Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 1 年 前同步成功

通知 2299
Star 20931
Fork 5422
P
Paddle
未验证 提交 445fce62 编写于 作者: T taixiurong 提交者： GitHub
浏览文件
操作

xpu-paddlepaddle-38 [任务] 迁移bilinear_interp，nearest_interp到phi test=kunlun (#45608)

上级 f9882854
隐藏空白更改
内联 并排
Showing with 478 addition and 616 deletion
paddle/fluid/operators/interpolate_op_xpu.cc 已删除 100644 → 0
浏览文件 @ f9882854
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
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 <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/interpolate_op.h"
#ifdef PADDLE_WITH_XPU
namespace paddle {
namespace operators {
using framework::Tensor;
using DataLayout = framework::DataLayout;
inline std::vector<int> get_new_shape_xpu(
const std::vector<const Tensor*>& list_new_shape_tensor) {
// get tensor from
std::vector<int> vec_new_shape;
for (size_t i = 0; i < list_new_shape_tensor.size(); ++i) {
auto tensor = list_new_shape_tensor[i];
PADDLE_ENFORCE_EQ(
tensor->dims(),
phi::make_ddim({1}),
platform::errors::InvalidArgument("shape of dim tensor should be [1]"));
if (platform::is_xpu_place(tensor->place())) {
framework::Tensor temp;
paddle::framework::TensorCopySync(*tensor, platform::CPUPlace(), &temp);
vec_new_shape.push_back(static_cast<int32_t>(*temp.data<int32_t>()));
} else {
vec_new_shape.push_back(static_cast<int32_t>(*tensor->data<int32_t>()));
}
}
return vec_new_shape;
}
template <typename T>
inline std::vector<T> get_new_data_from_tensor_xpu(
const Tensor* new_data_tensor) {
std::vector<T> vec_new_data;
auto* new_data = new_data_tensor->data<T>();
framework::Tensor cpu_starts_tensor;
if (platform::is_xpu_place(new_data_tensor->place())) {
paddle::framework::TensorCopySync(
*new_data_tensor, platform::CPUPlace(), &cpu_starts_tensor);
new_data = cpu_starts_tensor.data<T>();
}
vec_new_data = std::vector<T>(new_data, new_data + new_data_tensor->numel());
return vec_new_data;
}
template <typename T>
class InterpolateXPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* input = ctx.Input<Tensor>("X");
auto* output = ctx.Output<Tensor>("Out");
auto input_dims = input->dims();
PADDLE_ENFORCE_EQ(
input_dims.size(),
4,
platform::errors::External("XPU Interpolate kernel only support 2d"));
const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
const DataLayout data_layout =
framework::StringToDataLayout(data_layout_str);
int n, c, in_d, in_h, in_w;
ExtractNCDWH(input_dims, data_layout, &n, &c, &in_d, &in_h, &in_w);
auto interp_method = ctx.Attr<std::string>("interp_method");
bool align_corners = ctx.Attr<bool>("align_corners");
int align_mode = ctx.Attr<int>("align_mode");
int out_h = ctx.Attr<int>("out_h");
int out_w = ctx.Attr<int>("out_w");
auto list_new_size_tensor = ctx.MultiInput<framework::Tensor>("SizeTensor");
if (list_new_size_tensor.size() > 0) {
// have size tensor
auto new_size = get_new_shape_xpu(list_new_size_tensor);
out_h = new_size[0];
out_w = new_size[1];
} else {
float scale;
auto scale_tensor = ctx.Input<Tensor>("Scale");
if (scale_tensor != nullptr) {
auto scale_data = get_new_data_from_tensor_xpu<float>(scale_tensor);
scale = scale_data[0];
} else {
scale = ctx.Attr<float>("scale");
}
if (scale > 0) {
out_h = static_cast<int>(in_h * scale);
out_w = static_cast<int>(in_w * scale);
}
auto out_size = ctx.Input<Tensor>("OutSize");
if (out_size != nullptr) {
