potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 07:43:03 +04:00
Code Packages Releases Wiki Activity

Merge pull request #24694 from fengyuentau:matmul_refactor

Browse Source 
dnn: refactor ONNX MatMul with fastGemm #24694

Done:
- [x] add backends
    - [x] CUDA
    - [x] OpenVINO
    - [x] CANN
    - [x] OpenCL
    - [x] Vulkan
- [x] add perf tests
- [x] const B case

### Benchmark

Tests are done on M1. All data is in milliseconds (ms).

| Configuration | MatMul (Prepacked) | MatMul | InnerProduct |
| - | - | - | - |
| A=[12, 197, 197], B=[12, 197, 64], trans_a=0, trans_b=0 | **0.39** | 0.41 | 1.33 |
| A=[12, 197, 64], B=[12, 64, 197], trans_a=0, trans_b=0  | **0.42** | 0.42 | 1.17 |
| A=[12, 50, 64], B=[12, 64, 50], trans_a=0, trans_b=0    | **0.13** | 0.15 | 0.33 |
| A=[12, 50, 50], B=[12, 50, 64], trans_a=0, trans_b=0    | **0.11** | 0.13 | 0.22 |
| A=[16, 197, 197], B=[16, 197, 64], trans_a=0, trans_b=0 | **0.46** | 0.54 | 1.46 |
| A=[16, 197, 64], B=[16, 64, 197], trans_a=0, trans_b=0  | **0.46** | 0.95 | 1.74 |
| A=[16, 50, 64], B=[16, 64, 50], trans_a=0, trans_b=0    | **0.18** | 0.32 | 0.43 |
| A=[16, 50, 50], B=[16, 50, 64], trans_a=0, trans_b=0    | **0.15** | 0.25 | 0.25 |

