200 lines
6.1 KiB
C++
200 lines
6.1 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.
|
|
|
|
#pragma once
|
|
|
|
#include "paddle/phi/core/dense_tensor.h"
|
|
#include "paddle/phi/core/tensor_utils.h"
|
|
#include "paddle/phi/kernels/atan2_grad_kernel.h"
|
|
#include "paddle/phi/kernels/broadcast_tensors_kernel.h"
|
|
#include "paddle/phi/kernels/full_kernel.h"
|
|
#include "paddle/phi/kernels/funcs/for_range.h"
|
|
#include "paddle/phi/kernels/reduce_sum_kernel.h"
|
|
|
|
namespace phi {
|
|
|
|
// dx1 = dout * x2 / ((x1)^2 + (x2)^2)
|
|
// dx2 = - dout * x1 / ((x1)^2 + (x2)^2)
|
|
template <typename T>
|
|
struct Atan2GradFunctor {
|
|
Atan2GradFunctor(
|
|
const T* x1, const T* x2, const T* dout, T* dx1, T* dx2, int64_t numel)
|
|
: x1_(x1), x2_(x2), dout_(dout), dx1_(dx1), dx2_(dx2), numel_(numel) {}
|
|
|
|
HOSTDEVICE void operator()(int64_t idx) const {
|
|
float x1 = static_cast<float>(x1_[idx]);
|
|
float x2 = static_cast<float>(x2_[idx]);
|
|
float x = x1 * x1 + x2 * x2;
|
|
if (dx1_) {
|
|
dx1_[idx] = static_cast<T>(static_cast<float>(dout_[idx]) * x2 / x);
|
|
}
|
|
if (dx2_) {
|
|
dx2_[idx] = static_cast<T>(-static_cast<float>(dout_[idx]) * x1 / x);
|
|
}
|
|
}
|
|
|
|
const T* x1_;
|
|
const T* x2_;
|
|
const T* dout_;
|
|
T* dx1_;
|
|
T* dx2_;
|
|
int64_t numel_;
|
|
};
|
|
|
|
template <>
|
|
struct Atan2GradFunctor<double> {
|
|
Atan2GradFunctor(const double* x1,
|
|
const double* x2,
|
|
const double* dout,
|
|
double* dx1,
|
|
double* dx2,
|
|
int64_t numel)
|
|
: x1_(x1), x2_(x2), dout_(dout), dx1_(dx1), dx2_(dx2), numel_(numel) {}
|
|
|
|
HOSTDEVICE void operator()(int64_t idx) const {
|
|
auto x = x1_[idx] * x1_[idx] + x2_[idx] * x2_[idx];
|
|
if (dx1_) {
|
|
dx1_[idx] = dout_[idx] * x2_[idx] / x;
|
|
}
|
|
if (dx2_) {
|
|
dx2_[idx] = -dout_[idx] * x1_[idx] / x;
|
|
}
|
|
}
|
|
|
|
const double* x1_;
|
|
const double* x2_;
|
|
const double* dout_;
|
|
double* dx1_;
|
|
double* dx2_;
|
|
int64_t numel_;
|
|
};
|
|
|
|
template <typename T, typename Context>
|
|
void Atan2GradKernel(const Context& dev_ctx,
|
|
const DenseTensor& x,
|
|
const DenseTensor& y,
|
|
const DenseTensor& out_grad,
|
|
DenseTensor* x_grad,
|
|
DenseTensor* y_grad) {
|
|
if (out_grad.numel() == 0) {
|
|
if (x_grad) {
|
|
dev_ctx.template Alloc<T>(x_grad);
|
|
if (x_grad->numel() != 0) {
|
|
Full<T, Context>(dev_ctx, x_grad->dims(), 0, x_grad);
|
|
}
|
|
}
|
|
if (y_grad) {
|
|
dev_ctx.template Alloc<T>(y_grad);
|
|
if (y_grad->numel() != 0) {
|
|
Full<T, Context>(dev_ctx, y_grad->dims(), 0, y_grad);
|
|
}
|
|
}
|
|
return;
|
|
}
|
|
|
|
if (x.dims() == y.dims() && x.dims() == out_grad.dims()) {
|
|
auto numel = x.numel();
|
