Files
2026-07-13 12:40:42 +08:00

278 lines
11 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 ResNetUnitKernel(const Context &dev_ctx,
const DenseTensor &x_in,
const DenseTensor &filter_x_in,
const DenseTensor &scale_x_in,
const DenseTensor &bias_x_in,
const DenseTensor &mean_x_in,
const DenseTensor &var_x_in,
const optional<DenseTensor> &z_in,
const optional<DenseTensor> &filter_z_in,
const optional<DenseTensor> &scale_z_in,
const optional<DenseTensor> &bias_z_in,
const optional<DenseTensor> &mean_z_in,
const optional<DenseTensor> &var_z_in,
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 *out,
DenseTensor *bit_mask,
DenseTensor *conv_x,
DenseTensor *saved_mean_x,
DenseTensor *saved_invstd_x,
DenseTensor *running_mean_x,
DenseTensor *running_var_x,
DenseTensor *conv_z,
DenseTensor *saved_mean_z,
DenseTensor *saved_invstd_z,
DenseTensor *running_mean_z,
DenseTensor *running_var_z) {
PADDLE_ENFORCE_EQ(backends::gpu::CudnnDataType<T>::type,
CUDNN_DATA_HALF,
common::errors::Unavailable(
"ResNetUnitOp only supports float16 for now."));
// input x
const DenseTensor *input_x = &x_in;
const DenseTensor *filter_x = &filter_x_in;
const DenseTensor *scale_x = &scale_x_in;
const DenseTensor *bias_x = &bias_x_in;
// norm conv
DenseTensor *conv_out_x = conv_x;
// sbar
DenseTensor *output = out;
DenseTensor *bitmask = bit_mask;
// attrs
double eps = static_cast<double>(epsilon);
double momentum = static_cast<double>(momentum_in);
bool is_train = !is_test && !use_global_stats;
auto input_x_shape = vectorize<int>(input_x->dims());
auto filter_x_shape = vectorize<int>(filter_x->dims());
// std::swap used to convert shape of filter from conv2d when kernel size is
// 1.
if (filter_x_shape[1] != filter_x_shape[2] && 1 == filter_x_shape[2]) {
std::swap(filter_x_shape[1], filter_x_shape[3]);
}
auto param_dims = scale_x->dims();
auto param_shape = vectorize<int>(scale_x->dims());
if (1 == param_shape.size()) {
param_shape = {1, 1, 1, param_shape[0]};
}
auto output_shape = vectorize<int>(output->dims());
auto bitmask_shape = vectorize<int>(bitmask->dims());
int output_channel = filter_x_shape[0];
int64_t ele_count =
std::accumulate(
output_shape.begin(), output_shape.end(), 1, std::multiplies<int>()) /
output_channel;
// 1. Conv
DenseTensor sum_x;
DenseTensor sum_of_squares_x;
sum_x.Resize(param_dims);
sum_of_squares_x.Resize(param_dims);
phi::fusion::CudnnNormConvolution<T> conv_x_op(dev_ctx,
input_x_shape,
filter_x_shape,
output_shape,
padding,
stride,
dilation,
group);
conv_x_op.Forward(
dev_ctx, *input_x, *filter_x, conv_out_x, &sum_x, &sum_of_squares_x);
// 2. BN
DenseTensor equiv_scale_x;
DenseTensor equiv_bias_x;
equiv_scale_x.Resize(param_dims);
equiv_bias_x.Resize(param_dims);
phi::fusion::CudnnBNStatsFinalize<T> bn_x_op(dev_ctx, param_shape);
bn_x_op.Forward(dev_ctx,
sum_x,
sum_of_squares_x,
*scale_x,
*bias_x,
saved_mean_x,
saved_invstd_x,
running_mean_x,
running_var_x,
&equiv_scale_x,
&equiv_bias_x,
eps,
momentum,
ele_count,
is_train);
// 3. scale + bias + add + relu
phi::fusion::CudnnScaleBiasAddRelu<T> sbar_op(dev_ctx,
act_type,
fuse_add,
has_shortcut,
output_shape,
param_shape,
bitmask_shape);
if (has_shortcut) {
// input z
const DenseTensor *input_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();
// norm conv
DenseTensor *conv_out_z = conv_z;
auto input_z_shape = vectorize<int>(input_z->dims());
auto filter_z_shape = vectorize<int>(filter_z->dims());
// 3.1 Conv for second input
DenseTensor sum_z;
DenseTensor sum_of_squares_z;
sum_z.Resize(param_dims);
sum_of_squares_z.Resize(param_dims);
phi::fusion::CudnnNormConvolution<T> conv_z_op(dev_ctx,
input_z_shape,
filter_z_shape,
output_shape,
padding,
stride_z,
dilation,
group);
conv_z_op.Forward(
dev_ctx, *input_z, *filter_z, conv_out_z, &sum_z, &sum_of_squares_z);
// 3.2 BN for second input
DenseTensor equiv_scale_z;
DenseTensor equiv_bias_z;
equiv_scale_z.Resize(param_dims);
equiv_bias_z.Resize(param_dims);
phi::fusion::CudnnBNStatsFinalize<T> bn_z_op(dev_ctx, param_shape);
bn_z_op.Forward(dev_ctx,
sum_z,
sum_of_squares_z,
*scale_z,
*bias_z,
saved_mean_z,
saved_invstd_z,
running_mean_z,
running_var_z,
&equiv_scale_z,
&equiv_bias_z,
eps,
momentum,
ele_count,
is_train);
// 3.3 sbar
sbar_op.Forward(dev_ctx,
*conv_out_x,
equiv_scale_x,
equiv_bias_x,
conv_out_z,
&equiv_scale_z,
&equiv_bias_z,
output,
bitmask);
} else {
const DenseTensor *input_z = fuse_add ? z_in.get_ptr() : nullptr;
sbar_op.Forward(dev_ctx,
*conv_out_x,
equiv_scale_x,
equiv_bias_x,
input_z,
nullptr,
nullptr,
output,
bitmask);
}
}
} // namespace phi
PD_REGISTER_KERNEL(
resnet_unit, GPU, ALL_LAYOUT, phi::ResNetUnitKernel, phi::float16) {}
#else
namespace phi {
template <typename T, typename Context>
void ResNetUnitEmptyKernel(const Context &dev_ctx,
const DenseTensor &x_in,
const DenseTensor &filter_x_in,
const DenseTensor &scale_x_in,
const DenseTensor &bias_x_in,
const DenseTensor &mean_x_in,
const DenseTensor &var_x_in,
const optional<DenseTensor> &z_in,
const optional<DenseTensor> &filter_z_in,
const optional<DenseTensor> &scale_z_in,
const optional<DenseTensor> &bias_z_in,
const optional<DenseTensor> &mean_z_in,
const optional<DenseTensor> &var_z_in,
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 *out,
DenseTensor *bit_mask,
DenseTensor *conv_x,
DenseTensor *saved_mean_x,
DenseTensor *saved_invstd_x,
DenseTensor *running_mean_x,
DenseTensor *running_var_x,
DenseTensor *conv_z,
DenseTensor *saved_mean_z,
DenseTensor *saved_invstd_z,
DenseTensor *running_mean_z,
DenseTensor *running_var_z) {
PADDLE_THROW(common::errors::Unavailable(
"ResNetUnitOp only supports CUDNN_VERSION >= 8000 for now."));
}
} // namespace phi
PD_REGISTER_KERNEL(
resnet_unit, GPU, ALL_LAYOUT, phi::ResNetUnitEmptyKernel, phi::float16) {}
#endif