Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 1 年 前同步成功

通知 2298
Star 20931
Fork 5422
P
Paddle
未验证 提交 13a0ea4c 编写于 作者: J james 提交者： GitHub
浏览文件
操作

[XPU] move pool/pool_grad xpu kernel to phi (#45480) 

* move pool/pool_grad xpu kernel to phi, test=kunlun

* replace mutable_data() with DeviceContext::Alloc()

* replace PADDLE_ENFORCE_EQ with PADDLE_ENFORCE_XDNN, test=kunlun

* adjust function param name & update include header

* remove pool_op_xpu.cc

* fire r200 test

* minor, test=kunlun
上级 4478389c
隐藏空白更改
内联 并排
Showing with 310 addition and 317 deletion
paddle/fluid/operators/pool_op_xpu.cc 已删除 100644 → 0
浏览文件 @ 4478389c
/* 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. */
#include <unordered_map>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/phi/kernels/funcs/pooling.h"
#ifdef PADDLE_WITH_XPU
namespace paddle {
namespace operators {
using framework::Tensor;
template <typename DeviceContext, typename T>
class PoolXPUKernel : public framework::OpKernel<T> {
using XPUType = typename XPUTypeTrait<T>::Type;
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* in_x = context.Input<Tensor>("X");
Tensor* out = context.Output<Tensor>("Out");
std::string pooling_type = context.Attr<std::string>("pooling_type");
std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
bool exclusive = context.Attr<bool>("exclusive");
bool adaptive = context.Attr<bool>("adaptive");
bool ceil_mode = context.Attr<bool>("ceil_mode");
std::string padding_algorithm =
context.Attr<std::string>("padding_algorithm");
PADDLE_ENFORCE_EQ(
ksize.size(),
2,
platform::errors::InvalidArgument(
"The Pool2d XPU OP only support 2 dimension pooling!"));
std::string data_format = context.Attr<std::string>("data_format");
PADDLE_ENFORCE_EQ(
data_format,
"NCHW",
platform::errors::InvalidArgument("The Pool2d XPU OP only support"
"data_format is 'NCHW', but received "
"%s",
data_format));
int* index_data = nullptr;
bool global_pooling = context.Attr<bool>("global_pooling") ||
(adaptive && (ksize[0] * ksize[1] == 1));
if (global_pooling) {
for (size_t i = 0; i < ksize.size(); ++i) {
paddings[i] = 0;
ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
}
}
const int n = in_x->dims()[0];
const int c = in_x->dims()[1];
const int in_h = in_x->dims()[2];
const int in_w = in_x->dims()[3];
const int out_h = out->dims()[2];
const int out_w = out->dims()[3];
framework::DDim data_dims;
data_dims = phi::slice_ddim(in_x->dims(), 2, in_x->dims().size());
phi::funcs::UpdatePadding(&paddings,
global_pooling,
adaptive,
padding_algorithm,
data_dims,
strides,
ksize);
if (ceil_mode) {
int in_h_ceil = (out_h - 1) * strides[0] + ksize[0] - 2 * paddings[0];
int in_w_ceil = (out_w - 1) * strides[1] + ksize[1] - 2 * paddings[2];
paddings[1] += (in_h_ceil - in_h);
paddings[3] += (in_w_ceil - in_w);
}
auto input = reinterpret_cast<const XPUType*>(in_x->data<T>());
out->mutable_data<T>(context.GetPlace());
auto output = reinterpret_cast<XPUType*>(out->data<T>());
auto& dev_ctx = context.template device_context<DeviceContext>();
int r = xpu::Error_t::SUCCESS;
if (!adaptive) {
if (pooling_type == "max") {
r = xpu::max_pool2d<XPUType>(dev_ctx.x_context(),
input,
output,
index_data,
n,
c,
in_h,
in_w,
ksize,
strides,
paddings,
true);
} else if (pooling_type == "avg") {
r = xpu::avg_pool2d<XPUType>(dev_ctx.x_context(),
input,
output,
n,
c,
in_h,
in_w,
ksize,
strides,
paddings,
!exclusive,
true);
} else {
PADDLE_THROW(platform::errors::InvalidArgument(
"Unsupported pooling type for kunlun ", pooling_type));
}
} else {
if (pooling_type == "max") {
r = xpu::adaptive_max_pool2d<XPUType>(dev_ctx.x_context(),
input,
output,
index_data,
n,
