Merge pull request #1202 from fengyuentau:dnn/fix_cuda_matmul_crash

opencv · Aug 16, 2024 · 042dd81 · 042dd81
2 parents 9d0d24d + c397223
commit 042dd81
Show file tree

Hide file tree

Showing 4 changed files with 10 additions and 0 deletions.
diff --git a/testdata/dnn/onnx/data/input_matmul_bcast.npy b/testdata/dnn/onnx/data/input_matmul_bcast.npy
diff --git a/testdata/dnn/onnx/data/output_matmul_bcast.npy b/testdata/dnn/onnx/data/output_matmul_bcast.npy
diff --git a/testdata/dnn/onnx/generate_onnx_models_with_onnxscript.py b/testdata/dnn/onnx/generate_onnx_models_with_onnxscript.py
@@ -380,3 +380,13 @@ def clip_div_shared_constant(x: ost.FLOAT[1, 8, 12, 10]) -> ost.FLOAT[1, 8, 12,
     clip = op.Clip(div, Constant_output_0, Constant_1_output_0)
     return clip
 make_model_and_data(clip_div_shared_constant, np.random.rand(1, 8, 12, 10).astype(np.float32))
+
+''' Subgraph [Input] -> MatMul<B> -> [Outpunt]
+'''
+
+B = np.random.randn(16, 8).astype(np.float32)
+
+@ost.script()
+def matmul_bcast(x: ost.FLOAT[64, 1, 16]) -> ost.FLOAT[64, 1, 8]:
+    return op.MatMul(x, op.Constant(value=onnx.numpy_helper.from_array(B)))
+make_model_and_data(matmul_bcast, np.random.randn(64, 1, 16).astype(np.float32))
diff --git a/testdata/dnn/onnx/models/matmul_bcast.onnx b/testdata/dnn/onnx/models/matmul_bcast.onnx