Copyright 2015-2023 The Khronos Group Inc.
 **
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
+** SPDX-License-Identifier: Apache-2.0
 */

 #include "vk_platform.h"
@@ -38,20 +28,16 @@
 #include "vulkan_macos.h"
 #endif

-
-#ifdef VK_USE_PLATFORM_MIR_KHR
-#include <mir_toolkit/client_types.h>
-#include "vulkan_mir.h"
+#ifdef VK_USE_PLATFORM_METAL_EXT
+#include "vulkan_metal.h"
 #endif

-
 #ifdef VK_USE_PLATFORM_VI_NN
 #include "vulkan_vi.h"
 #endif


 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
-#include <wayland-client.h>
 #include "vulkan_wayland.h"
 #endif

@@ -74,10 +60,40 @@
 #endif


+#ifdef VK_USE_PLATFORM_DIRECTFB_EXT
+#include <directfb.h>
+#include "vulkan_directfb.h"
+#endif
+
+
 #ifdef VK_USE_PLATFORM_XLIB_XRANDR_EXT
 #include <X11/Xlib.h>
 #include <X11/extensions/Xrandr.h>
 #include "vulkan_xlib_xrandr.h"
 #endif

+
+#ifdef VK_USE_PLATFORM_GGP
+#include <ggp_c/vulkan_types.h>
+#include "vulkan_ggp.h"
+#endif
+
+
+#ifdef VK_USE_PLATFORM_SCREEN_QNX
+#include <screen/screen.h>
+#include "vulkan_screen.h"
+#endif
+
+
+#ifdef VK_USE_PLATFORM_SCI
+#include <nvscisync.h>
+#include <nvscibuf.h>
+#include "vulkan_sci.h"
+#endif
+
+
+#ifdef VK_ENABLE_BETA_EXTENSIONS
+#include "vulkan_beta.h"
+#endif
+
 #endif // VULKAN_H_
@@ -1,24 +1,10 @@
 #ifndef VULKAN_ANDROID_H_
 #define VULKAN_ANDROID_H_ 1

-#ifdef __cplusplus
-extern "C" {
-#endif
-
 /*
-** Copyright (c) 2015-2018 The Khronos Group Inc.
+** Copyright 2015-2023 The Khronos Group Inc.
 **
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
+** SPDX-License-Identifier: Apache-2.0
 */

 /*
@@ -27,14 +13,17 @@ extern "C" {
 */


+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
 #define VK_KHR_android_surface 1
 struct ANativeWindow;
-
 #define VK_KHR_ANDROID_SURFACE_SPEC_VERSION 6
 #define VK_KHR_ANDROID_SURFACE_EXTENSION_NAME "VK_KHR_android_surface"
-
 typedef VkFlags VkAndroidSurfaceCreateFlagsKHR;
-
 typedef struct VkAndroidSurfaceCreateInfoKHR {
    VkStructureType                   sType;
    const void*                       pNext;
@@ -42,7 +31,6 @@ typedef struct VkAndroidSurfaceCreateInfoKHR {
    struct ANativeWindow*             window;
 } VkAndroidSurfaceCreateInfoKHR;

-
 typedef VkResult (VKAPI_PTR *PFN_vkCreateAndroidSurfaceKHR)(VkInstance instance, const VkAndroidSurfaceCreateInfoKHR* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);

 #ifndef VK_NO_PROTOTYPES
@@ -53,12 +41,11 @@ VKAPI_ATTR VkResult VKAPI_CALL vkCreateAndroidSurfaceKHR(
    VkSurfaceKHR*                               pSurface);
 #endif

+
 #define VK_ANDROID_external_memory_android_hardware_buffer 1
 struct AHardwareBuffer;
-
-#define VK_ANDROID_EXTERNAL_MEMORY_ANDROID_HARDWARE_BUFFER_SPEC_VERSION 3
+#define VK_ANDROID_EXTERNAL_MEMORY_ANDROID_HARDWARE_BUFFER_SPEC_VERSION 5
 #define VK_ANDROID_EXTERNAL_MEMORY_ANDROID_HARDWARE_BUFFER_EXTENSION_NAME "VK_ANDROID_external_memory_android_hardware_buffer"
-
 typedef struct VkAndroidHardwareBufferUsageANDROID {
    VkStructureType    sType;
    void*              pNext;
@@ -103,6 +90,18 @@ typedef struct VkExternalFormatANDROID {
    uint64_t           externalFormat;
 } VkExternalFormatANDROID;

+typedef struct VkAndroidHardwareBufferFormatProperties2ANDROID {
+    VkStructureType                  sType;
+    void*                            pNext;
+    VkFormat                         format;
+    uint64_t                         externalFormat;
+    VkFormatFeatureFlags2            formatFeatures;
+    VkComponentMapping               samplerYcbcrConversionComponents;
+    VkSamplerYcbcrModelConversion    suggestedYcbcrModel;
+    VkSamplerYcbcrRange              suggestedYcbcrRange;
+    VkChromaLocation                 suggestedXChromaOffset;
+    VkChromaLocation                 suggestedYChromaOffset;
+} VkAndroidHardwareBufferFormatProperties2ANDROID;

 typedef VkResult (VKAPI_PTR *PFN_vkGetAndroidHardwareBufferPropertiesANDROID)(VkDevice device, const struct AHardwareBuffer* buffer, VkAndroidHardwareBufferPropertiesANDROID* pProperties);
 typedef VkResult (VKAPI_PTR *PFN_vkGetMemoryAndroidHardwareBufferANDROID)(VkDevice device, const VkMemoryGetAndroidHardwareBufferInfoANDROID* pInfo, struct AHardwareBuffer** pBuffer);
@@ -0,0 +1,492 @@
+#ifndef VULKAN_BETA_H_
+#define VULKAN_BETA_H_ 1
+
+/*
+** Copyright 2015-2023 The Khronos Group Inc.
+**
+** SPDX-License-Identifier: Apache-2.0
+*/
+
+/*
+** This header is generated from the Khronos Vulkan XML API Registry.
+**
+*/
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
+#define VK_KHR_portability_subset 1
+#define VK_KHR_PORTABILITY_SUBSET_SPEC_VERSION 1
+#define VK_KHR_PORTABILITY_SUBSET_EXTENSION_NAME "VK_KHR_portability_subset"
+typedef struct VkPhysicalDevicePortabilitySubsetFeaturesKHR {
+    VkStructureType    sType;
+    void*              pNext;
+    VkBool32           constantAlphaColorBlendFactors;
+    VkBool32           events;
+    VkBool32           imageViewFormatReinterpretation;
+    VkBool32           imageViewFormatSwizzle;
+    VkBool32           imageView2DOn3DImage;
+    VkBool32           multisampleArrayImage;
+    VkBool32           mutableComparisonSamplers;
+    VkBool32           pointPolygons;
+    VkBool32           samplerMipLodBias;
+    VkBool32           separateStencilMaskRef;
+    VkBool32           shaderSampleRateInterpolationFunctions;
+    VkBool32           tessellationIsolines;
+    VkBool32           tessellationPointMode;
+    VkBool32           triangleFans;
+    VkBool32           vertexAttributeAccessBeyondStride;
+} VkPhysicalDevicePortabilitySubsetFeaturesKHR;
+
+typedef struct VkPhysicalDevicePortabilitySubsetPropertiesKHR {
+    VkStructureType    sType;
+    void*              pNext;
+    uint32_t           minVertexInputBindingStrideAlignment;
+} VkPhysicalDevicePortabilitySubsetPropertiesKHR;
+
+
+
+#define VK_KHR_video_encode_queue 1
+#define VK_KHR_VIDEO_ENCODE_QUEUE_SPEC_VERSION 8
+#define VK_KHR_VIDEO_ENCODE_QUEUE_EXTENSION_NAME "VK_KHR_video_encode_queue"
+
+typedef enum VkVideoEncodeTuningModeKHR {
+    VK_VIDEO_ENCODE_TUNING_MODE_DEFAULT_KHR = 0,
+    VK_VIDEO_ENCODE_TUNING_MODE_HIGH_QUALITY_KHR = 1,
+    VK_VIDEO_ENCODE_TUNING_MODE_LOW_LATENCY_KHR = 2,
+    VK_VIDEO_ENCODE_TUNING_MODE_ULTRA_LOW_LATENCY_KHR = 3,
+    VK_VIDEO_ENCODE_TUNING_MODE_LOSSLESS_KHR = 4,
+    VK_VIDEO_ENCODE_TUNING_MODE_MAX_ENUM_KHR = 0x7FFFFFFF
+} VkVideoEncodeTuningModeKHR;
+typedef VkFlags VkVideoEncodeFlagsKHR;
+
+typedef enum VkVideoEncodeCapabilityFlagBitsKHR {
+    VK_VIDEO_ENCODE_CAPABILITY_PRECEDING_EXTERNALLY_ENCODED_BYTES_BIT_KHR = 0x00000001,
+    VK_VIDEO_ENCODE_CAPABILITY_FLAG_BITS_MAX_ENUM_KHR = 0x7FFFFFFF
+} VkVideoEncodeCapabilityFlagBitsKHR;
+typedef VkFlags VkVideoEncodeCapabilityFlagsKHR;
+
+typedef enum VkVideoEncodeRateControlModeFlagBitsKHR {
+    VK_VIDEO_ENCODE_RATE_CONTROL_MODE_DEFAULT_KHR = 0,
+    VK_VIDEO_ENCODE_RATE_CONTROL_MODE_DISABLED_BIT_KHR = 0x00000001,
+    VK_VIDEO_ENCODE_RATE_CONTROL_MODE_CBR_BIT_KHR = 0x00000002,
+    VK_VIDEO_ENCODE_RATE_CONTROL_MODE_VBR_BIT_KHR = 0x00000004,
+    VK_VIDEO_ENCODE_RATE_CONTROL_MODE_FLAG_BITS_MAX_ENUM_KHR = 0x7FFFFFFF
+} VkVideoEncodeRateControlModeFlagBitsKHR;
+typedef VkFlags VkVideoEncodeRateControlModeFlagsKHR;
+
+typedef enum VkVideoEncodeFeedbackFlagBitsKHR {
+    VK_VIDEO_ENCODE_FEEDBACK_BITSTREAM_BUFFER_OFFSET_BIT_KHR = 0x00000001,
+    VK_VIDEO_ENCODE_FEEDBACK_BITSTREAM_BYTES_WRITTEN_BIT_KHR = 0x00000002,
+    VK_VIDEO_ENCODE_FEEDBACK_FLAG_BITS_MAX_ENUM_KHR = 0x7FFFFFFF
+} VkVideoEncodeFeedbackFlagBitsKHR;
+typedef VkFlags VkVideoEncodeFeedbackFlagsKHR;
+
+typedef enum VkVideoEncodeUsageFlagBitsKHR {
+    VK_VIDEO_ENCODE_USAGE_DEFAULT_KHR = 0,
+    VK_VIDEO_ENCODE_USAGE_TRANSCODING_BIT_KHR = 0x00000001,
+    VK_VIDEO_ENCODE_USAGE_STREAMING_BIT_KHR = 0x00000002,
+    VK_VIDEO_ENCODE_USAGE_RECORDING_BIT_KHR = 0x00000004,
+    VK_VIDEO_ENCODE_USAGE_CONFERENCING_BIT_KHR = 0x00000008,
+    VK_VIDEO_ENCODE_USAGE_FLAG_BITS_MAX_ENUM_KHR = 0x7FFFFFFF
+} VkVideoEncodeUsageFlagBitsKHR;
+typedef VkFlags VkVideoEncodeUsageFlagsKHR;
+
+typedef enum VkVideoEncodeContentFlagBitsKHR {
+    VK_VIDEO_ENCODE_CONTENT_DEFAULT_KHR = 0,
+    VK_VIDEO_ENCODE_CONTENT_CAMERA_BIT_KHR = 0x00000001,
+    VK_VIDEO_ENCODE_CONTENT_DESKTOP_BIT_KHR = 0x00000002,
+    VK_VIDEO_ENCODE_CONTENT_RENDERED_BIT_KHR = 0x00000004,
+    VK_VIDEO_ENCODE_CONTENT_FLAG_BITS_MAX_ENUM_KHR = 0x7FFFFFFF
+} VkVideoEncodeContentFlagBitsKHR;
+typedef VkFlags VkVideoEncodeContentFlagsKHR;
+typedef VkFlags VkVideoEncodeRateControlFlagsKHR;
+typedef struct VkVideoEncodeInfoKHR {
+    VkStructureType                       sType;
+    const void*                           pNext;
+    VkVideoEncodeFlagsKHR                 flags;
+    uint32_t                              qualityLevel;
+    VkBuffer                              dstBuffer;
+    VkDeviceSize                          dstBufferOffset;
+    VkDeviceSize                          dstBufferRange;
+    VkVideoPictureResourceInfoKHR         srcPictureResource;
+    const VkVideoReferenceSlotInfoKHR*    pSetupReferenceSlot;
+    uint32_t                              referenceSlotCount;
+    const VkVideoReferenceSlotInfoKHR*    pReferenceSlots;
+    uint32_t                              precedingExternallyEncodedBytes;
+} VkVideoEncodeInfoKHR;
+
+typedef struct VkVideoEncodeCapabilitiesKHR {
+    VkStructureType                         sType;
+    void*                                   pNext;
+    VkVideoEncodeCapabilityFlagsKHR         flags;
+    VkVideoEncodeRateControlModeFlagsKHR    rateControlModes;
+    uint32_t                                maxRateControlLayers;
+    uint32_t                                maxQualityLevels;
+    VkExtent2D                              inputImageDataFillAlignment;
+    VkVideoEncodeFeedbackFlagsKHR           supportedEncodeFeedbackFlags;
+} VkVideoEncodeCapabilitiesKHR;
+
+typedef struct VkQueryPoolVideoEncodeFeedbackCreateInfoKHR {
+    VkStructureType                  sType;
+    const void*                      pNext;
+    VkVideoEncodeFeedbackFlagsKHR    encodeFeedbackFlags;
+} VkQueryPoolVideoEncodeFeedbackCreateInfoKHR;
+
+typedef struct VkVideoEncodeUsageInfoKHR {
+    VkStructureType                 sType;
+    const void*                     pNext;
+    VkVideoEncodeUsageFlagsKHR      videoUsageHints;
+    VkVideoEncodeContentFlagsKHR    videoContentHints;
+    VkVideoEncodeTuningModeKHR      tuningMode;
+} VkVideoEncodeUsageInfoKHR;
+
+typedef struct VkVideoEncodeRateControlLayerInfoKHR {
+    VkStructureType    sType;
+    const void*        pNext;
+    uint64_t           averageBitrate;
+    uint64_t           maxBitrate;
+    uint32_t           frameRateNumerator;
+    uint32_t           frameRateDenominator;
+    uint32_t           virtualBufferSizeInMs;
+    uint32_t           initialVirtualBufferSizeInMs;
+} VkVideoEncodeRateControlLayerInfoKHR;
+
+typedef struct VkVideoEncodeRateControlInfoKHR {
+    VkStructureType                                sType;
+    const void*                                    pNext;
+    VkVideoEncodeRateControlFlagsKHR               flags;
+    VkVideoEncodeRateControlModeFlagBitsKHR        rateControlMode;
+    uint32_t                                       layerCount;
+    const VkVideoEncodeRateControlLayerInfoKHR*    pLayers;
+} VkVideoEncodeRateControlInfoKHR;
+
+typedef void (VKAPI_PTR *PFN_vkCmdEncodeVideoKHR)(VkCommandBuffer commandBuffer, const VkVideoEncodeInfoKHR* pEncodeInfo);
+
+#ifndef VK_NO_PROTOTYPES
+VKAPI_ATTR void VKAPI_CALL vkCmdEncodeVideoKHR(
+    VkCommandBuffer                             commandBuffer,
+    const VkVideoEncodeInfoKHR*                 pEncodeInfo);
+#endif
+
+
+#define VK_EXT_video_encode_h264 1
+#include "vk_video/vulkan_video_codec_h264std.h"
+#include "vk_video/vulkan_video_codec_h264std_encode.h"
+#define VK_EXT_VIDEO_ENCODE_H264_SPEC_VERSION 10
+#define VK_EXT_VIDEO_ENCODE_H264_EXTENSION_NAME "VK_EXT_video_encode_h264"
+
+typedef enum VkVideoEncodeH264RateControlStructureEXT {
+    VK_VIDEO_ENCODE_H264_RATE_CONTROL_STRUCTURE_UNKNOWN_EXT = 0,
+    VK_VIDEO_ENCODE_H264_RATE_CONTROL_STRUCTURE_FLAT_EXT = 1,
+    VK_VIDEO_ENCODE_H264_RATE_CONTROL_STRUCTURE_DYADIC_EXT = 2,
+    VK_VIDEO_ENCODE_H264_RATE_CONTROL_STRUCTURE_MAX_ENUM_EXT = 0x7FFFFFFF
+} VkVideoEncodeH264RateControlStructureEXT;
+
+typedef enum VkVideoEncodeH264CapabilityFlagBitsEXT {
+    VK_VIDEO_ENCODE_H264_CAPABILITY_DIRECT_8X8_INFERENCE_ENABLED_BIT_EXT = 0x00000001,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_DIRECT_8X8_INFERENCE_DISABLED_BIT_EXT = 0x00000002,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_SEPARATE_COLOUR_PLANE_BIT_EXT = 0x00000004,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_QPPRIME_Y_ZERO_TRANSFORM_BYPASS_BIT_EXT = 0x00000008,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_SCALING_LISTS_BIT_EXT = 0x00000010,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_HRD_COMPLIANCE_BIT_EXT = 0x00000020,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_CHROMA_QP_OFFSET_BIT_EXT = 0x00000040,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_SECOND_CHROMA_QP_OFFSET_BIT_EXT = 0x00000080,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_PIC_INIT_QP_MINUS26_BIT_EXT = 0x00000100,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_WEIGHTED_PRED_BIT_EXT = 0x00000200,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_WEIGHTED_BIPRED_EXPLICIT_BIT_EXT = 0x00000400,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_WEIGHTED_BIPRED_IMPLICIT_BIT_EXT = 0x00000800,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_WEIGHTED_PRED_NO_TABLE_BIT_EXT = 0x00001000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_TRANSFORM_8X8_BIT_EXT = 0x00002000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_CABAC_BIT_EXT = 0x00004000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_CAVLC_BIT_EXT = 0x00008000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_DEBLOCKING_FILTER_DISABLED_BIT_EXT = 0x00010000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_DEBLOCKING_FILTER_ENABLED_BIT_EXT = 0x00020000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_DEBLOCKING_FILTER_PARTIAL_BIT_EXT = 0x00040000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_DISABLE_DIRECT_SPATIAL_MV_PRED_BIT_EXT = 0x00080000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_MULTIPLE_SLICE_PER_FRAME_BIT_EXT = 0x00100000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_SLICE_MB_COUNT_BIT_EXT = 0x00200000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_ROW_UNALIGNED_SLICE_BIT_EXT = 0x00400000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_DIFFERENT_SLICE_TYPE_BIT_EXT = 0x00800000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_B_FRAME_IN_L1_LIST_BIT_EXT = 0x01000000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_DIFFERENT_REFERENCE_FINAL_LISTS_BIT_EXT = 0x02000000,
+    VK_VIDEO_ENCODE_H264_CAPABILITY_FLAG_BITS_MAX_ENUM_EXT = 0x7FFFFFFF
+} VkVideoEncodeH264CapabilityFlagBitsEXT;
+typedef VkFlags VkVideoEncodeH264CapabilityFlagsEXT;
+typedef struct VkVideoEncodeH264CapabilitiesEXT {
+    VkStructureType                        sType;
+    void*                                  pNext;
+    VkVideoEncodeH264CapabilityFlagsEXT    flags;
+    uint32_t                               maxPPictureL0ReferenceCount;
+    uint32_t                               maxBPictureL0ReferenceCount;
+    uint32_t                               maxL1ReferenceCount;
+    VkBool32                               motionVectorsOverPicBoundariesFlag;
+    uint32_t                               maxBytesPerPicDenom;
+    uint32_t                               maxBitsPerMbDenom;
+    uint32_t                               log2MaxMvLengthHorizontal;
+    uint32_t                               log2MaxMvLengthVertical;
+} VkVideoEncodeH264CapabilitiesEXT;
+
+typedef struct VkVideoEncodeH264SessionParametersAddInfoEXT {
+    VkStructureType                            sType;
+    const void*                                pNext;
+    uint32_t                                   stdSPSCount;
+    const StdVideoH264SequenceParameterSet*    pStdSPSs;
+    uint32_t                                   stdPPSCount;
+    const StdVideoH264PictureParameterSet*     pStdPPSs;
+} VkVideoEncodeH264SessionParametersAddInfoEXT;
+
+typedef struct VkVideoEncodeH264SessionParametersCreateInfoEXT {
+    VkStructureType                                        sType;
+    const void*                                            pNext;
+    uint32_t                                               maxStdSPSCount;
+    uint32_t                                               maxStdPPSCount;
+    const VkVideoEncodeH264SessionParametersAddInfoEXT*    pParametersAddInfo;
+} VkVideoEncodeH264SessionParametersCreateInfoEXT;
+
+typedef struct VkVideoEncodeH264NaluSliceInfoEXT {
+    VkStructureType                                sType;
+    const void*                                    pNext;
+    uint32_t                                       mbCount;
+    const StdVideoEncodeH264ReferenceListsInfo*    pStdReferenceFinalLists;
+    const StdVideoEncodeH264SliceHeader*           pStdSliceHeader;
+} VkVideoEncodeH264NaluSliceInfoEXT;
+
+typedef struct VkVideoEncodeH264VclFrameInfoEXT {
+    VkStructureType                                sType;
+    const void*                                    pNext;
+    const StdVideoEncodeH264ReferenceListsInfo*    pStdReferenceFinalLists;
+    uint32_t                                       naluSliceEntryCount;
+    const VkVideoEncodeH264NaluSliceInfoEXT*       pNaluSliceEntries;
+    const StdVideoEncodeH264PictureInfo*           pStdPictureInfo;
+} VkVideoEncodeH264VclFrameInfoEXT;
+
+typedef struct VkVideoEncodeH264DpbSlotInfoEXT {
+    VkStructureType                           sType;
+    const void*                               pNext;
+    const StdVideoEncodeH264ReferenceInfo*    pStdReferenceInfo;
+} VkVideoEncodeH264DpbSlotInfoEXT;
+
+typedef struct VkVideoEncodeH264ProfileInfoEXT {
+    VkStructureType           sType;
+    const void*               pNext;
+    StdVideoH264ProfileIdc    stdProfileIdc;
+} VkVideoEncodeH264ProfileInfoEXT;
+
+typedef struct VkVideoEncodeH264RateControlInfoEXT {
+    VkStructureType                             sType;
+    const void*                                 pNext;
+    uint32_t                                    gopFrameCount;
+    uint32_t                                    idrPeriod;
+    uint32_t                                    consecutiveBFrameCount;
+    VkVideoEncodeH264RateControlStructureEXT    rateControlStructure;
+    uint32_t                                    temporalLayerCount;
+} VkVideoEncodeH264RateControlInfoEXT;
+
+typedef struct VkVideoEncodeH264QpEXT {
+    int32_t    qpI;
+    int32_t    qpP;
+    int32_t    qpB;
+} VkVideoEncodeH264QpEXT;
+
+typedef struct VkVideoEncodeH264FrameSizeEXT {
+    uint32_t    frameISize;
+    uint32_t    framePSize;
+    uint32_t    frameBSize;
+} VkVideoEncodeH264FrameSizeEXT;
+
+typedef struct VkVideoEncodeH264RateControlLayerInfoEXT {
+    VkStructureType                  sType;
+    const void*                      pNext;
+    uint32_t                         temporalLayerId;
+    VkBool32                         useInitialRcQp;
+    VkVideoEncodeH264QpEXT           initialRcQp;
+    VkBool32                         useMinQp;
+    VkVideoEncodeH264QpEXT           minQp;
+    VkBool32                         useMaxQp;
+    VkVideoEncodeH264QpEXT           maxQp;
+    VkBool32                         useMaxFrameSize;
+    VkVideoEncodeH264FrameSizeEXT    maxFrameSize;
+} VkVideoEncodeH264RateControlLayerInfoEXT;
+
+
+
+#define VK_EXT_video_encode_h265 1
+#include "vk_video/vulkan_video_codec_h265std.h"
+#include "vk_video/vulkan_video_codec_h265std_encode.h"
+#define VK_EXT_VIDEO_ENCODE_H265_SPEC_VERSION 10
+#define VK_EXT_VIDEO_ENCODE_H265_EXTENSION_NAME "VK_EXT_video_encode_h265"
+
+typedef enum VkVideoEncodeH265RateControlStructureEXT {
+    VK_VIDEO_ENCODE_H265_RATE_CONTROL_STRUCTURE_UNKNOWN_EXT = 0,
+    VK_VIDEO_ENCODE_H265_RATE_CONTROL_STRUCTURE_FLAT_EXT = 1,
+    VK_VIDEO_ENCODE_H265_RATE_CONTROL_STRUCTURE_DYADIC_EXT = 2,
+    VK_VIDEO_ENCODE_H265_RATE_CONTROL_STRUCTURE_MAX_ENUM_EXT = 0x7FFFFFFF
+} VkVideoEncodeH265RateControlStructureEXT;
+
+typedef enum VkVideoEncodeH265CapabilityFlagBitsEXT {
+    VK_VIDEO_ENCODE_H265_CAPABILITY_SEPARATE_COLOUR_PLANE_BIT_EXT = 0x00000001,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_SCALING_LISTS_BIT_EXT = 0x00000002,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_SAMPLE_ADAPTIVE_OFFSET_ENABLED_BIT_EXT = 0x00000004,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_PCM_ENABLE_BIT_EXT = 0x00000008,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_SPS_TEMPORAL_MVP_ENABLED_BIT_EXT = 0x00000010,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_HRD_COMPLIANCE_BIT_EXT = 0x00000020,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_INIT_QP_MINUS26_BIT_EXT = 0x00000040,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_LOG2_PARALLEL_MERGE_LEVEL_MINUS2_BIT_EXT = 0x00000080,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_SIGN_DATA_HIDING_ENABLED_BIT_EXT = 0x00000100,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_TRANSFORM_SKIP_ENABLED_BIT_EXT = 0x00000200,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_TRANSFORM_SKIP_DISABLED_BIT_EXT = 0x00000400,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_PPS_SLICE_CHROMA_QP_OFFSETS_PRESENT_BIT_EXT = 0x00000800,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_WEIGHTED_PRED_BIT_EXT = 0x00001000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_WEIGHTED_BIPRED_BIT_EXT = 0x00002000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_WEIGHTED_PRED_NO_TABLE_BIT_EXT = 0x00004000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_TRANSQUANT_BYPASS_ENABLED_BIT_EXT = 0x00008000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_ENTROPY_CODING_SYNC_ENABLED_BIT_EXT = 0x00010000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_DEBLOCKING_FILTER_OVERRIDE_ENABLED_BIT_EXT = 0x00020000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_MULTIPLE_TILE_PER_FRAME_BIT_EXT = 0x00040000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_MULTIPLE_SLICE_PER_TILE_BIT_EXT = 0x00080000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_MULTIPLE_TILE_PER_SLICE_BIT_EXT = 0x00100000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_SLICE_SEGMENT_CTB_COUNT_BIT_EXT = 0x00200000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_ROW_UNALIGNED_SLICE_SEGMENT_BIT_EXT = 0x00400000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_DEPENDENT_SLICE_SEGMENT_BIT_EXT = 0x00800000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_DIFFERENT_SLICE_TYPE_BIT_EXT = 0x01000000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_B_FRAME_IN_L1_LIST_BIT_EXT = 0x02000000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_DIFFERENT_REFERENCE_FINAL_LISTS_BIT_EXT = 0x04000000,
+    VK_VIDEO_ENCODE_H265_CAPABILITY_FLAG_BITS_MAX_ENUM_EXT = 0x7FFFFFFF
+} VkVideoEncodeH265CapabilityFlagBitsEXT;
+typedef VkFlags VkVideoEncodeH265CapabilityFlagsEXT;
+
+typedef enum VkVideoEncodeH265CtbSizeFlagBitsEXT {
+    VK_VIDEO_ENCODE_H265_CTB_SIZE_16_BIT_EXT = 0x00000001,
+    VK_VIDEO_ENCODE_H265_CTB_SIZE_32_BIT_EXT = 0x00000002,
+    VK_VIDEO_ENCODE_H265_CTB_SIZE_64_BIT_EXT = 0x00000004,
+    VK_VIDEO_ENCODE_H265_CTB_SIZE_FLAG_BITS_MAX_ENUM_EXT = 0x7FFFFFFF
+} VkVideoEncodeH265CtbSizeFlagBitsEXT;
+typedef VkFlags VkVideoEncodeH265CtbSizeFlagsEXT;
+
+typedef enum VkVideoEncodeH265TransformBlockSizeFlagBitsEXT {
+    VK_VIDEO_ENCODE_H265_TRANSFORM_BLOCK_SIZE_4_BIT_EXT = 0x00000001,
+    VK_VIDEO_ENCODE_H265_TRANSFORM_BLOCK_SIZE_8_BIT_EXT = 0x00000002,
+    VK_VIDEO_ENCODE_H265_TRANSFORM_BLOCK_SIZE_16_BIT_EXT = 0x00000004,
+    VK_VIDEO_ENCODE_H265_TRANSFORM_BLOCK_SIZE_32_BIT_EXT = 0x00000008,
+    VK_VIDEO_ENCODE_H265_TRANSFORM_BLOCK_SIZE_FLAG_BITS_MAX_ENUM_EXT = 0x7FFFFFFF
+} VkVideoEncodeH265TransformBlockSizeFlagBitsEXT;
+typedef VkFlags VkVideoEncodeH265TransformBlockSizeFlagsEXT;
+typedef struct VkVideoEncodeH265CapabilitiesEXT {
+    VkStructureType                                sType;
+    void*                                          pNext;
+    VkVideoEncodeH265CapabilityFlagsEXT            flags;
+    VkVideoEncodeH265CtbSizeFlagsEXT               ctbSizes;
+    VkVideoEncodeH265TransformBlockSizeFlagsEXT    transformBlockSizes;
+    uint32_t                                       maxPPictureL0ReferenceCount;
+    uint32_t                                       maxBPictureL0ReferenceCount;
+    uint32_t                                       maxL1ReferenceCount;
+    uint32_t                                       maxSubLayersCount;
+    uint32_t                                       minLog2MinLumaCodingBlockSizeMinus3;
+    uint32_t                                       maxLog2MinLumaCodingBlockSizeMinus3;
+    uint32_t                                       minLog2MinLumaTransformBlockSizeMinus2;
+    uint32_t                                       maxLog2MinLumaTransformBlockSizeMinus2;
+    uint32_t                                       minMaxTransformHierarchyDepthInter;
+    uint32_t                                       maxMaxTransformHierarchyDepthInter;
+    uint32_t                                       minMaxTransformHierarchyDepthIntra;
+    uint32_t                                       maxMaxTransformHierarchyDepthIntra;
+    uint32_t                                       maxDiffCuQpDeltaDepth;
+    uint32_t                                       minMaxNumMergeCand;
+    uint32_t                                       maxMaxNumMergeCand;
+} VkVideoEncodeH265CapabilitiesEXT;
+
+typedef struct VkVideoEncodeH265SessionParametersAddInfoEXT {
+    VkStructureType                            sType;
+    const void*                                pNext;
+    uint32_t                                   stdVPSCount;
+    const StdVideoH265VideoParameterSet*       pStdVPSs;
+    uint32_t                                   stdSPSCount;
+    const StdVideoH265SequenceParameterSet*    pStdSPSs;
+    uint32_t                                   stdPPSCount;
+    const StdVideoH265PictureParameterSet*     pStdPPSs;
+} VkVideoEncodeH265SessionParametersAddInfoEXT;
+
+typedef struct VkVideoEncodeH265SessionParametersCreateInfoEXT {
+    VkStructureType                                        sType;
+    const void*                                            pNext;
+    uint32_t                                               maxStdVPSCount;
+    uint32_t                                               maxStdSPSCount;
+    uint32_t                                               maxStdPPSCount;
+    const VkVideoEncodeH265SessionParametersAddInfoEXT*    pParametersAddInfo;
+} VkVideoEncodeH265SessionParametersCreateInfoEXT;
+
+typedef struct VkVideoEncodeH265NaluSliceSegmentInfoEXT {
+    VkStructureType                                sType;
+    const void*                                    pNext;
+    uint32_t                                       ctbCount;
+    const StdVideoEncodeH265ReferenceListsInfo*    pStdReferenceFinalLists;
+    const StdVideoEncodeH265SliceSegmentHeader*    pStdSliceSegmentHeader;
+} VkVideoEncodeH265NaluSliceSegmentInfoEXT;
+
+typedef struct VkVideoEncodeH265VclFrameInfoEXT {
+    VkStructureType                                    sType;
+    const void*                                        pNext;
+    const StdVideoEncodeH265ReferenceListsInfo*        pStdReferenceFinalLists;
+    uint32_t                                           naluSliceSegmentEntryCount;
+    const VkVideoEncodeH265NaluSliceSegmentInfoEXT*    pNaluSliceSegmentEntries;
+    const StdVideoEncodeH265PictureInfo*               pStdPictureInfo;
+} VkVideoEncodeH265VclFrameInfoEXT;
+
+typedef struct VkVideoEncodeH265DpbSlotInfoEXT {
+    VkStructureType                           sType;
+    const void*                               pNext;
+    const StdVideoEncodeH265ReferenceInfo*    pStdReferenceInfo;
+} VkVideoEncodeH265DpbSlotInfoEXT;
+
+typedef struct VkVideoEncodeH265ProfileInfoEXT {
+    VkStructureType           sType;
+    const void*               pNext;
+    StdVideoH265ProfileIdc    stdProfileIdc;
+} VkVideoEncodeH265ProfileInfoEXT;
+
+typedef struct VkVideoEncodeH265RateControlInfoEXT {
+    VkStructureType                             sType;
+    const void*                                 pNext;
+    uint32_t                                    gopFrameCount;
+    uint32_t                                    idrPeriod;
+    uint32_t                                    consecutiveBFrameCount;
+    VkVideoEncodeH265RateControlStructureEXT    rateControlStructure;
+    uint32_t                                    subLayerCount;
+} VkVideoEncodeH265RateControlInfoEXT;
+
+typedef struct VkVideoEncodeH265QpEXT {
+    int32_t    qpI;
+    int32_t    qpP;
+    int32_t    qpB;
+} VkVideoEncodeH265QpEXT;
+
+typedef struct VkVideoEncodeH265FrameSizeEXT {
+    uint32_t    frameISize;
+    uint32_t    framePSize;
+    uint32_t    frameBSize;
+} VkVideoEncodeH265FrameSizeEXT;
+
+typedef struct VkVideoEncodeH265RateControlLayerInfoEXT {
+    VkStructureType                  sType;
+    const void*                      pNext;
+    uint32_t                         temporalId;
+    VkBool32                         useInitialRcQp;
+    VkVideoEncodeH265QpEXT           initialRcQp;
+    VkBool32                         useMinQp;
+    VkVideoEncodeH265QpEXT           minQp;
+    VkBool32                         useMaxQp;
+    VkVideoEncodeH265QpEXT           maxQp;
+    VkBool32                         useMaxFrameSize;
+    VkVideoEncodeH265FrameSizeEXT    maxFrameSize;
+} VkVideoEncodeH265RateControlLayerInfoEXT;
+
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
@@ -0,0 +1,54 @@
+#ifndef VULKAN_DIRECTFB_H_
+#define VULKAN_DIRECTFB_H_ 1
+
+/*
+** Copyright 2015-2023 The Khronos Group Inc.
+**
+** SPDX-License-Identifier: Apache-2.0
+*/
+
+/*
+** This header is generated from the Khronos Vulkan XML API Registry.
+**
+*/
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
+#define VK_EXT_directfb_surface 1
+#define VK_EXT_DIRECTFB_SURFACE_SPEC_VERSION 1
+#define VK_EXT_DIRECTFB_SURFACE_EXTENSION_NAME "VK_EXT_directfb_surface"
+typedef VkFlags VkDirectFBSurfaceCreateFlagsEXT;
+typedef struct VkDirectFBSurfaceCreateInfoEXT {
+    VkStructureType                    sType;
+    const void*                        pNext;
+    VkDirectFBSurfaceCreateFlagsEXT    flags;
+    IDirectFB*                         dfb;
+    IDirectFBSurface*                  surface;
+} VkDirectFBSurfaceCreateInfoEXT;
+
+typedef VkResult (VKAPI_PTR *PFN_vkCreateDirectFBSurfaceEXT)(VkInstance instance, const VkDirectFBSurfaceCreateInfoEXT* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);
+typedef VkBool32 (VKAPI_PTR *PFN_vkGetPhysicalDeviceDirectFBPresentationSupportEXT)(VkPhysicalDevice physicalDevice, uint32_t queueFamilyIndex, IDirectFB* dfb);
+
+#ifndef VK_NO_PROTOTYPES
+VKAPI_ATTR VkResult VKAPI_CALL vkCreateDirectFBSurfaceEXT(
+    VkInstance                                  instance,
+    const VkDirectFBSurfaceCreateInfoEXT*       pCreateInfo,
+    const VkAllocationCallbacks*                pAllocator,
+    VkSurfaceKHR*                               pSurface);
+
+VKAPI_ATTR VkBool32 VKAPI_CALL vkGetPhysicalDeviceDirectFBPresentationSupportEXT(
+    VkPhysicalDevice                            physicalDevice,
+    uint32_t                                    queueFamilyIndex,
+    IDirectFB*                                  dfb);
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
@@ -1,24 +1,10 @@
 #ifndef VULKAN_FUCHSIA_H_
 #define VULKAN_FUCHSIA_H_ 1

-#ifdef __cplusplus
-extern "C" {
-#endif
-
 /*
-** Copyright (c) 2015-2018 The Khronos Group Inc.
+** Copyright 2015-2023 The Khronos Group Inc.
 **
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
+** SPDX-License-Identifier: Apache-2.0
 */

 /*
@@ -27,12 +13,16 @@ extern "C" {
 */


+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
 #define VK_FUCHSIA_imagepipe_surface 1
 #define VK_FUCHSIA_IMAGEPIPE_SURFACE_SPEC_VERSION 1
 #define VK_FUCHSIA_IMAGEPIPE_SURFACE_EXTENSION_NAME "VK_FUCHSIA_imagepipe_surface"
-
 typedef VkFlags VkImagePipeSurfaceCreateFlagsFUCHSIA;
-
 typedef struct VkImagePipeSurfaceCreateInfoFUCHSIA {
    VkStructureType                         sType;
    const void*                             pNext;
@@ -40,7 +30,6 @@ typedef struct VkImagePipeSurfaceCreateInfoFUCHSIA {
    zx_handle_t                             imagePipeHandle;
 } VkImagePipeSurfaceCreateInfoFUCHSIA;

-
 typedef VkResult (VKAPI_PTR *PFN_vkCreateImagePipeSurfaceFUCHSIA)(VkInstance instance, const VkImagePipeSurfaceCreateInfoFUCHSIA* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);

 #ifndef VK_NO_PROTOTYPES
@@ -51,6 +40,217 @@ VKAPI_ATTR VkResult VKAPI_CALL vkCreateImagePipeSurfaceFUCHSIA(
    VkSurfaceKHR*                               pSurface);
 #endif

+
+#define VK_FUCHSIA_external_memory 1
+#define VK_FUCHSIA_EXTERNAL_MEMORY_SPEC_VERSION 1
+#define VK_FUCHSIA_EXTERNAL_MEMORY_EXTENSION_NAME "VK_FUCHSIA_external_memory"
+typedef struct VkImportMemoryZirconHandleInfoFUCHSIA {
+    VkStructureType                       sType;
+    const void*                           pNext;
+    VkExternalMemoryHandleTypeFlagBits    handleType;
+    zx_handle_t                           handle;
+} VkImportMemoryZirconHandleInfoFUCHSIA;
+
+typedef struct VkMemoryZirconHandlePropertiesFUCHSIA {
+    VkStructureType    sType;
+    void*              pNext;
+    uint32_t           memoryTypeBits;
+} VkMemoryZirconHandlePropertiesFUCHSIA;
+
+typedef struct VkMemoryGetZirconHandleInfoFUCHSIA {
+    VkStructureType                       sType;
+    const void*                           pNext;
+    VkDeviceMemory                        memory;
+    VkExternalMemoryHandleTypeFlagBits    handleType;
+} VkMemoryGetZirconHandleInfoFUCHSIA;
+
+typedef VkResult (VKAPI_PTR *PFN_vkGetMemoryZirconHandleFUCHSIA)(VkDevice device, const VkMemoryGetZirconHandleInfoFUCHSIA* pGetZirconHandleInfo, zx_handle_t* pZirconHandle);
+typedef VkResult (VKAPI_PTR *PFN_vkGetMemoryZirconHandlePropertiesFUCHSIA)(VkDevice device, VkExternalMemoryHandleTypeFlagBits handleType, zx_handle_t zirconHandle, VkMemoryZirconHandlePropertiesFUCHSIA* pMemoryZirconHandleProperties);
+
+#ifndef VK_NO_PROTOTYPES
+VKAPI_ATTR VkResult VKAPI_CALL vkGetMemoryZirconHandleFUCHSIA(
+    VkDevice                                    device,
+    const VkMemoryGetZirconHandleInfoFUCHSIA*   pGetZirconHandleInfo,
+    zx_handle_t*                                pZirconHandle);
+
+VKAPI_ATTR VkResult VKAPI_CALL vkGetMemoryZirconHandlePropertiesFUCHSIA(
+    VkDevice                                    device,
+    VkExternalMemoryHandleTypeFlagBits          handleType,
+    zx_handle_t                                 zirconHandle,
+    VkMemoryZirconHandlePropertiesFUCHSIA*      pMemoryZirconHandleProperties);
+#endif
+
+
+#define VK_FUCHSIA_external_semaphore 1
+#define VK_FUCHSIA_EXTERNAL_SEMAPHORE_SPEC_VERSION 1
+#define VK_FUCHSIA_EXTERNAL_SEMAPHORE_EXTENSION_NAME "VK_FUCHSIA_external_semaphore"
+typedef struct VkImportSemaphoreZirconHandleInfoFUCHSIA {
+    VkStructureType                          sType;
+    const void*                              pNext;
+    VkSemaphore                              semaphore;
+    VkSemaphoreImportFlags                   flags;
+    VkExternalSemaphoreHandleTypeFlagBits    handleType;
+    zx_handle_t                              zirconHandle;
+} VkImportSemaphoreZirconHandleInfoFUCHSIA;
+
+typedef struct VkSemaphoreGetZirconHandleInfoFUCHSIA {
+    VkStructureType                          sType;
+    const void*                              pNext;
+    VkSemaphore                              semaphore;
+    VkExternalSemaphoreHandleTypeFlagBits    handleType;
+} VkSemaphoreGetZirconHandleInfoFUCHSIA;
+
+typedef VkResult (VKAPI_PTR *PFN_vkImportSemaphoreZirconHandleFUCHSIA)(VkDevice device, const VkImportSemaphoreZirconHandleInfoFUCHSIA* pImportSemaphoreZirconHandleInfo);
+typedef VkResult (VKAPI_PTR *PFN_vkGetSemaphoreZirconHandleFUCHSIA)(VkDevice device, const VkSemaphoreGetZirconHandleInfoFUCHSIA* pGetZirconHandleInfo, zx_handle_t* pZirconHandle);
+
+#ifndef VK_NO_PROTOTYPES
+VKAPI_ATTR VkResult VKAPI_CALL vkImportSemaphoreZirconHandleFUCHSIA(
+    VkDevice                                    device,
+    const VkImportSemaphoreZirconHandleInfoFUCHSIA* pImportSemaphoreZirconHandleInfo);
+
+VKAPI_ATTR VkResult VKAPI_CALL vkGetSemaphoreZirconHandleFUCHSIA(
+    VkDevice                                    device,
+    const VkSemaphoreGetZirconHandleInfoFUCHSIA* pGetZirconHandleInfo,
+    zx_handle_t*                                pZirconHandle);
+#endif
+
+
+#define VK_FUCHSIA_buffer_collection 1
+VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkBufferCollectionFUCHSIA)
+#define VK_FUCHSIA_BUFFER_COLLECTION_SPEC_VERSION 2
+#define VK_FUCHSIA_BUFFER_COLLECTION_EXTENSION_NAME "VK_FUCHSIA_buffer_collection"
+typedef VkFlags VkImageFormatConstraintsFlagsFUCHSIA;
+
+typedef enum VkImageConstraintsInfoFlagBitsFUCHSIA {
+    VK_IMAGE_CONSTRAINTS_INFO_CPU_READ_RARELY_FUCHSIA = 0x00000001,
+    VK_IMAGE_CONSTRAINTS_INFO_CPU_READ_OFTEN_FUCHSIA = 0x00000002,
+    VK_IMAGE_CONSTRAINTS_INFO_CPU_WRITE_RARELY_FUCHSIA = 0x00000004,
+    VK_IMAGE_CONSTRAINTS_INFO_CPU_WRITE_OFTEN_FUCHSIA = 0x00000008,
+    VK_IMAGE_CONSTRAINTS_INFO_PROTECTED_OPTIONAL_FUCHSIA = 0x00000010,
+    VK_IMAGE_CONSTRAINTS_INFO_FLAG_BITS_MAX_ENUM_FUCHSIA = 0x7FFFFFFF
+} VkImageConstraintsInfoFlagBitsFUCHSIA;
+typedef VkFlags VkImageConstraintsInfoFlagsFUCHSIA;
+typedef struct VkBufferCollectionCreateInfoFUCHSIA {
+    VkStructureType    sType;
+    const void*        pNext;
+    zx_handle_t        collectionToken;
+} VkBufferCollectionCreateInfoFUCHSIA;
+
+typedef struct VkImportMemoryBufferCollectionFUCHSIA {
+    VkStructureType              sType;
+    const void*                  pNext;
+    VkBufferCollectionFUCHSIA    collection;
+    uint32_t                     index;
+} VkImportMemoryBufferCollectionFUCHSIA;
+
+typedef struct VkBufferCollectionImageCreateInfoFUCHSIA {
+    VkStructureType              sType;
+    const void*                  pNext;
+    VkBufferCollectionFUCHSIA    collection;
+    uint32_t                     index;
+} VkBufferCollectionImageCreateInfoFUCHSIA;
+
+typedef struct VkBufferCollectionConstraintsInfoFUCHSIA {
+    VkStructureType    sType;
+    const void*        pNext;
+    uint32_t           minBufferCount;
+    uint32_t           maxBufferCount;
+    uint32_t           minBufferCountForCamping;
+    uint32_t           minBufferCountForDedicatedSlack;
+    uint32_t           minBufferCountForSharedSlack;
+} VkBufferCollectionConstraintsInfoFUCHSIA;
+
+typedef struct VkBufferConstraintsInfoFUCHSIA {
+    VkStructureType                             sType;
+    const void*                                 pNext;
+    VkBufferCreateInfo                          createInfo;
+    VkFormatFeatureFlags                        requiredFormatFeatures;
+    VkBufferCollectionConstraintsInfoFUCHSIA    bufferCollectionConstraints;
+} VkBufferConstraintsInfoFUCHSIA;
+
+typedef struct VkBufferCollectionBufferCreateInfoFUCHSIA {
+    VkStructureType              sType;
+    const void*                  pNext;
+    VkBufferCollectionFUCHSIA    collection;
+    uint32_t                     index;
+} VkBufferCollectionBufferCreateInfoFUCHSIA;
+
+typedef struct VkSysmemColorSpaceFUCHSIA {
+    VkStructureType    sType;
+    const void*        pNext;
+    uint32_t           colorSpace;
+} VkSysmemColorSpaceFUCHSIA;
+
+typedef struct VkBufferCollectionPropertiesFUCHSIA {
+    VkStructureType                  sType;
+    void*                            pNext;
+    uint32_t                         memoryTypeBits;
+    uint32_t                         bufferCount;
+    uint32_t                         createInfoIndex;
+    uint64_t                         sysmemPixelFormat;
+    VkFormatFeatureFlags             formatFeatures;
+    VkSysmemColorSpaceFUCHSIA        sysmemColorSpaceIndex;
+    VkComponentMapping               samplerYcbcrConversionComponents;
+    VkSamplerYcbcrModelConversion    suggestedYcbcrModel;
+    VkSamplerYcbcrRange              suggestedYcbcrRange;
+    VkChromaLocation                 suggestedXChromaOffset;
+    VkChromaLocation                 suggestedYChromaOffset;
+} VkBufferCollectionPropertiesFUCHSIA;
+
+typedef struct VkImageFormatConstraintsInfoFUCHSIA {
+    VkStructureType                         sType;
+    const void*                             pNext;
+    VkImageCreateInfo                       imageCreateInfo;
+    VkFormatFeatureFlags                    requiredFormatFeatures;
+    VkImageFormatConstraintsFlagsFUCHSIA    flags;
+    uint64_t                                sysmemPixelFormat;
+    uint32_t                                colorSpaceCount;
+    const VkSysmemColorSpaceFUCHSIA*        pColorSpaces;
+} VkImageFormatConstraintsInfoFUCHSIA;
+
+typedef struct VkImageConstraintsInfoFUCHSIA {
+    VkStructureType                               sType;
+    const void*                                   pNext;
+    uint32_t                                      formatConstraintsCount;
+    const VkImageFormatConstraintsInfoFUCHSIA*    pFormatConstraints;
+    VkBufferCollectionConstraintsInfoFUCHSIA      bufferCollectionConstraints;
+    VkImageConstraintsInfoFlagsFUCHSIA            flags;
+} VkImageConstraintsInfoFUCHSIA;
+
+typedef VkResult (VKAPI_PTR *PFN_vkCreateBufferCollectionFUCHSIA)(VkDevice device, const VkBufferCollectionCreateInfoFUCHSIA* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkBufferCollectionFUCHSIA* pCollection);
+typedef VkResult (VKAPI_PTR *PFN_vkSetBufferCollectionImageConstraintsFUCHSIA)(VkDevice device, VkBufferCollectionFUCHSIA collection, const VkImageConstraintsInfoFUCHSIA* pImageConstraintsInfo);
+typedef VkResult (VKAPI_PTR *PFN_vkSetBufferCollectionBufferConstraintsFUCHSIA)(VkDevice device, VkBufferCollectionFUCHSIA collection, const VkBufferConstraintsInfoFUCHSIA* pBufferConstraintsInfo);
+typedef void (VKAPI_PTR *PFN_vkDestroyBufferCollectionFUCHSIA)(VkDevice device, VkBufferCollectionFUCHSIA collection, const VkAllocationCallbacks* pAllocator);
+typedef VkResult (VKAPI_PTR *PFN_vkGetBufferCollectionPropertiesFUCHSIA)(VkDevice device, VkBufferCollectionFUCHSIA collection, VkBufferCollectionPropertiesFUCHSIA* pProperties);
+
+#ifndef VK_NO_PROTOTYPES
+VKAPI_ATTR VkResult VKAPI_CALL vkCreateBufferCollectionFUCHSIA(
+    VkDevice                                    device,
+    const VkBufferCollectionCreateInfoFUCHSIA*  pCreateInfo,
+    const VkAllocationCallbacks*                pAllocator,
+    VkBufferCollectionFUCHSIA*                  pCollection);
+
+VKAPI_ATTR VkResult VKAPI_CALL vkSetBufferCollectionImageConstraintsFUCHSIA(
+    VkDevice                                    device,
+    VkBufferCollectionFUCHSIA                   collection,
+    const VkImageConstraintsInfoFUCHSIA*        pImageConstraintsInfo);
+
+VKAPI_ATTR VkResult VKAPI_CALL vkSetBufferCollectionBufferConstraintsFUCHSIA(
+    VkDevice                                    device,
+    VkBufferCollectionFUCHSIA                   collection,
+    const VkBufferConstraintsInfoFUCHSIA*       pBufferConstraintsInfo);
+
+VKAPI_ATTR void VKAPI_CALL vkDestroyBufferCollectionFUCHSIA(
+    VkDevice                                    device,
+    VkBufferCollectionFUCHSIA                   collection,
+    const VkAllocationCallbacks*                pAllocator);
+
+VKAPI_ATTR VkResult VKAPI_CALL vkGetBufferCollectionPropertiesFUCHSIA(
+    VkDevice                                    device,
+    VkBufferCollectionFUCHSIA                   collection,
+    VkBufferCollectionPropertiesFUCHSIA*        pProperties);
+#endif
+
 #ifdef __cplusplus
 }
 #endif
@@ -0,0 +1,58 @@
+#ifndef VULKAN_GGP_H_
+#define VULKAN_GGP_H_ 1
+
+/*
+** Copyright 2015-2023 The Khronos Group Inc.
+**
+** SPDX-License-Identifier: Apache-2.0
+*/
+
+/*
+** This header is generated from the Khronos Vulkan XML API Registry.
+**
+*/
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
+#define VK_GGP_stream_descriptor_surface 1
+#define VK_GGP_STREAM_DESCRIPTOR_SURFACE_SPEC_VERSION 1
+#define VK_GGP_STREAM_DESCRIPTOR_SURFACE_EXTENSION_NAME "VK_GGP_stream_descriptor_surface"
+typedef VkFlags VkStreamDescriptorSurfaceCreateFlagsGGP;
+typedef struct VkStreamDescriptorSurfaceCreateInfoGGP {
+    VkStructureType                            sType;
+    const void*                                pNext;
+    VkStreamDescriptorSurfaceCreateFlagsGGP    flags;
+    GgpStreamDescriptor                        streamDescriptor;
+} VkStreamDescriptorSurfaceCreateInfoGGP;
+
+typedef VkResult (VKAPI_PTR *PFN_vkCreateStreamDescriptorSurfaceGGP)(VkInstance instance, const VkStreamDescriptorSurfaceCreateInfoGGP* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);
+
+#ifndef VK_NO_PROTOTYPES
+VKAPI_ATTR VkResult VKAPI_CALL vkCreateStreamDescriptorSurfaceGGP(
+    VkInstance                                  instance,
+    const VkStreamDescriptorSurfaceCreateInfoGGP* pCreateInfo,
+    const VkAllocationCallbacks*                pAllocator,
+    VkSurfaceKHR*                               pSurface);
+#endif
+
+
+#define VK_GGP_frame_token 1
+#define VK_GGP_FRAME_TOKEN_SPEC_VERSION   1
+#define VK_GGP_FRAME_TOKEN_EXTENSION_NAME "VK_GGP_frame_token"
+typedef struct VkPresentFrameTokenGGP {
+    VkStructureType    sType;
+    const void*        pNext;
+    GgpFrameToken      frameToken;
+} VkPresentFrameTokenGGP;
+
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
@@ -1,24 +1,10 @@
 #ifndef VULKAN_IOS_H_
 #define VULKAN_IOS_H_ 1

