Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 8add11a0 编写于 作者: L Leo Guo 提交者： GitHub
浏览文件
操作

Migrate roi_align and roi_align_grad to phi. test=kunlun (#45858)

上级 e56a2853
隐藏空白更改
内联 并排
Showing with 263 addition and 278 deletion
paddle/fluid/operators/roi_align_op_xpu.cc 已删除 100644 → 0
浏览文件 @ e56a2853
/* Copyright (c) 2016 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. */
#ifdef PADDLE_WITH_XPU
#include <memory>
#include <string>
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
template <typename DeviceContext, typename T>
class XPUROIAlignOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* in = ctx.Input<Tensor>("X");
auto* rois = ctx.Input<LoDTensor>("ROIs");
auto* out = ctx.Output<Tensor>("Out");
auto pooled_height = ctx.Attr<int>("pooled_height");
auto pooled_width = ctx.Attr<int>("pooled_width");
auto spatial_scale = ctx.Attr<float>("spatial_scale");
auto sampling_ratio = ctx.Attr<int>("sampling_ratio");
auto aligned = ctx.Attr<bool>("aligned");
const auto& in_dims = in->dims();
int batch_size = in_dims[0];
int channels = in_dims[1];
int height = in_dims[2];
int width = in_dims[3];
int rois_num = rois->dims()[0];
if (rois_num == 0) return;
Tensor roi_batch_id_list;
roi_batch_id_list.Resize({rois_num});
auto cplace = platform::CPUPlace();
int* roi_batch_id_data = roi_batch_id_list.mutable_data<int>(cplace);
auto& dev_ctx = ctx.template device_context<DeviceContext>();
auto xplace = ctx.GetPlace();
int rois_batch_size = 0;
int* cpu_lod = nullptr;
if (ctx.HasInput("RoisNum")) {
auto* rois_num_t = ctx.Input<Tensor>("RoisNum");
rois_batch_size = rois_num_t->numel();
PADDLE_ENFORCE_EQ(
rois_batch_size,
batch_size,
platform::errors::InvalidArgument(
"The rois_batch_size and imgs "
"batch_size must be the same. But received rois_batch_size = %d, "
"batch_size = %d",
rois_batch_size,
batch_size));
std::vector<int> rois_num_list(rois_batch_size);
memory::Copy(cplace,
rois_num_list.data(),
xplace,
rois_num_t->data<int>(),
sizeof(int) * rois_batch_size);
cpu_lod = new int[rois_batch_size + 1];
cpu_lod[0] = 0;
for (int i = 0; i < rois_batch_size; i++) {
cpu_lod[i + 1] = cpu_lod[i] + rois_num_list[i];
}
} else {
auto lod = rois->lod();
PADDLE_ENFORCE_EQ(
lod.empty(),
false,
platform::errors::InvalidArgument("Input(ROIs) in ROIAlignOp does "
"not contain LoD information."));
auto rois_lod = lod.back();
rois_batch_size = rois_lod.size() - 1;
PADDLE_ENFORCE_EQ(
rois_batch_size,
batch_size,
platform::errors::InvalidArgument(
"The batch size of rois and batch size "
"of images must be the same. But received rois batch size = %d, "
"and images batch size = %d",
rois_batch_size,
batch_size));
int rois_num_with_lod = rois_lod[rois_batch_size];
PADDLE_ENFORCE_EQ(
rois_num,
rois_num_with_lod,
platform::errors::InvalidArgument(
"The actual number of rois and the number of rois "
"provided from Input(RoIsLoD) in RoIAlign must be the same."
