dnn(cuda): transformer layer coverage and CUDA Graph replay on the new engine - #30439
Follow-up to #30415 (ResNet layer coverage and cuDNN JIT conv/pooling on the new engine). Adds CUDA support on the new DNN engine
for the layers transformer models need, and brings the CUDA path to parity with ONNX Runtime.
**Companion PR** opencv/opencv_extra#1448
**Changes**
- CUDA execs for Attention, Cast, ConstantOfShape, Expand, Gather, Not, Reduce, Shape, Slice, Softmax, Tile, Transpose, Concat,
Gelu, LayerNormalization, MatMul, Split, Squeeze, Unsqueeze and the element-wise activations.
- A graph with static shapes is recorded as one CUDA Graph on the second `forward()` and replayed from then on.
- The cuDNN JIT convolution reads NCHW buffers in place; NHWC staging is kept only for FP16 grouped convolutions.
- TF32 for cuBLAS GEMMs (`OPENCV_DNN_CUDA_ALLOW_TF32`), FP32 accumulation for FP16 GEMMs
(`OPENCV_DNN_CUDA_FP16_HALF_ACCUMULATE`), and fused MatMul/Gemm post-ops.
- Batch-broadcast MatMul and Shape are safe to capture in a CUDA Graph; the unused pointer-array `cublas::gemmBatched` is
removed.
- Tests: `forward()` runs three times on batch-broadcast MatMul and Shape models; `MatMul_init_bcast` runs on CUDA again;
`AttentionSharedShapeReshape` uses the FP16 default tolerance on CUDA FP16.
- Perf: new CUDA-only `*_CUDA_pinned` tests pin the input shapes.
**Performance** (ms, median of 50, FP32 with TF32, RTX PRO 6000 Blackwell, idle GPU)
| Model | CPU | CUDA (cuDNN JIT) | CUDA (full cuDNN) | ORT 1.31 CUDA EP |
|---|---|---|---|---|
| ResNet-50 | 8.02 | 0.80 | 0.80 | 0.88 |
| ResNet-18 | 3.27 | 0.44 | 0.44 | 0.48 |
| MobileNetV2 | 1.99 | 0.28 | 0.27 | 0.40 |
| DenseNet-121 | 17.42 | 1.48 | 1.58 | 1.58 |
| EfficientNet-Lite4 | 5.03 | 0.53 | 0.54 | 0.73 |
| ViT-B/32 | 11.06 | 1.02 | 1.02 | 0.95 |
| ViT-Base/16 | 24.38 | 1.16 | 1.17 | 1.13 |
| DeiT-Tiny | 4.58 | 0.65 | 0.65 | 0.68 |
| MobileViT-XS | 6.93 | 0.72 | 0.72 | 0.77 |
| YOLOv3 | 47.21 | 2.60 | 2.59 | 2.72 |
| YOLOv4 | 109.38 | 3.58 | 3.61 | 4.18 |
| YOLOv4-tiny | 7.31 | 0.64 | 0.64 | 0.66 |
| BEiT-Base | 25.30 | 1.24 | 1.24 | 1.19 |
| BERT | 10.00 | 0.97 | 0.97 | 0.99 |
| SigLIP | 41.14 | 2.09 | 2.09 | 2.44
### 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
dnn(onnx): add support for ONNX Compress operator - #30576
### Description
This PR adds support for the ONNX `Compress` operator (opset 9-24) to the OpenCV DNN dynamic engine:
- Implements `CompressLayer` in `modules/dnn/src/layers/compress_layer.cpp` supporting dynamic output shapes (`dynamicOutputShapes() { return true; }`).
- Implements full ONNX `Compress` semantics:
- Slices along a specified axis when the `axis` attribute is provided (supports negative axis indexing via `normalize_axis`).
- Flattens the input tensor to 1D before compression when the `axis` attribute is omitted.
- Discards elements exceeding condition length when condition length is shorter than target axis dimension (`std::min(dimAxis, condTotal)`).
- Handles all-false condition ($K=0$) cleanly with early return after fitting the valid empty output shape (`[..., 0, ...]`).
- Enforces continuous tensor invariants with `CV_Assert(data.isContinuous())` and validates `condition` as `CV_Bool` per ONNX spec.
- Slicing optimized via OpenCV native `parallel_for_` flattening multi-dimensional iteration domains to 1D (`totalSlices = outer * K`).
- Un-denylists and passes all 4 official ONNX conformance tests in `test_onnx_conformance_layer_parser_denylist.inl.hpp`.
- Marks `Compress` conformance tests with `SKIP;` in `test_onnx_conformance_layer_filter__openvino.inl.hpp` to ensure OpenVINO CI remains green.
- Adds focused Google Test unit tests in `modules/dnn/test/test_layers.cpp` verifying empty output behavior on all-false condition both with and without `axis`.
### Test Results
All 4 ONNX conformance tests + 2 all-false unit tests pass locally:
```text
[==========] Running 6 tests from 2 test cases.
[----------] Global test environment set-up.
[----------] 2 tests from Test_Compress
[ RUN ] Test_Compress.AllFalse_WithAxis
[ OK ] Test_Compress.AllFalse_WithAxis (29 ms)
[ RUN ] Test_Compress.AllFalse_WithoutAxis
[ OK ] Test_Compress.AllFalse_WithoutAxis (0 ms)
[----------] 2 tests from Test_Compress (30 ms total)
[----------] 4 tests from Test_ONNX_conformance
[ RUN ] Test_ONNX_conformance.Layer_Test/test_compress_0_OCV_CPU
[ OK ] Test_ONNX_conformance.Layer_Test/test_compress_0_OCV_CPU (901 ms)
[ RUN ] Test_ONNX_conformance.Layer_Test/test_compress_1_OCV_CPU
[ OK ] Test_ONNX_conformance.Layer_Test/test_compress_1_OCV_CPU (465 ms)
[ RUN ] Test_ONNX_conformance.Layer_Test/test_compress_default_axis_OCV_CPU
[ OK ] Test_ONNX_conformance.Layer_Test/test_compress_default_axis_OCV_CPU (456 ms)
[ RUN ] Test_ONNX_conformance.Layer_Test/test_compress_negative_axis_OCV_CPU
[ OK ] Test_ONNX_conformance.Layer_Test/test_compress_negative_axis_OCV_CPU (559 ms)
[----------] 4 tests from Test_ONNX_conformance (2383 ms total)
[----------] Global test environment tear-down
[==========] 6 tests from 2 test cases ran. (2413 ms total)
[ PASSED ] 6 tests.
```
### Checklist
- [x] I have read the [OpenCV contribution guidelines](https://github.com/opencv/opencv/wiki/How_to_contribute).
- [x] My code follows the OpenCV coding style guidelines.
- [x] I have tested my changes locally and all relevant unit tests pass.
- [x] No unrelated files or whitespace changes are included.
Added flag for bitexact resize in blobFromImage and overloads. - #30131
Current INTER_LINEAR resize allows small deviations up to rounding. It affects inference result for networks that produce coordinates (object detectors, features) and descriptors. The new option allow to get the same result independently on CPU arch and used HALs.
Companion PR for references update: https://github.com/opencv/opencv_extra/pull/1422
### 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
Remove Tensorflow Importer - #29474
[PR OpenCV Extra](https://github.com/opencv/opencv_extra/pull/1390)
[PR OpenCV Contrib](https://github.com/opencv/opencv_contrib/pull/4171)
Removed tensorflow importers
### 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): move the CUDA backend to UMat device buffers - #29807
This PR is in continuation with #29658 and #30259
### 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
dnn: fix blobRectsToImageRects when the image is not resized (5.x) - #30287
### 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. 5.x port of #30243, as requested there. Related: #30042.
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable. Accuracy test added; no test data needed.