-#ifdef __cplusplus
-extern "C" {
-#endif
-
 /*
-** Copyright (c) 2015-2018 The Khronos Group Inc.
+** Copyright 2015-2023 The Khronos Group Inc.
 **
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
+** SPDX-License-Identifier: Apache-2.0
 */

 /*
@@ -27,12 +13,16 @@ extern "C" {
 */


+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
 #define VK_MVK_ios_surface 1
-#define VK_MVK_IOS_SURFACE_SPEC_VERSION   2
+#define VK_MVK_IOS_SURFACE_SPEC_VERSION   3
 #define VK_MVK_IOS_SURFACE_EXTENSION_NAME "VK_MVK_ios_surface"
-
 typedef VkFlags VkIOSSurfaceCreateFlagsMVK;
-
 typedef struct VkIOSSurfaceCreateInfoMVK {
    VkStructureType               sType;
    const void*                   pNext;
@@ -40,7 +30,6 @@ typedef struct VkIOSSurfaceCreateInfoMVK {
    const void*                   pView;
 } VkIOSSurfaceCreateInfoMVK;

-
 typedef VkResult (VKAPI_PTR *PFN_vkCreateIOSSurfaceMVK)(VkInstance instance, const VkIOSSurfaceCreateInfoMVK* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);

 #ifndef VK_NO_PROTOTYPES
@@ -1,24 +1,10 @@
 #ifndef VULKAN_MACOS_H_
 #define VULKAN_MACOS_H_ 1

-#ifdef __cplusplus
-extern "C" {
-#endif
-
 /*
-** Copyright (c) 2015-2018 The Khronos Group Inc.
+** Copyright 2015-2023 The Khronos Group Inc.
 **
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
+** SPDX-License-Identifier: Apache-2.0
 */

 /*
@@ -27,12 +13,16 @@ extern "C" {
 */


+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
 #define VK_MVK_macos_surface 1
-#define VK_MVK_MACOS_SURFACE_SPEC_VERSION 2
+#define VK_MVK_MACOS_SURFACE_SPEC_VERSION 3
 #define VK_MVK_MACOS_SURFACE_EXTENSION_NAME "VK_MVK_macos_surface"
-
 typedef VkFlags VkMacOSSurfaceCreateFlagsMVK;
-
 typedef struct VkMacOSSurfaceCreateInfoMVK {
    VkStructureType                 sType;
    const void*                     pNext;
@@ -40,7 +30,6 @@ typedef struct VkMacOSSurfaceCreateInfoMVK {
    const void*                     pView;
 } VkMacOSSurfaceCreateInfoMVK;

-
 typedef VkResult (VKAPI_PTR *PFN_vkCreateMacOSSurfaceMVK)(VkInstance instance, const VkMacOSSurfaceCreateInfoMVK* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);

 #ifndef VK_NO_PROTOTYPES
@@ -0,0 +1,193 @@
+#ifndef VULKAN_METAL_H_
+#define VULKAN_METAL_H_ 1
+
+/*
+** Copyright 2015-2023 The Khronos Group Inc.
+**
+** SPDX-License-Identifier: Apache-2.0
+*/
+
+/*
+** This header is generated from the Khronos Vulkan XML API Registry.
+**
+*/
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
+#define VK_EXT_metal_surface 1
+#ifdef __OBJC__
+@class CAMetalLayer;
+#else
+typedef void CAMetalLayer;
+#endif
+
+#define VK_EXT_METAL_SURFACE_SPEC_VERSION 1
+#define VK_EXT_METAL_SURFACE_EXTENSION_NAME "VK_EXT_metal_surface"
+typedef VkFlags VkMetalSurfaceCreateFlagsEXT;
+typedef struct VkMetalSurfaceCreateInfoEXT {
+    VkStructureType                 sType;
+    const void*                     pNext;
+    VkMetalSurfaceCreateFlagsEXT    flags;
+    const CAMetalLayer*             pLayer;
+} VkMetalSurfaceCreateInfoEXT;
+
+typedef VkResult (VKAPI_PTR *PFN_vkCreateMetalSurfaceEXT)(VkInstance instance, const VkMetalSurfaceCreateInfoEXT* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);
+
+#ifndef VK_NO_PROTOTYPES
+VKAPI_ATTR VkResult VKAPI_CALL vkCreateMetalSurfaceEXT(
+    VkInstance                                  instance,
+    const VkMetalSurfaceCreateInfoEXT*          pCreateInfo,
+    const VkAllocationCallbacks*                pAllocator,
+    VkSurfaceKHR*                               pSurface);
+#endif
+
+
+#define VK_EXT_metal_objects 1
+#ifdef __OBJC__
+@protocol MTLDevice;
+typedef id<MTLDevice> MTLDevice_id;
+#else
+typedef void* MTLDevice_id;
+#endif
+
+#ifdef __OBJC__
+@protocol MTLCommandQueue;
+typedef id<MTLCommandQueue> MTLCommandQueue_id;
+#else
+typedef void* MTLCommandQueue_id;
+#endif
+
+#ifdef __OBJC__
+@protocol MTLBuffer;
+typedef id<MTLBuffer> MTLBuffer_id;
+#else
+typedef void* MTLBuffer_id;
+#endif
+
+#ifdef __OBJC__
+@protocol MTLTexture;
+typedef id<MTLTexture> MTLTexture_id;
+#else
+typedef void* MTLTexture_id;
+#endif
+
+typedef struct __IOSurface* IOSurfaceRef;
+#ifdef __OBJC__
+@protocol MTLSharedEvent;
+typedef id<MTLSharedEvent> MTLSharedEvent_id;
+#else
+typedef void* MTLSharedEvent_id;
+#endif
+
+#define VK_EXT_METAL_OBJECTS_SPEC_VERSION 1
+#define VK_EXT_METAL_OBJECTS_EXTENSION_NAME "VK_EXT_metal_objects"
+
+typedef enum VkExportMetalObjectTypeFlagBitsEXT {
+    VK_EXPORT_METAL_OBJECT_TYPE_METAL_DEVICE_BIT_EXT = 0x00000001,
+    VK_EXPORT_METAL_OBJECT_TYPE_METAL_COMMAND_QUEUE_BIT_EXT = 0x00000002,
+    VK_EXPORT_METAL_OBJECT_TYPE_METAL_BUFFER_BIT_EXT = 0x00000004,
+    VK_EXPORT_METAL_OBJECT_TYPE_METAL_TEXTURE_BIT_EXT = 0x00000008,
+    VK_EXPORT_METAL_OBJECT_TYPE_METAL_IOSURFACE_BIT_EXT = 0x00000010,
+    VK_EXPORT_METAL_OBJECT_TYPE_METAL_SHARED_EVENT_BIT_EXT = 0x00000020,
+    VK_EXPORT_METAL_OBJECT_TYPE_FLAG_BITS_MAX_ENUM_EXT = 0x7FFFFFFF
+} VkExportMetalObjectTypeFlagBitsEXT;
+typedef VkFlags VkExportMetalObjectTypeFlagsEXT;
+typedef struct VkExportMetalObjectCreateInfoEXT {
+    VkStructureType                       sType;
+    const void*                           pNext;
+    VkExportMetalObjectTypeFlagBitsEXT    exportObjectType;
+} VkExportMetalObjectCreateInfoEXT;
+
+typedef struct VkExportMetalObjectsInfoEXT {
+    VkStructureType    sType;
+    const void*        pNext;
+} VkExportMetalObjectsInfoEXT;
+
+typedef struct VkExportMetalDeviceInfoEXT {
+    VkStructureType    sType;
+    const void*        pNext;
+    MTLDevice_id       mtlDevice;
+} VkExportMetalDeviceInfoEXT;
+
+typedef struct VkExportMetalCommandQueueInfoEXT {
+    VkStructureType       sType;
+    const void*           pNext;
+    VkQueue               queue;
+    MTLCommandQueue_id    mtlCommandQueue;
+} VkExportMetalCommandQueueInfoEXT;
+
+typedef struct VkExportMetalBufferInfoEXT {
+    VkStructureType    sType;
+    const void*        pNext;
+    VkDeviceMemory     memory;
+    MTLBuffer_id       mtlBuffer;
+} VkExportMetalBufferInfoEXT;
+
+typedef struct VkImportMetalBufferInfoEXT {
+    VkStructureType    sType;
+    const void*        pNext;
+    MTLBuffer_id       mtlBuffer;
+} VkImportMetalBufferInfoEXT;
+
+typedef struct VkExportMetalTextureInfoEXT {
+    VkStructureType          sType;
+    const void*              pNext;
+    VkImage                  image;
+    VkImageView              imageView;
+    VkBufferView             bufferView;
+    VkImageAspectFlagBits    plane;
+    MTLTexture_id            mtlTexture;
+} VkExportMetalTextureInfoEXT;
+
+typedef struct VkImportMetalTextureInfoEXT {
+    VkStructureType          sType;
+    const void*              pNext;
+    VkImageAspectFlagBits    plane;
+    MTLTexture_id            mtlTexture;
+} VkImportMetalTextureInfoEXT;
+
+typedef struct VkExportMetalIOSurfaceInfoEXT {
+    VkStructureType    sType;
+    const void*        pNext;
+    VkImage            image;
+    IOSurfaceRef       ioSurface;
+} VkExportMetalIOSurfaceInfoEXT;
+
+typedef struct VkImportMetalIOSurfaceInfoEXT {
+    VkStructureType    sType;
+    const void*        pNext;
+    IOSurfaceRef       ioSurface;
+} VkImportMetalIOSurfaceInfoEXT;
+
+typedef struct VkExportMetalSharedEventInfoEXT {
+    VkStructureType      sType;
+    const void*          pNext;
+    VkSemaphore          semaphore;
+    VkEvent              event;
+    MTLSharedEvent_id    mtlSharedEvent;
+} VkExportMetalSharedEventInfoEXT;
+
+typedef struct VkImportMetalSharedEventInfoEXT {
+    VkStructureType      sType;
+    const void*          pNext;
+    MTLSharedEvent_id    mtlSharedEvent;
+} VkImportMetalSharedEventInfoEXT;
+
+typedef void (VKAPI_PTR *PFN_vkExportMetalObjectsEXT)(VkDevice device, VkExportMetalObjectsInfoEXT* pMetalObjectsInfo);
+
+#ifndef VK_NO_PROTOTYPES
+VKAPI_ATTR void VKAPI_CALL vkExportMetalObjectsEXT(
+    VkDevice                                    device,
+    VkExportMetalObjectsInfoEXT*                pMetalObjectsInfo);
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
@@ -1,65 +0,0 @@
-#ifndef VULKAN_MIR_H_
-#define VULKAN_MIR_H_ 1
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/*
-** Copyright (c) 2015-2018 The Khronos Group Inc.
-**
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
-*/
-
-/*
-** This header is generated from the Khronos Vulkan XML API Registry.
-**
-*/
-
-
-#define VK_KHR_mir_surface 1
-#define VK_KHR_MIR_SURFACE_SPEC_VERSION   4
-#define VK_KHR_MIR_SURFACE_EXTENSION_NAME "VK_KHR_mir_surface"
-
-typedef VkFlags VkMirSurfaceCreateFlagsKHR;
-
-typedef struct VkMirSurfaceCreateInfoKHR {
-    VkStructureType               sType;
-    const void*                   pNext;
-    VkMirSurfaceCreateFlagsKHR    flags;
-    MirConnection*                connection;
-    MirSurface*                   mirSurface;
-} VkMirSurfaceCreateInfoKHR;
-
-
-typedef VkResult (VKAPI_PTR *PFN_vkCreateMirSurfaceKHR)(VkInstance instance, const VkMirSurfaceCreateInfoKHR* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);
-typedef VkBool32 (VKAPI_PTR *PFN_vkGetPhysicalDeviceMirPresentationSupportKHR)(VkPhysicalDevice physicalDevice, uint32_t queueFamilyIndex, MirConnection* connection);
-
-#ifndef VK_NO_PROTOTYPES
-VKAPI_ATTR VkResult VKAPI_CALL vkCreateMirSurfaceKHR(
-    VkInstance                                  instance,
-    const VkMirSurfaceCreateInfoKHR*            pCreateInfo,
-    const VkAllocationCallbacks*                pAllocator,
-    VkSurfaceKHR*                               pSurface);
-
-VKAPI_ATTR VkBool32 VKAPI_CALL vkGetPhysicalDeviceMirPresentationSupportKHR(
-    VkPhysicalDevice                            physicalDevice,
-    uint32_t                                    queueFamilyIndex,
-    MirConnection*                              connection);
-#endif
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif
@@ -0,0 +1,54 @@
+#ifndef VULKAN_SCREEN_H_
+#define VULKAN_SCREEN_H_ 1
+
+/*
+** Copyright 2015-2023 The Khronos Group Inc.
+**
+** SPDX-License-Identifier: Apache-2.0
+*/
+
+/*
+** This header is generated from the Khronos Vulkan XML API Registry.
+**
+*/
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
+#define VK_QNX_screen_surface 1
+#define VK_QNX_SCREEN_SURFACE_SPEC_VERSION 1
+#define VK_QNX_SCREEN_SURFACE_EXTENSION_NAME "VK_QNX_screen_surface"
+typedef VkFlags VkScreenSurfaceCreateFlagsQNX;
+typedef struct VkScreenSurfaceCreateInfoQNX {
+    VkStructureType                  sType;
+    const void*                      pNext;
+    VkScreenSurfaceCreateFlagsQNX    flags;
+    struct _screen_context*          context;
+    struct _screen_window*           window;
+} VkScreenSurfaceCreateInfoQNX;
+
+typedef VkResult (VKAPI_PTR *PFN_vkCreateScreenSurfaceQNX)(VkInstance instance, const VkScreenSurfaceCreateInfoQNX* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);
+typedef VkBool32 (VKAPI_PTR *PFN_vkGetPhysicalDeviceScreenPresentationSupportQNX)(VkPhysicalDevice physicalDevice, uint32_t queueFamilyIndex, struct _screen_window* window);
+
+#ifndef VK_NO_PROTOTYPES
+VKAPI_ATTR VkResult VKAPI_CALL vkCreateScreenSurfaceQNX(
+    VkInstance                                  instance,
+    const VkScreenSurfaceCreateInfoQNX*         pCreateInfo,
+    const VkAllocationCallbacks*                pAllocator,
+    VkSurfaceKHR*                               pSurface);
+
+VKAPI_ATTR VkBool32 VKAPI_CALL vkGetPhysicalDeviceScreenPresentationSupportQNX(
+    VkPhysicalDevice                            physicalDevice,
+    uint32_t                                    queueFamilyIndex,
+    struct _screen_window*                      window);
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
@@ -1,24 +1,10 @@
 #ifndef VULKAN_VI_H_
 #define VULKAN_VI_H_ 1

-#ifdef __cplusplus
-extern "C" {
-#endif
-
 /*
-** Copyright (c) 2015-2018 The Khronos Group Inc.
+** Copyright 2015-2023 The Khronos Group Inc.
 **
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
+** SPDX-License-Identifier: Apache-2.0
 */

 /*
@@ -27,12 +13,16 @@ extern "C" {
 */


+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
 #define VK_NN_vi_surface 1
 #define VK_NN_VI_SURFACE_SPEC_VERSION     1
 #define VK_NN_VI_SURFACE_EXTENSION_NAME   "VK_NN_vi_surface"
-
 typedef VkFlags VkViSurfaceCreateFlagsNN;
-
 typedef struct VkViSurfaceCreateInfoNN {
    VkStructureType             sType;
    const void*                 pNext;
@@ -40,7 +30,6 @@ typedef struct VkViSurfaceCreateInfoNN {
    void*                       window;
 } VkViSurfaceCreateInfoNN;

-
 typedef VkResult (VKAPI_PTR *PFN_vkCreateViSurfaceNN)(VkInstance instance, const VkViSurfaceCreateInfoNN* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);

 #ifndef VK_NO_PROTOTYPES
@@ -1,24 +1,10 @@
 #ifndef VULKAN_WAYLAND_H_
 #define VULKAN_WAYLAND_H_ 1

-#ifdef __cplusplus
-extern "C" {
-#endif
-
 /*
-** Copyright (c) 2015-2018 The Khronos Group Inc.
+** Copyright 2015-2023 The Khronos Group Inc.
 **
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
+** SPDX-License-Identifier: Apache-2.0
 */

 /*
@@ -27,12 +13,16 @@ extern "C" {
 */


+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
 #define VK_KHR_wayland_surface 1
 #define VK_KHR_WAYLAND_SURFACE_SPEC_VERSION 6
 #define VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME "VK_KHR_wayland_surface"
-
 typedef VkFlags VkWaylandSurfaceCreateFlagsKHR;
-
 typedef struct VkWaylandSurfaceCreateInfoKHR {
    VkStructureType                   sType;
    const void*                       pNext;
@@ -41,7 +31,6 @@ typedef struct VkWaylandSurfaceCreateInfoKHR {
    struct wl_surface*                surface;
 } VkWaylandSurfaceCreateInfoKHR;

-
 typedef VkResult (VKAPI_PTR *PFN_vkCreateWaylandSurfaceKHR)(VkInstance instance, const VkWaylandSurfaceCreateInfoKHR* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);
 typedef VkBool32 (VKAPI_PTR *PFN_vkGetPhysicalDeviceWaylandPresentationSupportKHR)(VkPhysicalDevice physicalDevice, uint32_t queueFamilyIndex, struct wl_display* display);

@@ -1,24 +1,10 @@
 #ifndef VULKAN_WIN32_H_
 #define VULKAN_WIN32_H_ 1

-#ifdef __cplusplus
-extern "C" {
-#endif
-
 /*
-** Copyright (c) 2015-2018 The Khronos Group Inc.
+** Copyright 2015-2023 The Khronos Group Inc.
 **
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
+** SPDX-License-Identifier: Apache-2.0
 */

 /*
@@ -27,12 +13,16 @@ extern "C" {
 */


+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
 #define VK_KHR_win32_surface 1
 #define VK_KHR_WIN32_SURFACE_SPEC_VERSION 6
 #define VK_KHR_WIN32_SURFACE_EXTENSION_NAME "VK_KHR_win32_surface"
-
 typedef VkFlags VkWin32SurfaceCreateFlagsKHR;
-
 typedef struct VkWin32SurfaceCreateInfoKHR {
    VkStructureType                 sType;
    const void*                     pNext;
@@ -41,7 +31,6 @@ typedef struct VkWin32SurfaceCreateInfoKHR {
    HWND                            hwnd;
 } VkWin32SurfaceCreateInfoKHR;

-
 typedef VkResult (VKAPI_PTR *PFN_vkCreateWin32SurfaceKHR)(VkInstance instance, const VkWin32SurfaceCreateInfoKHR* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);
 typedef VkBool32 (VKAPI_PTR *PFN_vkGetPhysicalDeviceWin32PresentationSupportKHR)(VkPhysicalDevice physicalDevice, uint32_t queueFamilyIndex);

@@ -57,10 +46,10 @@ VKAPI_ATTR VkBool32 VKAPI_CALL vkGetPhysicalDeviceWin32PresentationSupportKHR(
    uint32_t                                    queueFamilyIndex);
 #endif

+
 #define VK_KHR_external_memory_win32 1
 #define VK_KHR_EXTERNAL_MEMORY_WIN32_SPEC_VERSION 1
 #define VK_KHR_EXTERNAL_MEMORY_WIN32_EXTENSION_NAME "VK_KHR_external_memory_win32"
-
 typedef struct VkImportMemoryWin32HandleInfoKHR {
    VkStructureType                       sType;
    const void*                           pNext;
@@ -90,7 +79,6 @@ typedef struct VkMemoryGetWin32HandleInfoKHR {
    VkExternalMemoryHandleTypeFlagBits    handleType;
 } VkMemoryGetWin32HandleInfoKHR;

-
 typedef VkResult (VKAPI_PTR *PFN_vkGetMemoryWin32HandleKHR)(VkDevice device, const VkMemoryGetWin32HandleInfoKHR* pGetWin32HandleInfo, HANDLE* pHandle);
 typedef VkResult (VKAPI_PTR *PFN_vkGetMemoryWin32HandlePropertiesKHR)(VkDevice device, VkExternalMemoryHandleTypeFlagBits handleType, HANDLE handle, VkMemoryWin32HandlePropertiesKHR* pMemoryWin32HandleProperties);

@@ -107,10 +95,10 @@ VKAPI_ATTR VkResult VKAPI_CALL vkGetMemoryWin32HandlePropertiesKHR(
    VkMemoryWin32HandlePropertiesKHR*           pMemoryWin32HandleProperties);
 #endif

+
 #define VK_KHR_win32_keyed_mutex 1
 #define VK_KHR_WIN32_KEYED_MUTEX_SPEC_VERSION 1
 #define VK_KHR_WIN32_KEYED_MUTEX_EXTENSION_NAME "VK_KHR_win32_keyed_mutex"
-
 typedef struct VkWin32KeyedMutexAcquireReleaseInfoKHR {
    VkStructureType          sType;
    const void*              pNext;
@@ -128,7 +116,6 @@ typedef struct VkWin32KeyedMutexAcquireReleaseInfoKHR {
 #define VK_KHR_external_semaphore_win32 1
 #define VK_KHR_EXTERNAL_SEMAPHORE_WIN32_SPEC_VERSION 1
 #define VK_KHR_EXTERNAL_SEMAPHORE_WIN32_EXTENSION_NAME "VK_KHR_external_semaphore_win32"
-
 typedef struct VkImportSemaphoreWin32HandleInfoKHR {
    VkStructureType                          sType;
    const void*                              pNext;
@@ -163,7 +150,6 @@ typedef struct VkSemaphoreGetWin32HandleInfoKHR {
    VkExternalSemaphoreHandleTypeFlagBits    handleType;
 } VkSemaphoreGetWin32HandleInfoKHR;

-
 typedef VkResult (VKAPI_PTR *PFN_vkImportSemaphoreWin32HandleKHR)(VkDevice device, const VkImportSemaphoreWin32HandleInfoKHR* pImportSemaphoreWin32HandleInfo);
 typedef VkResult (VKAPI_PTR *PFN_vkGetSemaphoreWin32HandleKHR)(VkDevice device, const VkSemaphoreGetWin32HandleInfoKHR* pGetWin32HandleInfo, HANDLE* pHandle);

@@ -178,10 +164,10 @@ VKAPI_ATTR VkResult VKAPI_CALL vkGetSemaphoreWin32HandleKHR(
    HANDLE*                                     pHandle);
 #endif

+
 #define VK_KHR_external_fence_win32 1
 #define VK_KHR_EXTERNAL_FENCE_WIN32_SPEC_VERSION 1
 #define VK_KHR_EXTERNAL_FENCE_WIN32_EXTENSION_NAME "VK_KHR_external_fence_win32"
-
 typedef struct VkImportFenceWin32HandleInfoKHR {
    VkStructureType                      sType;
    const void*                          pNext;
@@ -207,7 +193,6 @@ typedef struct VkFenceGetWin32HandleInfoKHR {
    VkExternalFenceHandleTypeFlagBits    handleType;
 } VkFenceGetWin32HandleInfoKHR;

-
 typedef VkResult (VKAPI_PTR *PFN_vkImportFenceWin32HandleKHR)(VkDevice device, const VkImportFenceWin32HandleInfoKHR* pImportFenceWin32HandleInfo);
 typedef VkResult (VKAPI_PTR *PFN_vkGetFenceWin32HandleKHR)(VkDevice device, const VkFenceGetWin32HandleInfoKHR* pGetWin32HandleInfo, HANDLE* pHandle);

@@ -222,10 +207,10 @@ VKAPI_ATTR VkResult VKAPI_CALL vkGetFenceWin32HandleKHR(
    HANDLE*                                     pHandle);
 #endif

+
 #define VK_NV_external_memory_win32 1
 #define VK_NV_EXTERNAL_MEMORY_WIN32_SPEC_VERSION 1
 #define VK_NV_EXTERNAL_MEMORY_WIN32_EXTENSION_NAME "VK_NV_external_memory_win32"
-
 typedef struct VkImportMemoryWin32HandleInfoNV {
    VkStructureType                      sType;
    const void*                          pNext;
@@ -240,7 +225,6 @@ typedef struct VkExportMemoryWin32HandleInfoNV {
    DWORD                         dwAccess;
 } VkExportMemoryWin32HandleInfoNV;

-
 typedef VkResult (VKAPI_PTR *PFN_vkGetMemoryWin32HandleNV)(VkDevice device, VkDeviceMemory memory, VkExternalMemoryHandleTypeFlagsNV handleType, HANDLE* pHandle);

 #ifndef VK_NO_PROTOTYPES
@@ -251,10 +235,10 @@ VKAPI_ATTR VkResult VKAPI_CALL vkGetMemoryWin32HandleNV(
    HANDLE*                                     pHandle);
 #endif

-#define VK_NV_win32_keyed_mutex 1
-#define VK_NV_WIN32_KEYED_MUTEX_SPEC_VERSION 1
-#define VK_NV_WIN32_KEYED_MUTEX_EXTENSION_NAME "VK_NV_win32_keyed_mutex"

+#define VK_NV_win32_keyed_mutex 1
+#define VK_NV_WIN32_KEYED_MUTEX_SPEC_VERSION 2
+#define VK_NV_WIN32_KEYED_MUTEX_EXTENSION_NAME "VK_NV_win32_keyed_mutex"
 typedef struct VkWin32KeyedMutexAcquireReleaseInfoNV {
    VkStructureType          sType;
    const void*              pNext;
@@ -269,6 +253,79 @@ typedef struct VkWin32KeyedMutexAcquireReleaseInfoNV {



+#define VK_EXT_full_screen_exclusive 1
+#define VK_EXT_FULL_SCREEN_EXCLUSIVE_SPEC_VERSION 4
+#define VK_EXT_FULL_SCREEN_EXCLUSIVE_EXTENSION_NAME "VK_EXT_full_screen_exclusive"
+
+typedef enum VkFullScreenExclusiveEXT {
+    VK_FULL_SCREEN_EXCLUSIVE_DEFAULT_EXT = 0,
+    VK_FULL_SCREEN_EXCLUSIVE_ALLOWED_EXT = 1,
+    VK_FULL_SCREEN_EXCLUSIVE_DISALLOWED_EXT = 2,
+    VK_FULL_SCREEN_EXCLUSIVE_APPLICATION_CONTROLLED_EXT = 3,
+    VK_FULL_SCREEN_EXCLUSIVE_MAX_ENUM_EXT = 0x7FFFFFFF
+} VkFullScreenExclusiveEXT;
+typedef struct VkSurfaceFullScreenExclusiveInfoEXT {
+    VkStructureType             sType;
+    void*                       pNext;
+    VkFullScreenExclusiveEXT    fullScreenExclusive;
+} VkSurfaceFullScreenExclusiveInfoEXT;
+
+typedef struct VkSurfaceCapabilitiesFullScreenExclusiveEXT {
+    VkStructureType    sType;
+    void*              pNext;
+    VkBool32           fullScreenExclusiveSupported;
+} VkSurfaceCapabilitiesFullScreenExclusiveEXT;
+
+typedef struct VkSurfaceFullScreenExclusiveWin32InfoEXT {
+    VkStructureType    sType;
+    const void*        pNext;
+    HMONITOR           hmonitor;
+} VkSurfaceFullScreenExclusiveWin32InfoEXT;
+
+typedef VkResult (VKAPI_PTR *PFN_vkGetPhysicalDeviceSurfacePresentModes2EXT)(VkPhysicalDevice physicalDevice, const VkPhysicalDeviceSurfaceInfo2KHR* pSurfaceInfo, uint32_t* pPresentModeCount, VkPresentModeKHR* pPresentModes);
+typedef VkResult (VKAPI_PTR *PFN_vkAcquireFullScreenExclusiveModeEXT)(VkDevice device, VkSwapchainKHR swapchain);
+typedef VkResult (VKAPI_PTR *PFN_vkReleaseFullScreenExclusiveModeEXT)(VkDevice device, VkSwapchainKHR swapchain);
+typedef VkResult (VKAPI_PTR *PFN_vkGetDeviceGroupSurfacePresentModes2EXT)(VkDevice device, const VkPhysicalDeviceSurfaceInfo2KHR* pSurfaceInfo, VkDeviceGroupPresentModeFlagsKHR* pModes);
+
+#ifndef VK_NO_PROTOTYPES
+VKAPI_ATTR VkResult VKAPI_CALL vkGetPhysicalDeviceSurfacePresentModes2EXT(
+    VkPhysicalDevice                            physicalDevice,
+    const VkPhysicalDeviceSurfaceInfo2KHR*      pSurfaceInfo,
+    uint32_t*                                   pPresentModeCount,
+    VkPresentModeKHR*                           pPresentModes);
+
+VKAPI_ATTR VkResult VKAPI_CALL vkAcquireFullScreenExclusiveModeEXT(
+    VkDevice                                    device,
+    VkSwapchainKHR                              swapchain);
+
+VKAPI_ATTR VkResult VKAPI_CALL vkReleaseFullScreenExclusiveModeEXT(
+    VkDevice                                    device,
+    VkSwapchainKHR                              swapchain);
+
+VKAPI_ATTR VkResult VKAPI_CALL vkGetDeviceGroupSurfacePresentModes2EXT(
+    VkDevice                                    device,
+    const VkPhysicalDeviceSurfaceInfo2KHR*      pSurfaceInfo,
+    VkDeviceGroupPresentModeFlagsKHR*           pModes);
+#endif
+
+
+#define VK_NV_acquire_winrt_display 1
+#define VK_NV_ACQUIRE_WINRT_DISPLAY_SPEC_VERSION 1
+#define VK_NV_ACQUIRE_WINRT_DISPLAY_EXTENSION_NAME "VK_NV_acquire_winrt_display"
+typedef VkResult (VKAPI_PTR *PFN_vkAcquireWinrtDisplayNV)(VkPhysicalDevice physicalDevice, VkDisplayKHR display);
+typedef VkResult (VKAPI_PTR *PFN_vkGetWinrtDisplayNV)(VkPhysicalDevice physicalDevice, uint32_t deviceRelativeId, VkDisplayKHR* pDisplay);
+
+#ifndef VK_NO_PROTOTYPES
+VKAPI_ATTR VkResult VKAPI_CALL vkAcquireWinrtDisplayNV(
+    VkPhysicalDevice                            physicalDevice,
+    VkDisplayKHR                                display);
+
+VKAPI_ATTR VkResult VKAPI_CALL vkGetWinrtDisplayNV(
+    VkPhysicalDevice                            physicalDevice,
+    uint32_t                                    deviceRelativeId,
+    VkDisplayKHR*                               pDisplay);
+#endif
+
 #ifdef __cplusplus
 }
 #endif
@@ -1,24 +1,10 @@
 #ifndef VULKAN_XCB_H_
 #define VULKAN_XCB_H_ 1

-#ifdef __cplusplus
-extern "C" {
-#endif
-
 /*
-** Copyright (c) 2015-2018 The Khronos Group Inc.
+** Copyright 2015-2023 The Khronos Group Inc.
 **
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
+** SPDX-License-Identifier: Apache-2.0
 */

 /*
@@ -27,12 +13,16 @@ extern "C" {
 */


+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
 #define VK_KHR_xcb_surface 1
 #define VK_KHR_XCB_SURFACE_SPEC_VERSION   6
 #define VK_KHR_XCB_SURFACE_EXTENSION_NAME "VK_KHR_xcb_surface"
-
 typedef VkFlags VkXcbSurfaceCreateFlagsKHR;
-
 typedef struct VkXcbSurfaceCreateInfoKHR {
    VkStructureType               sType;
    const void*                   pNext;
@@ -41,7 +31,6 @@ typedef struct VkXcbSurfaceCreateInfoKHR {
    xcb_window_t                  window;
 } VkXcbSurfaceCreateInfoKHR;

-
 typedef VkResult (VKAPI_PTR *PFN_vkCreateXcbSurfaceKHR)(VkInstance instance, const VkXcbSurfaceCreateInfoKHR* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);
 typedef VkBool32 (VKAPI_PTR *PFN_vkGetPhysicalDeviceXcbPresentationSupportKHR)(VkPhysicalDevice physicalDevice, uint32_t queueFamilyIndex, xcb_connection_t* connection, xcb_visualid_t visual_id);

@@ -1,24 +1,10 @@
 #ifndef VULKAN_XLIB_H_
 #define VULKAN_XLIB_H_ 1

-#ifdef __cplusplus
-extern "C" {
-#endif
-
 /*
-** Copyright (c) 2015-2018 The Khronos Group Inc.
+** Copyright 2015-2023 The Khronos Group Inc.
 **
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
+** SPDX-License-Identifier: Apache-2.0
 */

 /*
@@ -27,12 +13,16 @@ extern "C" {
 */


+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
 #define VK_KHR_xlib_surface 1
 #define VK_KHR_XLIB_SURFACE_SPEC_VERSION  6
 #define VK_KHR_XLIB_SURFACE_EXTENSION_NAME "VK_KHR_xlib_surface"
-
 typedef VkFlags VkXlibSurfaceCreateFlagsKHR;
-
 typedef struct VkXlibSurfaceCreateInfoKHR {
    VkStructureType                sType;
    const void*                    pNext;
@@ -41,7 +31,6 @@ typedef struct VkXlibSurfaceCreateInfoKHR {
    Window                         window;
 } VkXlibSurfaceCreateInfoKHR;

-
 typedef VkResult (VKAPI_PTR *PFN_vkCreateXlibSurfaceKHR)(VkInstance instance, const VkXlibSurfaceCreateInfoKHR* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);
 typedef VkBool32 (VKAPI_PTR *PFN_vkGetPhysicalDeviceXlibPresentationSupportKHR)(VkPhysicalDevice physicalDevice, uint32_t queueFamilyIndex, Display* dpy, VisualID visualID);

@@ -1,24 +1,10 @@
 #ifndef VULKAN_XLIB_XRANDR_H_
 #define VULKAN_XLIB_XRANDR_H_ 1

-#ifdef __cplusplus
-extern "C" {
-#endif
-
 /*
-** Copyright (c) 2015-2018 The Khronos Group Inc.
+** Copyright 2015-2023 The Khronos Group Inc.
 **
-** Licensed under the Apache License, Version 2.0 (the "License");
-** you may not use this file except in compliance with the License.
-** You may obtain a copy of the License at
-**
-**     http://www.apache.org/licenses/LICENSE-2.0
-**
-** Unless required by applicable law or agreed to in writing, software
-** distributed under the License is distributed on an "AS IS" BASIS,
-** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-** See the License for the specific language governing permissions and
-** limitations under the License.
+** SPDX-License-Identifier: Apache-2.0
 */

 /*
@@ -27,10 +13,15 @@ extern "C" {
 */


+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
 #define VK_EXT_acquire_xlib_display 1
 #define VK_EXT_ACQUIRE_XLIB_DISPLAY_SPEC_VERSION 1
 #define VK_EXT_ACQUIRE_XLIB_DISPLAY_EXTENSION_NAME "VK_EXT_acquire_xlib_display"
-
 typedef VkResult (VKAPI_PTR *PFN_vkAcquireXlibDisplayEXT)(VkPhysicalDevice physicalDevice, Display* dpy, VkDisplayKHR display);
 typedef VkResult (VKAPI_PTR *PFN_vkGetRandROutputDisplayEXT)(VkPhysicalDevice physicalDevice, Display* dpy, RROutput rrOutput, VkDisplayKHR* pDisplay);

@@ -24,7 +24,7 @@ if(UNIX)
    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Wno-unused-function -Wno-missing-braces -Wno-missing-field-initializers")
  endif()
  if(CV_CLANG)
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Wno-self-assign")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Wno-self-assign -Wno-strict-prototypes")
  endif()
 endif()

@@ -2,32 +2,32 @@ function(download_ippicv root_var)
  set(${root_var} "" PARENT_SCOPE)

  # Commit SHA in the opencv_3rdparty repo
-  set(IPPICV_COMMIT "a56b6ac6f030c312b2dce17430eef13aed9af274")
+  set(IPPICV_COMMIT "1224f78da6684df04397ac0f40c961ed37f79ccb")
  # Define actual ICV versions
  if(APPLE)
    set(OPENCV_ICV_PLATFORM "macosx")
    set(OPENCV_ICV_PACKAGE_SUBDIR "ippicv_mac")
-    set(OPENCV_ICV_NAME "ippicv_2020_mac_intel64_20191018_general.tgz")
-    set(OPENCV_ICV_HASH "1c3d675c2a2395d094d523024896e01b")
+    set(OPENCV_ICV_NAME "ippicv_2021.8_mac_intel64_20230330_general.tgz")
+    set(OPENCV_ICV_HASH "d2b234a86af1b616958619a4560356d9")
  elseif((UNIX AND NOT ANDROID) OR (UNIX AND ANDROID_ABI MATCHES "x86"))
    set(OPENCV_ICV_PLATFORM "linux")
    set(OPENCV_ICV_PACKAGE_SUBDIR "ippicv_lnx")
    if(X86_64)
-      set(OPENCV_ICV_NAME "ippicv_2020_lnx_intel64_20191018_general.tgz")
-      set(OPENCV_ICV_HASH "7421de0095c7a39162ae13a6098782f9")
+      set(OPENCV_ICV_NAME "ippicv_2021.8_lnx_intel64_20230330_general.tgz")
+      set(OPENCV_ICV_HASH "43219bdc7e3805adcbe3a1e2f1f3ef3b")
    else()
-      set(OPENCV_ICV_NAME "ippicv_2020_lnx_ia32_20191018_general.tgz")
-      set(OPENCV_ICV_HASH "ad189a940fb60eb71f291321322fe3e8")
+      set(OPENCV_ICV_NAME "ippicv_2021.8_lnx_ia32_20230330_general.tgz")
+      set(OPENCV_ICV_HASH "165875443d72faa3fd2146869da90d07")
    endif()
  elseif(WIN32 AND NOT ARM)
    set(OPENCV_ICV_PLATFORM "windows")
    set(OPENCV_ICV_PACKAGE_SUBDIR "ippicv_win")
    if(X86_64)
-      set(OPENCV_ICV_NAME "ippicv_2020_win_intel64_20191018_general.zip")
-      set(OPENCV_ICV_HASH "879741a7946b814455eee6c6ffde2984")
+      set(OPENCV_ICV_NAME "ippicv_2021.8_win_intel64_20230330_general.zip")
+      set(OPENCV_ICV_HASH "71e4f58de939f0348ec7fb58ffb17dbf")
    else()
-      set(OPENCV_ICV_NAME "ippicv_2020_win_ia32_20191018_general.zip")
-      set(OPENCV_ICV_HASH "cd39bdf0c2e1cac9a61101dad7a2413e")
+      set(OPENCV_ICV_NAME "ippicv_2021.8_win_ia32_20230330_general.zip")
+      set(OPENCV_ICV_HASH "57fd4648cfe64eae9e2ad9d50173a553")
    endif()
  else()
    return()
@@ -9,8 +9,8 @@ if(ANDROID)
  ocv_include_directories(${CPUFEATURES_INCLUDE_DIRS})
 endif()

-file(GLOB lib_srcs src/dec/*.c src/demux/*.c src/dsp/*.c src/enc/*.c src/mux/*.c src/utils/*.c src/webp/*.c)
-file(GLOB lib_hdrs src/dec/*.h src/demux/*.h src/dsp/*.h src/enc/*.h src/mux/*.h src/utils/*.h src/webp/*.h)
+file(GLOB lib_srcs sharpyuv/*.c src/dec/*.c src/demux/*.c src/dsp/*.c src/enc/*.c src/mux/*.c src/utils/*.c src/webp/*.c)
+file(GLOB lib_hdrs sharpyuv/*.h src/dec/*.h src/demux/*.h src/dsp/*.h src/enc/*.h src/mux/*.h src/utils/*.h src/webp/*.h)

 # FIXIT
 if(ANDROID AND ARMEABI_V7A AND NOT NEON)
@@ -21,12 +21,6 @@ if(ANDROID AND ARMEABI_V7A AND NOT NEON)
  endforeach()
 endif()

-# FIX for quant.h - requires C99 for() loops
-ocv_check_flag_support(C "-std=c99" _varname "${CMAKE_C_FLAGS}")
-if(${_varname})
-  set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -std=c99")
-endif()
-

