mirror of
https://github.com/opencv/opencv.git
synced 2026-11-21 00:35:35 -06:00
dnn: faster ConvTranspose2 on CPU - #30396 Merge pull request #30395 from uwezkhan:onnx-qconv-scale-bias-bound - With MLAS, every fp32 ConvTranspose2 runs as a packed GEMM per group, followed by a col2im that adds each kernel tap into the output. Threads own bands of output rows, so col2im needs no locking. When taps overlap, output channels are processed in chunks so the column buffer stays in cache. When kernel == stride, each output pixel gets one tap and value + bias is written directly. - conv2_deconv is now CPU-dispatched (AVX, AVX2, NEON) and uses four accumulators, so the FMAs no longer wait on each other. It is still used when MLAS is not available or the input is not fp32. - Attention accepts [..., S, C] and treats the leading dims as batch; the SAM3 decoder feeds it a 4D input and failed before. opencv_extra: https://github.com/opencv/opencv_extra/pull/1442 (test models for the single-tap path, channel chunks, and groups that start inside a channel block) Measured on an i9-14900KS, times in ms. **Whole decoder** | Model | Threads | 5.x | This PR | ORT 1.30 | |---|---:|---:|---:|---:| | SAM2 decoder | 1 | 68.5 | **36.8** | 36.5 | | SAM2 decoder | 8 | 14.0 | **9.7** | 8.2 | | SAM2 decoder | 32 | 12.8 | **10.5** | 9.8 | | SAM3 decoder | 1 | 87.9* | **51.1** | 47.1 | | SAM3 decoder | 8 | 19.0* | **14.1** | 12.6 | | SAM3 decoder | 32 | 16.4* | **14.4** | 13.0 | \* SAM3 does not run on 5.x; this column includes the Attention change from this PR. Best of several interleaved rounds, measured on the first commit. The col2im commit leaves the SAM ConvTranspose layers within a few percent. **ConvTranspose layer** | Layer | Threads | 5.x | This PR | ORT 1.30 | |---|---:|---:|---:|---:| | SAM2 256→64 ch, 2×2 s2, 64×64 | 1 | 24.0 | **3.35** | 3.54 | | | 8 | 3.30 | **0.59** | 0.92 | | | 32 | 1.53 | **0.40** | 0.95 | | SAM2 64→32 ch, 2×2 s2, 128×128 | 1 | 13.2 | **2.10** | 2.52 | | | 8 | 1.76 | **0.31** | 1.15 | | | 32 | 0.85 | **0.24** | 1.88 | | SAM3 256→64 ch, 2×2 s2, 72×72 | 1 | 27.4 | **4.25** | 4.50 | | | 8 | 3.85 | **0.59** | 1.16 | | | 32 | 1.88 | **0.45** | 1.22 | | SAM3 64→32 ch, 2×2 s2, 144×144 | 1 | 15.5 | **2.72** | 3.41 | | | 8 | 2.08 | **0.41** | 1.44 | | | 32 | 1.00 | **0.32** | 1.62 | | U-Net 256→128 ch, 4×4 s2 p1, 32×32 | 1 | 46.6 | **7.63** | 6.70 | | | 8 | 6.49 | **1.19** | 1.45 | | | 32 | 4.50 | **0.97** | 1.68 | | GAN 128→64 ch, 4×4 s2 p1, 64×64 | 1 | 50.1 | **8.57** | 7.51 | | | 8 | 7.01 | **1.37** | 2.22 | | | 32 | 4.28 | **1.16** | 2.53 | | 128→64 ch, 3×3 s2 p1 op1, 64×64 | 1 | 28.3 | **4.70** | 4.24 | | | 8 | 3.95 | **0.76** | 1.31 | | | 32 | 1.79 | **0.66** | 1.75 | | 64→64 ch, 3×3 s1 p1, 64×64 | 1 | 13.4 | **2.52** | 2.35 | | | 8 | 1.91 | **0.45** | 0.97 | | | 32 | 0.80 | **0.35** | 0.98 | | 128→128 ch, 4 groups, 4×4 s2 p1, 32×32 | 1 | 6.73 | **1.65** | 1.30 | | | 8 | 0.97 | **0.26** | 0.73 | | | 32 | 0.44 | **0.21** | 1.00 | | 64→64 ch, depthwise, 4×4 s2 p1, 64×64 | 1 | 30.1 | **7.54** | 1.54 | | | 8 | 4.21 | **2.51** | 1.89 | | | 32 | 1.92 | **1.96** | 2.77 | ### 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
49 lines
1.8 KiB
C++
49 lines
1.8 KiB
C++
// This file is part of OpenCV project.
|
|
// It is subject to the license terms in the LICENSE file found in the top-level directory
|
|
// of this distribution and at http://opencv.org/license.html.
|
|
// Copyright (C) 2026, BigVision LLC, all rights reserved.
|
|
// Third party copyrights are property of their respective owners.
|
|
|
|
// ConvTranspose2 (a layer of the new graph engine) on single-node ONNX models.
|
|
|
|
#include "perf_precomp.hpp"
|
|
|
|
namespace opencv_test {
|
|
|
|
// model name, input channels, input height and width
|
|
typedef TestBaseWithParam<tuple<std::string, int, int> > Layer_ConvTranspose2;
|
|
|
|
PERF_TEST_P_(Layer_ConvTranspose2, forward)
|
|
{
|
|
const std::string name = get<0>(GetParam());
|
|
const int channels = get<1>(GetParam()), size = get<2>(GetParam());
|
|
|
|
const std::string model = "dnn/onnx/models/convtranspose_perf_" + name + ".onnx";
|
|
Net net = readNetFromONNX(findDataFile(model, true), ENGINE_OPENCV);
|
|
Mat input({1, channels, size, size}, CV_32F);
|
|
randu(input, -1.f, 1.f);
|
|
net.setInput(input);
|
|
net.forward();
|
|
|
|
TEST_CYCLE()
|
|
{
|
|
Mat res = net.forward();
|
|
}
|
|
SANITY_CHECK_NOTHING();
|
|
}
|
|
|
|
INSTANTIATE_TEST_CASE_P(/**/, Layer_ConvTranspose2, testing::Values(
|
|
make_tuple(std::string("sam2_256_64"), 256, 64),
|
|
make_tuple(std::string("sam2_64_32"), 64, 128),
|
|
make_tuple(std::string("sam3_256_64"), 256, 72),
|
|
make_tuple(std::string("sam3_64_32"), 64, 144),
|
|
make_tuple(std::string("unet_256_128_k4s2p1"), 256, 32),
|
|
make_tuple(std::string("gan_128_64_k4s2p1"), 128, 64),
|
|
make_tuple(std::string("128_64_k3s2p1op1"), 128, 64),
|
|
make_tuple(std::string("64_64_k3s1p1"), 64, 64),
|
|
make_tuple(std::string("group4_128_128_k4s2p1"), 128, 32),
|
|
make_tuple(std::string("depthwise_64_k4s2p1"), 64, 64)
|
|
));
|
|
|
|
} // namespace opencv_test
|