Files
paddlepaddle--paddle/paddle/phi/kernels/impl/lu_grad_kernel_impl.h
T
2026-07-13 12:40:42 +08:00

308 lines
12 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/funcs/blas/blas.h"
#include "paddle/phi/kernels/funcs/math_function.h"
#include "paddle/phi/kernels/triangular_solve_kernel.h"
#include "paddle/phi/kernels/impl/lu_kernel_impl.h"
namespace phi {
template <typename T, typename Context>
void LUGradKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& out,
const DenseTensor& pivots,
const DenseTensor& out_grad,
bool pivot UNUSED,
DenseTensor* x_grad) {
dev_ctx.template Alloc<T>(x_grad);
if (x_grad->numel() == 0) return;
auto blas = funcs::GetBlas<Context, T>(dev_ctx);
auto xdims = x.dims();
int xrank = xdims.size();
int64_t m = xdims[xrank - 2];
int64_t n = xdims[xrank - 1];
int64_t k = std::min(m, n);
DenseTensor L, U, L_narrow, U_narrow, L_narrow_mH, U_narrow_mH, grad_narrow;
LU_Unpack<Context, T>(dev_ctx, &out, &L, &U);
Tensor_narrow<Context, T>(dev_ctx, &L, &L_narrow, 0, k, 0, k);
Tensor_narrow<Context, T>(dev_ctx, &U, &U_narrow, 0, k, 0, k);
Tensor_narrow<Context, T>(dev_ctx, &out_grad, &grad_narrow, 0, k, 0, k);
auto graddims = grad_narrow.dims();
Tensor_Conj<Context, T>(dev_ctx, L_narrow, &L_narrow_mH);
Tensor_Conj<Context, T>(dev_ctx, U_narrow, &U_narrow_mH);
L_narrow_mH = Transpose2DTo6D<Context, T>(dev_ctx, L_narrow_mH);
U_narrow_mH = Transpose2DTo6D<Context, T>(dev_ctx, U_narrow_mH);
auto LmHdims = L_narrow_mH.dims();
auto UmHdims = U_narrow_mH.dims();
DenseTensor phi_L, phi_U, phi, psi;
phi_L.Resize(LmHdims);
dev_ctx.template Alloc<T>(&phi_L);
phi_U.Resize(UmHdims);
dev_ctx.template Alloc<T>(&phi_U);
auto mat_dim_l = funcs::CreateMatrixDescriptor(LmHdims, 0, false);
auto mat_dim_u = funcs::CreateMatrixDescriptor(UmHdims, 0, false);
auto mat_dim_g = funcs::CreateMatrixDescriptor(graddims, 0, false);
blas.MatMul(L_narrow_mH,
mat_dim_l,
grad_narrow,
mat_dim_g,
static_cast<T>(1),
&phi_L,
static_cast<T>(0));
blas.MatMul(grad_narrow,
mat_dim_g,
U_narrow_mH,
mat_dim_u,
static_cast<T>(1),
&phi_U,
static_cast<T>(0));
auto phil_rank = LmHdims.size();
auto phiu_rank = UmHdims.size();
funcs::ForRange<Context> l_for_range(dev_ctx, phi_L.numel());
funcs::TrilTriuCompute<T> tril_computer(phi_L.data<T>(),
-1,
true,
LmHdims[phil_rank - 2],
LmHdims[phil_rank - 1],
phi_L.data<T>());
l_for_range(tril_computer);
funcs::ForRange<Context> u_for_range(dev_ctx, phi_U.numel());
funcs::TrilTriuCompute<T> triu_computer(phi_U.data<T>(),
0,
false,
UmHdims[phiu_rank - 2],
UmHdims[phiu_rank - 1],
phi_U.data<T>());
u_for_range(triu_computer);
Tensor_Add<Context, T>(dev_ctx, phi_L, phi_U, &phi);
psi.Resize(xdims);
dev_ctx.template Alloc<T>(&psi);