 # ----------------------------------------------------------------------------------
 #         Define the library target:
@@ -1,22 +0,0 @@
-diff --git a/3rdparty/libwebp/src/dsp/msa_macro.h b/3rdparty/libwebp/src/dsp/msa_macro.h
-index de026a1d9e..a16c0bb300 100644
--- a/3rdparty/libwebp/src/dsp/msa_macro.h
-+++ b/3rdparty/libwebp/src/dsp/msa_macro.h
-@@ -73,7 +73,7 @@
-   static inline TYPE FUNC_NAME(const void* const psrc) {  \
-     const uint8_t* const psrc_m = (const uint8_t*)psrc;   \
-     TYPE val_m;                                           \
-    asm volatile (                                        \
-+    __asm__ volatile (                                        \
-       "" #INSTR " %[val_m], %[psrc_m]  \n\t"              \
-       : [val_m] "=r" (val_m)                              \
-       : [psrc_m] "m" (*psrc_m));                          \
-@@ -86,7 +86,7 @@
-   static inline void FUNC_NAME(TYPE val, void* const pdst) { \
-     uint8_t* const pdst_m = (uint8_t*)pdst;                  \
-     TYPE val_m = val;                                        \
-    asm volatile (                                           \
-+    __asm__ volatile (                                           \
-       " " #INSTR "  %[val_m],  %[pdst_m]  \n\t"              \
-       : [pdst_m] "=m" (*pdst_m)                              \
-       : [val_m] "r" (val_m));                                \
@@ -0,0 +1,565 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Sharp RGB to YUV conversion.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "sharpyuv/sharpyuv.h"
+
+#include <assert.h>
+#include <limits.h>
+#include <stddef.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "src/webp/types.h"
+#include "sharpyuv/sharpyuv_cpu.h"
+#include "sharpyuv/sharpyuv_dsp.h"
+#include "sharpyuv/sharpyuv_gamma.h"
+
+//------------------------------------------------------------------------------
+
+int SharpYuvGetVersion(void) {
+  return SHARPYUV_VERSION;
+}
+
+//------------------------------------------------------------------------------
+// Sharp RGB->YUV conversion
+
+static const int kNumIterations = 4;
+
+#define YUV_FIX 16  // fixed-point precision for RGB->YUV
+static const int kYuvHalf = 1 << (YUV_FIX - 1);
+
+// Max bit depth so that intermediate calculations fit in 16 bits.
+static const int kMaxBitDepth = 14;
+
+// Returns the precision shift to use based on the input rgb_bit_depth.
+static int GetPrecisionShift(int rgb_bit_depth) {
+  // Try to add 2 bits of precision if it fits in kMaxBitDepth. Otherwise remove
+  // bits if needed.
+  return ((rgb_bit_depth + 2) <= kMaxBitDepth) ? 2
+                                               : (kMaxBitDepth - rgb_bit_depth);
+}
+
+typedef int16_t fixed_t;      // signed type with extra precision for UV
+typedef uint16_t fixed_y_t;   // unsigned type with extra precision for W
+
+//------------------------------------------------------------------------------
+
+static uint8_t clip_8b(fixed_t v) {
+  return (!(v & ~0xff)) ? (uint8_t)v : (v < 0) ? 0u : 255u;
+}
+
+static uint16_t clip(fixed_t v, int max) {
+  return (v < 0) ? 0 : (v > max) ? max : (uint16_t)v;
+}
+
+static fixed_y_t clip_bit_depth(int y, int bit_depth) {
+  const int max = (1 << bit_depth) - 1;
+  return (!(y & ~max)) ? (fixed_y_t)y : (y < 0) ? 0 : max;
+}
+
+//------------------------------------------------------------------------------
+
+static int RGBToGray(int64_t r, int64_t g, int64_t b) {
+  const int64_t luma = 13933 * r + 46871 * g + 4732 * b + kYuvHalf;
+  return (int)(luma >> YUV_FIX);
+}
+
+static uint32_t ScaleDown(uint16_t a, uint16_t b, uint16_t c, uint16_t d,
+                          int rgb_bit_depth,
+                          SharpYuvTransferFunctionType transfer_type) {
+  const int bit_depth = rgb_bit_depth + GetPrecisionShift(rgb_bit_depth);
+  const uint32_t A = SharpYuvGammaToLinear(a, bit_depth, transfer_type);
+  const uint32_t B = SharpYuvGammaToLinear(b, bit_depth, transfer_type);
+  const uint32_t C = SharpYuvGammaToLinear(c, bit_depth, transfer_type);
+  const uint32_t D = SharpYuvGammaToLinear(d, bit_depth, transfer_type);
+  return SharpYuvLinearToGamma((A + B + C + D + 2) >> 2, bit_depth,
+                               transfer_type);
+}
+
+static WEBP_INLINE void UpdateW(const fixed_y_t* src, fixed_y_t* dst, int w,
+                                int rgb_bit_depth,
+                                SharpYuvTransferFunctionType transfer_type) {
+  const int bit_depth = rgb_bit_depth + GetPrecisionShift(rgb_bit_depth);
+  int i;
+  for (i = 0; i < w; ++i) {
+    const uint32_t R =
+        SharpYuvGammaToLinear(src[0 * w + i], bit_depth, transfer_type);
+    const uint32_t G =
+        SharpYuvGammaToLinear(src[1 * w + i], bit_depth, transfer_type);
+    const uint32_t B =
+        SharpYuvGammaToLinear(src[2 * w + i], bit_depth, transfer_type);
+    const uint32_t Y = RGBToGray(R, G, B);
+    dst[i] = (fixed_y_t)SharpYuvLinearToGamma(Y, bit_depth, transfer_type);
+  }
+}
+
+static void UpdateChroma(const fixed_y_t* src1, const fixed_y_t* src2,
+                         fixed_t* dst, int uv_w, int rgb_bit_depth,
+                         SharpYuvTransferFunctionType transfer_type) {
+  int i;
+  for (i = 0; i < uv_w; ++i) {
+    const int r =
+        ScaleDown(src1[0 * uv_w + 0], src1[0 * uv_w + 1], src2[0 * uv_w + 0],
+                  src2[0 * uv_w + 1], rgb_bit_depth, transfer_type);
+    const int g =
+        ScaleDown(src1[2 * uv_w + 0], src1[2 * uv_w + 1], src2[2 * uv_w + 0],
+                  src2[2 * uv_w + 1], rgb_bit_depth, transfer_type);
+    const int b =
+        ScaleDown(src1[4 * uv_w + 0], src1[4 * uv_w + 1], src2[4 * uv_w + 0],
+                  src2[4 * uv_w + 1], rgb_bit_depth, transfer_type);
+    const int W = RGBToGray(r, g, b);
+    dst[0 * uv_w] = (fixed_t)(r - W);
+    dst[1 * uv_w] = (fixed_t)(g - W);
+    dst[2 * uv_w] = (fixed_t)(b - W);
+    dst  += 1;
+    src1 += 2;
+    src2 += 2;
+  }
+}
+
+static void StoreGray(const fixed_y_t* rgb, fixed_y_t* y, int w) {
+  int i;
+  assert(w > 0);
+  for (i = 0; i < w; ++i) {
+    y[i] = RGBToGray(rgb[0 * w + i], rgb[1 * w + i], rgb[2 * w + i]);
+  }
+}
+
+//------------------------------------------------------------------------------
+
+static WEBP_INLINE fixed_y_t Filter2(int A, int B, int W0, int bit_depth) {
+  const int v0 = (A * 3 + B + 2) >> 2;
+  return clip_bit_depth(v0 + W0, bit_depth);
+}
+
+//------------------------------------------------------------------------------
+
+static WEBP_INLINE int Shift(int v, int shift) {
+  return (shift >= 0) ? (v << shift) : (v >> -shift);
+}
+
+static void ImportOneRow(const uint8_t* const r_ptr,
+                         const uint8_t* const g_ptr,
+                         const uint8_t* const b_ptr,
+                         int rgb_step,
+                         int rgb_bit_depth,
+                         int pic_width,
+                         fixed_y_t* const dst) {
+  // Convert the rgb_step from a number of bytes to a number of uint8_t or
+  // uint16_t values depending the bit depth.
+  const int step = (rgb_bit_depth > 8) ? rgb_step / 2 : rgb_step;
+  int i;
+  const int w = (pic_width + 1) & ~1;
+  for (i = 0; i < pic_width; ++i) {
+    const int off = i * step;
+    const int shift = GetPrecisionShift(rgb_bit_depth);
+    if (rgb_bit_depth == 8) {
+      dst[i + 0 * w] = Shift(r_ptr[off], shift);
+      dst[i + 1 * w] = Shift(g_ptr[off], shift);
+      dst[i + 2 * w] = Shift(b_ptr[off], shift);
+    } else {
+      dst[i + 0 * w] = Shift(((uint16_t*)r_ptr)[off], shift);
+      dst[i + 1 * w] = Shift(((uint16_t*)g_ptr)[off], shift);
+      dst[i + 2 * w] = Shift(((uint16_t*)b_ptr)[off], shift);
+    }
+  }
+  if (pic_width & 1) {  // replicate rightmost pixel
+    dst[pic_width + 0 * w] = dst[pic_width + 0 * w - 1];
+    dst[pic_width + 1 * w] = dst[pic_width + 1 * w - 1];
+    dst[pic_width + 2 * w] = dst[pic_width + 2 * w - 1];
+  }
+}
+
+static void InterpolateTwoRows(const fixed_y_t* const best_y,
+                               const fixed_t* prev_uv,
+                               const fixed_t* cur_uv,
+                               const fixed_t* next_uv,
+                               int w,
+                               fixed_y_t* out1,
+                               fixed_y_t* out2,
+                               int rgb_bit_depth) {
+  const int uv_w = w >> 1;
+  const int len = (w - 1) >> 1;   // length to filter
+  int k = 3;
+  const int bit_depth = rgb_bit_depth + GetPrecisionShift(rgb_bit_depth);
+  while (k-- > 0) {   // process each R/G/B segments in turn
+    // special boundary case for i==0
+    out1[0] = Filter2(cur_uv[0], prev_uv[0], best_y[0], bit_depth);
+    out2[0] = Filter2(cur_uv[0], next_uv[0], best_y[w], bit_depth);
+
+    SharpYuvFilterRow(cur_uv, prev_uv, len, best_y + 0 + 1, out1 + 1,
+                      bit_depth);
+    SharpYuvFilterRow(cur_uv, next_uv, len, best_y + w + 1, out2 + 1,
+                      bit_depth);
+
+    // special boundary case for i == w - 1 when w is even
+    if (!(w & 1)) {
+      out1[w - 1] = Filter2(cur_uv[uv_w - 1], prev_uv[uv_w - 1],
+                            best_y[w - 1 + 0], bit_depth);
+      out2[w - 1] = Filter2(cur_uv[uv_w - 1], next_uv[uv_w - 1],
+                            best_y[w - 1 + w], bit_depth);
+    }
+    out1 += w;
+    out2 += w;
+    prev_uv += uv_w;
+    cur_uv  += uv_w;
+    next_uv += uv_w;
+  }
+}
+
+static WEBP_INLINE int RGBToYUVComponent(int r, int g, int b,
+                                         const int coeffs[4], int sfix) {
+  const int srounder = 1 << (YUV_FIX + sfix - 1);
+  const int luma = coeffs[0] * r + coeffs[1] * g + coeffs[2] * b +
+                   coeffs[3] + srounder;
+  return (luma >> (YUV_FIX + sfix));
+}
+
+static int ConvertWRGBToYUV(const fixed_y_t* best_y, const fixed_t* best_uv,
+                            uint8_t* y_ptr, int y_stride, uint8_t* u_ptr,
+                            int u_stride, uint8_t* v_ptr, int v_stride,
+                            int rgb_bit_depth,
+                            int yuv_bit_depth, int width, int height,
+                            const SharpYuvConversionMatrix* yuv_matrix) {
+  int i, j;
+  const fixed_t* const best_uv_base = best_uv;
+  const int w = (width + 1) & ~1;
+  const int h = (height + 1) & ~1;
+  const int uv_w = w >> 1;
+  const int uv_h = h >> 1;
+  const int sfix = GetPrecisionShift(rgb_bit_depth);
+  const int yuv_max = (1 << yuv_bit_depth) - 1;
+
+  for (best_uv = best_uv_base, j = 0; j < height; ++j) {
+    for (i = 0; i < width; ++i) {
+      const int off = (i >> 1);
+      const int W = best_y[i];
+      const int r = best_uv[off + 0 * uv_w] + W;
+      const int g = best_uv[off + 1 * uv_w] + W;
+      const int b = best_uv[off + 2 * uv_w] + W;
+      const int y = RGBToYUVComponent(r, g, b, yuv_matrix->rgb_to_y, sfix);
+      if (yuv_bit_depth <= 8) {
+        y_ptr[i] = clip_8b(y);
+      } else {
+        ((uint16_t*)y_ptr)[i] = clip(y, yuv_max);
+      }
+    }
+    best_y += w;
+    best_uv += (j & 1) * 3 * uv_w;
+    y_ptr += y_stride;
+  }
+  for (best_uv = best_uv_base, j = 0; j < uv_h; ++j) {
+    for (i = 0; i < uv_w; ++i) {
+      const int off = i;
+      // Note r, g and b values here are off by W, but a constant offset on all
+      // 3 components doesn't change the value of u and v with a YCbCr matrix.
+      const int r = best_uv[off + 0 * uv_w];
+      const int g = best_uv[off + 1 * uv_w];
+      const int b = best_uv[off + 2 * uv_w];
+      const int u = RGBToYUVComponent(r, g, b, yuv_matrix->rgb_to_u, sfix);
+      const int v = RGBToYUVComponent(r, g, b, yuv_matrix->rgb_to_v, sfix);
+      if (yuv_bit_depth <= 8) {
+        u_ptr[i] = clip_8b(u);
+        v_ptr[i] = clip_8b(v);
+      } else {
+        ((uint16_t*)u_ptr)[i] = clip(u, yuv_max);
+        ((uint16_t*)v_ptr)[i] = clip(v, yuv_max);
+      }
+    }
+    best_uv += 3 * uv_w;
+    u_ptr += u_stride;
+    v_ptr += v_stride;
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// Main function
+
+static void* SafeMalloc(uint64_t nmemb, size_t size) {
+  const uint64_t total_size = nmemb * (uint64_t)size;
+  if (total_size != (size_t)total_size) return NULL;
+  return malloc((size_t)total_size);
+}
+
+#define SAFE_ALLOC(W, H, T) ((T*)SafeMalloc((W) * (H), sizeof(T)))
+
+static int DoSharpArgbToYuv(const uint8_t* r_ptr, const uint8_t* g_ptr,
+                            const uint8_t* b_ptr, int rgb_step, int rgb_stride,
+                            int rgb_bit_depth, uint8_t* y_ptr, int y_stride,
+                            uint8_t* u_ptr, int u_stride, uint8_t* v_ptr,
+                            int v_stride, int yuv_bit_depth, int width,
+                            int height,
+                            const SharpYuvConversionMatrix* yuv_matrix,
+                            SharpYuvTransferFunctionType transfer_type) {
+  // we expand the right/bottom border if needed
+  const int w = (width + 1) & ~1;
+  const int h = (height + 1) & ~1;
+  const int uv_w = w >> 1;
+  const int uv_h = h >> 1;
+  uint64_t prev_diff_y_sum = ~0;
+  int j, iter;
+
+  // TODO(skal): allocate one big memory chunk. But for now, it's easier
+  // for valgrind debugging to have several chunks.
+  fixed_y_t* const tmp_buffer = SAFE_ALLOC(w * 3, 2, fixed_y_t);   // scratch
+  fixed_y_t* const best_y_base = SAFE_ALLOC(w, h, fixed_y_t);
+  fixed_y_t* const target_y_base = SAFE_ALLOC(w, h, fixed_y_t);
+  fixed_y_t* const best_rgb_y = SAFE_ALLOC(w, 2, fixed_y_t);
+  fixed_t* const best_uv_base = SAFE_ALLOC(uv_w * 3, uv_h, fixed_t);
+  fixed_t* const target_uv_base = SAFE_ALLOC(uv_w * 3, uv_h, fixed_t);
+  fixed_t* const best_rgb_uv = SAFE_ALLOC(uv_w * 3, 1, fixed_t);
+  fixed_y_t* best_y = best_y_base;
+  fixed_y_t* target_y = target_y_base;
+  fixed_t* best_uv = best_uv_base;
+  fixed_t* target_uv = target_uv_base;
+  const uint64_t diff_y_threshold = (uint64_t)(3.0 * w * h);
+  int ok;
+  assert(w > 0);
+  assert(h > 0);
+
+  if (best_y_base == NULL || best_uv_base == NULL ||
+      target_y_base == NULL || target_uv_base == NULL ||
+      best_rgb_y == NULL || best_rgb_uv == NULL ||
+      tmp_buffer == NULL) {
+    ok = 0;
+    goto End;
+  }
+
+  // Import RGB samples to W/RGB representation.
+  for (j = 0; j < height; j += 2) {
+    const int is_last_row = (j == height - 1);
+    fixed_y_t* const src1 = tmp_buffer + 0 * w;
+    fixed_y_t* const src2 = tmp_buffer + 3 * w;
+
+    // prepare two rows of input
+    ImportOneRow(r_ptr, g_ptr, b_ptr, rgb_step, rgb_bit_depth, width,
+                 src1);
+    if (!is_last_row) {
+      ImportOneRow(r_ptr + rgb_stride, g_ptr + rgb_stride, b_ptr + rgb_stride,
+                   rgb_step, rgb_bit_depth, width, src2);
+    } else {
+      memcpy(src2, src1, 3 * w * sizeof(*src2));
+    }
+    StoreGray(src1, best_y + 0, w);
+    StoreGray(src2, best_y + w, w);
+
+    UpdateW(src1, target_y, w, rgb_bit_depth, transfer_type);
+    UpdateW(src2, target_y + w, w, rgb_bit_depth, transfer_type);
+    UpdateChroma(src1, src2, target_uv, uv_w, rgb_bit_depth, transfer_type);
+    memcpy(best_uv, target_uv, 3 * uv_w * sizeof(*best_uv));
+    best_y += 2 * w;
+    best_uv += 3 * uv_w;
+    target_y += 2 * w;
+    target_uv += 3 * uv_w;
+    r_ptr += 2 * rgb_stride;
+    g_ptr += 2 * rgb_stride;
+    b_ptr += 2 * rgb_stride;
+  }
+
+  // Iterate and resolve clipping conflicts.
+  for (iter = 0; iter < kNumIterations; ++iter) {
+    const fixed_t* cur_uv = best_uv_base;
+    const fixed_t* prev_uv = best_uv_base;
+    uint64_t diff_y_sum = 0;
+
+    best_y = best_y_base;
+    best_uv = best_uv_base;
+    target_y = target_y_base;
+    target_uv = target_uv_base;
+    for (j = 0; j < h; j += 2) {
+      fixed_y_t* const src1 = tmp_buffer + 0 * w;
+      fixed_y_t* const src2 = tmp_buffer + 3 * w;
+      {
+        const fixed_t* const next_uv = cur_uv + ((j < h - 2) ? 3 * uv_w : 0);
+        InterpolateTwoRows(best_y, prev_uv, cur_uv, next_uv, w,
+                           src1, src2, rgb_bit_depth);
+        prev_uv = cur_uv;
+        cur_uv = next_uv;
+      }
+
+      UpdateW(src1, best_rgb_y + 0 * w, w, rgb_bit_depth, transfer_type);
+      UpdateW(src2, best_rgb_y + 1 * w, w, rgb_bit_depth, transfer_type);
+      UpdateChroma(src1, src2, best_rgb_uv, uv_w, rgb_bit_depth, transfer_type);
+
+      // update two rows of Y and one row of RGB
+      diff_y_sum +=
+          SharpYuvUpdateY(target_y, best_rgb_y, best_y, 2 * w,
+                          rgb_bit_depth + GetPrecisionShift(rgb_bit_depth));
+      SharpYuvUpdateRGB(target_uv, best_rgb_uv, best_uv, 3 * uv_w);
+
+      best_y += 2 * w;
+      best_uv += 3 * uv_w;
+      target_y += 2 * w;
+      target_uv += 3 * uv_w;
+    }
+    // test exit condition
+    if (iter > 0) {
+      if (diff_y_sum < diff_y_threshold) break;
+      if (diff_y_sum > prev_diff_y_sum) break;
+    }
+    prev_diff_y_sum = diff_y_sum;
+  }
+
+  // final reconstruction
+  ok = ConvertWRGBToYUV(best_y_base, best_uv_base, y_ptr, y_stride, u_ptr,
+                        u_stride, v_ptr, v_stride, rgb_bit_depth, yuv_bit_depth,
+                        width, height, yuv_matrix);
+
+ End:
+  free(best_y_base);
+  free(best_uv_base);
+  free(target_y_base);
+  free(target_uv_base);
+  free(best_rgb_y);
+  free(best_rgb_uv);
+  free(tmp_buffer);
+  return ok;
+}
+#undef SAFE_ALLOC
+
+#if defined(WEBP_USE_THREAD) && !defined(_WIN32)
+#include <pthread.h>  // NOLINT
+
+#define LOCK_ACCESS \
+    static pthread_mutex_t sharpyuv_lock = PTHREAD_MUTEX_INITIALIZER; \
+    if (pthread_mutex_lock(&sharpyuv_lock)) return
+#define UNLOCK_ACCESS_AND_RETURN                  \
+    do {                                          \
+      (void)pthread_mutex_unlock(&sharpyuv_lock); \
+      return;                                     \
+    } while (0)
+#else  // !(defined(WEBP_USE_THREAD) && !defined(_WIN32))
+#define LOCK_ACCESS do {} while (0)
+#define UNLOCK_ACCESS_AND_RETURN return
+#endif  // defined(WEBP_USE_THREAD) && !defined(_WIN32)
+
+// Hidden exported init function.
+// By default SharpYuvConvert calls it with SharpYuvGetCPUInfo. If needed,
+// users can declare it as extern and call it with an alternate VP8CPUInfo
+// function.
+extern VP8CPUInfo SharpYuvGetCPUInfo;
+SHARPYUV_EXTERN void SharpYuvInit(VP8CPUInfo cpu_info_func);
+void SharpYuvInit(VP8CPUInfo cpu_info_func) {
+  static volatile VP8CPUInfo sharpyuv_last_cpuinfo_used =
+      (VP8CPUInfo)&sharpyuv_last_cpuinfo_used;
+  LOCK_ACCESS;
+  // Only update SharpYuvGetCPUInfo when called from external code to avoid a
+  // race on reading the value in SharpYuvConvert().
+  if (cpu_info_func != (VP8CPUInfo)&SharpYuvGetCPUInfo) {
+    SharpYuvGetCPUInfo = cpu_info_func;
+  }
+  if (sharpyuv_last_cpuinfo_used == SharpYuvGetCPUInfo) {
+    UNLOCK_ACCESS_AND_RETURN;
+  }
+
+  SharpYuvInitDsp();
+  SharpYuvInitGammaTables();
+
+  sharpyuv_last_cpuinfo_used = SharpYuvGetCPUInfo;
+  UNLOCK_ACCESS_AND_RETURN;
+}
+
+int SharpYuvConvert(const void* r_ptr, const void* g_ptr, const void* b_ptr,
+                    int rgb_step, int rgb_stride, int rgb_bit_depth,
+                    void* y_ptr, int y_stride, void* u_ptr, int u_stride,
+                    void* v_ptr, int v_stride, int yuv_bit_depth, int width,
+                    int height, const SharpYuvConversionMatrix* yuv_matrix) {
+  SharpYuvOptions options;
+  options.yuv_matrix = yuv_matrix;
+  options.transfer_type = kSharpYuvTransferFunctionSrgb;
+  return SharpYuvConvertWithOptions(
+      r_ptr, g_ptr, b_ptr, rgb_step, rgb_stride, rgb_bit_depth, y_ptr, y_stride,
+      u_ptr, u_stride, v_ptr, v_stride, yuv_bit_depth, width, height, &options);
+}
+
+int SharpYuvOptionsInitInternal(const SharpYuvConversionMatrix* yuv_matrix,
+                                SharpYuvOptions* options, int version) {
+  const int major = (version >> 24);
+  const int minor = (version >> 16) & 0xff;
+  if (options == NULL || yuv_matrix == NULL ||
+      (major == SHARPYUV_VERSION_MAJOR && major == 0 &&
+       minor != SHARPYUV_VERSION_MINOR) ||
+      (major != SHARPYUV_VERSION_MAJOR)) {
+    return 0;
+  }
+  options->yuv_matrix = yuv_matrix;
+  options->transfer_type = kSharpYuvTransferFunctionSrgb;
+  return 1;
+}
+
+int SharpYuvConvertWithOptions(const void* r_ptr, const void* g_ptr,
+                               const void* b_ptr, int rgb_step, int rgb_stride,
+                               int rgb_bit_depth, void* y_ptr, int y_stride,
+                               void* u_ptr, int u_stride, void* v_ptr,
+                               int v_stride, int yuv_bit_depth, int width,
+                               int height, const SharpYuvOptions* options) {
+  const SharpYuvConversionMatrix* yuv_matrix = options->yuv_matrix;
+  SharpYuvTransferFunctionType transfer_type = options->transfer_type;
+  SharpYuvConversionMatrix scaled_matrix;
+  const int rgb_max = (1 << rgb_bit_depth) - 1;
+  const int rgb_round = 1 << (rgb_bit_depth - 1);
+  const int yuv_max = (1 << yuv_bit_depth) - 1;
+  const int sfix = GetPrecisionShift(rgb_bit_depth);
+
+  if (width < 1 || height < 1 || width == INT_MAX || height == INT_MAX ||
+      r_ptr == NULL || g_ptr == NULL || b_ptr == NULL || y_ptr == NULL ||
+      u_ptr == NULL || v_ptr == NULL) {
+    return 0;
+  }
+  if (rgb_bit_depth != 8 && rgb_bit_depth != 10 && rgb_bit_depth != 12 &&
+      rgb_bit_depth != 16) {
+    return 0;
+  }
+  if (yuv_bit_depth != 8 && yuv_bit_depth != 10 && yuv_bit_depth != 12) {
+    return 0;
+  }
+  if (rgb_bit_depth > 8 && (rgb_step % 2 != 0 || rgb_stride %2 != 0)) {
+    // Step/stride should be even for uint16_t buffers.
+    return 0;
+  }
+  if (yuv_bit_depth > 8 &&
+      (y_stride % 2 != 0 || u_stride % 2 != 0 || v_stride % 2 != 0)) {
+    // Stride should be even for uint16_t buffers.
+    return 0;
+  }
+  // The address of the function pointer is used to avoid a read race.
+  SharpYuvInit((VP8CPUInfo)&SharpYuvGetCPUInfo);
+
+  // Add scaling factor to go from rgb_bit_depth to yuv_bit_depth, to the
+  // rgb->yuv conversion matrix.
+  if (rgb_bit_depth == yuv_bit_depth) {
+    memcpy(&scaled_matrix, yuv_matrix, sizeof(scaled_matrix));
+  } else {
+    int i;
+    for (i = 0; i < 3; ++i) {
+      scaled_matrix.rgb_to_y[i] =
+          (yuv_matrix->rgb_to_y[i] * yuv_max + rgb_round) / rgb_max;
+      scaled_matrix.rgb_to_u[i] =
+          (yuv_matrix->rgb_to_u[i] * yuv_max + rgb_round) / rgb_max;
+      scaled_matrix.rgb_to_v[i] =
+          (yuv_matrix->rgb_to_v[i] * yuv_max + rgb_round) / rgb_max;
+    }
+  }
+  // Also incorporate precision change scaling.
+  scaled_matrix.rgb_to_y[3] = Shift(yuv_matrix->rgb_to_y[3], sfix);
+  scaled_matrix.rgb_to_u[3] = Shift(yuv_matrix->rgb_to_u[3], sfix);
+  scaled_matrix.rgb_to_v[3] = Shift(yuv_matrix->rgb_to_v[3], sfix);
+
+  return DoSharpArgbToYuv(r_ptr, g_ptr, b_ptr, rgb_step, rgb_stride,
+                          rgb_bit_depth, y_ptr, y_stride, u_ptr, u_stride,
+                          v_ptr, v_stride, yuv_bit_depth, width, height,
+                          &scaled_matrix, transfer_type);
+}
+
+//------------------------------------------------------------------------------
@@ -0,0 +1,174 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Sharp RGB to YUV conversion.
+
+#ifndef WEBP_SHARPYUV_SHARPYUV_H_
+#define WEBP_SHARPYUV_SHARPYUV_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef SHARPYUV_EXTERN
+#ifdef WEBP_EXTERN
+#define SHARPYUV_EXTERN WEBP_EXTERN
+#else
+// This explicitly marks library functions and allows for changing the
+// signature for e.g., Windows DLL builds.
+#if defined(__GNUC__) && __GNUC__ >= 4
+#define SHARPYUV_EXTERN extern __attribute__((visibility("default")))
+#else
+#if defined(_MSC_VER) && defined(WEBP_DLL)
+#define SHARPYUV_EXTERN __declspec(dllexport)
+#else
+#define SHARPYUV_EXTERN extern
+#endif /* _MSC_VER && WEBP_DLL */
+#endif /* __GNUC__ >= 4 */
+#endif /* WEBP_EXTERN */
+#endif /* SHARPYUV_EXTERN */
+
+#ifndef SHARPYUV_INLINE
+#ifdef WEBP_INLINE
+#define SHARPYUV_INLINE WEBP_INLINE
+#else
+#ifndef _MSC_VER
+#if defined(__cplusplus) || !defined(__STRICT_ANSI__) || \
+    (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L)
+#define SHARPYUV_INLINE inline
+#else
+#define SHARPYUV_INLINE
+#endif
+#else
+#define SHARPYUV_INLINE __forceinline
+#endif /* _MSC_VER */
+#endif /* WEBP_INLINE */
+#endif /* SHARPYUV_INLINE */
+
+// SharpYUV API version following the convention from semver.org
+#define SHARPYUV_VERSION_MAJOR 0
+#define SHARPYUV_VERSION_MINOR 4
+#define SHARPYUV_VERSION_PATCH 0
+// Version as a uint32_t. The major number is the high 8 bits.
+// The minor number is the middle 8 bits. The patch number is the low 16 bits.
+#define SHARPYUV_MAKE_VERSION(MAJOR, MINOR, PATCH) \
+  (((MAJOR) << 24) | ((MINOR) << 16) | (PATCH))
+#define SHARPYUV_VERSION                                                \
+  SHARPYUV_MAKE_VERSION(SHARPYUV_VERSION_MAJOR, SHARPYUV_VERSION_MINOR, \
+                        SHARPYUV_VERSION_PATCH)
+
+// Returns the library's version number, packed in hexadecimal. See
+// SHARPYUV_VERSION.
+SHARPYUV_EXTERN int SharpYuvGetVersion(void);
+
+// RGB to YUV conversion matrix, in 16 bit fixed point.
+// y = rgb_to_y[0] * r + rgb_to_y[1] * g + rgb_to_y[2] * b + rgb_to_y[3]
+// u = rgb_to_u[0] * r + rgb_to_u[1] * g + rgb_to_u[2] * b + rgb_to_u[3]
+// v = rgb_to_v[0] * r + rgb_to_v[1] * g + rgb_to_v[2] * b + rgb_to_v[3]
+// Then y, u and v values are divided by 1<<16 and rounded.
+typedef struct {
+  int rgb_to_y[4];
+  int rgb_to_u[4];
+  int rgb_to_v[4];
+} SharpYuvConversionMatrix;
+
+typedef struct SharpYuvOptions SharpYuvOptions;
+
+// Enums for transfer functions, as defined in H.273,
+// https://www.itu.int/rec/T-REC-H.273-202107-I/en
+typedef enum SharpYuvTransferFunctionType {
+  // 0 is reserved
+  kSharpYuvTransferFunctionBt709 = 1,
+  // 2 is unspecified
+  // 3 is reserved
+  kSharpYuvTransferFunctionBt470M = 4,
+  kSharpYuvTransferFunctionBt470Bg = 5,
+  kSharpYuvTransferFunctionBt601 = 6,
+  kSharpYuvTransferFunctionSmpte240 = 7,
+  kSharpYuvTransferFunctionLinear = 8,
+  kSharpYuvTransferFunctionLog100 = 9,
+  kSharpYuvTransferFunctionLog100_Sqrt10 = 10,
+  kSharpYuvTransferFunctionIec61966 = 11,
+  kSharpYuvTransferFunctionBt1361 = 12,
+  kSharpYuvTransferFunctionSrgb = 13,
+  kSharpYuvTransferFunctionBt2020_10Bit = 14,
+  kSharpYuvTransferFunctionBt2020_12Bit = 15,
+  kSharpYuvTransferFunctionSmpte2084 = 16,  // PQ
+  kSharpYuvTransferFunctionSmpte428 = 17,
+  kSharpYuvTransferFunctionHlg = 18,
+  kSharpYuvTransferFunctionNum
+} SharpYuvTransferFunctionType;
+
+// Converts RGB to YUV420 using a downsampling algorithm that minimizes
+// artefacts caused by chroma subsampling.
+// This is slower than standard downsampling (averaging of 4 UV values).
+// Assumes that the image will be upsampled using a bilinear filter. If nearest
+// neighbor is used instead, the upsampled image might look worse than with
+// standard downsampling.
+// r_ptr, g_ptr, b_ptr: pointers to the source r, g and b channels. Should point
+//     to uint8_t buffers if rgb_bit_depth is 8, or uint16_t buffers otherwise.
+// rgb_step: distance in bytes between two horizontally adjacent pixels on the
+//     r, g and b channels. If rgb_bit_depth is > 8, it should be a
+//     multiple of 2.
+// rgb_stride: distance in bytes between two vertically adjacent pixels on the
+//     r, g, and b channels. If rgb_bit_depth is > 8, it should be a
+//     multiple of 2.
+// rgb_bit_depth: number of bits for each r/g/b value. One of: 8, 10, 12, 16.
+//     Note: 16 bit input is truncated to 14 bits before conversion to yuv.
+// yuv_bit_depth: number of bits for each y/u/v value. One of: 8, 10, 12.
+// y_ptr, u_ptr, v_ptr: pointers to the destination y, u and v channels.  Should
+//     point to uint8_t buffers if yuv_bit_depth is 8, or uint16_t buffers
+//     otherwise.
+// y_stride, u_stride, v_stride: distance in bytes between two vertically
+//     adjacent pixels on the y, u and v channels. If yuv_bit_depth > 8, they
+//     should be multiples of 2.
+// width, height: width and height of the image in pixels
+// This function calls SharpYuvConvertWithOptions with a default transfer
+// function of kSharpYuvTransferFunctionSrgb.
+SHARPYUV_EXTERN int SharpYuvConvert(const void* r_ptr, const void* g_ptr,
+                                    const void* b_ptr, int rgb_step,
+                                    int rgb_stride, int rgb_bit_depth,
+                                    void* y_ptr, int y_stride, void* u_ptr,
+                                    int u_stride, void* v_ptr, int v_stride,
+                                    int yuv_bit_depth, int width, int height,
+                                    const SharpYuvConversionMatrix* yuv_matrix);
+
+struct SharpYuvOptions {
+  // This matrix cannot be NULL and can be initialized by
+  // SharpYuvComputeConversionMatrix.
+  const SharpYuvConversionMatrix* yuv_matrix;
+  SharpYuvTransferFunctionType transfer_type;
+};
+
+// Internal, version-checked, entry point
+SHARPYUV_EXTERN int SharpYuvOptionsInitInternal(const SharpYuvConversionMatrix*,
+                                                SharpYuvOptions*, int);
+
+// Should always be called, to initialize a fresh SharpYuvOptions
+// structure before modification. SharpYuvOptionsInit() must have succeeded
+// before using the 'options' object.
+static SHARPYUV_INLINE int SharpYuvOptionsInit(
+    const SharpYuvConversionMatrix* yuv_matrix, SharpYuvOptions* options) {
+  return SharpYuvOptionsInitInternal(yuv_matrix, options, SHARPYUV_VERSION);
+}
+
+SHARPYUV_EXTERN int SharpYuvConvertWithOptions(
+    const void* r_ptr, const void* g_ptr, const void* b_ptr, int rgb_step,
+    int rgb_stride, int rgb_bit_depth, void* y_ptr, int y_stride, void* u_ptr,
+    int u_stride, void* v_ptr, int v_stride, int yuv_bit_depth, int width,
+    int height, const SharpYuvOptions* options);
+
+// TODO(b/194336375): Add YUV444 to YUV420 conversion. Maybe also add 422
+// support (it's rarely used in practice, especially for images).
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // WEBP_SHARPYUV_SHARPYUV_H_
@@ -0,0 +1,14 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+#include "sharpyuv/sharpyuv_cpu.h"
+
+// Include src/dsp/cpu.c to create SharpYuvGetCPUInfo from VP8GetCPUInfo. The
+// function pointer is renamed in sharpyuv_cpu.h.
+#include "src/dsp/cpu.c"
@@ -0,0 +1,22 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+#ifndef WEBP_SHARPYUV_SHARPYUV_CPU_H_
+#define WEBP_SHARPYUV_SHARPYUV_CPU_H_
+
+#include "sharpyuv/sharpyuv.h"
+
+// Avoid exporting SharpYuvGetCPUInfo in shared object / DLL builds.
+// SharpYuvInit() replaces the use of the function pointer.
+#undef WEBP_EXTERN
+#define WEBP_EXTERN extern
+#define VP8GetCPUInfo SharpYuvGetCPUInfo
+#include "src/dsp/cpu.h"
+
+#endif  // WEBP_SHARPYUV_SHARPYUV_CPU_H_
@@ -0,0 +1,110 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Colorspace utilities.
+
+#include "sharpyuv/sharpyuv_csp.h"
+
+#include <assert.h>
+#include <math.h>
+#include <stddef.h>
+
+static int ToFixed16(float f) { return (int)floor(f * (1 << 16) + 0.5f); }
+
+void SharpYuvComputeConversionMatrix(const SharpYuvColorSpace* yuv_color_space,
+                                     SharpYuvConversionMatrix* matrix) {
+  const float kr = yuv_color_space->kr;
+  const float kb = yuv_color_space->kb;
+  const float kg = 1.0f - kr - kb;
+  const float cr = 0.5f / (1.0f - kb);
+  const float cb = 0.5f / (1.0f - kr);
+
+  const int shift = yuv_color_space->bit_depth - 8;
+
+  const float denom = (float)((1 << yuv_color_space->bit_depth) - 1);
+  float scale_y = 1.0f;
+  float add_y = 0.0f;
+  float scale_u = cr;
+  float scale_v = cb;
+  float add_uv = (float)(128 << shift);
+  assert(yuv_color_space->bit_depth >= 8);
+
+  if (yuv_color_space->range == kSharpYuvRangeLimited) {
+    scale_y *= (219 << shift) / denom;
+    scale_u *= (224 << shift) / denom;
+    scale_v *= (224 << shift) / denom;
+    add_y = (float)(16 << shift);
+  }
+
+  matrix->rgb_to_y[0] = ToFixed16(kr * scale_y);
+  matrix->rgb_to_y[1] = ToFixed16(kg * scale_y);
+  matrix->rgb_to_y[2] = ToFixed16(kb * scale_y);
+  matrix->rgb_to_y[3] = ToFixed16(add_y);
+
+  matrix->rgb_to_u[0] = ToFixed16(-kr * scale_u);
+  matrix->rgb_to_u[1] = ToFixed16(-kg * scale_u);
+  matrix->rgb_to_u[2] = ToFixed16((1 - kb) * scale_u);
+  matrix->rgb_to_u[3] = ToFixed16(add_uv);
+
+  matrix->rgb_to_v[0] = ToFixed16((1 - kr) * scale_v);
+  matrix->rgb_to_v[1] = ToFixed16(-kg * scale_v);
+  matrix->rgb_to_v[2] = ToFixed16(-kb * scale_v);
+  matrix->rgb_to_v[3] = ToFixed16(add_uv);
+}
+
+// Matrices are in YUV_FIX fixed point precision.
+// WebP's matrix, similar but not identical to kRec601LimitedMatrix.
+static const SharpYuvConversionMatrix kWebpMatrix = {
+  {16839, 33059, 6420, 16 << 16},
+  {-9719, -19081, 28800, 128 << 16},
+  {28800, -24116, -4684, 128 << 16},
+};
+// Kr=0.2990f Kb=0.1140f bits=8 range=kSharpYuvRangeLimited
+static const SharpYuvConversionMatrix kRec601LimitedMatrix = {
+  {16829, 33039, 6416, 16 << 16},
+  {-9714, -19071, 28784, 128 << 16},
+  {28784, -24103, -4681, 128 << 16},
+};
+// Kr=0.2990f Kb=0.1140f bits=8 range=kSharpYuvRangeFull
+static const SharpYuvConversionMatrix kRec601FullMatrix = {
+  {19595, 38470, 7471, 0},
+  {-11058, -21710, 32768, 128 << 16},
+  {32768, -27439, -5329, 128 << 16},
+};
+// Kr=0.2126f Kb=0.0722f bits=8 range=kSharpYuvRangeLimited
+static const SharpYuvConversionMatrix kRec709LimitedMatrix = {
+  {11966, 40254, 4064, 16 << 16},
+  {-6596, -22189, 28784, 128 << 16},
+  {28784, -26145, -2639, 128 << 16},
+};
+// Kr=0.2126f Kb=0.0722f bits=8 range=kSharpYuvRangeFull
+static const SharpYuvConversionMatrix kRec709FullMatrix = {
+  {13933, 46871, 4732, 0},
+  {-7509, -25259, 32768, 128 << 16},
+  {32768, -29763, -3005, 128 << 16},
+};
+
+const SharpYuvConversionMatrix* SharpYuvGetConversionMatrix(
+    SharpYuvMatrixType matrix_type) {
+  switch (matrix_type) {
+    case kSharpYuvMatrixWebp:
+      return &kWebpMatrix;
+    case kSharpYuvMatrixRec601Limited:
+      return &kRec601LimitedMatrix;
+    case kSharpYuvMatrixRec601Full:
+      return &kRec601FullMatrix;
+    case kSharpYuvMatrixRec709Limited:
+      return &kRec709LimitedMatrix;
+    case kSharpYuvMatrixRec709Full:
+      return &kRec709FullMatrix;
+    case kSharpYuvMatrixNum:
+      return NULL;
+  }
+  return NULL;
+}
@@ -0,0 +1,60 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Colorspace utilities.
+
+#ifndef WEBP_SHARPYUV_SHARPYUV_CSP_H_
+#define WEBP_SHARPYUV_SHARPYUV_CSP_H_
+
+#include "sharpyuv/sharpyuv.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Range of YUV values.
+typedef enum {
+  kSharpYuvRangeFull,     // YUV values between [0;255] (for 8 bit)
+  kSharpYuvRangeLimited   // Y in [16;235], YUV in [16;240] (for 8 bit)
+} SharpYuvRange;
+
+// Constants that define a YUV color space.
+typedef struct {
+  // Kr and Kb are defined such that:
+  // Y = Kr * r + Kg * g + Kb * b where Kg = 1 - Kr - Kb.
+  float kr;
+  float kb;
+  int bit_depth;  // 8, 10 or 12
+  SharpYuvRange range;
+} SharpYuvColorSpace;
+
+// Fills in 'matrix' for the given YUVColorSpace.
+SHARPYUV_EXTERN void SharpYuvComputeConversionMatrix(
+    const SharpYuvColorSpace* yuv_color_space,
+    SharpYuvConversionMatrix* matrix);
+
+// Enums for precomputed conversion matrices.
+typedef enum {
+  kSharpYuvMatrixWebp = 0,
+  kSharpYuvMatrixRec601Limited,
+  kSharpYuvMatrixRec601Full,
+  kSharpYuvMatrixRec709Limited,
+  kSharpYuvMatrixRec709Full,
+  kSharpYuvMatrixNum
+} SharpYuvMatrixType;
+
+// Returns a pointer to a matrix for one of the predefined colorspaces.
+SHARPYUV_EXTERN const SharpYuvConversionMatrix* SharpYuvGetConversionMatrix(
+    SharpYuvMatrixType matrix_type);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // WEBP_SHARPYUV_SHARPYUV_CSP_H_
@@ -0,0 +1,104 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Speed-critical functions for Sharp YUV.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "sharpyuv/sharpyuv_dsp.h"
+
+#include <assert.h>
+#include <stdlib.h>
+
+#include "sharpyuv/sharpyuv_cpu.h"
+
+//-----------------------------------------------------------------------------
+
+#if !WEBP_NEON_OMIT_C_CODE
+static uint16_t clip(int v, int max) {
+  return (v < 0) ? 0 : (v > max) ? max : (uint16_t)v;
+}
+
+static uint64_t SharpYuvUpdateY_C(const uint16_t* ref, const uint16_t* src,
+                                  uint16_t* dst, int len, int bit_depth) {
+  uint64_t diff = 0;
+  int i;
+  const int max_y = (1 << bit_depth) - 1;
+  for (i = 0; i < len; ++i) {
+    const int diff_y = ref[i] - src[i];
+    const int new_y = (int)dst[i] + diff_y;
+    dst[i] = clip(new_y, max_y);
+    diff += (uint64_t)abs(diff_y);
+  }
+  return diff;
+}
+
+static void SharpYuvUpdateRGB_C(const int16_t* ref, const int16_t* src,
+                                int16_t* dst, int len) {
+  int i;
+  for (i = 0; i < len; ++i) {
+    const int diff_uv = ref[i] - src[i];
+    dst[i] += diff_uv;
+  }
+}
+
+static void SharpYuvFilterRow_C(const int16_t* A, const int16_t* B, int len,
+                                const uint16_t* best_y, uint16_t* out,
+                                int bit_depth) {
+  int i;
+  const int max_y = (1 << bit_depth) - 1;
+  for (i = 0; i < len; ++i, ++A, ++B) {
+    const int v0 = (A[0] * 9 + A[1] * 3 + B[0] * 3 + B[1] + 8) >> 4;
+    const int v1 = (A[1] * 9 + A[0] * 3 + B[1] * 3 + B[0] + 8) >> 4;
+    out[2 * i + 0] = clip(best_y[2 * i + 0] + v0, max_y);
+    out[2 * i + 1] = clip(best_y[2 * i + 1] + v1, max_y);
+  }
+}
+#endif  // !WEBP_NEON_OMIT_C_CODE
+
+//-----------------------------------------------------------------------------
+
+uint64_t (*SharpYuvUpdateY)(const uint16_t* src, const uint16_t* ref,
+                            uint16_t* dst, int len, int bit_depth);
+void (*SharpYuvUpdateRGB)(const int16_t* src, const int16_t* ref, int16_t* dst,
+                          int len);
+void (*SharpYuvFilterRow)(const int16_t* A, const int16_t* B, int len,
+                          const uint16_t* best_y, uint16_t* out,
+                          int bit_depth);
+
+extern VP8CPUInfo SharpYuvGetCPUInfo;
+extern void InitSharpYuvSSE2(void);
+extern void InitSharpYuvNEON(void);
+
+void SharpYuvInitDsp(void) {
+#if !WEBP_NEON_OMIT_C_CODE
+  SharpYuvUpdateY = SharpYuvUpdateY_C;
+  SharpYuvUpdateRGB = SharpYuvUpdateRGB_C;
+  SharpYuvFilterRow = SharpYuvFilterRow_C;
+#endif
+
+  if (SharpYuvGetCPUInfo != NULL) {
+#if defined(WEBP_HAVE_SSE2)
+    if (SharpYuvGetCPUInfo(kSSE2)) {
+      InitSharpYuvSSE2();
+    }
+#endif  // WEBP_HAVE_SSE2
+  }
+
+#if defined(WEBP_HAVE_NEON)
+  if (WEBP_NEON_OMIT_C_CODE ||
+      (SharpYuvGetCPUInfo != NULL && SharpYuvGetCPUInfo(kNEON))) {
+    InitSharpYuvNEON();
+  }
+#endif  // WEBP_HAVE_NEON
+
+  assert(SharpYuvUpdateY != NULL);
+  assert(SharpYuvUpdateRGB != NULL);
+  assert(SharpYuvFilterRow != NULL);
+}
@@ -0,0 +1,28 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Speed-critical functions for Sharp YUV.
+
+#ifndef WEBP_SHARPYUV_SHARPYUV_DSP_H_
+#define WEBP_SHARPYUV_SHARPYUV_DSP_H_
+
+#include "sharpyuv/sharpyuv_cpu.h"
+#include "src/webp/types.h"
+
+extern uint64_t (*SharpYuvUpdateY)(const uint16_t* src, const uint16_t* ref,
+                                   uint16_t* dst, int len, int bit_depth);
+extern void (*SharpYuvUpdateRGB)(const int16_t* src, const int16_t* ref,
+                                 int16_t* dst, int len);
+extern void (*SharpYuvFilterRow)(const int16_t* A, const int16_t* B, int len,
+                                 const uint16_t* best_y, uint16_t* out,
+                                 int bit_depth);
+
+void SharpYuvInitDsp(void);
+
+#endif  // WEBP_SHARPYUV_SHARPYUV_DSP_H_
@@ -0,0 +1,419 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Gamma correction utilities.
+
+#include "sharpyuv/sharpyuv_gamma.h"
+
+#include <assert.h>
+#include <float.h>
+#include <math.h>
+
+#include "src/webp/types.h"
+
+// Gamma correction compensates loss of resolution during chroma subsampling.
+// Size of pre-computed table for converting from gamma to linear.
+#define GAMMA_TO_LINEAR_TAB_BITS 10
+#define GAMMA_TO_LINEAR_TAB_SIZE (1 << GAMMA_TO_LINEAR_TAB_BITS)
+static uint32_t kGammaToLinearTabS[GAMMA_TO_LINEAR_TAB_SIZE + 2];
+#define LINEAR_TO_GAMMA_TAB_BITS 9
+#define LINEAR_TO_GAMMA_TAB_SIZE (1 << LINEAR_TO_GAMMA_TAB_BITS)
+static uint32_t kLinearToGammaTabS[LINEAR_TO_GAMMA_TAB_SIZE + 2];
+
+static const double kGammaF = 1. / 0.45;
+#define GAMMA_TO_LINEAR_BITS 16
+
+static volatile int kGammaTablesSOk = 0;
+void SharpYuvInitGammaTables(void) {
+  assert(GAMMA_TO_LINEAR_BITS <= 16);
+  if (!kGammaTablesSOk) {
+    int v;
+    const double a = 0.09929682680944;
+    const double thresh = 0.018053968510807;
+    const double final_scale = 1 << GAMMA_TO_LINEAR_BITS;
+    // Precompute gamma to linear table.
+    {
+      const double norm = 1. / GAMMA_TO_LINEAR_TAB_SIZE;
+      const double a_rec = 1. / (1. + a);
+      for (v = 0; v <= GAMMA_TO_LINEAR_TAB_SIZE; ++v) {
+        const double g = norm * v;
+        double value;
+        if (g <= thresh * 4.5) {
+          value = g / 4.5;
+        } else {
+          value = pow(a_rec * (g + a), kGammaF);
+        }
+        kGammaToLinearTabS[v] = (uint32_t)(value * final_scale + .5);
+      }
+      // to prevent small rounding errors to cause read-overflow:
+      kGammaToLinearTabS[GAMMA_TO_LINEAR_TAB_SIZE + 1] =
+          kGammaToLinearTabS[GAMMA_TO_LINEAR_TAB_SIZE];
+    }
+    // Precompute linear to gamma table.
+    {
+      const double scale = 1. / LINEAR_TO_GAMMA_TAB_SIZE;
+      for (v = 0; v <= LINEAR_TO_GAMMA_TAB_SIZE; ++v) {
+        const double g = scale * v;
+        double value;
+        if (g <= thresh) {
+          value = 4.5 * g;
+        } else {
+          value = (1. + a) * pow(g, 1. / kGammaF) - a;
+        }
+        kLinearToGammaTabS[v] =
+            (uint32_t)(final_scale * value + 0.5);
+      }
+      // to prevent small rounding errors to cause read-overflow:
+      kLinearToGammaTabS[LINEAR_TO_GAMMA_TAB_SIZE + 1] =
+          kLinearToGammaTabS[LINEAR_TO_GAMMA_TAB_SIZE];
+    }
+    kGammaTablesSOk = 1;
+  }
+}
+
+static WEBP_INLINE int Shift(int v, int shift) {
+  return (shift >= 0) ? (v << shift) : (v >> -shift);
+}
+
+static WEBP_INLINE uint32_t FixedPointInterpolation(int v, uint32_t* tab,
+                                                    int tab_pos_shift_right,
+                                                    int tab_value_shift) {
+  const uint32_t tab_pos = Shift(v, -tab_pos_shift_right);
+  // fractional part, in 'tab_pos_shift' fixed-point precision
+  const uint32_t x = v - (tab_pos << tab_pos_shift_right);  // fractional part
+  // v0 / v1 are in kGammaToLinearBits fixed-point precision (range [0..1])
+  const uint32_t v0 = Shift(tab[tab_pos + 0], tab_value_shift);
+  const uint32_t v1 = Shift(tab[tab_pos + 1], tab_value_shift);
+  // Final interpolation.
+  const uint32_t v2 = (v1 - v0) * x;  // note: v1 >= v0.
+  const int half =
+      (tab_pos_shift_right > 0) ? 1 << (tab_pos_shift_right - 1) : 0;
+  const uint32_t result = v0 + ((v2 + half) >> tab_pos_shift_right);
+  return result;
+}
+
+static uint32_t ToLinearSrgb(uint16_t v, int bit_depth) {
+  const int shift = GAMMA_TO_LINEAR_TAB_BITS - bit_depth;
+  if (shift > 0) {
+    return kGammaToLinearTabS[v << shift];
+  }
+  return FixedPointInterpolation(v, kGammaToLinearTabS, -shift, 0);
+}
+
+static uint16_t FromLinearSrgb(uint32_t value, int bit_depth) {
+  return FixedPointInterpolation(
+      value, kLinearToGammaTabS,
+      (GAMMA_TO_LINEAR_BITS - LINEAR_TO_GAMMA_TAB_BITS),
+      bit_depth - GAMMA_TO_LINEAR_BITS);
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+#define CLAMP(x, low, high) \
+  (((x) < (low)) ? (low) : (((high) < (x)) ? (high) : (x)))
+#define MIN(a, b) (((a) < (b)) ? (a) : (b))
+#define MAX(a, b) (((a) > (b)) ? (a) : (b))
+
+static WEBP_INLINE float Roundf(float x) {
+  if (x < 0)
+    return (float)ceil((double)(x - 0.5f));
+  else
+    return (float)floor((double)(x + 0.5f));
+}
+
+static WEBP_INLINE float Powf(float base, float exp) {
+  return (float)pow((double)base, (double)exp);
+}
+
+static WEBP_INLINE float Log10f(float x) { return (float)log10((double)x); }
+
+static float ToLinear709(float gamma) {
+  if (gamma < 0.f) {
+    return 0.f;
+  } else if (gamma < 4.5f * 0.018053968510807f) {
+    return gamma / 4.5f;
+  } else if (gamma < 1.f) {
+    return Powf((gamma + 0.09929682680944f) / 1.09929682680944f, 1.f / 0.45f);
+  }
+  return 1.f;
+}
+
+static float FromLinear709(float linear) {
+  if (linear < 0.f) {
+    return 0.f;
+  } else if (linear < 0.018053968510807f) {
+    return linear * 4.5f;
+  } else if (linear < 1.f) {
+    return 1.09929682680944f * Powf(linear, 0.45f) - 0.09929682680944f;
+  }
+  return 1.f;
+}
+
+static float ToLinear470M(float gamma) {
+  return Powf(CLAMP(gamma, 0.f, 1.f), 1.f / 2.2f);
+}
+
+static float FromLinear470M(float linear) {
+  return Powf(CLAMP(linear, 0.f, 1.f), 2.2f);
+}
+
+static float ToLinear470Bg(float gamma) {
+  return Powf(CLAMP(gamma, 0.f, 1.f), 1.f / 2.8f);
+}
+
+static float FromLinear470Bg(float linear) {
+  return Powf(CLAMP(linear, 0.f, 1.f), 2.8f);
+}
+
+static float ToLinearSmpte240(float gamma) {
+  if (gamma < 0.f) {
+    return 0.f;
+  } else if (gamma < 4.f * 0.022821585529445f) {
+    return gamma / 4.f;
+  } else if (gamma < 1.f) {
+    return Powf((gamma + 0.111572195921731f) / 1.111572195921731f, 1.f / 0.45f);
+  }
+  return 1.f;
+}
+
+static float FromLinearSmpte240(float linear) {
+  if (linear < 0.f) {
+    return 0.f;
+  } else if (linear < 0.022821585529445f) {
+    return linear * 4.f;
+  } else if (linear < 1.f) {
+    return 1.111572195921731f * Powf(linear, 0.45f) - 0.111572195921731f;
+  }
+  return 1.f;
+}
+
+static float ToLinearLog100(float gamma) {
+  return (gamma < 0.01f) ? 0.0f : 1.0f + Log10f(MIN(gamma, 1.f)) / 2.0f;
+}
+
+static float FromLinearLog100(float linear) {
+  // The function is non-bijective so choose the middle of [0, 0.01].
+  const float mid_interval = 0.01f / 2.f;
+  return (linear <= 0.0f) ? mid_interval
+                          : Powf(10.0f, 2.f * (MIN(linear, 1.f) - 1.0f));
+}
+
+static float ToLinearLog100Sqrt10(float gamma) {
+  return (gamma < 0.00316227766f) ? 0.0f
+                                  : 1.0f + Log10f(MIN(gamma, 1.f)) / 2.5f;
+}
+
+static float FromLinearLog100Sqrt10(float linear) {
+  // The function is non-bijective so choose the middle of [0, 0.00316227766f[.
+  const float mid_interval = 0.00316227766f / 2.f;
+  return (linear < 0.0f) ? mid_interval
+                         : Powf(10.0f, 2.5f * (MIN(linear, 1.f) - 1.0f));
+}
+
+static float ToLinearIec61966(float gamma) {
+  if (gamma <= -4.5f * 0.018053968510807f) {
+    return Powf((-gamma + 0.09929682680944f) / -1.09929682680944f, 1.f / 0.45f);
+  } else if (gamma < 4.5f * 0.018053968510807f) {
+    return gamma / 4.5f;
+  }
+  return Powf((gamma + 0.09929682680944f) / 1.09929682680944f, 1.f / 0.45f);
+}
+
+static float FromLinearIec61966(float linear) {
+  if (linear <= -0.018053968510807f) {
+    return -1.09929682680944f * Powf(-linear, 0.45f) + 0.09929682680944f;
+  } else if (linear < 0.018053968510807f) {
+    return linear * 4.5f;
+  }
+  return 1.09929682680944f * Powf(linear, 0.45f) - 0.09929682680944f;
+}
+
+static float ToLinearBt1361(float gamma) {
+  if (gamma < -0.25f) {
+    return -0.25f;
+  } else if (gamma < 0.f) {
+    return Powf((gamma - 0.02482420670236f) / -0.27482420670236f, 1.f / 0.45f) /
+           -4.f;
+  } else if (gamma < 4.5f * 0.018053968510807f) {
+    return gamma / 4.5f;
+  } else if (gamma < 1.f) {
+    return Powf((gamma + 0.09929682680944f) / 1.09929682680944f, 1.f / 0.45f);
+  }
+  return 1.f;
+}
+
+static float FromLinearBt1361(float linear) {
+  if (linear < -0.25f) {
+    return -0.25f;
+  } else if (linear < 0.f) {
+    return -0.27482420670236f * Powf(-4.f * linear, 0.45f) + 0.02482420670236f;
+  } else if (linear < 0.018053968510807f) {
+    return linear * 4.5f;
+  } else if (linear < 1.f) {
+    return 1.09929682680944f * Powf(linear, 0.45f) - 0.09929682680944f;
+  }
+  return 1.f;
+}
+
+static float ToLinearPq(float gamma) {
+  if (gamma > 0.f) {
+    const float pow_gamma = Powf(gamma, 32.f / 2523.f);
+    const float num = MAX(pow_gamma - 107.f / 128.f, 0.0f);
+    const float den = MAX(2413.f / 128.f - 2392.f / 128.f * pow_gamma, FLT_MIN);
+    return Powf(num / den, 4096.f / 653.f);
+  }
+  return 0.f;
+}
+
+static float FromLinearPq(float linear) {
+  if (linear > 0.f) {
+    const float pow_linear = Powf(linear, 653.f / 4096.f);
+    const float num = 107.f / 128.f + 2413.f / 128.f * pow_linear;
+    const float den = 1.0f + 2392.f / 128.f * pow_linear;
+    return Powf(num / den, 2523.f / 32.f);
+  }
+  return 0.f;
+}
+
+static float ToLinearSmpte428(float gamma) {
+  return Powf(0.91655527974030934f * MAX(gamma, 0.f), 1.f / 2.6f);
+}
+
+static float FromLinearSmpte428(float linear) {
+  return Powf(MAX(linear, 0.f), 2.6f) / 0.91655527974030934f;
+}
+
+// Conversion in BT.2100 requires RGB info. Simplify to gamma correction here.
+static float ToLinearHlg(float gamma) {
+  if (gamma < 0.f) {
+    return 0.f;
+  } else if (gamma <= 0.5f) {
+    return Powf((gamma * gamma) * (1.f / 3.f), 1.2f);
+  }
+  return Powf((expf((gamma - 0.55991073f) / 0.17883277f) + 0.28466892f) / 12.0f,
+              1.2f);
+}
+
+static float FromLinearHlg(float linear) {
+  linear = Powf(linear, 1.f / 1.2f);
+  if (linear < 0.f) {
+    return 0.f;
+  } else if (linear <= (1.f / 12.f)) {
+    return sqrtf(3.f * linear);
+  }
+  return 0.17883277f * logf(12.f * linear - 0.28466892f) + 0.55991073f;
+}
+
+uint32_t SharpYuvGammaToLinear(uint16_t v, int bit_depth,
+                               SharpYuvTransferFunctionType transfer_type) {
+  float v_float, linear;
+  if (transfer_type == kSharpYuvTransferFunctionSrgb) {
+    return ToLinearSrgb(v, bit_depth);
+  }
+  v_float = (float)v / ((1 << bit_depth) - 1);
+  switch (transfer_type) {
+    case kSharpYuvTransferFunctionBt709:
+    case kSharpYuvTransferFunctionBt601:
+    case kSharpYuvTransferFunctionBt2020_10Bit:
+    case kSharpYuvTransferFunctionBt2020_12Bit:
+      linear = ToLinear709(v_float);
+      break;
+    case kSharpYuvTransferFunctionBt470M:
+      linear = ToLinear470M(v_float);
+      break;
+    case kSharpYuvTransferFunctionBt470Bg:
+      linear = ToLinear470Bg(v_float);
+      break;
+    case kSharpYuvTransferFunctionSmpte240:
+      linear = ToLinearSmpte240(v_float);
+      break;
+    case kSharpYuvTransferFunctionLinear:
+      return v;
+    case kSharpYuvTransferFunctionLog100:
+      linear = ToLinearLog100(v_float);
+      break;
+    case kSharpYuvTransferFunctionLog100_Sqrt10:
+      linear = ToLinearLog100Sqrt10(v_float);
+      break;
+    case kSharpYuvTransferFunctionIec61966:
+      linear = ToLinearIec61966(v_float);
+      break;
+    case kSharpYuvTransferFunctionBt1361:
+      linear = ToLinearBt1361(v_float);
+      break;
+    case kSharpYuvTransferFunctionSmpte2084:
+      linear = ToLinearPq(v_float);
+      break;
+    case kSharpYuvTransferFunctionSmpte428:
+      linear = ToLinearSmpte428(v_float);
+      break;
+    case kSharpYuvTransferFunctionHlg:
+      linear = ToLinearHlg(v_float);
+      break;
+    default:
+      assert(0);
+      linear = 0;
+      break;
+  }
+  return (uint32_t)Roundf(linear * ((1 << 16) - 1));
+}
+
+uint16_t SharpYuvLinearToGamma(uint32_t v, int bit_depth,
+                               SharpYuvTransferFunctionType transfer_type) {
+  float v_float, linear;
+  if (transfer_type == kSharpYuvTransferFunctionSrgb) {
+    return FromLinearSrgb(v, bit_depth);
+  }
+  v_float = (float)v / ((1 << 16) - 1);
+  switch (transfer_type) {
+    case kSharpYuvTransferFunctionBt709:
+    case kSharpYuvTransferFunctionBt601:
+    case kSharpYuvTransferFunctionBt2020_10Bit:
+    case kSharpYuvTransferFunctionBt2020_12Bit:
+      linear = FromLinear709(v_float);
+      break;
+    case kSharpYuvTransferFunctionBt470M:
+      linear = FromLinear470M(v_float);
+      break;
+    case kSharpYuvTransferFunctionBt470Bg:
+      linear = FromLinear470Bg(v_float);
+      break;
+    case kSharpYuvTransferFunctionSmpte240:
+      linear = FromLinearSmpte240(v_float);
+      break;
+    case kSharpYuvTransferFunctionLinear:
+      return v;
+    case kSharpYuvTransferFunctionLog100:
+      linear = FromLinearLog100(v_float);
+      break;
+    case kSharpYuvTransferFunctionLog100_Sqrt10:
+      linear = FromLinearLog100Sqrt10(v_float);
+      break;
+    case kSharpYuvTransferFunctionIec61966:
+      linear = FromLinearIec61966(v_float);
+      break;
+    case kSharpYuvTransferFunctionBt1361:
+      linear = FromLinearBt1361(v_float);
+      break;
+    case kSharpYuvTransferFunctionSmpte2084:
+      linear = FromLinearPq(v_float);
+      break;
+    case kSharpYuvTransferFunctionSmpte428:
+      linear = FromLinearSmpte428(v_float);
+      break;
+    case kSharpYuvTransferFunctionHlg:
+      linear = FromLinearHlg(v_float);
+      break;
+    default:
+      assert(0);
+      linear = 0;
+      break;
+  }
+  return (uint16_t)Roundf(linear * ((1 << bit_depth) - 1));
+}
@@ -0,0 +1,38 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Gamma correction utilities.
+
+#ifndef WEBP_SHARPYUV_SHARPYUV_GAMMA_H_
+#define WEBP_SHARPYUV_SHARPYUV_GAMMA_H_
+
+#include "sharpyuv/sharpyuv.h"
+#include "src/webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Initializes precomputed tables. Must be called once before calling
+// SharpYuvGammaToLinear or SharpYuvLinearToGamma.
+void SharpYuvInitGammaTables(void);
+
+// Converts a 'bit_depth'-bit gamma color value to a 16-bit linear value.
+uint32_t SharpYuvGammaToLinear(uint16_t v, int bit_depth,
+                               SharpYuvTransferFunctionType transfer_type);
+
+// Converts a 16-bit linear color value to a 'bit_depth'-bit gamma value.
+uint16_t SharpYuvLinearToGamma(uint32_t value, int bit_depth,
+                               SharpYuvTransferFunctionType transfer_type);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // WEBP_SHARPYUV_SHARPYUV_GAMMA_H_
@@ -0,0 +1,181 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Speed-critical functions for Sharp YUV.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "sharpyuv/sharpyuv_dsp.h"
+
+#if defined(WEBP_USE_NEON)
+#include <assert.h>
+#include <stdlib.h>
+#include <arm_neon.h>
+
+static uint16_t clip_NEON(int v, int max) {
+  return (v < 0) ? 0 : (v > max) ? max : (uint16_t)v;
+}
+
+static uint64_t SharpYuvUpdateY_NEON(const uint16_t* ref, const uint16_t* src,
+                                     uint16_t* dst, int len, int bit_depth) {
+  const int max_y = (1 << bit_depth) - 1;
+  int i;
+  const int16x8_t zero = vdupq_n_s16(0);
+  const int16x8_t max = vdupq_n_s16(max_y);
+  uint64x2_t sum = vdupq_n_u64(0);
+  uint64_t diff;
+
+  for (i = 0; i + 8 <= len; i += 8) {
+    const int16x8_t A = vreinterpretq_s16_u16(vld1q_u16(ref + i));
+    const int16x8_t B = vreinterpretq_s16_u16(vld1q_u16(src + i));
+    const int16x8_t C = vreinterpretq_s16_u16(vld1q_u16(dst + i));
+    const int16x8_t D = vsubq_s16(A, B);       // diff_y
+    const int16x8_t F = vaddq_s16(C, D);       // new_y
+    const uint16x8_t H =
+        vreinterpretq_u16_s16(vmaxq_s16(vminq_s16(F, max), zero));
+    const int16x8_t I = vabsq_s16(D);          // abs(diff_y)
+    vst1q_u16(dst + i, H);
+    sum = vpadalq_u32(sum, vpaddlq_u16(vreinterpretq_u16_s16(I)));
+  }
+  diff = vgetq_lane_u64(sum, 0) + vgetq_lane_u64(sum, 1);
+  for (; i < len; ++i) {
+    const int diff_y = ref[i] - src[i];
+    const int new_y = (int)(dst[i]) + diff_y;
+    dst[i] = clip_NEON(new_y, max_y);
+    diff += (uint64_t)(abs(diff_y));
+  }
+  return diff;
+}
+
+static void SharpYuvUpdateRGB_NEON(const int16_t* ref, const int16_t* src,
+                                   int16_t* dst, int len) {
+  int i;
+  for (i = 0; i + 8 <= len; i += 8) {
+    const int16x8_t A = vld1q_s16(ref + i);
+    const int16x8_t B = vld1q_s16(src + i);
+    const int16x8_t C = vld1q_s16(dst + i);
+    const int16x8_t D = vsubq_s16(A, B);   // diff_uv
+    const int16x8_t E = vaddq_s16(C, D);   // new_uv
+    vst1q_s16(dst + i, E);
+  }
+  for (; i < len; ++i) {
+    const int diff_uv = ref[i] - src[i];
+    dst[i] += diff_uv;
+  }
+}
+
+static void SharpYuvFilterRow16_NEON(const int16_t* A, const int16_t* B,
+                                     int len, const uint16_t* best_y,
+                                     uint16_t* out, int bit_depth) {
+  const int max_y = (1 << bit_depth) - 1;
+  int i;
+  const int16x8_t max = vdupq_n_s16(max_y);
+  const int16x8_t zero = vdupq_n_s16(0);
+  for (i = 0; i + 8 <= len; i += 8) {
+    const int16x8_t a0 = vld1q_s16(A + i + 0);
+    const int16x8_t a1 = vld1q_s16(A + i + 1);
+    const int16x8_t b0 = vld1q_s16(B + i + 0);
+    const int16x8_t b1 = vld1q_s16(B + i + 1);
+    const int16x8_t a0b1 = vaddq_s16(a0, b1);
+    const int16x8_t a1b0 = vaddq_s16(a1, b0);
+    const int16x8_t a0a1b0b1 = vaddq_s16(a0b1, a1b0);  // A0+A1+B0+B1
+    const int16x8_t a0b1_2 = vaddq_s16(a0b1, a0b1);    // 2*(A0+B1)
+    const int16x8_t a1b0_2 = vaddq_s16(a1b0, a1b0);    // 2*(A1+B0)
+    const int16x8_t c0 = vshrq_n_s16(vaddq_s16(a0b1_2, a0a1b0b1), 3);
+    const int16x8_t c1 = vshrq_n_s16(vaddq_s16(a1b0_2, a0a1b0b1), 3);
+    const int16x8_t e0 = vrhaddq_s16(c1, a0);
+    const int16x8_t e1 = vrhaddq_s16(c0, a1);
+    const int16x8x2_t f = vzipq_s16(e0, e1);
+    const int16x8_t g0 = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i + 0));
+    const int16x8_t g1 = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i + 8));
+    const int16x8_t h0 = vaddq_s16(g0, f.val[0]);
+    const int16x8_t h1 = vaddq_s16(g1, f.val[1]);
+    const int16x8_t i0 = vmaxq_s16(vminq_s16(h0, max), zero);
+    const int16x8_t i1 = vmaxq_s16(vminq_s16(h1, max), zero);
+    vst1q_u16(out + 2 * i + 0, vreinterpretq_u16_s16(i0));
+    vst1q_u16(out + 2 * i + 8, vreinterpretq_u16_s16(i1));
+  }
+  for (; i < len; ++i) {
+    const int a0b1 = A[i + 0] + B[i + 1];
+    const int a1b0 = A[i + 1] + B[i + 0];
+    const int a0a1b0b1 = a0b1 + a1b0 + 8;
+    const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4;
+    const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4;
+    out[2 * i + 0] = clip_NEON(best_y[2 * i + 0] + v0, max_y);
+    out[2 * i + 1] = clip_NEON(best_y[2 * i + 1] + v1, max_y);
+  }
+}
+
+static void SharpYuvFilterRow32_NEON(const int16_t* A, const int16_t* B,
+                                     int len, const uint16_t* best_y,
+                                     uint16_t* out, int bit_depth) {
+  const int max_y = (1 << bit_depth) - 1;
+  int i;
+  const uint16x8_t max = vdupq_n_u16(max_y);
+  for (i = 0; i + 4 <= len; i += 4) {
+    const int16x4_t a0 = vld1_s16(A + i + 0);
+    const int16x4_t a1 = vld1_s16(A + i + 1);
+    const int16x4_t b0 = vld1_s16(B + i + 0);
+    const int16x4_t b1 = vld1_s16(B + i + 1);
+    const int32x4_t a0b1 = vaddl_s16(a0, b1);
+    const int32x4_t a1b0 = vaddl_s16(a1, b0);
+    const int32x4_t a0a1b0b1 = vaddq_s32(a0b1, a1b0);  // A0+A1+B0+B1
+    const int32x4_t a0b1_2 = vaddq_s32(a0b1, a0b1);    // 2*(A0+B1)
+    const int32x4_t a1b0_2 = vaddq_s32(a1b0, a1b0);    // 2*(A1+B0)
+    const int32x4_t c0 = vshrq_n_s32(vaddq_s32(a0b1_2, a0a1b0b1), 3);
+    const int32x4_t c1 = vshrq_n_s32(vaddq_s32(a1b0_2, a0a1b0b1), 3);
+    const int32x4_t e0 = vrhaddq_s32(c1, vmovl_s16(a0));
+    const int32x4_t e1 = vrhaddq_s32(c0, vmovl_s16(a1));
+    const int32x4x2_t f = vzipq_s32(e0, e1);
+
+    const int16x8_t g = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i));
+    const int32x4_t h0 = vaddw_s16(f.val[0], vget_low_s16(g));
+    const int32x4_t h1 = vaddw_s16(f.val[1], vget_high_s16(g));
+    const uint16x8_t i_16 = vcombine_u16(vqmovun_s32(h0), vqmovun_s32(h1));
+    const uint16x8_t i_clamped = vminq_u16(i_16, max);
+    vst1q_u16(out + 2 * i + 0, i_clamped);
+  }
+  for (; i < len; ++i) {
+    const int a0b1 = A[i + 0] + B[i + 1];
+    const int a1b0 = A[i + 1] + B[i + 0];
+    const int a0a1b0b1 = a0b1 + a1b0 + 8;
+    const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4;
+    const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4;
+    out[2 * i + 0] = clip_NEON(best_y[2 * i + 0] + v0, max_y);
+    out[2 * i + 1] = clip_NEON(best_y[2 * i + 1] + v1, max_y);
+  }
+}
+
+static void SharpYuvFilterRow_NEON(const int16_t* A, const int16_t* B, int len,
+                                   const uint16_t* best_y, uint16_t* out,
+                                   int bit_depth) {
+  if (bit_depth <= 10) {
+    SharpYuvFilterRow16_NEON(A, B, len, best_y, out, bit_depth);
+  } else {
+    SharpYuvFilterRow32_NEON(A, B, len, best_y, out, bit_depth);
+  }
+}
+
+//------------------------------------------------------------------------------
+
+extern void InitSharpYuvNEON(void);
+
+WEBP_TSAN_IGNORE_FUNCTION void InitSharpYuvNEON(void) {
+  SharpYuvUpdateY = SharpYuvUpdateY_NEON;
+  SharpYuvUpdateRGB = SharpYuvUpdateRGB_NEON;
+  SharpYuvFilterRow = SharpYuvFilterRow_NEON;
+}
+
+#else  // !WEBP_USE_NEON
+
+extern void InitSharpYuvNEON(void);
+
+void InitSharpYuvNEON(void) {}
+
+#endif  // WEBP_USE_NEON
@@ -0,0 +1,201 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Speed-critical functions for Sharp YUV.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "sharpyuv/sharpyuv_dsp.h"
+
+#if defined(WEBP_USE_SSE2)
+#include <stdlib.h>
+#include <emmintrin.h>
+
+static uint16_t clip_SSE2(int v, int max) {
+  return (v < 0) ? 0 : (v > max) ? max : (uint16_t)v;
+}
+
+static uint64_t SharpYuvUpdateY_SSE2(const uint16_t* ref, const uint16_t* src,
+                                     uint16_t* dst, int len, int bit_depth) {
+  const int max_y = (1 << bit_depth) - 1;
+  uint64_t diff = 0;
+  uint32_t tmp[4];
+  int i;
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i max = _mm_set1_epi16(max_y);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i sum = zero;
+
+  for (i = 0; i + 8 <= len; i += 8) {
+    const __m128i A = _mm_loadu_si128((const __m128i*)(ref + i));
+    const __m128i B = _mm_loadu_si128((const __m128i*)(src + i));
+    const __m128i C = _mm_loadu_si128((const __m128i*)(dst + i));
+    const __m128i D = _mm_sub_epi16(A, B);       // diff_y
+    const __m128i E = _mm_cmpgt_epi16(zero, D);  // sign (-1 or 0)
+    const __m128i F = _mm_add_epi16(C, D);       // new_y
+    const __m128i G = _mm_or_si128(E, one);      // -1 or 1
+    const __m128i H = _mm_max_epi16(_mm_min_epi16(F, max), zero);
+    const __m128i I = _mm_madd_epi16(D, G);      // sum(abs(...))
+    _mm_storeu_si128((__m128i*)(dst + i), H);
+    sum = _mm_add_epi32(sum, I);
+  }
+  _mm_storeu_si128((__m128i*)tmp, sum);
+  diff = tmp[3] + tmp[2] + tmp[1] + tmp[0];
+  for (; i < len; ++i) {
+    const int diff_y = ref[i] - src[i];
+    const int new_y = (int)dst[i] + diff_y;
+    dst[i] = clip_SSE2(new_y, max_y);
+    diff += (uint64_t)abs(diff_y);
+  }
+  return diff;
+}
+
+static void SharpYuvUpdateRGB_SSE2(const int16_t* ref, const int16_t* src,
+                                   int16_t* dst, int len) {
+  int i = 0;
+  for (i = 0; i + 8 <= len; i += 8) {
+    const __m128i A = _mm_loadu_si128((const __m128i*)(ref + i));
+    const __m128i B = _mm_loadu_si128((const __m128i*)(src + i));
+    const __m128i C = _mm_loadu_si128((const __m128i*)(dst + i));
+    const __m128i D = _mm_sub_epi16(A, B);   // diff_uv
+    const __m128i E = _mm_add_epi16(C, D);   // new_uv
+    _mm_storeu_si128((__m128i*)(dst + i), E);
+  }
+  for (; i < len; ++i) {
+    const int diff_uv = ref[i] - src[i];
+    dst[i] += diff_uv;
+  }
+}
+
+static void SharpYuvFilterRow16_SSE2(const int16_t* A, const int16_t* B,
+                                     int len, const uint16_t* best_y,
+                                     uint16_t* out, int bit_depth) {
+  const int max_y = (1 << bit_depth) - 1;
+  int i;
+  const __m128i kCst8 = _mm_set1_epi16(8);
+  const __m128i max = _mm_set1_epi16(max_y);
+  const __m128i zero = _mm_setzero_si128();
+  for (i = 0; i + 8 <= len; i += 8) {
+    const __m128i a0 = _mm_loadu_si128((const __m128i*)(A + i + 0));
+    const __m128i a1 = _mm_loadu_si128((const __m128i*)(A + i + 1));
+    const __m128i b0 = _mm_loadu_si128((const __m128i*)(B + i + 0));
+    const __m128i b1 = _mm_loadu_si128((const __m128i*)(B + i + 1));
+    const __m128i a0b1 = _mm_add_epi16(a0, b1);
+    const __m128i a1b0 = _mm_add_epi16(a1, b0);
+    const __m128i a0a1b0b1 = _mm_add_epi16(a0b1, a1b0);  // A0+A1+B0+B1
+    const __m128i a0a1b0b1_8 = _mm_add_epi16(a0a1b0b1, kCst8);
+    const __m128i a0b1_2 = _mm_add_epi16(a0b1, a0b1);    // 2*(A0+B1)
+    const __m128i a1b0_2 = _mm_add_epi16(a1b0, a1b0);    // 2*(A1+B0)
+    const __m128i c0 = _mm_srai_epi16(_mm_add_epi16(a0b1_2, a0a1b0b1_8), 3);
+    const __m128i c1 = _mm_srai_epi16(_mm_add_epi16(a1b0_2, a0a1b0b1_8), 3);
+    const __m128i d0 = _mm_add_epi16(c1, a0);
+    const __m128i d1 = _mm_add_epi16(c0, a1);
+    const __m128i e0 = _mm_srai_epi16(d0, 1);
+    const __m128i e1 = _mm_srai_epi16(d1, 1);
+    const __m128i f0 = _mm_unpacklo_epi16(e0, e1);
+    const __m128i f1 = _mm_unpackhi_epi16(e0, e1);
+    const __m128i g0 = _mm_loadu_si128((const __m128i*)(best_y + 2 * i + 0));
+    const __m128i g1 = _mm_loadu_si128((const __m128i*)(best_y + 2 * i + 8));
+    const __m128i h0 = _mm_add_epi16(g0, f0);
+    const __m128i h1 = _mm_add_epi16(g1, f1);
+    const __m128i i0 = _mm_max_epi16(_mm_min_epi16(h0, max), zero);
+    const __m128i i1 = _mm_max_epi16(_mm_min_epi16(h1, max), zero);
+    _mm_storeu_si128((__m128i*)(out + 2 * i + 0), i0);
+    _mm_storeu_si128((__m128i*)(out + 2 * i + 8), i1);
+  }
+  for (; i < len; ++i) {
+    //   (9 * A0 + 3 * A1 + 3 * B0 + B1 + 8) >> 4 =
+    // = (8 * A0 + 2 * (A1 + B0) + (A0 + A1 + B0 + B1 + 8)) >> 4
+    // We reuse the common sub-expressions.
+    const int a0b1 = A[i + 0] + B[i + 1];
+    const int a1b0 = A[i + 1] + B[i + 0];
+    const int a0a1b0b1 = a0b1 + a1b0 + 8;
+    const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4;
+    const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4;
+    out[2 * i + 0] = clip_SSE2(best_y[2 * i + 0] + v0, max_y);
+    out[2 * i + 1] = clip_SSE2(best_y[2 * i + 1] + v1, max_y);
+  }
+}
+
+static WEBP_INLINE __m128i s16_to_s32(__m128i in) {
+  return _mm_srai_epi32(_mm_unpacklo_epi16(in, in), 16);
+}
+
+static void SharpYuvFilterRow32_SSE2(const int16_t* A, const int16_t* B,
+                                     int len, const uint16_t* best_y,
+                                     uint16_t* out, int bit_depth) {
+  const int max_y = (1 << bit_depth) - 1;
+  int i;
+  const __m128i kCst8 = _mm_set1_epi32(8);
+  const __m128i max = _mm_set1_epi16(max_y);
+  const __m128i zero = _mm_setzero_si128();
+  for (i = 0; i + 4 <= len; i += 4) {
+    const __m128i a0 = s16_to_s32(_mm_loadl_epi64((const __m128i*)(A + i + 0)));
+    const __m128i a1 = s16_to_s32(_mm_loadl_epi64((const __m128i*)(A + i + 1)));
+    const __m128i b0 = s16_to_s32(_mm_loadl_epi64((const __m128i*)(B + i + 0)));
+    const __m128i b1 = s16_to_s32(_mm_loadl_epi64((const __m128i*)(B + i + 1)));
+    const __m128i a0b1 = _mm_add_epi32(a0, b1);
+    const __m128i a1b0 = _mm_add_epi32(a1, b0);
+    const __m128i a0a1b0b1 = _mm_add_epi32(a0b1, a1b0);  // A0+A1+B0+B1
+    const __m128i a0a1b0b1_8 = _mm_add_epi32(a0a1b0b1, kCst8);
+    const __m128i a0b1_2 = _mm_add_epi32(a0b1, a0b1);  // 2*(A0+B1)
+    const __m128i a1b0_2 = _mm_add_epi32(a1b0, a1b0);  // 2*(A1+B0)
+    const __m128i c0 = _mm_srai_epi32(_mm_add_epi32(a0b1_2, a0a1b0b1_8), 3);
+    const __m128i c1 = _mm_srai_epi32(_mm_add_epi32(a1b0_2, a0a1b0b1_8), 3);
+    const __m128i d0 = _mm_add_epi32(c1, a0);
+    const __m128i d1 = _mm_add_epi32(c0, a1);
+    const __m128i e0 = _mm_srai_epi32(d0, 1);
+    const __m128i e1 = _mm_srai_epi32(d1, 1);
+    const __m128i f0 = _mm_unpacklo_epi32(e0, e1);
+    const __m128i f1 = _mm_unpackhi_epi32(e0, e1);
+    const __m128i g = _mm_loadu_si128((const __m128i*)(best_y + 2 * i + 0));
+    const __m128i h_16 = _mm_add_epi16(g, _mm_packs_epi32(f0, f1));
+    const __m128i final = _mm_max_epi16(_mm_min_epi16(h_16, max), zero);
+    _mm_storeu_si128((__m128i*)(out + 2 * i + 0), final);
+  }
+  for (; i < len; ++i) {
+    //   (9 * A0 + 3 * A1 + 3 * B0 + B1 + 8) >> 4 =
+    // = (8 * A0 + 2 * (A1 + B0) + (A0 + A1 + B0 + B1 + 8)) >> 4
+    // We reuse the common sub-expressions.
+    const int a0b1 = A[i + 0] + B[i + 1];
+    const int a1b0 = A[i + 1] + B[i + 0];
+    const int a0a1b0b1 = a0b1 + a1b0 + 8;
+    const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4;
+    const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4;
+    out[2 * i + 0] = clip_SSE2(best_y[2 * i + 0] + v0, max_y);
+    out[2 * i + 1] = clip_SSE2(best_y[2 * i + 1] + v1, max_y);
+  }
+}
+
+static void SharpYuvFilterRow_SSE2(const int16_t* A, const int16_t* B, int len,
+                                   const uint16_t* best_y, uint16_t* out,
+                                   int bit_depth) {
+  if (bit_depth <= 10) {
+    SharpYuvFilterRow16_SSE2(A, B, len, best_y, out, bit_depth);
+  } else {
+    SharpYuvFilterRow32_SSE2(A, B, len, best_y, out, bit_depth);
+  }
+}
+
+//------------------------------------------------------------------------------
+
+extern void InitSharpYuvSSE2(void);
+
+WEBP_TSAN_IGNORE_FUNCTION void InitSharpYuvSSE2(void) {
+  SharpYuvUpdateY = SharpYuvUpdateY_SSE2;
+  SharpYuvUpdateRGB = SharpYuvUpdateRGB_SSE2;
+  SharpYuvFilterRow = SharpYuvFilterRow_SSE2;
+}
+#else  // !WEBP_USE_SSE2
+
+extern void InitSharpYuvSSE2(void);
+
+void InitSharpYuvSSE2(void) {}
+
+#endif  // WEBP_USE_SSE2
@@ -117,21 +117,12 @@ static int ALPHDecode(VP8Decoder* const dec, int row, int num_rows) {
    const uint8_t* deltas = dec->alpha_data_ + ALPHA_HEADER_LEN + row * width;
    uint8_t* dst = dec->alpha_plane_ + row * width;
    assert(deltas <= &dec->alpha_data_[dec->alpha_data_size_]);
-    if (alph_dec->filter_ != WEBP_FILTER_NONE) {
-      assert(WebPUnfilters[alph_dec->filter_] != NULL);
-      for (y = 0; y < num_rows; ++y) {
-        WebPUnfilters[alph_dec->filter_](prev_line, deltas, dst, width);
-        prev_line = dst;
-        dst += width;
-        deltas += width;
-      }
-    } else {
-      for (y = 0; y < num_rows; ++y) {
-        memcpy(dst, deltas, width * sizeof(*dst));
-        prev_line = dst;
-        dst += width;
-        deltas += width;
-      }
+    assert(WebPUnfilters[alph_dec->filter_] != NULL);
+    for (y = 0; y < num_rows; ++y) {
+      WebPUnfilters[alph_dec->filter_](prev_line, deltas, dst, width);
+      prev_line = dst;
+      dst += width;
+      deltas += width;
    }
    dec->alpha_prev_line_ = prev_line;
  } else {  // alph_dec->method_ == ALPHA_LOSSLESS_COMPRESSION
@@ -155,7 +146,8 @@ static int AllocateAlphaPlane(VP8Decoder* const dec, const VP8Io* const io) {
  dec->alpha_plane_mem_ =
      (uint8_t*)WebPSafeMalloc(alpha_size, sizeof(*dec->alpha_plane_));
  if (dec->alpha_plane_mem_ == NULL) {
-    return 0;
+    return VP8SetError(dec, VP8_STATUS_OUT_OF_MEMORY,
+                       "Alpha decoder initialization failed.");
  }
  dec->alpha_plane_ = dec->alpha_plane_mem_;
  dec->alpha_prev_line_ = NULL;
@@ -183,16 +175,25 @@ const uint8_t* VP8DecompressAlphaRows(VP8Decoder* const dec,
  assert(dec != NULL && io != NULL);

