Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 2b88057f 编写于 作者: L Li Min 提交者： GitHub
浏览文件
操作

Refactor dropout cuda impl for code reuse. (#35621)

上级 e26a2504
develop Ligoml-patch-1 ZHUI-patch-1 add_some_yaml_config cherry_undefined_var delete_disable_iterable_dataset_unittest delete_fix_undefined_var delete_revert-36057-dev/read_flags_in_ut dingjiaweiww-patch-1 disable_iterable_dataset_unittest dy2static enable_eager_model_test final_state_gen_python_c final_state_intermediate fix-numpy-issue fix_concat_slice fix_op_flops fix_rnn_docs fix_tensor_type fix_undefined_var incubate/infrt inplace_addto make_flag_adding_easier move_embedding_to_phi move_histogram_to_pten move_sgd_to_phi move_slice_to_pten move_temporal_shift_to_phi move_yolo_box_to_phi npu_fix_alloc preln_ernie prv-md-even-more prv-onednn-2.5 pten_tensor_refactor release/2.2 release/2.3 release/2.3-fc-ernie-fix release/2.4 revert-36057-dev/read_flags_in_ut revert-36201-refine_fast_threaded_ssa_graph_executor revert-36985-add_license revert-37318-refactor_dygraph_to_eager revert-37926-eager_coreops_500 revert-37956-revert-37727-pylayer_support_tuple revert-38100-mingdong revert-38301-allocation_rearrange_pr revert-38703-numpy_bf16_package_reupload revert-38732-remove_useless_header_in_elementwise_mul_grad revert-38959-Reduce_Grad revert-39143-adjust_empty revert-39227-move_trace_op_to_pten revert-39268-dev/remove_concat_fluid_kernel revert-40170-support_partial_grad revert-41056-revert-40727-move_some_activaion_to_phi revert-41065-revert-40993-mv_ele_floordiv_pow revert-41068-revert-40790-phi_new revert-41944-smaller_inference_api_test revert-42149-do-not-reset-default-stream-for-stream-safe-cuda-allocator revert-43155-fix_ut_tempfile revert-43882-revert-41944-smaller_inference_api_test revert-45808-phi/simplify_size_op revert-46827-deform_comment support_weight_transpose zhiqiu-patch-1 v2.4.0-rc0 v2.3.2 v2.3.1 v2.3.0 v2.3.0-rc0 v2.2.2 v2.2.1 v2.2.0 v2.2.0-rc0 v2.2.0-bak0
无相关合并请求
隐藏空白更改
内联 并排
Showing with 336 addition and 251 deletion
paddle/fluid/operators/dropout_impl.cu.h 0 → 100644
浏览文件 @ 2b88057f
/* Copyright (c) 2021 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. */
#pragma once
#include <string>
#ifdef PADDLE_WITH_CUDA
#include <cuda.h>
#include <curand_kernel.h>
#include "paddle/fluid/platform/dynload/curand.h"
#endif
#ifdef PADDLE_WITH_HIP
#include <hip/hip_runtime.h>
#include <hiprand_kernel.h>
#include "paddle/fluid/platform/dynload/hiprand.h"
#endif
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/generator.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/operators/dropout_op.h"
#include "paddle/fluid/platform/aligned_vector.h"
#include "paddle/fluid/platform/gpu_launch_config.h"
namespace paddle {
namespace operators {
template <typename T, typename MaskType>
__global__ void RandomGenerator(const size_t n, uint64_t seed,
const float dropout_prob, const T* src,
MaskType* mask, T* dst,
bool is_upscale_in_train, uint64_t increment) {
int idx = blockDim.x * blockIdx.x + threadIdx.x;
#ifdef PADDLE_WITH_HIP
hiprandStatePhilox4_32_10_t state;
hiprand_init(seed, idx, increment, &state);
#else
curandStatePhilox4_32_10_t state;
curand_init(seed, idx, increment, &state);
#endif
MaskType mask_val;
T dst_val;
T factor = static_cast<T>(1.0f / (1.0f - dropout_prob));
for (; idx < n; idx += blockDim.x * gridDim.x) {
T src_val = src[idx];
#ifdef PADDLE_WITH_HIP
if (hiprand_uniform(&state) < dropout_prob) {
#else