funcs::SetConstant<Context, T> setter;
setter(dev_ctx, &psi, static_cast<T>(0));
std::vector<int64_t> axes = {xrank - 2, xrank - 1};
std::vector<int64_t> slice_starts(2, 0);
std::vector<int64_t> slice_ends(2, 0);
auto valuedims = vectorize(xdims);
DenseTensor Pmat;
Unpack_Pivot<Context, T>(dev_ctx, pivots, &Pmat, m, k);
if (m <= n) {
if (k < n) {
DenseTensor U_complement, U_grad_complement, phi_complement,
phi_complement_l;
Tensor_narrow<Context, T>(dev_ctx, &U, &U_complement, 0, k, k, n);
Tensor_narrow<Context, T>(
dev_ctx, &out_grad, &U_grad_complement, 0, k, k, n);
DenseTensor U_complement_mH =
Transpose2DTo6D<Context, T>(dev_ctx, U_complement);
Tensor_Conj<Context, T>(dev_ctx, U_complement_mH, &U_complement_mH);
auto mat_dim_g =
funcs::CreateMatrixDescriptor(U_grad_complement.dims(), 0, false);
auto mat_dim_u =
funcs::CreateMatrixDescriptor(U_complement_mH.dims(), 0, false);
auto phidims = UmHdims;
phidims[UmHdims.size() - 2] = k;
phidims[UmHdims.size() - 1] = k;
phi_complement.Resize(phidims);
dev_ctx.template Alloc<T>(&phi_complement);
blas.MatMul(U_grad_complement,
mat_dim_g,
U_complement_mH,
mat_dim_u,
static_cast<T>(1),
&phi_complement,
static_cast<T>(0));
phi_complement_l.Resize(phidims);
dev_ctx.template Alloc<T>(&phi_complement_l);
const auto H = phidims[phidims.size() - 2];
const auto W = phidims[phidims.size() - 1];
funcs::ForRange<Context> x_for_range(dev_ctx, phi_complement.numel());
funcs::TrilTriuCompute<T> tril_computer(
phi_complement.data<T>(), -1, true, H, W, phi_complement_l.data<T>());
x_for_range(tril_computer);
Tensor_Sub<Context, T>(dev_ctx, phi, phi_complement_l, &phi);
slice_starts[0] = 0;
slice_starts[1] = k;
slice_ends[0] = k;
slice_ends[1] = n;
valuedims[xrank - 2] = k;
valuedims[xrank - 1] = n - k;
SetValueCompute_dispatch<Context, T>(dev_ctx,
&psi,
&U_grad_complement,
&psi,
axes,
&slice_starts,
&slice_ends,
valuedims,
xrank);
}
DenseTensor psi_principal, phi_mH, psi_tmp;
Tensor_Conj<Context, T>(dev_ctx, phi, &phi_mH);
phi_mH = Transpose2DTo6D<Context, T>(dev_ctx, phi_mH);
TriangularSolveKernel<T, Context>(
dev_ctx, U_narrow, phi_mH, true, false, false, &psi_principal);
Tensor_Conj<Context, T>(dev_ctx, psi_principal, &psi_principal);
psi_principal = Transpose2DTo6D<Context, T>(dev_ctx, psi_principal);
slice_starts[0] = 0;
slice_starts[1] = 0;
slice_ends[0] = k;
slice_ends[1] = k;
valuedims[xrank - 2] = k;
valuedims[xrank - 1] = k;
SetValueCompute_dispatch<Context, T>(dev_ctx,
&psi,
&psi_principal,
&psi,
axes,
&slice_starts,
&slice_ends,
valuedims,
xrank);
TriangularSolveKernel<T, Context>(
dev_ctx, L_narrow_mH, psi, true, false, true, &psi_tmp);
auto mat_dim_p = funcs::CreateMatrixDescriptor(Pmat.dims(), 0, false);