  if (row < 0 || num_rows <= 0 || row + num_rows > height) {
-    return NULL;    // sanity check.
+    return NULL;
  }

  if (!dec->is_alpha_decoded_) {
    if (dec->alph_dec_ == NULL) {    // Initialize decoder.
      dec->alph_dec_ = ALPHNew();
-      if (dec->alph_dec_ == NULL) return NULL;
+      if (dec->alph_dec_ == NULL) {
+        VP8SetError(dec, VP8_STATUS_OUT_OF_MEMORY,
+                    "Alpha decoder initialization failed.");
+        return NULL;
+      }
      if (!AllocateAlphaPlane(dec, io)) goto Error;
      if (!ALPHInit(dec->alph_dec_, dec->alpha_data_, dec->alpha_data_size_,
                    io, dec->alpha_plane_)) {
+        VP8LDecoder* const vp8l_dec = dec->alph_dec_->vp8l_dec_;
+        VP8SetError(dec,
+                    (vp8l_dec == NULL) ? VP8_STATUS_OUT_OF_MEMORY
+                                       : vp8l_dec->status_,
+                    "Alpha decoder initialization failed.");
        goto Error;
      }
      // if we allowed use of alpha dithering, check whether it's needed at all
@@ -75,7 +75,7 @@ static VP8StatusCode CheckDecBuffer(const WebPDecBuffer* const buffer) {
    const WebPRGBABuffer* const buf = &buffer->u.RGBA;
    const int stride = abs(buf->stride);
    const uint64_t size =
-        MIN_BUFFER_SIZE(width * kModeBpp[mode], height, stride);
+        MIN_BUFFER_SIZE((uint64_t)width * kModeBpp[mode], height, stride);
    ok &= (size <= buf->size);
    ok &= (stride >= width * kModeBpp[mode]);
    ok &= (buf->rgba != NULL);
@@ -102,7 +102,7 @@ static VP8StatusCode AllocateBuffer(WebPDecBuffer* const buffer) {
    int stride;
    uint64_t size;

-    if ((uint64_t)w * kModeBpp[mode] >= (1ull << 32)) {
+    if ((uint64_t)w * kModeBpp[mode] >= (1ull << 31)) {
      return VP8_STATUS_INVALID_PARAM;
    }
    stride = w * kModeBpp[mode];
@@ -117,7 +117,6 @@ static VP8StatusCode AllocateBuffer(WebPDecBuffer* const buffer) {
    }
    total_size = size + 2 * uv_size + a_size;

-    // Security/sanity checks
    output = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*output));
    if (output == NULL) {
      return VP8_STATUS_OUT_OF_MEMORY;
@@ -156,11 +155,11 @@ VP8StatusCode WebPFlipBuffer(WebPDecBuffer* const buffer) {
  }
  if (WebPIsRGBMode(buffer->colorspace)) {
    WebPRGBABuffer* const buf = &buffer->u.RGBA;
-    buf->rgba += (buffer->height - 1) * buf->stride;
+    buf->rgba += (int64_t)(buffer->height - 1) * buf->stride;
    buf->stride = -buf->stride;
  } else {
    WebPYUVABuffer* const buf = &buffer->u.YUVA;
-    const int H = buffer->height;
+    const int64_t H = buffer->height;
    buf->y += (H - 1) * buf->y_stride;
    buf->y_stride = -buf->y_stride;
    buf->u += ((H - 1) >> 1) * buf->u_stride;
@@ -188,8 +187,7 @@ VP8StatusCode WebPAllocateDecBuffer(int width, int height,
      const int ch = options->crop_height;
      const int x = options->crop_left & ~1;
      const int y = options->crop_top & ~1;
-      if (x < 0 || y < 0 || cw <= 0 || ch <= 0 ||
-          x + cw > width || y + ch > height) {
+      if (!WebPCheckCropDimensions(width, height, x, y, cw, ch)) {
        return VP8_STATUS_INVALID_PARAM;   // out of frame boundary.
      }
      width = cw;
@@ -705,7 +705,7 @@ static int AllocateMemory(VP8Decoder* const dec) {
                        + cache_size + alpha_size + WEBP_ALIGN_CST;
  uint8_t* mem;

-  if (needed != (size_t)needed) return 0;  // check for overflow
+  if (!CheckSizeOverflow(needed)) return 0;  // check for overflow
  if (needed > dec->mem_size_) {
    WebPSafeFree(dec->mem_);
    dec->mem_size_ = 0;
@@ -298,46 +298,57 @@ static int InitYUVRescaler(const VP8Io* const io, WebPDecParams* const p) {
  const int uv_out_height = (out_height + 1) >> 1;
  const int uv_in_width  = (io->mb_w + 1) >> 1;
  const int uv_in_height = (io->mb_h + 1) >> 1;
-  const size_t work_size = 2 * out_width;   // scratch memory for luma rescaler
+  // scratch memory for luma rescaler
+  const size_t work_size = 2 * (size_t)out_width;
  const size_t uv_work_size = 2 * uv_out_width;  // and for each u/v ones
-  size_t tmp_size, rescaler_size;
+  uint64_t total_size;
+  size_t rescaler_size;
  rescaler_t* work;
  WebPRescaler* scalers;
  const int num_rescalers = has_alpha ? 4 : 3;

-  tmp_size = (work_size + 2 * uv_work_size) * sizeof(*work);
+  total_size = ((uint64_t)work_size + 2 * uv_work_size) * sizeof(*work);
  if (has_alpha) {
-    tmp_size += work_size * sizeof(*work);
+    total_size += (uint64_t)work_size * sizeof(*work);
  }
  rescaler_size = num_rescalers * sizeof(*p->scaler_y) + WEBP_ALIGN_CST;
+  total_size += rescaler_size;
+  if (!CheckSizeOverflow(total_size)) {
+    return 0;
+  }

-  p->memory = WebPSafeMalloc(1ULL, tmp_size + rescaler_size);
+  p->memory = WebPSafeMalloc(1ULL, (size_t)total_size);
  if (p->memory == NULL) {
    return 0;   // memory error
  }
  work = (rescaler_t*)p->memory;

-  scalers = (WebPRescaler*)WEBP_ALIGN((const uint8_t*)work + tmp_size);
+  scalers = (WebPRescaler*)WEBP_ALIGN(
+      (const uint8_t*)work + total_size - rescaler_size);
  p->scaler_y = &scalers[0];
  p->scaler_u = &scalers[1];
  p->scaler_v = &scalers[2];
  p->scaler_a = has_alpha ? &scalers[3] : NULL;

-  WebPRescalerInit(p->scaler_y, io->mb_w, io->mb_h,
-                   buf->y, out_width, out_height, buf->y_stride, 1,
-                   work);
-  WebPRescalerInit(p->scaler_u, uv_in_width, uv_in_height,
-                   buf->u, uv_out_width, uv_out_height, buf->u_stride, 1,
-                   work + work_size);
-  WebPRescalerInit(p->scaler_v, uv_in_width, uv_in_height,
-                   buf->v, uv_out_width, uv_out_height, buf->v_stride, 1,
-                   work + work_size + uv_work_size);
+  if (!WebPRescalerInit(p->scaler_y, io->mb_w, io->mb_h,
+                        buf->y, out_width, out_height, buf->y_stride, 1,
+                        work) ||
+      !WebPRescalerInit(p->scaler_u, uv_in_width, uv_in_height,
+                        buf->u, uv_out_width, uv_out_height, buf->u_stride, 1,
+                        work + work_size) ||
+      !WebPRescalerInit(p->scaler_v, uv_in_width, uv_in_height,
+                        buf->v, uv_out_width, uv_out_height, buf->v_stride, 1,
+                        work + work_size + uv_work_size)) {
+    return 0;
+  }
  p->emit = EmitRescaledYUV;

  if (has_alpha) {
-    WebPRescalerInit(p->scaler_a, io->mb_w, io->mb_h,
-                     buf->a, out_width, out_height, buf->a_stride, 1,
-                     work + work_size + 2 * uv_work_size);
+    if (!WebPRescalerInit(p->scaler_a, io->mb_w, io->mb_h,
+                          buf->a, out_width, out_height, buf->a_stride, 1,
+                          work + work_size + 2 * uv_work_size)) {
+      return 0;
+    }
    p->emit_alpha = EmitRescaledAlphaYUV;
    WebPInitAlphaProcessing();
  }
@@ -480,51 +491,58 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
  const int out_height = io->scaled_height;
  const int uv_in_width  = (io->mb_w + 1) >> 1;
  const int uv_in_height = (io->mb_h + 1) >> 1;
-  const size_t work_size = 2 * out_width;   // scratch memory for one rescaler
+  // scratch memory for one rescaler
+  const size_t work_size = 2 * (size_t)out_width;
  rescaler_t* work;  // rescalers work area
  uint8_t* tmp;   // tmp storage for scaled YUV444 samples before RGB conversion
-  size_t tmp_size1, tmp_size2, total_size, rescaler_size;
+  uint64_t tmp_size1, tmp_size2, total_size;
+  size_t rescaler_size;
  WebPRescaler* scalers;
  const int num_rescalers = has_alpha ? 4 : 3;

-  tmp_size1 = 3 * work_size;
-  tmp_size2 = 3 * out_width;
-  if (has_alpha) {
-    tmp_size1 += work_size;
-    tmp_size2 += out_width;
-  }
+  tmp_size1 = (uint64_t)num_rescalers * work_size;
+  tmp_size2 = (uint64_t)num_rescalers * out_width;
  total_size = tmp_size1 * sizeof(*work) + tmp_size2 * sizeof(*tmp);
  rescaler_size = num_rescalers * sizeof(*p->scaler_y) + WEBP_ALIGN_CST;
+  total_size += rescaler_size;
+  if (!CheckSizeOverflow(total_size)) {
+    return 0;
+  }

-  p->memory = WebPSafeMalloc(1ULL, total_size + rescaler_size);
+  p->memory = WebPSafeMalloc(1ULL, (size_t)total_size);
  if (p->memory == NULL) {
    return 0;   // memory error
  }
  work = (rescaler_t*)p->memory;
  tmp = (uint8_t*)(work + tmp_size1);

-  scalers = (WebPRescaler*)WEBP_ALIGN((const uint8_t*)work + total_size);
+  scalers = (WebPRescaler*)WEBP_ALIGN(
+      (const uint8_t*)work + total_size - rescaler_size);
  p->scaler_y = &scalers[0];
  p->scaler_u = &scalers[1];
  p->scaler_v = &scalers[2];
  p->scaler_a = has_alpha ? &scalers[3] : NULL;

-  WebPRescalerInit(p->scaler_y, io->mb_w, io->mb_h,
-                   tmp + 0 * out_width, out_width, out_height, 0, 1,
-                   work + 0 * work_size);
-  WebPRescalerInit(p->scaler_u, uv_in_width, uv_in_height,
-                   tmp + 1 * out_width, out_width, out_height, 0, 1,
-                   work + 1 * work_size);
-  WebPRescalerInit(p->scaler_v, uv_in_width, uv_in_height,
-                   tmp + 2 * out_width, out_width, out_height, 0, 1,
-                   work + 2 * work_size);
+  if (!WebPRescalerInit(p->scaler_y, io->mb_w, io->mb_h,
+                        tmp + 0 * out_width, out_width, out_height, 0, 1,
+                        work + 0 * work_size) ||
+      !WebPRescalerInit(p->scaler_u, uv_in_width, uv_in_height,
+                        tmp + 1 * out_width, out_width, out_height, 0, 1,
+                        work + 1 * work_size) ||
+      !WebPRescalerInit(p->scaler_v, uv_in_width, uv_in_height,
+                        tmp + 2 * out_width, out_width, out_height, 0, 1,
+                        work + 2 * work_size)) {
+    return 0;
+  }
  p->emit = EmitRescaledRGB;
  WebPInitYUV444Converters();

  if (has_alpha) {
-    WebPRescalerInit(p->scaler_a, io->mb_w, io->mb_h,
-                     tmp + 3 * out_width, out_width, out_height, 0, 1,
-                     work + 3 * work_size);
+    if (!WebPRescalerInit(p->scaler_a, io->mb_w, io->mb_h,
+                          tmp + 3 * out_width, out_width, out_height, 0, 1,
+                          work + 3 * work_size)) {
+      return 0;
+    }
    p->emit_alpha = EmitRescaledAlphaRGB;
    if (p->output->colorspace == MODE_RGBA_4444 ||
        p->output->colorspace == MODE_rgbA_4444) {
@@ -12,10 +12,11 @@
 // Author: Skal (pascal.massimino@gmail.com)

 #include "src/dec/vp8i_dec.h"
+#include "src/dsp/cpu.h"
 #include "src/utils/bit_reader_inl_utils.h"

 #if !defined(USE_GENERIC_TREE)
-#if !defined(__arm__) && !defined(_M_ARM) && !defined(__aarch64__)
+#if !defined(__arm__) && !defined(_M_ARM) && !WEBP_AARCH64
 // using a table is ~1-2% slower on ARM. Prefer the coded-tree approach then.
 #define USE_GENERIC_TREE 1   // ALTERNATE_CODE
 #else
@@ -335,7 +335,7 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
    io->scaled_width = io->width;
    io->scaled_height = io->height;

-    io->mb_w = io->width;   // sanity check
+    io->mb_w = io->width;   // for soundness
    io->mb_h = io->height;  // ditto

    VP8ResetProba(&dec->proba_);
@@ -403,7 +403,7 @@ static const uint8_t kZigzag[16] = {
  0, 1, 4, 8,  5, 2, 3, 6,  9, 12, 13, 10,  7, 11, 14, 15
 };

-// See section 13-2: http://tools.ietf.org/html/rfc6386#section-13.2
+// See section 13-2: https://datatracker.ietf.org/doc/html/rfc6386#section-13.2
 static int GetLargeValue(VP8BitReader* const br, const uint8_t* const p) {
  int v;
  if (!VP8GetBit(br, p[3], "coeffs")) {
@@ -494,6 +494,8 @@ static int GetCoeffsAlt(VP8BitReader* const br,
  return 16;
 }

