318 lines
10 KiB
C++
318 lines
10 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/kernels/full_kernel.h"
|
|
#include "paddle/phi/kernels/impl/kron_kernel_impl.h"
|
|
#include "paddle/phi/kernels/reduce_sum_kernel.h"
|
|
namespace phi {
|
|
|
|
template <typename T>
|
|
struct KronGradElemFunctor {
|
|
KronGradElemFunctor(const T *dout,
|
|
const T *A,
|
|
const T *B,
|
|
T *dout_a,
|
|
T *dout_b,
|
|
const int64_t *stride_dout,
|
|
const int64_t *stride_a,
|
|
const int64_t *stride_b,
|
|
const int64_t *shape_b,
|
|
const int64_t numel_a,
|
|
const int64_t numel_b,
|
|
const int ndims)
|
|
: dout_(dout),
|
|
A_(A),
|
|
B_(B),
|
|
dout_a_(dout_a),
|
|
dout_b_(dout_b),
|
|
stride_dout_(stride_dout),
|
|
stride_a_(stride_a),
|
|
stride_b_(stride_b),
|
|
shape_b_(shape_b),
|
|
numel_a_(numel_a),
|
|
numel_b_(numel_b),
|
|
ndims_(ndims) {}
|
|
|
|
HOSTDEVICE void operator()(int64_t idx) {
|
|
int64_t index = idx;
|
|
int64_t index_a = 0;
|
|
int64_t index_b = 0;
|
|
for (int i = 0; i < ndims_; i++) {
|
|
auto pos_i = index / stride_dout_[i];
|
|
index = index % stride_dout_[i];
|
|
auto pos_ai = pos_i / shape_b_[i];
|
|
auto pos_bi = pos_i % shape_b_[i];
|
|
index_a += stride_a_[i] * pos_ai;
|
|
index_b += stride_b_[i] * pos_bi;
|
|
}
|
|
|
|
if (dout_a_) {
|
|
size_t index_out_a = index_a * numel_b_ + index_b;
|
|
dout_a_[index_out_a] = dout_[idx] * B_[index_b];
|
|
}
|
|
if (dout_b_) {
|
|
size_t index_out_b = index_b * numel_a_ + index_a;
|
|
dout_b_[index_out_b] = dout_[idx] * A_[index_a];
|
|
}
|
|
}
|
|
|
|
private:
|
|
const T *dout_;
|
|
const T *A_;
|
|
const T *B_;
|
|
T *dout_a_;
|
|
T *dout_b_;
|
|
const int64_t *stride_dout_;
|
|
const int64_t *stride_a_;
|
|
const int64_t *stride_b_;
|
|
const int64_t *shape_b_;
|
|
const int64_t numel_a_;
|
|
const int64_t numel_b_;
|
|
const int ndims_;
|
|
};
|
|
|
|
template <typename T>
|
|
struct KronGradElemFunctor<dtype::complex<T>> {
|
|
KronGradElemFunctor(const dtype::complex<T> *dout,
|
|
const dtype::complex<T> *A,
|
|
const dtype::complex<T> *B,
|
|
dtype::complex<T> *dout_a,
|
|
dtype::complex<T> *dout_b,
|
|
const int64_t *stride_dout,
|
|
const int64_t *stride_a,
|
|
const int64_t *stride_b,
|
|
const int64_t *shape_b,
|
|
const int64_t numel_a,
|
|
const int64_t numel_b,
|
|
const int ndims)
|
|
: dout_(dout),
|
|
A_(A),
|
|
B_(B),
|
|
dout_a_(dout_a),
|
|
dout_b_(dout_b),
|
|
stride_dout_(stride_dout),
|
|
stride_a_(stride_a),
|
|
stride_b_(stride_b),
|
|
shape_b_(shape_b),
|
|
numel_a_(numel_a),
|
|
numel_b_(numel_b),
|
|
ndims_(ndims) {}
|
|
|
|
HOSTDEVICE void operator()(int64_t idx) {
|
|
int64_t index = idx;
|
|
int64_t index_a = 0;
|
|
int64_t index_b = 0;
|
|
for (int i = 0; i < ndims_; i++) {
|
|
auto pos_i = index / stride_dout_[i];
|
|
index = index % stride_dout_[i];
|
|
auto pos_ai = pos_i / shape_b_[i];
|
|
auto pos_bi = pos_i % shape_b_[i];
|
|
index_a += stride_a_[i] * pos_ai;
