Skip to content

[NPU] Fix the performance problem when 'axis' is not specified (#35116) #35301

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
wangxicoding merged 1 commit into from
Aug 31, 2021
Merged

[NPU] Fix the performance problem when 'axis' is not specified (#35116) #35301

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
21 changes: 8 additions & 13 deletions paddle/fluid/operators/elementwise/elementwise_add_op_npu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,27 +42,22 @@ class ElementwiseAddNPUKernel : public framework::OpKernel<T> {
auto y_dims = y->dims();
axis = (axis == -1 ? std::abs(x_dims.size() - y_dims.size()) : axis);
if (x_dims.size() >= y_dims.size()) {
direct_compute =
y_dims == framework::slice_ddim(x_dims, axis, x_dims.size());
direct_compute = x_dims.size() == (y_dims.size() + axis);
} else {
direct_compute =
x_dims == framework::slice_ddim(y_dims, axis, y_dims.size());
direct_compute = y_dims.size() == (x_dims.size() + axis);
}

Tensor transformed_x, transformed_y;
if (direct_compute) {
transformed_x.ShareDataWith(*x);
transformed_y.ShareDataWith(*y);
const auto& runner = NpuOpRunner("Add", {*x, *y}, {*out}, {});
runner.Run(dev_ctx.stream());
} else {
Tensor transformed_x, transformed_y;
NpuElementWiseOpBroadcast<T>(dev_ctx, x, y, axis, &transformed_x,
&transformed_y);
const auto& runner =
NpuOpRunner("Add", {transformed_x, transformed_y}, {*out}, {});
runner.Run(dev_ctx.stream());
}
const auto& runner =
NpuOpRunner("Add", {transformed_x, transformed_y}, {*out}, {});
auto stream =
ctx.template device_context<paddle::platform::NPUDeviceContext>()
.stream();
runner.Run(stream);
}
};

Expand Down