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

dnn: parameterize fastGemmThin accuracy test (TEST_P)

Browse Source
This commit is contained in:
Teddy-Yangjiale
2026-07-23 23:30:48 +08:00
parent 6b04d456b3
commit 1af832250b
1 changed files with 34 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/test/test_fast_gemm.cpp
+34 -42
View File
@@ -30,56 +30,48 @@ static Mat referenceMatMul(const Mat& A, const Mat& B, bool trans_a, float alpha
return expected;
}
// Exercises the fastGemmThin path (constant-B MatMul): M covers every register
// Exercises the fastGemmThin path (constant-B MatMul). M covers every register
// block width and the multi-block remainder, N covers full strips and partial
// column tails for any VLEN <= 512.
TEST(DNN_FastGemmThin, MatMulAccuracy)
typedef testing::TestWithParam<tuple<int, int, int, int, float>> DNN_FastGemmThin;
TEST_P(DNN_FastGemmThin, MatMulAccuracy)
{
static const int m_values[] = { 1, 2, 3, 4, 5 };
static const int k_values[] = { 1, 3, 4, 7, 16, 64 };
static const int n_values[] = { 16, 19, 64, 67, 80, 83 };
static const float alpha_values[] = { 1.f, -0.5f };
int M = get<0>(GetParam()), N = get<1>(GetParam()), K = get<2>(GetParam());
int trans_a = get<3>(GetParam());
float alpha = get<4>(GetParam());
RNG rng(0x5EED);
for (int M : m_values)
{
for (int N : n_values)
{
for (int K : k_values)
{
Mat B(K, N, CV_32F);
rng.fill(B, RNG::UNIFORM, -1.f, 1.f);
Mat B(K, N, CV_32F); rng.fill(B, RNG::UNIFORM, -1.f, 1.f);
Mat A(trans_a ? K : M, trans_a ? M : K, CV_32F);
rng.fill(A, RNG::UNIFORM, -1.f, 1.f);
for (int trans_a = 0; trans_a <= 1; trans_a++)
{
Mat A(trans_a ? K : M, trans_a ? M : K, CV_32F);
rng.fill(A, RNG::UNIFORM, -1.f, 1.f);
LayerParams lp;
lp.type = "MatMul";
lp.name = "thin_matmul";
lp.set("transA", trans_a != 0);
lp.set("transB", false);
lp.set("alpha", alpha);
lp.blobs.push_back(B);
for (float alpha : alpha_values)
{
LayerParams lp;
lp.type = "MatMul";
lp.name = "thin_matmul";
lp.set("transA", trans_a != 0);
lp.set("transB", false);
lp.set("alpha", alpha);
lp.blobs.push_back(B);
Net net;
net.addLayerToPrev(lp.name, lp.type, lp);
net.setInputsNames(std::vector<String>{ "A" });
net.setInput(A, "A");
Mat actual = net.forward();
Mat expected = referenceMatMul(A, B, trans_a != 0, alpha);
Net net;
net.addLayerToPrev(lp.name, lp.type, lp);
net.setInputsNames(std::vector<String>{ "A" });
net.setInput(A, "A");
Mat actual = net.forward();
Mat expected = referenceMatMul(A, B, trans_a != 0, alpha);
EXPECT_LE(cv::norm(expected, actual, NORM_INF), 2e-5f * std::max(K, 1))
<< "M=" << M << ", N=" << N << ", K=" << K
<< ", trans_a=" << trans_a << ", alpha=" << alpha;
}
}
}
}
}
EXPECT_LE(cv::norm(expected, actual, NORM_INF), 2e-5f * std::max(K, 1))
<< "M=" << M << ", N=" << N << ", K=" << K
<< ", trans_a=" << trans_a << ", alpha=" << alpha;
}
INSTANTIATE_TEST_CASE_P(/*nothing*/, DNN_FastGemmThin, Combine(
Values(1, 2, 3, 4, 5),
Values(16, 19, 64, 67, 80, 83),
Values(1, 3, 4, 7, 16, 64),
Values(0, 1),
Values(1.f, -0.5f)
));
}} // namespace