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

Flip Kernel Optimization #46119

Merged
merged 4 commits into from
Sep 20, 2022
Merged

Flip Kernel Optimization #46119

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
87824ce
Optimize flip kernel by eliminating H2D data transfer, test=develop
Courtesy-Xs Sep 14, 2022
a57ea9b
Optimize flip kernel by eliminating H2D data transfer, test=develop
Courtesy-Xs Sep 15, 2022
e8acfda
fix function name
Courtesy-Xs Sep 15, 2022
71e6bbe
fix extra flip_dims_v
Courtesy-Xs Sep 16, 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
66 changes: 35 additions & 31 deletions paddle/phi/kernels/gpu/flip_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,14 +16,15 @@
#include "paddle/fluid/memory/malloc.h"
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/backends/gpu/gpu_launch_config.h"
#include "paddle/phi/common/place.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/core/utils/array.h"

namespace phi {

template <typename T, size_t Rank>
__global__ void flip_cuda_kernel(const int N,
__global__ void flip_cuda_kernel(const int64_t N,
const T* in_data,
T* out_data,
phi::Array<int64_t, Rank> shape,
Expand Down Expand Up @@ -53,41 +54,44 @@ __global__ void flip_cuda_kernel(const int N,
}

template <typename T, typename Context, size_t N>
void launch_flip_cuda_kernel(const Context& dev_ctx,
const DenseTensor& x,
const std::vector<int>& axis,
DenseTensor* out) {
std::vector<int> flip_dims_v = axis;
void LaunchFlipCudaKernel(const Context& dev_ctx,
const DenseTensor& x,
const std::vector<int>& axis,
DenseTensor* out) {
auto* in_data = x.data<T>();
auto* out_data = dev_ctx.template Alloc<T>(out);

auto x_dims = x.dims();
const int total_dims = x_dims.size();
const int numel = x.numel();

int block_size = 512;
dim3 dim_block(block_size);
dim3 dim_grid((numel + block_size - 1) / block_size);

for (size_t i = 0; i < flip_dims_v.size(); ++i) {
if (flip_dims_v[i] < 0) {
flip_dims_v[i] += total_dims;
}
}

const int64_t numel = x.numel();
auto config = phi::backends::gpu::GetGpuLaunchConfig1D(dev_ctx, numel);
auto x_stride = phi::stride(x_dims);

phi::Array<int64_t, N> stride_a;
phi::Array<int64_t, N> shape_a;
phi::Array<int, N> flip_dims_a;
size_t flip_dims_size = flip_dims_v.size();
size_t flip_dims_size = axis.size();

for (size_t idx = 0; idx < N; ++idx) {
stride_a[idx] = x_stride[idx];
shape_a[idx] = x_dims[idx];
flip_dims_a[idx] = idx < flip_dims_size ? flip_dims_v[idx] : 0;
flip_dims_a[idx] = idx < flip_dims_size ? axis[idx] : 0;
}

for (size_t i = 0; i < flip_dims_a.size(); ++i) {
if (flip_dims_a[i] < 0) {
flip_dims_a[i] += total_dims;
}
}
flip_cuda_kernel<T, N><<<dim_grid, dim_block, 0, dev_ctx.stream()>>>(
numel, in_data, out_data, shape_a, stride_a, flip_dims_a, flip_dims_size);
flip_cuda_kernel<T, N>
<<<config.block_per_grid, config.thread_per_block, 0, dev_ctx.stream()>>>(
numel,
in_data,
out_data,
shape_a,
stride_a,
flip_dims_a,
flip_dims_size);
}

template <typename T, typename Context>
Expand All @@ -98,31 +102,31 @@ void FlipKernel(const Context& dev_ctx,
const size_t total_dims = x.dims().size();
switch (total_dims) {
case 1:
launch_flip_cuda_kernel<T, Context, 1>(dev_ctx, x, axis, out);
LaunchFlipCudaKernel<T, Context, 1>(dev_ctx, x, axis, out);
break;
case 2:
launch_flip_cuda_kernel<T, Context, 2>(dev_ctx, x, axis, out);
LaunchFlipCudaKernel<T, Context, 2>(dev_ctx, x, axis, out);
break;
case 3:
launch_flip_cuda_kernel<T, Context, 3>(dev_ctx, x, axis, out);
LaunchFlipCudaKernel<T, Context, 3>(dev_ctx, x, axis, out);
break;
case 4:
launch_flip_cuda_kernel<T, Context, 4>(dev_ctx, x, axis, out);
LaunchFlipCudaKernel<T, Context, 4>(dev_ctx, x, axis, out);
break;
case 5:
launch_flip_cuda_kernel<T, Context, 5>(dev_ctx, x, axis, out);
LaunchFlipCudaKernel<T, Context, 5>(dev_ctx, x, axis, out);
break;
case 6:
launch_flip_cuda_kernel<T, Context, 6>(dev_ctx, x, axis, out);
LaunchFlipCudaKernel<T, Context, 6>(dev_ctx, x, axis, out);
break;
case 7:
launch_flip_cuda_kernel<T, Context, 7>(dev_ctx, x, axis, out);
LaunchFlipCudaKernel<T, Context, 7>(dev_ctx, x, axis, out);
break;
case 8:
launch_flip_cuda_kernel<T, Context, 8>(dev_ctx, x, axis, out);
LaunchFlipCudaKernel<T, Context, 8>(dev_ctx, x, axis, out);
break;
case 9:
launch_flip_cuda_kernel<T, Context, 9>(dev_ctx, x, axis, out);
LaunchFlipCudaKernel<T, Context, 9>(dev_ctx, x, axis, out);
break;
default:
PADDLE_THROW(phi::errors::InvalidArgument(
Expand Down