if (curand_uniform(&state) < dropout_prob) {
#endif
mask_val = 0;
dst_val = 0;
} else {
mask_val = 1;
dst_val = is_upscale_in_train ? src_val * factor : src_val;
}
mask[idx] = mask_val;
dst[idx] = dst_val;
}
}
template <typename T, typename MaskType, int VecSize>
__global__ void VectorizedRandomGenerator(const size_t n, uint64_t seed,
const float dropout_prob,
const T* src, MaskType* mask, T* dst,
bool is_upscale_in_train,
uint64_t increment) {
using LoadT = platform::AlignedVector<T, VecSize>;
using MaskLoadT = platform::AlignedVector<MaskType, VecSize>;
#ifdef PADDLE_WITH_HIP
int64_t idx = hipBlockDim_x * hipBlockIdx_x + hipThreadIdx_x;
hiprandStatePhilox4_32_10_t state;
hiprand_init(seed, idx, increment, &state);
#else
int64_t idx = blockDim.x * blockIdx.x + threadIdx.x;
curandStatePhilox4_32_10_t state;
curand_init(seed, idx, increment, &state);
#endif
T factor = static_cast<T>(1.0f / (1.0f - dropout_prob));
for (int i = idx * VecSize; i < n; i += blockDim.x * gridDim.x * VecSize) {
LoadT src_val;
platform::Load<T, VecSize>(&src[i], &src_val);
#ifdef PADDLE_WITH_HIP
float4 rand = hiprand_uniform4(&state);
#else
float4 rand = curand_uniform4(&state);
#endif
LoadT dst_val;
MaskLoadT mask_val;
#pragma unroll
for (int j = 0; j < VecSize; j++) {
if ((&rand.x)[j] < dropout_prob) {
dst_val[j] = 0;
mask_val[j] = 0;
} else {
dst_val[j] = is_upscale_in_train ? src_val[j] * factor : src_val[j];
mask_val[j] = 1;
}
}
platform::Store<T, VecSize>(dst_val, &dst[i]);
platform::Store<MaskType, VecSize>(mask_val, &mask[i]);
}
}
template <typename T, typename MaskType, int VecSize>
__global__ void DropoutGradCUDAKernel(const T* dout, const MaskType* mask,
const T factor, const int64_t size,
T* dx) {
using LoadT = platform::AlignedVector<T, VecSize>;
using MaskLoadT = platform::AlignedVector<MaskType, VecSize>;
int64_t idx = blockDim.x * blockIdx.x + threadIdx.x;
for (int i = idx * VecSize; i < size; i += blockDim.x * gridDim.x * VecSize) {
LoadT dout_val;
platform::Load<T, VecSize>(&dout[i], &dout_val);
MaskLoadT mask_val;
platform::Load<MaskType, VecSize>(&mask[i], &mask_val);
LoadT dx_val;
#pragma unroll
for (int j = 0; j < VecSize; j++) {
dx_val[j] = dout_val[j] * static_cast<T>(mask_val[j]) * factor;
}
platform::Store<T, VecSize>(dx_val, &dx[i]);
}
}
template <typename T>
void DropoutFwGPUKernelDriver(const platform::CUDADeviceContext& dev_ctx,
bool is_test,
const std::string dropout_implementation,
float dropout_prob, bool upscale_in_train,
bool is_fix_seed, int seed_val, const Tensor& x,
const Tensor* seed, Tensor* mask, Tensor* y) {
auto& place = *dev_ctx.eigen_device();
if (!is_test) {
int64_t x_numel = x.numel();
auto stream = dev_ctx.stream();
auto* mask_data = mask->data<uint8_t>();
size_t size = framework::product(mask->dims());
auto* x_data = x.data<T>();
auto* y_data = y->data<T>();
if (dropout_prob == 1.0f) {
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_CUDA_SUCCESS(
hipMemsetAsync(y_data, 0, x_numel * sizeof(T), stream));
PADDLE_ENFORCE_CUDA_SUCCESS(
hipMemsetAsync(mask_data, 0, x_numel * sizeof(*mask_data), stream));
#else
PADDLE_ENFORCE_CUDA_SUCCESS(
cudaMemsetAsync(y_data, 0, x_numel * sizeof(T), stream));
PADDLE_ENFORCE_CUDA_SUCCESS(
cudaMemsetAsync(mask_data, 0, x_numel * sizeof(*mask_data), stream));
#endif
return;
}
platform::GpuLaunchConfig config =
platform::GetGpuLaunchConfig1D(dev_ctx, size);
// increment is used to set the args(offset) of curand_init, which defines
// offset in subsequence.
