wehub-resource-sync
145 lines
4.7 KiB
C++
145 lines
4.7 KiB
C++
// 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/temporal_shift_grad_kernel.h"
|
|
|
|
#include "paddle/common/layout.h"
|
|
#include "paddle/phi/backends/cpu/cpu_context.h"
|
|
#include "paddle/phi/core/kernel_registry.h"
|
|
|
|
namespace phi {
|
|
|
|
template <typename T>
|
|
void TemporalShiftBwNCHW(const T* output_grad,
|
|
T* input_grad,
|
|
const int64_t ntchw,
|
|
const int64_t tchw,
|
|
const int64_t chw,
|
|
const int64_t hw,
|
|
const int t,
|
|
const int c1,
|
|
const int c2) {
|
|
int src_it = 0;
|
|
for (int64_t i = 0; i < ntchw; i++) {
|
|
int64_t it = (i % tchw) / chw;
|
|
int64_t ic = (i % chw) / hw;
|
|
|
|
if (ic < c1) {
|
|
src_it = it + 1;
|
|
} else if (ic < c2) {
|
|
src_it = it - 1;
|
|
} else {
|
|
src_it = it;
|
|
}
|
|
|
|
if (src_it >= 0 && src_it < t) {
|
|
input_grad[i] = output_grad[i + (src_it - it) * chw];
|
|
} else {
|
|
input_grad[i] = 0;
|
|
}
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
void TemporalShiftBwNHWC(const T* output_grad,
|
|
T* input_grad,
|
|
const int64_t nthwc,
|
|
const int64_t thwc,
|
|
const int64_t hwc,
|
|
const int t,
|
|
const int c,
|
|
const int c1,
|
|
const int c2) {
|
|
int src_it = 0;
|
|
for (int64_t i = 0; i < nthwc; i++) {
|
|
int64_t it = (i % thwc) / hwc;
|
|
int64_t ic = i % c;
|
|
|
|
if (ic < c1) {
|
|
src_it = it + 1;
|
|
} else if (ic < c2) {
|
|
src_it = it - 1;
|
|
} else {
|
|
src_it = it;
|
|
}
|
|
|
|
if (src_it >= 0 && src_it < t) {
|
|
input_grad[i] = output_grad[i + (src_it - it) * hwc];
|
|
} else {
|
|
input_grad[i] = 0;
|
|
}
|
|
}
|
|
}
|
|
|
|
template <typename T, typename Context>
|
|
void TemporalShiftGradKernel(const Context& dev_ctx,
|
|
const DenseTensor& out_grad,
|
|
int seg_num,
|
|
float shift_ratio,
|
|
const std::string& data_format_str,
|
|
DenseTensor* x_grad) {
|
|
if (x_grad && x_grad->numel() == 0) {
|
|
dev_ctx.template Alloc<T>(x_grad);
|
|
return;
|
|
}
|
|
auto* input_grad = x_grad;
|
|
auto* output_grad = &out_grad;
|
|
int t = seg_num;
|
|
const DataLayout data_layout = StringToDataLayout(data_format_str);
|
|
|
|
const int nt = static_cast<int>(output_grad->dims()[0]);
|
|
const int c = static_cast<int>(data_layout == DataLayout::NCHW
|
|
? output_grad->dims()[1]
|
|
: output_grad->dims()[3]);
|
|
const int h = static_cast<int>(data_layout == DataLayout::NCHW
|
|
? output_grad->dims()[2]
|
|
: output_grad->dims()[1]);
|
|
const int w = static_cast<int>(data_layout == DataLayout::NCHW
|
|
? output_grad->dims()[3]
|
|
: output_grad->dims()[2]);
|
|
|
|
const int64_t hw = static_cast<int64_t>(h) * w;
|
|
const int64_t chw = static_cast<int64_t>(c) * hw;
|
|
const int64_t tchw = static_cast<int64_t>(t) * chw;
|
|
const int64_t ntchw = static_cast<int64_t>(nt) * chw;
|
|
|
|
const int c1 = static_cast<int>(static_cast<float>(c) * shift_ratio);
|
|
const int c2 = static_cast<int>(static_cast<float>(c) * 2.f * shift_ratio);
|
|
|
|
DDim in_grad_dims =
|
|
(data_layout == DataLayout::NCHW ? make_ddim({nt, c, h, w})
|
|
: make_ddim({nt, h, w, c}));
|
|
const T* output_grad_data = output_grad->data<T>();
|
|
input_grad->Resize(in_grad_dims);
|
|
|
|
T* input_grad_data = dev_ctx.template Alloc<T>(input_grad);
|
|
|
|
if (data_layout == DataLayout::NCHW) {
|
|
TemporalShiftBwNCHW<T>(
|
|
output_grad_data, input_grad_data, ntchw, tchw, chw, hw, t, c1, c2);
|
|
} else {
|
|
TemporalShiftBwNHWC<T>(
|
|
output_grad_data, input_grad_data, ntchw, tchw, chw, t, c, c1, c2);
|
|
}
|
|
}
|
|
|
|
} // namespace phi
|
|
|
|
PD_REGISTER_KERNEL(temporal_shift_grad,
|
|
CPU,
|
|
ALL_LAYOUT,
|
|
phi::TemporalShiftGradKernel,
|
|
float,
|
|
double) {}
|