auto out_size_data = get_new_data_from_tensor_xpu<int>(out_size);
out_h = out_size_data[0];
out_w = out_size_data[1];
}
}
PADDLE_ENFORCE_GT(
out_h,
0,
platform::errors::InvalidArgument("out_h in Attr(out_shape) of "
"Op(interpolate) "
"should be greater than 0."));
PADDLE_ENFORCE_GT(
out_w,
0,
platform::errors::InvalidArgument("out_w in Attr(out_shape) of "
"Op(interpolate) "
"should be greater than 0."));
framework::DDim dim_out;
if (data_layout == DataLayout::kNCHW) {
dim_out = {n, c, out_h, out_w};
} else {
dim_out = {n, out_h, out_w, c};
}
output->mutable_data<T>(dim_out, ctx.GetPlace());
if (in_h == out_h && in_w == out_w) {
framework::TensorCopy(*input, ctx.GetPlace(), output);
return;
}
bool nearest = "nearest" == interp_method;
int trans_mode = (align_corners) ? (0) : ((align_mode == 0) ? (1) : (2));
auto& dev_ctx = ctx.template device_context<platform::XPUDeviceContext>();
if (nearest) {
PADDLE_ENFORCE_EQ((data_layout == DataLayout::kNCHW),
true,
platform::errors::InvalidArgument(
"XPU nearest is only support NCHW"));
}
int r = xpu::interpolate2d<float>(dev_ctx.x_context(),
input->data<float>(),
output->data<float>(),
n,
c,
in_h,
in_w,
out_h,
out_w,
nearest,
trans_mode,
(data_layout == DataLayout::kNCHW));
PADDLE_ENFORCE_EQ(r,
XPU_SUCCESS,
platform::errors::External("XPU interpolate2d kernel "
"return wrong value[%d %s]",
r,
XPUAPIErrorMsg[r]));
}
};
template <typename T>
class InterpolateGradXPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* output_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
auto output_grad_dims = output_grad->dims();
PADDLE_ENFORCE_EQ(output_grad_dims.size(),
4,
platform::errors::External(
"XPU Interpolategrad kernel only support 2d"));
auto* input = ctx.Input<Tensor>("X");
const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
const DataLayout data_layout =
framework::StringToDataLayout(data_layout_str);
int n, c, in_d, in_h, in_w;
ExtractNCDWH(input->dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);
auto interp_method = ctx.Attr<std::string>("interp_method");
bool align_corners = ctx.Attr<bool>("align_corners");
int align_mode = ctx.Attr<int>("align_mode");
int out_h = ctx.Attr<int>("out_h");
int out_w = ctx.Attr<int>("out_w");
float scale;
auto scale_tensor = ctx.Input<Tensor>("Scale");
if (scale_tensor != nullptr) {
auto scale_data = get_new_data_from_tensor_xpu<float>(scale_tensor);
scale = scale_data[0];
} else {
scale = ctx.Attr<float>("scale");
}
if (scale > 0) {
out_h = static_cast<int>(in_h * scale);
out_w = static_cast<int>(in_w * scale);
}
auto out_size = ctx.Input<Tensor>("OutSize");
if (out_size != nullptr) {
auto out_size_data = get_new_data_from_tensor_xpu<int>(out_size);
out_h = out_size_data[0];
out_w = out_size_data[1];
}
auto list_new_size_tensor = ctx.MultiInput<framework::Tensor>("SizeTensor");
if (list_new_size_tensor.size() > 0) {
// have size tensor
auto new_size = get_new_shape_xpu(list_new_size_tensor);
out_h = new_size[0];
out_w = new_size[1];
}
framework::DDim dim_grad;
if (data_layout == DataLayout::kNCHW) {
dim_grad = {n, c, in_h, in_w};
} else {
dim_grad = {n, in_h, in_w, c};
}
input_grad->mutable_data<T>(dim_grad, ctx.GetPlace());
auto& dev_ctx = ctx.template device_context<platform::XPUDeviceContext>();
int r = XPU_SUCCESS;
r = xpu::constant<T>(dev_ctx.x_context(),
input_grad->data<T>(),
input_grad->numel(),
static_cast<T>(0.0));
PADDLE_ENFORCE_EQ(r,
