dnn: parallelize nary elementwise forward implementation & enable related conformance tests #25630
This PR introduces the following changes:
- [x] Parallelize binary forward impl
- [x] Parallelize ternary forward impl (Where)
- [x] Parallelize nary (Operator that can take >=1 operands)
- [x] Enable conformance tests if workable
## Performance
### i7-12700K, RAM 64GB, Ubuntu 22.04
```
Geometric mean (ms)
Name of Test opencv opencv opencv
perf perf perf
core.x64.0606 core.x64.0606 core.x64.0606
vs
opencv
perf
core.x64.0606
(x-factor)
NCHW_C_sum::Layer_NaryEltwise::OCV/CPU 16.116 11.161 1.44
NCHW_NCHW_add::Layer_NaryEltwise::OCV/CPU 17.469 11.446 1.53
NCHW_NCHW_div::Layer_NaryEltwise::OCV/CPU 17.531 11.469 1.53
NCHW_NCHW_equal::Layer_NaryEltwise::OCV/CPU 28.653 13.682 2.09
NCHW_NCHW_greater::Layer_NaryEltwise::OCV/CPU 21.899 13.422 1.63
NCHW_NCHW_less::Layer_NaryEltwise::OCV/CPU 21.738 13.185 1.65
NCHW_NCHW_max::Layer_NaryEltwise::OCV/CPU 16.172 11.473 1.41
NCHW_NCHW_mean::Layer_NaryEltwise::OCV/CPU 16.309 11.565 1.41
NCHW_NCHW_min::Layer_NaryEltwise::OCV/CPU 16.166 11.454 1.41
NCHW_NCHW_mul::Layer_NaryEltwise::OCV/CPU 16.157 11.443 1.41
NCHW_NCHW_pow::Layer_NaryEltwise::OCV/CPU 163.459 15.234 10.73
NCHW_NCHW_ref_div::Layer_NaryEltwise::OCV/CPU 10.880 10.868 1.00
NCHW_NCHW_ref_max::Layer_NaryEltwise::OCV/CPU 10.947 11.058 0.99
NCHW_NCHW_ref_min::Layer_NaryEltwise::OCV/CPU 10.948 10.910 1.00
NCHW_NCHW_ref_mul::Layer_NaryEltwise::OCV/CPU 10.874 10.871 1.00
NCHW_NCHW_ref_sum::Layer_NaryEltwise::OCV/CPU 10.971 10.920 1.00
NCHW_NCHW_sub::Layer_NaryEltwise::OCV/CPU 17.546 11.462 1.53
NCHW_NCHW_sum::Layer_NaryEltwise::OCV/CPU 16.175 11.475 1.41
NHWC_C::Layer_NaryEltwise::OCV/CPU 11.339 11.333 1.00
NHWC_H::Layer_NaryEltwise::OCV/CPU 16.154 11.102 1.46
```
### Apple M1, RAM 16GB, macOS 14.4.1
```
Geometric mean (ms)
Name of Test opencv opencv opencv
perf perf perf
core.m1.0606 core.m1.0606.patch core.m1.0606.patch
vs
opencv
perf
core.m1.0606
(x-factor)
NCHW_C_sum::Layer_NaryEltwise::OCV/CPU 28.418 3.768 7.54
NCHW_NCHW_add::Layer_NaryEltwise::OCV/CPU 6.942 5.679 1.22
NCHW_NCHW_div::Layer_NaryEltwise::OCV/CPU 5.822 5.653 1.03
NCHW_NCHW_equal::Layer_NaryEltwise::OCV/CPU 5.751 5.628 1.02
NCHW_NCHW_greater::Layer_NaryEltwise::OCV/CPU 5.797 5.599 1.04
NCHW_NCHW_less::Layer_NaryEltwise::OCV/CPU 7.272 5.578 1.30
NCHW_NCHW_max::Layer_NaryEltwise::OCV/CPU 5.777 5.562 1.04
NCHW_NCHW_mean::Layer_NaryEltwise::OCV/CPU 5.819 5.559 1.05
NCHW_NCHW_min::Layer_NaryEltwise::OCV/CPU 5.830 5.574 1.05
NCHW_NCHW_mul::Layer_NaryEltwise::OCV/CPU 5.759 5.567 1.03
NCHW_NCHW_pow::Layer_NaryEltwise::OCV/CPU 342.260 74.655 4.58
NCHW_NCHW_ref_div::Layer_NaryEltwise::OCV/CPU 8.338 8.280 1.01
NCHW_NCHW_ref_max::Layer_NaryEltwise::OCV/CPU 8.359 8.309 1.01
NCHW_NCHW_ref_min::Layer_NaryEltwise::OCV/CPU 8.412 8.295 1.01
NCHW_NCHW_ref_mul::Layer_NaryEltwise::OCV/CPU 8.380 8.297 1.01
NCHW_NCHW_ref_sum::Layer_NaryEltwise::OCV/CPU 8.356 8.323 1.00
NCHW_NCHW_sub::Layer_NaryEltwise::OCV/CPU 6.818 5.561 1.23
NCHW_NCHW_sum::Layer_NaryEltwise::OCV/CPU 5.805 5.570 1.04
NHWC_C::Layer_NaryEltwise::OCV/CPU 3.834 4.817 0.80
