wehub-resource-sync
269 lines
12 KiB
Plaintext
269 lines
12 KiB
Plaintext
// Copyright (c) 2024 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 "paddle/phi/core/kernel_registry.h"
|
|
#include "paddle/phi/kernels/fusion/gpu/cudnn_bn_stats_finalize.cu.h"
|
|
#include "paddle/phi/kernels/fusion/gpu/cudnn_norm_conv.cu.h"
|
|
#include "paddle/phi/kernels/fusion/gpu/cudnn_scale_bias_add_relu.cu.h"
|
|
#include "paddle/utils/optional.h"
|
|
|
|
#if CUDNN_VERSION >= 8000
|
|
namespace phi {
|
|
|
|
template <typename T, typename Context>
|
|
void ResNetUnitGradKernel(const Context &dev_ctx,
|
|
const DenseTensor &x_in,
|
|
const DenseTensor &filter_x_in,
|
|
const DenseTensor &conv_x_in,
|
|
const DenseTensor &scale_x_in,
|
|
const DenseTensor &bias_x_in,
|
|
const DenseTensor &saved_mean_x_in,
|
|
const DenseTensor &saved_invstd_x_in,
|
|
const optional<DenseTensor> &z_in,
|
|
const optional<DenseTensor> &filter_z_in,
|
|
const optional<DenseTensor> &conv_z_in,
|
|
const optional<DenseTensor> &scale_z_in,
|
|
const optional<DenseTensor> &bias_z_in,
|
|
const optional<DenseTensor> &saved_mean_z_in,
|
|
const optional<DenseTensor> &saved_invstd_z_in,
|
|
const DenseTensor &out,
|
|
const DenseTensor &bit_mask,
|
|
const DenseTensor &out_grad,
|
|
int stride,
|
|
int stride_z,
|
|
int padding,
|
|
int dilation,
|
|
int group,
|
|
float momentum_in,
|
|
float epsilon,
|
|
const std::string &data_format,
|
|
bool fuse_add,
|
|
bool has_shortcut,
|
|
bool use_global_stats,
|
|
bool is_test,
|
|
bool use_addto,
|
|
const std::string &act_type,
|
|
DenseTensor *x_grad,
|
|
DenseTensor *filter_x_grad,
|
|
DenseTensor *scale_x_grad,
|
|
DenseTensor *bias_x_grad,
|
|
DenseTensor *z_grad,
|
|
DenseTensor *filter_z_grad,
|
|
DenseTensor *scale_z_grad,
|
|
DenseTensor *bias_z_grad) {
|
|
PADDLE_ENFORCE_EQ(backends::gpu::CudnnDataType<T>::type,
|
|
CUDNN_DATA_HALF,
|
|
common::errors::Unavailable(
|
|
"ResNetUnitOp only supports float16 for now."));
|
|
|
|
const DenseTensor *y_grad = &out_grad;
|
|
|
|
const DenseTensor *x = &x_in;
|
|
const DenseTensor *filter_x = &filter_x_in;
|
|
const DenseTensor *scale_x = &scale_x_in;
|
|
const DenseTensor *bias_x = &bias_x_in;
|
|
const DenseTensor *saved_mean_x = &saved_mean_x_in;
|
|
const DenseTensor *saved_invstd_x = &saved_invstd_x_in;
|
|
|
|
const DenseTensor *conv_out_x = &conv_x_in;
|
|
const DenseTensor *output = &out;
|
|
const DenseTensor *bitmask = &bit_mask;
|
|
|
|
double eps = static_cast<double>(epsilon);
|
|
double momentum = static_cast<double>(momentum_in);
|
|
|
|
auto x_shape = vectorize<int>(x->dims());
|
|
auto filter_x_shape = vectorize<int>(filter_x->dims());
|
|
auto param_shape = vectorize<int>(scale_x->dims());
|
|
auto output_shape = vectorize<int>(output->dims());
|
|
auto bitmask_shape = vectorize<int>(bitmask->dims());
|
|
|
|
// 1. Backward of BN (+ Add + Relu) for x, get conv_out_x_grad,
|
|
// scale_x_grad, bias_x_grad
|
|
DenseTensor conv_out_x_grad;
|
|
conv_out_x_grad.Resize(conv_out_x->dims());
|
|
phi::fusion::CudnnScaleBiasAddRelu<T> sbar_x_op(dev_ctx,
|
|
act_type,
|
|
fuse_add,
|
|
has_shortcut,
|
|
output_shape,
|
|
param_shape,
|
|
bitmask_shape);
|
|
if (has_shortcut) {
|
|
// X Z
|
|
// | |
|
|
// NormConv NormConv
|
|
// | |
|
|
// BNStatsFinalize BNStatsFinalize
|
|
// \ /
|
|
// ScaleBiasAddRelu
|
|
// |
|
|
// Y
|
|
const DenseTensor *z = z_in.get_ptr();
|
|
const DenseTensor *filter_z = filter_z_in.get_ptr();
|
|
const DenseTensor *scale_z = scale_z_in.get_ptr();
|
|
const DenseTensor *bias_z = bias_z_in.get_ptr();
|
|
const DenseTensor *saved_mean_z = saved_mean_z_in.get_ptr();
|
|
const DenseTensor *saved_invstd_z = saved_invstd_z_in.get_ptr();
|
|
const DenseTensor *conv_out_z = conv_z_in.get_ptr();
|
|
|
|
