SIMD Kernel speedup for DNN layers #28889
This PR add following speedups for Grounding Dino tiny model.
For Device: Intel(R) Core(TM) i9-14900KS, x86, 32 Cores, ubuntu 24.04,
| Model | `ENGINE_NEW (Before)` | `ENGINE_NEW (After)` | `ENGINE_ORT` |
| :--- | :--- | :--- | :--- |
| **Grounding Dino Tiny** | 3130 ms | 1872.06 ms| 1800.18 ms|
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
Previously, OpenCL in-place flip kernels (rows/cols/both) could produce
incorrect results compared to the CPU implementation on strict OpenCL
drivers such as Mesa. The kernels relied on implicit load–load–store–store
(LLSS) ordering when src and dst alias, which is not guaranteed by the
OpenCL memory model and may be reordered.
Some vendor drivers happened to preserve the expected ordering, masking
the issue, but Mesa correctly exposes the undefined behavior.
This change introduces dedicated in-place flip kernels that:
- Explicitly detect in-place execution (src == dst)
- Stage data through local memory tiles
- Enforce correct ordering with work-group barriers
- Avoid global memory read/write aliasing hazards
The non in-place path is unchanged.
With this fix, OpenCL in-place flip produces correct and consistent results
across drivers, matches CPU behavior, and complies with the OpenCL memory
model.
Signed-off-by: jiajia Qian <jiajia.qian@nxp.com>
core(opencl): fix inplace transpose race by enforcing LLSS ordering via local barrier #28686
The former inplace transpose implementation allowed a reordering of global-memory operations across work-items. Specifically, the intended LLSS (Load–Load–Store–Store) access pattern could be reordered by the GPU into LSLS (Load–Store–Load–Store), causing partially written tiles to be observed by other work-items and producing incorrect output.
This patch introduces a tiled LDS-based algorithm and adds an explicit:
barrier(CLK_LOCAL_MEM_FENCE);
between the load and store phases.
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [ ] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
Nms empty detection fix#28749
Requires opencv_extra: https://github.com/opencv/opencv_extra/pull/1330
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
core: add NEON implementation for rotate function #28609
- This PR adds a NEON intrinsics-based implementation for the rotate function in matrix_transform.cpp for Windows-ARM64.
- The optimized implementation uses ARM NEON intrinsics to accelerate the internal transpose step used by the rotate function.
- In the x64 architecture, the rotate operation benefits from IPP-based optimized implementations. However, on ARM64, the execution falls back to the scalar implementation, which results in lower performance.
- To achieve performance parity with x64, a NEON-based SIMD implementation has been added for ARM64.
- After introducing these changes, the rotate function showed noticeable performance improvements on ARM64 platforms.
<img width="1009" height="817" alt="image" src="https://github.com/user-attachments/assets/8bec0041-b19c-4fc8-9103-532746224515" />
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
dnn: fix dst_dp assertion in broadcast for size-1 dims causing crash in lightglue.onnx model #28692
The original assertion CV_Assert(dst_dp == 1) does not handle valid cases where the innermost dimension size is 1 like [10, 5, 1], resulting in dst_dp == 0.
This occurs during broadcasting in LightGlue ONNX model and leads to assertion failure.
Allow dst_dp == 0 for size-1 dimensions to handle this edge case correctly.
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
Optimized flip Horizontal #28614
- Refactor flipHoriz implementation.
- Optimizations for horizontal image flipping added.
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
New dnn engine #26056
This is the 1st PR with the new engine; CI is green and PR is ready to be merged, I think.
Merge together with https://github.com/opencv/opencv_contrib/pull/3794
---
**Known limitations:**
* [solved] OpenVINO is temporarily disabled, but is probably easy to restore (it's not a deal breaker to merge this PR, I guess)
* The new engine does not support any backends nor any targets except for the default CPU implementation. But it's possible to choose the old engine when loading a model, then all the functionality is available.
* [Caffe patch is here: #26208] The new engine only supports ONNX. When a model is constructed manually or is loaded from a file of different format (.tf, .tflite, .caffe, .darknet), the old engine is used.
* Even in the case of ONNX some layers are not supported by the new engine, such as all quantized layers (including DequantizeLinear, QuantizeLinear, QLinearConv etc.), LSTM, GRU, .... It's planned, of course, to have full support for ONNX by OpenCV 5.0 gold release. When a loaded model contains unsupported layers, we switch to the old engine automatically (at ONNX parsing time, not at `forward()` time).
* Some layers , e.g. Expat, are only partially supported by the new engine. In the case of unsupported flavours it switches to the old engine automatically (at ONNX parsing time, not at `forward()` time).
* 'Concat' graph optimization is disabled. The optimization eliminates Concat layer and instead makes the layers that generate tensors to be concatenated to write the outputs to the final destination. Of course, it's only possible when `axis=0` or `axis=N=1`. The optimization is not compatible with dynamic shapes since we need to know in advance where to store the tensors. Because some of the layer implementations have been modified to become more compatible with the new engine, the feature appears to be broken even when the old engine is used.
* Some `dnn::Net` API is not available with the new engine. Also, shape inference may return false if some of the output or intermediate tensors' shapes cannot be inferred without running the model. Probably this can be fixed by a dummy run of the model with zero inputs.
* Some overloads of `dnn::Net::getFLOPs()` and `dnn::Net::getMemoryConsumption()` are not exposed any longer in wrapper generators; but the most useful overloads are exposed (and checked by Java tests).
* [in progress] A few Einsum tests related to empty shapes have been disabled due to crashes in the tests and in Einsum implementations. The code and the tests need to be repaired.