c,
in_h,
in_w,
out_h,
out_w,
true);
} else if (pooling_type == "avg") {
r = xpu::adaptive_avg_pool2d<XPUType>(dev_ctx.x_context(),
input,
output,
n,
c,
in_h,
in_w,
out_h,
out_w,
true);
} else {
PADDLE_THROW(platform::errors::InvalidArgument(
"Unsupported pooling type for kunlun ", pooling_type));
}
}
PADDLE_ENFORCE_EQ(r,
xpu::Error_t::SUCCESS,
platform::errors::External(
"The pool2d XPU API return wrong value[%d %s]",
r,
XPUAPIErrorMsg[r]));
}
};
template <typename DeviceContext, typename T>
class PoolGradXPUKernel : public framework::OpKernel<T> {
using XPUType = typename XPUTypeTrait<T>::Type;
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* in_x = context.Input<Tensor>("X");
const Tensor* out = context.Input<Tensor>("Out");
const Tensor* out_grad =
context.Input<Tensor>(framework::GradVarName("Out"));
Tensor* in_x_grad = context.Output<Tensor>(framework::GradVarName("X"));
std::string pooling_type = context.Attr<std::string>("pooling_type");
std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
bool exclusive = context.Attr<bool>("exclusive");
bool adaptive = context.Attr<bool>("adaptive");
bool ceil_mode = context.Attr<bool>("ceil_mode");
std::string data_format = context.Attr<std::string>("data_format");
PADDLE_ENFORCE_EQ(
data_format,
"NCHW",
platform::errors::InvalidArgument("The Pool2d_grad XPU OP only support"
"data_format is 'NCHW', but received "
"%s",
data_format));
std::string padding_algorithm =
context.Attr<std::string>("padding_algorithm");
const int* index_data = nullptr;
PADDLE_ENFORCE_EQ(
ksize.size(),
2,
platform::errors::InvalidArgument("The Pool2d XPU OP only support 2 "
"dimension pooling!, but received "
"%d-dimension pool kernel size",
ksize.size()));
bool global_pooling = context.Attr<bool>("global_pooling") ||
(adaptive && (ksize[0] * ksize[1] == 1));
if (global_pooling) {
for (size_t i = 0; i < ksize.size(); ++i) {
paddings[i] = 0;
ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
}
}
if (!in_x_grad) {
return;
}
const int n = in_x->dims()[0];
const int c = in_x->dims()[1];
const int in_h = in_x->dims()[2];
const int in_w = in_x->dims()[3];
const int out_h = out->dims()[2];
const int out_w = out->dims()[3];
framework::DDim data_dims;
data_dims = phi::slice_ddim(in_x->dims(), 2, in_x->dims().size());
phi::funcs::UpdatePadding(&paddings,
global_pooling,
adaptive,
padding_algorithm,
data_dims,
strides,
ksize);
if (ceil_mode) {
int in_h_ceil = (out_h - 1) * strides[0] + ksize[0] - 2 * paddings[0];
int in_w_ceil = (out_w - 1) * strides[1] + ksize[1] - 2 * paddings[2];
paddings[1] += (in_h_ceil - in_h);
paddings[3] += (in_w_ceil - in_w);
}
auto input = reinterpret_cast<const XPUType*>(in_x->data<T>());
auto output = reinterpret_cast<const XPUType*>(out->data<T>());
auto output_grad = reinterpret_cast<const XPUType*>(out_grad->data<T>());
in_x_grad->mutable_data<T>(context.GetPlace());
auto input_grad = reinterpret_cast<XPUType*>(in_x_grad->data<T>());
auto& dev_ctx = context.template device_context<DeviceContext>();
int r = xpu::Error_t::SUCCESS;
if (adaptive) {
// floor for stride
strides = {in_h / out_h, in_w / out_w};
int kh = in_h - (out_h - 1) * strides[0];
int kw = in_w - (out_w - 1) * strides[1];
ksize = {kh, kw};
paddings = {0, 0, 0, 0};
}
if (pooling_type == "max") {
// TODO(zhanghuan05) to bind max_pool2d_grad_indices xpu api
r = xpu::max_pool2d_grad<XPUType>(dev_ctx.x_context(),
input,
output,
index_data,
output_grad,
input_grad,
n,
c,
in_h,
in_w,
ksize,
strides,
paddings,
true);
} else if (pooling_type == "avg") {
r = xpu::avg_pool2d_grad<XPUType>(dev_ctx.x_context(),
input,
output,
output_grad,
input_grad,
n,
c,
in_h,
in_w,
ksize,
