wehub-resource-sync
117 lines
3.6 KiB
C++
117 lines
3.6 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/bmm_grad_kernel.h"
|
|
|
|
#include "paddle/phi/kernels/full_kernel.h"
|
|
#include "paddle/phi/kernels/xpu/bmm_xpu_utils.h"
|
|
|
|
namespace phi {
|
|
|
|
template <typename T, typename Context>
|
|
void MatMul(const Context& dev_ctx,
|
|
const DenseTensor& a,
|
|
bool trans_a,
|
|
const DenseTensor& b,
|
|
bool trans_b,
|
|
DenseTensor* out) {
|
|
using XPUType = typename XPUTypeTrait<T>::Type;
|
|
dev_ctx.template Alloc<T>(out);
|
|
xpu::Context* xpu_ctx = dev_ctx.x_context();
|
|
int fc_calc_type = FCCalcType<XPUType>();
|
|
if (fc_calc_type == XPUFCCalcType::FC_INT32) {
|
|
MatMulXPUFunction<T, int32_t>(a, b, out, trans_a, trans_b, xpu_ctx);
|
|
} else if (fc_calc_type == XPUFCCalcType::FC_FLOAT) {
|
|
MatMulXPUFunction<T, float>(a, b, out, trans_a, trans_b, xpu_ctx);
|
|
} else if (fc_calc_type == XPUFCCalcType::FC_INT32_WITH_LL) {
|
|
MatMulXPUFunction<T, int_with_ll_t>(a, b, out, trans_a, trans_b, xpu_ctx);
|
|
} else if (fc_calc_type == XPUFCCalcType::FC_FLOAT16) {
|
|
MatMulXPUFunction<T, float16>(a, b, out, trans_a, trans_b, xpu_ctx);
|
|
} else {
|
|
MatMulXPUFunction<T, int16_t>(a, b, out, trans_a, trans_b, xpu_ctx);
|
|
}
|
|
}
|
|
|
|
template <typename T, typename Context>
|
|
void CalcInputGrad(const Context& dev_ctx,
|
|
const DenseTensor& a,
|
|
bool trans_a,
|
|
const DenseTensor& b,
|
|
bool trans_b,
|
|
DenseTensor* out) {
|
|
if (out == nullptr) return;
|
|
MatMul<T, Context>(dev_ctx, a, trans_a, b, trans_b, out);
|
|
}
|
|
|
|
template <typename T, typename Context>
|
|
void BmmGradKernel(const Context& dev_ctx,
|
|
const DenseTensor& x,
|
|
const DenseTensor& y,
|
|
const DenseTensor& out_grad,
|
|
DenseTensor* x_grad,
|
|
DenseTensor* y_grad) {
|
|
if (x_grad && x_grad->numel() == 0) {
|
|
dev_ctx.template Alloc<T>(x_grad);
|
|
Full<T, Context>(dev_ctx, y.dims(), 0, y_grad);
|
|
return;
|
|
}
|
|
if (y_grad && y_grad->numel() == 0) {
|
|
dev_ctx.template Alloc<T>(y_grad);
|
|
Full<T, Context>(dev_ctx, x.dims(), 0, x_grad);
|
|
return;
|
|
}
|
|
DenseTensor x_help = x;
|
|
DenseTensor y_help = y;
|
|
DenseTensor out_grad_help = out_grad;
|
|
ReshapeXYOutIntoMatrixSequence(
|
|
&x_help, &y_help, &out_grad_help, false, false);
|
|
|
|
DDim dx_dims;
|
|
if (x_grad) {
|
|
dx_dims = x_grad->dims();
|
|
if (dx_dims != x_help.dims()) {
|
|
x_grad->Resize(x_help.dims());
|
|
}
|
|
}
|
|
|
|
DDim dy_dims;
|
|
if (y_grad) {
|
|
dy_dims = y_grad->dims();
|
|
if (dy_dims != y_help.dims()) {
|
|
y_grad->Resize(y_help.dims());
|
|
}
|
|
}
|
|
|
|
CalcInputGrad<T, Context>(
|
|
dev_ctx, out_grad_help, false, y_help, true, x_grad);
|
|
CalcInputGrad<T, Context>(
|
|
dev_ctx, x_help, true, out_grad_help, false, y_grad);
|
|
|
|
if (x_grad) {
|
|
if (dx_dims != x_help.dims()) {
|
|
x_grad->Resize(dx_dims);
|
|
}
|
|
}
|
|
if (y_grad) {
|
|
if (dy_dims != y_help.dims()) {
|
|
y_grad->Resize(dy_dims);
|
|
}
|
|
}
|
|
}
|
|
|
|
} // namespace phi
|
|
|
|
PD_REGISTER_KERNEL(
|
|
bmm_grad, XPU, ALL_LAYOUT, phi::BmmGradKernel, float, phi::float16) {}
|