### Pull Request Readiness Checklist

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
      Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
This commit is contained in:
Yuantao Feng
2023-12-20 00:36:41 +08:00
committed by GitHub
parent 465e601e10
commit fa5ed62a66
13 changed files with 1340 additions and 107 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/perf/perf_gemm.cpp
+168 -2
View File
@@ -5,6 +5,8 @@
#include "perf_precomp.hpp"
#include <opencv2/dnn/shape_utils.hpp>
#include <numeric>
namespace opencv_test {
struct GemmParam_t {
@@ -71,6 +73,18 @@ static const GemmParam_t test_gemm_configs[] = {
*/
};
static const GemmParam_t test_matmul_configs[] = {
// vision transformer cases
{ {12, 197, 197}, {12, 197, 64} },
{ {12, 197, 64 }, {12, 64, 197} },
{ {12, 50, 64}, {12, 64, 50} },
{ {12, 50, 50}, {12, 50, 64} },
{ {16, 197, 197}, {16, 197, 64} },
{ {16, 197, 64 }, {16, 64, 197} },
{ {16, 50, 64}, {16, 64, 50} },
{ {16, 50, 50}, {16, 50, 64} },
};
struct GemmParamId
{
enum {
@@ -88,6 +102,21 @@ struct GemmParamId
}
};
struct MatMulParamId {
enum {
MATMUL_0 = 0,
MATMUL_LAST = sizeof(test_matmul_configs) / sizeof(test_matmul_configs[0])
};
int val_;
MatMulParamId(int val = 0) : val_(val) {}
operator int() const { return val_; }
static ::testing::internal::ParamGenerator<MatMulParamId> all() {
enum { NUM = (int)MATMUL_LAST };
MatMulParamId v_[NUM]; for (int i = 0; i < NUM; i++) { v_[i] = MatMulParamId(i); }
return ::testing::ValuesIn(v_, v_ + NUM);
}
};
static inline void PrintTo(const GemmParamId& v, std::ostream* os)
{
CV_Assert((int)v >= 0); CV_Assert((int)v < GemmParamId::GEMM_LAST);
@@ -138,7 +167,7 @@ PERF_TEST_P_(Gemm, gemm)
Mat A(static_cast<int>(a_shape.size()), a_shape.data(), CV_32F);
randu(A, -1.0f, 1.0f);
Mat B(static_cast<int>(b_shape.size()), b_shape.data(), CV_32F);
randu(A, -1.0f, 1.0f);
randu(B, -1.0f, 1.0f);
LayerParams lp;
lp.type = "Gemm";
@@ -197,7 +226,7 @@ PERF_TEST_P_(Gemm, innerproduct)
Mat A(static_cast<int>(a_shape.size()), a_shape.data(), CV_32F);
randu(A, -1.0f, 1.0f);
Mat B(static_cast<int>(b_shape.size()), b_shape.data(), CV_32F);
randu(A, -1.0f, 1.0f);
randu(B, -1.0f, 1.0f);
LayerParams lp;
lp.type = "InnerProduct";
@@ -241,9 +270,146 @@ PERF_TEST_P_(Gemm, innerproduct)
SANITY_CHECK_NOTHING();
}
static inline void PrintTo(const MatMulParamId& v, std::ostream* os)
{
CV_Assert((int)v >= 0); CV_Assert((int)v < MatMulParamId::MATMUL_LAST);
const GemmParam_t& p = test_matmul_configs[(int)v];
auto print_shape = [os](const std::vector<int>& shape, const std::string tag) {
if (shape.empty()) {
return ;
}
*os << tag << "=[";
for (size_t i = 0; i < shape.size(); ++i) {
if (i == shape.size() - 1) {
*os << shape[i] << "]";
break;
}
*os << shape[i] << ", ";
}
};
print_shape(p.a_shape, "A");
print_shape(p.b_shape, ", B");
print_shape(p.c_shape, ", C");
*os << ", trans_a=" << p.trans_a << ", trans_b=" << p.trans_b;
}
using MatMulTestParam_t = tuple<MatMulParamId, tuple<Backend, Target>>;
using MatMul = TestBaseWithParam<MatMulTestParam_t>;
PERF_TEST_P_(MatMul, matmul)
{
int test_id = (int)get<0>(GetParam());
ASSERT_GE(test_id, 0); ASSERT_LT(test_id, MatMulParamId::MATMUL_LAST);
const GemmParam_t& params = test_matmul_configs[test_id];
auto a_shape = params.a_shape;
auto b_shape = params.b_shape;
auto trans_a = params.trans_a;
auto trans_b = params.trans_b;
float alpha = 1.f;
float beta = 1.f;
Backend backend_id = get<0>(get<1>(GetParam()));
Target target_id = get<1>(get<1>(GetParam()));
Mat A(a_shape, CV_32F);
randu(A, -1.0f, 1.0f);
Mat B(b_shape, CV_32F);
randu(B, -1.0f, 1.0f);
LayerParams lp;
lp.type = "MatMul";
lp.name = "testLayer";
lp.set("transA", trans_a);
lp.set("transB", trans_b);
lp.set("alpha", alpha);
lp.set("beta", beta);
lp.blobs.push_back(B);
Net net;
net.addLayerToPrev(lp.name, lp.type, lp);
net.setPreferableBackend(backend_id);
net.setPreferableTarget(target_id);
// warmup
{
std::vector<std::string> input_names{"A"};
net.setInputsNames(input_names);
net.setInput(A, input_names[0]);
Mat out = net.forward();
}
TEST_CYCLE()
{
Mat res = net.forward();
}
SANITY_CHECK_NOTHING();
}
PERF_TEST_P_(MatMul, innerproduct)
{
int test_id = (int)get<0>(GetParam());
ASSERT_GE(test_id, 0); ASSERT_LT(test_id, MatMulParamId::MATMUL_LAST);
const GemmParam_t& params = test_matmul_configs[test_id];
auto a_shape = params.a_shape;
auto b_shape = params.b_shape;
Backend backend_id = get<0>(get<1>(GetParam()));
Target target_id = get<1>(get<1>(GetParam()));
Mat A(a_shape, CV_32F);
randu(A, -1.0f, 1.0f);
Mat B(b_shape, CV_32F);
randu(B, -1.0f, 1.0f);
LayerParams lp;
lp.type = "InnerProduct";
lp.name = "testLayer";
lp.set("axis", (int)(a_shape.size() - 1));
lp.set("bias_term", false);
// pre-transpose
std::vector<int> order(b_shape.size());
std::iota(order.begin(), order.end(), 0);
std::swap(order.back(), order[b_shape.size() - 2]);
Mat B_transposed;
transposeND(B, order, B_transposed);
lp.blobs.push_back(B_transposed);
lp.set("num_output", int(B_transposed.total(0, b_shape.size() - 1)));
lp.set("is_matmul", true);
Net net;
net.addLayerToPrev(lp.name, lp.type, lp);
net.setPreferableBackend(backend_id);
net.setPreferableTarget(target_id);
// warmup
{
std::vector<std::string> input_names{"A"};
net.setInputsNames(input_names);
net.setInput(A, input_names[0]);
Mat out = net.forward();
}
TEST_CYCLE()
{
Mat res = net.forward();
}
SANITY_CHECK_NOTHING();
}
INSTANTIATE_TEST_CASE_P(/**/, Gemm, Combine(
GemmParamId::all(),
dnnBackendsAndTargets(false, false) // defined in ../test/test_common.hpp
));
INSTANTIATE_TEST_CASE_P(/**/, MatMul, Combine(
MatMulParamId::all(),
dnnBackendsAndTargets(false, false) // defined in ../test/test_common.hpp
));
} // namespace