wehub-resource-sync
372 lines
14 KiB
C++
372 lines
14 KiB
C++
/* Copyright (c) 2018 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 <vector>
|
|
|
|
#include "paddle/common/enforce.h"
|
|
#include "paddle/phi/backends/context_pool.h"
|
|
#include "paddle/phi/backends/gpu/gpu_dnn.h"
|
|
#include "paddle/phi/core/dense_tensor.h"
|
|
namespace phi {
|
|
namespace funcs {
|
|
|
|
class CudnnRNNCache {
|
|
public:
|
|
CudnnRNNCache() {
|
|
x_desc_ = NULL;
|
|
y_desc_ = NULL;
|
|
}
|
|
~CudnnRNNCache() { release(); }
|
|
|
|
cudnnRNNDescriptor_t rnn_desc_;
|
|
#if CUDNN_VERSION >= 90000
|
|
cudnnRNNDataDescriptor_t x_desc_;
|
|
cudnnRNNDataDescriptor_t y_desc_;
|
|
#else
|
|
cudnnTensorDescriptor_t *x_desc_;
|
|
cudnnTensorDescriptor_t *y_desc_;
|
|
#endif
|
|
|
|
cudnnTensorDescriptor_t hx_desc_;
|
|
cudnnTensorDescriptor_t cx_desc_;
|
|
cudnnTensorDescriptor_t hy_desc_;
|
|
cudnnTensorDescriptor_t cy_desc_;
|
|
|
|
cudnnTensorDescriptor_t dhx_desc_;
|
|
cudnnTensorDescriptor_t dcx_desc_;
|
|
cudnnTensorDescriptor_t dhy_desc_;
|
|
cudnnTensorDescriptor_t dcy_desc_;
|
|
|
|
cudnnTensorDescriptor_t output_x_desc_;
|
|
cudnnTensorDescriptor_t output_y_desc_;
|
|
|
|
cudnnDropoutDescriptor_t dropout_desc_;
|
|
|
|
size_t weights_size_;
|
|
cudnnFilterDescriptor_t w_desc_;
|
|
cudnnFilterDescriptor_t dw_desc_;
|
|
|
|
size_t workspace_size_;
|
|
DenseTensor workspace_data_;
|
|
|
|
size_t seq_length_;
|
|
|
|
float dropout_prob_;
|
|
bool is_bidirec_;
|
|
|
|
int batch_size_;
|
|
int input_size_;
|
|
int hidden_size_;
|
|
int num_layers_;
|
|
int seed_;
|
|
|
|
void init(cudnnHandle_t handle,
|
|
const phi::Place &place,
|
|
size_t seq_len,
|
|
int batch_size,
|
|
int input_size,
|
|
int hidden_size,
|
|
int num_layers,
|
|
float dropout_prob,
|
|
bool is_bidirec,
|
|
int seed,
|
|
int weight_numel,
|
|
size_t *reserve_size_,
|
|
DenseTensor *dropout_state_,
|
|
bool initialized,
|
|
cudnnDataType_t cudnn_type) {
|
|
seq_length_ = seq_len;
|
|
batch_size_ = batch_size;
|
|
input_size_ = input_size;
|
|
hidden_size_ = hidden_size;
|
|
num_layers_ = num_layers;
|
|
dropout_prob_ = dropout_prob;
|
|
is_bidirec_ = is_bidirec;
|
|
seed_ = seed;
|
|
|
|
const auto numDirections = is_bidirec_ ? 2 : 1;
|
|
const int64_t hidden_size_directions_64 =
|
|
static_cast<int64_t>(hidden_size_) * numDirections;
|
|
PADDLE_ENFORCE_LE_INT_MAX(hidden_size_directions_64,
|
|
"RNN hidden size times directions");
|
|
const int hidden_size_directions =
|
|
static_cast<int>(hidden_size_directions_64);
|
|
const int64_t num_layers_directions_64 =
|
|
static_cast<int64_t>(num_layers_) * numDirections;
|
|
PADDLE_ENFORCE_LE_INT_MAX(num_layers_directions_64,
|
|
"RNN num layers times directions");
|
|
const int num_layers_directions =
|
|
static_cast<int>(num_layers_directions_64);
|
|
const int64_t hidden_size_batch_64 =
|
|
static_cast<int64_t>(hidden_size_) * batch_size_;
|
|
PADDLE_ENFORCE_LE_INT_MAX(hidden_size_batch_64,
|
|
"RNN hidden size times batch size");
|
|
const int hidden_size_batch = static_cast<int>(hidden_size_batch_64);
|
|
auto cudnn_size =
|
|
cudnn_type == CUDNN_DATA_FLOAT ? sizeof(float) : sizeof(double);
|
|
#if CUDNN_VERSION >= 90000
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateRNNDataDescriptor(&x_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateRNNDataDescriptor(&y_desc_));
|
|
|
|
PADDLE_ENFORCE_LE_INT_MAX(seq_length_, "RNN sequence length");
|
|
const int seq_length_int = static_cast<int>(seq_length_);
|
|
std::vector<int> seq_length_array(batch_size_);
|
|
for (int i = 0; i < batch_size_; ++i) {
|
|
seq_length_array[i] = seq_length_int;
|
|
}
|
|
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetRNNDataDescriptor(
|
|
x_desc_,
|
|
cudnn_type,
|
|
CUDNN_RNN_DATA_LAYOUT_BATCH_MAJOR_UNPACKED,
|
|
seq_length_int,
|
|
batch_size_,
|
|
input_size_,
|
|
reinterpret_cast<const int *>(seq_length_array.data()),
|
|
nullptr));
|
|
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetRNNDataDescriptor(
|
|
y_desc_,
|
|
cudnn_type,
|
|
CUDNN_RNN_DATA_LAYOUT_BATCH_MAJOR_UNPACKED,
|
|
seq_length_int,
|
|
batch_size_,
|
|
