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

Skip to content

dnn: fuse matmul and add #25271

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 15 commits into from
Mar 29, 2024
Merged

dnn: fuse matmul and add #25271

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
15 commits
Select commit Hold shift + click to select a range
d7a2dd2
initial implementation
fengyuentau Mar 25, 2024
8372d2d
fix attention subgraph
fengyuentau Mar 27, 2024
0092696
remove trailing whitespace
fengyuentau Mar 27, 2024
c484ff0
do not fuse when matmul has input A constant
fengyuentau Mar 27, 2024
b0b481d
fix logic ensuring input A is not constant
fengyuentau Mar 27, 2024
494c4f1
fix biased matmul subgraph recognition logic
fengyuentau Mar 27, 2024
05b889b
fix openvino backend
fengyuentau Mar 28, 2024
65bd9e6
fix cann backend
fengyuentau Mar 28, 2024
e598ee7
pre-broadcast constant bias; fix opencl, openvino and cann backend
fengyuentau Mar 28, 2024
6511004
fix cann backend part 2 - BatchMatMulV2 does not support already broa…
fengyuentau Mar 28, 2024
27fad3a
fix logic detecting whether broadcast is needed
fengyuentau Mar 28, 2024
e65a585
fix cuda backend
fengyuentau Mar 28, 2024
3f3df9f
remove trailing whitespace
fengyuentau Mar 28, 2024
fb97eff
fix vulkan backend
fengyuentau Mar 28, 2024
ee7e6c7
resolve comments
fengyuentau Mar 29, 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
26 changes: 22 additions & 4 deletions modules/dnn/src/cuda4dnn/primitives/matmul_broadcast.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,8 @@
#include "../csl/tensor.hpp"
#include "../csl/tensor_ops.hpp"

#include "../kernels/eltwise_ops.hpp" // for adding bias

#include <opencv2/core.hpp>

#include <utility>
Expand All @@ -23,7 +25,7 @@ namespace cv { namespace dnn { namespace cuda4dnn {
public:
using wrapper_type = GetCUDABackendWrapperType<T>;

MatMulBroadcastOp(csl::Stream stream_, csl::cublas::Handle handle, const Mat &B, bool _transA, bool _transB,
MatMulBroadcastOp(csl::Stream stream_, csl::cublas::Handle handle, const Mat &B, const Mat &bias, bool _transA, bool _transB,
const std::vector<size_t> &A_offsets_, const std::vector<size_t> &B_offsets_, std::vector<size_t> &C_offsets_,
size_t batch_)
: stream(std::move(stream_)), cublasHandle(std::move(handle)), A_offsets(A_offsets_), B_offsets(B_offsets_), C_offsets(C_offsets_), batch(batch_)
Expand All @@ -33,6 +35,11 @@ namespace cv { namespace dnn { namespace cuda4dnn {
csl::copyMatToTensor<T>(B, input_B_tensor, stream);
}

if (!bias.empty()) {
bias_tensor = csl::makeTensorHeader<T>(bias);
csl::copyMatToTensor<T>(bias, bias_tensor, stream);
}

transA = _transA;
transB = _transB;
}
Expand All @@ -42,9 +49,6 @@ namespace cv { namespace dnn { namespace cuda4dnn {
const std::vector<cv::Ptr<BackendWrapper>>& outputs,
csl::Workspace& workspace) override
{
CV_Assert(((inputs.size() == 2 && input_B_tensor.empty()) ||
(inputs.size() == 1 && !input_B_tensor.empty())) && outputs.size() == 1);

auto input_A_wrapper = inputs[0].dynamicCast<wrapper_type>();
auto input_A = input_A_wrapper->getView();

Expand All @@ -60,12 +64,26 @@ namespace cv { namespace dnn { namespace cuda4dnn {
auto output = output_wrapper->getSpan();

csl::tensor_ops::gemmBatched<T>(cublasHandle, batch, 0.f, output, C_offsets, 1.f, transA, input_A, A_offsets, transB, input_B, B_offsets);

// add bias if exists
if (!bias_tensor.empty() || inputs.size() >= 3) {
csl::TensorView<T> bias;
if (bias_tensor.empty()) {
auto bias_wrapper = inputs[2].dynamicCast<wrapper_type>();
bias = bias_wrapper->getView();
} else {
bias = csl::TensorView<T>(bias_tensor);
}

kernels::eltwise_sum_2<T>(stream, output, output, bias);
}
}

private:
csl::Stream stream;
csl::cublas::Handle cublasHandle;
csl::Tensor<T> input_B_tensor;
csl::Tensor<T> bias_tensor;
bool transA, transB;

std::vector<size_t> A_offsets;
Expand Down
Loading