|
|
index_b += stride_b_[i] * pos_bi;
|
|
}
|
|
|
|
if (dout_a_) {
|
|
size_t index_out_a = index_a * numel_b_ + index_b;
|
|
dout_a_[index_out_a] =
|
|
dout_[idx] * dtype::complex<T>(B_[index_b].real, -B_[index_b].imag);
|
|
}
|
|
if (dout_b_) {
|
|
size_t index_out_b = index_b * numel_a_ + index_a;
|
|
dout_b_[index_out_b] =
|
|
dout_[idx] * dtype::complex<T>(A_[index_a].real, -A_[index_a].imag);
|
|
}
|
|
}
|
|
|
|
private:
|
|
const dtype::complex<T> *dout_;
|
|
const dtype::complex<T> *A_;
|
|
const dtype::complex<T> *B_;
|
|
dtype::complex<T> *dout_a_;
|
|
dtype::complex<T> *dout_b_;
|
|
const int64_t *stride_dout_;
|
|
const int64_t *stride_a_;
|
|
const int64_t *stride_b_;
|
|
const int64_t *shape_b_;
|
|
const int64_t numel_a_;
|
|
const int64_t numel_b_;
|
|
const int ndims_;
|
|
};
|
|
|
|
template <typename Context, typename T>
|
|
struct KronGradOpFunctor {
|
|
void operator()(const Context &dev_ctx,
|
|
const DenseTensor &dout,
|
|
const DenseTensor &x,
|
|
const DenseTensor &y,
|
|
DenseTensor *dx,
|
|
DenseTensor *dy) {
|
|
int ndims = dout.dims().size();
|
|
int64_t numel = dout.numel();
|
|
int64_t numel_x = x.numel();
|
|
int64_t numel_y = y.numel();
|
|
|
|
const DDim &dim_x = x.dims();
|
|
const DDim &dim_y = y.dims();
|
|
const DDim &dim_dout = dout.dims();
|
|
const DDim stride_x =
|
|
dim_x.size() == 0 ? DDim(dim_x) : common::stride(dim_x);
|
|
const DDim stride_y =
|
|
dim_y.size() == 0 ? DDim(dim_y) : common::stride(dim_y);
|
|
const DDim stride_dout =
|
|
dim_dout.size() == 0 ? DDim(dim_dout) : common::stride(dim_dout);
|
|
|
|
const int64_t *p_stride_x = nullptr;
|
|
const int64_t *p_stride_y = nullptr;
|
|
const int64_t *p_stride_dout = nullptr;
|
|
const int64_t *p_shape_y = nullptr;
|
|
#if defined(__NVCC__) || defined(__HIPCC__)
|
|
thrust::device_vector<int64_t> d_stride_x(ndims);
|
|
thrust::device_vector<int64_t> d_stride_y(ndims);
|
|
thrust::device_vector<int64_t> d_stride_dout(ndims);
|
|
thrust::device_vector<int64_t> d_shape_y(ndims);
|
|
thrust::copy(stride_x.Get(), stride_x.Get() + ndims, d_stride_x.begin());
|
|
thrust::copy(stride_y.Get(), stride_y.Get() + ndims, d_stride_y.begin());
|
|
thrust::copy(
|
|
stride_dout.Get(), stride_dout.Get() + ndims, d_stride_dout.begin());
|
|
thrust::copy(dim_y.Get(), dim_y.Get() + ndims, d_shape_y.begin());
|
|
|
|
p_stride_x = thrust::raw_pointer_cast(d_stride_x.data());
|
|
p_stride_y = thrust::raw_pointer_cast(d_stride_y.data());
|
|
p_stride_dout = thrust::raw_pointer_cast(d_stride_dout.data());
|
|
p_shape_y = thrust::raw_pointer_cast(d_shape_y.data());
|
|
#else
|
|
p_stride_x = stride_x.Get();
|
|
p_stride_y = stride_y.Get();
|
|
p_stride_dout = stride_dout.Get();
|
|
p_shape_y = dim_y.Get();
|
|
#endif
|
|
// dout_x: dout * kron(ones(X), Y) re-arranged in shape (numel_x, numel_y)
|
|
// dout_y: dout * kron(X, ones(Y)) re-arranged in shape (numel_y, numel_x)