// 1.1 Backward of BN + Add (+ Relu) for x, get conv_out_x_grad,
|
|
// scale_x_grad, bias_x_grad and z_grad_temp
|
|
DenseTensor z_grad_temp;
|
|
z_grad_temp.Resize(conv_out_z->dims());
|
|
sbar_x_op.Backward(dev_ctx,
|
|
*y_grad,
|
|
*conv_out_x,
|
|
*scale_x,
|
|
*bias_x,
|
|
*saved_mean_x,
|
|
*saved_invstd_x,
|
|
bitmask,
|
|
&conv_out_x_grad,
|
|
&z_grad_temp,
|
|
scale_x_grad,
|
|
bias_x_grad,
|
|
eps);
|
|
|
|
// 1.2 bn backward for z, get conv_out_z_grad, dscale_z, dbias_z
|
|
DenseTensor conv_out_z_grad;
|
|
conv_out_z_grad.Resize(conv_out_z->dims());
|
|
phi::fusion::CudnnScaleBiasAddRelu<T> sbar_z_op(
|
|
dev_ctx, "", false, false, output_shape, param_shape, bitmask_shape);
|
|
sbar_z_op.Backward(dev_ctx,
|
|
z_grad_temp,
|
|
*conv_out_z,
|
|
*scale_z,
|
|
*bias_z,
|
|
*saved_mean_z,
|
|
*saved_invstd_z,
|
|
nullptr,
|
|
&conv_out_z_grad,
|
|
nullptr,
|
|
scale_z_grad,
|
|
bias_z_grad,
|
|
eps);
|
|
|
|
// 1.3 Backward of Conv for z, get z_grad and filter_z_grad
|
|
auto z_shape = vectorize<int>(z->dims());
|
|
auto filter_z_shape = vectorize<int>(filter_z->dims());
|
|
phi::fusion::CudnnNormConvolutionGrad<T> conv_z_op(dev_ctx,
|
|
z_shape,
|
|
filter_z_shape,
|
|
output_shape,
|
|
padding,
|
|
stride_z,
|
|
dilation,
|
|
group);
|
|
conv_z_op.Backward(
|
|
dev_ctx, *z, *filter_z, conv_out_z_grad, z_grad, filter_z_grad);
|
|
} else {
|
|
// 1.1 Backward of BN (+ Add + Relu) for x, get conv_out_x_grad,
|
|
// scale_x_grad, bias_x_grad (and z_grad)
|
|
DenseTensor *z_grad_tmp = fuse_add ? z_grad : nullptr;
|
|
sbar_x_op.Backward(dev_ctx,
|
|
*y_grad,
|
|
*conv_out_x,
|
|
*scale_x,
|
|
*bias_x,
|
|
*saved_mean_x,
|
|
*saved_invstd_x,
|
|
bitmask,
|
|
&conv_out_x_grad,
|
|
z_grad_tmp,
|
|
scale_x_grad,
|
|
bias_x_grad,
|
|
eps);
|
|
}
|
|
|
|
// 2. Backward of Conv for x, get x_grad and filter_x_grad
|
|
phi::fusion::CudnnNormConvolutionGrad<T> conv_x_op(dev_ctx,
|
|
x_shape,
|
|
filter_x_shape,
|
|
output_shape,
|
|
padding,
|
|
stride,
|
|
dilation,
|
|
group);
|
|
conv_x_op.Backward(dev_ctx,
|
|
*x,
|
|
*filter_x,
|
|
conv_out_x_grad,
|
|
x_grad,
|
|
filter_x_grad,
|
|
use_addto);
|
|
}
|
|
|
|
} // namespace phi
|
|
|
|
PD_REGISTER_KERNEL(resnet_unit_grad,
|
|
GPU,
|
|
ALL_LAYOUT,
|
|
phi::ResNetUnitGradKernel,
|
|
phi::float16) {}
|
|
#else
|
|
namespace phi {
|
|
|
|
template <typename T, typename Context>
|
|
void ResNetUnitGradEmptyKernel(const Context &dev_ctx,
|
|
const DenseTensor &x_in,
|
|
const DenseTensor &filter_x_in,
|
|
const DenseTensor &conv_x_in,
|
|
const DenseTensor &scale_x_in,
|
|
const DenseTensor &bias_x_in,
|
|
const DenseTensor &saved_mean_x_in,
|
|
const DenseTensor &saved_invstd_x_in,
|
|
const optional<DenseTensor> &z_in,
|
|
const optional<DenseTensor> &filter_z_in,
|
|
const optional<DenseTensor> &conv_z_in,
|
|
const optional<DenseTensor> &scale_z_in,
|
|
const optional<DenseTensor> &bias_z_in,
|
|
const optional<DenseTensor> &saved_mean_z_in,
|
|
const optional<DenseTensor> &saved_invstd_z_in,
|
|
const DenseTensor &out,
|
|
const DenseTensor &bit_mask,
|
|
const DenseTensor &out_grad,
|
|
int stride,
|
|
int stride_z,
|
|
int padding,
|
|
int dilation,
|
|
int group,
|
|
float momentum_in,
|
|
float epsilon,
|
|
const std::string &data_format,
|
|
bool fuse_add,
|
|
bool has_shortcut,
|
|
bool use_global_stats,
|
|
bool is_test,
|
|
bool use_addto,
|
|
const std::string &act_type,
|
|
DenseTensor *x_grad,
|
|
DenseTensor *filter_x_grad,
|
|
DenseTensor *scale_x_grad,
|
|
DenseTensor *bias_x_grad,
|
|
DenseTensor *z_grad,
|
|
DenseTensor *filter_z_grad,
|
|
DenseTensor *scale_z_grad,
|
|
DenseTensor *bias_z_grad) {}
|
|
} // namespace phi
|
|
|
|
PD_REGISTER_KERNEL(resnet_unit_grad,
|
|
GPU,
|
|
ALL_LAYOUT,
|
|
phi::ResNetUnitGradEmptyKernel,
|
|
phi::float16) {}
|
|
#endif
|