// The detail:
// https://docs.nvidia.com/cuda/curand/device-api-overview.html
// Increment should be at least the number of curand() random numbers used
// in each thread to avoid the random number generated this time being the
// same as the previous calls.
uint64_t seed_data;
uint64_t increment;
int vec_size = platform::GetVectorizedSize<T>(x_data);
auto offset = ((x_numel - 1) / (config.block_per_grid.x *
config.thread_per_block.x * vec_size) +
1) *
vec_size;
int device_id =
BOOST_GET_CONST(platform::CUDAPlace, dev_ctx.GetPlace()).GetDeviceId();
auto gen_cuda = framework::GetDefaultCUDAGenerator(device_id);
if ((seed) && platform::is_gpu_place(seed->place())) {
framework::Tensor seed_cpu_tensor;
TensorCopySync(*seed, platform::CPUPlace(), &seed_cpu_tensor);
seed_data = static_cast<uint64_t>(seed_cpu_tensor.data<int>()[0]);
increment = offset;
} else if (gen_cuda->GetIsInitPy() && (!is_fix_seed)) {
auto seed_offset = gen_cuda->IncrementOffset(offset);
seed_data = seed_offset.first;
increment = seed_offset.second;
} else {
if (seed) {
seed_data = *(seed->data<int>());
} else {
std::random_device rnd;
seed_data = is_fix_seed ? seed_val : rnd();
}
increment = offset;
}
#ifdef __HIPCC__
if (vec_size == 4 && size % 4 == 0) {
hipLaunchKernelGGL(
HIP_KERNEL_NAME(VectorizedRandomGenerator<T, uint8_t, 4>),
config.block_per_grid, config.thread_per_block, 0, stream, size,
seed_data, dropout_prob, x_data, mask_data, y_data, upscale_in_train,
increment);
} else {
hipLaunchKernelGGL(HIP_KERNEL_NAME(RandomGenerator<T, uint8_t>),
config.block_per_grid, config.thread_per_block, 0,
stream, size, seed_data, dropout_prob, x_data,
mask_data, y_data, upscale_in_train, increment);
}
#else
if (vec_size == 4 && size % 4 == 0) {
VectorizedRandomGenerator<
T, uint8_t,
4><<<config.block_per_grid, config.thread_per_block, 0, stream>>>(
size, seed_data, dropout_prob, x_data, mask_data, y_data,
upscale_in_train, increment);
} else {
RandomGenerator<T, uint8_t><<<config.block_per_grid,
config.thread_per_block, 0, stream>>>(
size, seed_data, dropout_prob, x_data, mask_data, y_data,
upscale_in_train, increment);
}
#endif
} else {
auto X = EigenMatrix<T>::Reshape(x, 1);
auto Y = EigenMatrix<T>::Reshape(*y, 1);
if (upscale_in_train) {
Y.device(place) = X;
} else {
Y.device(place) = X * static_cast<T>(1.0f - dropout_prob);
}
}
}
template <typename T>
void DropoutGradGPUKernelDriver(const platform::CUDADeviceContext& dev_ctx,
const std::string dropout_implementation,
float dropout_prob, const Tensor& grad_y,
const Tensor& mask, int64_t size,
Tensor* grad_x) {
auto M = EigenVector<uint8_t>::Flatten(mask);
auto dX = EigenVector<T>::Flatten(*grad_x);
auto dY = EigenVector<T>::Flatten(grad_y);
auto& place = *dev_ctx.eigen_device();
if (dropout_implementation == "upscale_in_train") {
if (dropout_prob == 1.0f) {