+extern VP8CPUInfo VP8GetCPUInfo;
+
 WEBP_DSP_INIT_FUNC(InitGetCoeffs) {
  if (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kSlowSSSE3)) {
    GetCoeffs = GetCoeffsAlt;
@@ -31,8 +31,8 @@ extern "C" {

 // version numbers
 #define DEC_MAJ_VERSION 1
-#define DEC_MIN_VERSION 2
-#define DEC_REV_VERSION 0
+#define DEC_MIN_VERSION 3
+#define DEC_REV_VERSION 1

 // YUV-cache parameters. Cache is 32-bytes wide (= one cacheline).
 // Constraints are: We need to store one 16x16 block of luma samples (y),
@@ -12,6 +12,7 @@
 // Authors: Vikas Arora (vikaas.arora@gmail.com)
 //          Jyrki Alakuijala (jyrki@google.com)

+#include <assert.h>
 #include <stdlib.h>

 #include "src/dec/alphai_dec.h"
@@ -84,7 +85,7 @@ static const uint8_t kCodeToPlane[CODE_TO_PLANE_CODES] = {
 // to 256 (green component values) + 24 (length prefix values)
 // + color_cache_size (between 0 and 2048).
 // All values computed for 8-bit first level lookup with Mark Adler's tool:
-// http://www.hdfgroup.org/ftp/lib-external/zlib/zlib-1.2.5/examples/enough.c
+// https://github.com/madler/zlib/blob/v1.2.5/examples/enough.c
 #define FIXED_TABLE_SIZE (630 * 3 + 410)
 static const uint16_t kTableSize[12] = {
  FIXED_TABLE_SIZE + 654,
@@ -101,6 +102,14 @@ static const uint16_t kTableSize[12] = {
  FIXED_TABLE_SIZE + 2704
 };

+static int VP8LSetError(VP8LDecoder* const dec, VP8StatusCode error) {
+  // The oldest error reported takes precedence over the new one.
+  if (dec->status_ == VP8_STATUS_OK || dec->status_ == VP8_STATUS_SUSPENDED) {
+    dec->status_ = error;
+  }
+  return 0;
+}
+
 static int DecodeImageStream(int xsize, int ysize,
                             int is_level0,
                             VP8LDecoder* const dec,
@@ -178,7 +187,7 @@ static WEBP_INLINE int PlaneCodeToDistance(int xsize, int plane_code) {

 //------------------------------------------------------------------------------
 // Decodes the next Huffman code from bit-stream.
-// FillBitWindow(br) needs to be called at minimum every second call
+// VP8LFillBitWindow(br) needs to be called at minimum every second call
 // to ReadSymbol, in order to pre-fetch enough bits.
 static WEBP_INLINE int ReadSymbol(const HuffmanCode* table,
                                  VP8LBitReader* const br) {
@@ -253,11 +262,11 @@ static int ReadHuffmanCodeLengths(
  int symbol;
  int max_symbol;
  int prev_code_len = DEFAULT_CODE_LENGTH;
-  HuffmanCode table[1 << LENGTHS_TABLE_BITS];
+  HuffmanTables tables;

-  if (!VP8LBuildHuffmanTable(table, LENGTHS_TABLE_BITS,
-                             code_length_code_lengths,
-                             NUM_CODE_LENGTH_CODES)) {
+  if (!VP8LHuffmanTablesAllocate(1 << LENGTHS_TABLE_BITS, &tables) ||
+      !VP8LBuildHuffmanTable(&tables, LENGTHS_TABLE_BITS,
+                             code_length_code_lengths, NUM_CODE_LENGTH_CODES)) {
    goto End;
  }

@@ -277,7 +286,7 @@ static int ReadHuffmanCodeLengths(
    int code_len;
    if (max_symbol-- == 0) break;
    VP8LFillBitWindow(br);
-    p = &table[VP8LPrefetchBits(br) & LENGTHS_TABLE_MASK];
+    p = &tables.curr_segment->start[VP8LPrefetchBits(br) & LENGTHS_TABLE_MASK];
    VP8LSetBitPos(br, br->bit_pos_ + p->bits);
    code_len = p->value;
    if (code_len < kCodeLengthLiterals) {
@@ -300,14 +309,16 @@ static int ReadHuffmanCodeLengths(
  ok = 1;

 End:
-  if (!ok) dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
+  VP8LHuffmanTablesDeallocate(&tables);
+  if (!ok) return VP8LSetError(dec, VP8_STATUS_BITSTREAM_ERROR);
  return ok;
 }

 // 'code_lengths' is pre-allocated temporary buffer, used for creating Huffman
 // tree.
 static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
-                           int* const code_lengths, HuffmanCode* const table) {
+                           int* const code_lengths,
+                           HuffmanTables* const table) {
  int ok = 0;
  int size = 0;
  VP8LBitReader* const br = &dec->br_;
@@ -321,7 +332,7 @@ static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
    // The first code is either 1 bit or 8 bit code.
    int symbol = VP8LReadBits(br, (first_symbol_len_code == 0) ? 1 : 8);
    code_lengths[symbol] = 1;
-    // The second code (if present), is always 8 bit long.
+    // The second code (if present), is always 8 bits long.
    if (num_symbols == 2) {
      symbol = VP8LReadBits(br, 8);
      code_lengths[symbol] = 1;
@@ -331,10 +342,7 @@ static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
    int i;
    int code_length_code_lengths[NUM_CODE_LENGTH_CODES] = { 0 };
    const int num_codes = VP8LReadBits(br, 4) + 4;
-    if (num_codes > NUM_CODE_LENGTH_CODES) {
-      dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
-      return 0;
-    }
+    assert(num_codes <= NUM_CODE_LENGTH_CODES);

    for (i = 0; i < num_codes; ++i) {
      code_length_code_lengths[kCodeLengthCodeOrder[i]] = VP8LReadBits(br, 3);
@@ -349,36 +357,35 @@ static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
                                 code_lengths, alphabet_size);
  }
  if (!ok || size == 0) {
-    dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
-    return 0;
+    return VP8LSetError(dec, VP8_STATUS_BITSTREAM_ERROR);
  }
  return size;
 }

 static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
                            int color_cache_bits, int allow_recursion) {
-  int i, j;
+  int i;
  VP8LBitReader* const br = &dec->br_;
  VP8LMetadata* const hdr = &dec->hdr_;
  uint32_t* huffman_image = NULL;
  HTreeGroup* htree_groups = NULL;
-  HuffmanCode* huffman_tables = NULL;
-  HuffmanCode* huffman_table = NULL;
+  HuffmanTables* huffman_tables = &hdr->huffman_tables_;
  int num_htree_groups = 1;
  int num_htree_groups_max = 1;
-  int max_alphabet_size = 0;
-  int* code_lengths = NULL;
-  const int table_size = kTableSize[color_cache_bits];
  int* mapping = NULL;
  int ok = 0;

+  // Check the table has been 0 initialized (through InitMetadata).
+  assert(huffman_tables->root.start == NULL);
+  assert(huffman_tables->curr_segment == NULL);
+
  if (allow_recursion && VP8LReadBits(br, 1)) {
    // use meta Huffman codes.
    const int huffman_precision = VP8LReadBits(br, 3) + 2;
    const int huffman_xsize = VP8LSubSampleSize(xsize, huffman_precision);
    const int huffman_ysize = VP8LSubSampleSize(ysize, huffman_precision);
    const int huffman_pixs = huffman_xsize * huffman_ysize;
-    if (!DecodeImageStream(huffman_xsize, huffman_ysize, 0, dec,
+    if (!DecodeImageStream(huffman_xsize, huffman_ysize, /*is_level0=*/0, dec,
                           &huffman_image)) {
      goto Error;
    }
@@ -402,7 +409,7 @@ static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
      // values [0, num_htree_groups)
      mapping = (int*)WebPSafeMalloc(num_htree_groups_max, sizeof(*mapping));
      if (mapping == NULL) {
-        dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
+        VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
        goto Error;
      }
      // -1 means a value is unmapped, and therefore unused in the Huffman
@@ -421,29 +428,55 @@ static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,

  if (br->eos_) goto Error;

-  // Find maximum alphabet size for the htree group.
-  for (j = 0; j < HUFFMAN_CODES_PER_META_CODE; ++j) {
-    int alphabet_size = kAlphabetSize[j];
-    if (j == 0 && color_cache_bits > 0) {
-      alphabet_size += 1 << color_cache_bits;
-    }
-    if (max_alphabet_size < alphabet_size) {
-      max_alphabet_size = alphabet_size;
-    }
+  if (!ReadHuffmanCodesHelper(color_cache_bits, num_htree_groups,
+                              num_htree_groups_max, mapping, dec,
+                              huffman_tables, &htree_groups)) {
+    goto Error;
  }
+  ok = 1;

-  code_lengths = (int*)WebPSafeCalloc((uint64_t)max_alphabet_size,
-                                      sizeof(*code_lengths));
-  huffman_tables = (HuffmanCode*)WebPSafeMalloc(num_htree_groups * table_size,
-                                                sizeof(*huffman_tables));
-  htree_groups = VP8LHtreeGroupsNew(num_htree_groups);
+  // All OK. Finalize pointers.
+  hdr->huffman_image_ = huffman_image;
+  hdr->num_htree_groups_ = num_htree_groups;
+  hdr->htree_groups_ = htree_groups;

-  if (htree_groups == NULL || code_lengths == NULL || huffman_tables == NULL) {
-    dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
+ Error:
+  WebPSafeFree(mapping);
+  if (!ok) {
+    WebPSafeFree(huffman_image);
+    VP8LHuffmanTablesDeallocate(huffman_tables);
+    VP8LHtreeGroupsFree(htree_groups);
+  }
+  return ok;
+}
+
+int ReadHuffmanCodesHelper(int color_cache_bits, int num_htree_groups,
+                           int num_htree_groups_max, const int* const mapping,
+                           VP8LDecoder* const dec,
+                           HuffmanTables* const huffman_tables,
+                           HTreeGroup** const htree_groups) {
+  int i, j, ok = 0;
+  const int max_alphabet_size =
+      kAlphabetSize[0] + ((color_cache_bits > 0) ? 1 << color_cache_bits : 0);
+  const int table_size = kTableSize[color_cache_bits];
+  int* code_lengths = NULL;
+
+  if ((mapping == NULL && num_htree_groups != num_htree_groups_max) ||
+      num_htree_groups > num_htree_groups_max) {
+    goto Error;
+  }
+
+  code_lengths =
+      (int*)WebPSafeCalloc((uint64_t)max_alphabet_size, sizeof(*code_lengths));
+  *htree_groups = VP8LHtreeGroupsNew(num_htree_groups);
+
+  if (*htree_groups == NULL || code_lengths == NULL ||
+      !VP8LHuffmanTablesAllocate(num_htree_groups * table_size,
+                                 huffman_tables)) {
+    VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
    goto Error;
  }

-  huffman_table = huffman_tables;
  for (i = 0; i < num_htree_groups_max; ++i) {
    // If the index "i" is unused in the Huffman image, just make sure the
    // coefficients are valid but do not store them.
@@ -460,7 +493,7 @@ static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
      }
    } else {
      HTreeGroup* const htree_group =
-          &htree_groups[(mapping == NULL) ? i : mapping[i]];
+          &(*htree_groups)[(mapping == NULL) ? i : mapping[i]];
      HuffmanCode** const htrees = htree_group->htrees;
      int size;
      int total_size = 0;
@@ -468,19 +501,20 @@ static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
      int max_bits = 0;
      for (j = 0; j < HUFFMAN_CODES_PER_META_CODE; ++j) {
        int alphabet_size = kAlphabetSize[j];
-        htrees[j] = huffman_table;
        if (j == 0 && color_cache_bits > 0) {
          alphabet_size += (1 << color_cache_bits);
        }
-        size = ReadHuffmanCode(alphabet_size, dec, code_lengths, huffman_table);
+        size =
+            ReadHuffmanCode(alphabet_size, dec, code_lengths, huffman_tables);
+        htrees[j] = huffman_tables->curr_segment->curr_table;
        if (size == 0) {
          goto Error;
        }
        if (is_trivial_literal && kLiteralMap[j] == 1) {
-          is_trivial_literal = (huffman_table->bits == 0);
+          is_trivial_literal = (htrees[j]->bits == 0);
        }
-        total_size += huffman_table->bits;
-        huffman_table += size;
+        total_size += htrees[j]->bits;
+        huffman_tables->curr_segment->curr_table += size;
        if (j <= ALPHA) {
          int local_max_bits = code_lengths[0];
          int k;
@@ -511,19 +545,12 @@ static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
  }
  ok = 1;

-  // All OK. Finalize pointers.
-  hdr->huffman_image_ = huffman_image;
-  hdr->num_htree_groups_ = num_htree_groups;
-  hdr->htree_groups_ = htree_groups;
-  hdr->huffman_tables_ = huffman_tables;
-
 Error:
  WebPSafeFree(code_lengths);
-  WebPSafeFree(mapping);
  if (!ok) {
-    WebPSafeFree(huffman_image);
-    WebPSafeFree(huffman_tables);
-    VP8LHtreeGroupsFree(htree_groups);
+    VP8LHuffmanTablesDeallocate(huffman_tables);
+    VP8LHtreeGroupsFree(*htree_groups);
+    *htree_groups = NULL;
  }
  return ok;
 }
@@ -547,8 +574,7 @@ static int AllocateAndInitRescaler(VP8LDecoder* const dec, VP8Io* const io) {
                               scaled_data_size * sizeof(*scaled_data);
  uint8_t* memory = (uint8_t*)WebPSafeMalloc(memory_size, sizeof(*memory));
  if (memory == NULL) {
-    dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
-    return 0;
+    return VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
  }
  assert(dec->rescaler_memory == NULL);
  dec->rescaler_memory = memory;
@@ -559,8 +585,11 @@ static int AllocateAndInitRescaler(VP8LDecoder* const dec, VP8Io* const io) {
  memory += work_size * sizeof(*work);
  scaled_data = (uint32_t*)memory;

-  WebPRescalerInit(dec->rescaler, in_width, in_height, (uint8_t*)scaled_data,
-                   out_width, out_height, 0, num_channels, work);
+  if (!WebPRescalerInit(dec->rescaler, in_width, in_height,
+                        (uint8_t*)scaled_data, out_width, out_height,
+                        0, num_channels, work)) {
+    return 0;
+  }
  return 1;
 }
 #endif   // WEBP_REDUCE_SIZE
@@ -574,13 +603,14 @@ static int AllocateAndInitRescaler(VP8LDecoder* const dec, VP8Io* const io) {
 static int Export(WebPRescaler* const rescaler, WEBP_CSP_MODE colorspace,
                  int rgba_stride, uint8_t* const rgba) {
  uint32_t* const src = (uint32_t*)rescaler->dst;
+  uint8_t* dst = rgba;
  const int dst_width = rescaler->dst_width;
  int num_lines_out = 0;
  while (WebPRescalerHasPendingOutput(rescaler)) {
-    uint8_t* const dst = rgba + num_lines_out * rgba_stride;
    WebPRescalerExportRow(rescaler);
    WebPMultARGBRow(src, dst_width, 1);
    VP8LConvertFromBGRA(src, dst_width, colorspace, dst);
+    dst += rgba_stride;
    ++num_lines_out;
  }
  return num_lines_out;
@@ -594,8 +624,8 @@ static int EmitRescaledRowsRGBA(const VP8LDecoder* const dec,
  int num_lines_in = 0;
  int num_lines_out = 0;
  while (num_lines_in < mb_h) {
-    uint8_t* const row_in = in + num_lines_in * in_stride;
-    uint8_t* const row_out = out + num_lines_out * out_stride;
+    uint8_t* const row_in = in + (uint64_t)num_lines_in * in_stride;
+    uint8_t* const row_out = out + (uint64_t)num_lines_out * out_stride;
    const int lines_left = mb_h - num_lines_in;
    const int needed_lines = WebPRescaleNeededLines(dec->rescaler, lines_left);
    int lines_imported;
@@ -796,7 +826,8 @@ static void ProcessRows(VP8LDecoder* const dec, int row) {
      const WebPDecBuffer* const output = dec->output_;
      if (WebPIsRGBMode(output->colorspace)) {  // convert to RGBA
        const WebPRGBABuffer* const buf = &output->u.RGBA;
-        uint8_t* const rgba = buf->rgba + dec->last_out_row_ * buf->stride;
+        uint8_t* const rgba =
+            buf->rgba + (int64_t)dec->last_out_row_ * buf->stride;
        const int num_rows_out =
 #if !defined(WEBP_REDUCE_SIZE)
         io->use_scaling ?
@@ -1077,12 +1108,10 @@ static int DecodeAlphaData(VP8LDecoder* const dec, uint8_t* const data,
 End:
  br->eos_ = VP8LIsEndOfStream(br);
  if (!ok || (br->eos_ && pos < end)) {
-    ok = 0;
-    dec->status_ = br->eos_ ? VP8_STATUS_SUSPENDED
-                            : VP8_STATUS_BITSTREAM_ERROR;
-  } else {
-    dec->last_pixel_ = pos;
+    return VP8LSetError(
+        dec, br->eos_ ? VP8_STATUS_SUSPENDED : VP8_STATUS_BITSTREAM_ERROR);
  }
+  dec->last_pixel_ = pos;
  return ok;
 }

@@ -1232,9 +1261,20 @@ static int DecodeImageData(VP8LDecoder* const dec, uint32_t* const data,
  }

  br->eos_ = VP8LIsEndOfStream(br);
-  if (dec->incremental_ && br->eos_ && src < src_end) {
+  // In incremental decoding:
+  // br->eos_ && src < src_last: if 'br' reached the end of the buffer and
+  // 'src_last' has not been reached yet, there is not enough data. 'dec' has to
+  // be reset until there is more data.
+  // !br->eos_ && src < src_last: this cannot happen as either the buffer is
+  // fully read, either enough has been read to reach 'src_last'.
+  // src >= src_last: 'src_last' is reached, all is fine. 'src' can actually go
+  // beyond 'src_last' in case the image is cropped and an LZ77 goes further.
+  // The buffer might have been enough or there is some left. 'br->eos_' does
+  // not matter.
+  assert(!dec->incremental_ || (br->eos_ && src < src_last) || src >= src_last);
+  if (dec->incremental_ && br->eos_ && src < src_last) {
    RestoreState(dec);
-  } else if (!br->eos_) {
+  } else if ((dec->incremental_ && src >= src_last) || !br->eos_) {
    // Process the remaining rows corresponding to last row-block.
    if (process_func != NULL) {
      process_func(dec, row > last_row ? last_row : row);
@@ -1249,8 +1289,7 @@ static int DecodeImageData(VP8LDecoder* const dec, uint32_t* const data,
  return 1;

 Error:
-  dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
-  return 0;
+  return VP8LSetError(dec, VP8_STATUS_BITSTREAM_ERROR);
 }

 // -----------------------------------------------------------------------------
@@ -1276,7 +1315,7 @@ static int ExpandColorMap(int num_colors, VP8LTransform* const transform) {
    uint8_t* const new_data = (uint8_t*)new_color_map;
    new_color_map[0] = transform->data_[0];
    for (i = 4; i < 4 * num_colors; ++i) {
-      // Equivalent to AddPixelEq(), on a byte-basis.
+      // Equivalent to VP8LAddPixels(), on a byte-basis.
      new_data[i] = (data[i] + new_data[i - 4]) & 0xff;
    }
    for (; i < 4 * final_num_colors; ++i) {
@@ -1317,7 +1356,7 @@ static int ReadTransform(int* const xsize, int const* ysize,
                                               transform->bits_),
                             VP8LSubSampleSize(transform->ysize_,
                                               transform->bits_),
-                             0, dec, &transform->data_);
+                             /*is_level0=*/0, dec, &transform->data_);
      break;
    case COLOR_INDEXING_TRANSFORM: {
       const int num_colors = VP8LReadBits(br, 8) + 1;
@@ -1327,11 +1366,14 @@ static int ReadTransform(int* const xsize, int const* ysize,
                      : 3;
       *xsize = VP8LSubSampleSize(transform->xsize_, bits);
       transform->bits_ = bits;
-       ok = DecodeImageStream(num_colors, 1, 0, dec, &transform->data_);
-       ok = ok && ExpandColorMap(num_colors, transform);
+       ok = DecodeImageStream(num_colors, /*ysize=*/1, /*is_level0=*/0, dec,
+                              &transform->data_);
+       if (ok && !ExpandColorMap(num_colors, transform)) {
+         return VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
+       }
      break;
    }
-    case SUBTRACT_GREEN:
+    case SUBTRACT_GREEN_TRANSFORM:
      break;
    default:
      assert(0);    // can't happen
@@ -1353,7 +1395,7 @@ static void ClearMetadata(VP8LMetadata* const hdr) {
  assert(hdr != NULL);

  WebPSafeFree(hdr->huffman_image_);
-  WebPSafeFree(hdr->huffman_tables_);
+  VP8LHuffmanTablesDeallocate(&hdr->huffman_tables_);
  VP8LHtreeGroupsFree(hdr->htree_groups_);
  VP8LColorCacheClear(&hdr->color_cache_);
  VP8LColorCacheClear(&hdr->saved_color_cache_);
@@ -1434,7 +1476,7 @@ static int DecodeImageStream(int xsize, int ysize,
    color_cache_bits = VP8LReadBits(br, 4);
    ok = (color_cache_bits >= 1 && color_cache_bits <= MAX_CACHE_BITS);
    if (!ok) {
-      dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
+      VP8LSetError(dec, VP8_STATUS_BITSTREAM_ERROR);
      goto End;
    }
  }
@@ -1443,7 +1485,7 @@ static int DecodeImageStream(int xsize, int ysize,
  ok = ok && ReadHuffmanCodes(dec, transform_xsize, transform_ysize,
                              color_cache_bits, is_level0);
  if (!ok) {
-    dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
+    VP8LSetError(dec, VP8_STATUS_BITSTREAM_ERROR);
    goto End;
  }

@@ -1451,8 +1493,7 @@ static int DecodeImageStream(int xsize, int ysize,
  if (color_cache_bits > 0) {
    hdr->color_cache_size_ = 1 << color_cache_bits;
    if (!VP8LColorCacheInit(&hdr->color_cache_, color_cache_bits)) {
-      dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
-      ok = 0;
+      ok = VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
      goto End;
    }
  } else {
@@ -1469,8 +1510,7 @@ static int DecodeImageStream(int xsize, int ysize,
    const uint64_t total_size = (uint64_t)transform_xsize * transform_ysize;
    data = (uint32_t*)WebPSafeMalloc(total_size, sizeof(*data));
    if (data == NULL) {
-      dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
-      ok = 0;
+      ok = VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
      goto End;
    }
  }
@@ -1514,9 +1554,8 @@ static int AllocateInternalBuffers32b(VP8LDecoder* const dec, int final_width) {
  assert(dec->width_ <= final_width);
  dec->pixels_ = (uint32_t*)WebPSafeMalloc(total_num_pixels, sizeof(uint32_t));
  if (dec->pixels_ == NULL) {
-    dec->argb_cache_ = NULL;    // for sanity check
-    dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
-    return 0;
+    dec->argb_cache_ = NULL;    // for soundness
+    return VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
  }
  dec->argb_cache_ = dec->pixels_ + num_pixels + cache_top_pixels;
  return 1;
@@ -1524,11 +1563,10 @@ static int AllocateInternalBuffers32b(VP8LDecoder* const dec, int final_width) {

 static int AllocateInternalBuffers8b(VP8LDecoder* const dec) {
  const uint64_t total_num_pixels = (uint64_t)dec->width_ * dec->height_;
-  dec->argb_cache_ = NULL;    // for sanity check
+  dec->argb_cache_ = NULL;    // for soundness
  dec->pixels_ = (uint32_t*)WebPSafeMalloc(total_num_pixels, sizeof(uint8_t));
  if (dec->pixels_ == NULL) {
-    dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
-    return 0;
+    return VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
  }
  return 1;
 }
@@ -1583,7 +1621,8 @@ int VP8LDecodeAlphaHeader(ALPHDecoder* const alph_dec,
  dec->status_ = VP8_STATUS_OK;
  VP8LInitBitReader(&dec->br_, data, data_size);

-  if (!DecodeImageStream(alph_dec->width_, alph_dec->height_, 1, dec, NULL)) {
+  if (!DecodeImageStream(alph_dec->width_, alph_dec->height_, /*is_level0=*/1,
+                         dec, /*decoded_data=*/NULL)) {
    goto Err;
  }

@@ -1638,22 +1677,24 @@ int VP8LDecodeHeader(VP8LDecoder* const dec, VP8Io* const io) {

  if (dec == NULL) return 0;
  if (io == NULL) {
-    dec->status_ = VP8_STATUS_INVALID_PARAM;
-    return 0;
+    return VP8LSetError(dec, VP8_STATUS_INVALID_PARAM);
  }

  dec->io_ = io;
  dec->status_ = VP8_STATUS_OK;
  VP8LInitBitReader(&dec->br_, io->data, io->data_size);
  if (!ReadImageInfo(&dec->br_, &width, &height, &has_alpha)) {
-    dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
+    VP8LSetError(dec, VP8_STATUS_BITSTREAM_ERROR);
    goto Error;
  }
  dec->state_ = READ_DIM;
  io->width = width;
  io->height = height;

-  if (!DecodeImageStream(width, height, 1, dec, NULL)) goto Error;
+  if (!DecodeImageStream(width, height, /*is_level0=*/1, dec,
+                         /*decoded_data=*/NULL)) {
+    goto Error;
+  }
  return 1;

 Error:
@@ -1666,10 +1707,9 @@ int VP8LDecodeImage(VP8LDecoder* const dec) {
  VP8Io* io = NULL;
  WebPDecParams* params = NULL;

-  // Sanity checks.
  if (dec == NULL) return 0;

-  assert(dec->hdr_.huffman_tables_ != NULL);
+  assert(dec->hdr_.huffman_tables_.root.start != NULL);
  assert(dec->hdr_.htree_groups_ != NULL);
  assert(dec->hdr_.num_htree_groups_ > 0);

@@ -1684,7 +1724,7 @@ int VP8LDecodeImage(VP8LDecoder* const dec) {
    assert(dec->output_ != NULL);

    if (!WebPIoInitFromOptions(params->options, io, MODE_BGRA)) {
-      dec->status_ = VP8_STATUS_INVALID_PARAM;
+      VP8LSetError(dec, VP8_STATUS_INVALID_PARAM);
      goto Err;
    }

@@ -1694,7 +1734,7 @@ int VP8LDecodeImage(VP8LDecoder* const dec) {
    if (io->use_scaling && !AllocateAndInitRescaler(dec, io)) goto Err;
 #else
    if (io->use_scaling) {
-      dec->status_ = VP8_STATUS_INVALID_PARAM;
+      VP8LSetError(dec, VP8_STATUS_INVALID_PARAM);
      goto Err;
    }
 #endif
@@ -1712,7 +1752,7 @@ int VP8LDecodeImage(VP8LDecoder* const dec) {
          dec->hdr_.saved_color_cache_.colors_ == NULL) {
        if (!VP8LColorCacheInit(&dec->hdr_.saved_color_cache_,
                                dec->hdr_.color_cache_.hash_bits_)) {
-          dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
+          VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
          goto Err;
        }
      }
@@ -51,7 +51,7 @@ typedef struct {
  uint32_t*       huffman_image_;
  int             num_htree_groups_;
  HTreeGroup*     htree_groups_;
-  HuffmanCode*    huffman_tables_;
+  HuffmanTables   huffman_tables_;
 } VP8LMetadata;

 typedef struct VP8LDecoder VP8LDecoder;
@@ -126,6 +126,19 @@ void VP8LClear(VP8LDecoder* const dec);
 // Clears and deallocate a lossless decoder instance.
 void VP8LDelete(VP8LDecoder* const dec);

+// Helper function for reading the different Huffman codes and storing them in
+// 'huffman_tables' and 'htree_groups'.
+// If mapping is NULL 'num_htree_groups_max' must equal 'num_htree_groups'.
+// If it is not NULL, it maps 'num_htree_groups_max' indices to the
+// 'num_htree_groups' groups. If 'num_htree_groups_max' > 'num_htree_groups',
+// some of those indices map to -1. This is used for non-balanced codes to
+// limit memory usage.
+int ReadHuffmanCodesHelper(int color_cache_bits, int num_htree_groups,
+                           int num_htree_groups_max, const int* const mapping,
+                           VP8LDecoder* const dec,
+                           HuffmanTables* const huffman_tables,
+                           HTreeGroup** const htree_groups);
+
 //------------------------------------------------------------------------------

 #ifdef __cplusplus
@@ -179,7 +179,7 @@ static VP8StatusCode ParseOptionalChunks(const uint8_t** const data,
      return VP8_STATUS_BITSTREAM_ERROR;          // Not a valid chunk size.
    }
    // For odd-sized chunk-payload, there's one byte padding at the end.
-    disk_chunk_size = (CHUNK_HEADER_SIZE + chunk_size + 1) & ~1;
+    disk_chunk_size = (CHUNK_HEADER_SIZE + chunk_size + 1) & ~1u;
    total_size += disk_chunk_size;

    // Check that total bytes skipped so far does not exceed riff_size.
@@ -658,19 +658,26 @@ uint8_t* WebPDecodeBGRA(const uint8_t* data, size_t data_size,
 uint8_t* WebPDecodeYUV(const uint8_t* data, size_t data_size,
                       int* width, int* height, uint8_t** u, uint8_t** v,
                       int* stride, int* uv_stride) {
-  WebPDecBuffer output;   // only to preserve the side-infos
-  uint8_t* const out = Decode(MODE_YUV, data, data_size,
-                              width, height, &output);
-
-  if (out != NULL) {
-    const WebPYUVABuffer* const buf = &output.u.YUVA;
-    *u = buf->u;
-    *v = buf->v;
-    *stride = buf->y_stride;
-    *uv_stride = buf->u_stride;
-    assert(buf->u_stride == buf->v_stride);
+  // data, width and height are checked by Decode().
+  if (u == NULL || v == NULL || stride == NULL || uv_stride == NULL) {
+    return NULL;
+  }
+
+  {
+    WebPDecBuffer output;   // only to preserve the side-infos
+    uint8_t* const out = Decode(MODE_YUV, data, data_size,
+                                width, height, &output);
+
+    if (out != NULL) {
+      const WebPYUVABuffer* const buf = &output.u.YUVA;
+      *u = buf->u;
+      *v = buf->v;
+      *stride = buf->y_stride;
+      *uv_stride = buf->u_stride;
+      assert(buf->u_stride == buf->v_stride);
+    }
+    return out;
  }
-  return out;
 }

 static void DefaultFeatures(WebPBitstreamFeatures* const features) {
@@ -785,6 +792,13 @@ VP8StatusCode WebPDecode(const uint8_t* data, size_t data_size,
 //------------------------------------------------------------------------------
 // Cropping and rescaling.

+int WebPCheckCropDimensions(int image_width, int image_height,
+                            int x, int y, int w, int h) {
+  return !(x < 0 || y < 0 || w <= 0 || h <= 0 ||
+           x >= image_width || w > image_width || w > image_width - x ||
+           y >= image_height || h > image_height || h > image_height - y);
+}
+
 int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
                          VP8Io* const io, WEBP_CSP_MODE src_colorspace) {
  const int W = io->width;
@@ -792,7 +806,7 @@ int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
  int x = 0, y = 0, w = W, h = H;

  // Cropping
-  io->use_cropping = (options != NULL) && (options->use_cropping > 0);
+  io->use_cropping = (options != NULL) && options->use_cropping;
  if (io->use_cropping) {
    w = options->crop_width;
    h = options->crop_height;
@@ -802,7 +816,7 @@ int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
      x &= ~1;
      y &= ~1;
    }
-    if (x < 0 || y < 0 || w <= 0 || h <= 0 || x + w > W || y + h > H) {
+    if (!WebPCheckCropDimensions(W, H, x, y, w, h)) {
      return 0;  // out of frame boundary error
    }
  }
@@ -814,7 +828,7 @@ int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
  io->mb_h = h;

  // Scaling
-  io->use_scaling = (options != NULL) && (options->use_scaling > 0);
+  io->use_scaling = (options != NULL) && options->use_scaling;
  if (io->use_scaling) {
    int scaled_width = options->scaled_width;
    int scaled_height = options->scaled_height;
@@ -835,8 +849,8 @@ int WebPIoInitFromOptions(const WebPDecoderOptions* const options,

  if (io->use_scaling) {
    // disable filter (only for large downscaling ratio).
-    io->bypass_filtering = (io->scaled_width < W * 3 / 4) &&
-                           (io->scaled_height < H * 3 / 4);
+    io->bypass_filtering |= (io->scaled_width < W * 3 / 4) &&
+                            (io->scaled_height < H * 3 / 4);
    io->fancy_upsampling = 0;
  }
  return 1;
@@ -77,6 +77,10 @@ VP8StatusCode WebPParseHeaders(WebPHeaderStructure* const headers);
 //------------------------------------------------------------------------------
 // Misc utils

+// Returns true if crop dimensions are within image bounds.
+int WebPCheckCropDimensions(int image_width, int image_height,
+                            int x, int y, int w, int h);
+
 // Initializes VP8Io with custom setup, io and teardown functions. The default
 // hooks will use the supplied 'params' as io->opaque handle.
 void WebPInitCustomIo(WebPDecParams* const params, VP8Io* const io);
@@ -23,6 +23,14 @@

 #define NUM_CHANNELS 4

+// Channel extraction from a uint32_t representation of a uint8_t RGBA/BGRA
+// buffer.
+#ifdef WORDS_BIGENDIAN
+#define CHANNEL_SHIFT(i) (24 - (i) * 8)
+#else
+#define CHANNEL_SHIFT(i) ((i) * 8)
+#endif
+
 typedef void (*BlendRowFunc)(uint32_t* const, const uint32_t* const, int);
 static void BlendPixelRowNonPremult(uint32_t* const src,
                                    const uint32_t* const dst, int num_pixels);
@@ -87,11 +95,19 @@ WebPAnimDecoder* WebPAnimDecoderNewInternal(
    int abi_version) {
  WebPAnimDecoderOptions options;
  WebPAnimDecoder* dec = NULL;
+  WebPBitstreamFeatures features;
  if (webp_data == NULL ||
      WEBP_ABI_IS_INCOMPATIBLE(abi_version, WEBP_DEMUX_ABI_VERSION)) {
    return NULL;
  }

+  // Validate the bitstream before doing expensive allocations. The demuxer may
+  // be more tolerant than the decoder.
+  if (WebPGetFeatures(webp_data->bytes, webp_data->size, &features) !=
+      VP8_STATUS_OK) {
+    return NULL;
+  }
+
  // Note: calloc() so that the pointer members are initialized to NULL.
  dec = (WebPAnimDecoder*)WebPSafeCalloc(1ULL, sizeof(*dec));
  if (dec == NULL) goto Error;
@@ -145,7 +161,7 @@ static int ZeroFillCanvas(uint8_t* buf, uint32_t canvas_width,
                          uint32_t canvas_height) {
  const uint64_t size =
      (uint64_t)canvas_width * canvas_height * NUM_CHANNELS * sizeof(*buf);
-  if (size != (size_t)size) return 0;
+  if (!CheckSizeOverflow(size)) return 0;
  memset(buf, 0, (size_t)size);
  return 1;
 }
@@ -166,7 +182,7 @@ static void ZeroFillFrameRect(uint8_t* buf, int buf_stride, int x_offset,
 static int CopyCanvas(const uint8_t* src, uint8_t* dst,
                      uint32_t width, uint32_t height) {
  const uint64_t size = (uint64_t)width * height * NUM_CHANNELS;
-  if (size != (size_t)size) return 0;
+  if (!CheckSizeOverflow(size)) return 0;
  assert(src != NULL && dst != NULL);
  memcpy(dst, src, (size_t)size);
  return 1;
@@ -201,35 +217,35 @@ static uint8_t BlendChannelNonPremult(uint32_t src, uint8_t src_a,
  const uint8_t dst_channel = (dst >> shift) & 0xff;
  const uint32_t blend_unscaled = src_channel * src_a + dst_channel * dst_a;
  assert(blend_unscaled < (1ULL << 32) / scale);
-  return (blend_unscaled * scale) >> 24;
+  return (blend_unscaled * scale) >> CHANNEL_SHIFT(3);
 }

 // Blend 'src' over 'dst' assuming they are NOT pre-multiplied by alpha.
 static uint32_t BlendPixelNonPremult(uint32_t src, uint32_t dst) {
-  const uint8_t src_a = (src >> 24) & 0xff;
+  const uint8_t src_a = (src >> CHANNEL_SHIFT(3)) & 0xff;

  if (src_a == 0) {
    return dst;
  } else {
-    const uint8_t dst_a = (dst >> 24) & 0xff;
+    const uint8_t dst_a = (dst >> CHANNEL_SHIFT(3)) & 0xff;
    // This is the approximate integer arithmetic for the actual formula:
    // dst_factor_a = (dst_a * (255 - src_a)) / 255.
    const uint8_t dst_factor_a = (dst_a * (256 - src_a)) >> 8;
    const uint8_t blend_a = src_a + dst_factor_a;
    const uint32_t scale = (1UL << 24) / blend_a;

-    const uint8_t blend_r =
-        BlendChannelNonPremult(src, src_a, dst, dst_factor_a, scale, 0);
-    const uint8_t blend_g =
-        BlendChannelNonPremult(src, src_a, dst, dst_factor_a, scale, 8);
-    const uint8_t blend_b =
-        BlendChannelNonPremult(src, src_a, dst, dst_factor_a, scale, 16);
+    const uint8_t blend_r = BlendChannelNonPremult(
+        src, src_a, dst, dst_factor_a, scale, CHANNEL_SHIFT(0));
+    const uint8_t blend_g = BlendChannelNonPremult(
+        src, src_a, dst, dst_factor_a, scale, CHANNEL_SHIFT(1));
+    const uint8_t blend_b = BlendChannelNonPremult(
+        src, src_a, dst, dst_factor_a, scale, CHANNEL_SHIFT(2));
    assert(src_a + dst_factor_a < 256);

-    return (blend_r << 0) |
-           (blend_g << 8) |
-           (blend_b << 16) |
-           ((uint32_t)blend_a << 24);
+    return ((uint32_t)blend_r << CHANNEL_SHIFT(0)) |
+           ((uint32_t)blend_g << CHANNEL_SHIFT(1)) |
+           ((uint32_t)blend_b << CHANNEL_SHIFT(2)) |
+           ((uint32_t)blend_a << CHANNEL_SHIFT(3));
  }
 }

@@ -239,7 +255,7 @@ static void BlendPixelRowNonPremult(uint32_t* const src,
                                    const uint32_t* const dst, int num_pixels) {
  int i;
  for (i = 0; i < num_pixels; ++i) {
-    const uint8_t src_alpha = (src[i] >> 24) & 0xff;
+    const uint8_t src_alpha = (src[i] >> CHANNEL_SHIFT(3)) & 0xff;
    if (src_alpha != 0xff) {
      src[i] = BlendPixelNonPremult(src[i], dst[i]);
    }
@@ -256,7 +272,7 @@ static WEBP_INLINE uint32_t ChannelwiseMultiply(uint32_t pix, uint32_t scale) {

 // Blend 'src' over 'dst' assuming they are pre-multiplied by alpha.
 static uint32_t BlendPixelPremult(uint32_t src, uint32_t dst) {
-  const uint8_t src_a = (src >> 24) & 0xff;
+  const uint8_t src_a = (src >> CHANNEL_SHIFT(3)) & 0xff;
  return src + ChannelwiseMultiply(dst, 256 - src_a);
 }

@@ -266,7 +282,7 @@ static void BlendPixelRowPremult(uint32_t* const src, const uint32_t* const dst,
                                 int num_pixels) {
  int i;
  for (i = 0; i < num_pixels; ++i) {
-    const uint8_t src_alpha = (src[i] >> 24) & 0xff;
+    const uint8_t src_alpha = (src[i] >> CHANNEL_SHIFT(3)) & 0xff;
    if (src_alpha != 0xff) {
      src[i] = BlendPixelPremult(src[i], dst[i]);
    }
@@ -24,8 +24,8 @@
 #include "src/webp/format_constants.h"

 #define DMUX_MAJ_VERSION 1
-#define DMUX_MIN_VERSION 2
-#define DMUX_REV_VERSION 0
+#define DMUX_MIN_VERSION 3
+#define DMUX_REV_VERSION 1

 typedef struct {
  size_t start_;        // start location of the data
@@ -221,12 +221,16 @@ static ParseStatus StoreFrame(int frame_num, uint32_t min_size,
    const size_t chunk_start_offset = mem->start_;
    const uint32_t fourcc = ReadLE32(mem);
    const uint32_t payload_size = ReadLE32(mem);
-    const uint32_t payload_size_padded = payload_size + (payload_size & 1);
-    const size_t payload_available = (payload_size_padded > MemDataSize(mem))
-                                   ? MemDataSize(mem) : payload_size_padded;
-    const size_t chunk_size = CHUNK_HEADER_SIZE + payload_available;
+    uint32_t payload_size_padded;
+    size_t payload_available;
+    size_t chunk_size;

    if (payload_size > MAX_CHUNK_PAYLOAD) return PARSE_ERROR;
+
+    payload_size_padded = payload_size + (payload_size & 1);
+    payload_available = (payload_size_padded > MemDataSize(mem))
+                      ? MemDataSize(mem) : payload_size_padded;
+    chunk_size = CHUNK_HEADER_SIZE + payload_available;
    if (SizeIsInvalid(mem, payload_size_padded)) return PARSE_ERROR;
    if (payload_size_padded > MemDataSize(mem)) status = PARSE_NEED_MORE_DATA;

@@ -451,9 +455,11 @@ static ParseStatus ParseVP8XChunks(WebPDemuxer* const dmux) {
    const size_t chunk_start_offset = mem->start_;
    const uint32_t fourcc = ReadLE32(mem);
    const uint32_t chunk_size = ReadLE32(mem);
-    const uint32_t chunk_size_padded = chunk_size + (chunk_size & 1);
+    uint32_t chunk_size_padded;

    if (chunk_size > MAX_CHUNK_PAYLOAD) return PARSE_ERROR;
+
+    chunk_size_padded = chunk_size + (chunk_size & 1);
    if (SizeIsInvalid(mem, chunk_size_padded)) return PARSE_ERROR;

    switch (fourcc) {
@@ -608,7 +614,6 @@ static int IsValidExtendedFormat(const WebPDemuxer* const dmux) {

  while (f != NULL) {
    const int cur_frame_set = f->frame_num_;
-    int frame_count = 0;

    // Check frame properties.
    for (; f != NULL && f->frame_num_ == cur_frame_set; f = f->next_) {
@@ -643,8 +648,6 @@ static int IsValidExtendedFormat(const WebPDemuxer* const dmux) {
                            dmux->canvas_width_, dmux->canvas_height_)) {
        return 0;
      }
-
-      ++frame_count;
    }
  }
  return 1;
@@ -157,7 +157,8 @@ void WebPMultARGBRow_C(uint32_t* const ptr, int width, int inverse) {
  }
 }

-void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha,
+void WebPMultRow_C(uint8_t* WEBP_RESTRICT const ptr,
+                   const uint8_t* WEBP_RESTRICT const alpha,
                   int width, int inverse) {
  int x;
  for (x = 0; x < width; ++x) {
@@ -178,7 +179,8 @@ void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha,
 #undef MFIX

 void (*WebPMultARGBRow)(uint32_t* const ptr, int width, int inverse);
-void (*WebPMultRow)(uint8_t* const ptr, const uint8_t* const alpha,
+void (*WebPMultRow)(uint8_t* WEBP_RESTRICT const ptr,
+                    const uint8_t* WEBP_RESTRICT const alpha,
                    int width, int inverse);

 //------------------------------------------------------------------------------
@@ -193,8 +195,8 @@ void WebPMultARGBRows(uint8_t* ptr, int stride, int width, int num_rows,
  }
 }

-void WebPMultRows(uint8_t* ptr, int stride,
-                  const uint8_t* alpha, int alpha_stride,
+void WebPMultRows(uint8_t* WEBP_RESTRICT ptr, int stride,
+                  const uint8_t* WEBP_RESTRICT alpha, int alpha_stride,
                  int width, int num_rows, int inverse) {
  int n;
  for (n = 0; n < num_rows; ++n) {
@@ -290,9 +292,9 @@ static void ApplyAlphaMultiply_16b_C(uint8_t* rgba4444,
 }

 #if !WEBP_NEON_OMIT_C_CODE
-static int DispatchAlpha_C(const uint8_t* alpha, int alpha_stride,
+static int DispatchAlpha_C(const uint8_t* WEBP_RESTRICT alpha, int alpha_stride,
                           int width, int height,
-                           uint8_t* dst, int dst_stride) {
+                           uint8_t* WEBP_RESTRICT dst, int dst_stride) {
  uint32_t alpha_mask = 0xff;
  int i, j;

@@ -309,9 +311,10 @@ static int DispatchAlpha_C(const uint8_t* alpha, int alpha_stride,
  return (alpha_mask != 0xff);
 }

-static void DispatchAlphaToGreen_C(const uint8_t* alpha, int alpha_stride,
-                                   int width, int height,
-                                   uint32_t* dst, int dst_stride) {
+static void DispatchAlphaToGreen_C(const uint8_t* WEBP_RESTRICT alpha,
+                                   int alpha_stride, int width, int height,
+                                   uint32_t* WEBP_RESTRICT dst,
+                                   int dst_stride) {
  int i, j;
  for (j = 0; j < height; ++j) {
    for (i = 0; i < width; ++i) {
@@ -322,9 +325,9 @@ static void DispatchAlphaToGreen_C(const uint8_t* alpha, int alpha_stride,
  }
 }

-static int ExtractAlpha_C(const uint8_t* argb, int argb_stride,
+static int ExtractAlpha_C(const uint8_t* WEBP_RESTRICT argb, int argb_stride,
                          int width, int height,
-                          uint8_t* alpha, int alpha_stride) {
+                          uint8_t* WEBP_RESTRICT alpha, int alpha_stride) {
  uint8_t alpha_mask = 0xff;
  int i, j;

@@ -340,7 +343,8 @@ static int ExtractAlpha_C(const uint8_t* argb, int argb_stride,
  return (alpha_mask == 0xff);
 }

-static void ExtractGreen_C(const uint32_t* argb, uint8_t* alpha, int size) {
+static void ExtractGreen_C(const uint32_t* WEBP_RESTRICT argb,
+                           uint8_t* WEBP_RESTRICT alpha, int size) {
  int i;
  for (i = 0; i < size; ++i) alpha[i] = argb[i] >> 8;
 }
@@ -372,8 +376,11 @@ static WEBP_INLINE uint32_t MakeARGB32(int a, int r, int g, int b) {
 }

 #ifdef WORDS_BIGENDIAN
-static void PackARGB_C(const uint8_t* a, const uint8_t* r, const uint8_t* g,
-                       const uint8_t* b, int len, uint32_t* out) {
+static void PackARGB_C(const uint8_t* WEBP_RESTRICT a,
+                       const uint8_t* WEBP_RESTRICT r,
+                       const uint8_t* WEBP_RESTRICT g,
+                       const uint8_t* WEBP_RESTRICT b,
+                       int len, uint32_t* WEBP_RESTRICT out) {
  int i;
  for (i = 0; i < len; ++i) {
    out[i] = MakeARGB32(a[4 * i], r[4 * i], g[4 * i], b[4 * i]);
@@ -381,8 +388,10 @@ static void PackARGB_C(const uint8_t* a, const uint8_t* r, const uint8_t* g,
 }
 #endif

-static void PackRGB_C(const uint8_t* r, const uint8_t* g, const uint8_t* b,
-                      int len, int step, uint32_t* out) {
+static void PackRGB_C(const uint8_t* WEBP_RESTRICT r,
+                      const uint8_t* WEBP_RESTRICT g,
+                      const uint8_t* WEBP_RESTRICT b,
+                      int len, int step, uint32_t* WEBP_RESTRICT out) {
  int i, offset = 0;
  for (i = 0; i < len; ++i) {
    out[i] = MakeARGB32(0xff, r[offset], g[offset], b[offset]);
@@ -392,16 +401,22 @@ static void PackRGB_C(const uint8_t* r, const uint8_t* g, const uint8_t* b,

 void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int);
 void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int);
-int (*WebPDispatchAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
-void (*WebPDispatchAlphaToGreen)(const uint8_t*, int, int, int, uint32_t*, int);
-int (*WebPExtractAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
-void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size);
+int (*WebPDispatchAlpha)(const uint8_t* WEBP_RESTRICT, int, int, int,
+                         uint8_t* WEBP_RESTRICT, int);
+void (*WebPDispatchAlphaToGreen)(const uint8_t* WEBP_RESTRICT, int, int, int,
+                                 uint32_t* WEBP_RESTRICT, int);
+int (*WebPExtractAlpha)(const uint8_t* WEBP_RESTRICT, int, int, int,
+                        uint8_t* WEBP_RESTRICT, int);
+void (*WebPExtractGreen)(const uint32_t* WEBP_RESTRICT argb,
+                         uint8_t* WEBP_RESTRICT alpha, int size);
 #ifdef WORDS_BIGENDIAN
 void (*WebPPackARGB)(const uint8_t* a, const uint8_t* r, const uint8_t* g,
                     const uint8_t* b, int, uint32_t*);
 #endif
-void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b,
-                    int len, int step, uint32_t* out);
+void (*WebPPackRGB)(const uint8_t* WEBP_RESTRICT r,
+                    const uint8_t* WEBP_RESTRICT g,
+                    const uint8_t* WEBP_RESTRICT b,
+                    int len, int step, uint32_t* WEBP_RESTRICT out);

 int (*WebPHasAlpha8b)(const uint8_t* src, int length);
 int (*WebPHasAlpha32b)(const uint8_t* src, int length);
@@ -410,6 +425,7 @@ void (*WebPAlphaReplace)(uint32_t* src, int length, uint32_t color);
 //------------------------------------------------------------------------------
 // Init function