- [ ] The feature is well documented and sample code can be built with the project CMake. Not applicable to this bug fix.
### Problem
Port of #30243 to 5.x. `Image2BlobParams::blobRectToImageRect()` and `blobRectsToImageRects()` are identical on 5.x and have the same bug:
- When `Image2BlobParams::size` equals the image size, every output rect is `Rect(0, 0, 0, 0)`, in all three padding modes.
- When `size` is empty (the default, which `blobFromImageWithParams()` treats as "keep the image size"), the mapping divides by zero and returns `INT_MIN` coordinates.
```python
import cv2 as cv
p = cv.dnn.Image2BlobParams()
p.size = (640, 640)
p.paddingmode = cv.dnn.DNN_PMODE_LETTERBOX
print(p.blobRectToImageRect((10, 20, 30, 40), (640, 640)))
# (0, 0, 0, 0), expected (10, 20, 30, 40)
p.size = (0, 0)
print(p.blobRectToImageRect((10, 20, 30, 40), (640, 480)))
# (-2147483648, -2147483648, -2147483648, -2147483648), expected (10, 20, 30, 40)
```
### Change
Same change as #30243, cherry-picked without conflicts. `blobRectsToImageRects()` now follows the same two rules as `blobFromImagesWithParams()` on 5.x: an empty `size` means the image size, and when the sizes match the input rects are returned unchanged. The resize, crop-center and letterbox mappings are unchanged. It also includes the Doxygen `@note` on both functions requested in the #30243 review.
The new `blobRectToImageRect.no_resize` test covers both cases for all three padding modes, and checks that the blob built with the same parameters is pixel-identical to the image.
#30042 contains an equivalent pass-through for the equal-size case as part of a larger Model API change. This PR is the standalone fix, including the empty-size case. If #30042 lands first, the conflict should be limited to `blobRectsToImageRects()`.
### Testing
Built `opencv_test_dnn` and the Python bindings from this branch (base 5.x `6dc5ff5c1f`) on Ubuntu 26.04 (WSL2), GCC 15.2, Ninja, `BUILD_LIST=dnn,python3,ts`.
- `opencv_test_dnn --gtest_filter=blobRectToImageRect*`: 4 passed. With only the `dnn_utils.cpp` change reverted, the new `blobRectToImageRect.no_resize` fails in all six mode/size combinations (`[0 x 0 from (0, 0)]` for equal sizes, `INT_MIN` for empty size).
- `opencv_test_dnn --gtest_filter='blob*:*Image2Blob*:*imagesFromBlob*'`: 11 passed.
- `modules/dnn/misc/python/test/test_dnn.py`: 8 passed, 2 skipped; 6 failed only because their opencv_extra files are not present locally (`dnn/googlenet_0.png`, `dnn/street.png`, ...).
- The Python reproducer above prints `(10, 20, 30, 40)` for both cases with this branch (`5.1.0-dev`).
- No compiler warnings in the build; `git diff --check` clean.
- Full `opencv_test_dnn` was not run (needs opencv_extra test data).
Added support for Granite Docling 258M and PaddleOCR-VL - #29618
Adds `com.microsoft::GroupQueryAttention` to the engine-new ONNX importer, required by GraniteDocling-258M's ,end-to-end samples for GraniteDocling-258M and PaddleOCR-VL-1.5. Verified on Linux x86_64 / GCC / CPU
**companion opencv_extra PR** [Vlm ocr support engine new opencv/opencv_extra#1412](https://github.com/opencv/opencv_extra/pull/1412) for test data
upstream/5.x already lowers `GroupQueryAttention` to `AttentionOnnxAi`, but that path rejects `do_rotary=1`:
```cpp
CV_CheckEQ(params.get<int>("do_rotary", 0), 0, "GroupQueryAttention: do_rotary=1 is not supported");
```
GraniteDocling-258M's export sets it. Confirmed two ways: building a Llama checkpoint from its published `text_config` (9 heads / 3 KV heads, `rope_theta=100000`) and running onnxruntime-genai's `model_builder`; and reading the community export at `onnx-community/granite-docling-258M-ONNX`, which carries 30 `GroupQueryAttention` nodes with `do_rotary=1`, `num_heads=9`, `kv_num_heads=3`, `head_dim=64`.
## Changes
**Importer** (`onnx_importer2.cpp`, +49)
Routing as above, plus the input validation the op needs: packed QKV rejected, past_key / past_value required as a pair, `do_rotary=1` requires cos/sin, unsupported trailing inputs rejected, and `seqlens_k` shorter than the buffer recognised as a shared-buffer export rather than silently mis-attended.
**Layer** (`group_query_attention_layer.cpp`, +380)
Attention is built on `fastGemmBatch` and `fused_softmax_softcap_mask`, the same kernels `attention_onnxai_layer.cpp` uses. Grouping falls out of the batched-GEMM offset table (query head → `h / groupSize`), so no broadcast copy is needed, and the second GEMM writes straight into the `[B, S, num_heads*D]` output. Supports rotary (full and partial), sliding window, softcap, growing and preallocated caches, and optional `present_*` outputs.
**Core** (`persistence_json.cpp`, +92)
`\uXXXX` escapes are UTF-16 code units, so anything above the BMP arrives as a surrogate pair. These were encoded as two separate 3-byte sequences CESU-8, which is not valid UTF-8 - for keys and values alike. Now recombined into one 4-byte sequence, with unpaired surrogates rejected.
**Tests / perf / samples**
`test_layer_group_query_attention.cpp` (+93), `test_tokenizer.cpp` (+28), `test_tokenizer.py` (+42), `perf_layer.cpp` (+125: `Layer_GroupQueryAttention` MHA_ShortCache and Grouped_WithCache, plus `Layer_Sign`), and two samples (`granite_docling_inference.py`, `paddleocr_vl_inference.py`).
### 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 TFLite importer to support more models on the new engine - #29569
Requires: https://github.com/opencv/opencv_extra/pull/1399
### 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
Unified DNN graph fusion behind fuseChains - #30088
Requires: https://github.com/opencv/opencv_extra/pull/1417
### What This PR Does
Follow-up to #29656, which added the backend-agnostic chain mechanism. That PR introduced it and made pointwise fusion its first consumer; this one moves the rest of the fusion onto it.
- `Net::Impl::fuseBasic()` is gone. Its conv+BatchNorm, conv+activation and conv+residual branches now go through the chain mechanism.
- `fuseBN` and the six other passes now run from inside `fuseChains()`, in their existing order.
- `graph_fusion_basic.cpp` keeps `fuseBN()` and `fuseInstanceNormAffine()`.
- New `FusionOps::foldInputScale` folds a scale sitting before a conv directly into its weights, across the dense, depthwise, grouped and MLAS-prepacked layouts.