dX.device(place) = static_cast<T>(0) * dY;
} else {
int vec_size = platform::GetVectorizedSize<T>(grad_y.data<T>());
if (vec_size == 4 && size % 4 == 0) {
auto factor = static_cast<T>(1.0f / (1.0f - dropout_prob));
auto stream = dev_ctx.stream();
platform::GpuLaunchConfig config =
platform::GetGpuLaunchConfig1D(dev_ctx, size);
DropoutGradCUDAKernel<
T, uint8_t,
4><<<config.block_per_grid, config.thread_per_block, 0, stream>>>(
grad_y.data<T>(), mask.data<uint8_t>(), factor, size,
grad_x->data<T>());
} else {
dX.device(place) =
dY * M.cast<T>() / static_cast<T>(1.0f - dropout_prob);
}
}
} else {
dX.device(place) = dY * M.cast<T>();
}
}
} // namespace operators
} // namespace paddle
paddle/fluid/operators/dropout_op.cu
浏览文件 @ 2b88057f
......@@ -12,113 +12,16 @@ 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. */
#ifdef PADDLE_WITH_CUDA
#include <cuda.h>
#include <curand_kernel.h>
#include "paddle/fluid/platform/dynload/curand.h"
#endif
#ifdef PADDLE_WITH_HIP
#include <hip/hip_runtime.h>
#include <hiprand_kernel.h>
#include "paddle/fluid/platform/dynload/hiprand.h"
#endif
#include <thrust/device_ptr.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/random.h>
#include <thrust/transform.h>
#include <algorithm>
#include <string>
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/dropout_impl.cu.h"
#include "paddle/fluid/operators/dropout_op.h"
#include "paddle/fluid/platform/float16.h"
namespace paddle {
namespace operators {
template <typename T, typename MaskType>
__global__ void RandomGenerator(const size_t n, uint64_t seed,
const float dropout_prob, const T* src,
MaskType* mask, T* dst,
bool is_upscale_in_train, uint64_t increment) {
int idx = blockDim.x * blockIdx.x + threadIdx.x;
#ifdef PADDLE_WITH_HIP
hiprandStatePhilox4_32_10_t state;
hiprand_init(seed, idx, increment, &state);
#else
curandStatePhilox4_32_10_t state;
curand_init(seed, idx, increment, &state);
#endif
MaskType mask_val;
T dst_val;
T factor = static_cast<T>(1.0f / (1.0f - dropout_prob));
for (; idx < n; idx += blockDim.x * gridDim.x) {
T src_val = src[idx];
#ifdef PADDLE_WITH_HIP
if (hiprand_uniform(&state) < dropout_prob) {
#else
if (curand_uniform(&state) < dropout_prob) {
#endif
mask_val = 0;
dst_val = 0;
} else {
mask_val = 1;
dst_val = is_upscale_in_train ? src_val * factor : src_val;
}
mask[idx] = mask_val;
dst[idx] = dst_val;
}
}
template <typename T, typename MaskType, int VecSize>
__global__ void VectorizedRandomGenerator(const size_t n, uint64_t seed,
const float dropout_prob,
const T* src, MaskType* mask, T* dst,
bool is_upscale_in_train,
uint64_t increment) {
using LoadT = platform::AlignedVector<T, VecSize>;
using MaskLoadT = platform::AlignedVector<MaskType, VecSize>;
#ifdef PADDLE_WITH_HIP
int64_t idx = hipBlockDim_x * hipBlockIdx_x + hipThreadIdx_x;
hiprandStatePhilox4_32_10_t state;