" But received actual number of rois is %d, and the number "
"of rois from RoIsLoD is %d",
rois_num,
rois_num_with_lod));
for (int n = 0; n < rois_batch_size; ++n) {
for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) {
roi_batch_id_data[i] = n;
}
}
cpu_lod = new int[rois_batch_size + 1];
for (int i = 0; i < rois_batch_size + 1; i++) {
cpu_lod[i] = rois_lod[i];
}
}
int* roi_id_data = nullptr;
int r = xpu_malloc(reinterpret_cast<void**>(&roi_id_data),
(rois_batch_size + 1) * sizeof(int));
PADDLE_ENFORCE_EQ(r,
xpu::Error_t::SUCCESS,
platform::errors::External("no enough memory in xpu"));
memory::Copy(xplace,
roi_id_data,
cplace,
cpu_lod,
(rois_batch_size + 1) * sizeof(int));
delete[] cpu_lod;
r = xpu::roi_align<T, int>(dev_ctx.x_context(),
in->data<T>(),
out->mutable_data<T>(ctx.GetPlace()),
rois->data<T>(),
roi_id_data,
batch_size,
channels,
height,
width,
out->dims()[0],
pooled_height,
pooled_width,
spatial_scale,
sampling_ratio,
true,
aligned);
PADDLE_ENFORCE_EQ(r,
xpu::Error_t::SUCCESS,
platform::errors::External(
"The roi_align XPU OP return wrong value[%d %s]",
r,
XPUAPIErrorMsg[r]));
if (dev_ctx.x_context()->xpu_stream) {
dev_ctx.Wait();
}
xpu_free(roi_id_data);
}
};
template <typename DeviceContext, typename T>
class XPUROIAlignGradOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* in = ctx.Input<Tensor>("X");
auto* rois = ctx.Input<LoDTensor>("ROIs");
auto* out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
auto* in_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
auto pooled_height = ctx.Attr<int>("pooled_height");
auto pooled_width = ctx.Attr<int>("pooled_width");
auto spatial_scale = ctx.Attr<float>("spatial_scale");
auto sampling_ratio = ctx.Attr<int>("sampling_ratio");
auto aligned = ctx.Attr<bool>("aligned");
int rois_num = rois->dims()[0];
int channels = in->dims()[1];
int height = in->dims()[2];
int width = in->dims()[3];
if (!in_grad) {
return;
}
Tensor roi_batch_id_list;
roi_batch_id_list.Resize({rois_num});
auto cplace = platform::CPUPlace();
auto& dev_ctx = ctx.template device_context<DeviceContext>();
auto xplace = ctx.GetPlace();
int rois_batch_size = 0;
int* cpu_lod = nullptr;
if (ctx.HasInput("RoisNum")) {
auto* rois_num_t = ctx.Input<Tensor>("RoisNum");
rois_batch_size = rois_num_t->numel();
std::vector<int> rois_num_list(rois_batch_size);
memory::Copy(cplace,
rois_num_list.data(),
xplace,
rois_num_t->data<int>(),
sizeof(int) * rois_batch_size);
cpu_lod = new int[rois_batch_size + 1];
cpu_lod[0] = 0;
for (int i = 0; i < rois_batch_size; i++) {
cpu_lod[i + 1] = cpu_lod[i] + rois_num_list[i];
}
} else {
auto rois_lod = rois->lod().back();
rois_batch_size = rois_lod.size() - 1;
cpu_lod = new int[rois_batch_size + 1];
for (int i = 0; i < rois_batch_size + 1; i++) {
cpu_lod[i] = rois_lod[i];
}
}
int* roi_id_data = nullptr;
int r = xpu_malloc(reinterpret_cast<void**>(&roi_id_data),
(rois_batch_size + 1) * sizeof(int));
PADDLE_ENFORCE_EQ(r,
xpu::Error_t::SUCCESS,
platform::errors::External("no enough memory in xpu"));
memory::Copy(xplace,
roi_id_data,
cplace,
cpu_lod,
(rois_batch_size + 1) * sizeof(int));
in_grad->mutable_data<T>(ctx.GetPlace());
int output_grad_size = out_grad->numel();
delete[] cpu_lod;
if (output_grad_size > 0) {
r = xpu::roi_align_grad<T, int>(dev_ctx.x_context(),