|
|
DenseTensor dout_x;
|
|
T *p_dout_x = nullptr;
|
|
if (dx) {
|
|
dout_x.Resize({numel_x, numel_y});
|
|
dev_ctx.template Alloc<T>(&dout_x);
|
|
p_dout_x = dout_x.data<T>();
|
|
}
|
|
DenseTensor dout_y;
|
|
T *p_dout_y = nullptr;
|
|
if (dy) {
|
|
dout_y.Resize({numel_y, numel_x});
|
|
dev_ctx.template Alloc<T>(&dout_y);
|
|
p_dout_y = dout_y.data<T>();
|
|
}
|
|
|
|
funcs::ForRange<Context> for_range(dev_ctx, numel);
|
|
KronGradElemFunctor<T> func(dout.data<T>(),
|
|
x.data<T>(),
|
|
y.data<T>(),
|
|
p_dout_x,
|
|
p_dout_y,
|
|
p_stride_dout,
|
|
p_stride_x,
|
|
p_stride_y,
|
|
p_shape_y,
|
|
numel_x,
|
|
numel_y,
|
|
ndims);
|
|
for_range(func);
|
|
|
|
// reduce_sum along axis 1
|
|
#if defined(__NVCC__) || defined(__HIPCC__)
|
|
auto stream = dev_ctx.stream(); // it is a cuda device_context
|
|
if (dx) {
|
|
SumKernel<T, Context>(dev_ctx, dout_x, {1}, dout_x.dtype(), false, dx);
|
|
}
|
|
if (dy) {
|
|
SumKernel<T, Context>(dev_ctx, dout_y, {1}, dout_y.dtype(), false, dy);
|
|
}
|
|
#else
|
|
auto *place = dev_ctx.eigen_device();
|
|
Eigen::array<int, 1> reduce_dim = {1};
|
|
if (dx) {
|
|
auto eigen_dout_x = EigenMatrix<T>::Reshape(dout_x, 1);
|
|
auto eigen_vec_dx = EigenVector<T>::Flatten(*dx);
|
|
if constexpr (std::is_same_v<T, float16> || std::is_same_v<T, bfloat16>) {
|
|
eigen_vec_dx.device(*place) = eigen_dout_x.template cast<float>()
|
|
.sum(reduce_dim)
|
|
.template cast<T>();
|
|
} else {
|
|
eigen_vec_dx.device(*place) = eigen_dout_x.sum(reduce_dim);
|
|
}
|
|
}
|
|
if (dy) {
|
|
auto eigen_dout_y = EigenMatrix<T>::Reshape(dout_y, 1);
|
|
auto eigen_vec_dy = EigenVector<T>::Flatten(*dy);
|
|
if constexpr (std::is_same_v<T, float16> || std::is_same_v<T, bfloat16>) {
|
|
eigen_vec_dy.device(*place) = eigen_dout_y.template cast<float>()
|
|
.sum(reduce_dim)
|
|
.template cast<T>();
|
|
} else {
|
|
eigen_vec_dy.device(*place) = eigen_dout_y.sum(reduce_dim);
|
|
}
|
|
}
|
|
#endif
|
|
}
|
|
};
|
|
|
|
template <typename T, typename Context>
|
|
void KronGradKernel(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) {
|
|
Full<T, Context>(dev_ctx, x_grad->dims(), 0, x_grad);
|
|
}
|
|
if (y_grad) {
|
|
Full<T, Context>(dev_ctx, y_grad->dims(), 0, y_grad);
|
|
}
|
|
return;
|
|
}
|
|
if (x_grad) {
|
|
dev_ctx.template Alloc<T>(x_grad);
|
|
}
|
|
if (y_grad) {
|
|
dev_ctx.template Alloc<T>(y_grad);
|
|
}
|
|
|
|
int ndims = out_grad.dims().size();
|
|
DenseTensor xx = UnsqueezeTo(x, ndims);
|
|
DenseTensor yy = UnsqueezeTo(y, ndims);
|
|
|
|
DenseTensor *pdxx = nullptr;
|
|
DenseTensor *pdyy = nullptr;
|
|
DenseTensor dxx;
|
|
DenseTensor dyy;
|
|
if (x_grad) {
|
|
dxx = UnsqueezeTo(*x_grad, ndims);
|
|
pdxx = &dxx;
|
|
}
|
|
|
|
if (y_grad) {
|
|
dyy = UnsqueezeTo(*y_grad, ndims);
|
|
pdyy = &dyy;
|
|
}
|
|
|
|
KronGradOpFunctor<Context, T> func;
|
|
func(dev_ctx, out_grad, xx, yy, pdxx, pdyy);
|
|
}
|
|
|
|
} // namespace phi
|