+extern VP8CPUInfo VP8GetCPUInfo;
 extern void WebPInitAlphaProcessingMIPSdspR2(void);
 extern void WebPInitAlphaProcessingSSE2(void);
 extern void WebPInitAlphaProcessingSSE41(void);
@@ -438,10 +454,10 @@ WEBP_DSP_INIT_FUNC(WebPInitAlphaProcessing) {

  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
  if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
    if (VP8GetCPUInfo(kSSE2)) {
      WebPInitAlphaProcessingSSE2();
-#if defined(WEBP_USE_SSE41)
+#if defined(WEBP_HAVE_SSE41)
      if (VP8GetCPUInfo(kSSE4_1)) {
        WebPInitAlphaProcessingSSE41();
      }
@@ -455,7 +471,7 @@ WEBP_DSP_INIT_FUNC(WebPInitAlphaProcessing) {
 #endif
  }

-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
  if (WEBP_NEON_OMIT_C_CODE ||
      (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
    WebPInitAlphaProcessingNEON();
@@ -80,10 +80,10 @@ static void ApplyAlphaMultiply_NEON(uint8_t* rgba, int alpha_first,

 //------------------------------------------------------------------------------

-static int DispatchAlpha_NEON(const uint8_t* alpha, int alpha_stride,
-                              int width, int height,
-                              uint8_t* dst, int dst_stride) {
-  uint32_t alpha_mask = 0xffffffffu;
+static int DispatchAlpha_NEON(const uint8_t* WEBP_RESTRICT alpha,
+                              int alpha_stride, int width, int height,
+                              uint8_t* WEBP_RESTRICT dst, int dst_stride) {
+  uint32_t alpha_mask = 0xffu;
  uint8x8_t mask8 = vdup_n_u8(0xff);
  uint32_t tmp[2];
  int i, j;
@@ -107,14 +107,16 @@ static int DispatchAlpha_NEON(const uint8_t* alpha, int alpha_stride,
    dst += dst_stride;
  }
  vst1_u8((uint8_t*)tmp, mask8);
+  alpha_mask *= 0x01010101;
  alpha_mask &= tmp[0];
  alpha_mask &= tmp[1];
  return (alpha_mask != 0xffffffffu);
 }

-static void DispatchAlphaToGreen_NEON(const uint8_t* alpha, int alpha_stride,
-                                      int width, int height,
-                                      uint32_t* dst, int dst_stride) {
+static void DispatchAlphaToGreen_NEON(const uint8_t* WEBP_RESTRICT alpha,
+                                      int alpha_stride, int width, int height,
+                                      uint32_t* WEBP_RESTRICT dst,
+                                      int dst_stride) {
  int i, j;
  uint8x8x4_t greens;   // leave A/R/B channels zero'd.
  greens.val[0] = vdup_n_u8(0);
@@ -131,10 +133,10 @@ static void DispatchAlphaToGreen_NEON(const uint8_t* alpha, int alpha_stride,
  }
 }

-static int ExtractAlpha_NEON(const uint8_t* argb, int argb_stride,
+static int ExtractAlpha_NEON(const uint8_t* WEBP_RESTRICT argb, int argb_stride,
                             int width, int height,
-                             uint8_t* alpha, int alpha_stride) {
-  uint32_t alpha_mask = 0xffffffffu;
+                             uint8_t* WEBP_RESTRICT alpha, int alpha_stride) {
+  uint32_t alpha_mask = 0xffu;
  uint8x8_t mask8 = vdup_n_u8(0xff);
  uint32_t tmp[2];
  int i, j;
@@ -156,13 +158,14 @@ static int ExtractAlpha_NEON(const uint8_t* argb, int argb_stride,
    alpha += alpha_stride;
  }
  vst1_u8((uint8_t*)tmp, mask8);
+  alpha_mask *= 0x01010101;
  alpha_mask &= tmp[0];
  alpha_mask &= tmp[1];
  return (alpha_mask == 0xffffffffu);
 }

-static void ExtractGreen_NEON(const uint32_t* argb,
-                              uint8_t* alpha, int size) {
+static void ExtractGreen_NEON(const uint32_t* WEBP_RESTRICT argb,
+                              uint8_t* WEBP_RESTRICT alpha, int size) {
  int i;
  for (i = 0; i + 16 <= size; i += 16) {
    const uint8x16x4_t rgbX = vld4q_u8((const uint8_t*)(argb + i));
@@ -18,16 +18,16 @@

 //------------------------------------------------------------------------------

-static int DispatchAlpha_SSE2(const uint8_t* alpha, int alpha_stride,
-                              int width, int height,
-                              uint8_t* dst, int dst_stride) {
+static int DispatchAlpha_SSE2(const uint8_t* WEBP_RESTRICT alpha,
+                              int alpha_stride, int width, int height,
+                              uint8_t* WEBP_RESTRICT dst, int dst_stride) {
  // alpha_and stores an 'and' operation of all the alpha[] values. The final
  // value is not 0xff if any of the alpha[] is not equal to 0xff.
  uint32_t alpha_and = 0xff;
  int i, j;
  const __m128i zero = _mm_setzero_si128();
-  const __m128i rgb_mask = _mm_set1_epi32(0xffffff00u);  // to preserve RGB
-  const __m128i all_0xff = _mm_set_epi32(0, 0, ~0u, ~0u);
+  const __m128i rgb_mask = _mm_set1_epi32((int)0xffffff00);  // to preserve RGB
+  const __m128i all_0xff = _mm_set_epi32(0, 0, ~0, ~0);
  __m128i all_alphas = all_0xff;

  // We must be able to access 3 extra bytes after the last written byte
@@ -72,9 +72,10 @@ static int DispatchAlpha_SSE2(const uint8_t* alpha, int alpha_stride,
  return (alpha_and != 0xff);
 }

-static void DispatchAlphaToGreen_SSE2(const uint8_t* alpha, int alpha_stride,
-                                      int width, int height,
-                                      uint32_t* dst, int dst_stride) {
+static void DispatchAlphaToGreen_SSE2(const uint8_t* WEBP_RESTRICT alpha,
+                                      int alpha_stride, int width, int height,
+                                      uint32_t* WEBP_RESTRICT dst,
+                                      int dst_stride) {
  int i, j;
  const __m128i zero = _mm_setzero_si128();
  const int limit = width & ~15;
@@ -98,15 +99,15 @@ static void DispatchAlphaToGreen_SSE2(const uint8_t* alpha, int alpha_stride,
  }
 }

-static int ExtractAlpha_SSE2(const uint8_t* argb, int argb_stride,
+static int ExtractAlpha_SSE2(const uint8_t* WEBP_RESTRICT argb, int argb_stride,
                             int width, int height,
-                             uint8_t* alpha, int alpha_stride) {
+                             uint8_t* WEBP_RESTRICT alpha, int alpha_stride) {
  // alpha_and stores an 'and' operation of all the alpha[] values. The final
  // value is not 0xff if any of the alpha[] is not equal to 0xff.
  uint32_t alpha_and = 0xff;
  int i, j;
-  const __m128i a_mask = _mm_set1_epi32(0xffu);  // to preserve alpha
-  const __m128i all_0xff = _mm_set_epi32(0, 0, ~0u, ~0u);
+  const __m128i a_mask = _mm_set1_epi32(0xff);  // to preserve alpha
+  const __m128i all_0xff = _mm_set_epi32(0, 0, ~0, ~0);
  __m128i all_alphas = all_0xff;

  // We must be able to access 3 extra bytes after the last written byte
@@ -143,6 +144,46 @@ static int ExtractAlpha_SSE2(const uint8_t* argb, int argb_stride,
  return (alpha_and == 0xff);
 }

+static void ExtractGreen_SSE2(const uint32_t* WEBP_RESTRICT argb,
+                              uint8_t* WEBP_RESTRICT alpha, int size) {
+  int i;
+  const __m128i mask = _mm_set1_epi32(0xff);
+  const __m128i* src = (const __m128i*)argb;
+
+  for (i = 0; i + 16 <= size; i += 16, src += 4) {
+    const __m128i a0 = _mm_loadu_si128(src + 0);
+    const __m128i a1 = _mm_loadu_si128(src + 1);
+    const __m128i a2 = _mm_loadu_si128(src + 2);
+    const __m128i a3 = _mm_loadu_si128(src + 3);
+    const __m128i b0 = _mm_srli_epi32(a0, 8);
+    const __m128i b1 = _mm_srli_epi32(a1, 8);
+    const __m128i b2 = _mm_srli_epi32(a2, 8);
+    const __m128i b3 = _mm_srli_epi32(a3, 8);
+    const __m128i c0 = _mm_and_si128(b0, mask);
+    const __m128i c1 = _mm_and_si128(b1, mask);
+    const __m128i c2 = _mm_and_si128(b2, mask);
+    const __m128i c3 = _mm_and_si128(b3, mask);
+    const __m128i d0 = _mm_packs_epi32(c0, c1);
+    const __m128i d1 = _mm_packs_epi32(c2, c3);
+    const __m128i e = _mm_packus_epi16(d0, d1);
+    // store
+    _mm_storeu_si128((__m128i*)&alpha[i], e);
+  }
+  if (i + 8 <= size) {
+    const __m128i a0 = _mm_loadu_si128(src + 0);
+    const __m128i a1 = _mm_loadu_si128(src + 1);
+    const __m128i b0 = _mm_srli_epi32(a0, 8);
+    const __m128i b1 = _mm_srli_epi32(a1, 8);
+    const __m128i c0 = _mm_and_si128(b0, mask);
+    const __m128i c1 = _mm_and_si128(b1, mask);
+    const __m128i d = _mm_packs_epi32(c0, c1);
+    const __m128i e = _mm_packus_epi16(d, d);
+    _mm_storel_epi64((__m128i*)&alpha[i], e);
+    i += 8;
+  }
+  for (; i < size; ++i) alpha[i] = argb[i] >> 8;
+}
+
 //------------------------------------------------------------------------------
 // Non-dither premultiplied modes

@@ -177,7 +218,7 @@ static int ExtractAlpha_SSE2(const uint8_t* argb, int argb_stride,
 static void ApplyAlphaMultiply_SSE2(uint8_t* rgba, int alpha_first,
                                    int w, int h, int stride) {
  const __m128i zero = _mm_setzero_si128();
-  const __m128i kMult = _mm_set1_epi16(0x8081u);
+  const __m128i kMult = _mm_set1_epi16((short)0x8081);
  const __m128i kMask = _mm_set_epi16(0, 0xff, 0xff, 0, 0, 0xff, 0xff, 0);
  const int kSpan = 4;
  while (h-- > 0) {
@@ -266,7 +307,7 @@ static int HasAlpha32b_SSE2(const uint8_t* src, int length) {
 }

 static void AlphaReplace_SSE2(uint32_t* src, int length, uint32_t color) {
-  const __m128i m_color = _mm_set1_epi32(color);
+  const __m128i m_color = _mm_set1_epi32((int)color);
  const __m128i zero = _mm_setzero_si128();
  int i = 0;
  for (; i + 8 <= length; i += 8) {
@@ -317,7 +358,8 @@ static void MultARGBRow_SSE2(uint32_t* const ptr, int width, int inverse) {
  if (width > 0) WebPMultARGBRow_C(ptr + x, width, inverse);
 }

-static void MultRow_SSE2(uint8_t* const ptr, const uint8_t* const alpha,
+static void MultRow_SSE2(uint8_t* WEBP_RESTRICT const ptr,
+                         const uint8_t* WEBP_RESTRICT const alpha,
                         int width, int inverse) {
  int x = 0;
  if (!inverse) {
@@ -352,6 +394,7 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingSSE2(void) {
  WebPDispatchAlpha = DispatchAlpha_SSE2;
  WebPDispatchAlphaToGreen = DispatchAlphaToGreen_SSE2;
  WebPExtractAlpha = ExtractAlpha_SSE2;
+  WebPExtractGreen = ExtractGreen_SSE2;

  WebPHasAlpha8b = HasAlpha8b_SSE2;
  WebPHasAlpha32b = HasAlpha32b_SSE2;
@@ -19,14 +19,14 @@

 //------------------------------------------------------------------------------

-static int ExtractAlpha_SSE41(const uint8_t* argb, int argb_stride,
-                              int width, int height,
-                              uint8_t* alpha, int alpha_stride) {
+static int ExtractAlpha_SSE41(const uint8_t* WEBP_RESTRICT argb,
+                              int argb_stride, int width, int height,
+                              uint8_t* WEBP_RESTRICT alpha, int alpha_stride) {
  // alpha_and stores an 'and' operation of all the alpha[] values. The final
  // value is not 0xff if any of the alpha[] is not equal to 0xff.
  uint32_t alpha_and = 0xff;
  int i, j;
-  const __m128i all_0xff = _mm_set1_epi32(~0u);
+  const __m128i all_0xff = _mm_set1_epi32(~0);
  __m128i all_alphas = all_0xff;

  // We must be able to access 3 extra bytes after the last written byte
@@ -374,6 +374,7 @@ static void SetResidualCoeffs_C(const int16_t* const coeffs,
 VP8GetResidualCostFunc VP8GetResidualCost;
 VP8SetResidualCoeffsFunc VP8SetResidualCoeffs;

+extern VP8CPUInfo VP8GetCPUInfo;
 extern void VP8EncDspCostInitMIPS32(void);
 extern void VP8EncDspCostInitMIPSdspR2(void);
 extern void VP8EncDspCostInitSSE2(void);
@@ -395,12 +396,12 @@ WEBP_DSP_INIT_FUNC(VP8EncDspCostInit) {
      VP8EncDspCostInitMIPSdspR2();
    }
 #endif
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
    if (VP8GetCPUInfo(kSSE2)) {
      VP8EncDspCostInitSSE2();
    }
 #endif
-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
    if (VP8GetCPUInfo(kNEON)) {
      VP8EncDspCostInitNEON();
    }
@@ -29,7 +29,7 @@ static void SetResidualCoeffs_NEON(const int16_t* const coeffs,
  const uint8x16_t eob = vcombine_u8(vqmovn_u16(eob_0), vqmovn_u16(eob_1));
  const uint8x16_t masked = vandq_u8(eob, vld1q_u8(position));

-#ifdef __aarch64__
+#if WEBP_AARCH64
  res->last = vmaxvq_u8(masked) - 1;
 #else
  const uint8x8_t eob_8x8 = vmax_u8(vget_low_u8(masked), vget_high_u8(masked));
@@ -43,7 +43,7 @@ static void SetResidualCoeffs_NEON(const int16_t* const coeffs,

  vst1_lane_s32(&res->last, vreinterpret_s32_u32(eob_32x2), 0);
  --res->last;
-#endif  // __aarch64__
+#endif  // WEBP_AARCH64

  res->coeffs = coeffs;
 }
@@ -11,7 +11,7 @@
 //
 // Author: Christian Duvivier (cduvivier@google.com)

-#include "src/dsp/dsp.h"
+#include "src/dsp/cpu.h"

 #if defined(WEBP_HAVE_NEON_RTCD)
 #include <stdio.h>
@@ -173,6 +173,7 @@ static int x86CPUInfo(CPUFeature feature) {
  }
  return 0;
 }
+WEBP_EXTERN VP8CPUInfo VP8GetCPUInfo;
 VP8CPUInfo VP8GetCPUInfo = x86CPUInfo;
 #elif defined(WEBP_ANDROID_NEON)  // NB: needs to be before generic NEON test.
 static int AndroidCPUInfo(CPUFeature feature) {
@@ -184,22 +185,23 @@ static int AndroidCPUInfo(CPUFeature feature) {
  }
  return 0;
 }
+WEBP_EXTERN VP8CPUInfo VP8GetCPUInfo;
 VP8CPUInfo VP8GetCPUInfo = AndroidCPUInfo;
 #elif defined(EMSCRIPTEN) // also needs to be before generic NEON test
 // Use compile flags as an indicator of SIMD support instead of a runtime check.
 static int wasmCPUInfo(CPUFeature feature) {
  switch (feature) {
-#ifdef WEBP_USE_SSE2
+#ifdef WEBP_HAVE_SSE2
    case kSSE2:
      return 1;
 #endif
-#ifdef WEBP_USE_SSE41
+#ifdef WEBP_HAVE_SSE41
    case kSSE3:
    case kSlowSSSE3:
    case kSSE4_1:
      return 1;
 #endif
-#ifdef WEBP_USE_NEON
+#ifdef WEBP_HAVE_NEON
    case kNEON:
      return 1;
 #endif
@@ -208,10 +210,12 @@ static int wasmCPUInfo(CPUFeature feature) {
  }
  return 0;
 }
+WEBP_EXTERN VP8CPUInfo VP8GetCPUInfo;
 VP8CPUInfo VP8GetCPUInfo = wasmCPUInfo;
-#elif defined(WEBP_USE_NEON)
-// define a dummy function to enable turning off NEON at runtime by setting
-// VP8DecGetCPUInfo = NULL
+#elif defined(WEBP_HAVE_NEON)
+// In most cases this function doesn't check for NEON support (it's assumed by
+// the configuration), but enables turning off NEON at runtime, for testing
+// purposes, by setting VP8GetCPUInfo = NULL.
 static int armCPUInfo(CPUFeature feature) {
  if (feature != kNEON) return 0;
 #if defined(__linux__) && defined(WEBP_HAVE_NEON_RTCD)
@@ -235,6 +239,7 @@ static int armCPUInfo(CPUFeature feature) {
  return 1;
 #endif
 }
+WEBP_EXTERN VP8CPUInfo VP8GetCPUInfo;
 VP8CPUInfo VP8GetCPUInfo = armCPUInfo;
 #elif defined(WEBP_USE_MIPS32) || defined(WEBP_USE_MIPS_DSP_R2) || \
      defined(WEBP_USE_MSA)
@@ -246,7 +251,9 @@ static int mipsCPUInfo(CPUFeature feature) {
  }

 }
+WEBP_EXTERN VP8CPUInfo VP8GetCPUInfo;
 VP8CPUInfo VP8GetCPUInfo = mipsCPUInfo;
 #else
+WEBP_EXTERN VP8CPUInfo VP8GetCPUInfo;
 VP8CPUInfo VP8GetCPUInfo = NULL;
 #endif
@@ -0,0 +1,266 @@
+// Copyright 2022 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//   CPU detection functions and macros.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_DSP_CPU_H_
+#define WEBP_DSP_CPU_H_
+
+#include <stddef.h>
+
+#ifdef HAVE_CONFIG_H
+#include "src/webp/config.h"
+#endif
+
+#include "src/webp/types.h"
+
+#if defined(__GNUC__)
+#define LOCAL_GCC_VERSION ((__GNUC__ << 8) | __GNUC_MINOR__)
+#define LOCAL_GCC_PREREQ(maj, min) (LOCAL_GCC_VERSION >= (((maj) << 8) | (min)))
+#else
+#define LOCAL_GCC_VERSION 0
+#define LOCAL_GCC_PREREQ(maj, min) 0
+#endif
+
+#if defined(__clang__)
+#define LOCAL_CLANG_VERSION ((__clang_major__ << 8) | __clang_minor__)
+#define LOCAL_CLANG_PREREQ(maj, min) \
+  (LOCAL_CLANG_VERSION >= (((maj) << 8) | (min)))
+#else
+#define LOCAL_CLANG_VERSION 0
+#define LOCAL_CLANG_PREREQ(maj, min) 0
+#endif
+
+#ifndef __has_builtin
+#define __has_builtin(x) 0
+#endif
+
+//------------------------------------------------------------------------------
+// x86 defines.
+
+#if !defined(HAVE_CONFIG_H)
+#if defined(_MSC_VER) && _MSC_VER > 1310 && \
+    (defined(_M_X64) || defined(_M_IX86))
+#define WEBP_MSC_SSE2  // Visual C++ SSE2 targets
+#endif
+
+#if defined(_MSC_VER) && _MSC_VER >= 1500 && \
+    (defined(_M_X64) || defined(_M_IX86))
+#define WEBP_MSC_SSE41  // Visual C++ SSE4.1 targets
+#endif
+#endif
+
+// WEBP_HAVE_* are used to indicate the presence of the instruction set in dsp
+// files without intrinsics, allowing the corresponding Init() to be called.
+// Files containing intrinsics will need to be built targeting the instruction
+// set so should succeed on one of the earlier tests.
+#if (defined(__SSE2__) || defined(WEBP_MSC_SSE2)) && \
+    (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_SSE2))
+#define WEBP_USE_SSE2
+#endif
+
+#if defined(WEBP_USE_SSE2) && !defined(WEBP_HAVE_SSE2)
+#define WEBP_HAVE_SSE2
+#endif
+
+#if (defined(__SSE4_1__) || defined(WEBP_MSC_SSE41)) && \
+    (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_SSE41))
+#define WEBP_USE_SSE41
+#endif
+
+#if defined(WEBP_USE_SSE41) && !defined(WEBP_HAVE_SSE41)
+#define WEBP_HAVE_SSE41
+#endif
+
+#undef WEBP_MSC_SSE41
+#undef WEBP_MSC_SSE2
+
+//------------------------------------------------------------------------------
+// Arm defines.
+
+// The intrinsics currently cause compiler errors with arm-nacl-gcc and the
+// inline assembly would need to be modified for use with Native Client.
+#if ((defined(__ARM_NEON__) || defined(__aarch64__)) &&       \
+     (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_NEON))) && \
+    !defined(__native_client__)
+#define WEBP_USE_NEON
+#endif
+
+#if !defined(WEBP_USE_NEON) && defined(__ANDROID__) && \
+    defined(__ARM_ARCH_7A__) && defined(HAVE_CPU_FEATURES_H)
+#define WEBP_ANDROID_NEON  // Android targets that may have NEON
+#define WEBP_USE_NEON
+#endif
+
+// Note: ARM64 is supported in Visual Studio 2017, but requires the direct
+// inclusion of arm64_neon.h; Visual Studio 2019 includes this file in
+// arm_neon.h. Compile errors were seen with Visual Studio 2019 16.4 with
+// vtbl4_u8(); a fix was made in 16.6.
+#if defined(_MSC_VER) && \
+    ((_MSC_VER >= 1700 && defined(_M_ARM)) || \
+     (_MSC_VER >= 1926 && (defined(_M_ARM64) || defined(_M_ARM64EC))))
+#define WEBP_USE_NEON
+#define WEBP_USE_INTRINSICS
+#endif
+
+#if defined(__aarch64__) || defined(_M_ARM64) || defined(_M_ARM64EC)
+#define WEBP_AARCH64 1
+#else
+#define WEBP_AARCH64 0
+#endif
+
+#if defined(WEBP_USE_NEON) && !defined(WEBP_HAVE_NEON)
+#define WEBP_HAVE_NEON
+#endif
+
+//------------------------------------------------------------------------------
+// MIPS defines.
+
+#if defined(__mips__) && !defined(__mips64) && defined(__mips_isa_rev) && \
+    (__mips_isa_rev >= 1) && (__mips_isa_rev < 6)
+#define WEBP_USE_MIPS32
+#if (__mips_isa_rev >= 2)
+#define WEBP_USE_MIPS32_R2
+#if defined(__mips_dspr2) || (defined(__mips_dsp_rev) && __mips_dsp_rev >= 2)
+#define WEBP_USE_MIPS_DSP_R2
+#endif
+#endif
+#endif
+
+#if defined(__mips_msa) && defined(__mips_isa_rev) && (__mips_isa_rev >= 5)
+#define WEBP_USE_MSA
+#endif
+
+//------------------------------------------------------------------------------
+
+#ifndef WEBP_DSP_OMIT_C_CODE
+#define WEBP_DSP_OMIT_C_CODE 1
+#endif
+
+#if defined(WEBP_USE_NEON) && WEBP_DSP_OMIT_C_CODE
+#define WEBP_NEON_OMIT_C_CODE 1
+#else
+#define WEBP_NEON_OMIT_C_CODE 0
+#endif
+
+#if !(LOCAL_CLANG_PREREQ(3, 8) || LOCAL_GCC_PREREQ(4, 8) || WEBP_AARCH64)
+#define WEBP_NEON_WORK_AROUND_GCC 1
+#else
+#define WEBP_NEON_WORK_AROUND_GCC 0
+#endif
+
+//------------------------------------------------------------------------------
+
+// This macro prevents thread_sanitizer from reporting known concurrent writes.
+#define WEBP_TSAN_IGNORE_FUNCTION
+#if defined(__has_feature)
+#if __has_feature(thread_sanitizer)
+#undef WEBP_TSAN_IGNORE_FUNCTION
+#define WEBP_TSAN_IGNORE_FUNCTION __attribute__((no_sanitize_thread))
+#endif
+#endif
+
+#if defined(__has_feature)
+#if __has_feature(memory_sanitizer)
+#define WEBP_MSAN
+#endif
+#endif
+
+#if defined(WEBP_USE_THREAD) && !defined(_WIN32)
+#include <pthread.h>  // NOLINT
+
+#define WEBP_DSP_INIT(func)                                         \
+  do {                                                              \
+    static volatile VP8CPUInfo func##_last_cpuinfo_used =           \
+        (VP8CPUInfo)&func##_last_cpuinfo_used;                      \
+    static pthread_mutex_t func##_lock = PTHREAD_MUTEX_INITIALIZER; \
+    if (pthread_mutex_lock(&func##_lock)) break;                    \
+    if (func##_last_cpuinfo_used != VP8GetCPUInfo) func();          \
+    func##_last_cpuinfo_used = VP8GetCPUInfo;                       \
+    (void)pthread_mutex_unlock(&func##_lock);                       \
+  } while (0)
+#else  // !(defined(WEBP_USE_THREAD) && !defined(_WIN32))
+#define WEBP_DSP_INIT(func)                               \
+  do {                                                    \
+    static volatile VP8CPUInfo func##_last_cpuinfo_used = \
+        (VP8CPUInfo)&func##_last_cpuinfo_used;            \
+    if (func##_last_cpuinfo_used == VP8GetCPUInfo) break; \
+    func();                                               \
+    func##_last_cpuinfo_used = VP8GetCPUInfo;             \
+  } while (0)
+#endif  // defined(WEBP_USE_THREAD) && !defined(_WIN32)
+
+// Defines an Init + helper function that control multiple initialization of
+// function pointers / tables.
+/* Usage:
+   WEBP_DSP_INIT_FUNC(InitFunc) {
+     ...function body
+   }
+*/
+#define WEBP_DSP_INIT_FUNC(name)                                            \
+  static WEBP_TSAN_IGNORE_FUNCTION void name##_body(void);                  \
+  WEBP_TSAN_IGNORE_FUNCTION void name(void) { WEBP_DSP_INIT(name##_body); } \
+  static WEBP_TSAN_IGNORE_FUNCTION void name##_body(void)
+
+#define WEBP_UBSAN_IGNORE_UNDEF
+#define WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW
+#if defined(__clang__) && defined(__has_attribute)
+#if __has_attribute(no_sanitize)
+// This macro prevents the undefined behavior sanitizer from reporting
+// failures. This is only meant to silence unaligned loads on platforms that
+// are known to support them.
+#undef WEBP_UBSAN_IGNORE_UNDEF
+#define WEBP_UBSAN_IGNORE_UNDEF __attribute__((no_sanitize("undefined")))
+
+// This macro prevents the undefined behavior sanitizer from reporting
+// failures related to unsigned integer overflows. This is only meant to
+// silence cases where this well defined behavior is expected.
+#undef WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW
+#define WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW \
+  __attribute__((no_sanitize("unsigned-integer-overflow")))
+#endif
+#endif
+
+// If 'ptr' is NULL, returns NULL. Otherwise returns 'ptr + off'.
+// Prevents undefined behavior sanitizer nullptr-with-nonzero-offset warning.
+#if !defined(WEBP_OFFSET_PTR)
+#define WEBP_OFFSET_PTR(ptr, off) (((ptr) == NULL) ? NULL : ((ptr) + (off)))
+#endif
+
+// Regularize the definition of WEBP_SWAP_16BIT_CSP (backward compatibility)
+#if !defined(WEBP_SWAP_16BIT_CSP)
+#define WEBP_SWAP_16BIT_CSP 0
+#endif
+
+// some endian fix (e.g.: mips-gcc doesn't define __BIG_ENDIAN__)
+#if !defined(WORDS_BIGENDIAN) &&                   \
+    (defined(__BIG_ENDIAN__) || defined(_M_PPC) || \
+     (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)))
+#define WORDS_BIGENDIAN
+#endif
+
+typedef enum {
+  kSSE2,
+  kSSE3,
+  kSlowSSSE3,  // special feature for slow SSSE3 architectures
+  kSSE4_1,
+  kAVX,
+  kAVX2,
+  kNEON,
+  kMIPS32,
+  kMIPSdspR2,
+  kMSA
+} CPUFeature;
+
+// returns true if the CPU supports the feature.
+typedef int (*VP8CPUInfo)(CPUFeature feature);
+
+#endif  // WEBP_DSP_CPU_H_
@@ -734,6 +734,7 @@ VP8SimpleFilterFunc VP8SimpleHFilter16i;
 void (*VP8DitherCombine8x8)(const uint8_t* dither, uint8_t* dst,
                            int dst_stride);

+extern VP8CPUInfo VP8GetCPUInfo;
 extern void VP8DspInitSSE2(void);
 extern void VP8DspInitSSE41(void);
 extern void VP8DspInitNEON(void);
@@ -807,10 +808,10 @@ WEBP_DSP_INIT_FUNC(VP8DspInit) {

  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
  if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
    if (VP8GetCPUInfo(kSSE2)) {
      VP8DspInitSSE2();
-#if defined(WEBP_USE_SSE41)
+#if defined(WEBP_HAVE_SSE41)
      if (VP8GetCPUInfo(kSSE4_1)) {
        VP8DspInitSSE41();
      }
@@ -834,7 +835,7 @@ WEBP_DSP_INIT_FUNC(VP8DspInit) {
 #endif
  }

-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
  if (WEBP_NEON_OMIT_C_CODE ||
      (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
    VP8DspInitNEON();
@@ -1428,7 +1428,7 @@ static WEBP_INLINE void DC8_NEON(uint8_t* dst, int do_top, int do_left) {

  if (do_top) {
    const uint8x8_t A = vld1_u8(dst - BPS);  // top row
-#if defined(__aarch64__)
+#if WEBP_AARCH64
    const uint16_t p2 = vaddlv_u8(A);
    sum_top = vdupq_n_u16(p2);
 #else
@@ -1511,7 +1511,7 @@ static WEBP_INLINE void DC16_NEON(uint8_t* dst, int do_top, int do_left) {

  if (do_top) {
    const uint8x16_t A = vld1q_u8(dst - BPS);  // top row
-#if defined(__aarch64__)
+#if WEBP_AARCH64
    const uint16_t p3 = vaddlvq_u8(A);
    sum_top = vdupq_n_u16(p3);
 #else
@@ -158,10 +158,10 @@ static void Transform_SSE2(const int16_t* in, uint8_t* dst, int do_two) {
      dst3 = _mm_loadl_epi64((__m128i*)(dst + 3 * BPS));
    } else {
      // Load four bytes/pixels per line.
-      dst0 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 0 * BPS));
-      dst1 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 1 * BPS));
-      dst2 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 2 * BPS));
-      dst3 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 3 * BPS));
+      dst0 = _mm_cvtsi32_si128(WebPMemToInt32(dst + 0 * BPS));
+      dst1 = _mm_cvtsi32_si128(WebPMemToInt32(dst + 1 * BPS));
+      dst2 = _mm_cvtsi32_si128(WebPMemToInt32(dst + 2 * BPS));
+      dst3 = _mm_cvtsi32_si128(WebPMemToInt32(dst + 3 * BPS));
    }
    // Convert to 16b.
    dst0 = _mm_unpacklo_epi8(dst0, zero);
@@ -187,10 +187,10 @@ static void Transform_SSE2(const int16_t* in, uint8_t* dst, int do_two) {
      _mm_storel_epi64((__m128i*)(dst + 3 * BPS), dst3);
    } else {
      // Store four bytes/pixels per line.
-      WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(dst0));
-      WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(dst1));
-      WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(dst2));
-      WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(dst3));
+      WebPInt32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(dst0));
+      WebPInt32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(dst1));
+      WebPInt32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(dst2));
+      WebPInt32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(dst3));
    }
  }
 }
@@ -213,10 +213,10 @@ static void TransformAC3(const int16_t* in, uint8_t* dst) {
  const __m128i m3 = _mm_subs_epi16(B, d4);
  const __m128i zero = _mm_setzero_si128();
  // Load the source pixels.
-  __m128i dst0 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 0 * BPS));
-  __m128i dst1 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 1 * BPS));
-  __m128i dst2 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 2 * BPS));
-  __m128i dst3 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 3 * BPS));
+  __m128i dst0 = _mm_cvtsi32_si128(WebPMemToInt32(dst + 0 * BPS));
+  __m128i dst1 = _mm_cvtsi32_si128(WebPMemToInt32(dst + 1 * BPS));
+  __m128i dst2 = _mm_cvtsi32_si128(WebPMemToInt32(dst + 2 * BPS));
+  __m128i dst3 = _mm_cvtsi32_si128(WebPMemToInt32(dst + 3 * BPS));
  // Convert to 16b.
  dst0 = _mm_unpacklo_epi8(dst0, zero);
  dst1 = _mm_unpacklo_epi8(dst1, zero);
@@ -233,10 +233,10 @@ static void TransformAC3(const int16_t* in, uint8_t* dst) {
  dst2 = _mm_packus_epi16(dst2, dst2);
  dst3 = _mm_packus_epi16(dst3, dst3);
  // Store the results.
-  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(dst0));
-  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(dst1));
-  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(dst2));
-  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(dst3));
+  WebPInt32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(dst0));
+  WebPInt32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(dst1));
+  WebPInt32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(dst2));
+  WebPInt32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(dst3));
 }
 #undef MUL
 #endif   // USE_TRANSFORM_AC3
@@ -477,11 +477,11 @@ static WEBP_INLINE void Load8x4_SSE2(const uint8_t* const b, int stride,
  // A0 = 63 62 61 60 23 22 21 20 43 42 41 40 03 02 01 00
  // A1 = 73 72 71 70 33 32 31 30 53 52 51 50 13 12 11 10
  const __m128i A0 = _mm_set_epi32(
-      WebPMemToUint32(&b[6 * stride]), WebPMemToUint32(&b[2 * stride]),
-      WebPMemToUint32(&b[4 * stride]), WebPMemToUint32(&b[0 * stride]));
+      WebPMemToInt32(&b[6 * stride]), WebPMemToInt32(&b[2 * stride]),
+      WebPMemToInt32(&b[4 * stride]), WebPMemToInt32(&b[0 * stride]));
  const __m128i A1 = _mm_set_epi32(
-      WebPMemToUint32(&b[7 * stride]), WebPMemToUint32(&b[3 * stride]),
-      WebPMemToUint32(&b[5 * stride]), WebPMemToUint32(&b[1 * stride]));
+      WebPMemToInt32(&b[7 * stride]), WebPMemToInt32(&b[3 * stride]),
+      WebPMemToInt32(&b[5 * stride]), WebPMemToInt32(&b[1 * stride]));

  // B0 = 53 43 52 42 51 41 50 40 13 03 12 02 11 01 10 00
  // B1 = 73 63 72 62 71 61 70 60 33 23 32 22 31 21 30 20
@@ -540,7 +540,7 @@ static WEBP_INLINE void Store4x4_SSE2(__m128i* const x,
                                      uint8_t* dst, int stride) {
  int i;
  for (i = 0; i < 4; ++i, dst += stride) {
-    WebPUint32ToMem(dst, _mm_cvtsi128_si32(*x));
+    WebPInt32ToMem(dst, _mm_cvtsi128_si32(*x));
    *x = _mm_srli_si128(*x, 4);
  }
 }
@@ -908,10 +908,10 @@ static void VE4_SSE2(uint8_t* dst) {    // vertical
  const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGH00), one);
  const __m128i b = _mm_subs_epu8(a, lsb);
  const __m128i avg = _mm_avg_epu8(b, BCDEFGH0);
-  const uint32_t vals = _mm_cvtsi128_si32(avg);
+  const int vals = _mm_cvtsi128_si32(avg);
  int i;
  for (i = 0; i < 4; ++i) {
-    WebPUint32ToMem(dst + i * BPS, vals);
+    WebPInt32ToMem(dst + i * BPS, vals);
  }
 }

@@ -925,10 +925,10 @@ static void LD4_SSE2(uint8_t* dst) {   // Down-Left
  const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGHH0), one);
  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
  const __m128i abcdefg = _mm_avg_epu8(avg2, BCDEFGH0);
-  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcdefg    ));
-  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
-  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
-  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
+  WebPInt32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcdefg    ));
+  WebPInt32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
+  WebPInt32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
+  WebPInt32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
 }

 static void VR4_SSE2(uint8_t* dst) {   // Vertical-Right
@@ -946,10 +946,10 @@ static void VR4_SSE2(uint8_t* dst) {   // Vertical-Right
  const __m128i lsb = _mm_and_si128(_mm_xor_si128(IXABCD, ABCD0), one);
  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
  const __m128i efgh = _mm_avg_epu8(avg2, XABCD);
-  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcd    ));
-  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               efgh    ));
-  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1)));
-  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1)));
+  WebPInt32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcd    ));
+  WebPInt32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               efgh    ));
+  WebPInt32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1)));
+  WebPInt32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1)));

  // these two are hard to implement in SSE2, so we keep the C-version:
  DST(0, 2) = AVG3(J, I, X);
@@ -970,11 +970,12 @@ static void VL4_SSE2(uint8_t* dst) {   // Vertical-Left
  const __m128i abbc = _mm_or_si128(ab, bc);
  const __m128i lsb2 = _mm_and_si128(abbc, lsb1);
  const __m128i avg4 = _mm_subs_epu8(avg3, lsb2);
-  const uint32_t extra_out = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 4));
-  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               avg1    ));
-  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               avg4    ));
-  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1)));
-  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1)));
+  const uint32_t extra_out =
+      (uint32_t)_mm_cvtsi128_si32(_mm_srli_si128(avg4, 4));
+  WebPInt32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               avg1    ));
+  WebPInt32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               avg4    ));
+  WebPInt32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1)));
+  WebPInt32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1)));

  // these two are hard to get and irregular
  DST(3, 2) = (extra_out >> 0) & 0xff;
@@ -990,7 +991,7 @@ static void RD4_SSE2(uint8_t* dst) {   // Down-right
  const uint32_t K = dst[-1 + 2 * BPS];
  const uint32_t L = dst[-1 + 3 * BPS];
  const __m128i LKJI_____ =
-      _mm_cvtsi32_si128(L | (K << 8) | (J << 16) | (I << 24));
+      _mm_cvtsi32_si128((int)(L | (K << 8) | (J << 16) | (I << 24)));
  const __m128i LKJIXABCD = _mm_or_si128(LKJI_____, ____XABCD);
  const __m128i KJIXABCD_ = _mm_srli_si128(LKJIXABCD, 1);
  const __m128i JIXABCD__ = _mm_srli_si128(LKJIXABCD, 2);
@@ -998,10 +999,10 @@ static void RD4_SSE2(uint8_t* dst) {   // Down-right
  const __m128i lsb = _mm_and_si128(_mm_xor_si128(JIXABCD__, LKJIXABCD), one);
  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
  const __m128i abcdefg = _mm_avg_epu8(avg2, KJIXABCD_);
-  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(               abcdefg    ));
-  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
-  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
-  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
+  WebPInt32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(               abcdefg    ));
+  WebPInt32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
+  WebPInt32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
+  WebPInt32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
 }

 #undef DST
@@ -1015,13 +1016,13 @@ static WEBP_INLINE void TrueMotion_SSE2(uint8_t* dst, int size) {
  const __m128i zero = _mm_setzero_si128();
  int y;
  if (size == 4) {
-    const __m128i top_values = _mm_cvtsi32_si128(WebPMemToUint32(top));
+    const __m128i top_values = _mm_cvtsi32_si128(WebPMemToInt32(top));
    const __m128i top_base = _mm_unpacklo_epi8(top_values, zero);
    for (y = 0; y < 4; ++y, dst += BPS) {
      const int val = dst[-1] - top[-1];
      const __m128i base = _mm_set1_epi16(val);
      const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero);
-      WebPUint32ToMem(dst, _mm_cvtsi128_si32(out));
+      WebPInt32ToMem(dst, _mm_cvtsi128_si32(out));
    }
  } else if (size == 8) {
    const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
@@ -1062,7 +1063,7 @@ static void VE16_SSE2(uint8_t* dst) {
 static void HE16_SSE2(uint8_t* dst) {     // horizontal
  int j;
  for (j = 16; j > 0; --j) {
-    const __m128i values = _mm_set1_epi8(dst[-1]);
+    const __m128i values = _mm_set1_epi8((char)dst[-1]);
    _mm_storeu_si128((__m128i*)dst, values);
    dst += BPS;
  }
@@ -1070,7 +1071,7 @@ static void HE16_SSE2(uint8_t* dst) {     // horizontal

 static WEBP_INLINE void Put16_SSE2(uint8_t v, uint8_t* dst) {
  int j;
-  const __m128i values = _mm_set1_epi8(v);
+  const __m128i values = _mm_set1_epi8((char)v);
  for (j = 0; j < 16; ++j) {
    _mm_storeu_si128((__m128i*)(dst + j * BPS), values);
  }
@@ -1130,7 +1131,7 @@ static void VE8uv_SSE2(uint8_t* dst) {    // vertical
 // helper for chroma-DC predictions
 static WEBP_INLINE void Put8x8uv_SSE2(uint8_t v, uint8_t* dst) {
  int j;
-  const __m128i values = _mm_set1_epi8(v);
+  const __m128i values = _mm_set1_epi8((char)v);
  for (j = 0; j < 8; ++j) {
    _mm_storel_epi64((__m128i*)(dst + j * BPS), values);
  }
@@ -23,7 +23,7 @@ static void HE16_SSE41(uint8_t* dst) {     // horizontal
  int j;
  const __m128i kShuffle3 = _mm_set1_epi8(3);
  for (j = 16; j > 0; --j) {
-    const __m128i in = _mm_cvtsi32_si128(WebPMemToUint32(dst - 4));
+    const __m128i in = _mm_cvtsi32_si128(WebPMemToInt32(dst - 4));
    const __m128i values = _mm_shuffle_epi8(in, kShuffle3);
    _mm_storeu_si128((__m128i*)dst, values);
    dst += BPS;
@@ -18,6 +18,7 @@
 #include "src/webp/config.h"
 #endif

+#include "src/dsp/cpu.h"
 #include "src/webp/types.h"

 #ifdef __cplusplus
@@ -27,205 +28,22 @@ extern "C" {
 #define BPS 32   // this is the common stride for enc/dec

 //------------------------------------------------------------------------------
-// CPU detection
+// WEBP_RESTRICT

+// Declares a pointer with the restrict type qualifier if available.
+// This allows code to hint to the compiler that only this pointer references a
+// particular object or memory region within the scope of the block in which it
+// is declared. This may allow for improved optimizations due to the lack of
+// pointer aliasing. See also:
+// https://en.cppreference.com/w/c/language/restrict
 #if defined(__GNUC__)
-# define LOCAL_GCC_VERSION ((__GNUC__ << 8) | __GNUC_MINOR__)
-# define LOCAL_GCC_PREREQ(maj, min) \
-    (LOCAL_GCC_VERSION >= (((maj) << 8) | (min)))
+#define WEBP_RESTRICT __restrict__
+#elif defined(_MSC_VER)
+#define WEBP_RESTRICT __restrict
 #else
-# define LOCAL_GCC_VERSION 0
-# define LOCAL_GCC_PREREQ(maj, min) 0
+#define WEBP_RESTRICT
 #endif

-#if defined(__clang__)
-# define LOCAL_CLANG_VERSION ((__clang_major__ << 8) | __clang_minor__)
-# define LOCAL_CLANG_PREREQ(maj, min) \
-    (LOCAL_CLANG_VERSION >= (((maj) << 8) | (min)))
-#else
-# define LOCAL_CLANG_VERSION 0
-# define LOCAL_CLANG_PREREQ(maj, min) 0
-#endif
-
-#ifndef __has_builtin
-# define __has_builtin(x) 0
-#endif
-
-#if !defined(HAVE_CONFIG_H)
-#if defined(_MSC_VER) && _MSC_VER > 1310 && \
-    (defined(_M_X64) || defined(_M_IX86))
-#define WEBP_MSC_SSE2  // Visual C++ SSE2 targets
-#endif
-
-#if defined(_MSC_VER) && _MSC_VER >= 1500 && \
-    (defined(_M_X64) || defined(_M_IX86))
-#define WEBP_MSC_SSE41  // Visual C++ SSE4.1 targets
-#endif
-#endif
-
-// WEBP_HAVE_* are used to indicate the presence of the instruction set in dsp
-// files without intrinsics, allowing the corresponding Init() to be called.
-// Files containing intrinsics will need to be built targeting the instruction
-// set so should succeed on one of the earlier tests.
-#if defined(__SSE2__) || defined(WEBP_MSC_SSE2) || defined(WEBP_HAVE_SSE2)
-#define WEBP_USE_SSE2
-#endif
-
-#if defined(__SSE4_1__) || defined(WEBP_MSC_SSE41) || defined(WEBP_HAVE_SSE41)
-#define WEBP_USE_SSE41
-#endif
-
-#undef WEBP_MSC_SSE41
-#undef WEBP_MSC_SSE2
-
-// The intrinsics currently cause compiler errors with arm-nacl-gcc and the
-// inline assembly would need to be modified for use with Native Client.
-#if (defined(__ARM_NEON__) || \
-     defined(__aarch64__) || defined(WEBP_HAVE_NEON)) && \
-    !defined(__native_client__)
-#define WEBP_USE_NEON
-#endif
-
-#if !defined(WEBP_USE_NEON) && defined(__ANDROID__) && \
-    defined(__ARM_ARCH_7A__) && defined(HAVE_CPU_FEATURES_H)
-#define WEBP_ANDROID_NEON  // Android targets that may have NEON
-#define WEBP_USE_NEON
-#endif
-
-#if defined(_MSC_VER) && _MSC_VER >= 1700 && defined(_M_ARM)
-#define WEBP_USE_NEON
-#define WEBP_USE_INTRINSICS
-#endif
-
-#if defined(__mips__) && !defined(__mips64) && \
-    defined(__mips_isa_rev) && (__mips_isa_rev >= 1) && (__mips_isa_rev < 6)
-#define WEBP_USE_MIPS32
-#if (__mips_isa_rev >= 2)
-#define WEBP_USE_MIPS32_R2
-#if defined(__mips_dspr2) || (defined(__mips_dsp_rev) && __mips_dsp_rev >= 2)
-#define WEBP_USE_MIPS_DSP_R2
-#endif
-#endif
-#endif
-
-#if defined(__mips_msa) && defined(__mips_isa_rev) && (__mips_isa_rev >= 5)
-#define WEBP_USE_MSA
-#endif
-
-#ifndef WEBP_DSP_OMIT_C_CODE
-#define WEBP_DSP_OMIT_C_CODE 1
-#endif
-
-#if (defined(__aarch64__) || defined(__ARM_NEON__)) && WEBP_DSP_OMIT_C_CODE
-#define WEBP_NEON_OMIT_C_CODE 1
-#else
-#define WEBP_NEON_OMIT_C_CODE 0
-#endif
-
-#if !(LOCAL_CLANG_PREREQ(3,8) || LOCAL_GCC_PREREQ(4,8) || defined(__aarch64__))
-#define WEBP_NEON_WORK_AROUND_GCC 1
-#else
-#define WEBP_NEON_WORK_AROUND_GCC 0
-#endif
-
-// This macro prevents thread_sanitizer from reporting known concurrent writes.
-#define WEBP_TSAN_IGNORE_FUNCTION
-#if defined(__has_feature)
-#if __has_feature(thread_sanitizer)
-#undef WEBP_TSAN_IGNORE_FUNCTION
-#define WEBP_TSAN_IGNORE_FUNCTION __attribute__((no_sanitize_thread))
-#endif
-#endif
-
-#if defined(WEBP_USE_THREAD) && !defined(_WIN32)
-#include <pthread.h>  // NOLINT
-
-#define WEBP_DSP_INIT(func) do {                                    \
-  static volatile VP8CPUInfo func ## _last_cpuinfo_used =           \
-      (VP8CPUInfo)&func ## _last_cpuinfo_used;                      \
-  static pthread_mutex_t func ## _lock = PTHREAD_MUTEX_INITIALIZER; \
-  if (pthread_mutex_lock(&func ## _lock)) break;                    \
-  if (func ## _last_cpuinfo_used != VP8GetCPUInfo) func();          \
-  func ## _last_cpuinfo_used = VP8GetCPUInfo;                       \
-  (void)pthread_mutex_unlock(&func ## _lock);                       \
-} while (0)
-#else  // !(defined(WEBP_USE_THREAD) && !defined(_WIN32))
-#define WEBP_DSP_INIT(func) do {                                    \
-  static volatile VP8CPUInfo func ## _last_cpuinfo_used =           \
-      (VP8CPUInfo)&func ## _last_cpuinfo_used;                      \
-  if (func ## _last_cpuinfo_used == VP8GetCPUInfo) break;           \
-  func();                                                           \
-  func ## _last_cpuinfo_used = VP8GetCPUInfo;                       \
-} while (0)
-#endif  // defined(WEBP_USE_THREAD) && !defined(_WIN32)
-
-// Defines an Init + helper function that control multiple initialization of
-// function pointers / tables.
-/* Usage:
-   WEBP_DSP_INIT_FUNC(InitFunc) {
-     ...function body
-   }
-*/
-#define WEBP_DSP_INIT_FUNC(name)                             \
-  static WEBP_TSAN_IGNORE_FUNCTION void name ## _body(void); \
-  WEBP_TSAN_IGNORE_FUNCTION void name(void) {                \
-    WEBP_DSP_INIT(name ## _body);                            \
-  }                                                          \
-  static WEBP_TSAN_IGNORE_FUNCTION void name ## _body(void)
-
-#define WEBP_UBSAN_IGNORE_UNDEF
-#define WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW
-#if defined(__clang__) && defined(__has_attribute)
-#if __has_attribute(no_sanitize)
-// This macro prevents the undefined behavior sanitizer from reporting
-// failures. This is only meant to silence unaligned loads on platforms that
-// are known to support them.
-#undef WEBP_UBSAN_IGNORE_UNDEF
-#define WEBP_UBSAN_IGNORE_UNDEF \
-  __attribute__((no_sanitize("undefined")))
-
-// This macro prevents the undefined behavior sanitizer from reporting
-// failures related to unsigned integer overflows. This is only meant to
-// silence cases where this well defined behavior is expected.
-#undef WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW
-#define WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW \
-  __attribute__((no_sanitize("unsigned-integer-overflow")))
-#endif
-#endif
-
-// If 'ptr' is NULL, returns NULL. Otherwise returns 'ptr + off'.
-// Prevents undefined behavior sanitizer nullptr-with-nonzero-offset warning.
-#if !defined(WEBP_OFFSET_PTR)
-#define WEBP_OFFSET_PTR(ptr, off) (((ptr) == NULL) ? NULL : ((ptr) + (off)))
-#endif
-
-// Regularize the definition of WEBP_SWAP_16BIT_CSP (backward compatibility)
-#if !defined(WEBP_SWAP_16BIT_CSP)
-#define WEBP_SWAP_16BIT_CSP 0
-#endif
-
-// some endian fix (e.g.: mips-gcc doesn't define __BIG_ENDIAN__)
-#if !defined(WORDS_BIGENDIAN) && \
-    (defined(__BIG_ENDIAN__) || defined(_M_PPC) || \
-     (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)))
-#define WORDS_BIGENDIAN
-#endif
-
-typedef enum {
-  kSSE2,
-  kSSE3,
-  kSlowSSSE3,  // special feature for slow SSSE3 architectures
-  kSSE4_1,
-  kAVX,
-  kAVX2,
-  kNEON,
-  kMIPS32,
-  kMIPSdspR2,
-  kMSA
-} CPUFeature;
-// returns true if the CPU supports the feature.
-typedef int (*VP8CPUInfo)(CPUFeature feature);
-WEBP_EXTERN VP8CPUInfo VP8GetCPUInfo;