out_grad->data<T>(),
in_grad->data<T>(),
rois->data<T>(),
roi_id_data,
in->dims()[0],
channels,
height,
width,
out_grad->dims()[0],
pooled_height,
pooled_width,
spatial_scale,
sampling_ratio,
true,
aligned);
PADDLE_ENFORCE_EQ(
r,
xpu::Error_t::SUCCESS,
platform::errors::External(
"The roi_align_grad XPU OP return wrong value[%d %s]",
r,
XPUAPIErrorMsg[r]));
}
if (dev_ctx.x_context()->xpu_stream) {
dev_ctx.Wait();
}
xpu_free(roi_id_data);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_XPU_KERNEL(
roi_align,
ops::XPUROIAlignOpKernel<paddle::platform::XPUDeviceContext, float>);
REGISTER_OP_XPU_KERNEL(
roi_align_grad,
ops::XPUROIAlignGradOpKernel<paddle::platform::XPUDeviceContext, float>);
#endif
paddle/phi/kernels/xpu/roi_align_grad_kernel.cc 0 → 100644
浏览文件 @ 8add11a0
// 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/roi_align_kernel.h"
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/backends/xpu/xpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
namespace phi {
template <typename T, typename Context>
void RoiAlignGradKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& boxes,
const paddle::optional<DenseTensor>& boxes_num,
const DenseTensor& out_grad,
int pooled_height,
int pooled_width,
float spatial_scale,
int sampling_ratio,
bool aligned,
DenseTensor* dx) {
int rois_num = boxes.dims()[0];
int channels = x.dims()[1];
int height = x.dims()[2];
int width = x.dims()[3];
if (!dx) {
return;
}
DenseTensor roi_batch_id_list;
roi_batch_id_list.Resize({rois_num});
auto cplace = phi::CPUPlace();
auto xplace = dev_ctx.GetPlace();
int rois_batch_size = 0;
int* cpu_lod = nullptr;
if (boxes_num) {
rois_batch_size = boxes_num->numel();
std::vector<int> rois_num_list(rois_batch_size);
paddle::memory::Copy(cplace,
rois_num_list.data(),
xplace,
boxes_num->data<int>(),
sizeof(int) * rois_batch_size);
cpu_lod = new int[rois_batch_size + 1];
cpu_lod[0] = 0;
for (int i = 0; i < rois_batch_size; i++) {
cpu_lod[i + 1] = cpu_lod[i] + rois_num_list[i];
}
} else {
auto rois_lod = boxes.lod().back();
rois_batch_size = rois_lod.size() - 1;
cpu_lod = new int[rois_batch_size + 1];
for (int i = 0; i < rois_batch_size + 1; i++) {
cpu_lod[i] = rois_lod[i];
}
}
int* roi_id_data = nullptr;
int r = xpu_malloc(reinterpret_cast<void**>(&roi_id_data),
(rois_batch_size + 1) * sizeof(int));
PADDLE_ENFORCE_XPU_SUCCESS(r);
paddle::memory::Copy(xplace,
roi_id_data,
cplace,
cpu_lod,
(rois_batch_size + 1) * sizeof(int));
dev_ctx.template Alloc<T>(dx);
int output_grad_size = out_grad.numel();
delete[] cpu_lod;
if (output_grad_size > 0) {
r = xpu::roi_align_grad<T, int>(dev_ctx.x_context(),
out_grad.data<T>(),
dx->data<T>(),
boxes.data<T>(),
roi_id_data,
x.dims()[0],
channels,
height,
width,
out_grad.dims()[0],
pooled_height,
pooled_width,
spatial_scale,
sampling_ratio,
true,
aligned);
PADDLE_ENFORCE_XDNN_SUCCESS(r, "roi_align_grad");
}
if (dev_ctx.x_context()->xpu_stream) {
dev_ctx.Wait();
}
xpu_free(roi_id_data);
}
} // namespace phi
PD_REGISTER_KERNEL(
roi_align_grad, XPU, ALL_LAYOUT, phi::RoiAlignGradKernel, float) {}
paddle/phi/kernels/xpu/roi_align_kernel.cc 0 → 100644
浏览文件 @ 8add11a0
// 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/roi_align_kernel.h"
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/backends/xpu/xpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
