Port of #29377 to 5.x. Wires Log, Erf, Exp, Sin, Cos, Sinh, Cosh, Tan,
Softplus, BNLL, Asinh, Acosh, Atanh into the dispatched activation_kernels
registry, plus a Layer_Activation perf test. 1.4-12.3x across
M4/A76/Threadripper/Xeon, correct to <=5e-7 vs scalar.
dnn: SIMD for transcendental activation layers (15 functors) 🧑💻🤖#2937
# dnn: SIMD for transcendental activation layers (15 functors) 🧑💻🤖
### Summary
Many `BaseDefaultFunctor`-based activation functors fall back to the scalar `BaseDefaultFunctor::apply`
— one `libm` call per element. This PR adds vectorized `apply()` overrides (universal intrinsics,
matching the existing `MishFunctor`/`GeluFunctor`/`SigmoidFunctor` pattern, each with a scalar tail) to
the **15 transcendental activations** that have a clean SIMD path:
| functor | uses | | functor | uses |
|---|---|---|---|---|
| TanH | `v_exp` (1−2/(e²ˣ+1)) | | Asinh | `v_log`,`v_sqrt` (sign·log(\|x\|+√(x²+1))) |
| Log | `v_log` | | Acosh | `v_log`,`v_sqrt` |
| Erf | `v_erf` | | Atanh | `v_log` |
| Exp | `v_exp` | | Softplus | `v_exp`,`v_log` |
| Sin | `v_sin` | | BNLL | `v_exp`,`v_log` (max(x,0)+log1p(e^−\|x\|)) |
| Cos | `v_cos` | | GeluApproximation | `v_exp` (tanh) |
| Tan | `v_sin`/`v_cos` | | Sinh / Cosh | `v_exp` |
Functors compile at the baseline SIMD width (SSE on x86, NEON on ARM — same as the existing SIMD
functors). `Sqrt`/`Floor`/`Ceil`/`Round`/`Abs`/etc. are intentionally **not** included — those are
single cheap ops the compiler already auto-vectorizes (measured `Sqrt` at 0.95×). `Asin`/`Acos`/`Atan`
are omitted (no `v_atan`/`v_asin`/`v_acos` intrinsic).
### Performance
Real `opencv_perf_dnn` `Layer_Activation` (added here), 8×256×128×100 = 26.2 M-elem CV_32F blob, A/B
vs the scalar fallback, median speedup:
| functor | M4 clang/NEON | A76 gcc/NEON | Threadripper gcc | Xeon W-2235 gcc | functor | M4 | A76 | TR | Xeon |
|---|---|---|---|---|---|---|---|---|---|
| TanH | 9.1× | 2.3× | 7.3× | 7.3× | Sinh | 3.3× | 3.2× | 5.2× | 5.0× |
| Cos | 8.8× | 3.0× | 3.3× | 3.0× | Acosh | 3.2× | 2.9× | 2.7× | 2.6× |
| Log | 7.8× | 3.0× | 2.5× | 2.4× | Cosh | 2.8× | 3.2× | 2.6× | 2.5× |
| Sin | 7.8× | 3.0× | 3.4× | 2.9× | Exp | 2.4× | 2.5× | 1.4× | 1.5× |
| GeluApprox | 7.6× | 2.4× | 6.6× | 5.8× | BNLL | 4.2× | 2.0× | 1.6× | 1.4× |
| Atanh | 7.0× | 3.6× | 6.0× | 5.2× | Softplus | 2.0× | 2.0× | 3.3× | 2.7× |
| Tan | 6.8× | 3.6× | 6.6× | 6.0× | Erf | 4.2× | 2.4× | 3.9× | 3.5× |
| Asinh | 3.9× | 2.4× | 5.8× | 5.7× | | | | | |
Every functor is a win on every tested out-of-order core (worst case 1.4×). On the **in-order
Cortex-A55** (a55-tuned build) all 15 measure 0.985–0.99× — neutral, within noise: the SIMD
polynomials are dependency-chain-bound, which an in-order pipeline can neither accelerate nor (here)
slow down. So: meaningful wins on out-of-order ARM + x86, no regression on in-order ARM.
