Merge pull request #25518 from alexlyulkov:al/fixed-gemm-openvino

alexlyulkov · web-flow · commit 03507e06b404 · 2024-05-14T17:41:19.000+03:00
Fixed OpenVINO gemm layer #25518 Fixed OpenVINO gemm layer The problem was that our layer didn't properly handle all the possible gemm options in OpenVINO mode Fixes #25472 ### 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
diff --git a/modules/dnn/src/layers/gemm_layer.cpp b/modules/dnn/src/layers/gemm_layer.cpp
@@ -289,48 +289,54 @@ class GemmLayerImpl CV_FINAL : public GemmLayer {
     virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs,
                                         const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
     {
-        auto ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
-        std::shared_ptr<ov::Node> matmul;
+        ov::Output<ov::Node> nodeA = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
+        ov::Output<ov::Node> nodeB;
+        if (const_B)
+            nodeB = std::make_shared<ov::op::v0::Constant>(ov::element::f32, getShape(blobs[0]), blobs[0].data);
+        else
+            nodeB = nodes[1].dynamicCast<InfEngineNgraphNode>()->node;
+
+        int flatten_axis = nodeA.get_shape().size() - nodeB.get_shape().size();
+        if (flatten_axis > 0) {
+            std::vector<int> shape(1 + flatten_axis, 0);
+            shape[shape.size() - 1] = -1;
+            nodeA = std::make_shared<ov::op::v1::Reshape>(
+                nodeA,
+                std::make_shared<ov::op::v0::Constant>(ov::element::i32, ov::Shape{shape.size()}, shape.data()),
+                true);
+        }
 
-        if (nodes.size() == 2)
+        std::shared_ptr<ov::Node> nodeAB = std::make_shared<ov::op::v0::MatMul>(nodeA, nodeB, trans_a, trans_b);
+        if (alpha != 1.0f)
         {
-            auto& inp2 = nodes[1].dynamicCast<InfEngineNgraphNode>()->node;
-            matmul = std::make_shared<ov::op::v0::MatMul>(ieInpNode, inp2, trans_a, trans_b);
+            nodeAB = std::make_shared<ov::op::v1::Multiply>(
+                nodeAB,
+                std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1}, &alpha));
         }
-        else
+
+        if (!have_bias)
+            return Ptr<BackendNode>(new InfEngineNgraphNode(nodeAB));
+
+        ov::Output<ov::Node> nodeC;
+        if (const_C)
         {
-            std::shared_ptr<ov::Node> ieWeights = std::make_shared<ov::op::v0::Constant>(ov::element::f32, getShape(blobs[0]), blobs[0].data);
-
-            int flatten_axis = ieInpNode.get_shape().size() - ieWeights->get_shape().size();
-            if (flatten_axis > 0) {
-                std::vector<int> shape(1 + flatten_axis, 0);
-                shape[shape.size() - 1] = -1;
-                ieInpNode = std::make_shared<ov::op::v1::Reshape>(
-                    ieInpNode,
-                    std::make_shared<ov::op::v0::Constant>(ov::element::i32, ov::Shape{shape.size()}, shape.data()),
-                    true
-                );
-            }
-            matmul = std::make_shared<ov::op::v0::MatMul>(ieInpNode, ieWeights, trans_a, trans_b);
+            auto shape_C = blobs.back().total() == blobs.back().size[0] ? ov::Shape{blobs.back().total()} : getShape(blobs.back());
+            nodeC = std::make_shared<ov::op::v0::Constant>(ov::element::f32, shape_C, blobs.back().data);
         }
-        if (alpha != 1.0f) {
-            matmul = std::make_shared<ov::op::v1::Multiply>(matmul,
-                std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1}, &alpha)
-            );
+        else
+        {
+            nodeC = nodes.back().dynamicCast<InfEngineNgraphNode>()->node;
         }
 
-        if (have_bias && const_C) {
-            Mat bias = blobs.back();
-            auto shape = bias.total() == bias.size[0] ? ov::Shape{bias.total()} : getShape(bias);
-            std::shared_ptr<ov::Node> bias_node = std::make_shared<ov::op::v0::Constant>(ov::element::f32, shape, bias.data);
-            if (beta != 1.0f) {
-                bias_node = std::make_shared<ov::op::v1::Multiply>(bias_node,
-                    std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1}, &beta)
-                );
-            }
-            matmul = std::make_shared<ov::op::v1::Add>(matmul, bias_node, ov::op::AutoBroadcastType::NUMPY);
+        if (beta != 1.0f)
+        {
+            nodeC = std::make_shared<ov::op::v1::Multiply>(
+                nodeC,
+                std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1}, &beta));
         }
-        return Ptr<BackendNode>(new InfEngineNgraphNode(matmul));
+
+        auto nodeGemm = std::make_shared<ov::op::v1::Add>(nodeAB, nodeC, ov::op::AutoBroadcastType::NUMPY);
+        return Ptr<BackendNode>(new InfEngineNgraphNode(nodeGemm));
     }
 #endif // HAVE_DNN_NGRAPH
 
diff --git a/modules/dnn/test/test_onnx_conformance_layer_parser_denylist.inl.hpp b/modules/dnn/test/test_onnx_conformance_layer_parser_denylist.inl.hpp
@@ -121,8 +121,6 @@
 "test_gemm_all_attributes",
 "test_gemm_alpha",
 "test_gemm_beta",
-"test_gemm_default_matrix_bias",
-"test_gemm_default_no_bias",
 "test_gemm_default_scalar_bias",
 "test_gemm_default_single_elem_vector_bias",
 "test_gemm_default_vector_bias",