auto mat_dim_b = funcs::CreateMatrixDescriptor(psi_tmp.dims(), 0, false);
blas.MatMul(Pmat,
mat_dim_p,
psi_tmp,
mat_dim_b,
static_cast<T>(1),
x_grad,
static_cast<T>(0));
} else {
DenseTensor L_complement, L_grad_complement, phi_complement,
phi_complement_u;
Tensor_narrow<Context, T>(dev_ctx, &L, &L_complement, k, m, 0, k);
Tensor_narrow<Context, T>(
dev_ctx, &out_grad, &L_grad_complement, k, m, 0, k);
DenseTensor L_complement_mH =
Transpose2DTo6D<Context, T>(dev_ctx, L_complement);
Tensor_Conj<Context, T>(dev_ctx, L_complement_mH, &L_complement_mH);
auto mat_dim_g =
funcs::CreateMatrixDescriptor(L_grad_complement.dims(), 0, false);
auto mat_dim_u =
funcs::CreateMatrixDescriptor(L_complement_mH.dims(), 0, false);
auto phidims = LmHdims;
phidims[LmHdims.size() - 2] = k;
phidims[LmHdims.size() - 1] = k;
phi_complement.Resize(phidims);
dev_ctx.template Alloc<T>(&phi_complement);
blas.MatMul(L_complement_mH,
mat_dim_u,
L_grad_complement,
mat_dim_g,
static_cast<T>(1),
&phi_complement,
static_cast<T>(0));
phi_complement_u.Resize(phidims);
dev_ctx.template Alloc<T>(&phi_complement_u);
const auto H = phidims[phidims.size() - 2];
const auto W = phidims[phidims.size() - 1];
funcs::ForRange<Context> x_for_range(dev_ctx, phi_complement.numel());
funcs::TrilTriuCompute<T> triu_computer(
phi_complement.data<T>(), 0, false, H, W, phi_complement_u.data<T>());
x_for_range(triu_computer);
Tensor_Sub<Context, T>(dev_ctx, phi, phi_complement_u, &phi);
slice_starts[0] = k;
slice_starts[1] = 0;
slice_ends[0] = m;
slice_ends[1] = k;
valuedims[xrank - 2] = m - k;
valuedims[xrank - 1] = k;
SetValueCompute_dispatch<Context, T>(dev_ctx,
&psi,
&L_grad_complement,
&psi,
axes,
&slice_starts,
&slice_ends,
valuedims,
xrank);
DenseTensor psi_principal, phi_mH, psi_tmp, U_narrow_mH;
TriangularSolveKernel<T, Context>(
dev_ctx, L_narrow_mH, phi, true, false, true, &psi_principal);
slice_starts[0] = 0;
slice_starts[1] = 0;
slice_ends[0] = k;
slice_ends[1] = k;
valuedims[xrank - 2] = k;
valuedims[xrank - 1] = k;
SetValueCompute_dispatch<Context, T>(dev_ctx,
&psi,
&psi_principal,
&psi,
axes,
&slice_starts,
&slice_ends,
valuedims,
xrank);
psi_tmp.Resize(psi.dims());
dev_ctx.template Alloc<T>(&psi_tmp);
auto mat_dim_p = funcs::CreateMatrixDescriptor(Pmat.dims(), 0, false);
auto mat_dim_b = funcs::CreateMatrixDescriptor(psi.dims(), 0, false);
blas.MatMul(Pmat,
mat_dim_p,
psi,
mat_dim_b,
static_cast<T>(1),
&psi_tmp,
static_cast<T>(0));
psi_tmp = Transpose2DTo6D<Context, T>(dev_ctx, psi_tmp);
Tensor_Conj<Context, T>(dev_ctx, U_narrow, &U_narrow_mH);
TriangularSolveKernel<T, Context>(
dev_ctx, U_narrow_mH, psi_tmp, true, false, false, &psi);
*x_grad = Transpose2DTo6D<Context, T>(dev_ctx, psi);
}
}
} // namespace phi