### Accuracy
Uses the `v_exp`/`v_log`/`v_erf`/`v_sin`/`v_cos` polynomials OpenCV already ships and relies on (Gelu
via `v_erf`, Mish/Sigmoid via `v_exp`). Verified vs scalar `libm` on 320 K random elements per functor
with domain-correct inputs (Acosh x≥1, Atanh |x|<1, …): **max abs error ≤ 3e-5** (most ≤ 2e-6), zero
elements exceeding rel>1e-3 & abs>1e-4. ONNX conformance and the layer accuracy tests exercise these ops.
### Pull Request Readiness Checklist
- [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 (4.x)
- [ ] There is a reference to the original bug report and related work — N/A (perf improvement)
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable — perf test added (`Layer_Activation`, `SANITY_CHECK_NOTHING`); accuracy covered by existing ONNX conformance/layer tests
- [ ] The feature is well documented and sample code can be built with the project CMake — N/A (internal optimization, no new API)
[FOLLOW UP] : Documentation optimizations for the new Sphinx structure #29220
### Pull Request Readiness Checklist
This PR serves as a follow-up to the new documentation system introduced in [#29206](https://github.com/opencv/opencv/pull/29206)
Co-authored by: @abhishek-gola @kirtijindal14 @Akansha-977 @Prasadayus @varun-jaiswal17
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
Caffe importer cleanup #28678
Merge with: https://github.com/opencv/opencv_extra/pull/1324
### 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
- [x] The feature is well documented and sample code can be built with the project CMake
Update default YuNet model to new dynamic inputs #29107
Update the default model in `face_detect.py` and `face_detect.cpp` to
`face_detection_yunet_2026may.onnx`, which has symbolic `height`/`width` input dims.
## Changes
- `samples/dnn/face_detect.py`: update default `--face_detection_model` to `face_detection_yunet_2026may.onnx`
- `samples/dnn/face_detect.cpp`: update default `fd_model` to `face_detection_yunet_2026may.onnx`
Companion PR :
- https://github.com/opencv/opencv_zoo/pull/310
- https://github.com/opencv/opencv_extra/pull/1373
Closes : https://github.com/opencv/opencv/issues/28769
### 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
requires (YOLO26n , RetinaFace model download) : https://github.com/opencv/opencv_extra/pull/1357
added performance tests for 14 modern deep learning models to `modules/dnn/perf/perf_net.cpp.`
**Tests added**
- RT_DETR_L
- RF_DETR
- Grounding_DINO
- BlazeFace
- RetinaFace
- YOLO26n
- YOLO26m_Seg
- SegFormer_B2_Clothes
- Depth_Anything_V2
- SigLIP
- OWLv2
- RAFT
- SAM2_Encoder
- SAM2_Decoder
Models are available at: https://drive.google.com/drive/folders/1lpYAWSLMtxcmfDtYjiw0qwzZX0bCH-i4
### 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
Extend several operations to support block layout #28839
Requires opencv_extra: https://github.com/opencv/opencv_extra/pull/1347
After this PR, we observed the following improvements in the listed models:
Model | Before (ENGINE_NEW) | After (ENGINE_NEW) | ENGINE_ORT | % Improvement (Before v/s After)
-- | -- | -- | -- | --
Face_Paint | 514.68ms | 384.40ms | 394.38ms | 25.31%
BlazeFace | 1.03ms | 0.83ms | 0.66ms | 19.42%
**Device details:**
Model name: Intel(R) Core(TM) i9-14900KS, x86_64, 32 Cores, Ubuntu 22.04.5 LTS
Operations covered:
- [x] Pad
- [x] Resize
- [x] Reshape
- [x] Shape
- [x] InstanceNorm
- [x] GroupNorm
### 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
Added Attention fusion and thin gemm support #28859
Merge with: https://github.com/opencv/opencv_extra/pull/1350
Performance numbers after these optimizations:
For Device: Intel(R) Core(TM) i9-14900KS, x86, 32 Cores, ubuntu 24.04,
| Model | `ENGINE_NEW (Before)` | `ENGINE_NEW (After)` | `ENGINE_ORT` |
| :--- | :--- | :--- | :--- |
| **BERT** |26.3 ms| 9.15 ms| 9.13 ms|
| **ViT** | 79.65 ms| 63.23 ms| 32.3 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
Parallelize DNN layers using chunking #28821
At a high level, I replaced tensor-level parallelism with chunk-level parallelism. Previously, parallel_for_ was dispatched over the number of input or output tensors i.e. one thread handled one whole tensor's copy.
