Rewrite Universal Intrinsic code: float related part #24325
The goal of this series of PRs is to modify the SIMD code blocks guarded by CV_SIMD macro: rewrite them by using the new Universal Intrinsic API.
The series of PRs is listed below:
#23885 First patch, an example
#23980 Core module
#24058 ImgProc module, part 1
#24132 ImgProc module, part 2
#24166 ImgProc module, part 3
#24301 Features2d and calib3d module
#24324 Gapi module
This patch (hopefully) is the last one in the series.
This patch mainly involves 3 parts
1. Add some modifications related to float (CV_SIMD_64F)
2. Use `#if (CV_SIMD || CV_SIMD_SCALABLE)` instead of `#if CV_SIMD || CV_SIMD_SCALABLE`,
then we can get the `CV_SIMD` module that is not enabled for `CV_SIMD_SCALABLE` by looking for `if CV_SIMD`
3. Summary of `CV_SIMD` blocks that remains unmodified: Updated comments
- Some blocks will cause test fail when enable for RVV, marked as `TODO: enable for CV_SIMD_SCALABLE, ....`
- Some blocks can not be rewrited directly. (Not commented in the source code, just listed here)
- ./modules/core/src/mathfuncs_core.simd.hpp (Vector type wrapped in class/struct)
- ./modules/imgproc/src/color_lab.cpp (Array of vector type)
- ./modules/imgproc/src/color_rgb.simd.hpp (Array of vector type)
- ./modules/imgproc/src/sumpixels.simd.hpp (fixed length algorithm, strongly ralated with `CV_SIMD_WIDTH`)
These algorithms will need to be redesigned to accommodate scalable backends.
### 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
Fix tests writing to current work dir #24343
Several tests were writing files in the current work directory and did not clean up after test. Moved all temporary files to the `/tmp` dir and added a cleanup code.
Fixed CumSum dnn layer #24353Fixes#20110
The algorithm had several errors, so I rewrote it.
Also the layer didn't work with non constant axis tensor. Fixed it.
Enabled CumSum layer tests from ONNX conformance.
OpenVINO backend for INT8 models #23987
### Pull Request Readiness Checklist
TODO:
- [x] DetectionOutput layer (https://github.com/opencv/opencv/pull/24069)
- [x] Less FP32 fallbacks (i.e. Sigmoid, eltwise sum)
- [x] Accuracy, performance tests (https://github.com/opencv/opencv/pull/24039)
- [x] Single layer tests (convolution)
- [x] ~~Fixes for OpenVINO 2022.1 (https://pullrequest.opencv.org/buildbot/builders/precommit_custom_linux/builds/100334)~~
Performace results for object detection model `coco_efficientdet_lite0_v1_1.0_quant_2021_09_06.tflite`:
| backend | performance (median time) |
|---|---|
| OpenCV | 77.42ms |
| OpenVINO 2023.0 | 10.90ms |
CPU: `11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz`
Serialized model per-layer stats (note that Convolution should use `*_I8` primitives if they are quantized correctly): https://gist.github.com/dkurt/7772bbf1907035441bb5454f19f0feef
---
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
Optimization for parallelization when large core number #24280
**Problem description：**
When the number of cores is large, OpenCV’s thread library may reduce performance when processing parallel jobs.
**The reason for this problem:**
When the number of cores (the thread pool initialized the threads, whose number is as same as the number of cores) is large, the main thread will spend too much time on waking up unnecessary threads.
When a parallel job needs to be executed, the main thread will wake up all threads in sequence, and then wait for the signal for the job completion after waking up all threads. When the number of threads is larger than the parallel number of a job slices, there will be a situation where the main thread wakes up the threads in sequence and the awakened threads have completed the job, but the main thread is still waking up the other threads. The threads woken up by the main thread after this have nothing to do, and the broadcasts made by the waking threads take a lot of time, which reduce the performance.
**Solution：**
Reduce the time for the process of main thread waking up the worker threads through the following two methods:
• The number of threads awakened by the main thread should be adjusted according to the parallel number of a job slices. If the number of threads is greater than the number of the parallel number of job slices, the total number of threads awakened should be reduced.
• In the process of waking up threads in sequence, if the main thread finds that all parallel job slices have been allocated, it will jump out of the loop in time and wait for the signal for the job completion.
**Performance Test:**
The tests were run in the manner described by https://github.com/opencv/opencv/wiki/HowToUsePerfTests.
At core number = 160, There are big performance gain in some cases.
Take the following cases in the video module as examples:
OpticalFlowPyrLK_self::Path_Idx_Cn_NPoints_WSize_Deriv::("cv/optflow/frames/VGA_%02d.png", 2, 1, (9, 9), 11, true)
Performance improves 191%:0.185405ms ->0.0636496ms
perf::DenseOpticalFlow_VariationalRefinement::(320x240, 10, 10)
Performance improves 112%:23.88938ms -> 11.2562ms
Among all the modules, the performance improvement is greatest on module video, and there are also certain improvements on other modules.
