From bc05b1c0709022e901c6fdb05fbe2cbdcd0861ba Mon Sep 17 00:00:00 2001
From: Alexander Lyulkov <alexander.lyulkov@opencv.ai>
Date: Wed, 1 May 2024 12:57:03 +0300
Subject: [PATCH 1/6] Fixed OpenVINO gemm layer

---
 modules/dnn/src/layers/gemm_layer.cpp | 85 ++++++++++++++++-----------
 1 file changed, 50 insertions(+), 35 deletions(-)

diff --git a/modules/dnn/src/layers/gemm_layer.cpp b/modules/dnn/src/layers/gemm_layer.cpp
index bbbb10d2dfb1..f770e4e8bda1 100644
--- a/modules/dnn/src/layers/gemm_layer.cpp
+++ b/modules/dnn/src/layers/gemm_layer.cpp
@@ -289,48 +289,63 @@ 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;
-
-        if (nodes.size() == 2)
-        {
-            auto& inp2 = nodes[1].dynamicCast<InfEngineNgraphNode>()->node;
-            matmul = std::make_shared<ov::op::v0::MatMul>(ieInpNode, inp2, trans_a, trans_b);
-        }
+        auto nodeA = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
+        std::shared_ptr<ov::Node> nodeAB;
+
+        if (const_B)
+            nodeAB = std::make_shared<ov::op::v0::MatMul>(
+                nodeA,
+                std::make_shared<ov::op::v0::Constant>(ov::element::f32, getShape(blobs[0]), blobs[0].data),
+                trans_a,
+                trans_b);
         else
+            nodeAB = std::make_shared<ov::op::v0::MatMul>(
+                nodeA,
+                nodes[1].dynamicCast<InfEngineNgraphNode>()->node,
+                trans_a,
+                trans_b);
+
+        if (alpha != 1.0f)
         {
-            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);
-        }
-        if (alpha != 1.0f) {
-            matmul = std::make_shared<ov::op::v1::Multiply>(matmul,
+            nodeAB = std::make_shared<ov::op::v1::Multiply>(
+                nodeAB,
                 std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1}, &alpha)
             );
         }
 
-        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 (!have_bias)
+            return Ptr<BackendNode>(new InfEngineNgraphNode(nodeAB));
+
+        std::shared_ptr<ov::Node> nodeGemm;
+        if (beta != 1.0f)
+        {
+            std::shared_ptr<ov::Node> nodeC;
+            if (const_C)
+                nodeC = std::make_shared<ov::op::v1::Multiply>(
+                    std::make_shared<ov::op::v0::Constant>(ov::element::f32, getShape(blobs.back()), blobs.back().data),
+                    std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1}, &beta));
+            else
+                nodeC = std::make_shared<ov::op::v1::Multiply>(
+                    nodes.back().dynamicCast<InfEngineNgraphNode>()->node,
+                    std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1}, &beta));
+
+            nodeGemm = std::make_shared<ov::op::v1::Add>(nodeAB, nodeC, ov::op::AutoBroadcastType::NUMPY);
+        }
+        else
+        {
+            if (const_C)
+                nodeGemm = std::make_shared<ov::op::v1::Add>(
+                    nodeAB,
+                    std::make_shared<ov::op::v0::Constant>(ov::element::f32, getShape(blobs.back()), blobs.back().data),
+                    ov::op::AutoBroadcastType::NUMPY);
+            else
+                nodeGemm = std::make_shared<ov::op::v1::Add>(
+                    nodeAB,
+                    nodes.back().dynamicCast<InfEngineNgraphNode>()->node,
+                    ov::op::AutoBroadcastType::NUMPY);
         }
-        return Ptr<BackendNode>(new InfEngineNgraphNode(matmul));
+
+        return Ptr<BackendNode>(new InfEngineNgraphNode(nodeGemm));
     }
 #endif // HAVE_DNN_NGRAPH
 

From 9a477eb930d5b7bc7ff426599ea96ca19fe2e19e Mon Sep 17 00:00:00 2001
From: Alexander Lyulkov <alexander.lyulkov@opencv.ai>
Date: Fri, 3 May 2024 10:09:07 +0300
Subject: [PATCH 2/6] Enabled more gemm conformance tests

---
 .../test/test_onnx_conformance_layer_parser_denylist.inl.hpp  | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

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
index 68f49e5fa4d1..76ed6c323b88 100644
--- 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 +121,8 @@
 "test_gemm_all_attributes",
 "test_gemm_alpha",
 "test_gemm_beta",
-"test_gemm_default_matrix_bias",
-"test_gemm_default_no_bias",
+// "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",

From f9852fdad2348d9e9d85630153e5daa4ffa4517b Mon Sep 17 00:00:00 2001
From: Alexander Lyulkov <alexander.lyulkov@opencv.ai>
Date: Fri, 3 May 2024 13:15:16 +0300
Subject: [PATCH 3/6] empty commit

---
 modules/dnn/src/layers/gemm_layer.cpp | 1 -
 1 file changed, 1 deletion(-)

diff --git a/modules/dnn/src/layers/gemm_layer.cpp b/modules/dnn/src/layers/gemm_layer.cpp
index f770e4e8bda1..7224e60e2fe9 100644
--- a/modules/dnn/src/layers/gemm_layer.cpp
+++ b/modules/dnn/src/layers/gemm_layer.cpp
@@ -315,7 +315,6 @@ class GemmLayerImpl CV_FINAL : public GemmLayer {
 
         if (!have_bias)
             return Ptr<BackendNode>(new InfEngineNgraphNode(nodeAB));
-
         std::shared_ptr<ov::Node> nodeGemm;
         if (beta != 1.0f)
         {

