Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Fixed OpenVINO gemm layer #25518

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
May 14, 2024
+40 −36
Merged

Fixed OpenVINO gemm layer #25518

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
bc05b1c
Fixed OpenVINO gemm layer
May 1, 2024
9a477eb
Enabled more gemm conformance tests
May 3, 2024
f9852fd
empty commit
May 3, 2024
7b25f39
empty commit
May 3, 2024
c2b8084
Fixed gemm layer
May 8, 2024
9ae67a8
Code review fixes.
asmorkalov May 13, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
74 changes: 40 additions & 34 deletions modules/dnn/src/layers/gemm_layer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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

Expand Down
2 changes: 0 additions & 2 deletions modules/dnn/test/test_onnx_conformance_layer_parser_denylist.inl.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -121,8 +121,6 @@
"test_gemm_all_attributes",
"test_gemm_alpha",
"test_gemm_beta",
"test_gemm_default_matrix_bias",
"test_gemm_default_no_bias",
Comment on lines -124 to -125
Copy link
Contributor

@asmorkalov asmorkalov May 13, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please remove.

"test_gemm_default_scalar_bias",
"test_gemm_default_single_elem_vector_bias",
"test_gemm_default_vector_bias",
Expand Down