hiprand_init(seed, idx, increment, &state);
#else
int64_t idx = blockDim.x * blockIdx.x + threadIdx.x;
curandStatePhilox4_32_10_t state;
curand_init(seed, idx, increment, &state);
#endif
T factor = static_cast<T>(1.0f / (1.0f - dropout_prob));
for (int i = idx * VecSize; i < n; i += blockDim.x * gridDim.x * VecSize) {
LoadT src_val;
platform::Load<T, VecSize>(&src[i], &src_val);
#ifdef PADDLE_WITH_HIP
float4 rand = hiprand_uniform4(&state);
#else
float4 rand = curand_uniform4(&state);
#endif
LoadT dst_val;
MaskLoadT mask_val;
#pragma unroll
for (int j = 0; j < VecSize; j++) {
if ((&rand.x)[j] < dropout_prob) {
dst_val[j] = 0;
mask_val[j] = 0;
} else {
dst_val[j] = is_upscale_in_train ? src_val[j] * factor : src_val[j];
mask_val[j] = 1;
}
}
platform::Store<T, VecSize>(dst_val, &dst[i]);
platform::Store<MaskType, VecSize>(mask_val, &mask[i]);
}
}
// It seems that Eigen::Tensor::setRandom in GPU will SEGFAULT.
// Use std::random and thrust::random(thrust is a std library in CUDA) to
// implement uniform random.
......@@ -137,109 +40,41 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
context.Attr<std::string>("dropout_implementation");
bool upscale_in_train = (dropout_implementation == "upscale_in_train");
auto& place = *context.template device_context<Place>().eigen_device();
if (!context.Attr<bool>("is_test")) {
int64_t x_numel = x->numel();
auto stream = context.cuda_device_context().stream();
auto* mask = context.Output<Tensor>("Mask");
auto* mask_data = mask->mutable_data<uint8_t>(context.GetPlace());
size_t size = framework::product(mask->dims());
auto* x_data = x->data<T>();
auto* y_data = y->mutable_data<T>(context.GetPlace());
if (dropout_prob == 1.0f) {
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_CUDA_SUCCESS(
hipMemsetAsync(y_data, 0, x_numel * sizeof(T), stream));
PADDLE_ENFORCE_CUDA_SUCCESS(
hipMemsetAsync(mask_data, 0, x_numel * sizeof(*mask_data), stream));
#else
PADDLE_ENFORCE_CUDA_SUCCESS(
cudaMemsetAsync(y_data, 0, x_numel * sizeof(T), stream));
PADDLE_ENFORCE_CUDA_SUCCESS(cudaMemsetAsync(
mask_data, 0, x_numel * sizeof(*mask_data), stream));
#endif
return;
}
bool is_test = context.Attr<bool>("is_test");
const auto& dev_ctx = context.cuda_device_context();
platform::GpuLaunchConfig config =
platform::GetGpuLaunchConfig1D(dev_ctx, size);
auto& dev_ctx = context.cuda_device_context();
auto* mask = context.Output<Tensor>("Mask");
mask->mutable_data<uint8_t>(context.GetPlace());
// increment is used to set the args(offset) of curand_init, which defines
// offset in subsequence.
// The detail:
// https://docs.nvidia.com/cuda/curand/device-api-overview.html
// Increment should be at least the number of curand() random numbers used
// in each thread to avoid the random number generated this time being the
// same as the previous calls.