### Results
Measured on Intel Core i9-11900K
| Model | 5.x | this PR | speedup |
|-----------------------------|-----------|-----------|---------|
| inception_v2 | 15.04 ms | 11.80 ms | 1.27 |
| efficientnet-lite4 | 14.83 ms | 12.28 ms | 1.21 |
| pose_estimation_mediapipe | 5.66 ms | 5.25 ms | 1.08 |
| siglip_base_patch16_224 | 143.04 ms | 135.27 ms | 1.06 |
| densenet121 | 25.99 ms | 24.76 ms | 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
Generalize tokenizer loading to support method-based family dispatch in DNN - #29675
Companion PR: https://github.com/opencv/opencv_extra/pull/1402
### Changes
Added ALBERT and BERT support end-to-end, with `samples/dnn/albert_inference.py `and `samples/dnn/bert_inference.py `as validation samples, plus expanded coverage in modules/dnn/test/test_tokenizer.cpp. Required changes to `cv::dnn::dnn.hpp`, `graph_fusion_attention.cpp`, and `unicode.cpp/unicode.hpp` to support Unigram and WordPiece tokenizers.
To back ALBERT/BERT, generalized `cv::dnn::Tokenizer `from a single BPE implementation into a method-dispatched frontend, adding `core_wordpiece.cpp/hpp` (WordPiece) and `core_unigram.cpp/hpp` (Unigram) as new backends. `tokenizer.cpp` now routes by method across BPE, Gemma,, SentencePiece, Unigram, and WordPiece behind one shared interface.
Tested against the following samples and the output matches to old tokenizer:
```
gpt2_inference.py
qwen_inference.py
gemma3_inference.py
```
GPT2:
```
Preparing GPT-2 model...
Inferencing GPT-2 model...
Hello, I'm a language model, not a programming language. I'm a language model. I'm a language model. I'm a language model. I'm a language model. I'm a
```
Gemma3:
```
Preparing Gemma3 model...
Prompt:
<start_of_turn>user
What is OpenCV?<end_of_turn>
<start_of_turn>model
Inferencing Gemma3 model...
Response:
Okay, let's break down what OpenCV is.
**What is OpenCV?**
OpenCV (Open Source Computer Vision Library) is a powerful and
```
Qwen2.5:
```
Preparing Qwen2.5 model...
Prompt:
<|im_start|>user
What is OpenCV?<|im_end|>
<|im_start|>assistant
Inferencing Qwen2.5 model...
Response:
OpenCV is a set of computer vision libraries in C++ designed to be used for image and video processing. It provides a wide range of tools and functions for
```
### Tokenizer References
Byte-level BPE: [tokenizers/src/pre_tokenizers/byte_level.rs](https://github.com/huggingface/tokenizers/blob/main/tokenizers/src/pre_tokenizers/byte_level.rs)
(This defines the byte-level mapping rules, which is used in conjunction with the [BPE model](https://www.google.com/search?q=https://github.com/huggingface/tokenizers/blob/main/tokenizers/src/models/bpe/mod.rs))
SentencePiece BPE (Metaspace): [tokenizers/src/pre_tokenizers/metaspace.rs](https://www.google.com/search?q=https://github.com/huggingface/tokenizers/blob/main/tokenizers/src/pre_tokenizers/metaspace.rs)
(This defines the rule for replacing whitespace with the U+2581 _ character and handling byte fallback)
Unigram: [tokenizers/src/models/unigram/mod.rs](https://www.google.com/search?q=https://github.com/huggingface/tokenizers/blob/main/tokenizers/src/models/unigram/mod.rs)
(This contains the core logic for the Unigram lattice scoring and probabilistic tokenization rules)
WordPiece: [tokenizers/src/models/wordpiece/mod.rs](https://www.google.com/search?q=https://github.com/huggingface/tokenizers/blob/main/tokenizers/src/models/wordpiece/mod.rs)
(This explicitly cites Schuster & Nakajima in the code comments and implements the greedy longest-match rule with the ## prefix)
### 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
code under GPL or another license incompatible with OpenCV.
- [x] The PR is proposed to the proper branch (`5.x`).
- [x] There is a reference to the original bug report and related work.
- [x] There is accuracy test and test data in `opencv_extra`, same
branch name (`generalized-tokenizer`) — `bert/`, `t5/` fixtures
back the new C++ tests.
- [x] The feature is documented and sample code builds with project CMake.
FP8 model support in DNN - #29834
ONNX coverage after this PR: 76.6%
co-authored by: @SavyaSanchi-Sharma
### 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 Backend Agnostic fusion in DNN - #29656
# What This Adds
A way for any layer, on any backend, to absorb trailing per-element math — with the pass knowing nothing about either layer and holding no list of op types.
- **Description**: `LayerMath` (one layer's math) → `AdjacencyGraph` (hash-consed chain)
- **Contract**: two `bool` virtuals on `Layer`
- **CPU implementations**: three, behind that contract
Pointwise fusion is the first consumer, not the point.
## The Problem
```cpp
ActivationLayer* activ = dynamic_cast<ActivationLayer*>(layer_ptr);
Conv2Layer* conv = getLayer<Conv2Layer>(newprog, conv_layer_idx);
if (conv) conv->fuseActivation(layer);
```
Guest type, host type, method name all hardcoded in the pass. `Clip` never fused (fails the cast), `Gemm` never fused (no virtual), no 2-op chain fused at all, and no backend could fuse anything without a CPU-specific method.
## Results
| Model | 5.x | this PR | speedup |
|-----------------------|----------|----------|---------|
| MPHand | 2.34 ms | 1.03 ms | 2.27 |
| EfficientNet | 9.82 ms | 5.00 ms | 1.96 |
| MPPose | 5.14 ms | 2.81 ms | 1.83 |
| MobileNet_SSD_v1_ONNX | 12.58 ms | 8.38 ms | 1.50 |
| BlazeFace | 0.95 ms | 0.77 ms | 1.23 |
| DenseNet_121 | 20.56 ms | 19.08 ms | 1.08 |
| MobileNetv2_ONNX | 2.07 ms | 1.99 ms | 1.04 |
| MobileViT_XS | 6.33 ms | 6.11 ms | 1.04 |
| YuNet_320 | 1.33 ms | 1.27 ms | 1.04 |
| PPOCRv3 | 43.05 ms | 42.15 ms | 1.02 |
| BERT | 9.78 ms | 9.68 ms | 1.01 |
| BEiT_Base_Patch16_224 | 27.48 ms | 27.08 ms | 1.01 |
| DeiT_Tiny_Patch16_224 | 4.74 ms | 4.72 ms | 1.01 |
| MPPalm | 1.88 ms | 1.85 ms | 1.01 |
| SSD | 50.19 ms | 49.82 ms | 1.01 |
| YOLOv4_tiny | 7.10 ms | 7.03 ms | 1.01 |
### 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
Merge pull request #29931 from abhishek-gola:onnx_conformance_remaining_fixes
fix ONNX auto_pad, PRelu broadcasting and LSTM peepholes - #29931
closes: https://github.com/opencv/opencv/issues/21078
Upated ONNX coverage: 77.8%
### 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 Bitcast layer & extended MatMul and DFT layers support - #29594
### 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 GridSample BiCubic, Dropout support - #29783
Updated ONNX coverage after this PR: 74.8%
### 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
Linear and Flex attention layers support - #29624
### 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
Add Image Decoder ONNX Layer - #29785
### 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
Dynamic KV-cache support - #29642
The core idea is: reserveKVCache() API to pre-allocate memory for attention caches upfront, which eliminates allocation overhead during token decoding. For LLM inference, simply call reserveKVCache(prompt_len + max_new_tokens) before the prefill stage so the decode loop runs without page allocations, significantly reducing per-token latency for models like Gemma3 and Qwen.