 //------------------------------------------------------------------------------
 // Init stub generator
@@ -514,15 +332,6 @@ extern void WebPConvertARGBToUV_C(const uint32_t* argb, uint8_t* u, uint8_t* v,
 extern void WebPConvertRGBA32ToUV_C(const uint16_t* rgb,
                                    uint8_t* u, uint8_t* v, int width);

-// utilities for accurate RGB->YUV conversion
-extern uint64_t (*WebPSharpYUVUpdateY)(const uint16_t* src, const uint16_t* ref,
-                                       uint16_t* dst, int len);
-extern void (*WebPSharpYUVUpdateRGB)(const int16_t* src, const int16_t* ref,
-                                     int16_t* dst, int len);
-extern void (*WebPSharpYUVFilterRow)(const int16_t* A, const int16_t* B,
-                                     int len,
-                                     const uint16_t* best_y, uint16_t* out);
-
 // Must be called before using the above.
 void WebPInitConvertARGBToYUV(void);

@@ -578,26 +387,29 @@ extern void (*WebPApplyAlphaMultiply4444)(

 // Dispatch the values from alpha[] plane to the ARGB destination 'dst'.
 // Returns true if alpha[] plane has non-trivial values different from 0xff.
-extern int (*WebPDispatchAlpha)(const uint8_t* alpha, int alpha_stride,
-                                int width, int height,
-                                uint8_t* dst, int dst_stride);
+extern int (*WebPDispatchAlpha)(const uint8_t* WEBP_RESTRICT alpha,
+                                int alpha_stride, int width, int height,
+                                uint8_t* WEBP_RESTRICT dst, int dst_stride);

 // Transfer packed 8b alpha[] values to green channel in dst[], zero'ing the
 // A/R/B values. 'dst_stride' is the stride for dst[] in uint32_t units.
-extern void (*WebPDispatchAlphaToGreen)(const uint8_t* alpha, int alpha_stride,
-                                        int width, int height,
-                                        uint32_t* dst, int dst_stride);
+extern void (*WebPDispatchAlphaToGreen)(const uint8_t* WEBP_RESTRICT alpha,
+                                        int alpha_stride, int width, int height,
+                                        uint32_t* WEBP_RESTRICT dst,
+                                        int dst_stride);

 // Extract the alpha values from 32b values in argb[] and pack them into alpha[]
 // (this is the opposite of WebPDispatchAlpha).
 // Returns true if there's only trivial 0xff alpha values.
-extern int (*WebPExtractAlpha)(const uint8_t* argb, int argb_stride,
-                               int width, int height,
-                               uint8_t* alpha, int alpha_stride);
+extern int (*WebPExtractAlpha)(const uint8_t* WEBP_RESTRICT argb,
+                               int argb_stride, int width, int height,
+                               uint8_t* WEBP_RESTRICT alpha,
+                               int alpha_stride);

 // Extract the green values from 32b values in argb[] and pack them into alpha[]
 // (this is the opposite of WebPDispatchAlphaToGreen).
-extern void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size);
+extern void (*WebPExtractGreen)(const uint32_t* WEBP_RESTRICT argb,
+                                uint8_t* WEBP_RESTRICT alpha, int size);

 // Pre-Multiply operation transforms x into x * A / 255  (where x=Y,R,G or B).
 // Un-Multiply operation transforms x into x * 255 / A.
@@ -610,29 +422,35 @@ void WebPMultARGBRows(uint8_t* ptr, int stride, int width, int num_rows,
                      int inverse);

 // Same for a row of single values, with side alpha values.
-extern void (*WebPMultRow)(uint8_t* const ptr, const uint8_t* const alpha,
+extern void (*WebPMultRow)(uint8_t* WEBP_RESTRICT const ptr,
+                           const uint8_t* WEBP_RESTRICT const alpha,
                           int width, int inverse);

 // Same a WebPMultRow(), but for several 'num_rows' rows.
-void WebPMultRows(uint8_t* ptr, int stride,
-                  const uint8_t* alpha, int alpha_stride,
+void WebPMultRows(uint8_t* WEBP_RESTRICT ptr, int stride,
+                  const uint8_t* WEBP_RESTRICT alpha, int alpha_stride,
                  int width, int num_rows, int inverse);

 // Plain-C versions, used as fallback by some implementations.
-void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha,
+void WebPMultRow_C(uint8_t* WEBP_RESTRICT const ptr,
+                   const uint8_t* WEBP_RESTRICT const alpha,
                   int width, int inverse);
 void WebPMultARGBRow_C(uint32_t* const ptr, int width, int inverse);

 #ifdef WORDS_BIGENDIAN
 // ARGB packing function: a/r/g/b input is rgba or bgra order.
-extern void (*WebPPackARGB)(const uint8_t* a, const uint8_t* r,
-                            const uint8_t* g, const uint8_t* b, int len,
-                            uint32_t* out);
+extern void (*WebPPackARGB)(const uint8_t* WEBP_RESTRICT a,
+                            const uint8_t* WEBP_RESTRICT r,
+                            const uint8_t* WEBP_RESTRICT g,
+                            const uint8_t* WEBP_RESTRICT b,
+                            int len, uint32_t* WEBP_RESTRICT out);
 #endif

 // RGB packing function. 'step' can be 3 or 4. r/g/b input is rgb or bgr order.
-extern void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b,
-                           int len, int step, uint32_t* out);
+extern void (*WebPPackRGB)(const uint8_t* WEBP_RESTRICT r,
+                           const uint8_t* WEBP_RESTRICT g,
+                           const uint8_t* WEBP_RESTRICT b,
+                           int len, int step, uint32_t* WEBP_RESTRICT out);

 // This function returns true if src[i] contains a value different from 0xff.
 extern int (*WebPHasAlpha8b)(const uint8_t* src, int length);
@@ -732,6 +732,7 @@ VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
 VP8BlockCopy VP8Copy4x4;
 VP8BlockCopy VP8Copy16x8;

+extern VP8CPUInfo VP8GetCPUInfo;
 extern void VP8EncDspInitSSE2(void);
 extern void VP8EncDspInitSSE41(void);
 extern void VP8EncDspInitNEON(void);
@@ -773,10 +774,10 @@ WEBP_DSP_INIT_FUNC(VP8EncDspInit) {

  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
  if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
    if (VP8GetCPUInfo(kSSE2)) {
      VP8EncDspInitSSE2();
-#if defined(WEBP_USE_SSE41)
+#if defined(WEBP_HAVE_SSE41)
      if (VP8GetCPUInfo(kSSE4_1)) {
        VP8EncDspInitSSE41();
      }
@@ -800,7 +801,7 @@ WEBP_DSP_INIT_FUNC(VP8EncDspInit) {
 #endif
  }

-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
  if (WEBP_NEON_OMIT_C_CODE ||
      (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
    VP8EncDspInitNEON();
@@ -9,7 +9,7 @@
 //
 // ARM NEON version of speed-critical encoding functions.
 //
-// adapted from libvpx (http://www.webmproject.org/code/)
+// adapted from libvpx (https://www.webmproject.org/code/)

 #include "src/dsp/dsp.h"

@@ -764,9 +764,14 @@ static WEBP_INLINE void AccumulateSSE16_NEON(const uint8_t* const a,

 // Horizontal sum of all four uint32_t values in 'sum'.
 static int SumToInt_NEON(uint32x4_t sum) {
+#if WEBP_AARCH64
+  return (int)vaddvq_u32(sum);
+#else
  const uint64x2_t sum2 = vpaddlq_u32(sum);
-  const uint64_t sum3 = vgetq_lane_u64(sum2, 0) + vgetq_lane_u64(sum2, 1);
-  return (int)sum3;
+  const uint32x2_t sum3 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(sum2)),
+                                   vreinterpret_u32_u64(vget_high_u64(sum2)));
+  return (int)vget_lane_u32(sum3, 0);
+#endif
 }

 static int SSE16x16_NEON(const uint8_t* a, const uint8_t* b) {
@@ -860,7 +865,7 @@ static int QuantizeBlock_NEON(int16_t in[16], int16_t out[16],
  uint8x8x4_t shuffles;
  // vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use
  // non-standard versions there.
-#if defined(__APPLE__) && defined(__aarch64__) && \
+#if defined(__APPLE__) && WEBP_AARCH64 && \
    defined(__apple_build_version__) && (__apple_build_version__< 6020037)
  uint8x16x2_t all_out;
  INIT_VECTOR2(all_out, vreinterpretq_u8_s16(out0), vreinterpretq_u8_s16(out1));
@@ -25,9 +25,160 @@
 //------------------------------------------------------------------------------
 // Transforms (Paragraph 14.4)

-// Does one or two inverse transforms.
-static void ITransform_SSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
-                            int do_two) {
+// Does one inverse transform.
+static void ITransform_One_SSE2(const uint8_t* ref, const int16_t* in,
+                                uint8_t* dst) {
+  // This implementation makes use of 16-bit fixed point versions of two
+  // multiply constants:
+  //    K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
+  //    K2 = sqrt(2) * sin (pi/8) ~= 35468 / 2^16
+  //
+  // To be able to use signed 16-bit integers, we use the following trick to
+  // have constants within range:
+  // - Associated constants are obtained by subtracting the 16-bit fixed point
+  //   version of one:
+  //      k = K - (1 << 16)  =>  K = k + (1 << 16)
+  //      K1 = 85267  =>  k1 =  20091
+  //      K2 = 35468  =>  k2 = -30068
+  // - The multiplication of a variable by a constant become the sum of the
+  //   variable and the multiplication of that variable by the associated
+  //   constant:
+  //      (x * K) >> 16 = (x * (k + (1 << 16))) >> 16 = ((x * k ) >> 16) + x
+  const __m128i k1k2 = _mm_set_epi16(-30068, -30068, -30068, -30068,
+                                     20091, 20091, 20091, 20091);
+  const __m128i k2k1 = _mm_set_epi16(20091, 20091, 20091, 20091,
+                                     -30068, -30068, -30068, -30068);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i zero_four = _mm_set_epi16(0, 0, 0, 0, 4, 4, 4, 4);
+  __m128i T01, T23;
+
+  // Load and concatenate the transform coefficients.
+  const __m128i in01 = _mm_loadu_si128((const __m128i*)&in[0]);
+  const __m128i in23 = _mm_loadu_si128((const __m128i*)&in[8]);
+  // a00 a10 a20 a30   a01 a11 a21 a31
+  // a02 a12 a22 a32   a03 a13 a23 a33
+
+  // Vertical pass and subsequent transpose.
+  {
+    const __m128i in1 = _mm_unpackhi_epi64(in01, in01);
+    const __m128i in3 = _mm_unpackhi_epi64(in23, in23);
+
+    // First pass, c and d calculations are longer because of the "trick"
+    // multiplications.
+    // c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3
+    // d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3
+    const __m128i a_d3 = _mm_add_epi16(in01, in23);
+    const __m128i b_c3 = _mm_sub_epi16(in01, in23);
+    const __m128i c1d1 = _mm_mulhi_epi16(in1, k2k1);
+    const __m128i c2d2 = _mm_mulhi_epi16(in3, k1k2);
+    const __m128i c3 = _mm_unpackhi_epi64(b_c3, b_c3);
+    const __m128i c4 = _mm_sub_epi16(c1d1, c2d2);
+    const __m128i c = _mm_add_epi16(c3, c4);
+    const __m128i d4u = _mm_add_epi16(c1d1, c2d2);
+    const __m128i du = _mm_add_epi16(a_d3, d4u);
+    const __m128i d = _mm_unpackhi_epi64(du, du);
+
+    // Second pass.
+    const __m128i comb_ab = _mm_unpacklo_epi64(a_d3, b_c3);
+    const __m128i comb_dc = _mm_unpacklo_epi64(d, c);
+
+    const __m128i tmp01 = _mm_add_epi16(comb_ab, comb_dc);
+    const __m128i tmp32 = _mm_sub_epi16(comb_ab, comb_dc);
+    const __m128i tmp23 = _mm_shuffle_epi32(tmp32, _MM_SHUFFLE(1, 0, 3, 2));
+
+    const __m128i transpose_0 = _mm_unpacklo_epi16(tmp01, tmp23);
+    const __m128i transpose_1 = _mm_unpackhi_epi16(tmp01, tmp23);
+    // a00 a20 a01 a21   a02 a22 a03 a23
+    // a10 a30 a11 a31   a12 a32 a13 a33
+
+    T01 = _mm_unpacklo_epi16(transpose_0, transpose_1);
+    T23 = _mm_unpackhi_epi16(transpose_0, transpose_1);
+    // a00 a10 a20 a30   a01 a11 a21 a31
+    // a02 a12 a22 a32   a03 a13 a23 a33
+  }
+
+  // Horizontal pass and subsequent transpose.
+  {
+    const __m128i T1 = _mm_unpackhi_epi64(T01, T01);
+    const __m128i T3 = _mm_unpackhi_epi64(T23, T23);
+
+    // First pass, c and d calculations are longer because of the "trick"
+    // multiplications.
+    const __m128i dc = _mm_add_epi16(T01, zero_four);
+
+    // c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3
+    // d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3
+    const __m128i a_d3 = _mm_add_epi16(dc, T23);
+    const __m128i b_c3 = _mm_sub_epi16(dc, T23);
+    const __m128i c1d1 = _mm_mulhi_epi16(T1, k2k1);
+    const __m128i c2d2 = _mm_mulhi_epi16(T3, k1k2);
+    const __m128i c3 = _mm_unpackhi_epi64(b_c3, b_c3);
+    const __m128i c4 = _mm_sub_epi16(c1d1, c2d2);
+    const __m128i c = _mm_add_epi16(c3, c4);
+    const __m128i d4u = _mm_add_epi16(c1d1, c2d2);
+    const __m128i du = _mm_add_epi16(a_d3, d4u);
+    const __m128i d = _mm_unpackhi_epi64(du, du);
+
+    // Second pass.
+    const __m128i comb_ab = _mm_unpacklo_epi64(a_d3, b_c3);
+    const __m128i comb_dc = _mm_unpacklo_epi64(d, c);
+
+    const __m128i tmp01 = _mm_add_epi16(comb_ab, comb_dc);
+    const __m128i tmp32 = _mm_sub_epi16(comb_ab, comb_dc);
+    const __m128i tmp23 = _mm_shuffle_epi32(tmp32, _MM_SHUFFLE(1, 0, 3, 2));
+
+    const __m128i shifted01 = _mm_srai_epi16(tmp01, 3);
+    const __m128i shifted23 = _mm_srai_epi16(tmp23, 3);
+    // a00 a01 a02 a03   a10 a11 a12 a13
+    // a20 a21 a22 a23   a30 a31 a32 a33
+
+    const __m128i transpose_0 = _mm_unpacklo_epi16(shifted01, shifted23);
+    const __m128i transpose_1 = _mm_unpackhi_epi16(shifted01, shifted23);
+    // a00 a20 a01 a21   a02 a22 a03 a23
+    // a10 a30 a11 a31   a12 a32 a13 a33
+
+    T01 = _mm_unpacklo_epi16(transpose_0, transpose_1);
+    T23 = _mm_unpackhi_epi16(transpose_0, transpose_1);
+    // a00 a10 a20 a30   a01 a11 a21 a31
+    // a02 a12 a22 a32   a03 a13 a23 a33
+  }
+
+  // Add inverse transform to 'ref' and store.
+  {
+    // Load the reference(s).
+    __m128i ref01, ref23, ref0123;
+    int32_t buf[4];
+
+    // Load four bytes/pixels per line.
+    const __m128i ref0 = _mm_cvtsi32_si128(WebPMemToInt32(&ref[0 * BPS]));
+    const __m128i ref1 = _mm_cvtsi32_si128(WebPMemToInt32(&ref[1 * BPS]));
+    const __m128i ref2 = _mm_cvtsi32_si128(WebPMemToInt32(&ref[2 * BPS]));
+    const __m128i ref3 = _mm_cvtsi32_si128(WebPMemToInt32(&ref[3 * BPS]));
+    ref01 = _mm_unpacklo_epi32(ref0, ref1);
+    ref23 = _mm_unpacklo_epi32(ref2, ref3);
+
+    // Convert to 16b.
+    ref01 = _mm_unpacklo_epi8(ref01, zero);
+    ref23 = _mm_unpacklo_epi8(ref23, zero);
+    // Add the inverse transform(s).
+    ref01 = _mm_add_epi16(ref01, T01);
+    ref23 = _mm_add_epi16(ref23, T23);
+    // Unsigned saturate to 8b.
+    ref0123 = _mm_packus_epi16(ref01, ref23);
+
+    _mm_storeu_si128((__m128i *)buf, ref0123);
+
+    // Store four bytes/pixels per line.
+    WebPInt32ToMem(&dst[0 * BPS], buf[0]);
+    WebPInt32ToMem(&dst[1 * BPS], buf[1]);
+    WebPInt32ToMem(&dst[2 * BPS], buf[2]);
+    WebPInt32ToMem(&dst[3 * BPS], buf[3]);
+  }
+}
+
+// Does two inverse transforms.
+static void ITransform_Two_SSE2(const uint8_t* ref, const int16_t* in,
+                                uint8_t* dst) {
  // This implementation makes use of 16-bit fixed point versions of two
  // multiply constants:
  //    K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
@@ -49,33 +200,21 @@ static void ITransform_SSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
  __m128i T0, T1, T2, T3;

  // Load and concatenate the transform coefficients (we'll do two inverse
-  // transforms in parallel). In the case of only one inverse transform, the
-  // second half of the vectors will just contain random value we'll never
-  // use nor store.
+  // transforms in parallel).
  __m128i in0, in1, in2, in3;
  {
-    in0 = _mm_loadl_epi64((const __m128i*)&in[0]);
-    in1 = _mm_loadl_epi64((const __m128i*)&in[4]);
-    in2 = _mm_loadl_epi64((const __m128i*)&in[8]);
-    in3 = _mm_loadl_epi64((const __m128i*)&in[12]);
-    // a00 a10 a20 a30   x x x x
-    // a01 a11 a21 a31   x x x x
-    // a02 a12 a22 a32   x x x x
-    // a03 a13 a23 a33   x x x x
-    if (do_two) {
-      const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]);
-      const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]);
-      const __m128i inB2 = _mm_loadl_epi64((const __m128i*)&in[24]);
-      const __m128i inB3 = _mm_loadl_epi64((const __m128i*)&in[28]);
-      in0 = _mm_unpacklo_epi64(in0, inB0);
-      in1 = _mm_unpacklo_epi64(in1, inB1);
-      in2 = _mm_unpacklo_epi64(in2, inB2);
-      in3 = _mm_unpacklo_epi64(in3, inB3);
-      // a00 a10 a20 a30   b00 b10 b20 b30
-      // a01 a11 a21 a31   b01 b11 b21 b31
-      // a02 a12 a22 a32   b02 b12 b22 b32
-      // a03 a13 a23 a33   b03 b13 b23 b33
-    }
+    const __m128i tmp0 = _mm_loadu_si128((const __m128i*)&in[0]);
+    const __m128i tmp1 = _mm_loadu_si128((const __m128i*)&in[8]);
+    const __m128i tmp2 = _mm_loadu_si128((const __m128i*)&in[16]);
+    const __m128i tmp3 = _mm_loadu_si128((const __m128i*)&in[24]);
+    in0 = _mm_unpacklo_epi64(tmp0, tmp2);
+    in1 = _mm_unpackhi_epi64(tmp0, tmp2);
+    in2 = _mm_unpacklo_epi64(tmp1, tmp3);
+    in3 = _mm_unpackhi_epi64(tmp1, tmp3);
+    // a00 a10 a20 a30   b00 b10 b20 b30
+    // a01 a11 a21 a31   b01 b11 b21 b31
+    // a02 a12 a22 a32   b02 b12 b22 b32
+    // a03 a13 a23 a33   b03 b13 b23 b33
  }

  // Vertical pass and subsequent transpose.
@@ -148,19 +287,11 @@ static void ITransform_SSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
    const __m128i zero = _mm_setzero_si128();
    // Load the reference(s).
    __m128i ref0, ref1, ref2, ref3;
-    if (do_two) {
-      // Load eight bytes/pixels per line.
-      ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]);
-      ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]);
-      ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]);
-      ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]);
-    } else {
-      // Load four bytes/pixels per line.
-      ref0 = _mm_cvtsi32_si128(WebPMemToUint32(&ref[0 * BPS]));
-      ref1 = _mm_cvtsi32_si128(WebPMemToUint32(&ref[1 * BPS]));
-      ref2 = _mm_cvtsi32_si128(WebPMemToUint32(&ref[2 * BPS]));
-      ref3 = _mm_cvtsi32_si128(WebPMemToUint32(&ref[3 * BPS]));
-    }
+    // Load eight bytes/pixels per line.
+    ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]);
+    ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]);
+    ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]);
+    ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]);
    // Convert to 16b.
    ref0 = _mm_unpacklo_epi8(ref0, zero);
    ref1 = _mm_unpacklo_epi8(ref1, zero);
@@ -176,20 +307,21 @@ static void ITransform_SSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
    ref1 = _mm_packus_epi16(ref1, ref1);
    ref2 = _mm_packus_epi16(ref2, ref2);
    ref3 = _mm_packus_epi16(ref3, ref3);
-    // Store the results.
-    if (do_two) {
-      // Store eight bytes/pixels per line.
-      _mm_storel_epi64((__m128i*)&dst[0 * BPS], ref0);
-      _mm_storel_epi64((__m128i*)&dst[1 * BPS], ref1);
-      _mm_storel_epi64((__m128i*)&dst[2 * BPS], ref2);
-      _mm_storel_epi64((__m128i*)&dst[3 * BPS], ref3);
-    } else {
-      // Store four bytes/pixels per line.
-      WebPUint32ToMem(&dst[0 * BPS], _mm_cvtsi128_si32(ref0));
-      WebPUint32ToMem(&dst[1 * BPS], _mm_cvtsi128_si32(ref1));
-      WebPUint32ToMem(&dst[2 * BPS], _mm_cvtsi128_si32(ref2));
-      WebPUint32ToMem(&dst[3 * BPS], _mm_cvtsi128_si32(ref3));
-    }
+    // Store eight bytes/pixels per line.
+    _mm_storel_epi64((__m128i*)&dst[0 * BPS], ref0);
+    _mm_storel_epi64((__m128i*)&dst[1 * BPS], ref1);
+    _mm_storel_epi64((__m128i*)&dst[2 * BPS], ref2);
+    _mm_storel_epi64((__m128i*)&dst[3 * BPS], ref3);
+  }
+}
+
+// Does one or two inverse transforms.
+static void ITransform_SSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
+                            int do_two) {
+  if (do_two) {
+    ITransform_Two_SSE2(ref, in, dst);
+  } else {
+    ITransform_One_SSE2(ref, in, dst);
  }
 }

@@ -481,7 +613,7 @@ static void CollectHistogram_SSE2(const uint8_t* ref, const uint8_t* pred,
 // helper for chroma-DC predictions
 static WEBP_INLINE void Put8x8uv_SSE2(uint8_t v, uint8_t* dst) {
  int j;
-  const __m128i values = _mm_set1_epi8(v);
+  const __m128i values = _mm_set1_epi8((char)v);
  for (j = 0; j < 8; ++j) {
    _mm_storel_epi64((__m128i*)(dst + j * BPS), values);
  }
@@ -489,7 +621,7 @@ static WEBP_INLINE void Put8x8uv_SSE2(uint8_t v, uint8_t* dst) {

 static WEBP_INLINE void Put16_SSE2(uint8_t v, uint8_t* dst) {
  int j;
-  const __m128i values = _mm_set1_epi8(v);
+  const __m128i values = _mm_set1_epi8((char)v);
  for (j = 0; j < 16; ++j) {
    _mm_store_si128((__m128i*)(dst + j * BPS), values);
  }
@@ -540,7 +672,7 @@ static WEBP_INLINE void VerticalPred_SSE2(uint8_t* dst,
 static WEBP_INLINE void HE8uv_SSE2(uint8_t* dst, const uint8_t* left) {
  int j;
  for (j = 0; j < 8; ++j) {
-    const __m128i values = _mm_set1_epi8(left[j]);
+    const __m128i values = _mm_set1_epi8((char)left[j]);
    _mm_storel_epi64((__m128i*)dst, values);
    dst += BPS;
  }
@@ -549,7 +681,7 @@ static WEBP_INLINE void HE8uv_SSE2(uint8_t* dst, const uint8_t* left) {
 static WEBP_INLINE void HE16_SSE2(uint8_t* dst, const uint8_t* left) {
  int j;
  for (j = 0; j < 16; ++j) {
-    const __m128i values = _mm_set1_epi8(left[j]);
+    const __m128i values = _mm_set1_epi8((char)left[j]);
    _mm_store_si128((__m128i*)dst, values);
    dst += BPS;
  }
@@ -722,10 +854,10 @@ static WEBP_INLINE void VE4_SSE2(uint8_t* dst,
  const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGH00), one);
  const __m128i b = _mm_subs_epu8(a, lsb);
  const __m128i avg = _mm_avg_epu8(b, BCDEFGH0);
-  const uint32_t vals = _mm_cvtsi128_si32(avg);
+  const int vals = _mm_cvtsi128_si32(avg);
  int i;
  for (i = 0; i < 4; ++i) {
-    WebPUint32ToMem(dst + i * BPS, vals);
+    WebPInt32ToMem(dst + i * BPS, vals);
  }
 }

@@ -760,10 +892,10 @@ static WEBP_INLINE void LD4_SSE2(uint8_t* dst,
  const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGHH0), one);
  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
  const __m128i abcdefg = _mm_avg_epu8(avg2, BCDEFGH0);
-  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcdefg    ));
-  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
-  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
-  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
+  WebPInt32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcdefg    ));
+  WebPInt32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
+  WebPInt32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
+  WebPInt32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
 }

 static WEBP_INLINE void VR4_SSE2(uint8_t* dst,
@@ -782,10 +914,10 @@ static WEBP_INLINE void VR4_SSE2(uint8_t* dst,
  const __m128i lsb = _mm_and_si128(_mm_xor_si128(IXABCD, ABCD0), one);
  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
  const __m128i efgh = _mm_avg_epu8(avg2, XABCD);
-  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcd    ));
-  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               efgh    ));
-  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1)));
-  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1)));
+  WebPInt32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcd    ));
+  WebPInt32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               efgh    ));
+  WebPInt32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1)));
+  WebPInt32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1)));

  // these two are hard to implement in SSE2, so we keep the C-version:
  DST(0, 2) = AVG3(J, I, X);
@@ -807,11 +939,12 @@ static WEBP_INLINE void VL4_SSE2(uint8_t* dst,
  const __m128i abbc = _mm_or_si128(ab, bc);
  const __m128i lsb2 = _mm_and_si128(abbc, lsb1);
  const __m128i avg4 = _mm_subs_epu8(avg3, lsb2);
-  const uint32_t extra_out = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 4));
-  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               avg1    ));
-  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               avg4    ));
-  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1)));
-  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1)));
+  const uint32_t extra_out =
+      (uint32_t)_mm_cvtsi128_si32(_mm_srli_si128(avg4, 4));
+  WebPInt32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               avg1    ));
+  WebPInt32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               avg4    ));
+  WebPInt32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1)));
+  WebPInt32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1)));

  // these two are hard to get and irregular
  DST(3, 2) = (extra_out >> 0) & 0xff;
@@ -829,10 +962,10 @@ static WEBP_INLINE void RD4_SSE2(uint8_t* dst,
  const __m128i lsb = _mm_and_si128(_mm_xor_si128(JIXABCD__, LKJIXABCD), one);
  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
  const __m128i abcdefg = _mm_avg_epu8(avg2, KJIXABCD_);
-  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(               abcdefg    ));
-  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
-  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
-  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
+  WebPInt32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(               abcdefg    ));
+  WebPInt32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
+  WebPInt32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
+  WebPInt32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
 }

 static WEBP_INLINE void HU4_SSE2(uint8_t* dst, const uint8_t* top) {
@@ -875,14 +1008,14 @@ static WEBP_INLINE void HD4_SSE2(uint8_t* dst, const uint8_t* top) {

 static WEBP_INLINE void TM4_SSE2(uint8_t* dst, const uint8_t* top) {
  const __m128i zero = _mm_setzero_si128();
-  const __m128i top_values = _mm_cvtsi32_si128(WebPMemToUint32(top));
+  const __m128i top_values = _mm_cvtsi32_si128(WebPMemToInt32(top));
  const __m128i top_base = _mm_unpacklo_epi8(top_values, zero);
  int y;
  for (y = 0; y < 4; ++y, dst += BPS) {
    const int val = top[-2 - y] - top[-1];
    const __m128i base = _mm_set1_epi16(val);
    const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero);
-    WebPUint32ToMem(dst, _mm_cvtsi128_si32(out));
+    WebPInt32ToMem(dst, _mm_cvtsi128_si32(out));
  }
 }

@@ -189,6 +189,12 @@ static void GradientFilter_C(const uint8_t* data, int width, int height,

 //------------------------------------------------------------------------------

+static void NoneUnfilter_C(const uint8_t* prev, const uint8_t* in,
+                           uint8_t* out, int width) {
+  (void)prev;
+  if (out != in) memcpy(out, in, width * sizeof(*out));
+}
+
 static void HorizontalUnfilter_C(const uint8_t* prev, const uint8_t* in,
                                 uint8_t* out, int width) {
  uint8_t pred = (prev == NULL) ? 0 : prev[0];
@@ -233,13 +239,14 @@ static void GradientUnfilter_C(const uint8_t* prev, const uint8_t* in,
 WebPFilterFunc WebPFilters[WEBP_FILTER_LAST];
 WebPUnfilterFunc WebPUnfilters[WEBP_FILTER_LAST];

+extern VP8CPUInfo VP8GetCPUInfo;
 extern void VP8FiltersInitMIPSdspR2(void);
 extern void VP8FiltersInitMSA(void);
 extern void VP8FiltersInitNEON(void);
 extern void VP8FiltersInitSSE2(void);

 WEBP_DSP_INIT_FUNC(VP8FiltersInit) {
-  WebPUnfilters[WEBP_FILTER_NONE] = NULL;
+  WebPUnfilters[WEBP_FILTER_NONE] = NoneUnfilter_C;
 #if !WEBP_NEON_OMIT_C_CODE
  WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter_C;
  WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter_C;
@@ -254,7 +261,7 @@ WEBP_DSP_INIT_FUNC(VP8FiltersInit) {
 #endif

  if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
    if (VP8GetCPUInfo(kSSE2)) {
      VP8FiltersInitSSE2();
    }
@@ -271,13 +278,14 @@ WEBP_DSP_INIT_FUNC(VP8FiltersInit) {
 #endif
  }

-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
  if (WEBP_NEON_OMIT_C_CODE ||
      (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
    VP8FiltersInitNEON();
  }
 #endif

+  assert(WebPUnfilters[WEBP_FILTER_NONE] != NULL);
  assert(WebPUnfilters[WEBP_FILTER_HORIZONTAL] != NULL);
  assert(WebPUnfilters[WEBP_FILTER_VERTICAL] != NULL);
  assert(WebPUnfilters[WEBP_FILTER_GRADIENT] != NULL);
@@ -320,7 +320,12 @@ extern void VP8FiltersInitSSE2(void);

 WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitSSE2(void) {
  WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter_SSE2;
+#if defined(CHROMIUM)
+  // TODO(crbug.com/654974)
+  (void)VerticalUnfilter_SSE2;
+#else
  WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter_SSE2;
+#endif
  WebPUnfilters[WEBP_FILTER_GRADIENT] = GradientUnfilter_SSE2;

  WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter_SSE2;
@@ -49,7 +49,7 @@ static WEBP_INLINE uint32_t Clip255(uint32_t a) {
 }

 static WEBP_INLINE int AddSubtractComponentFull(int a, int b, int c) {
-  return Clip255(a + b - c);
+  return Clip255((uint32_t)(a + b - c));
 }

 static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1,
@@ -66,7 +66,7 @@ static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1,
 }

 static WEBP_INLINE int AddSubtractComponentHalf(int a, int b) {
-  return Clip255(a + (a - b) / 2);
+  return Clip255((uint32_t)(a + (a - b) / 2));
 }

 static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1,
@@ -107,63 +107,77 @@ static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) {
 //------------------------------------------------------------------------------
 // Predictors

-uint32_t VP8LPredictor0_C(uint32_t left, const uint32_t* const top) {
+uint32_t VP8LPredictor0_C(const uint32_t* const left,
+                          const uint32_t* const top) {
  (void)top;
  (void)left;
  return ARGB_BLACK;
 }
-uint32_t VP8LPredictor1_C(uint32_t left, const uint32_t* const top) {
+uint32_t VP8LPredictor1_C(const uint32_t* const left,
+                          const uint32_t* const top) {
  (void)top;
-  return left;
+  return *left;
 }
-uint32_t VP8LPredictor2_C(uint32_t left, const uint32_t* const top) {
+uint32_t VP8LPredictor2_C(const uint32_t* const left,
+                          const uint32_t* const top) {
  (void)left;
  return top[0];
 }
-uint32_t VP8LPredictor3_C(uint32_t left, const uint32_t* const top) {
+uint32_t VP8LPredictor3_C(const uint32_t* const left,
+                          const uint32_t* const top) {
  (void)left;
  return top[1];
 }
-uint32_t VP8LPredictor4_C(uint32_t left, const uint32_t* const top) {
+uint32_t VP8LPredictor4_C(const uint32_t* const left,
+                          const uint32_t* const top) {
  (void)left;
  return top[-1];
 }
-uint32_t VP8LPredictor5_C(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = Average3(left, top[0], top[1]);
+uint32_t VP8LPredictor5_C(const uint32_t* const left,
+                          const uint32_t* const top) {
+  const uint32_t pred = Average3(*left, top[0], top[1]);
  return pred;
 }
-uint32_t VP8LPredictor6_C(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = Average2(left, top[-1]);
+uint32_t VP8LPredictor6_C(const uint32_t* const left,
+                          const uint32_t* const top) {
+  const uint32_t pred = Average2(*left, top[-1]);
  return pred;
 }
-uint32_t VP8LPredictor7_C(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = Average2(left, top[0]);
+uint32_t VP8LPredictor7_C(const uint32_t* const left,
+                          const uint32_t* const top) {
+  const uint32_t pred = Average2(*left, top[0]);
  return pred;
 }
-uint32_t VP8LPredictor8_C(uint32_t left, const uint32_t* const top) {
+uint32_t VP8LPredictor8_C(const uint32_t* const left,
+                          const uint32_t* const top) {
  const uint32_t pred = Average2(top[-1], top[0]);
  (void)left;
  return pred;
 }
-uint32_t VP8LPredictor9_C(uint32_t left, const uint32_t* const top) {
+uint32_t VP8LPredictor9_C(const uint32_t* const left,
+                          const uint32_t* const top) {
  const uint32_t pred = Average2(top[0], top[1]);
  (void)left;
  return pred;
 }
-uint32_t VP8LPredictor10_C(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = Average4(left, top[-1], top[0], top[1]);
+uint32_t VP8LPredictor10_C(const uint32_t* const left,
+                           const uint32_t* const top) {
+  const uint32_t pred = Average4(*left, top[-1], top[0], top[1]);
  return pred;
 }
-uint32_t VP8LPredictor11_C(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = Select(top[0], left, top[-1]);
+uint32_t VP8LPredictor11_C(const uint32_t* const left,
+                           const uint32_t* const top) {
+  const uint32_t pred = Select(top[0], *left, top[-1]);
  return pred;
 }
-uint32_t VP8LPredictor12_C(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = ClampedAddSubtractFull(left, top[0], top[-1]);
+uint32_t VP8LPredictor12_C(const uint32_t* const left,
+                           const uint32_t* const top) {
+  const uint32_t pred = ClampedAddSubtractFull(*left, top[0], top[-1]);
  return pred;
 }
-uint32_t VP8LPredictor13_C(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = ClampedAddSubtractHalf(left, top[0], top[-1]);
+uint32_t VP8LPredictor13_C(const uint32_t* const left,
+                           const uint32_t* const top) {
+  const uint32_t pred = ClampedAddSubtractHalf(*left, top[0], top[-1]);
  return pred;
 }

@@ -279,10 +293,10 @@ void VP8LTransformColorInverse_C(const VP8LMultipliers* const m,
    const uint32_t red = argb >> 16;
    int new_red = red & 0xff;
    int new_blue = argb & 0xff;
-    new_red += ColorTransformDelta(m->green_to_red_, green);
+    new_red += ColorTransformDelta((int8_t)m->green_to_red_, green);
    new_red &= 0xff;
-    new_blue += ColorTransformDelta(m->green_to_blue_, green);
-    new_blue += ColorTransformDelta(m->red_to_blue_, (int8_t)new_red);
+    new_blue += ColorTransformDelta((int8_t)m->green_to_blue_, green);
+    new_blue += ColorTransformDelta((int8_t)m->red_to_blue_, (int8_t)new_red);
    new_blue &= 0xff;
    dst[i] = (argb & 0xff00ff00u) | (new_red << 16) | (new_blue);
  }
@@ -381,7 +395,7 @@ void VP8LInverseTransform(const VP8LTransform* const transform,
  assert(row_start < row_end);
  assert(row_end <= transform->ysize_);
  switch (transform->type_) {
-    case SUBTRACT_GREEN:
+    case SUBTRACT_GREEN_TRANSFORM:
      VP8LAddGreenToBlueAndRed(in, (row_end - row_start) * width, out);
      break;
    case PREDICTOR_TRANSFORM:
@@ -574,7 +588,9 @@ VP8LConvertFunc VP8LConvertBGRAToBGR;
 VP8LMapARGBFunc VP8LMapColor32b;
 VP8LMapAlphaFunc VP8LMapColor8b;

+extern VP8CPUInfo VP8GetCPUInfo;
 extern void VP8LDspInitSSE2(void);
+extern void VP8LDspInitSSE41(void);
 extern void VP8LDspInitNEON(void);
 extern void VP8LDspInitMIPSdspR2(void);
 extern void VP8LDspInitMSA(void);
@@ -621,9 +637,14 @@ WEBP_DSP_INIT_FUNC(VP8LDspInit) {

  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
  if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
    if (VP8GetCPUInfo(kSSE2)) {
      VP8LDspInitSSE2();
+#if defined(WEBP_HAVE_SSE41)
+      if (VP8GetCPUInfo(kSSE4_1)) {
+        VP8LDspInitSSE41();
+      }
+#endif
    }
 #endif
 #if defined(WEBP_USE_MIPS_DSP_R2)
@@ -638,7 +659,7 @@ WEBP_DSP_INIT_FUNC(VP8LDspInit) {
 #endif
  }

-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
  if (WEBP_NEON_OMIT_C_CODE ||
      (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
    VP8LDspInitNEON();
@@ -28,23 +28,38 @@ extern "C" {
 //------------------------------------------------------------------------------
 // Decoding

-typedef uint32_t (*VP8LPredictorFunc)(uint32_t left, const uint32_t* const top);
+typedef uint32_t (*VP8LPredictorFunc)(const uint32_t* const left,
+                                      const uint32_t* const top);
 extern VP8LPredictorFunc VP8LPredictors[16];

-uint32_t VP8LPredictor0_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor1_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor2_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor3_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor4_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor5_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor6_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor7_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor8_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor9_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor10_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor11_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor12_C(uint32_t left, const uint32_t* const top);
-uint32_t VP8LPredictor13_C(uint32_t left, const uint32_t* const top);
+uint32_t VP8LPredictor0_C(const uint32_t* const left,
+                          const uint32_t* const top);
+uint32_t VP8LPredictor1_C(const uint32_t* const left,
+                          const uint32_t* const top);
+uint32_t VP8LPredictor2_C(const uint32_t* const left,
+                          const uint32_t* const top);
+uint32_t VP8LPredictor3_C(const uint32_t* const left,
+                          const uint32_t* const top);
+uint32_t VP8LPredictor4_C(const uint32_t* const left,
+                          const uint32_t* const top);
+uint32_t VP8LPredictor5_C(const uint32_t* const left,
+                          const uint32_t* const top);
+uint32_t VP8LPredictor6_C(const uint32_t* const left,
+                          const uint32_t* const top);
+uint32_t VP8LPredictor7_C(const uint32_t* const left,
+                          const uint32_t* const top);
+uint32_t VP8LPredictor8_C(const uint32_t* const left,
+                          const uint32_t* const top);
+uint32_t VP8LPredictor9_C(const uint32_t* const left,
+                          const uint32_t* const top);
+uint32_t VP8LPredictor10_C(const uint32_t* const left,
+                           const uint32_t* const top);
+uint32_t VP8LPredictor11_C(const uint32_t* const left,
+                           const uint32_t* const top);
+uint32_t VP8LPredictor12_C(const uint32_t* const left,
+                           const uint32_t* const top);
+uint32_t VP8LPredictor13_C(const uint32_t* const left,
+                           const uint32_t* const top);

 // These Add/Sub function expects upper[-1] and out[-1] to be readable.
 typedef void (*VP8LPredictorAddSubFunc)(const uint32_t* in,
@@ -167,9 +182,9 @@ extern VP8LPredictorAddSubFunc VP8LPredictorsSub_C[16];
 // -----------------------------------------------------------------------------
 // Huffman-cost related functions.

-typedef double (*VP8LCostFunc)(const uint32_t* population, int length);
-typedef double (*VP8LCostCombinedFunc)(const uint32_t* X, const uint32_t* Y,
-                                       int length);
+typedef uint32_t (*VP8LCostFunc)(const uint32_t* population, int length);
+typedef uint32_t (*VP8LCostCombinedFunc)(const uint32_t* X, const uint32_t* Y,
+                                         int length);
 typedef float (*VP8LCombinedShannonEntropyFunc)(const int X[256],
                                                const int Y[256]);

@@ -183,7 +198,7 @@ typedef struct {        // small struct to hold counters
 } VP8LStreaks;

 typedef struct {            // small struct to hold bit entropy results
-  double entropy;           // entropy
+  float entropy;            // entropy
  uint32_t sum;             // sum of the population
  int nonzeros;             // number of non-zero elements in the population
  uint32_t max_val;         // maximum value in the population
@@ -16,9 +16,9 @@
 #ifndef WEBP_DSP_LOSSLESS_COMMON_H_
 #define WEBP_DSP_LOSSLESS_COMMON_H_

-#include "src/webp/types.h"
-
+#include "src/dsp/cpu.h"
 #include "src/utils/utils.h"
+#include "src/webp/types.h"

 #ifdef __cplusplus
 extern "C" {
@@ -166,7 +166,7 @@ uint32_t VP8LSubPixels(uint32_t a, uint32_t b) {
 }

 //------------------------------------------------------------------------------
-// Transform-related functions use din both encoding and decoding.
+// Transform-related functions used in both encoding and decoding.