uint64_t seed_data;
uint64_t increment;
int vec_size = platform::GetVectorizedSize<T>(x_data);
auto offset = ((x_numel - 1) / (config.block_per_grid.x *
config.thread_per_block.x * vec_size) +
1) *
vec_size;
int device_id = BOOST_GET_CONST(platform::CUDAPlace, context.GetPlace())
.GetDeviceId();
auto gen_cuda = framework::GetDefaultCUDAGenerator(device_id);
bool is_fix_seed = context.Attr<bool>("fix_seed");
int seed_val = context.Attr<int>("seed");
DropoutFwGPUKernelDriver<T>(dev_ctx, is_test, dropout_implementation,
dropout_prob, upscale_in_train, is_fix_seed,
seed_val, *x, seed, mask, y);
}
};
if (seed && platform::is_gpu_place(seed->place())) {
framework::Tensor seed_cpu_tensor;
TensorCopySync(*seed, platform::CPUPlace(), &seed_cpu_tensor);
seed_data = static_cast<uint64_t>(seed_cpu_tensor.data<int>()[0]);
increment = offset;
} else if (gen_cuda->GetIsInitPy() && (!context.Attr<bool>("fix_seed"))) {
auto seed_offset = gen_cuda->IncrementOffset(offset);
seed_data = seed_offset.first;
increment = seed_offset.second;
} else {
if (seed) {
seed_data = *(seed->data<int>());
} else {
std::random_device rnd;
seed_data = context.Attr<bool>("fix_seed") ? context.Attr<int>("seed")
: rnd();
}
increment = offset;
}
template <typename DeviceContext, typename T>
class GPUDropoutGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
PADDLE_ENFORCE_EQ(!context.Attr<bool>("is_test"), true,
platform::errors::PreconditionNotMet(
"GradOp is only callable when is_test is false"));
auto* grad_x = context.Output<Tensor>(framework::GradVarName("X"));
auto* grad_y = context.Input<Tensor>(framework::GradVarName("Out"));
auto* mask = context.Input<Tensor>("Mask");
grad_x->mutable_data<T>(context.GetPlace());
auto size = grad_x->numel();
auto& dropout_implementation =
context.Attr<std::string>("dropout_implementation");
float dropout_prob = context.Attr<float>("dropout_prob");
#ifdef __HIPCC__
if (vec_size == 4 && size % 4 == 0) {
hipLaunchKernelGGL(
HIP_KERNEL_NAME(VectorizedRandomGenerator<T, uint8_t, 4>),
config.block_per_grid, config.thread_per_block, 0, stream, size,
seed_data, dropout_prob, x_data, mask_data, y_data,
upscale_in_train, increment);
} else {
hipLaunchKernelGGL(HIP_KERNEL_NAME(RandomGenerator<T, uint8_t>),
config.block_per_grid, config.thread_per_block, 0,
stream, size, seed_data, dropout_prob, x_data,
mask_data, y_data, upscale_in_train, increment);
}
#else
if (vec_size == 4 && size % 4 == 0) {
VectorizedRandomGenerator<
T, uint8_t,
4><<<config.block_per_grid, config.thread_per_block, 0, stream>>>(
size, seed_data, dropout_prob, x_data, mask_data, y_data,
upscale_in_train, increment);
} else {
RandomGenerator<T, uint8_t><<<config.block_per_grid,
config.thread_per_block, 0, stream>>>(
size, seed_data, dropout_prob, x_data, mask_data, y_data,
upscale_in_train, increment);
}
#endif
} else {
auto X = EigenMatrix<T>::Reshape(*x, 1);
auto Y = EigenMatrix<T>::Reshape(*y, 1);
if (upscale_in_train) {
Y.device(place) = X;
} else {
Y.device(place) = X * static_cast<T>(1.0f - dropout_prob);
}
}
auto& dev_ctx =
context.template device_context<platform::CUDADeviceContext>();
DropoutGradGPUKernelDriver<T>(dev_ctx, dropout_implementation, dropout_prob,
*grad_y, *mask, size, grad_x);
}
};
......@@ -253,6 +88,6 @@ REGISTER_OP_CUDA_KERNEL(
ops::GPUDropoutKernel<plat::CUDADeviceContext, plat::float16>,