At core number = 160, the times labeled below are the geometric mean of the average time of all cases for one module. The optimization is available on each module.
overall | time(ms) | | | | | | |
-- | -- | -- | -- | -- | -- | -- | -- | --
module name | gapi | dnn | features2d | objdetect | core | imgproc | stitching | video
original | 0.185 | 1.586 | 9.998 | 11.846 | 0.205 | 0.215 | 164.409 | 0.803
optimized | 0.174 | 1.353 | 9.535 | 11.105 | 0.199 | 0.185 | 153.972 | 0.489
Performance improves | 6% | 17% | 5% | 7% | 3% | 16% | 7% | 64%
Meanwhile, It is found that adjusting the order of test cases will have an impact on some test cases. For example, we used option --gtest-shuffle to run opencv_perf_gapi, the performance of TestPerformance::CmpWithScalarPerfTestFluid/CmpWithScalarPerfTest::(compare_f, CMP_GE, 1920x1080, 32FC1, { gapi.kernel_package }) case had 30% changes compared to the case without shuffle. I would like to ask if you have also encountered such a situation and could you share your experience?
### 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
dnn: merge tests from test_halide_layers to test_backends #24283
Context: https://github.com/opencv/opencv/pull/24231#pullrequestreview-1628649980
### 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
cmake: Fix riscv-gnu toolchain file. #24305
cmake(3.22.1) failed without the keyword `PATHS` on my device when I manually set `TOOLCHAIN_COMPILER_LOCATION_HINT` in command. And this patch is going to fix this issue.
[CMake Doc](https://cmake.org/cmake/help/latest/command/find_program.html):
> find_program (
> <VAR>
> name | NAMES name1 [name2 ...] [NAMES_PER_DIR]
> [HINTS [path | ENV var]... ]
> [PATHS [path | ENV var]... ]
### 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
Rewrite Universal Intrinsic code: features2d and calib3d module. #24301
The goal of this series of PRs is to modify the SIMD code blocks guarded by CV_SIMD macro: rewrite them by using the new Universal Intrinsic API.
This is the modification to the features2d module and calib3d module.
Test with clang 16 and QEMU v7.0.0. `AP3P.ctheta1p_nan_23607` failed beacuse of a small calculation error. But this patch does not touch the relevant code, and this error always reproduce on QEMU, regardless of whether the patch is applied or not. I think we can ignore it
```
[ RUN ] AP3P.ctheta1p_nan_23607
/home/hanliutong/project/opencv/modules/calib3d/test/test_solvepnp_ransac.cpp:2319: Failure
Expected: (cvtest::norm(res.colRange(0, 2), expected, NORM_INF)) <= (3e-16), actual: 3.33067e-16 vs 3e-16
[ FAILED ] AP3P.ctheta1p_nan_23607 (26 ms)
...
[==========] 148 tests from 64 test cases ran. (1147114 ms total)
[ PASSED ] 147 tests.
[ FAILED ] 1 test, listed below:
[ FAILED ] AP3P.ctheta1p_nan_23607
```
Note: There are 2 test cases failed with GCC 13.2.1 without this patch, seems like there are someting wrong with RVV part on GCC.
```
[----------] Global test environment tear-down
[==========] 148 tests from 64 test cases ran. (1511399 ms total)
[ PASSED ] 146 tests.
[ FAILED ] 2 tests, listed below:
[ FAILED ] Calib3d_StereoSGBM.regression
[ FAILED ] Calib3d_StereoSGBM_HH4.regression
```
The patch is partially auto-generated by using the [rewriter](https://github.com/hanliutong/rewriter).
### 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
Add Support for Einsum Layer #24037
### This PR adding support for [Einsum Layer](https://pytorch.org/docs/stable/generated/torch.einsum.html) (in progress).
