Skip to content
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

dnn: add more CANN operators to support SAM #23550

Closed
wants to merge 14 commits into from
Closed

dnn: add more CANN operators to support SAM #23550

Changes from 1 commit
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
370b50c
support broadcast on axis > 1 for Expand
fengyuentau Apr 14, 2023
381d2a5
allow null constant_value in Pad and ignore it when loading
fengyuentau Apr 14, 2023
7cb5d46
add cann layernorm op
fengyuentau Apr 26, 2023
efa45b5
add gelu
fengyuentau Apr 27, 2023
d4b43b1
support pow
fengyuentau Apr 27, 2023
1fddf4a
add sqrt
fengyuentau Apr 27, 2023
1c070be
add reduce
fengyuentau Apr 28, 2023
a949861
make error msg informative with dtype and node name
fengyuentau Apr 28, 2023
4452498
fix empty bias_mat when the node type is matmul
fengyuentau Apr 28, 2023
854c696
mat type should not have channel for CannConstOp
fengyuentau Apr 28, 2023
11cfffd
fix number of Split outputs
fengyuentau Apr 28, 2023
0ae42db
support batched matmul
fengyuentau May 5, 2023
015c9c0
support 1D const; fix desc_gamma and desc_beta in layernorm; support …
fengyuentau May 5, 2023
f38c4e8
revert some out-of-date changes
fengyuentau Dec 23, 2023
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
Prev Previous commit
Next Next commit
support batched matmul
  • Loading branch information
@fengyuentau
fengyuentau committed May 5, 2023
commit 0ae42db4f468ce88f6589b07ea9616198a145e2f
109 changes: 75 additions & 34 deletions modules/dnn/src/layers/fully_connected_layer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -666,59 +666,100 @@ class FullyConnectedLayerImpl CV_FINAL : public InnerProductLayer
const std::vector<Ptr<BackendWrapper> > &outputs,
const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
{
CV_CheckGE(inputs.size(), (size_t)1, "DNN/CANN: InnerProduct should have at least one input");
auto x1 = inputs[0].dynamicCast<CannBackendWrapper>();
auto x1_desc = x1->getTensorDesc();
auto op_x1 = nodes[0].dynamicCast<CannBackendNode>()->getOp();
auto output_desc = std::make_shared<ge::TensorDesc>(ge::Shape(), ge::FORMAT_NCHW, ge::DT_FLOAT);
auto dims_x1 = x1->host->dims;
int dims_x2;
if (!blobs.empty()) // operand B is a const
{
dims_x2 = weightsMat.dims;
}
else
{
CV_CheckGE(inputs.size(), (size_t)2, "DNN/CANN: InnerProduct should have two inputs");
auto x2 = inputs[1].dynamicCast<CannBackendWrapper>();
dims_x2 = x2->host->dims;
}
// support 2D & 3D MatMat for now
CV_CheckEQ(dims_x1, dims_x2, "DNN/CANN: InnerProduct should have inputs of the same dimension");
CV_CheckGE(dims_x1, 2, "DNN/CANN: InnerProduct should have inputs of at least 2 dimensions");
CV_CheckLE(dims_x1, 3, "DNN/CANN: InnerProduct should have inputs of at most 3 dimensions");
bool isBatchMatMul = dims_x1 == 3;

auto op = std::make_shared<ge::op::MatMulV2>(name);
// std::cout << cv::format("Node name: %s, isMatmul=%d, bias=%d, isBatchMatMul=%d\n", name.c_str(), isMatMul, bias, isBatchMatMul);
auto op_x1 = nodes[0].dynamicCast<CannBackendNode>()->getOp();
auto desc_x1 = x1->getTensorDesc();
std::shared_ptr<ge::Operator> op_x2;
std::shared_ptr<ge::TensorDesc> desc_x2;
bool trans_x1 = blobs.empty() ? transA : false;
bool trans_x2 = blobs.empty() ? transB : true;
if (!blobs.empty())
{
auto op_const_w = std::make_shared<CannConstOp>(weightsMat.data, weightsMat.type(), shape(weightsMat), cv::format("%s_w", name.c_str()));
op_x2 = op_const_w->getOp();
desc_x2 = op_const_w->getTensorDesc();
}
else
{
op_x2 = nodes[1].dynamicCast<CannBackendNode>()->getOp();
desc_x2 = inputs[1].dynamicCast<CannBackendWrapper>()->getTensorDesc();
}
auto bias_mat = bias ? biasMat : (blobs.empty() ? Mat::zeros(1, 1, CV_32F) : Mat::zeros(1, weightsMat.size[0], weightsMat.type()));
std::vector<int> bias_shape{weightsMat.size[0]};
auto op_const_bias = std::make_shared<CannConstOp>(bias_mat.data, bias_mat.type(), bias_shape, cv::format("%s_b", name.c_str()));
auto op_b = op_const_bias->getOp();
auto desc_b = op_const_bias->getTensorDesc();
auto desc_output = std::make_shared<ge::TensorDesc>(ge::Shape(), ge::FORMAT_NCHW, ge::DT_FLOAT);