ops::GPUDropoutKernel<plat::CUDADeviceContext, double>);
REGISTER_OP_CUDA_KERNEL(
dropout_grad, ops::DropoutGradKernel<plat::CUDADeviceContext, float>,
ops::DropoutGradKernel<plat::CUDADeviceContext, plat::float16>,
ops::DropoutGradKernel<plat::CUDADeviceContext, double>);
dropout_grad, ops::GPUDropoutGradKernel<plat::CUDADeviceContext, float>,
ops::GPUDropoutGradKernel<plat::CUDADeviceContext, plat::float16>,
ops::GPUDropoutGradKernel<plat::CUDADeviceContext, double>);
paddle/fluid/operators/dropout_op.h
浏览文件 @ 2b88057f
......@@ -21,40 +21,10 @@ limitations under the License. */
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/generator.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/platform/aligned_vector.h"
#include "paddle/fluid/platform/gpu_launch_config.h"
namespace paddle {
namespace operators {
#if defined(__NVCC__) || defined(__HIPCC__)
template <typename T, typename MaskType, int VecSize>
__global__ void DropoutGradCUDAKernel(const T* dout, const MaskType* mask,
const T factor, const int64_t size,
T* dx) {
using LoadT = platform::AlignedVector<T, VecSize>;
using MaskLoadT = platform::AlignedVector<MaskType, VecSize>;
int64_t idx = blockDim.x * blockIdx.x + threadIdx.x;
for (int i = idx * VecSize; i < size; i += blockDim.x * gridDim.x * VecSize) {
LoadT dout_val;
platform::Load<T, VecSize>(&dout[i], &dout_val);
MaskLoadT mask_val;
platform::Load<MaskType, VecSize>(&mask[i], &mask_val);
LoadT dx_val;
#pragma unroll
for (int j = 0; j < VecSize; j++) {
dx_val[j] = dout_val[j] * static_cast<T>(mask_val[j]) * factor;
}
platform::Store<T, VecSize>(dx_val, &dx[i]);
}
}
#endif
using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
......@@ -137,7 +107,6 @@ class CPUDropoutKernel : public framework::OpKernel<T> {
}
}
};
template <typename DeviceContext, typename T>
class DropoutGradKernel : public framework::OpKernel<T> {
public:
......@@ -146,7 +115,6 @@ class DropoutGradKernel : public framework::OpKernel<T> {
auto* grad_y = context.Input<Tensor>(framework::GradVarName("Out"));
auto* mask = context.Input<Tensor>("Mask");
grad_x->mutable_data<T>(context.GetPlace());
auto size = grad_x->numel();
auto dX = EigenVector<T>::Flatten(*grad_x);
auto dY = EigenVector<T>::Flatten(*grad_y);
......@@ -169,23 +137,8 @@ class DropoutGradKernel : public framework::OpKernel<T> {
if (dropout_prob == 1.0f) {
dX.device(place) = static_cast<T>(0) * dY;
} else {
int vec_size = platform::GetVectorizedSize<T>(grad_y->data<T>());
if (platform::is_gpu_place(context.GetPlace()) && vec_size == 4 &&
size % 4 == 0) {
#if defined(__NVCC__) || defined(__HIPCC__)
auto factor = static_cast<T>(1.0f / (1.0f - dropout_prob));
auto stream = context.cuda_device_context().stream();
platform::GpuLaunchConfig config = platform::GetGpuLaunchConfig1D(
context.cuda_device_context(), size);
DropoutGradCUDAKernel<T, uint8_t, 4><<<
config.block_per_grid, config.thread_per_block, 0, stream>>>(
grad_y->data<T>(), mask->data<uint8_t>(), factor, size,
grad_x->data<T>());
#endif
} else {
dX.device(place) =
dY * M.cast<T>() / static_cast<T>(1.0f - dropout_prob);
}
dX.device(place) =
dY * M.cast<T>() / static_cast<T>(1.0f - dropout_prob);
}
} else {
dX.device(place) = dY * M.cast<T>();
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或
反馈
建议 客服 返回
顶部