XPU_SUCCESS,
platform::errors::External(
"XPU constant in interpolate2d_grad kernel return "
"wrong value[%d %s]",
r,
XPUAPIErrorMsg[r]));
if (in_h == out_h && in_w == out_w) {
framework::TensorCopy(*output_grad, ctx.GetPlace(), input_grad);
return;
}
bool nearest = "nearest" == interp_method;
int trans_mode = (align_corners) ? (0) : ((align_mode == 0) ? (1) : (2));
if (nearest) {
trans_mode = (align_corners) ? (0) : (2);
}
r = xpu::interpolate2d_grad<T>(dev_ctx.x_context(),
output_grad->data<T>(),
input_grad->data<T>(),
n,
c,
in_h,
in_w,
out_h,
out_w,
nearest,
trans_mode,
(data_layout == DataLayout::kNCHW));
PADDLE_ENFORCE_EQ(
r,
XPU_SUCCESS,
platform::errors::External("XPU interpolate2d_grad kernel return "
"wrong value[%d %s]",
r,
XPUAPIErrorMsg[r]));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_XPU_KERNEL(bilinear_interp, ops::InterpolateXPUKernel<float>);
REGISTER_OP_XPU_KERNEL(nearest_interp, ops::InterpolateXPUKernel<float>);
REGISTER_OP_XPU_KERNEL(bilinear_interp_grad,
ops::InterpolateGradXPUKernel<float>);
REGISTER_OP_XPU_KERNEL(nearest_interp_grad,
ops::InterpolateGradXPUKernel<float>);
#endif
paddle/fluid/operators/interpolate_v2_op_xpu.cc 已删除 100644 → 0
浏览文件 @ f9882854
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
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 <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/phi/kernels/funcs/interpolate_function.h"
#ifdef PADDLE_WITH_XPU
namespace paddle {
namespace operators {
using framework::Tensor;
using DataLayout = framework::DataLayout;
inline std::vector<int> get_new_shape_xpu(
const std::vector<const Tensor*>& list_new_shape_tensor) {
// get tensor from
std::vector<int> vec_new_shape;
for (size_t i = 0; i < list_new_shape_tensor.size(); ++i) {
auto tensor = list_new_shape_tensor[i];
PADDLE_ENFORCE_EQ(
tensor->dims(),
phi::make_ddim({1}),
platform::errors::InvalidArgument("shape of dim tensor should be [1]"));
framework::Tensor temp;
paddle::framework::TensorCopySync(*tensor, platform::CPUPlace(), &temp);
vec_new_shape.push_back(static_cast<int32_t>(*temp.data<int32_t>()));
}
return vec_new_shape;
}
template <typename T>
class InterpolateV2XPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* input = ctx.Input<Tensor>("X");
auto* output = ctx.Output<Tensor>("Out");
auto input_dims = input->dims();
PADDLE_ENFORCE_EQ(
input_dims.size(),
4,
platform::errors::External("XPU Interpolate kernel only support 2d"));
const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
const DataLayout data_layout =
framework::StringToDataLayout(data_layout_str);
int n, c, in_d, in_h, in_w;
phi::funcs::ExtractNCDWH(
input_dims, data_layout, &n, &c, &in_d, &in_h, &in_w);
auto interp_method = ctx.Attr<std::string>("interp_method");
bool align_corners = ctx.Attr<bool>("align_corners");
int align_mode = ctx.Attr<int>("align_mode");
int out_h = ctx.Attr<int>("out_h");
int out_w = ctx.Attr<int>("out_w");
float scale_h = -1;
float scale_w = -1;
auto list_new_size_tensor = ctx.MultiInput<framework::Tensor>("SizeTensor");
if (list_new_size_tensor.size() > 0) {
// have size tensor
auto new_size = get_new_shape_xpu(list_new_size_tensor);
out_h = new_size[0];
out_w = new_size[1];
} else {
auto scale_tensor = ctx.Input<Tensor>("Scale");
auto scale = ctx.Attr<std::vector<float>>("scale");
if (scale_tensor != nullptr) {
auto scale_data =