Speedups after this PR on AMD Ryzen 9 9950X 16-Core Processor device:
Qwen2.5-0.5B-Instruct, fp32, CPU, tok/s:
```
Tokens Before After Speedup
64 12.49 23.72 1.90×
128 10.37 23.14 2.23×
256 7.20 22.40 3.11×
512 4.25 21.03 4.95×
```
Gemma 3 1B-it, fp32, CPU, 512 tokens :
```
Tokens Before After Speedup
64 6.99 11.84 1.69×
128 5.84 11.72 2.01×
256 4.17 11.50 2.76×
512 2.47 11.15 4.51×
```
### 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
Add MatMulNBits layer and extend onnx coverage - #29666
Requires:https://github.com/opencv/opencv_extra/pull/1401
### 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
Add Scan layer to the new engine #29577
### 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 Cumprod and Causalconv layers in new dnn engine #29579
### 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
* experimental new arithmetics; work-in-progress
* continue working on new-gen arithmetic expressions
* * improved performance of the new add on small arrays
* added sub
* extended tests
* fixed potential bug when adding multi-channel array and a single-channel scalar
* improved const handling
* * added copyMask
* added mul/dev (without scale so far)
* * accelerated mul
* addedd scale to mul and div
* * done substantial refactoring; however a few more rounds of refactoring are ahead.
* added min, max, absdiff, addweighted.
* improved performance of the new arithmetic functions, but some of them are still slow, e.g. operations with mask have some bugs (that affect speed, not accuracy).
* * further (significantly) accelerated several functions, especially on small arrays: mul, binary ops with mask
* further polished the new arithmetic engine
* started integration of the new element-wise arithmetic engine into core
* big step forward. We now use the new engine inside cv::add, subtract, multiply, divide, absdiff, min and max.
* big progress:
* added bitwise operations
* fixed and accelerated compare
* ported regression tests to test new broadcasting behaviour of arithmetic functions
* lot's of improvements in compare, divide, addWeighted!
* lot's of small and big performance improvements in the new arithmetics
* * some more optimizations; parsing texpr-expressions is now faster as well
* port new_arithm to Linux/x86: dispatch guards, scalar-Mat compat fallback, dnn shape-contract fixes
Core:
- arithm.simd.hpp: CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY guards (the file is included
once per dispatched mode on x86), vx_load_expand instead of the 128-bit v_load_expand,
VTraits::vlanes() instead of ::nlanes
- arithm.cpp/precomp.hpp: compat fallback for scalar-like Mat operands (1x1, 1xcn/cnx1,
4x1 CV_64F - java/python tuples, operator-(Mat, Matx)): treated as a per-channel scalar
ONLY when the shapes are not broadcast-compatible, so every valid numpy-style broadcast
keeps its meaning and calls that would otherwise throw get the 4.x semantics
DNN (fallout of the stricter shape semantics, found by the new engine):
- dict.hpp: DictValue relied on fresh AutoBuffer having size()==fixed_size; allocate explicitly
- batch_norm: weights_/bias_ are 1-D [n] now; 0/1-D forward runs on exact-shape 1-D views
- net_impl2: extend the post-forward sanity check to non-temp outputs - a layer that
reallocates its preallocated output tensor now fails loudly instead of silently
detaching the result from the graph
- LSTM/LSTM2 batchwise (layout=1): getMemoryShapes now matches what forward() writes
(ONNX: Y=(batch,seq,dirs,hid), Yh/Yc=(batch,dirs,hid)); forward assembles seq-major
results in a local buffer and transposes INTO the preallocated outputs in place
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* u8/s8 multiply: exact integer SIMD path on non-FP16 builds (3-10x vs 5.x)
The unit-scale branch of vecBinaryKernel already supported a separate work-vector
type Wvec1 (used by the ARM f16 build and by u16/s16 everywhere), but on x86 the
u8/s8 same-type multiply still went through the f32 hub. Route it through
v_uint16/v_int16: products of 8-bit values fit exactly (255^2 < 2^16), the
saturating pack on store gives bit-exact results at half the vector traffic.
Also fix a latent kernel bug this exposed: the unit-scale branch stepped by
Wvec's lane count while loading/storing Wvec1 vectors. All previous Wvec1
instantiations had equal lane counts, but u8's v_uint16 has 2x the lanes of
v_float32 - the pairs overlapped (50% redundant work) and the tail backoff
could write VECSZ bytes past the row end. The branch now derives its step,
offsets and tail condition from Wvec1 itself.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* vecBinaryKernel: constexpr Op::useScalar instead of a runtime-only scale check
Every binary op functor now declares whether it consumes the scale scalar
(params[0]): true only for mul and the two div variants. Ops that ignore it
(add/sub/min/max/absdiff) take the fast 2-arg branch unconditionally - the
'scalar == 1' check used to fail for them (their params[0] is 0), sending them
through the preproc branch, and the condition now folds at compile time.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* restore cv::hal::mul8u as a wrapper over the element-wise engine
The symbol is still declared in core/hal/hal.hpp and called directly by external
code (the G-API fluid backend in opencv_contrib), but its implementation went
away with the old arithm kernels. Forward it to getMulFunc(CV_8U, CV_8U) - with
scale==1 it lands on the new exact integer SIMD path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* silence every new-arithm warning reported by CI (ARM64/Mac) and gcc 15
- arithm.cpp: bitwise_op_ocl and the actualScalarDepth/coerceTypes helpers are
consumed only by the OpenCL paths - guard them with HAVE_OPENCL; haveScalar in
cv::compare is read only inside CV_OCL_RUN - CV_UNUSED for OpenCL-less builds
- arithm_expr.hpp: declare getBitwiseFunc/getNotFunc/getAddWeightedFunc next to
the other per-op entry points (-Wmissing-prototypes in arithm.dispatch.cpp)
- arithm.simd.hpp: define CV_SIMD_16F to 0 when FP16 SIMD is absent (-Wundef);
{}-init the expandScalar staging buffers (-Wmaybe-uninitialized: they are
fully written before use, but the compiler cannot prove it with runtime
vector widths); rename the compare kernel's lambda parameter (-Wshadow)
- arithm_expr.cpp: rename the exec tile-lambda's hot-field locals that shadowed
TExpr members and outer locals (-Wshadow)
- test_new_arithm_extensive.cpp: rename the name-generator lambdas' parameter
shadowing the INSTANTIATE macro's own (-Wshadow), drop an unused variable
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* restore 4.x scalar semantics for the bindings' 4x1 CV_64F Scalar columns
The python/java bindings materialize numbers and tuples as a 2-D 4x1 CV_64F
Mat - or UMat, when the call carries UMat arguments. Three CI-reported python
failures came from those pseudo-scalars reaching the engine as arrays:
- absdiff(int_arr, 0): the (4,1) column is broadcast-COMPATIBLE with a 1-D
array, so numpy semantics silently won - an outer-product f64 result instead
of the int per-channel-scalar one;
- subtract(u8 4x8x4, (40,)): same, by the rows==4 coincidence;
- multiply(UMat, 2., dst=UMat): the scalar arrives as a UMAT, which the
scalar detection did not recognize at all.
isScalarArg now treats the exact bindings shape - 2-D 4x1 CV_64F single-channel
Mat/UMat against a <=4-channel array - as a scalar UNCONDITIONALLY (a 1-D [4]
array has dims==1 and still broadcasts). One exception, decided in arithm_op:
when the partner is itself a tiny scalar-shaped array, both are honest data and
ride the broadcast (compare(Mat 4x1, Mat 1x1) - issue #8999 - stays elementwise).