The new approach precomputes each tensor's destination offset and per-slice size upfront, then slices the total byte work into fixed 64 KB chunks. The full chunk count is handed to parallel_for_ as a single flat range, and each worker decodes its chunk index back into (tensor, slice, byte_offset) using a prefix-sum table before running a plain memcpy on its piece. A small-size threshold falls back to the sequential path so we don't get threading overhead on small tensors.
Performance numbers after these optimizations:
For Device: Intel(R) Core(TM) i9-14900KS, x86, 32 Cores, ubuntu 24.04,
| Model | `ENGINE_NEW` | `ENGINE_ORT` |
| :--- | :--- | :--- |
| **YOLOv8n** |10.9 ms| 12.15 ms|
| **YOLOv5n** | 8.36 ms| 9.23 ms|
| **YOLOX-S** | 23.46 ms| 25.16 ms|
For Device: Macbook M1 Air
| Model | `ENGINE_NEW` | `ENGINE_ORT` |
| :--- | :--- | :--- |
| **YOLOv8n** |34.45 ms| 42.52 ms|
| **YOLOv5n** | 31.62 ms| 25.52 ms|
| **YOLOX-S** | 88.9 ms| 116.7 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
Added getFLOPS support in new DNN engine #28634
closes: https://github.com/opencv/opencv/issues/26199
### 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
optimize(dnn): parallelize Slice layer implementation (4.x) #28511
### Summary
Backport of PR #28447 to 4.x branch.
### Description
This PR optimizes the SliceLayer implementation for strided inputs (where step > 1). The original implementation used a recursive element-wise copy (getSliceRecursive) for any strided slice, which was extremely inefficient.
This PR introduces:
- **Parallelization**: Uses `cv::parallel_for_` to parallelize the outermost dimension of the slice operation.
- **Memcpy Optimization**: Automatically detects "pseudo-contiguous" blocks in strided slices (e.g., slicing an outer dimension but keeping inner dimensions intact) and uses `std::memcpy` instead of scalar loops.
- **Refactoring**: Replaces the recursive function with a dedicated `ParallelSlice` loop body.
### Impact
Significant performance improvement for strided slice operations (common in detection heads, strided sampling, etc.).
### Benchmark Results
Tested on CPU with 20 threads.
| Test Case | Baseline (ms) | Optimized (ms) | Speedup |
| :--- | :--- | :--- | :--- |
| Strided Axis 0 [::2, ...] | 1.10 | 0.02 | **~55x** |
| Strided Axis 2 [..., ::2] | 1.15 | 0.11 | **~10.5x** |
### 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
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
- [ ] The feature is well documented and sample code can be built with the project CMake
optimize(dnn): optimize Slice layer for strided inputs #28447
## Description
This PR optimizes the `SliceLayer` implementation for strided inputs (where `step > 1`). The original implementation used a recursive element-wise copy (`getSliceRecursive`) for any strided slice, which was extremely inefficient.
This PR introduces:
1. **Parallelization**: Uses `cv::parallel_for_` to parallelize the outermost dimension of the slice operation.
2. **Memcpy Optimization**: Automatically detects "pseudo-contiguous" blocks in strided slices (e.g., slicing an outer dimension but keeping inner dimensions intact) and uses `std::memcpy` instead of scalar loops.
3. **Refactoring**: Replaces the recursive function with a dedicated `ParallelSlice` loop body.
## Impact
Significant performance improvement for strided slice operations (common in detection heads, strided sampling, etc.).