phi::funcs::get_new_data_from_tensor<float>(scale_tensor);
if (scale_data.size() > 1) {
scale_h = scale_data[0];
scale_w = scale_data[1];
} else {
scale_h = scale_data[0];
scale_w = scale_data[0];
}
PADDLE_ENFORCE_EQ(
scale_w > 0 && scale_h > 0,
true,
platform::errors::InvalidArgument("scale of Op(interpolate) "
"should be greater than 0."));
} else {
if (scale.size() > 1) {
scale_h = scale[0];
scale_w = scale[1];
PADDLE_ENFORCE_EQ(
scale_w > 0 && scale_h > 0,
true,
platform::errors::InvalidArgument("scale of Op(interpolate) "
"should be greater than 0."));
}
}
if (scale_h > 0. && scale_w > 0.) {
out_h = static_cast<int>(in_h * scale_h);
out_w = static_cast<int>(in_w * scale_w);
}
auto out_size = ctx.Input<Tensor>("OutSize");
if (out_size != nullptr) {
auto out_size_data =
phi::funcs::get_new_data_from_tensor<int>(out_size);
out_h = out_size_data[0];
out_w = out_size_data[1];
}
}
PADDLE_ENFORCE_GT(
out_h,
0,
platform::errors::InvalidArgument("out_h in Attr(out_shape) of "
"Op(interpolate) "
"should be greater than 0."));
PADDLE_ENFORCE_GT(
out_w,
0,
platform::errors::InvalidArgument("out_w in Attr(out_shape) of "
"Op(interpolate) "
"should be greater than 0."));
framework::DDim dim_out;
if (data_layout == DataLayout::kNCHW) {
dim_out = {n, c, out_h, out_w};
} else {
dim_out = {n, out_h, out_w, c};
}
output->mutable_data<T>(dim_out, ctx.GetPlace());
if (in_h == out_h && in_w == out_w) {
framework::TensorCopy(*input, ctx.GetPlace(), output);
return;
}
bool nearest = "nearest" == interp_method;
int trans_mode = (align_corners) ? (0) : ((align_mode == 0) ? (1) : (2));
auto& dev_ctx = ctx.template device_context<platform::XPUDeviceContext>();
if (nearest) {
PADDLE_ENFORCE_EQ((data_layout == DataLayout::kNCHW),
true,
platform::errors::InvalidArgument(
"XPU nearest is only support NCHW"));
}
int r = xpu::interpolate2d<T>(dev_ctx.x_context(),
input->data<T>(),
output->data<T>(),
n,
c,
in_h,
in_w,
out_h,
out_w,
nearest,
trans_mode,
(data_layout == DataLayout::kNCHW));
PADDLE_ENFORCE_EQ(r,
XPU_SUCCESS,
platform::errors::External("XPU interpolate2d kernel "
"return wrong value[%d %s]",
r,
XPUAPIErrorMsg[r]));
}
};
template <typename T>
class InterpolateV2GradXPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* output_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
auto output_grad_dims = output_grad->dims();
PADDLE_ENFORCE_EQ(output_grad_dims.size(),
4,
platform::errors::External(
"XPU Interpolategrad kernel only support 2d"));
auto* input = ctx.Input<Tensor>("X");
const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
const DataLayout data_layout =
framework::StringToDataLayout(data_layout_str);
int n, c, in_d, in_h, in_w;
phi::funcs::ExtractNCDWH(
input->dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);
auto interp_method = ctx.Attr<std::string>("interp_method");
bool align_corners = ctx.Attr<bool>("align_corners");
int align_mode = ctx.Attr<int>("align_mode");
int out_h = ctx.Attr<int>("out_h");
int out_w = ctx.Attr<int>("out_w");
float scale_h = -1;
float scale_w = -1;
auto list_new_size_tensor = ctx.MultiInput<framework::Tensor>("SizeTensor");
if (list_new_size_tensor.size() > 0) {
// have size tensor
auto new_size = get_new_shape_xpu(list_new_size_tensor);
out_h = new_size[0];
out_w = new_size[1];
} else {
auto scale_tensor = ctx.Input<Tensor>("Scale");
auto scale = ctx.Attr<std::vector<float>>("scale");
if (scale_tensor != nullptr) {