NHWC_H::Layer_NaryEltwise::OCV/CPU 28.402 3.771 7.53
```
### 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: add DepthToSpace and SpaceToDepth #25779
We are working on updating WeChat QRCode module. One of the new models is a fully convolutional model and hence it should be able to run with different input shapes. However, it has an operator `DepthToSpace`, which is parsed as a subgraph of `Reshape -> Permute -> Reshape` with a fixed shape getting during parsing. The subgraph itself is not a problem, but the true problem is the subgraph with a fixed input and output shape regardless input changes. This does not allow the model to run with different input shapes.
Solution is to add a dedicated layer for DepthtoSpace and SpaceToDepth.
Backend support:
- [x] CPU
- [x] CUDA
- [x] OpenCL
- [x] OpenVINO
- [x] CANN
- [x] TIMVX
- ~Vulkan~ (missing fundamental tools, like permutation and reshape)
### 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
Refactor DNN module to build with cudnn 9 #25412
A lot of APIs that are currently being used in the dnn module have been removed in cudnn 9. They were deprecated in 8.
This PR updates said code accordingly to the newer API.
Some key notes:
1) This is my first PR. I am new to openCV.
2) `opencv_test_core` tests pass
3) On a 3080, cuda 12.4(should be irrelevant since I didn't build the `opencv_modules`, gcc 11.4, WSL 2.
4) For brevity I will avoid including macro code that will allow for older versions of cudnn to build.
I was unable to get the tests working for `opencv_test_dnn` and `opencv_perf_dnn`. The errors I get are of the following:
```
OpenCV tests: Can't find required data file: dnn/onnx/conformance/node/test_reduce_prod_default_axes_keepdims_example/model.onnx in function 'findData'
" thrown in the test body.
```
So before I spend more time investigating I was hoping to get a maintainer to point me in the right direction here. I would like to run these tests and confirm things are working as intended. I may have missed some details.
### Pull Request Readiness Checklist
relevant issue
(https://github.com/opencv/opencv/issues/24983
- [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
- [x] 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
Merge with https://github.com/opencv/opencv_extra/pull/1158
Todo:
- [x] Fix Attention pattern recognition.
- [x] Handle other backends.
Benchmark:
"VIT_B_32 OCV/CPU", M1, results in milliseconds.
| Model | 4.x | This PR |
| - | - | - |
| VIT_B_32 OCV/CPU | 87.66 | **83.83** |
### 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
dnn: avoid const layer forwarding in layer norm layer and attention layer #25238
While profiling ViTs with dnn, I found `ConstLayer` can take a proportion of the inference time, which is weird. This comes from the data copy during the inference of `ConstLayer`. There is a chance that we can improve the efficiency of data copying but the easiest and most convenient way is to avoid `ConstLayer`. This PR change the way how we handle constants in layer normalization layer and attention layer, which is storing in the layer blobs instead of making constant layers for them.