**Benchmark Results:**
| Test Case | Before (ms) | After (ms) | Speedup |
| :--- | :--- | :--- | :--- |
| **Strided Axis 0** `[::2, ...]` | 1.10 | **0.02** | **~55x** |
| **Strided Axis 2** `[..., ::2]` | 1.10 | **0.06** | **~18x** |
| **Contiguous** (Baseline) | 0.10 | 0.10 | 1.0x (Unchanged) |
### 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
- [ ] 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
optimize(dnn): parallelize Resize layer implementation #28442
### optimize(dnn): parallelize Resize layer implementation
This PR addresses the `TODO` in `modules/dnn/src/layers/resize_layer.cpp` regarding the slow implementation of the `Resize` layer when using `opencv_linear` or `nearest` interpolation with specific configurations.
### Changes
- Replaced the serial nested loop (batch x channels) with `cv::parallel_for_`.
- This allows OpenCV to utilize multi-threading for the resizing operation, which was previously single-threaded for these specific paths.
### Performance Results
Benchmark run on 20-thread CPU (AVX2):
| Test Case | Resolution Change | Before (ms) | After (ms) | Speedup |
| :--- | :--- | :--- | :--- | :--- |
| **Upsample Linear** | 64x64 -> 128x128 | 1.47 ms | **0.18 ms** | **~8.1x** |
| **Downsample Nearest** | 128x128 -> 64x64 | 1.55 ms | **0.45 ms** | **~3.4x** |
**Test Configuration:**
- **Upsample**: `[4, 64, 64, 64]` -> `[4, 64, 128, 128]` (Factor: 2.0, Linear)
- **Downsample**: `[4, 128, 128, 128]` -> `[4, 128, 64, 64]` (Factor: 0.5, Nearest)
### 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
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
- Added `modules/dnn/perf/perf_resize.cpp` to verify performance gains.
- [ ] The feature is well documented and sample code can be built with the project CMake
Switch to the new Github actions pipeline for Windows in 5.x too #28035
### 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
Faster implementation of blobFromImages for cpu nchw output #26127
Faster implementation of blobFromImage and blobFromImages for
HWC cv::Mat images -> NCHW cv::Mat
case
Running time on my pc in ms:
**blobFromImage**
```
image size old new speed-up
32x32x3 0.008 0.002 4.0x
64x64x3 0.021 0.009 2.3x
128x128x3 0.164 0.037 4.4x
256x256x3 0.728 0.158 4.6x
512x512x3 3.310 0.628 5.2x
1024x1024x3 14.503 3.124 4.6x
2048x2048x3 61.647 28.049 2.2x
```
**blobFromImages**
```
image size old new speed-up
16x32x32x3 0.122 0.041 3.0x
16x64x64x3 0.790 0.165 4.8x
16x128x128x3 3.313 0.652 5.1x
16x256x256x3 13.495 3.127 4.3x
16x512x512x3 58.795 28.127 2.1x
16x1024x1024x3 251.135 121.955 2.1x
16x2048x2048x3 1023.570 487.188 2.1x
```
### 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.
- [x] The feature is well documented and sample code can be built with the project CMake
Modified tensorflow parser for the new dnn engine #26394
### 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
dnn: add ONNX TopK #23279
Merge with https://github.com/opencv/opencv_extra/pull/1200
Partially fixes#22890 and #20258
To-do:
- [x] TopK forward impl
- [x] add tests
- [x] support Opset 1 & 10 if possible
- [ ] ~Support other backends~ (TopK has two outputs, which is not supported by other backends, such as openvino)
Perf:
M1 (time in millisecond)
| input shape | axis | dnn | ort |
| --------------- | ---- | ---- | ---- |
| (1000, 100) | 0 | 1.68 | 4.07 |
| (1000, 100) K5 | 0 | 1.13 | 0.12 |
| (1000, 100) | 1 | 0.96 | 0.77 |
| (100, 100, 100) | 0 | 10.00 | 31.13 |
| (100, 100, 100) | 1 | 7.33 | 9.17 |
| (100, 100, 100) | 2 | 7.52 | 9.48 |
M2 (time in milisecond)
| input shape | axis | dnn | ort |
| --------------- | ---- | ---- | ---- |
| (1000, 100) | 0 | 0.76 | 2.44 |
| (1000, 100) K5 | 0 | 0.68 | 0.07 |
| (1000, 100) | 1 | 0.41 | 0.50 |
| (100, 100, 100) | 0 | 4.83 | 17.52|
| (100, 100, 100) | 1 | 3.60 | 5.08 |
| (100, 100, 100) | 2 | 3.73 | 5.10 |
ONNXRuntime performance testing script: https://gist.github.com/fengyuentau/a119f94fd16721ec9974b8c7b0a45d4c
### 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: optimize activations with v_exp #25881
Merge with https://github.com/opencv/opencv_extra/pull/1191.