auto scale_data =
phi::funcs::get_new_data_from_tensor<float>(scale_tensor);
if (scale_data.size() > 1) {
scale_h = scale_data[0];
scale_w = scale_data[1];
} else {
scale_h = scale_data[0];
scale_w = scale_data[0];
}
PADDLE_ENFORCE_EQ(
scale_w > 0 && scale_h > 0,
true,
platform::errors::InvalidArgument("scale of Op(interpolate) "
"should be greater than 0."));
} else {
if (scale.size() > 1) {
scale_h = scale[0];
scale_w = scale[1];
PADDLE_ENFORCE_EQ(
scale_w > 0 && scale_h > 0,
true,
platform::errors::InvalidArgument("scale of Op(interpolate) "
"should be greater than 0."));
}
}
if (scale_h > 0. && scale_w > 0.) {
out_h = static_cast<int>(in_h * scale_h);
out_w = static_cast<int>(in_w * scale_w);
}
auto out_size = ctx.Input<Tensor>("OutSize");
if (out_size != nullptr) {
auto out_size_data =
phi::funcs::get_new_data_from_tensor<int>(out_size);
out_h = out_size_data[0];
out_w = out_size_data[1];
}
}
framework::DDim dim_grad;
if (data_layout == DataLayout::kNCHW) {
dim_grad = {n, c, in_h, in_w};
} else {
dim_grad = {n, in_h, in_w, c};
}
input_grad->mutable_data<T>(dim_grad, ctx.GetPlace());
auto& dev_ctx = ctx.template device_context<platform::XPUDeviceContext>();
int r = XPU_SUCCESS;
r = xpu::constant<T>(dev_ctx.x_context(),
input_grad->data<T>(),
input_grad->numel(),
static_cast<T>(0.0));
PADDLE_ENFORCE_EQ(r,
XPU_SUCCESS,
platform::errors::External(
"XPU constant in interpolate2d_grad kernel return "
"wrong value[%d %s]",
r,
XPUAPIErrorMsg[r]));
if (in_h == out_h && in_w == out_w) {
framework::TensorCopy(*output_grad, ctx.GetPlace(), input_grad);
return;
}
bool nearest = "nearest" == interp_method;
int trans_mode = (align_corners) ? (0) : ((align_mode == 0) ? (1) : (2));
if (nearest) {
trans_mode = (align_corners) ? (0) : (2);
}
r = xpu::interpolate2d_grad<T>(dev_ctx.x_context(),
output_grad->data<T>(),
input_grad->data<T>(),
n,
c,
in_h,
in_w,
out_h,
out_w,
nearest,
trans_mode,
(data_layout == DataLayout::kNCHW));
PADDLE_ENFORCE_EQ(
r,
XPU_SUCCESS,
platform::errors::External("XPU interpolate2d_grad kernel return "
"wrong value[%d %s]",
r,
XPUAPIErrorMsg[r]));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_XPU_KERNEL(bilinear_interp_v2, ops::InterpolateV2XPUKernel<float>);
REGISTER_OP_XPU_KERNEL(nearest_interp_v2, ops::InterpolateV2XPUKernel<float>);
REGISTER_OP_XPU_KERNEL(bilinear_interp_v2_grad,
ops::InterpolateV2GradXPUKernel<float>);
REGISTER_OP_XPU_KERNEL(nearest_interp_v2_grad,
ops::InterpolateV2GradXPUKernel<float>);
#endif
paddle/phi/kernels/funcs/interpolate_function.h
浏览文件 @ 445fce62
...@@ -91,6 +91,14 @@ inline std::vector<int> get_new_shape( ...@@ -91,6 +91,14 @@ inline std::vector<int> get_new_shape(
errors::InvalidArgument("The shape of dimension tensor should be [1]," errors::InvalidArgument("The shape of dimension tensor should be [1],"
"but received d%.", "but received d%.",
tensor->dims())); tensor->dims()));
#ifdef PADDLE_WITH_XPU
if (tensor->place().GetType() == phi::AllocationType::XPU) {
DenseTensor temp;
paddle::framework::TensorCopySync(*tensor, phi::CPUPlace(), &temp);
vec_new_shape.push_back(static_cast<int32_t>(*temp.data<int32_t>()));
continue;
}
#endif
if (paddle::platform::is_gpu_place(tensor->place())) { if (paddle::platform::is_gpu_place(tensor->place())) {
DenseTensor temp; DenseTensor temp;
paddle::framework::TensorCopySync( paddle::framework::TensorCopySync(
......