strides,
paddings,
!exclusive,
true);
} else {
PADDLE_THROW(platform::errors::InvalidArgument(
"Unsupported pooling type for kunlun ", pooling_type));
}
PADDLE_ENFORCE_EQ(r,
xpu::Error_t::SUCCESS,
platform::errors::External(
"The Pool2dGrad XPU OP return wrong value[%d %s]",
r,
XPUAPIErrorMsg[r]));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_XPU_KERNEL(
pool2d,
ops::PoolXPUKernel<paddle::platform::XPUDeviceContext, float>,
ops::PoolXPUKernel<paddle::platform::XPUDeviceContext,
paddle::platform::float16>);
REGISTER_OP_XPU_KERNEL(
pool2d_grad,
ops::PoolGradXPUKernel<paddle::platform::XPUDeviceContext, float>,
ops::PoolGradXPUKernel<paddle::platform::XPUDeviceContext,
paddle::platform::float16>);
#endif
paddle/phi/kernels/xpu/pool_grad_kernel.cc 0 → 100644
浏览文件 @ 13a0ea4c
// 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/pool_grad_kernel.h"
#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/pooling.h"
namespace phi {
template <typename T, typename Context>
void Pool2dGradKernel(const Context& ctx,
const DenseTensor& x,
const DenseTensor& out,
const DenseTensor& dout,
const std::vector<int>& kernel_size_t,
const std::vector<int>& strides_t,
const std::vector<int>& paddings_t,
bool ceil_mode,
bool exclusive,
const std::string& data_format,
const std::string& pooling_type,
bool global_pooling,
bool adaptive,
const std::string& padding_algorithm,
DenseTensor* dx) {
using XPUType = typename XPUTypeTrait<T>::Type;
std::vector<int> paddings(paddings_t);
std::vector<int> kernel_size(kernel_size_t);
std::vector<int> strides(strides_t);
PADDLE_ENFORCE_EQ(
data_format,
"NCHW",
phi::errors::InvalidArgument("The Pool2d_grad XPU OP only support"
"data_format is 'NCHW', but received %s",
data_format));
PADDLE_ENFORCE_EQ(
kernel_size.size(),
2,
phi::errors::InvalidArgument("The Pool2d XPU OP only support 2 "
"dimension pooling!, but received "
"%d-dimension pool kernel size",
kernel_size.size()));
if (global_pooling) {
for (size_t i = 0; i < kernel_size.size(); ++i) {
paddings[i] = 0;
kernel_size[i] = static_cast<int>(x.dims()[i + 2]);
}
}
if (!dx) {
return;
}
const int n = x.dims()[0];
const int c = x.dims()[1];
const int in_h = x.dims()[2];
const int in_w = x.dims()[3];
const int out_h = out.dims()[2];
const int out_w = out.dims()[3];
DDim data_dims;
data_dims = slice_ddim(x.dims(), 2, x.dims().size());
funcs::UpdatePadding(&paddings,
global_pooling,
adaptive,
padding_algorithm,
data_dims,
strides,
kernel_size);
if (ceil_mode) {
int in_h_ceil = (out_h - 1) * strides[0] + kernel_size[0] - 2 * paddings[0];
int in_w_ceil = (out_w - 1) * strides[1] + kernel_size[1] - 2 * paddings[2];
paddings[1] += (in_h_ceil - in_h);
paddings[3] += (in_w_ceil - in_w);
}
ctx.template Alloc<T>(dx);
const int* index_data = nullptr;
int r = xpu::Error_t::SUCCESS;
if (adaptive) {
// floor for stride
strides = {in_h / out_h, in_w / out_w};
int kh = in_h - (out_h - 1) * strides[0];
int kw = in_w - (out_w - 1) * strides[1];
kernel_size = {kh, kw};
paddings = {0, 0, 0, 0};
}
if (pooling_type == "max") {
// TODO(zhanghuan05) to bind max_pool2d_grad_indices xpu api
r = xpu::max_pool2d_grad<XPUType>(
ctx.x_context(),
reinterpret_cast<const XPUType*>(x.data<T>()),
reinterpret_cast<const XPUType*>(out.data<T>()),
index_data,
reinterpret_cast<const XPUType*>(dout.data<T>()),
reinterpret_cast<XPUType*>(dx->data<T>()),
n,
c,
in_h,
in_w,
kernel_size,
strides,
paddings,
true);
} else if (pooling_type == "avg") {
r = xpu::avg_pool2d_grad<XPUType>(
ctx.x_context(),
reinterpret_cast<const XPUType*>(x.data<T>()),
reinterpret_cast<const XPUType*>(out.data<T>()),
reinterpret_cast<const XPUType*>(dout.data<T>()),