This PR optimizes the following activations:
- [x] Swish
- [x] Mish
- [x] Elu
- [x] Celu
- [x] Selu
- [x] HardSwish
### Performance (Updated on 2024-07-18)
#### AmLogic A311D2 (ARM Cortex A73 + A53)
```
Geometric mean (ms)
Name of Test activations activations.patch activations.patch
vs
activations
(x-factor)
Celu::Layer_Elementwise::OCV/CPU 115.859 27.930 4.15
Elu::Layer_Elementwise::OCV/CPU 27.846 27.003 1.03
Gelu::Layer_Elementwise::OCV/CPU 0.657 0.602 1.09
HardSwish::Layer_Elementwise::OCV/CPU 31.885 6.781 4.70
Mish::Layer_Elementwise::OCV/CPU 35.729 32.089 1.11
Selu::Layer_Elementwise::OCV/CPU 61.955 27.850 2.22
Swish::Layer_Elementwise::OCV/CPU 30.819 26.688 1.15
```
#### Apple M1
```
Geometric mean (ms)
Name of Test activations activations.patch activations.patch
vs
activations
(x-factor)
Celu::Layer_Elementwise::OCV/CPU 16.184 2.118 7.64
Celu::Layer_Elementwise::OCV/CPU_FP16 16.280 2.123 7.67
Elu::Layer_Elementwise::OCV/CPU 9.123 1.878 4.86
Elu::Layer_Elementwise::OCV/CPU_FP16 9.085 1.897 4.79
Gelu::Layer_Elementwise::OCV/CPU 0.089 0.081 1.11
Gelu::Layer_Elementwise::OCV/CPU_FP16 0.086 0.074 1.17
HardSwish::Layer_Elementwise::OCV/CPU 1.560 1.555 1.00
HardSwish::Layer_Elementwise::OCV/CPU_FP16 1.536 1.523 1.01
Mish::Layer_Elementwise::OCV/CPU 6.077 2.476 2.45
Mish::Layer_Elementwise::OCV/CPU_FP16 5.990 2.496 2.40
Selu::Layer_Elementwise::OCV/CPU 11.351 1.976 5.74
Selu::Layer_Elementwise::OCV/CPU_FP16 11.533 1.985 5.81
Swish::Layer_Elementwise::OCV/CPU 4.687 1.890 2.48
Swish::Layer_Elementwise::OCV/CPU_FP16 4.715 1.873 2.52
```
#### Intel i7-12700K
```
Geometric mean (ms)
Name of Test activations activations.patch activations.patch
vs
activations
(x-factor)
Celu::Layer_Elementwise::OCV/CPU 17.106 3.560 4.81
Elu::Layer_Elementwise::OCV/CPU 5.064 3.478 1.46
Gelu::Layer_Elementwise::OCV/CPU 0.036 0.035 1.04
HardSwish::Layer_Elementwise::OCV/CPU 2.914 2.893 1.01
Mish::Layer_Elementwise::OCV/CPU 3.820 3.529 1.08
Selu::Layer_Elementwise::OCV/CPU 10.799 3.593 3.01
Swish::Layer_Elementwise::OCV/CPU 3.651 3.473 1.05
```
### 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
* added v_erf and implemented gelu acceleration via vectorization
* remove anonymous v_erf and use v_erf from intrin_math
* enable perf for ov and cuda backend
Change opencv_face_detector related tests and samples from caffe to onnx #25463
Part of https://github.com/opencv/opencv/issues/25314
This PR aims to change the tests related to opencv_face_detector from caffe framework to onnx. Tests in `test_int8_layer.cpp` and `test_caffe_importer.cpp` will be removed in https://github.com/opencv/opencv/pull/25323
### 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
Added int32, int64 support and type inference to dnn #24411
**Added a type inference to dnn similar to the shape inference, added int32 and int64 support.**
- Added getTypes method for layers that calculates layer outputs types and internals types from inputs types (Similar to getMemoryShapes). By default outputs and internals types = input[0] type
- Added type inference pipeline similar to shape inference pipeline. LayersShapes struct (that is used in shape inference pipeline) now contains both shapes and types
- All layers output blobs are now allocated using the calculated types from the type inference.