hidden_size_directions,
|
|
reinterpret_cast<const int *>(seq_length_array.data()),
|
|
nullptr));
|
|
#else
|
|
PADDLE_ENFORCE_LE_INT_MAX(seq_length_, "RNN sequence length");
|
|
const int seq_length_int = static_cast<int>(seq_length_);
|
|
x_desc_ = new cudnnTensorDescriptor_t[seq_length_];
|
|
y_desc_ = new cudnnTensorDescriptor_t[seq_length_];
|
|
std::vector<int> dims = {batch_size_, input_size_, 1};
|
|
std::vector<int> strides = {input_size_, 1, 1};
|
|
|
|
std::vector<int> dims_y = {batch_size_, hidden_size_directions, 1};
|
|
std::vector<int> strides_y = {hidden_size_directions, 1, 1};
|
|
|
|
for (size_t i = 0; i < seq_length_; ++i) {
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateTensorDescriptor(&x_desc_[i]));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateTensorDescriptor(&y_desc_[i]));
|
|
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetTensorNdDescriptor(
|
|
x_desc_[i], cudnn_type, 3, dims.data(), strides.data()));
|
|
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetTensorNdDescriptor(
|
|
y_desc_[i], cudnn_type, 3, dims_y.data(), strides_y.data()));
|
|
}
|
|
#endif
|
|
|
|
std::vector<int> dims_hx = {
|
|
num_layers_directions, batch_size_, hidden_size_};
|
|
std::vector<int> strides_hx = {hidden_size_batch, hidden_size_, 1};
|
|
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateTensorDescriptor(&hx_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateTensorDescriptor(&cx_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateTensorDescriptor(&hy_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateTensorDescriptor(&cy_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateTensorDescriptor(&dhx_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateTensorDescriptor(&dcx_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateTensorDescriptor(&dhy_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateTensorDescriptor(&dcy_desc_));
|
|
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetTensorNdDescriptor(
|
|
hx_desc_, cudnn_type, 3, dims_hx.data(), strides_hx.data()));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetTensorNdDescriptor(
|
|
cx_desc_, cudnn_type, 3, dims_hx.data(), strides_hx.data()));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetTensorNdDescriptor(
|
|
hy_desc_, cudnn_type, 3, dims_hx.data(), strides_hx.data()));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetTensorNdDescriptor(
|
|
cy_desc_, cudnn_type, 3, dims_hx.data(), strides_hx.data()));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetTensorNdDescriptor(
|
|
dhx_desc_, cudnn_type, 3, dims_hx.data(), strides_hx.data()));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetTensorNdDescriptor(
|
|
dcx_desc_, cudnn_type, 3, dims_hx.data(), strides_hx.data()));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetTensorNdDescriptor(
|
|
dhy_desc_, cudnn_type, 3, dims_hx.data(), strides_hx.data()));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetTensorNdDescriptor(
|
|
dcy_desc_, cudnn_type, 3, dims_hx.data(), strides_hx.data()));
|
|
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateDropoutDescriptor(&dropout_desc_));
|
|
|
|
size_t state_size;
|
|
auto *dev_ctx = DeviceContextPool::Instance().Get(place);
|
|
if (!initialized) {
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDropoutGetStatesSize(handle, &state_size));
|
|
dropout_state_->Resize({static_cast<int64_t>(state_size)});
|
|
uint8_t *dropout_state_data = dev_ctx->Alloc<uint8_t>(dropout_state_);
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnSetDropoutDescriptor(dropout_desc_,
|
|
handle,
|
|
dropout_prob_,
|
|
dropout_state_data,
|
|
state_size,
|
|
seed_));
|
|
} else {
|
|
uint8_t *dropout_state_data = dropout_state_->data<uint8_t>();
|
|
auto dropout_state_dims = dropout_state_->dims();
|
|
state_size = dropout_state_dims[0];
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnRestoreDropoutDescriptor(dropout_desc_,
|
|
handle,
|
|
dropout_prob_,
|
|
dropout_state_data,
|
|
state_size,
|
|
0));
|
|
}
|
|
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateRNNDescriptor(&rnn_desc_));