|
auto x_data = x.data<T>();
|
|
auto y_data = y.data<T>();
|
|
auto out_grad_data = out_grad.data<T>();
|
|
|
|
auto* x_grad_data = x_grad ? dev_ctx.template Alloc<T>(
|
|
x_grad, size_t(x.numel() * sizeof(T)))
|
|
: nullptr;
|
|
auto* y_grad_data = y_grad ? dev_ctx.template Alloc<T>(
|
|
y_grad, size_t(y.numel() * sizeof(T)))
|
|
: nullptr;
|
|
|
|
funcs::ForRange<Context> for_range(dev_ctx, numel);
|
|
Atan2GradFunctor<T> functor(
|
|
x_data, y_data, out_grad_data, x_grad_data, y_grad_data, numel);
|
|
for_range(functor);
|
|
} else {
|
|
DenseTensor b_x, b_y;
|
|
b_x.Resize(out_grad.dims());
|
|
b_y.Resize(out_grad.dims());
|
|
|
|
std::vector<const DenseTensor*> inputs = {&x, &y};
|
|
std::vector<DenseTensor*> outputs = {&b_x, &b_y};
|
|
BroadcastTensorsKernel<T, Context>(dev_ctx, inputs, outputs);
|
|
|
|
DenseTensor dx_b, dy_b;
|
|
T* dx_b_data = nullptr;
|
|
T* dy_b_data = nullptr;
|
|
std::vector<int64_t> x_axes, y_axes;
|
|
|
|
if (x_grad) {
|
|
int in_rank = x.dims().size();
|
|
int out_rank = out_grad.dims().size();
|
|
int diff = out_rank - in_rank;
|
|
for (int i = 0; i < diff; ++i) x_axes.push_back(i);
|
|
for (int i = 0; i < in_rank; ++i) {
|
|
if (x.dims()[i] == 1 && out_grad.dims()[i + diff] > 1) {
|
|
x_axes.push_back(i + diff);
|
|
}
|
|
}
|
|
if (x_axes.empty()) {
|
|
dev_ctx.template Alloc<T>(x_grad);
|
|
dx_b_data = x_grad->data<T>();
|
|
} else {
|
|
dx_b.Resize(out_grad.dims());
|
|
dx_b_data = dev_ctx.template Alloc<T>(&dx_b);
|
|
}
|
|
}
|
|
|
|
if (y_grad) {
|
|
int in_rank = y.dims().size();
|
|
int out_rank = out_grad.dims().size();
|
|
int diff = out_rank - in_rank;
|
|
for (int i = 0; i < diff; ++i) y_axes.push_back(i);
|
|
for (int i = 0; i < in_rank; ++i) {
|
|
if (y.dims()[i] == 1 && out_grad.dims()[i + diff] > 1) {
|
|
y_axes.push_back(i + diff);
|
|
}
|
|
}
|
|
if (y_axes.empty()) {
|
|
dev_ctx.template Alloc<T>(y_grad);
|
|
dy_b_data = y_grad->data<T>();
|
|
} else {
|
|
dy_b.Resize(out_grad.dims());
|
|
dy_b_data = dev_ctx.template Alloc<T>(&dy_b);
|
|
}
|
|
}
|
|
|
|
auto numel = out_grad.numel();
|
|
funcs::ForRange<Context> for_range(dev_ctx, numel);
|
|
Atan2GradFunctor<T> functor(b_x.data<T>(),
|
|
b_y.data<T>(),
|
|
out_grad.data<T>(),
|
|
dx_b_data,
|
|
dy_b_data,
|
|
numel);
|
|
for_range(functor);
|
|
|
|
if (x_grad && !x_axes.empty()) {
|
|
SumKernel<T, Context>(
|
|
dev_ctx, dx_b, IntArray(x_axes), x_grad->dtype(), false, x_grad);
|
|
x_grad->Resize(x.dims());
|
|
}
|
|
|
|
if (y_grad && !y_axes.empty()) {
|
|
SumKernel<T, Context>(
|
|
dev_ctx, dy_b, IntArray(y_axes), y_grad->dtype(), false, y_grad);
|
|
y_grad->Resize(y.dims());
|
|
}
|
|
}
|
|
}
|
|
|
|
} // namespace phi
|