paddle/phi/kernels/xpu/interpolate_grad_kernel.cc 0 → 100644
浏览文件 @ 445fce62
// 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/interpolate_grad_kernel.h"
#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/backends/xpu/xpu_context.h"
#include "paddle/phi/common/layout.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/interpolate_function.h"
#include "paddle/phi/kernels/funcs/math_function.h"
namespace phi {
template <typename T, typename Context>
void InterpolateGradKernel(
const Context& dev_ctx,
const DenseTensor& x,
const paddle::optional<DenseTensor>& out_size,
const paddle::optional<std::vector<const DenseTensor*>>& size_tensor,
const paddle::optional<DenseTensor>& scale_tensor,
const DenseTensor& output_grad,
const std::string& data_layout_str,
int out_d,
int out_h,
int out_w,
const std::vector<float>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* x_grad) {
const DataLayout data_layout =
paddle::framework::StringToDataLayout(data_layout_str);
int n, c, in_d, in_h, in_w;
funcs::ExtractNCDWH(x.dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);
float scale_h = -1;
float scale_w = -1;
if (scale_tensor) {
auto scale_data =
funcs::get_new_data_from_tensor<float>(scale_tensor.get_ptr());
if (scale_data.size() > 1) {
scale_h = scale_data[0];
scale_w = scale_data[1];
} else {
scale_w = scale_data[0];
scale_h = scale_data[0];
}
PADDLE_ENFORCE_EQ(
scale_w > 0,
true,
errors::InvalidArgument(
"The scale_w in input 'Scale' Tensor of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_w));
PADDLE_ENFORCE_EQ(
scale_h > 0,
true,
errors::InvalidArgument(
"The scale_h in input 'Scale' Tensor of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_h));
} else {
if (scale.size() > 1) {
scale_h = scale[0];
scale_w = scale[1];
PADDLE_ENFORCE_EQ(
scale_w > 0,
true,
errors::InvalidArgument(
"The scale_w in Attr(scale) of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_w));
PADDLE_ENFORCE_EQ(
scale_h > 0,
true,
errors::InvalidArgument(
"The scale_h in Attr(scale) of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_h));
}
}
if (scale_h > 0. && scale_w > 0.) {
out_h = static_cast<int>(in_h * scale_h);
out_w = static_cast<int>(in_w * scale_w);
}
if (out_size) {
auto out_size_data =
funcs::get_new_data_from_tensor<int>(out_size.get_ptr());
out_h = out_size_data[0];
out_w = out_size_data[1];
}
if (size_tensor && size_tensor->size() > 0) {
// have size tensor
auto new_size = funcs::get_new_shape(size_tensor.get());
out_h = new_size[0];
out_w = new_size[1];
}
phi::DDim dim_grad;
if (data_layout == DataLayout::kNCHW) {
dim_grad = {n, c, in_h, in_w};
} else {
dim_grad = {n, in_h, in_w, c};
}
x_grad->Resize(dim_grad);
dev_ctx.template Alloc<T>(x_grad);
int r = XPU_SUCCESS;
r = xpu::constant<T>(dev_ctx.x_context(),
x_grad->data<T>(),
x_grad->numel(),
static_cast<T>(0.0));
PADDLE_ENFORCE_XDNN_SUCCESS(r, "constant");
if (in_h == out_h && in_w == out_w) {
phi::Copy<Context>(dev_ctx, output_grad, dev_ctx.GetPlace(), false, x_grad);
return;
}
bool nearest = "nearest" == interp_method;
int trans_mode = (align_corners) ? (0) : ((align_mode == 0) ? (1) : (2));
if (nearest) {
trans_mode = (align_corners) ? (0) : (2);
}
r = xpu::interpolate2d_grad<T>(dev_ctx.x_context(),
output_grad.data<T>(),
x_grad->data<T>(),
n,
c,
in_h,
in_w,
out_h,
out_w,
nearest,
trans_mode,
(data_layout == DataLayout::kNCHW));
PADDLE_ENFORCE_XDNN_SUCCESS(r, "interpolate2d_grad");
}
template <typename T, typename Context>