- Inputs and constants with int32 and int64 types are not automatically converted into float32 now.
- Added int32 and int64 support for all the layers with indexing and for all the layers required in tests.
Added int32 and int64 support for CUDA:
- Added host<->device data moving for int32 and int64
- Added int32 and int64 support for several layers (just slightly modified CUDA C++ templates)
Passed all the accuracy tests on CPU, OCL, OCL_FP16, CUDA, CUDA_FP16. (except RAFT model)
**CURRENT PROBLEMS**:
- ONNX parser always converts int64 constants and layers attributes to int32, so some models with int64 constants doesn't work (e.g. RAFT). The solution is to disable int64->int32 conversion and fix attributes reading in a lot of ONNX layers parsers (https://github.com/opencv/opencv/issues/25102)
- I didn't add type inference and int support to VULCAN, so it doesn't work at all now.
- Some layers don't support int yet, so some unknown models may not work.
**CURRENT WORKAROUNDS**:
- CPU arg_layer indides are implemented in int32 followed by a int32->int64 conversion (the master branch has the same workaround with int32->float conversion)
- CPU and OCL pooling_layer indices are implemented in float followed by a float->int64 conversion
- CPU gather_layer indices are implemented in int32, so int64 indices are converted to int32 (the master branch has the same workaround with float->int32 conversion)
**DISABLED TESTS**:
- RAFT model
**REMOVED TESTS**:
- Greater_input_dtype_int64 (because it doesn't fit ONNX rules, the whole test is just comparing float tensor with int constant)
**TODO IN NEXT PULL REQUESTS**:
- Add int64 support for ONNX parser
- Add int support for more layers
- Add int support for OCL (currently int layers just run on CPU)
- Add int tests
- Add int support for other backends
Vulkan backend for NaryEltwiseLayer in DNN module #24768
We improve Vulkan backend for ``NaryEltwiseLayer`` in DNN module by:
- add a basic framework for Vulkan backend in ``NaryEltwiseLayer``
- add a compute shader for binary forwarding (an imitation of what has been done in native OpenCV backend including broadcasting and eltwise-operation)
- typo fixed:
- Wrong info output in ``context.cpp``
Currently, our implementation (or all layers supporting Vulkan backend) runs pretty slow on discrete GPUs basically due to IO cost in function ``copyToHost``, and we are going to fix that by
- find out the best ``VkMemoryProperty`` for various discrete GPUs
- prevent ``copyToHost`` in middle layers during forwarding, (i.e keep data in GPU memory)
### 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
Co-authored-by: IskXCr <IskXCr@outlook.com>
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>
Abhishek Gola
Teddy-Yangjiale
Alexander Smorkalov
Prasadayus
Tonu Samuel
Tõnu Samuel
Andrei Fedorov
omrope79
Alexander Smorkalov
Varun Jaiswal
vrooomy
Prasad Ayush Kumar
Hanbin Bae
Hanbin Bae
alexlyulkov
Maksim Shabunin
Vadim Pisarevsky
Alexander Smorkalov
Yuantao Feng
Yuantao Feng
Wanli
Haosonn
jimmylaw21