wehub-resource-sync
89 lines
3.0 KiB
Plaintext
89 lines
3.0 KiB
Plaintext
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
|
|
//
|
|
// Licensed under the Apache License, Version 2.0 (the "License");
|
|
// you may not use this file except in compliance with the License.
|
|
// You may obtain a copy of the License at
|
|
//
|
|
// http://www.apache.org/licenses/LICENSE-2.0
|
|
//
|
|
// Unless required by applicable law or agreed to in writing, software
|
|
// distributed under the License is distributed on an "AS IS" BASIS,
|
|
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
// See the License for the specific language governing permissions and
|
|
// limitations under the License.
|
|
|
|
#include "paddle/phi/kernels/where_grad_kernel.h"
|
|
|
|
#include "paddle/phi/backends/gpu/gpu_launch_config.h"
|
|
#include "paddle/phi/core/kernel_registry.h"
|
|
#include "paddle/phi/kernels/full_kernel.h"
|
|
namespace phi {
|
|
|
|
template <typename T, typename IndexT>
|
|
__global__ void WhereGradCUDAKernel(
|
|
const IndexT N, const T* dout, const bool* cond, T* dx, T* dy) {
|
|
CUDA_KERNEL_LOOP_TYPE(idx, N, IndexT) {
|
|
if (dx != nullptr) {
|
|
dx[idx] = cond[idx] ? dout[idx] : static_cast<T>(0.);
|
|
}
|
|
if (dy != nullptr) {
|
|
dy[idx] = cond[idx] ? static_cast<T>(0.) : dout[idx];
|
|
}
|
|
}
|
|
}
|
|
|
|
template <typename T, typename Context>
|
|
void WhereGradKernel(const Context& dev_ctx,
|
|
const DenseTensor& condition,
|
|
const DenseTensor& x,
|
|
const DenseTensor& y,
|
|
const DenseTensor& out_grad,
|
|
DenseTensor* x_grad,
|
|
DenseTensor* y_grad) {
|
|
const bool* cond_data = condition.data<bool>();
|
|
auto numel = condition.numel();
|
|
auto* dout = out_grad.data<T>();
|
|
if (out_grad.numel() == 0) {
|
|
if (x_grad) {
|
|
Full<T, Context>(dev_ctx, x_grad->dims(), static_cast<T>(0), x_grad);
|
|
}
|
|
if (y_grad) {
|
|
Full<T, Context>(dev_ctx, y_grad->dims(), static_cast<T>(0), y_grad);
|
|
}
|
|
return;
|
|
}
|
|
T* dx = (x_grad != nullptr) ? dev_ctx.template Alloc<T>(x_grad) : nullptr;
|
|
T* dy = (y_grad != nullptr) ? dev_ctx.template Alloc<T>(y_grad) : nullptr;
|
|
|
|
auto stream = dev_ctx.stream();
|
|
auto config = backends::gpu::GetGpuLaunchConfig1D(dev_ctx, numel);
|
|
if (numel <= std::numeric_limits<int>::max()) {
|
|
WhereGradCUDAKernel<T, int>
|
|
<<<config.block_per_grid.x, config.thread_per_block.x, 0, stream>>>(
|
|
numel, dout, cond_data, dx, dy);
|
|
} else {
|
|
WhereGradCUDAKernel<T, int64_t>
|
|
<<<config.block_per_grid.x, config.thread_per_block.x, 0, stream>>>(
|
|
numel, dout, cond_data, dx, dy);
|
|
}
|
|
}
|
|
|
|
} // namespace phi
|
|
|
|
PD_REGISTER_KERNEL(where_grad,
|
|
GPU,
|
|
ALL_LAYOUT,
|
|
phi::WhereGradKernel,
|
|
bool,
|
|
float,
|
|
double,
|
|
int,
|
|
int8_t,
|
|
int64_t,
|
|
int16_t,
|
|
uint8_t,
|
|
phi::float16,
|
|
phi::bfloat16,
|
|
phi::complex64,
|
|
phi::complex128) {}
|