* OpenCL implementation of Deconvolution is disabled. It's very bad and very slow anyway; need to be completely revised.
* Deconvolution3D test is now skipped, because it was only supported by CUDA and OpenVINO backends, both of which are not supported by the new engine.
* Some tests, such as FastNeuralStyle, checked that the in the case of CUDA backend there is no fallback to CPU. Currently all layers in the new engine are processed on CPU, so there are many fallbacks. The checks, therefore, have been temporarily disabled.
---
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
Add HAL implementation hooks to cv::flip() and cv::rotate() functions from core module #24233
Hello,
This change proposes the addition of HAL hooks for cv::flip() and cv::rotate() functions from OpenCV core module.
Flip and rotation are functions commonly available from 2D hardware accelerators. This is convenient provision to enable custom optimized implementation of image flip/rotation on systems embedding such accelerator.
Thank you
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
* add broadcast_to with tests
* change name
* fix test
* fix implicit type conversion
* replace type of shape with InputArray
* add perf test
* add perf tests which takes care of axis
* v2 from ficus expand
* rename to broadcast
* use randu in place of declare
* doc improvement; smaller scale in perf
* capture get_index by reference
Rewrite Universal Intrinsic code by using new API: Core module. #23980
The goal of this PR is to match and modify all SIMD code blocks guarded by `CV_SIMD` macro in the `opencv/modules/core` folder and rewrite them by using the new Universal Intrinsic API.
The patch is almost auto-generated by using the [rewriter](https://github.com/hanliutong/rewriter), related PR #23885.
Most of the files have been rewritten, but I marked this PR as draft because, the `CV_SIMD` macro also exists in the following files, and the reasons why they are not rewrited are:
1. ~~code design for fixed-size SIMD (v_int16x8, v_float32x4, etc.), need to manually rewrite.~~ Rewrited
- ./modules/core/src/stat.simd.hpp
- ./modules/core/src/matrix_transform.cpp
- ./modules/core/src/matmul.simd.hpp
2. Vector types are wrapped in other class/struct, that are not supported by the compiler in variable-length backends. Can not be rewrited directly.
- ./modules/core/src/mathfuncs_core.simd.hpp
```cpp
struct v_atan_f32
{
explicit v_atan_f32(const float& scale)
{
...
}
v_float32 compute(const v_float32& y, const v_float32& x)
{
...
}
...
v_float32 val90; // sizeless type can not used in a class
v_float32 val180;
v_float32 val360;
v_float32 s;
};
```
3. The API interface does not support/does not match
- ./modules/core/src/norm.cpp
Use `v_popcount`, ~~waiting for #23966~~ Fixed
- ./modules/core/src/has_non_zero.simd.hpp
Use illegal Universal Intrinsic API: For float type, there is no logical operation `|`. Further discussion needed
```cpp
/** @brief Bitwise OR
Only for integer types. */
template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n> operator|(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n>& operator|=(v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
```
```cpp
#if CV_SIMD
typedef v_float32 v_type;
const v_type v_zero = vx_setzero_f32();
constexpr const int unrollCount = 8;
int step = v_type::nlanes * unrollCount;
int len0 = len & -step;
const float* srcSimdEnd = src+len0;
int countSIMD = static_cast<int>((srcSimdEnd-src)/step);
while(!res && countSIMD--)
{
v_type v0 = vx_load(src);
src += v_type::nlanes;
v_type v1 = vx_load(src);
src += v_type::nlanes;
....
src += v_type::nlanes;
v0 |= v1; //Illegal ?
....
//res = v_check_any(((v0 | v4) != v_zero));//beware : (NaN != 0) returns "false" since != is mapped to _CMP_NEQ_OQ and not _CMP_NEQ_UQ
res = !v_check_all(((v0 | v4) == v_zero));
}
v_cleanup();
#endif
```
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [ ] I agree to contribute to the project under Apache 2 License.
- [ ] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [ ] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
- to reduce binaries size of FFmpeg Windows wrapper
- MinGW linker doesn't support -ffunction-sections (used for FFmpeg Windows wrapper)
- move code to improve locality with its used dependencies
- move UMat::dot() to matmul.dispatch.cpp (Mat::dot() is already there)
- move UMat::inv() to lapack.cpp
- move UMat::mul() to arithm.cpp
- move UMat:eye() to matrix_operations.cpp (near setIdentity() implementation)
- move normalize(): convert_scale.cpp => norm.cpp
- move convertAndUnrollScalar(): arithm.cpp => copy.cpp
- move scalarToRawData(): array.cpp => copy.cpp
- move transpose(): matrix_operations.cpp => matrix_transform.cpp
- move flip(), rotate(): copy.cpp => matrix_transform.cpp (rotate90 uses flip and transpose)
- add 'OPENCV_CORE_EXCLUDE_C_API' CMake variable to exclude compilation of C-API functions from the core module
- matrix_wrap.cpp: add compile-time checks for CUDA/OpenGL calls
- the steps above allow to reduce FFmpeg wrapper size for ~1.5Mb (initial size of OpenCV part is about 3Mb)
backport is done to improve merge experience (less conflicts)
backport of commit: 65eb946756
Abhishek Gola
Alexander Smorkalov
jiajia Qian
Lurie97
Prasad Ayush Kumar
pratham-mcw
Yang Guanyuhan
Madan mohan Manokar
Vadim Pisarevsky
Alexander Smorkalov
Liutong HAN
jvuillaumier
Yuantao Feng
Maksim Shabunin
rogday
Vitaly Tuzov
Alexander Alekhin