if (!blobs.empty()) // if B is const
if (isBatchMatMul)
{
auto op = std::make_shared<ge::op::BatchMatMul>(name);

// set attributes
op->set_attr_transpose_x1(false);
op->set_attr_adj_x1(false);
// weightMat always needs to be transposed, since CPU backend
// implementation is input * weight.im2row
op->set_attr_transpose_x2(true);
op->set_attr_adj_x2(trans_x2);

// set inputs
// set inputs:
// set inputs : x1 (input)
op->set_input_x1_by_name(*op_x1, x1->name.c_str());
op->update_input_desc_x1(*desc_x1);
// set inputs : x2 (weight)
auto op_const_weight = std::make_shared<CannConstOp>(weightsMat.data, weightsMat.type(), shape(weightsMat), cv::format("%s_w", name.c_str()));
op->set_input_x2_by_name(*(op_const_weight->getOp()), "y");
op->update_input_desc_x2(*(op_const_weight->getTensorDesc()));
op->set_input_x2_by_name(*op_x2, "y");
op->update_input_desc_x2(*desc_x2);

// set outputs
op->update_output_desc_y(*desc_output);

return Ptr<BackendNode>(new CannBackendNode(op));
}
else
{
// A and B are variable inputs; non-const bias is not considered
CV_Assert(inputs.size() == 2);
CV_Assert(nodes.size() == 2);
auto op = std::make_shared<ge::op::MatMulV2>(name);

// set attributes
op->set_attr_transpose_x1(transA);
op->set_attr_transpose_x2(transB);
op->set_attr_transpose_x1(trans_x1);
// weightMat always needs to be transposed, since CPU backend
// implementation is input * weight.im2row
op->set_attr_transpose_x2(trans_x2);

// set inputs:
// set inputs : x1 (input)
op->set_input_x1_by_name(*op_x1, x1->name.c_str());
op->update_input_desc_x1(*desc_x1);
// set inputs : x2 (weight)
auto op_x2 = nodes[1].dynamicCast<CannBackendNode>()->getOp();
auto x2_desc = inputs[1].dynamicCast<CannBackendWrapper>()->getTensorDesc();
op->set_input_x2_by_name(*op_x2, "y");
op->update_input_desc_x2(*x2_desc);
}
op->update_input_desc_x2(*desc_x2);
// set inputs : bias
op->set_input_bias(*op_b);
op->update_input_desc_bias(*desc_b);

// set inputs
// set inputs : x1 (input)
op->set_input_x1_by_name(*op_x1, x1->name.c_str());
op->update_input_desc_x1(*x1_desc);
// set inputs : bias (bias)
auto bias_mat = bias ? biasMat : (blobs.empty() ? Mat::zeros(1, 1, CV_32F) : Mat::zeros(1, weightsMat.size[0], weightsMat.type()));
std::vector<int> bias_shape{weightsMat.size[0]};
auto op_const_bias = std::make_shared<CannConstOp>(bias_mat.data, bias_mat.type(), bias_shape, cv::format("%s_b", name.c_str()));
op->set_input_bias(*(op_const_bias->getOp()));
op->update_input_desc_bias(*(op_const_bias->getTensorDesc()));

// set outputs
op->update_output_desc_y(*output_desc);
// set outputs
op->update_output_desc_y(*desc_output);

return Ptr<BackendNode>(new CannBackendNode(op));
return Ptr<BackendNode>(new CannBackendNode(op));
}
}
#endif

Expand Down