From 7b25f3995e22f385d809d66533dfab7dfde95c8c Mon Sep 17 00:00:00 2001
From: Alexander Lyulkov <alexander.lyulkov@opencv.ai>
Date: Fri, 3 May 2024 13:15:29 +0300
Subject: [PATCH 4/6] empty commit

---
 modules/dnn/src/layers/gemm_layer.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/modules/dnn/src/layers/gemm_layer.cpp b/modules/dnn/src/layers/gemm_layer.cpp
index 7224e60e2fe9..f770e4e8bda1 100644
--- a/modules/dnn/src/layers/gemm_layer.cpp
+++ b/modules/dnn/src/layers/gemm_layer.cpp
@@ -315,6 +315,7 @@ class GemmLayerImpl CV_FINAL : public GemmLayer {
 
         if (!have_bias)
             return Ptr<BackendNode>(new InfEngineNgraphNode(nodeAB));
+
         std::shared_ptr<ov::Node> nodeGemm;
         if (beta != 1.0f)
         {

From c2b8084ec5df903e873f7a0eb6567168f46a32ac Mon Sep 17 00:00:00 2001
From: Alexander Lyulkov <alexander.lyulkov@opencv.ai>
Date: Wed, 8 May 2024 16:50:50 +0300
Subject: [PATCH 5/6] Fixed gemm layer

---
 modules/dnn/src/layers/gemm_layer.cpp | 65 ++++++++++++---------------
 1 file changed, 28 insertions(+), 37 deletions(-)

diff --git a/modules/dnn/src/layers/gemm_layer.cpp b/modules/dnn/src/layers/gemm_layer.cpp
index f770e4e8bda1..ac0914c2c210 100644
--- a/modules/dnn/src/layers/gemm_layer.cpp
+++ b/modules/dnn/src/layers/gemm_layer.cpp
@@ -289,62 +289,53 @@ 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 nodeA = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
-        std::shared_ptr<ov::Node> nodeAB;
-
+        ov::Output<ov::Node> nodeA = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
+        ov::Output<ov::Node> nodeB;
         if (const_B)
-            nodeAB = std::make_shared<ov::op::v0::MatMul>(
-                nodeA,
-                std::make_shared<ov::op::v0::Constant>(ov::element::f32, getShape(blobs[0]), blobs[0].data),
-                trans_a,
-                trans_b);
+            nodeB = std::make_shared<ov::op::v0::Constant>(ov::element::f32, getShape(blobs[0]), blobs[0].data);
         else
-            nodeAB = std::make_shared<ov::op::v0::MatMul>(
+            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,
-                nodes[1].dynamicCast<InfEngineNgraphNode>()->node,
-                trans_a,
-                trans_b);
+                std::make_shared<ov::op::v0::Constant>(ov::element::i32, ov::Shape{shape.size()}, shape.data()),
+                true);
+        }
 
+        std::shared_ptr<ov::Node> nodeAB = std::make_shared<ov::op::v0::MatMul>(nodeA, nodeB, trans_a, trans_b);
         if (alpha != 1.0f)
         {
             nodeAB = std::make_shared<ov::op::v1::Multiply>(
                 nodeAB,
-                std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1}, &alpha)
-            );
+                std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1}, &alpha));
         }
 
         if (!have_bias)
             return Ptr<BackendNode>(new InfEngineNgraphNode(nodeAB));
 
-        std::shared_ptr<ov::Node> nodeGemm;
-        if (beta != 1.0f)
+        ov::Output<ov::Node> nodeC;
+        if (const_C)
         {
-            std::shared_ptr<ov::Node> nodeC;
-            if (const_C)
-                nodeC = std::make_shared<ov::op::v1::Multiply>(
-                    std::make_shared<ov::op::v0::Constant>(ov::element::f32, getShape(blobs.back()), blobs.back().data),
-                    std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1}, &beta));
-            else
-                nodeC = std::make_shared<ov::op::v1::Multiply>(
-                    nodes.back().dynamicCast<InfEngineNgraphNode>()->node,
-                    std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1}, &beta));
-
-            nodeGemm = std::make_shared<ov::op::v1::Add>(nodeAB, nodeC, ov::op::AutoBroadcastType::NUMPY);
+            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);
         }
         else
         {
-            if (const_C)
-                nodeGemm = std::make_shared<ov::op::v1::Add>(
-                    nodeAB,
-                    std::make_shared<ov::op::v0::Constant>(ov::element::f32, getShape(blobs.back()), blobs.back().data),
-                    ov::op::AutoBroadcastType::NUMPY);
-            else
-                nodeGemm = std::make_shared<ov::op::v1::Add>(
-                    nodeAB,
-                    nodes.back().dynamicCast<InfEngineNgraphNode>()->node,
-                    ov::op::AutoBroadcastType::NUMPY);
+            nodeC = nodes.back().dynamicCast<InfEngineNgraphNode>()->node;
+        }
+
+        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));
         }
 
+        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

From 9ae67a882866ebe939e90f1a76e724cbf8d2fc07 Mon Sep 17 00:00:00 2001
From: Alexander Smorkalov <alexander.smorkalov@xperience.ai>
Date: Mon, 13 May 2024 16:40:22 +0300
Subject: [PATCH 6/6] Code review fixes.

---
 .../test/test_onnx_conformance_layer_parser_denylist.inl.hpp    | 2 --
 1 file changed, 2 deletions(-)

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
index 76ed6c323b88..cb008e9670c4 100644
--- 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",