reinterpret_cast<XPUType*>(dx->data<T>()),
n,
c,
in_h,
in_w,
kernel_size,
strides,
paddings,
!exclusive,
true);
} else {
PADDLE_THROW(phi::errors::InvalidArgument(
"Unsupported pooling type for kunlun ", pooling_type));
}
PADDLE_ENFORCE_XDNN_SUCCESS(r, "pool2dgrad");
}
} // namespace phi
PD_REGISTER_KERNEL(pool2d_grad,
XPU,
ALL_LAYOUT,
phi::Pool2dGradKernel,
float,
phi::dtype::float16) {}
paddle/phi/kernels/xpu/pool_kernel.cc 0 → 100644
浏览文件 @ 13a0ea4c
// 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/pool_kernel.h"
#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/pooling.h"
namespace phi {
template <typename T, typename Context>
void Pool2dKernel(const Context& ctx,
const DenseTensor& x,
const std::vector<int>& kernel_size_t,
const std::vector<int>& strides,
const std::vector<int>& paddings_t,
bool ceil_mode,
bool exclusive,
const std::string& data_format,
const std::string& pooling_type,
bool global_pooling,
bool adaptive,
const std::string& padding_algorithm,
DenseTensor* out) {
using XPUType = typename XPUTypeTrait<T>::Type;
std::vector<int> paddings(paddings_t);
std::vector<int> kernel_size(kernel_size_t);
PADDLE_ENFORCE_EQ(kernel_size.size(),
2,
phi::errors::InvalidArgument(
"The Pool2d XPU OP only support 2 dimension pooling!"));
PADDLE_ENFORCE_EQ(
data_format,
"NCHW",
phi::errors::InvalidArgument("The Pool2d XPU OP only support "
"data_format is 'NCHW', but received %s",
data_format));
if (global_pooling) {
for (size_t i = 0; i < kernel_size.size(); ++i) {
paddings[i] = 0;
kernel_size[i] = static_cast<int>(x.dims()[i + 2]);
}
}
const int n = x.dims()[0];
const int c = x.dims()[1];
const int in_h = x.dims()[2];
const int in_w = x.dims()[3];
const int out_h = out->dims()[2];
const int out_w = out->dims()[3];
DDim data_dims;
data_dims = slice_ddim(x.dims(), 2, x.dims().size());
funcs::UpdatePadding(&paddings,
global_pooling,
adaptive,
padding_algorithm,
data_dims,
strides,
kernel_size);
if (ceil_mode) {
int in_h_ceil = (out_h - 1) * strides[0] + kernel_size[0] - 2 * paddings[0];
int in_w_ceil = (out_w - 1) * strides[1] + kernel_size[1] - 2 * paddings[2];
paddings[1] += (in_h_ceil - in_h);
paddings[3] += (in_w_ceil - in_w);
}
ctx.template Alloc<T>(out);
int* index_data = nullptr;
int r = xpu::Error_t::SUCCESS;
if (!adaptive) {
if (pooling_type == "max") {
r = xpu::max_pool2d<XPUType>(
ctx.x_context(),
reinterpret_cast<const XPUType*>(x.data<T>()),
reinterpret_cast<XPUType*>(out->data<T>()),
index_data,
n,
c,
in_h,
in_w,
kernel_size,
strides,
paddings,
true);
} else if (pooling_type == "avg") {
r = xpu::avg_pool2d<XPUType>(
ctx.x_context(),
reinterpret_cast<const XPUType*>(x.data<T>()),
reinterpret_cast<XPUType*>(out->data<T>()),
n,
c,
in_h,
in_w,
kernel_size,
strides,
paddings,
!exclusive,
true);
} else {
PADDLE_THROW(phi::errors::InvalidArgument(
"Unsupported pooling type for kunlun ", pooling_type));
}
} else {
if (pooling_type == "max") {
r = xpu::adaptive_max_pool2d<XPUType>(
ctx.x_context(),
reinterpret_cast<const XPUType*>(x.data<T>()),
reinterpret_cast<XPUType*>(out->data<T>()),
index_data,
n,
c,
in_h,
in_w,
out_h,
out_w,
true);
} else if (pooling_type == "avg") {
r = xpu::adaptive_avg_pool2d<XPUType>(
ctx.x_context(),
reinterpret_cast<const XPUType*>(x.data<T>()),
reinterpret_cast<XPUType*>(out->data<T>()),
n,
c,
in_h,
in_w,
out_h,
out_w,
true);
} else {
PADDLE_THROW(phi::errors::InvalidArgument(
"Unsupported pooling type for kunlun ", pooling_type));
}
}
PADDLE_ENFORCE_XDNN_SUCCESS(r, "pool2d");
}
} // namespace phi
PD_REGISTER_KERNEL(
pool2d, XPU, ALL_LAYOUT, phi::Pool2dKernel, float, phi::dtype::float16) {}
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