Small-array discipline, this all runs per engine call: the probes read Mat/UMat
fields directly (rows == 4 alone rejects almost everything, no _InputArray
getter dispatch), and a UMAT scalar's 32 bytes are copied into a caller-stack
buffer - no heap, no getMat mapping. Measured: no latency change on 4x4/16x16
element-wise calls.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* cv::texpr: std::string_view -> const std::string& in the public API
string_view in an exported signature breaks some CUDA toolchain builds, and for
short expression strings the difference is immaterial (SSO, parsed once). The
parser internals keep string_view - the argument converts implicitly. Also drop
the now-unused <string_view> include from cvstd.hpp, so the header does not
reach every nvcc TU.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* compare boundary-rewrite: fixed 4-slot kind/bound arrays instead of AutoBuffers
A CONST operand is capped at 4 channels (addConst), so the per-channel
kind/bound staging needs no dynamic buffers - plain int[4]/double[4], with a
CV_Assert on the contract. This is also what gcc's -Wmaybe-uninitialized was
flagging (it could not see the AutoBuffer's inline storage get filled).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* element-wise engine: unary math kernels (sqrt/exp/log/sin/cos/tanh/erf/relu) + select
New dispatched pair math.simd.hpp / math.dispatch.cpp - the unary/ternary sibling
of arithm.simd.hpp:
- vecUnaryKernel: T -> T over f16/bf16/f32/f64 on top of the intrin_math
primitives (v_exp/v_log/v_sin/v_cos/v_sqrt/v_erf/v_max). f32/f64 compute
natively, f16/bf16 ride the f32 hub inside the kernel (vx_load_pair_as /
v_store_pair_as) - no materialized casts. Continuity collapse + the halide
right-edge backoff, suppressed in-place (it would re-apply Op to
already-written values). tanh = (e^2x-1)/(e^2x+1) with the input clamped to
+/-10 (f32) / +/-20 (f64) - unclamped saturation hits inf/inf = NaN. erf has
no f64 SIMD primitive: std::erf per lane.
- selectKernel(mask, x, y): 1-byte mask expanded to lane width and tested
against zero in the INTEGER domain (immune to DAZ/FTZ), branches of any
depth by element size, broadcast branches supported.
- emitUnary: math over a float input is T -> T now (f16 in -> f16 out, native
kernel when input and result depths match); integer inputs still compute in
the float domain and land in f32.
- emitTernary/select: literal branches are typed via typedConstFrom (an
OP_CAST of a depth-less flex const crashed); a non-1-byte mask is normalized
by an explicit "mask != 0" compare, never a value cast.
texpr already parsed the function names - they now execute. Tests: per-depth
accuracy of all 8 ops against the double std:: reference, integer input,
in-place, select over 4 depths / const branch / float mask.
Perf vs the classic kernels (1920x1080 f32, 16 threads, AVX2): exp 3.7x,
log 2.6x, sqrt 5.7x faster; polarToCart expressed as (r*cos(a), r*sin(a)) 4.0x.
Accuracy improves too (max rel err vs f64 reference): exp 8.1e-8 vs 2.1e-7,
log 8.0e-8 vs 1.5e-7.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* engine: OP_COPY_MASK folded into OP_SELECT; selectKernel moved to arithm.simd.hpp
copyMask(dst, mask, src) is select(mask, src, dst) - one masking primitive
instead of two. The compiler emits the masked-op tail as
addInsn(OP_SELECT, mask, r, out, out): the output slot rides as both arg2 and
the result, so unmasked elements are preserved by reading them back through
the b-branch. OP_COPY_MASK, copyMaskKernel and getCopyMaskFunc are gone.
selectKernel (moved from math.simd.hpp to arithm.simd.hpp) inherits every
copyMaskKernel optimization:
- the interleaved multichannel fast path (2..4 channels under a per-pixel
mask: expand the mask once per VECSZ rows, v_store_interleave across lanes);
- the per-row scalar path with the row-skip when the selected source row IS
dst (the "leave the output untouched" half of copyMask);
- plus the select-specific ones: branch broadcasts (stepx == 0) and the
right-edge tail backoff under dst-aliases-a-branch - safe because re-running
select over already-blended elements is idempotent; only dst == mask keeps
the backoff off (the store would rewrite mask bytes before the re-read).
Masked-add perf is on par with the old copyMask (1280x720, 1 thread: 8UC3
204 -> 198 us, 32FC3 1395 -> 1344, 8UC1/32FC1 within noise). Full core suite
24117 green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* engine: dedicated vectorized pow kernel (moved from arithm to math.simd.hpp)
pow was the last scalar-only binary op (scalarBinaryKernel + std::pow, f32/f64
only). The new powKernel keeps exact std::pow semantics and is T x T -> T over
all four float depths (f16/bf16 via the f32 hub):
- scalar exponent (the dominant call shape - texpr literals ride as 0-dim
broadcast consts) is dispatched PER ROW to the special cases:
y==2 -> x*x, y==3 -> x*x*x, y==0.5 -> v_sqrt, y==1 -> copy, y==0 -> fill 1;
- everything else - including a per-element exponent array - runs the general
vectorized exp(y * log(x)) path, valid for x > 0; a vector pair containing
any x <= 0 lane falls back to scalar std::pow for that pair (v_check_any),
which preserves every std::pow subtlety: signed results for integer y on
negative bases, NaN for fractional y, the x == 0 family;
- no right-edge tail backoff: pow is not idempotent, in-place calls finish
rows in the scalar tail.
Perf vs the classic cv::pow (1920x1080 f32, 16 threads, AVX2): p=2 1.4x
(classic special-cases it too), p=3 6.1x, p=0.5 6.3x, fractional p 4.5x with
slightly better accuracy (6.1e-7 vs 7.4e-7 max rel err). Tests: exponent
sweep 2/3/0.5/1/0/2.5/-1.5 vs the double std::pow reference on f32/f64,
negative bases (exact signed cubes, NaN for fractional), array exponent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* powKernel: halide right-edge tail backoff in every SIMD loop
Same shape as vecBinaryKernel: the final partial vector re-processes
[width - VECSZ*2, width) instead of finishing scalar, suppressed when dst
aliases an input (pow is not idempotent - the overlap region must be
recomputed from an untouched source, which the no-alias case guarantees).
Modest measured win (~2% on ROI rows for the special-cased exponents; the
general path tail was already cheap - modern libm powf is fast), no
regressions; mainly aligns the kernel with the house style, where every
SIMD loop ends vector-wide.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix MSVC 2019 C2975: function-local constexpr as a template argument inside a lambda
MSVC 2019 loses the constexpr-ness of function-local constants (LOCAL_OPS,
MAX_DIMS, ...) when they are used as template arguments inside a lambda body
(AutoBuffer<Slice, LOCAL_OPS> / std::array<int, MAX_DIMS> in the parallel
bodies of BroadcastOp::run and TExpr::exec). Hoist them to namespace scope -
no behavior change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* ocl_arithm_op: route 16U multiply to the CPU engine on Apple OpenCL
The Apple OpenCL driver miscompiles the 16U multiply kernel: products near
the top of the u16 range come back wrapped instead of saturated (CPU vs GPU
NORM_INF up to 65535 in OCL_Arithm/Mul.Mat CV_16U cases). The same arithm.cl
kernel is correct on Intel NEO and NVIDIA drivers - verified not to reproduce
on Linux/Intel iGPU - so gate the decline to __APPLE__ only; the CPU engine
computes 16u multiply exactly (integer SIMD path).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* engine: neg/abs as compositions, clamp kernel, ** and ?: operators, abs(a-b) peephole
- OP_NEG and OP_ABS need no kernels: neg = sub(0, a), abs = absdiff(a, 0) -
including the engine absdiff auto-type rule (signed |a| lands in the
UNSIGNED type of the same width: |SHRT_MIN| fits u16 exactly instead of
saturating; NB the public cv::absdiff auto depth keeps the source type for
4.x compatibility - values agree, the depth rule is the engine own).