|
|
#if CUDNN_VERSION >= 90000
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetRNNDescriptor_v8(
|
|
rnn_desc_,
|
|
CUDNN_RNN_ALGO_STANDARD,
|
|
CUDNN_LSTM,
|
|
CUDNN_RNN_DOUBLE_BIAS,
|
|
is_bidirec_ ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL,
|
|
CUDNN_LINEAR_INPUT,
|
|
cudnn_type,
|
|
cudnn_type,
|
|
CUDNN_DEFAULT_MATH,
|
|
input_size_,
|
|
hidden_size_,
|
|
hidden_size_,
|
|
num_layers_,
|
|
dropout_desc_,
|
|
CUDNN_RNN_PADDED_IO_ENABLED));
|
|
#else
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetRNNDescriptor_v6(
|
|
handle,
|
|
rnn_desc_,
|
|
hidden_size_,
|
|
num_layers_,
|
|
dropout_desc_,
|
|
CUDNN_LINEAR_INPUT,
|
|
is_bidirec_ ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL,
|
|
CUDNN_LSTM,
|
|
CUDNN_RNN_ALGO_STANDARD,
|
|
cudnn_type));
|
|
#endif
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateFilterDescriptor(&w_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnCreateFilterDescriptor(&dw_desc_));
|
|
|
|
#if CUDNN_VERSION >= 90000
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnGetRNNWeightSpaceSize(
|
|
handle, rnn_desc_, &weights_size_));
|
|
#else
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnGetRNNParamsSize(
|
|
handle, rnn_desc_, x_desc_[0], &weights_size_, cudnn_type));
|
|
#endif
|
|
PADDLE_ENFORCE_EQ(
|
|
weights_size_,
|
|
cudnn_size * weight_numel,
|
|
common::errors::InvalidArgument(
|
|
"The cudnn lstm and setting weight size should be same."));
|
|
|
|
PADDLE_ENFORCE_LE_INT_MAX(weights_size_ / cudnn_size,
|
|
"weights_size_ / cudnn_size");
|
|
int dim_w[3];
|
|
dim_w[0] = static_cast<int>(weights_size_ / cudnn_size);
|
|
dim_w[1] = 1;
|
|
dim_w[2] = 1;
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetFilterNdDescriptor(
|
|
w_desc_, cudnn_type, CUDNN_TENSOR_NCHW, 3, dim_w));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnSetFilterNdDescriptor(
|
|
dw_desc_, cudnn_type, CUDNN_TENSOR_NCHW, 3, dim_w));
|
|
#if CUDNN_VERSION >= 90000
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnGetRNNTempSpaceSizes(handle,
|
|
rnn_desc_,
|
|
CUDNN_FWD_MODE_TRAINING,
|
|
x_desc_,
|
|
&workspace_size_,
|
|
reserve_size_));
|
|
#else
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnGetRNNWorkspaceSize(
|
|
handle, rnn_desc_, seq_length_int, x_desc_, &workspace_size_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::cudnnGetRNNTrainingReserveSize(
|
|
handle, rnn_desc_, seq_length_int, x_desc_, reserve_size_));
|
|
#endif
|
|
workspace_data_.Resize({static_cast<int64_t>(workspace_size_)});
|
|
dev_ctx->Alloc<uint8_t>(&workspace_data_);
|
|
}
|
|
|
|
void release() {
|
|
#if CUDNN_VERSION >= 90000
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyRNNDataDescriptor(x_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyRNNDataDescriptor(y_desc_));
|
|
#else
|
|
for (size_t i = 0; i < seq_length_; ++i) {
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyTensorDescriptor(x_desc_[i]));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyTensorDescriptor(y_desc_[i]));
|
|
}
|
|
|
|
delete[] x_desc_;
|
|
delete[] y_desc_;
|
|
#endif
|
|
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyTensorDescriptor(hx_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyTensorDescriptor(cx_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyTensorDescriptor(hy_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyTensorDescriptor(cy_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyTensorDescriptor(dhx_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyTensorDescriptor(dcx_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyTensorDescriptor(dhy_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyTensorDescriptor(dcy_desc_));
|
|
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyDropoutDescriptor(dropout_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyRNNDescriptor(rnn_desc_));
|
|
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyFilterDescriptor(w_desc_));
|
|
PADDLE_ENFORCE_GPU_SUCCESS(
|
|
phi::dynload::cudnnDestroyFilterDescriptor(dw_desc_));
|
|
}
|
|
};
|
|
|
|
} // namespace funcs
|
|
} // namespace phi
|