Checklists:
- [x] Backend compatibility in layer normalization layer.
### 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
dnn onnx: add group norm layer #24610
dnn onnx: add group norm layer
Todo:
- [x] speed up by multi-threading
- [x] add perf
- [x] add backend: OpenVINO
- [x] add backend: CUDA
- [x] add backend: OpenCL (no fp16)
- [ ] add backend: CANN
### 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
Co-authored-by: fengyuentau <yuantao.feng@opencv.org.cn>
dnn onnx: add mod #24765
Resolves https://github.com/opencv/opencv/issues/23174
TODO:
- [x] enable some conformance tests
- [x] add backends
- [x] CANN
- [x] OpenVINO
- [x] CUDA
### 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
dnn: refactor ONNX MatMul with fastGemm #24694
Done:
- [x] add backends
- [x] CUDA
- [x] OpenVINO
- [x] CANN
- [x] OpenCL
- [x] Vulkan
- [x] add perf tests
- [x] const B case
### Benchmark
Tests are done on M1. All data is in milliseconds (ms).
| Configuration | MatMul (Prepacked) | MatMul | InnerProduct |
| - | - | - | - |
| A=[12, 197, 197], B=[12, 197, 64], trans_a=0, trans_b=0 | **0.39** | 0.41 | 1.33 |
| A=[12, 197, 64], B=[12, 64, 197], trans_a=0, trans_b=0 | **0.42** | 0.42 | 1.17 |
| A=[12, 50, 64], B=[12, 64, 50], trans_a=0, trans_b=0 | **0.13** | 0.15 | 0.33 |
| A=[12, 50, 50], B=[12, 50, 64], trans_a=0, trans_b=0 | **0.11** | 0.13 | 0.22 |
| A=[16, 197, 197], B=[16, 197, 64], trans_a=0, trans_b=0 | **0.46** | 0.54 | 1.46 |
| A=[16, 197, 64], B=[16, 64, 197], trans_a=0, trans_b=0 | **0.46** | 0.95 | 1.74 |
| A=[16, 50, 64], B=[16, 64, 50], trans_a=0, trans_b=0 | **0.18** | 0.32 | 0.43 |
| A=[16, 50, 50], B=[16, 50, 64], trans_a=0, trans_b=0 | **0.15** | 0.25 | 0.25 |
### 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
dnn cuda: support Sub #24647
Related https://github.com/opencv/opencv/issues/24606#issuecomment-1837390257
### 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
dnn: add openvino, opencl and cuda backends for layer normalization layer #24552
Merge after https://github.com/opencv/opencv/pull/24544.
Todo:
- [x] openvino
- [x] opencl
- [x] cuda
### 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
dnn onnx: add instance norm layer #24378
Resolves https://github.com/opencv/opencv/issues/24377
Relates https://github.com/opencv/opencv/pull/24092#discussion_r1349841644
| Perf | multi-thread | single-thread |
| - | - | - |
| x: [2, 64, 180, 240] | 3.95ms | 11.12ms |
Todo:
- [x] speed up by multi-threading
- [x] add perf
- [x] add backend: OpenVINO
- [x] add backend: CUDA
- [x] add backend: OpenCL (no fp16)
- [ ] add backend: CANN (will be done via https://github.com/opencv/opencv/pull/24462)
### 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
```
force_builders=Linux OpenCL,Win64 OpenCL,Custom
buildworker:Custom=linux-4
build_image:Custom=ubuntu:18.04
modules_filter:Custom=none
disable_ipp:Custom=ON
```
Resolve uncovered CUDA dnn layer #24080
### Pull Request Readiness Checklist
* Gelu activation layer on CUDA
* Try to relax GEMM from ONNX
resolves https://github.com/opencv/opencv/issues/24064
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
Conv1D and Pool1D for CUDA backend
* CUDA-independent changes
* Add Conv1D and Pool1D for CUDA backend
* CUDA-independent changes
* Fix typo
* fix comment
* Update fix
* make changes more correct for pooling layer
* Minor fixes for review
* Split skip blocks