- peephole: abs(x - y) rewrites to absdiff(x, y) ALWAYS. On integers the
literal semantics differ (the subtract saturates first: u8 gives
max(x-y, 0)), but whoever writes abs(a - b) means absdiff - we deliberately
hand out the useful semantics instead of the saturation artifact. The just-
emitted OP_SUB is retired via the moveToOutput manoeuvre, so the program
shrinks to the single absdiff instruction. abs(x), abs(x - 0) and
absdiff(x, 0) all give one result.
- OP_CLAMP kernel (arithm.simd.hpp): v_min(v_max(x, lo), hi) over
u8/s8/u16/s16/u32/s32/f32 (+f64 with 64-bit SIMD), scalar f16/bf16/64-bit
ints; lo/hi may broadcast (the common clamp(img, a, b) shape) or be full
arrays; the tail backoff stays on under dst-aliases-x (clamp is idempotent).
emitTernary types literal bounds via typedConstFrom (same flex-const crash
select had) and keeps the auto result type pinned to x.
- parser: "a ** b" == pow(a, b), precedence above * /, RIGHT-associative
(a ** 2 ** 3 == a ** 8); "cond ? a : b" == select(cond, a, b), precedence
below everything, right-associative chains (f1 ? a : f2 ? b : c) work
without parentheses. parseTernary() is the expression entry point now.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* cv::exp/log/sqrt on the engine via math_op; hal functions wrapped as engine kernels
math_op is the master function of the unary math family (the arithm_op
analogue): same-shape same-type output over f16/bf16/f32/f64 (classic
exp/log accepted f32/f64 only - the half floats are new), two tiers:
- small (<= 100000 elements) and continuous: call the kernel DIRECTLY over
the flattened data - no TExpr, no broadcastOp, no parallel_for setup;
- everything else: the usual single-instruction program via compile()/exec()
(parallelism for large arrays, real steps for ROIs).
getMathFunc routes OP_EXP/OP_LOG at f32/f64 through the full cv::hal stack -
an external vendor HAL (CALL_HAL), IPP, or the built-in table kernels,
whichever is installed - by wrapping hal::exp32f/exp64f/log32f/log64f as
engine kernels with the function pointer in TKernel::userdata, the same
mechanism castKernel uses for core BinaryFuncs. The engine adds tiling and
parallelism on top, so every tier gets the best available scalar-span
implementation. v_exp/v_log remain for f16/bf16 (the f32 hub) and the ops
hal has no entry points for.
v_log_default_32f: the degree-8 polynomial is evaluated by Estrin pairing
(4 dependent levels) instead of an 8-FMA Horner chain (~5% on the f16 hub
path). An exp64 Taylor-without-division rewrite was tried and benched SLOWER
than the Cephes Pade scheme (the evaluation is FMA-throughput-bound, and
vdivpd pipelines well enough) - reverted; a table-based reduction is the only
way further there.
cv::exp f32 (16 threads, AVX2+IPP build), old -> new: 640 elements
0.16 -> 0.15 us, 16k 2.79 -> 2.49, 640x480 47 -> 20, 1920x1080 452 -> 60 us
(the old CPU loops were single-threaded); f64 exp 1080p 1295 -> 161 us.
No size regresses; small arrays now run at installed-HAL speed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* getMathFunc: IPP tier + raw-HAL probing; the exp/log table kernels are deleted
The raw cv_hal_* entry points return int for a reason: without an installed
HAL they are stubs returning CV_HAL_ERROR_NOT_IMPLEMENTED. getMathFunc now
selects the exp/log implementation in three tiers:
1. HAVE_IPP && ipp::useIPP(): ippsExp/ippsLn through thin int adapters (IPP
is not routed through the cv_hal_ hooks, so it needs its own tier);
2. the raw cv_hal_exp32f/... hook, PROBED once with a 1-element call on the
safe input 1.0 (cached in magic statics): implemented -> wrapped as an
engine kernel with the function pointer in TKernel::userdata;
3. the engine own v_exp/v_log kernels.
Whichever wins, the engine adds tiling and parallelism on top.
The EXPTAB/LOGTAB table kernels and their tables (~790 lines in
mathfuncs_core.simd.hpp + mathfuncs.cpp) are DELETED: they benched within
~15% of v_exp/v_log, not worth a second implementation. The public
cv::hal::exp32f/exp64f/log32f/log64f keep their contract - CALL_HAL, then
IPP, then the built-in implementation - but the built-in is now the engine
vector kernel via ew::mathSpanEngine (one contiguous span, exported from
math.dispatch.cpp).
All unary math kernels (vec/scalar/hal wrappers, pow) also handle the
vertical-broadcast tile (s0y == 0, a row expanded into a matrix): the first
row is computed, the rest are memcpy of it - transcendentals cost far more
than a row copy.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* texpr: hypot(x, y) binary op (alias: mag)
hypot = sqrt(x^2 + y^2), NAIVE like cv::magnitude (not the overflow-safe
std::hypot), computed in the float work type; kernels for the four float
depths only (T x T -> T; integer inputs ride the usual f32-compute + cast).
A 10-line EwHypot functor on top of vecBinaryKernel in arithm.simd.hpp -
broadcast branches, continuity collapse and the tail backoff come for free.
Registered in the parser as both "hypot" (the C/numpy name) and "mag" (the
cv::magnitude-flavored alias). A building block for the future cartToPolar.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* texpr: atan2(y, x) binary op - radians, standard C range
v_atan2 (arithm.simd.hpp, generic over the universal-intrinsic float vector):
the fastAtan2 minimax polynomial from mathfuncs_core v_atan_f32 reworked to
plain radians - the 180/pi factor dropped from the coefficients and the C
quadrant logic instead of the [0, 360) wrap, so the result matches std::atan2
over (-pi, pi]. Measured absolute accuracy ~1.6e-4 rad. (v_atan_f32 itself is
untouched - cv::phase/fastAtan2 keep their degree semantics.)
EwAtan2 rides vecBinaryKernel: f16/bf16/f32 through v_atan2 (the f32 hub),
f64 through exact scalar std::atan2. arg0 = y, arg1 = x, like std::atan2;
float depths only, same emitBinary policy as pow/hypot. Parser name "atan2".