void BilinearInterpGradKernel(
const Context& dev_ctx,
const DenseTensor& x,
const paddle::optional<DenseTensor>& out_size,
const paddle::optional<std::vector<const DenseTensor*>>& size_tensor,
const paddle::optional<DenseTensor>& scale_tensor,
const DenseTensor& out_grad,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
const std::vector<float>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* x_grad) {
InterpolateGradKernel<T, Context>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
out_grad,
data_layout,
out_d,
out_h,
out_w,
scale,
interp_method,
align_corners,
align_mode,
x_grad);
}
template <typename T, typename Context>
void NearestInterpGradKernel(
const Context& dev_ctx,
const DenseTensor& x,
const paddle::optional<DenseTensor>& out_size,
const paddle::optional<std::vector<const DenseTensor*>>& size_tensor,
const paddle::optional<DenseTensor>& scale_tensor,
const DenseTensor& out_grad,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
const std::vector<float>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* x_grad) {
InterpolateGradKernel<T, Context>(dev_ctx,
x,
out_size,
size_tensor,
scale_tensor,
out_grad,
data_layout,
out_d,
out_h,
out_w,
scale,
interp_method,
align_corners,
align_mode,
x_grad);
}
} // namespace phi
PD_REGISTER_KERNEL(bilinear_interp_grad,
XPU,
ALL_LAYOUT,
phi::BilinearInterpGradKernel,
float) {
kernel->InputAt(2).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(3).SetBackend(phi::Backend::ALL_BACKEND);
}
PD_REGISTER_KERNEL(
nearest_interp_grad, XPU, ALL_LAYOUT, phi::NearestInterpGradKernel, float) {
kernel->InputAt(2).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(3).SetBackend(phi::Backend::ALL_BACKEND);
}
paddle/phi/kernels/xpu/interpolate_kernel.cc 0 → 100644
浏览文件 @ 445fce62
// 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/interpolate_kernel.h"
#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/backends/xpu/xpu_context.h"
#include "paddle/phi/common/layout.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/interpolate_function.h"
namespace phi {
template <typename T, typename Context>
void InterpolateKernel(
const Context& ctx,
const DenseTensor& x,
const paddle::optional<DenseTensor>& out_size,
const paddle::optional<std::vector<const DenseTensor*>>& size_tensor,
const paddle::optional<DenseTensor>& scale_tensor,
const std::string& data_layout_str,
int out_d,
int out_h,
int out_w,
const std::vector<float>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
const DataLayout data_layout =
paddle::framework::StringToDataLayout(data_layout_str);
int n, c, in_d, in_h, in_w;
phi::funcs::ExtractNCDWH(x.dims(), data_layout, &n, &c, &in_d, &in_h, &in_w);
float scale_h = -1;
float scale_w = -1;
if (size_tensor && size_tensor->size() > 0) {
// have size tensor
auto new_size = funcs::get_new_shape(size_tensor.get());
out_h = new_size[0];
out_w = new_size[1];
} else {
if (scale_tensor) {
auto scale_data =
funcs::get_new_data_from_tensor<float>(scale_tensor.get_ptr());
if (scale_data.size() > 1) {
scale_h = scale_data[0];
scale_w = scale_data[1];
} else {
scale_h = scale_data[0];
scale_w = scale_data[0];
}
PADDLE_ENFORCE_EQ(
scale_w > 0,
true,
errors::InvalidArgument(
"The scale_w in input 'Scale' Tensor of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_w));
PADDLE_ENFORCE_EQ(
scale_h > 0,
true,
errors::InvalidArgument(
"The scale_h in input 'Scale' Tensor of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_h));