namespace phi {
template <typename T, typename Context>
void RoiAlignKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& boxes,
const paddle::optional<DenseTensor>& boxes_num,
int pooled_height,
int pooled_width,
float spatial_scale,
int sampling_ratio,
bool aligned,
DenseTensor* out) {
const auto& in_dims = x.dims();
int batch_size = in_dims[0];
int channels = in_dims[1];
int height = in_dims[2];
int width = in_dims[3];
int rois_num = boxes.dims()[0];
if (rois_num == 0) return;
DenseTensor roi_batch_id_list;
roi_batch_id_list.Resize({rois_num});
auto cplace = phi::CPUPlace();
int* roi_batch_id_data = dev_ctx.template HostAlloc<int>(&roi_batch_id_list);
auto xplace = dev_ctx.GetPlace();
int rois_batch_size = 0;
int* cpu_lod = nullptr;
if (boxes_num) {
rois_batch_size = boxes_num->numel();
PADDLE_ENFORCE_EQ(
rois_batch_size,
batch_size,
errors::InvalidArgument(
"The rois_batch_size and imgs "
"batch_size must be the same. But received rois_batch_size = %d, "
"batch_size = %d",
rois_batch_size,
batch_size));
std::vector<int> rois_num_list(rois_batch_size);
paddle::memory::Copy(cplace,
rois_num_list.data(),
xplace,
boxes_num->data<int>(),
sizeof(int) * rois_batch_size);
cpu_lod = new int[rois_batch_size + 1];
cpu_lod[0] = 0;
for (int i = 0; i < rois_batch_size; i++) {
cpu_lod[i + 1] = cpu_lod[i] + rois_num_list[i];
}
} else {
auto lod = boxes.lod();
PADDLE_ENFORCE_EQ(lod.empty(),
false,
errors::InvalidArgument("Input(ROIs) in ROIAlignOp does "
"not contain LoD information."));
auto rois_lod = lod.back();
rois_batch_size = rois_lod.size() - 1;
PADDLE_ENFORCE_EQ(
rois_batch_size,
batch_size,
errors::InvalidArgument(
"The batch size of rois and batch size "
"of images must be the same. But received rois batch size = %d, "
"and images batch size = %d",
rois_batch_size,
batch_size));
int rois_num_with_lod = rois_lod[rois_batch_size];
PADDLE_ENFORCE_EQ(
rois_num,
rois_num_with_lod,
errors::InvalidArgument(
"The actual number of rois and the number of rois "
"provided from Input(RoIsLoD) in RoIAlign must be the same."
" But received actual number of rois is %d, and the number "
"of rois from RoIsLoD is %d",
rois_num,
rois_num_with_lod));
for (int n = 0; n < rois_batch_size; ++n) {
for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) {
roi_batch_id_data[i] = n;
}
}
cpu_lod = new int[rois_batch_size + 1];
for (int i = 0; i < rois_batch_size + 1; i++) {
cpu_lod[i] = rois_lod[i];
}
}
int* roi_id_data = nullptr;
int r = xpu_malloc(reinterpret_cast<void**>(&roi_id_data),
(rois_batch_size + 1) * sizeof(int));
PADDLE_ENFORCE_XPU_SUCCESS(r);
paddle::memory::Copy(xplace,
roi_id_data,
cplace,
cpu_lod,
(rois_batch_size + 1) * sizeof(int));
delete[] cpu_lod;
r = xpu::roi_align<T, int>(dev_ctx.x_context(),
x.data<T>(),
dev_ctx.template Alloc<T>(out),
boxes.data<T>(),
roi_id_data,
batch_size,
channels,
height,
width,
out->dims()[0],
pooled_height,
pooled_width,
spatial_scale,
sampling_ratio,
true,
aligned);
PADDLE_ENFORCE_XDNN_SUCCESS(r, "roi_align_grad");
if (dev_ctx.x_context()->xpu_stream) {
dev_ctx.Wait();
}
xpu_free(roi_id_data);
}
} // namespace phi
PD_REGISTER_KERNEL(roi_align, XPU, ALL_LAYOUT, phi::RoiAlignKernel, float) {}
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或