Together with hypot this completes the cartToPolar building blocks.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix the RISC-V RVV build and two ARM64 warnings
The new f64 kernel registrations (hypot, pow, the unary math family) gate on
CV_SIMD_64F || CV_SIMD_SCALABLE_64F, but the vx_setall_as(const double*,
v_float64&) helper family in arithm.simd.hpp was still CV_SIMD_64F-only -
scalable platforms (RVV) have v_float64 with CV_SIMD_64F == 0, so
vecBinaryKernel<double, ...> failed to instantiate there. Widen the helper
gate to match (verified with a riscv64 rv64gcv cross-build of opencv_core -
the engine f64 paths now vectorize on RVV instead of not compiling).
cv::exp/cv::log: the depth local is consumed by CV_OCL_RUN only - CV_UNUSED
for OpenCL-less builds (ARM64 -Wunused-variable).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* silence the remaining ARM64 gcc warnings
- compare boundary-rewrite: {}-init the fixed kind/bound arrays (filled for
every channel used below, but gcc cannot prove it across the cn <= 4 loop);
- cv::exp/log: [[maybe_unused]] on the depth local (consumed by CV_OCL_RUN
only), instead of the CV_UNUSED idiom;
- AutoBuffer::reserve: a targeted -Wmaybe-uninitialized suppression around
the live-element copy loop - only [0, sz) is read, all written before, but
gcc inlining a grow-from-inline-storage chain cannot see that. An
annotation for the analyzer, no behavior change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* saturating 32-bit add/sub kernels; cv::texpr python binding; two CI warnings
- v_add_sat/v_sub_sat for v_int32/v_uint32, local to arithm.simd.hpp for now
(the plan is to grow them into proper universal intrinsics later): NEON
single-instruction vqadd/vqsub, elsewhere the Hacker Delight bit tricks
over universal intrinsics (u32 add is 2 ops: or with the wrapped-compare
mask). EwAdd/EwSub overload vec() for the 32-bit lanes and getAddSubFunc
routes 32S/32U T->T through vecBinaryKernel instead of the former pure
scalar kernel. Semantics unchanged - the scalar int64 tail already
saturated; directed boundary tests added (both rails, 0 - INT_MIN, u32
cases, a full-range random block vs an exact int64 reference).
640x480 32S add: 0.48x of 5.x -> parity (memory-bound); 1080p: 6-8x.
- cv::texpr becomes CV_EXPORTS_W: python gets cv.texpr(expr, [inputs]) ->
tuple of ndarrays, so `res, = cv.texpr(...)` and `mag, ang = cv.texpr(...)`
unpacking both work. modules/python/test/test_expr.py covers arithmetic,
the fused abs(a-b), casts, broadcasting, ?: and ** operators, math
functions vs numpy, clamp, named temporaries, tuple outputs, the one-line
cartToPolar and the int32 saturation cases.
- warnings: {}-init the parser args array (gcc -Wmaybe-uninitialized on
Ubuntu 20/22); the compare short-row block gates sizeof(T) <= 4 as
constexpr so the f64 instantiation does not leave set-but-unused locals
(gcc 9 -Wunused-but-set-variable).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* arithm_op: direct-kernel fast path for small continuous arrays
Building and compiling the 1-instruction program plus the BroadcastOp setup
costs ~40-250ns per call - negligible on big images, dominant at 127x61-class
sizes where the classic 5.x functions were 1.3-2x faster. Mirror math_op two
tiers in arithm_op: two same-type same-shape continuous arrays, no mask, no
scalar, result depth == input depth, <= 100k elements -> call the T x T -> T
kernel directly over the flattened elements (checks ordered cheapest-first).
Applies to add/subtract/min/max/absdiff/multiply/addWeighted/and/or/xor;
compare and divide lower to more than a single kernel (boundary rewrites, int
guards) and keep the ordinary path. addWeighted falls through automatically
for the 32/64-bit int types whose lowering is wide-compute + cast
(getElemwiseFunc returns no direct kernel there).
127x61 vs 5.x, was -> now: add/subtract 8UC1 0.76x -> 1.3x, min/max u8
0.6x -> ~1x, addWeighted 1.0x -> 1.1-1.4x (32SC1 stays 4.5x); the one
remaining laggard is add/sub 32SC1 (0.74-0.82x) - the price of the new
SATURATING semantics (7-instruction AVX2 emulation vs the wrapping single
add of 5.x; single-instruction on NEON).
The 127x61 size is ADDED PERMANENTLY to the arithmetic/addWeighted/compare
perf grids: per-call overhead regressions in these base functions must be
caught by CI, not discovered by users.
dst creation goes through createSameSize (whole-shape transfer including
layout and future metadata, not piecemeal dims+sizes) here and in math_op.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* cv::pow rebuilt on the engine; integer-exponent and 1/sqrt(x) kernel branches
Routing: p = 0/1/2 keep their early special cases (fill/copy/multiply); an
INTEGER array with an INTEGER power keeps the classic iPow multiply chain -
bit-exact compatibility, including its wrap-around quirks (iPow squares in
int, so e.g. pow(255,4) on u8 wraps negative and saturates to 0 - somebody
may rely on that). Everything else goes through the engine with the math_op
two-tier scheme: small continuous arrays call powKernel directly (the
exponent rides as a broadcast T scalar), the rest run the tiled parallel
program. Integer arrays with fractional powers compute in the float domain
and saturate back; the 32U/64-bit depths (classic iPow asserted on them) and
f16/bf16 (the classic float path misread them) now just work.
powKernel gets two new per-row exponent branches:
- p == -0.5: 1/v_sqrt(x) (the classic path used IPP ippsInvSqrt_A21, a
21-bit approximation; ours is exact - slightly slower on small arrays,
4.5x faster at 1080p via parallelism);
- any other INTEGER p (|p| <= 65536): LSB-first binary exponentiation, the
same multiply chain and order as iPow, fully vectorized - a few ulp
accurate vs ~2e-7 of exp(p*log x), and exact on non-positive bases (the
sign falls out of the multiplies, 0^negative divides to inf) - no scalar
patching.
cv::pow f32 vs 5.x: p=0.5 365 -> 61 us at 1080p (6x), p=3 7.7x, p=5 7x AND
faster at every size (the old scalar chain: 0.81 -> 0.47 us at 127x61),
p=2.5 5.3x. The s16^5 iPow path is untouched (118.7 == 118.5 us).
pow_exponents accuracy tests extended to 11 exponents x f32/f64 against the
double std::pow reference.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* perf: SANITY_CHECK_NOTHING for the tests whose grids got the 127x61 size
The 127x61 entry added to guard per-call overhead has no regression data in
opencv_extra, so the legacy SANITY_CHECK in addWeighted/compare failed on CI
(locally it passes silently without the test-data path). Accuracy of both
functions is covered by the accuracy suite; the perf tests should measure
time. PatchNaNs/finiteMask keep their SANITY_CHECK - their grids are
untouched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix a temp-buffer double release in the liveness pass ("d*d" crash)
When the same temp is passed as SEVERAL arguments of its last-use
instruction (e.g. the named-intermediate expression "d = {0} - {1}; d*d",
where the MUL consumes slot d twice), the buffer-reuse scan pushed the
temp's physical buffer onto the free list once per argument. That
overflows the ntemps-sized freeBufs array (caught by the AutoBuffer range
check in Debug: python test_expr.py::test_named_temporary) and, in larger
programs, would hand the same physical buffer to two live temps.
Release the buffer once by retiring lastUse[t] after the first hit.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* executor: recalibrate opCost to vectorized cycles + clamp the stripe count
The per-element op costs fed into the parallel_for_ stripe hint were
scalar-era estimates (~20x above the vectorized reality: the atan2/exp
polynomials run at ~1.5 cycles/element, not 30). The hint therefore split
transcendental/divide work into hundreds of ~1 us jobs, which the
macOS/GCD backend dispatches poorly under sustained load, on top of the
P/E-core equal-share straggler effect. Measured on M4 Max (12P+4E),
sustained medians @1920x1080 f32: texpr atan2 320 -> 133 us, cv::exp
280 -> 141 us, cv::log 301 -> ~200 us, cv::pow(x,2.5) 388 -> 376 us;
the PR tables' math rows improved ~1.5-2x across the board.