 // Macros used to create a batch predictor that iteratively uses a
 // one-pixel predictor.
@@ -179,7 +179,7 @@ static void PREDICTOR_ADD(const uint32_t* in, const uint32_t* upper, \
  int x;                                                             \
  assert(upper != NULL);                                             \
  for (x = 0; x < num_pixels; ++x) {                                 \
-    const uint32_t pred = (PREDICTOR)(out[x - 1], upper + x);        \
+    const uint32_t pred = (PREDICTOR)(&out[x - 1], upper + x);       \
    out[x] = VP8LAddPixels(in[x], pred);                             \
  }                                                                  \
 }
@@ -329,6 +329,15 @@ const uint8_t kPrefixEncodeExtraBitsValue[PREFIX_LOOKUP_IDX_MAX] = {
 static float FastSLog2Slow_C(uint32_t v) {
  assert(v >= LOG_LOOKUP_IDX_MAX);
  if (v < APPROX_LOG_WITH_CORRECTION_MAX) {
+#if !defined(WEBP_HAVE_SLOW_CLZ_CTZ)
+    // use clz if available
+    const int log_cnt = BitsLog2Floor(v) - 7;
+    const uint32_t y = 1 << log_cnt;
+    int correction = 0;
+    const float v_f = (float)v;
+    const uint32_t orig_v = v;
+    v >>= log_cnt;
+#else
    int log_cnt = 0;
    uint32_t y = 1;
    int correction = 0;
@@ -339,6 +348,7 @@ static float FastSLog2Slow_C(uint32_t v) {
      v = v >> 1;
      y = y << 1;
    } while (v >= LOG_LOOKUP_IDX_MAX);
+#endif
    // vf = (2^log_cnt) * Xf; where y = 2^log_cnt and Xf < 256
    // Xf = floor(Xf) * (1 + (v % y) / v)
    // log2(Xf) = log2(floor(Xf)) + log2(1 + (v % y) / v)
@@ -355,6 +365,14 @@ static float FastSLog2Slow_C(uint32_t v) {
 static float FastLog2Slow_C(uint32_t v) {
  assert(v >= LOG_LOOKUP_IDX_MAX);
  if (v < APPROX_LOG_WITH_CORRECTION_MAX) {
+#if !defined(WEBP_HAVE_SLOW_CLZ_CTZ)
+    // use clz if available
+    const int log_cnt = BitsLog2Floor(v) - 7;
+    const uint32_t y = 1 << log_cnt;
+    const uint32_t orig_v = v;
+    double log_2;
+    v >>= log_cnt;
+#else
    int log_cnt = 0;
    uint32_t y = 1;
    const uint32_t orig_v = v;
@@ -364,6 +382,7 @@ static float FastLog2Slow_C(uint32_t v) {
      v = v >> 1;
      y = y << 1;
    } while (v >= LOG_LOOKUP_IDX_MAX);
+#endif
    log_2 = kLog2Table[v] + log_cnt;
    if (orig_v >= APPROX_LOG_MAX) {
      // Since the division is still expensive, add this correction factor only
@@ -383,7 +402,7 @@ static float FastLog2Slow_C(uint32_t v) {
 // Compute the combined Shanon's entropy for distribution {X} and {X+Y}
 static float CombinedShannonEntropy_C(const int X[256], const int Y[256]) {
  int i;
-  double retval = 0.;
+  float retval = 0.f;
  int sumX = 0, sumXY = 0;
  for (i = 0; i < 256; ++i) {
    const int x = X[i];
@@ -399,7 +418,7 @@ static float CombinedShannonEntropy_C(const int X[256], const int Y[256]) {
    }
  }
  retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY);
-  return (float)retval;
+  return retval;
 }

 void VP8LBitEntropyInit(VP8LBitEntropy* const entropy) {
@@ -503,11 +522,11 @@ static void GetCombinedEntropyUnrefined_C(const uint32_t X[],
 void VP8LSubtractGreenFromBlueAndRed_C(uint32_t* argb_data, int num_pixels) {
  int i;
  for (i = 0; i < num_pixels; ++i) {
-    const int argb = argb_data[i];
+    const int argb = (int)argb_data[i];
    const int green = (argb >> 8) & 0xff;
    const uint32_t new_r = (((argb >> 16) & 0xff) - green) & 0xff;
    const uint32_t new_b = (((argb >>  0) & 0xff) - green) & 0xff;
-    argb_data[i] = (argb & 0xff00ff00u) | (new_r << 16) | new_b;
+    argb_data[i] = ((uint32_t)argb & 0xff00ff00u) | (new_r << 16) | new_b;
  }
 }

@@ -528,10 +547,10 @@ void VP8LTransformColor_C(const VP8LMultipliers* const m, uint32_t* data,
    const int8_t red   = U32ToS8(argb >> 16);
    int new_red = red & 0xff;
    int new_blue = argb & 0xff;
-    new_red -= ColorTransformDelta(m->green_to_red_, green);
+    new_red -= ColorTransformDelta((int8_t)m->green_to_red_, green);
    new_red &= 0xff;
-    new_blue -= ColorTransformDelta(m->green_to_blue_, green);
-    new_blue -= ColorTransformDelta(m->red_to_blue_, red);
+    new_blue -= ColorTransformDelta((int8_t)m->green_to_blue_, green);
+    new_blue -= ColorTransformDelta((int8_t)m->red_to_blue_, red);
    new_blue &= 0xff;
    data[i] = (argb & 0xff00ff00u) | (new_red << 16) | (new_blue);
  }
@@ -541,7 +560,7 @@ static WEBP_INLINE uint8_t TransformColorRed(uint8_t green_to_red,
                                             uint32_t argb) {
  const int8_t green = U32ToS8(argb >> 8);
  int new_red = argb >> 16;
-  new_red -= ColorTransformDelta(green_to_red, green);
+  new_red -= ColorTransformDelta((int8_t)green_to_red, green);
  return (new_red & 0xff);
 }

@@ -550,9 +569,9 @@ static WEBP_INLINE uint8_t TransformColorBlue(uint8_t green_to_blue,
                                              uint32_t argb) {
  const int8_t green = U32ToS8(argb >>  8);
  const int8_t red   = U32ToS8(argb >> 16);
-  uint8_t new_blue = argb & 0xff;
-  new_blue -= ColorTransformDelta(green_to_blue, green);
-  new_blue -= ColorTransformDelta(red_to_blue, red);
+  int new_blue = argb & 0xff;
+  new_blue -= ColorTransformDelta((int8_t)green_to_blue, green);
+  new_blue -= ColorTransformDelta((int8_t)red_to_blue, red);
  return (new_blue & 0xff);
 }

@@ -617,20 +636,25 @@ void VP8LBundleColorMap_C(const uint8_t* const row, int width, int xbits,

 //------------------------------------------------------------------------------

-static double ExtraCost_C(const uint32_t* population, int length) {
+static uint32_t ExtraCost_C(const uint32_t* population, int length) {
  int i;
-  double cost = 0.;
-  for (i = 2; i < length - 2; ++i) cost += (i >> 1) * population[i + 2];
+  uint32_t cost = population[4] + population[5];
+  assert(length % 2 == 0);
+  for (i = 2; i < length / 2 - 1; ++i) {
+    cost += i * (population[2 * i + 2] + population[2 * i + 3]);
+  }
  return cost;
 }

-static double ExtraCostCombined_C(const uint32_t* X, const uint32_t* Y,
-                                  int length) {
+static uint32_t ExtraCostCombined_C(const uint32_t* X, const uint32_t* Y,
+                                    int length) {
  int i;
-  double cost = 0.;
-  for (i = 2; i < length - 2; ++i) {
-    const int xy = X[i + 2] + Y[i + 2];
-    cost += (i >> 1) * xy;
+  uint32_t cost = X[4] + Y[4] + X[5] + Y[5];
+  assert(length % 2 == 0);
+  for (i = 2; i < length / 2 - 1; ++i) {
+    const int xy0 = X[2 * i + 2] + Y[2 * i + 2];
+    const int xy1 = X[2 * i + 3] + Y[2 * i + 3];
+    cost += i * (xy0 + xy1);
  }
  return cost;
 }
@@ -726,7 +750,7 @@ static void PredictorSub##PREDICTOR_I##_C(const uint32_t* in,              \
  assert(upper != NULL);                                                   \
  for (x = 0; x < num_pixels; ++x) {                                       \
    const uint32_t pred =                                                  \
-        VP8LPredictor##PREDICTOR_I##_C(in[x - 1], upper + x);              \
+        VP8LPredictor##PREDICTOR_I##_C(&in[x - 1], upper + x);             \
    out[x] = VP8LSubPixels(in[x], pred);                                   \
  }                                                                        \
 }
@@ -772,6 +796,7 @@ VP8LBundleColorMapFunc VP8LBundleColorMap;
 VP8LPredictorAddSubFunc VP8LPredictorsSub[16];
 VP8LPredictorAddSubFunc VP8LPredictorsSub_C[16];

+extern VP8CPUInfo VP8GetCPUInfo;
 extern void VP8LEncDspInitSSE2(void);
 extern void VP8LEncDspInitSSE41(void);
 extern void VP8LEncDspInitNEON(void);
@@ -843,10 +868,10 @@ WEBP_DSP_INIT_FUNC(VP8LEncDspInit) {

  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
  if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
    if (VP8GetCPUInfo(kSSE2)) {
      VP8LEncDspInitSSE2();
-#if defined(WEBP_USE_SSE41)
+#if defined(WEBP_HAVE_SSE41)
      if (VP8GetCPUInfo(kSSE4_1)) {
        VP8LEncDspInitSSE41();
      }
@@ -870,7 +895,7 @@ WEBP_DSP_INIT_FUNC(VP8LEncDspInit) {
 #endif
  }

-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
  if (WEBP_NEON_OMIT_C_CODE ||
      (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
    VP8LEncDspInitNEON();
@@ -103,8 +103,8 @@ static float FastLog2Slow_MIPS32(uint32_t v) {
 //     cost += i * *(pop + 1);
 //     pop += 2;
 //   }
-//   return (double)cost;
-static double ExtraCost_MIPS32(const uint32_t* const population, int length) {
+//   return cost;
+static uint32_t ExtraCost_MIPS32(const uint32_t* const population, int length) {
  int i, temp0, temp1;
  const uint32_t* pop = &population[4];
  const uint32_t* const LoopEnd = &population[length];
@@ -130,7 +130,7 @@ static double ExtraCost_MIPS32(const uint32_t* const population, int length) {
    : "memory", "hi", "lo"
  );

-  return (double)((int64_t)temp0 << 32 | temp1);
+  return ((int64_t)temp0 << 32 | temp1);
 }

 // C version of this function:
@@ -148,9 +148,9 @@ static double ExtraCost_MIPS32(const uint32_t* const population, int length) {
 //     pX += 2;
 //     pY += 2;
 //   }
-//   return (double)cost;
-static double ExtraCostCombined_MIPS32(const uint32_t* const X,
-                                       const uint32_t* const Y, int length) {
+//   return cost;
+static uint32_t ExtraCostCombined_MIPS32(const uint32_t* const X,
+                                         const uint32_t* const Y, int length) {
  int i, temp0, temp1, temp2, temp3;
  const uint32_t* pX = &X[4];
  const uint32_t* pY = &Y[4];
@@ -183,7 +183,7 @@ static double ExtraCostCombined_MIPS32(const uint32_t* const X,
    : "memory", "hi", "lo"
  );

-  return (double)((int64_t)temp0 << 32 | temp1);
+  return ((int64_t)temp0 << 32 | temp1);
 }

 #define HUFFMAN_COST_PASS                                 \
@@ -347,24 +347,24 @@ static void GetCombinedEntropyUnrefined_MIPS32(const uint32_t X[],
 static void AddVector_MIPS32(const uint32_t* pa, const uint32_t* pb,
                             uint32_t* pout, int size) {
  uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
-  const uint32_t end = ((size) / 4) * 4;
+  const int end = ((size) / 4) * 4;
  const uint32_t* const LoopEnd = pa + end;
  int i;
  ASM_START
  ADD_TO_OUT(0, 4, 8, 12, 1, pa, pb, pout)
  ASM_END_0
-  for (i = end; i < size; ++i) pout[i] = pa[i] + pb[i];
+  for (i = 0; i < size - end; ++i) pout[i] = pa[i] + pb[i];
 }

 static void AddVectorEq_MIPS32(const uint32_t* pa, uint32_t* pout, int size) {
  uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
-  const uint32_t end = ((size) / 4) * 4;
+  const int end = ((size) / 4) * 4;
  const uint32_t* const LoopEnd = pa + end;
  int i;
  ASM_START
  ADD_TO_OUT(0, 4, 8, 12, 0, pa, pout, pout)
  ASM_END_1
-  for (i = end; i < size; ++i) pout[i] += pa[i];
+  for (i = 0; i < size - end; ++i) pout[i] += pa[i];
 }

 #undef ASM_END_1
@@ -25,7 +25,7 @@

 // vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use
 // non-standard versions there.
-#if defined(__APPLE__) && defined(__aarch64__) && \
+#if defined(__APPLE__) && WEBP_AARCH64 && \
    defined(__apple_build_version__) && (__apple_build_version__< 6020037)
 #define USE_VTBLQ
 #endif
@@ -54,8 +54,8 @@ static void TransformColor_SSE2(const VP8LMultipliers* const m,
  const __m128i mults_rb = MK_CST_16(CST_5b(m->green_to_red_),
                                     CST_5b(m->green_to_blue_));
  const __m128i mults_b2 = MK_CST_16(CST_5b(m->red_to_blue_), 0);
-  const __m128i mask_ag = _mm_set1_epi32(0xff00ff00);  // alpha-green masks
-  const __m128i mask_rb = _mm_set1_epi32(0x00ff00ff);  // red-blue masks
+  const __m128i mask_ag = _mm_set1_epi32((int)0xff00ff00);  // alpha-green masks
+  const __m128i mask_rb = _mm_set1_epi32(0x00ff00ff);       // red-blue masks
  int i;
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb
@@ -232,79 +232,55 @@ static void AddVectorEq_SSE2(const uint32_t* a, uint32_t* out, int size) {
 //------------------------------------------------------------------------------
 // Entropy

-// Checks whether the X or Y contribution is worth computing and adding.
-// Used in loop unrolling.
-#define ANALYZE_X_OR_Y(x_or_y, j)                                           \
-  do {                                                                      \
-    if ((x_or_y)[i + (j)] != 0) retval -= VP8LFastSLog2((x_or_y)[i + (j)]); \
-  } while (0)
+// TODO(https://crbug.com/webp/499): this function produces different results
+// from the C code due to use of double/float resulting in output differences
+// when compared to -noasm.
+#if !(defined(WEBP_HAVE_SLOW_CLZ_CTZ) || defined(__i386__) || defined(_M_IX86))

-// Checks whether the X + Y contribution is worth computing and adding.
-// Used in loop unrolling.
-#define ANALYZE_XY(j)                  \
-  do {                                 \
-    if (tmp[j] != 0) {                 \
-      retval -= VP8LFastSLog2(tmp[j]); \
-      ANALYZE_X_OR_Y(X, j);            \
-    }                                  \
-  } while (0)
-
-#if !(defined(__i386__) || defined(_M_IX86))
 static float CombinedShannonEntropy_SSE2(const int X[256], const int Y[256]) {
  int i;
-  double retval = 0.;
-  int sumX, sumXY;
-  int32_t tmp[4];
-  __m128i zero = _mm_setzero_si128();
-  // Sums up X + Y, 4 ints at a time (and will merge it at the end for sumXY).
-  __m128i sumXY_128 = zero;
-  __m128i sumX_128 = zero;
+  float retval = 0.f;
+  int sumX = 0, sumXY = 0;
+  const __m128i zero = _mm_setzero_si128();

-  for (i = 0; i < 256; i += 4) {
-    const __m128i x = _mm_loadu_si128((const __m128i*)(X + i));
-    const __m128i y = _mm_loadu_si128((const __m128i*)(Y + i));
-
-    // Check if any X is non-zero: this actually provides a speedup as X is
-    // usually sparse.
-    if (_mm_movemask_epi8(_mm_cmpeq_epi32(x, zero)) != 0xFFFF) {
-      const __m128i xy_128 = _mm_add_epi32(x, y);
-      sumXY_128 = _mm_add_epi32(sumXY_128, xy_128);
-
-      sumX_128 = _mm_add_epi32(sumX_128, x);
-
-      // Analyze the different X + Y.
-      _mm_storeu_si128((__m128i*)tmp, xy_128);
-
-      ANALYZE_XY(0);
-      ANALYZE_XY(1);
-      ANALYZE_XY(2);
-      ANALYZE_XY(3);
-    } else {
-      // X is fully 0, so only deal with Y.
-      sumXY_128 = _mm_add_epi32(sumXY_128, y);
-
-      ANALYZE_X_OR_Y(Y, 0);
-      ANALYZE_X_OR_Y(Y, 1);
-      ANALYZE_X_OR_Y(Y, 2);
-      ANALYZE_X_OR_Y(Y, 3);
+  for (i = 0; i < 256; i += 16) {
+    const __m128i x0 = _mm_loadu_si128((const __m128i*)(X + i +  0));
+    const __m128i y0 = _mm_loadu_si128((const __m128i*)(Y + i +  0));
+    const __m128i x1 = _mm_loadu_si128((const __m128i*)(X + i +  4));
+    const __m128i y1 = _mm_loadu_si128((const __m128i*)(Y + i +  4));
+    const __m128i x2 = _mm_loadu_si128((const __m128i*)(X + i +  8));
+    const __m128i y2 = _mm_loadu_si128((const __m128i*)(Y + i +  8));
+    const __m128i x3 = _mm_loadu_si128((const __m128i*)(X + i + 12));
+    const __m128i y3 = _mm_loadu_si128((const __m128i*)(Y + i + 12));
+    const __m128i x4 = _mm_packs_epi16(_mm_packs_epi32(x0, x1),
+                                       _mm_packs_epi32(x2, x3));
+    const __m128i y4 = _mm_packs_epi16(_mm_packs_epi32(y0, y1),
+                                       _mm_packs_epi32(y2, y3));
+    const int32_t mx = _mm_movemask_epi8(_mm_cmpgt_epi8(x4, zero));
+    int32_t my = _mm_movemask_epi8(_mm_cmpgt_epi8(y4, zero)) | mx;
+    while (my) {
+      const int32_t j = BitsCtz(my);
+      int xy;
+      if ((mx >> j) & 1) {
+        const int x = X[i + j];
+        sumXY += x;
+        retval -= VP8LFastSLog2(x);
+      }
+      xy = X[i + j] + Y[i + j];
+      sumX += xy;
+      retval -= VP8LFastSLog2(xy);
+      my &= my - 1;
    }
  }
-
-  // Sum up sumX_128 to get sumX.
-  _mm_storeu_si128((__m128i*)tmp, sumX_128);
-  sumX = tmp[3] + tmp[2] + tmp[1] + tmp[0];
-
-  // Sum up sumXY_128 to get sumXY.
-  _mm_storeu_si128((__m128i*)tmp, sumXY_128);
-  sumXY = tmp[3] + tmp[2] + tmp[1] + tmp[0];
-
  retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY);
-  return (float)retval;
+  return retval;
 }
-#endif  // !(defined(__i386__) || defined(_M_IX86))

-#undef ANALYZE_X_OR_Y
-#undef ANALYZE_XY
+#else
+
+#define DONT_USE_COMBINED_SHANNON_ENTROPY_SSE2_FUNC   // won't be faster
+
+#endif

 //------------------------------------------------------------------------------

@@ -400,7 +376,7 @@ static void BundleColorMap_SSE2(const uint8_t* const row, int width, int xbits,
      break;
    }
    case 2: {
-      const __m128i mask_or = _mm_set1_epi32(0xff000000);
+      const __m128i mask_or = _mm_set1_epi32((int)0xff000000);
      const __m128i mul_cst = _mm_set1_epi16(0x0104);
      const __m128i mask_mul = _mm_set1_epi16(0x0f00);
      for (x = 0; x + 16 <= width; x += 16, dst += 4) {
@@ -451,7 +427,7 @@ static WEBP_INLINE void Average2_m128i(const __m128i* const a0,
 static void PredictorSub0_SSE2(const uint32_t* in, const uint32_t* upper,
                               int num_pixels, uint32_t* out) {
  int i;
-  const __m128i black = _mm_set1_epi32(ARGB_BLACK);
+  const __m128i black = _mm_set1_epi32((int)ARGB_BLACK);
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
    const __m128i res = _mm_sub_epi8(src, black);
@@ -662,10 +638,7 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) {
  VP8LCollectColorRedTransforms = CollectColorRedTransforms_SSE2;
  VP8LAddVector = AddVector_SSE2;
  VP8LAddVectorEq = AddVectorEq_SSE2;
-  // TODO(https://crbug.com/webp/499): this function produces different results
-  // from the C code due to use of double/float resulting in output differences
-  // when compared to -noasm.
-#if !(defined(__i386__) || defined(_M_IX86))
+#if !defined(DONT_USE_COMBINED_SHANNON_ENTROPY_SSE2_FUNC)
  VP8LCombinedShannonEntropy = CombinedShannonEntropy_SSE2;
 #endif
  VP8LVectorMismatch = VectorMismatch_SSE2;
@@ -18,8 +18,53 @@
 #include <smmintrin.h>
 #include "src/dsp/lossless.h"

-// For sign-extended multiplying constants, pre-shifted by 5:
-#define CST_5b(X)  (((int16_t)((uint16_t)(X) << 8)) >> 5)
+//------------------------------------------------------------------------------
+// Cost operations.
+
+static WEBP_INLINE uint32_t HorizontalSum_SSE41(__m128i cost) {
+  cost = _mm_add_epi32(cost, _mm_srli_si128(cost, 8));
+  cost = _mm_add_epi32(cost, _mm_srli_si128(cost, 4));
+  return _mm_cvtsi128_si32(cost);
+}
+
+static uint32_t ExtraCost_SSE41(const uint32_t* const a, int length) {
+  int i;
+  __m128i cost = _mm_set_epi32(2 * a[7], 2 * a[6], a[5], a[4]);
+  assert(length % 8 == 0);
+
+  for (i = 8; i + 8 <= length; i += 8) {
+    const int j = (i - 2) >> 1;
+    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i]);
+    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
+    const __m128i w = _mm_set_epi32(j + 3, j + 2, j + 1, j);
+    const __m128i a2 = _mm_hadd_epi32(a0, a1);
+    const __m128i mul = _mm_mullo_epi32(a2, w);
+    cost = _mm_add_epi32(mul, cost);
+  }
+  return HorizontalSum_SSE41(cost);
+}
+
+static uint32_t ExtraCostCombined_SSE41(const uint32_t* const a,
+                                        const uint32_t* const b, int length) {
+  int i;
+  __m128i cost = _mm_add_epi32(_mm_set_epi32(2 * a[7], 2 * a[6], a[5], a[4]),
+                               _mm_set_epi32(2 * b[7], 2 * b[6], b[5], b[4]));
+  assert(length % 8 == 0);
+
+  for (i = 8; i + 8 <= length; i += 8) {
+    const int j = (i - 2) >> 1;
+    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i]);
+    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
+    const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i]);
+    const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i + 4]);
+    const __m128i w = _mm_set_epi32(j + 3, j + 2, j + 1, j);
+    const __m128i a2 = _mm_hadd_epi32(a0, a1);
+    const __m128i b2 = _mm_hadd_epi32(b0, b1);
+    const __m128i mul = _mm_mullo_epi32(_mm_add_epi32(a2, b2), w);
+    cost = _mm_add_epi32(mul, cost);
+  }
+  return HorizontalSum_SSE41(cost);
+}

 //------------------------------------------------------------------------------
 // Subtract-Green Transform
@@ -44,46 +89,50 @@ static void SubtractGreenFromBlueAndRed_SSE41(uint32_t* argb_data,
 //------------------------------------------------------------------------------
 // Color Transform

-#define SPAN 8
+// For sign-extended multiplying constants, pre-shifted by 5:
+#define CST_5b(X) (((int16_t)((uint16_t)(X) << 8)) >> 5)
+
+#define MK_CST_16(HI, LO) \
+  _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff)))
+
 static void CollectColorBlueTransforms_SSE41(const uint32_t* argb, int stride,
                                             int tile_width, int tile_height,
                                             int green_to_blue, int red_to_blue,
                                             int histo[]) {
-  const __m128i mults_r = _mm_set1_epi16(CST_5b(red_to_blue));
-  const __m128i mults_g = _mm_set1_epi16(CST_5b(green_to_blue));
-  const __m128i mask_g = _mm_set1_epi16((short)0xff00);   // green mask
-  const __m128i mask_gb = _mm_set1_epi32(0xffff);         // green/blue mask
-  const __m128i mask_b = _mm_set1_epi16(0x00ff);          // blue mask
-  const __m128i shuffler_lo = _mm_setr_epi8(-1, 2, -1, 6, -1, 10, -1, 14, -1,
-                                            -1, -1, -1, -1, -1, -1, -1);
-  const __m128i shuffler_hi = _mm_setr_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1,
-                                            2, -1, 6, -1, 10, -1, 14);
-  int y;
-  for (y = 0; y < tile_height; ++y) {
-    const uint32_t* const src = argb + y * stride;
-    int i, x;
-    for (x = 0; x + SPAN <= tile_width; x += SPAN) {
-      uint16_t values[SPAN];
-      const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]);
-      const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]);
-      const __m128i r0 = _mm_shuffle_epi8(in0, shuffler_lo);
-      const __m128i r1 = _mm_shuffle_epi8(in1, shuffler_hi);
-      const __m128i r = _mm_or_si128(r0, r1);         // r 0
-      const __m128i gb0 = _mm_and_si128(in0, mask_gb);
-      const __m128i gb1 = _mm_and_si128(in1, mask_gb);
-      const __m128i gb = _mm_packus_epi32(gb0, gb1);  // g b
-      const __m128i g = _mm_and_si128(gb, mask_g);    // g 0
-      const __m128i A = _mm_mulhi_epi16(r, mults_r);  // x dbr
-      const __m128i B = _mm_mulhi_epi16(g, mults_g);  // x dbg
-      const __m128i C = _mm_sub_epi8(gb, B);          // x b'
-      const __m128i D = _mm_sub_epi8(C, A);           // x b''
-      const __m128i E = _mm_and_si128(D, mask_b);     // 0 b''
-      _mm_storeu_si128((__m128i*)values, E);
-      for (i = 0; i < SPAN; ++i) ++histo[values[i]];
+  const __m128i mult =
+      MK_CST_16(CST_5b(red_to_blue) + 256,CST_5b(green_to_blue));
+  const __m128i perm =
+      _mm_setr_epi8(-1, 1, -1, 2, -1, 5, -1, 6, -1, 9, -1, 10, -1, 13, -1, 14);
+  if (tile_width >= 4) {
+    int y;
+    for (y = 0; y < tile_height; ++y) {
+      const uint32_t* const src = argb + y * stride;
+      const __m128i A1 = _mm_loadu_si128((const __m128i*)src);
+      const __m128i B1 = _mm_shuffle_epi8(A1, perm);
+      const __m128i C1 = _mm_mulhi_epi16(B1, mult);
+      const __m128i D1 = _mm_sub_epi16(A1, C1);
+      __m128i E = _mm_add_epi16(_mm_srli_epi32(D1, 16), D1);
+      int x;
+      for (x = 4; x + 4 <= tile_width; x += 4) {
+        const __m128i A2 = _mm_loadu_si128((const __m128i*)(src + x));
+        __m128i B2, C2, D2;
+        ++histo[_mm_extract_epi8(E,  0)];
+        B2 = _mm_shuffle_epi8(A2, perm);
+        ++histo[_mm_extract_epi8(E,  4)];
+        C2 = _mm_mulhi_epi16(B2, mult);
+        ++histo[_mm_extract_epi8(E,  8)];
+        D2 = _mm_sub_epi16(A2, C2);
+        ++histo[_mm_extract_epi8(E, 12)];
+        E = _mm_add_epi16(_mm_srli_epi32(D2, 16), D2);
+      }
+      ++histo[_mm_extract_epi8(E,  0)];
+      ++histo[_mm_extract_epi8(E,  4)];
+      ++histo[_mm_extract_epi8(E,  8)];
+      ++histo[_mm_extract_epi8(E, 12)];
    }
  }
  {
-    const int left_over = tile_width & (SPAN - 1);
+    const int left_over = tile_width & 3;
    if (left_over > 0) {
      VP8LCollectColorBlueTransforms_C(argb + tile_width - left_over, stride,
                                       left_over, tile_height,
@@ -95,33 +144,37 @@ static void CollectColorBlueTransforms_SSE41(const uint32_t* argb, int stride,
 static void CollectColorRedTransforms_SSE41(const uint32_t* argb, int stride,
                                            int tile_width, int tile_height,
                                            int green_to_red, int histo[]) {
-  const __m128i mults_g = _mm_set1_epi16(CST_5b(green_to_red));
-  const __m128i mask_g = _mm_set1_epi32(0x00ff00);  // green mask
-  const __m128i mask = _mm_set1_epi16(0xff);

-  int y;
-  for (y = 0; y < tile_height; ++y) {
-    const uint32_t* const src = argb + y * stride;
-    int i, x;
-    for (x = 0; x + SPAN <= tile_width; x += SPAN) {
-      uint16_t values[SPAN];
-      const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]);
-      const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]);
-      const __m128i g0 = _mm_and_si128(in0, mask_g);  // 0 0  | g 0
-      const __m128i g1 = _mm_and_si128(in1, mask_g);
-      const __m128i g = _mm_packus_epi32(g0, g1);     // g 0
-      const __m128i A0 = _mm_srli_epi32(in0, 16);     // 0 0  | x r
-      const __m128i A1 = _mm_srli_epi32(in1, 16);
-      const __m128i A = _mm_packus_epi32(A0, A1);     // x r
-      const __m128i B = _mm_mulhi_epi16(g, mults_g);  // x dr
-      const __m128i C = _mm_sub_epi8(A, B);           // x r'
-      const __m128i D = _mm_and_si128(C, mask);       // 0 r'
-      _mm_storeu_si128((__m128i*)values, D);
-      for (i = 0; i < SPAN; ++i) ++histo[values[i]];
+  const __m128i mult = MK_CST_16(0, CST_5b(green_to_red));
+  const __m128i mask_g = _mm_set1_epi32(0x0000ff00);
+  if (tile_width >= 4) {
+    int y;
+    for (y = 0; y < tile_height; ++y) {
+      const uint32_t* const src = argb + y * stride;
+      const __m128i A1 = _mm_loadu_si128((const __m128i*)src);
+      const __m128i B1 = _mm_and_si128(A1, mask_g);
+      const __m128i C1 = _mm_madd_epi16(B1, mult);
+      __m128i D = _mm_sub_epi16(A1, C1);
+      int x;
+      for (x = 4; x + 4 <= tile_width; x += 4) {
+        const __m128i A2 = _mm_loadu_si128((const __m128i*)(src + x));
+        __m128i B2, C2;
+        ++histo[_mm_extract_epi8(D,  2)];
+        B2 = _mm_and_si128(A2, mask_g);
+        ++histo[_mm_extract_epi8(D,  6)];
+        C2 = _mm_madd_epi16(B2, mult);
+        ++histo[_mm_extract_epi8(D, 10)];
+        ++histo[_mm_extract_epi8(D, 14)];
+        D = _mm_sub_epi16(A2, C2);
+      }
+      ++histo[_mm_extract_epi8(D,  2)];
+      ++histo[_mm_extract_epi8(D,  6)];
+      ++histo[_mm_extract_epi8(D, 10)];
+      ++histo[_mm_extract_epi8(D, 14)];
    }
  }
  {
-    const int left_over = tile_width & (SPAN - 1);
+    const int left_over = tile_width & 3;
    if (left_over > 0) {
      VP8LCollectColorRedTransforms_C(argb + tile_width - left_over, stride,
                                      left_over, tile_height, green_to_red,
@@ -130,12 +183,16 @@ static void CollectColorRedTransforms_SSE41(const uint32_t* argb, int stride,
  }
 }

+#undef MK_CST_16
+
 //------------------------------------------------------------------------------
 // Entry point

 extern void VP8LEncDspInitSSE41(void);

 WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE41(void) {
+  VP8LExtraCost = ExtraCost_SSE41;
+  VP8LExtraCostCombined = ExtraCostCombined_SSE41;
  VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_SSE41;
  VP8LCollectColorBlueTransforms = CollectColorBlueTransforms_SSE41;
  VP8LCollectColorRedTransforms = CollectColorRedTransforms_SSE41;
@@ -188,46 +188,51 @@ static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1,
  return Average2(Average2(a0, a1), Average2(a2, a3));
 }

-static uint32_t Predictor5_MIPSdspR2(uint32_t left, const uint32_t* const top) {
-  return Average3(left, top[0], top[1]);
+static uint32_t Predictor5_MIPSdspR2(const uint32_t* const left,
+                                     const uint32_t* const top) {
+  return Average3(*left, top[0], top[1]);
 }

-static uint32_t Predictor6_MIPSdspR2(uint32_t left, const uint32_t* const top) {
-  return Average2(left, top[-1]);
+static uint32_t Predictor6_MIPSdspR2(const uint32_t* const left,
+                                     const uint32_t* const top) {
+  return Average2(*left, top[-1]);
 }

-static uint32_t Predictor7_MIPSdspR2(uint32_t left, const uint32_t* const top) {
-  return Average2(left, top[0]);
+static uint32_t Predictor7_MIPSdspR2(const uint32_t* const left,
+                                     const uint32_t* const top) {
+  return Average2(*left, top[0]);
 }

-static uint32_t Predictor8_MIPSdspR2(uint32_t left, const uint32_t* const top) {
+static uint32_t Predictor8_MIPSdspR2(const uint32_t* const left,
+                                     const uint32_t* const top) {
  (void)left;
  return Average2(top[-1], top[0]);
 }

-static uint32_t Predictor9_MIPSdspR2(uint32_t left, const uint32_t* const top) {
+static uint32_t Predictor9_MIPSdspR2(const uint32_t* const left,
+                                     const uint32_t* const top) {
  (void)left;
  return Average2(top[0], top[1]);
 }

-static uint32_t Predictor10_MIPSdspR2(uint32_t left,
+static uint32_t Predictor10_MIPSdspR2(const uint32_t* const left,
                                      const uint32_t* const top) {
-  return Average4(left, top[-1], top[0], top[1]);
+  return Average4(*left, top[-1], top[0], top[1]);
 }

-static uint32_t Predictor11_MIPSdspR2(uint32_t left,
+static uint32_t Predictor11_MIPSdspR2(const uint32_t* const left,
                                      const uint32_t* const top) {
-  return Select(top[0], left, top[-1]);
+  return Select(top[0], *left, top[-1]);
 }

-static uint32_t Predictor12_MIPSdspR2(uint32_t left,
+static uint32_t Predictor12_MIPSdspR2(const uint32_t* const left,
                                      const uint32_t* const top) {
-  return ClampedAddSubtractFull(left, top[0], top[-1]);
+  return ClampedAddSubtractFull(*left, top[0], top[-1]);
 }

-static uint32_t Predictor13_MIPSdspR2(uint32_t left,
+static uint32_t Predictor13_MIPSdspR2(const uint32_t* const left,
                                      const uint32_t* const top) {
-  return ClampedAddSubtractHalf(left, top[0], top[-1]);
+  return ClampedAddSubtractHalf(*left, top[0], top[-1]);
 }

 // Add green to blue and red channels (i.e. perform the inverse transform of
@@ -188,17 +188,21 @@ static WEBP_INLINE uint32_t Average3_NEON(uint32_t a0, uint32_t a1,
  return avg;
 }

-static uint32_t Predictor5_NEON(uint32_t left, const uint32_t* const top) {
-  return Average3_NEON(left, top[0], top[1]);
+static uint32_t Predictor5_NEON(const uint32_t* const left,
+                                const uint32_t* const top) {
+  return Average3_NEON(*left, top[0], top[1]);
 }
-static uint32_t Predictor6_NEON(uint32_t left, const uint32_t* const top) {
-  return Average2_NEON(left, top[-1]);
+static uint32_t Predictor6_NEON(const uint32_t* const left,
+                                const uint32_t* const top) {
+  return Average2_NEON(*left, top[-1]);
 }
-static uint32_t Predictor7_NEON(uint32_t left, const uint32_t* const top) {
-  return Average2_NEON(left, top[0]);
+static uint32_t Predictor7_NEON(const uint32_t* const left,
+                                const uint32_t* const top) {
+  return Average2_NEON(*left, top[0]);
 }
-static uint32_t Predictor13_NEON(uint32_t left, const uint32_t* const top) {
-  return ClampedAddSubtractHalf_NEON(left, top[0], top[-1]);
+static uint32_t Predictor13_NEON(const uint32_t* const left,
+                                 const uint32_t* const top) {
+  return ClampedAddSubtractHalf_NEON(*left, top[0], top[-1]);
 }

 // Batch versions of those functions.
@@ -494,7 +498,7 @@ static void PredictorAdd13_NEON(const uint32_t* in, const uint32_t* upper,

 // vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use
 // non-standard versions there.
-#if defined(__APPLE__) && defined(__aarch64__) && \
+#if defined(__APPLE__) && WEBP_AARCH64 && \
    defined(__apple_build_version__) && (__apple_build_version__< 6020037)
 #define USE_VTBLQ
 #endif
@@ -18,7 +18,6 @@
 #include "src/dsp/common_sse2.h"
 #include "src/dsp/lossless.h"
 #include "src/dsp/lossless_common.h"
-#include <assert.h>
 #include <emmintrin.h>

 //------------------------------------------------------------------------------
@@ -28,23 +27,22 @@ static WEBP_INLINE uint32_t ClampedAddSubtractFull_SSE2(uint32_t c0,
                                                        uint32_t c1,
                                                        uint32_t c2) {
  const __m128i zero = _mm_setzero_si128();
-  const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero);
-  const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero);
-  const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero);
+  const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c0), zero);
+  const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c1), zero);
+  const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c2), zero);
  const __m128i V1 = _mm_add_epi16(C0, C1);
  const __m128i V2 = _mm_sub_epi16(V1, C2);
  const __m128i b = _mm_packus_epi16(V2, V2);
-  const uint32_t output = _mm_cvtsi128_si32(b);
-  return output;
+  return (uint32_t)_mm_cvtsi128_si32(b);
 }

 static WEBP_INLINE uint32_t ClampedAddSubtractHalf_SSE2(uint32_t c0,
                                                        uint32_t c1,
                                                        uint32_t c2) {
  const __m128i zero = _mm_setzero_si128();
-  const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero);
-  const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero);
-  const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero);
+  const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c0), zero);
+  const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c1), zero);
+  const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c2), zero);
  const __m128i avg = _mm_add_epi16(C1, C0);
  const __m128i A0 = _mm_srli_epi16(avg, 1);
  const __m128i A1 = _mm_sub_epi16(A0, B0);
@@ -53,16 +51,15 @@ static WEBP_INLINE uint32_t ClampedAddSubtractHalf_SSE2(uint32_t c0,
  const __m128i A3 = _mm_srai_epi16(A2, 1);
  const __m128i A4 = _mm_add_epi16(A0, A3);
  const __m128i A5 = _mm_packus_epi16(A4, A4);
-  const uint32_t output = _mm_cvtsi128_si32(A5);
-  return output;
+  return (uint32_t)_mm_cvtsi128_si32(A5);
 }

 static WEBP_INLINE uint32_t Select_SSE2(uint32_t a, uint32_t b, uint32_t c) {
  int pa_minus_pb;
  const __m128i zero = _mm_setzero_si128();
-  const __m128i A0 = _mm_cvtsi32_si128(a);
-  const __m128i B0 = _mm_cvtsi32_si128(b);
-  const __m128i C0 = _mm_cvtsi32_si128(c);
+  const __m128i A0 = _mm_cvtsi32_si128((int)a);
+  const __m128i B0 = _mm_cvtsi32_si128((int)b);
+  const __m128i C0 = _mm_cvtsi32_si128((int)c);
  const __m128i AC0 = _mm_subs_epu8(A0, C0);
  const __m128i CA0 = _mm_subs_epu8(C0, A0);
  const __m128i BC0 = _mm_subs_epu8(B0, C0);
@@ -95,8 +92,8 @@ static WEBP_INLINE void Average2_uint32_SSE2(const uint32_t a0,
                                             __m128i* const avg) {
  // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1)
  const __m128i ones = _mm_set1_epi8(1);
-  const __m128i A0 = _mm_cvtsi32_si128(a0);
-  const __m128i A1 = _mm_cvtsi32_si128(a1);
+  const __m128i A0 = _mm_cvtsi32_si128((int)a0);
+  const __m128i A1 = _mm_cvtsi32_si128((int)a1);
  const __m128i avg1 = _mm_avg_epu8(A0, A1);
  const __m128i one = _mm_and_si128(_mm_xor_si128(A0, A1), ones);
  *avg = _mm_sub_epi8(avg1, one);
@@ -104,8 +101,8 @@ static WEBP_INLINE void Average2_uint32_SSE2(const uint32_t a0,

 static WEBP_INLINE __m128i Average2_uint32_16_SSE2(uint32_t a0, uint32_t a1) {
  const __m128i zero = _mm_setzero_si128();
-  const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a0), zero);
-  const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero);
+  const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)a0), zero);
+  const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)a1), zero);
  const __m128i sum = _mm_add_epi16(A1, A0);
  return _mm_srli_epi16(sum, 1);
 }
@@ -113,19 +110,18 @@ static WEBP_INLINE __m128i Average2_uint32_16_SSE2(uint32_t a0, uint32_t a1) {
 static WEBP_INLINE uint32_t Average2_SSE2(uint32_t a0, uint32_t a1) {
  __m128i output;
  Average2_uint32_SSE2(a0, a1, &output);
-  return _mm_cvtsi128_si32(output);
+  return (uint32_t)_mm_cvtsi128_si32(output);
 }

 static WEBP_INLINE uint32_t Average3_SSE2(uint32_t a0, uint32_t a1,
                                          uint32_t a2) {
  const __m128i zero = _mm_setzero_si128();
  const __m128i avg1 = Average2_uint32_16_SSE2(a0, a2);
-  const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero);
+  const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)a1), zero);
  const __m128i sum = _mm_add_epi16(avg1, A1);
  const __m128i avg2 = _mm_srli_epi16(sum, 1);
  const __m128i A2 = _mm_packus_epi16(avg2, avg2);
-  const uint32_t output = _mm_cvtsi128_si32(A2);
-  return output;
+  return (uint32_t)_mm_cvtsi128_si32(A2);
 }

 static WEBP_INLINE uint32_t Average4_SSE2(uint32_t a0, uint32_t a1,
@@ -135,46 +131,54 @@ static WEBP_INLINE uint32_t Average4_SSE2(uint32_t a0, uint32_t a1,
  const __m128i sum = _mm_add_epi16(avg2, avg1);
  const __m128i avg3 = _mm_srli_epi16(sum, 1);
  const __m128i A0 = _mm_packus_epi16(avg3, avg3);
-  const uint32_t output = _mm_cvtsi128_si32(A0);
-  return output;
+  return (uint32_t)_mm_cvtsi128_si32(A0);
 }

-static uint32_t Predictor5_SSE2(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = Average3_SSE2(left, top[0], top[1]);
+static uint32_t Predictor5_SSE2(const uint32_t* const left,
+                                const uint32_t* const top) {
+  const uint32_t pred = Average3_SSE2(*left, top[0], top[1]);
  return pred;
 }
-static uint32_t Predictor6_SSE2(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = Average2_SSE2(left, top[-1]);
+static uint32_t Predictor6_SSE2(const uint32_t* const left,
+                                const uint32_t* const top) {
+  const uint32_t pred = Average2_SSE2(*left, top[-1]);
  return pred;
 }
-static uint32_t Predictor7_SSE2(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = Average2_SSE2(left, top[0]);
+static uint32_t Predictor7_SSE2(const uint32_t* const left,
+                                const uint32_t* const top) {
+  const uint32_t pred = Average2_SSE2(*left, top[0]);
  return pred;
 }
-static uint32_t Predictor8_SSE2(uint32_t left, const uint32_t* const top) {
+static uint32_t Predictor8_SSE2(const uint32_t* const left,
+                                const uint32_t* const top) {
  const uint32_t pred = Average2_SSE2(top[-1], top[0]);
  (void)left;
  return pred;
 }
-static uint32_t Predictor9_SSE2(uint32_t left, const uint32_t* const top) {
+static uint32_t Predictor9_SSE2(const uint32_t* const left,
+                                const uint32_t* const top) {
  const uint32_t pred = Average2_SSE2(top[0], top[1]);
  (void)left;
  return pred;
 }
-static uint32_t Predictor10_SSE2(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = Average4_SSE2(left, top[-1], top[0], top[1]);
+static uint32_t Predictor10_SSE2(const uint32_t* const left,
+                                 const uint32_t* const top) {
+  const uint32_t pred = Average4_SSE2(*left, top[-1], top[0], top[1]);
  return pred;
 }
-static uint32_t Predictor11_SSE2(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = Select_SSE2(top[0], left, top[-1]);
+static uint32_t Predictor11_SSE2(const uint32_t* const left,
+                                 const uint32_t* const top) {
+  const uint32_t pred = Select_SSE2(top[0], *left, top[-1]);
  return pred;
 }
-static uint32_t Predictor12_SSE2(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = ClampedAddSubtractFull_SSE2(left, top[0], top[-1]);
+static uint32_t Predictor12_SSE2(const uint32_t* const left,
+                                 const uint32_t* const top) {
+  const uint32_t pred = ClampedAddSubtractFull_SSE2(*left, top[0], top[-1]);
  return pred;
 }
-static uint32_t Predictor13_SSE2(uint32_t left, const uint32_t* const top) {
-  const uint32_t pred = ClampedAddSubtractHalf_SSE2(left, top[0], top[-1]);
+static uint32_t Predictor13_SSE2(const uint32_t* const left,
+                                 const uint32_t* const top) {
+  const uint32_t pred = ClampedAddSubtractHalf_SSE2(*left, top[0], top[-1]);
  return pred;
 }

@@ -184,7 +188,7 @@ static uint32_t Predictor13_SSE2(uint32_t left, const uint32_t* const top) {
 static void PredictorAdd0_SSE2(const uint32_t* in, const uint32_t* upper,
                               int num_pixels, uint32_t* out) {
  int i;
-  const __m128i black = _mm_set1_epi32(ARGB_BLACK);
+  const __m128i black = _mm_set1_epi32((int)ARGB_BLACK);
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
    const __m128i res = _mm_add_epi8(src, black);
@@ -200,7 +204,7 @@ static void PredictorAdd0_SSE2(const uint32_t* in, const uint32_t* upper,
 static void PredictorAdd1_SSE2(const uint32_t* in, const uint32_t* upper,
                               int num_pixels, uint32_t* out) {
  int i;
-  __m128i prev = _mm_set1_epi32(out[-1]);
+  __m128i prev = _mm_set1_epi32((int)out[-1]);
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    // a | b | c | d
    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
@@ -277,12 +281,12 @@ GENERATE_PREDICTOR_2(9, upper[i + 1])
 #undef GENERATE_PREDICTOR_2

 // Predictor10: average of (average of (L,TL), average of (T, TR)).
-#define DO_PRED10(OUT) do {               \
-  __m128i avgLTL, avg;                    \
-  Average2_m128i(&L, &TL, &avgLTL);       \
-  Average2_m128i(&avgTTR, &avgLTL, &avg); \
-  L = _mm_add_epi8(avg, src);             \
-  out[i + (OUT)] = _mm_cvtsi128_si32(L);  \
+#define DO_PRED10(OUT) do {                         \
+  __m128i avgLTL, avg;                              \
+  Average2_m128i(&L, &TL, &avgLTL);                 \
+  Average2_m128i(&avgTTR, &avgLTL, &avg);           \
+  L = _mm_add_epi8(avg, src);                       \
+  out[i + (OUT)] = (uint32_t)_mm_cvtsi128_si32(L);  \
 } while (0)

 #define DO_PRED10_SHIFT do {                                  \
@@ -295,7 +299,7 @@ GENERATE_PREDICTOR_2(9, upper[i + 1])
 static void PredictorAdd10_SSE2(const uint32_t* in, const uint32_t* upper,
                                int num_pixels, uint32_t* out) {
  int i;
-  __m128i L = _mm_cvtsi32_si128(out[-1]);
+  __m128i L = _mm_cvtsi32_si128((int)out[-1]);
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
    __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
@@ -328,7 +332,7 @@ static void PredictorAdd10_SSE2(const uint32_t* in, const uint32_t* upper,
  const __m128i B = _mm_andnot_si128(mask, T);                         \
  const __m128i pred = _mm_or_si128(A, B); /* pred = (pa > b)? L : T*/ \
  L = _mm_add_epi8(src, pred);                                         \
-  out[i + (OUT)] = _mm_cvtsi128_si32(L);                               \
+  out[i + (OUT)] = (uint32_t)_mm_cvtsi128_si32(L);                     \
 } while (0)

 #define DO_PRED11_SHIFT do {                                \
@@ -343,7 +347,7 @@ static void PredictorAdd11_SSE2(const uint32_t* in, const uint32_t* upper,
                                int num_pixels, uint32_t* out) {
  int i;
  __m128i pa;
-  __m128i L = _mm_cvtsi32_si128(out[-1]);
+  __m128i L = _mm_cvtsi32_si128((int)out[-1]);
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
    __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
@@ -376,12 +380,12 @@ static void PredictorAdd11_SSE2(const uint32_t* in, const uint32_t* upper,
 #undef DO_PRED11_SHIFT

 // Predictor12: ClampedAddSubtractFull.
-#define DO_PRED12(DIFF, LANE, OUT) do {            \
-  const __m128i all = _mm_add_epi16(L, (DIFF));    \
-  const __m128i alls = _mm_packus_epi16(all, all); \
-  const __m128i res = _mm_add_epi8(src, alls);     \
-  out[i + (OUT)] = _mm_cvtsi128_si32(res);         \
-  L = _mm_unpacklo_epi8(res, zero);                \
+#define DO_PRED12(DIFF, LANE, OUT) do {              \
+  const __m128i all = _mm_add_epi16(L, (DIFF));      \
+  const __m128i alls = _mm_packus_epi16(all, all);   \
+  const __m128i res = _mm_add_epi8(src, alls);       \
+  out[i + (OUT)] = (uint32_t)_mm_cvtsi128_si32(res); \
+  L = _mm_unpacklo_epi8(res, zero);                  \
 } while (0)

 #define DO_PRED12_SHIFT(DIFF, LANE) do {                    \
@@ -394,7 +398,7 @@ static void PredictorAdd12_SSE2(const uint32_t* in, const uint32_t* upper,
                                int num_pixels, uint32_t* out) {
  int i;
  const __m128i zero = _mm_setzero_si128();
-  const __m128i L8 = _mm_cvtsi32_si128(out[-1]);
+  const __m128i L8 = _mm_cvtsi32_si128((int)out[-1]);
  __m128i L = _mm_unpacklo_epi8(L8, zero);
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    // Load 4 pixels at a time.
@@ -460,7 +464,7 @@ static void TransformColorInverse_SSE2(const VP8LMultipliers* const m,
  const __m128i mults_b2 = MK_CST_16(CST(red_to_blue_), 0);
 #undef MK_CST_16
 #undef CST
-  const __m128i mask_ag = _mm_set1_epi32(0xff00ff00);  // alpha-green masks
+  const __m128i mask_ag = _mm_set1_epi32((int)0xff00ff00);  // alpha-green masks
  int i;
  for (i = 0; i + 4 <= num_pixels; i += 4) {
    const __m128i in = _mm_loadu_si128((const __m128i*)&src[i]); // argb
@@ -524,7 +528,7 @@ static void ConvertBGRAToRGB_SSE2(const uint32_t* src, int num_pixels,

 static void ConvertBGRAToRGBA_SSE2(const uint32_t* src,
                                   int num_pixels, uint8_t* dst) {
-  const __m128i red_blue_mask = _mm_set1_epi32(0x00ff00ffu);
+  const __m128i red_blue_mask = _mm_set1_epi32(0x00ff00ff);
  const __m128i* in = (const __m128i*)src;
  __m128i* out = (__m128i*)dst;
  while (num_pixels >= 8) {
@@ -553,7 +557,7 @@ static void ConvertBGRAToRGBA_SSE2(const uint32_t* src,
 static void ConvertBGRAToRGBA4444_SSE2(const uint32_t* src,
                                       int num_pixels, uint8_t* dst) {
  const __m128i mask_0x0f = _mm_set1_epi8(0x0f);
-  const __m128i mask_0xf0 = _mm_set1_epi8(0xf0);
+  const __m128i mask_0xf0 = _mm_set1_epi8((char)0xf0);
  const __m128i* in = (const __m128i*)src;
  __m128i* out = (__m128i*)dst;
  while (num_pixels >= 8) {
@@ -588,8 +592,8 @@ static void ConvertBGRAToRGBA4444_SSE2(const uint32_t* src,

 static void ConvertBGRAToRGB565_SSE2(const uint32_t* src,
                                     int num_pixels, uint8_t* dst) {
-  const __m128i mask_0xe0 = _mm_set1_epi8(0xe0);
-  const __m128i mask_0xf8 = _mm_set1_epi8(0xf8);
+  const __m128i mask_0xe0 = _mm_set1_epi8((char)0xe0);
+  const __m128i mask_0xf8 = _mm_set1_epi8((char)0xf8);
  const __m128i mask_0x07 = _mm_set1_epi8(0x07);
  const __m128i* in = (const __m128i*)src;
  __m128i* out = (__m128i*)dst;
--- a/Show More
+++ b/Show More
				`@@ -0,0 +1 @@`
				`Origin: https://github.com/google/flatbuffers/tree/v23.5.9`