This PR is currently not to be merged but only reviewed. Test cases are located in [#1090](https://github.com/opencv/opencv_extra/pull/1090)RP in OpenCV extra
**DONE**:
- [x] 2-5D GMM support added
- [x] Matrix transpose support added
- [x] Reduction type comupte 'ij->j'
- [x] 2nd shape computation - during forward
**Next PRs**:
- [ ] Broadcasting reduction "...ii ->...i"
- [ ] Add lazy shape deduction. "...ij, ...jk->...ik"
- [ ] Add implicit output computation support. "bij,bjk ->" (output subscripts should be "bik")
- [ ] Add support for CUDA backend
- [ ] BatchWiseMultiply optimize
**Later in 5.x version (requires support for 1D matrices)**:
- [ ] Add 1D vector multiplication support
- [ ] Inter product "i, i" (problems with 1D shapes)
### 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
* first commit
* turned C from input to constant; force C constant in impl; better handling 0d/1d cases
* integrate with gemm from ficus nn
* fix const inputs
* adjust threshold for int8 tryQuantize
* adjust threshold for int8 quantized 2
* support batched gemm and matmul; tune threshold for rcnn_ilsvrc13; update googlenet
* add gemm perf against innerproduct
* add perf tests for innerproduct with bias
* fix perf
* add memset
* renamings for next step
* add dedicated perf gemm
* add innerproduct in perf_gemm
* remove gemm and innerproduct perf tests from perf_layer
* add perf cases for vit sizes; prepack constants
* remove batched gemm; fix wrong trans; optimize KC
* remove prepacking for const A; several fixes for const B prepacking
* add todos and gemm expression
* add optimized branch for avx/avx2
* trigger build
* update macros and signature
* update signature
* fix macro
* fix bugs for neon aarch64 & x64
* add backends: cuda, cann, inf_ngraph and vkcom
* fix cuda backend
* test commit for cuda
* test cuda backend
* remove debug message from cuda backend
* use cpu dispatcher
* fix neon macro undef in dispatcher
* fix dispatcher
* fix inner kernel for neon aarch64
* fix compiling issue on armv7; try fixing accuracy issue on other platforms
* broadcast C with beta multiplied; improve func namings
* fix bug for avx and avx2
* put all platform-specific kernels in dispatcher
* fix typos
* attempt to fix compile issues on x64
* run old gemm when neon, avx, avx2 are all not available; add kernel for armv7 neon
* fix typo
* quick fix: add macros for pack4
* quick fix: use vmlaq_f32 for armv7
* quick fix for missing macro of fast gemm pack f32 4
* disable conformance tests when optimized branches are not supported
* disable perf tests when optimized branches are not supported
* decouple cv_try_neon and cv_neon_aarch64
* drop googlenet_2023; add fastGemmBatched
* fix step in fastGemmBatched
* cpu: fix initialization ofb; gpu: support batch
* quick followup fix for cuda
* add default kernels
* quick followup fix to avoid macro redef
* optmized kernels for lasx
* resolve mis-alignment; remove comments
* tune performance for x64 platform
* tune performance for neon aarch64
* tune for armv7
* comment time consuming tests
* quick follow-up fix
VIT track(gsoc realtime object tracking model) #24201
Vit tracker(vision transformer tracker) is a much better model for real-time object tracking. Vit tracker can achieve speeds exceeding nanotrack by 20% in single-threaded mode with ARM chip, and the advantage becomes even more pronounced in multi-threaded mode. In addition, on the dataset, vit tracker demonstrates better performance compared to nanotrack. Moreover, vit trackerprovides confidence values during the tracking process, which can be used to determine if the tracking is currently lost.
opencv_zoo: https://github.com/opencv/opencv_zoo/pull/194
opencv_extra: [https://github.com/opencv/opencv_extra/pull/1088](https://github.com/opencv/opencv_extra/pull/1088)
# Performance comparison is as follows:
NOTE: The speed below is tested by **onnxruntime** because opencv has poor support for the transformer architecture for now.
ONNX speed test on ARM platform(apple M2)(ms):
| thread nums | 1| 2| 3| 4|
|--------|--------|--------|--------|--------|
| nanotrack| 5.25| 4.86| 4.72| 4.49|
| vit tracker| 4.18| 2.41| 1.97| **1.46 (3X)**|
ONNX speed test on x86 platform(intel i3 10105)(ms):
| thread nums | 1| 2| 3| 4|
|--------|--------|--------|--------|--------|
| nanotrack|3.20|2.75|2.46|2.55|
| vit tracker|3.84|2.37|2.10|2.01|
opencv speed test on x86 platform(intel i3 10105)(ms):
| thread nums | 1| 2| 3| 4|
|--------|--------|--------|--------|--------|
| vit tracker|31.3|31.4|31.4|31.4|
preformance test on lasot dataset(AUC is the most important data. Higher AUC means better tracker):
|LASOT | AUC| P| Pnorm|
|--------|--------|--------|--------|
| nanotrack| 46.8| 45.0| 43.3|
| vit tracker| 48.6| 44.8| 54.7|
[https://youtu.be/MJiPnu1ZQRI](https://youtu.be/MJiPnu1ZQRI)
In target tracking tasks, the score is an important indicator that can indicate whether the current target is lost. In the video, vit tracker can track the target and display the current score in the upper left corner of the video. When the target is lost, the score drops significantly. While nanotrack will only return 0.9 score in any situation, so that we cannot determine whether the target is lost.
### 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.
- [ ] The feature is well documented and sample code can be built with the project CMake
HAN Liutong
Alexander Smorkalov
Dmitry Kurtaev
Alexander Smorkalov
Wanli
Maksim Shabunin
Yuriy Chernyshov
Sérgio M. Basto
Vadim Levin
alexlyulkov
Alex
Alexander Alekhin
Emmanuel Ferdman
fengyuentau
casualwinds
Alexander Lyulkov
ashadrina
Abduragim Shtanchaev
lpylpy0514