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

Jump to bottom

speed_up for deformable conv #46997

Merged
merged 2 commits into from
Oct 14, 2022
Merged

speed_up for deformable conv #46997

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8f3c955
speed_up for deformable conv
Rayman96 Oct 13, 2022
c08cdfe
codestyle
Rayman96 Oct 13, 2022
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
44 changes: 39 additions & 5 deletions paddle/phi/kernels/impl/deformable_conv_kernel_impl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@
#include "paddle/phi/kernels/empty_kernel.h"
#include "paddle/phi/kernels/funcs/blas/blas.h"
#include "paddle/phi/kernels/funcs/deformable_conv_functor.h"
#include "paddle/phi/kernels/transpose_kernel.h"
#include "paddle/utils/optional.h"

namespace phi {
Expand All @@ -38,6 +39,11 @@ void DeformableConvKernel(const Context& dev_ctx,
DenseTensor* out) {
const int batch_size = static_cast<int>(x.dims()[0]);

int temp_step = std::min(64, batch_size);
if (batch_size % temp_step == 0) {
im2col_step = temp_step;
}

std::vector<int64_t> filter_shape_vec(phi::vectorize(filter.dims()));
std::vector<int64_t> output_shape_vec(phi::vectorize(out->dims()));

Expand Down Expand Up @@ -101,17 +107,23 @@ void DeformableConvKernel(const Context& dev_ctx,
dilations,
deformable_groups,
col_buffer_ptr);
DenseTensor output_3d = output_4d.Slice(i, i + 1).Resize(
phi::slice_ddim(output_4d.dims(), 1, output_4d.dims().size()));
DenseTensor output_3d = output_4d.Slice(i, i + 1).Resize(phi::slice_ddim(
output_4d.dims(),
1,
output_4d.dims().size())); // group * C/group * (im2step * H * W)

// get the product of pixel and weight
for (int g = 0; g < groups; ++g) {
DenseTensor weight_3d_slice = weight_3d.Slice(g, g + 1).Resize(
phi::slice_ddim(weight_3d.dims(), 1, weight_3d.dims().size()));
DenseTensor col_buffer_3d_slice =
col_buffer_3d.Slice(g, g + 1).Resize(phi::slice_ddim(
col_buffer_3d.dims(), 1, col_buffer_3d.dims().size()));
DenseTensor output_3d_slice = output_3d.Slice(g, g + 1).Resize(
phi::slice_ddim(output_3d.dims(), 1, output_3d.dims().size()));
DenseTensor output_3d_slice =
output_3d.Slice(g, g + 1).Resize(phi::slice_ddim(
output_3d.dims(),
1,
output_3d.dims().size())); // C * ((im2col_step)*H*W))
blas.MatMul(weight_3d_slice,
false,
col_buffer_3d_slice,
Expand All @@ -121,7 +133,29 @@ void DeformableConvKernel(const Context& dev_ctx,
T(0.0));
}
}
out->ShareDataWith(output_buffer).Resize(phi::make_ddim(output_shape_vec));

// swap axis to get the right result when im2col_step is greater than 1
if (im2col_step > 1) {
std::vector<int> axis(4);
axis[0] = 0;
axis[1] = 2;
axis[2] = 1;
axis[3] = 3;

DenseTensor real_output_buffer = phi::Transpose<T, Context>(
dev_ctx,
output_4d.Resize(
phi::make_ddim({batch_size / im2col_step,
output_shape_vec[1],
im2col_step,
output_shape_vec[2] * output_shape_vec[3]})),
axis);

out->ShareDataWith(real_output_buffer)
.Resize(phi::make_ddim(output_shape_vec));
} else {
out->ShareDataWith(output_buffer).Resize(phi::make_ddim(output_shape_vec));
}
}

} // namespace phi