- opCost is now in units of ~1/4 cycle/element of the SIMD kernels:
cheap ops 1 (unchanged), div/sqrt/hypot/convert_scale 10 -> 3,
transcendentals 30 -> 6.
- the stripe hint is clamped by min(4*nthreads, max(32, 3*nthreads)):
~4 stripes/thread is plenty of granularity for element-wise work, and
the ceiling is 32 pieces except on machines with many (heterogeneous)
cores, where anything coarser than ~3 pieces/worker turns the slow
cores into equal-share bottlenecks (measured: 32 stripes on 16 threads
is the worst point of the curve - 193 us vs 137 us at 48 for atan2).
getNumThreads() is clamped from below (WINRT/plugin backends may
report 0).
Not addressed here (needs cross-machine data, M2/M3 Ultra): streaming
memory-bound ops saturate the M4 Max fabric at ~8 fat stripes and E-core
participation only adds contention - a cost model cannot express that;
candidate follow-up is a bytes-aware clamp or a GCD-backend-level fix.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* multiply: v_mul_sat integer kernels (full product clamped to the type)
v_mul_sat(V, V) -> V for u8/s8/u16/s16/u32/s32 - the full-precision
product clamped to the lane type, which is exactly cv::multiply's integer
semantics at scale == 1. Local to arithm.simd.hpp for now, next to
v_add_sat/v_sub_sat, to be promoted into proper universal intrinsics
later. NEON: widening vmull + saturating narrow (vqmovn); other backends:
the portable v_mul_expand + saturating v_pack composition for 8/16-bit
lanes. 32-bit lanes have no universal widening multiply (no v_mul_expand
for s32), so the 32-bit integer fast path is NEON-only for now and the
other backends keep the previous f64 work-vector kernels (which measure
well on x86 with IPP-free AVX2).
EwMul::vec() now routes the integer lane types through v_mul_sat, and
getMulFunc_ passes the NATIVE lane vector as the scale==1 fast-path type:
whole-register loads/stores, the widening happens inside the multiply.
Replaces both the half-register widening loads (u8/s8/u16/s16) and the
scalar-equivalent f64 path for 32S/32U on NEON.
M4 Max, 640x480 (the sizes where the old kernels lost to carotene):
8U 22.0 -> 13.6 us, 8S 15.0 -> 11.1, 16S 21.2 -> 17.4, 32S 84.8 -> 30.5
(parity with the classic path everywhere, 32S was 0.38x). 1920x1080:
8S 1.30x -> 1.78x, 16S 2.57x -> 2.98x, 32S 2.20x -> 4.02x vs 5.x.
Correctness: exhaustive 8-bit (all 65536 pairs per sign), directed
saturation corners for 16/32-bit (46341^2, INT_MIN*-1, 65536*65536, all
sign combinations) against an exact int64 reference.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* NEON: make v_cvt_f64(v_int32) exact (was via f32, losing bits > 2^24)
The NEON implementations of v_cvt_f64/v_cvt_f64_high for v_int32 did
s32 -> f32 -> f64 (vcvt_f32_s32 + vcvt_f64_f32), silently rounding any
|x| > 2^24. Every vectorized f64 work path with int32 inputs on AArch64
was affected: the engine's addWeighted/divide 32S kernels, convertTo
32S -> 64F, etc. Found via addWeighted 32SC1 on values ~1e9: max error
was 32 vs the exact double reference (the classic carotene path is worse
still - it computes in f32 end-to-end with f32-truncated weights, max
error 96 on the same data).
The exact sequence sxtl + scvtf (vmovl_s32 + vcvtq_f64_s64) is the same
2 instructions, so there is no cost. addWeighted 32SC1 on the engine now
matches the exact-double reference bit-for-bit and stays at parity/1.4x
vs the classic path (640x480/1920x1080).
Pre-existing upstream bug (same code in 4.x) - worth a standalone
backport with directed large-value tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* math_op: drop the temporary direct-IPP tier for exp/log
Upstream moved the IPP math wrappers into the hal/ipp HAL module (cv_hal_exp32f/
log32f & co now resolve to ipp_hal_* which honor cv::ipp::useIPP via
CV_HAL_CHECK_USE_IPP). The engine's single probeHalUnary(cv_hal_*) probe already
picks that up uniformly, so the stopgap #ifdef HAVE_IPP ippExp/ippLog tier and its
ipp::useIPP() branch in getMathFunc are now redundant - removed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* arithm: restore cv::hal::and8u/or8u/xor8u/not8u as engine wrappers
These public CV_EXPORTS entry points (core/hal/hal.hpp) lost their definitions when
the bitwise ops moved to the element-wise engine, but they are still declared and called
by other modules (opencv_objdetect's aruco) and external code - the link broke with
undefined references to cv::hal::and8u/xor8u. Restore them as thin forwarders over the
engine's byte-wise bitwise kernels (getBitwiseFunc / getNotFunc), mirroring the existing
mul8u wrapper. CV_Assert guards the kernel lookup.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* blobdetector: guard empty contour when computing blob radius
The new AutoBuffer leaves its tail uninitialized for trivial types, which surfaced a
-Wmaybe-uninitialized in findBlobs where the median of per-point distances is read. Use a
std::vector, default the radius to 0, and compute the median only for a non-empty contour -
no unproven size invariant.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Replace "static inline" by "inline" in headers #29453
This fixes#29436
I also replaced CV_INLINE which can be removed in a later PR.
I can make a similar PR for 4.x if you want.
### 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
dnn: implement LpPool ONNX operator #29217
## Summary
Implements the `LpPool` ONNX operator (opset 1–18), which was previously unregistered and caused a parse failure. `LpPool` computes the Lp-norm pooling: `(sum(|x|^p))^(1/p)` over a sliding window.
## Changes
- New `LpPoolLayer` in `modules/dnn/src/layers/lppool_layer.cpp`
- Supports `kernel_shape`, `strides`, `dilations`, `pads`, `auto_pad` (NOTSET/SAME_UPPER), `ceil_mode`, and `p` (default 2)
- SIMD fast paths for p=1 (abs + accumulate) and p=2 (square + accumulate + sqrt); scalar fallback for other values of p
- Registered `LpPool` dispatch entry in both `onnx_importer.cpp` (classic engine) and `onnx_importer2.cpp` (new graph engine)
- Added `LpPoolLayer` declaration to `modules/dnn/include/opencv2/dnn/all_layers.hpp`
- Registered layer class in `modules/dnn/src/init.cpp`
- Re-enabled 8 lppool conformance tests in `test_onnx_conformance.cpp` (previously in parser denylist)
- `test_lppool_2d_same_lower` added to the global conformance denylist — same known SAME_LOWER padding bug that affects `averagepool` and `maxpool`
## Testing
All applicable ONNX conformance tests pass:
| Test | Result |
|------|--------|
| test_lppool_1d_default | PASSED |
| test_lppool_2d_default | PASSED |
| test_lppool_2d_dilations | PASSED |
| test_lppool_2d_pads | PASSED |
| test_lppool_2d_same_lower | SKIPPED (known SAME_LOWER padding bug, consistent with avgpool/maxpool) |
| test_lppool_2d_same_upper | PASSED |
| test_lppool_2d_strides | PASSED |
| test_lppool_3d_default | PASSED |
Tested on: macOS (x86_64/SSE4, Rosetta 2) and Linux x86_64 (AVX2/AVX-512, GCC 13.3.0), Release build
OpenCV version: 5.0.0-pre
## Related Issues
None
---
### 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
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.
fixed Dynamic quantized linear layer error #29386
### 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