} else {
if (scale.size() > 1) {
scale_h = scale[0];
scale_w = scale[1];
PADDLE_ENFORCE_EQ(
scale_w > 0,
true,
errors::InvalidArgument(
"The scale_w in Attr(scale) of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_w));
PADDLE_ENFORCE_EQ(
scale_h > 0,
true,
errors::InvalidArgument(
"The scale_h in Attr(scale) of Operator(interpolate) "
"should be greater than 0, but received value is %d.",
scale_h));
}
}
if (scale_h > 0. && scale_w > 0.) {
out_h = static_cast<int>(in_h * scale_h);
out_w = static_cast<int>(in_w * scale_w);
}
if (out_size) {
auto out_size_data =
funcs::get_new_data_from_tensor<int>(out_size.get_ptr());
out_h = out_size_data[0];
out_w = out_size_data[1];
}
}
PADDLE_ENFORCE_GT(
out_h,
0,
errors::InvalidArgument("out_h in Attr(out_shape) of Op(interpolate) "
"should be greater than 0."));
PADDLE_ENFORCE_GT(
out_w,
0,
errors::InvalidArgument("out_w in Attr(out_shape) of Op(interpolate) "
"should be greater than 0."));
phi::DDim dim_out;
if (data_layout == DataLayout::kNCHW) {
dim_out = {n, c, out_h, out_w};
} else {
dim_out = {n, out_h, out_w, c};
}
output->Resize(dim_out);
ctx.template Alloc<T>(output);
if (in_h == out_h && in_w == out_w) {
phi::Copy<Context>(ctx, x, ctx.GetPlace(), false, output);
return;
}
bool nearest = "nearest" == interp_method;
int trans_mode = (align_corners) ? (0) : ((align_mode == 0) ? (1) : (2));
if (nearest) {
PADDLE_ENFORCE_EQ(
(data_layout == DataLayout::kNCHW),
true,
errors::InvalidArgument("XPU nearest is only support NCHW"));
}
int r = xpu::interpolate2d<T>(ctx.x_context(),
x.data<T>(),
output->data<T>(),
n,
c,
in_h,
in_w,
out_h,
out_w,
nearest,
trans_mode,
(data_layout == DataLayout::kNCHW));
PADDLE_ENFORCE_XDNN_SUCCESS(r, "interpolate2d");
}
template <typename T, typename Context>
void BilinearInterpKernel(
const Context& ctx,
const DenseTensor& x,
const paddle::optional<DenseTensor>& out_size,
const paddle::optional<std::vector<const DenseTensor*>>& size_tensor,
const paddle::optional<DenseTensor>& scale_tensor,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
const std::vector<float>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
InterpolateKernel<T, Context>(ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_d,
out_h,
out_w,
scale,
interp_method,
align_corners,
align_mode,
output);
}
template <typename T, typename Context>
void NearestInterpKernel(
const Context& ctx,
const DenseTensor& x,
const paddle::optional<DenseTensor>& out_size,
const paddle::optional<std::vector<const DenseTensor*>>& size_tensor,
const paddle::optional<DenseTensor>& scale_tensor,
const std::string& data_layout,
int out_d,
int out_h,
int out_w,
const std::vector<float>& scale,
const std::string& interp_method,
bool align_corners,
int align_mode,
DenseTensor* output) {
InterpolateKernel<T, Context>(ctx,
x,
out_size,
size_tensor,
scale_tensor,
data_layout,
out_d,
out_h,
out_w,
scale,
interp_method,
align_corners,
align_mode,
output);
}
} // namespace phi
PD_REGISTER_KERNEL(
bilinear_interp, XPU, ALL_LAYOUT, phi::BilinearInterpKernel, float) {
kernel->InputAt(2).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(3).SetBackend(phi::Backend::ALL_BACKEND);
}
PD_REGISTER_KERNEL(
nearest_interp, XPU, ALL_LAYOUT, phi::NearestInterpKernel, float) {
kernel->InputAt(2).SetBackend(phi::Backend::ALL_BACKEND);
kernel->InputAt(3).SetBackend(phi::Backend::ALL_BACKEND);
}
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