Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
PaddlePaddle
Paddle
提交
1137677a
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
1137677a
编写于
9月 06, 2022
作者:
H
houj04
提交者:
GitHub
9月 06, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[XPU] rmsprop to phi. (#45734)
上级
269bd1fe
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
85 addition
and
145 deletion
+85
-145
paddle/fluid/operators/optimizers/rmsprop_op_xpu.cc
paddle/fluid/operators/optimizers/rmsprop_op_xpu.cc
+0
-145
paddle/phi/kernels/xpu/rmsprop_kernel.cc
paddle/phi/kernels/xpu/rmsprop_kernel.cc
+85
-0
未找到文件。
paddle/fluid/operators/optimizers/rmsprop_op_xpu.cc
已删除
100644 → 0
浏览文件 @
269bd1fe
/* 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. */
#ifdef PADDLE_WITH_XPU
#include <gflags/gflags.h>
#include <iostream>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/platform/device/device_wrapper.h"
namespace
paddle
{
namespace
operators
{
static
inline
float
GetAttrFromTensor
(
const
framework
::
Tensor
*
tensor
)
{
const
float
*
tensor_data
=
tensor
->
data
<
float
>
();
framework
::
Tensor
cpu_tensor
;
if
(
platform
::
is_gpu_place
(
tensor
->
place
())
||
platform
::
is_xpu_place
(
tensor
->
place
()))
{
paddle
::
framework
::
TensorCopySync
(
*
tensor
,
platform
::
CPUPlace
(),
&
cpu_tensor
);
tensor_data
=
cpu_tensor
.
data
<
float
>
();
}
return
tensor_data
[
0
];
}
using
framework
::
OpKernelType
;
using
framework
::
Tensor
;
template
<
typename
DeviceContext
,
typename
T
>
class
RmspropOpXPUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
using
paddle
::
framework
::
LoDTensor
;
// check Param & Grad tensor type
const
auto
*
param_var
=
ctx
.
InputVar
(
"Param"
);
PADDLE_ENFORCE_EQ
(
param_var
->
IsType
<
LoDTensor
>
(),
true
,
platform
::
errors
::
InvalidArgument
(
"Tensor holds the wrong type,Expected Var(%s)'s "
"type is LoDTensor, "
"but the received is %s"
,
ctx
.
InputNames
(
"Param"
).
front
(),
framework
::
ToTypeName
(
param_var
->
Type
())));
const
auto
*
grad_var
=
ctx
.
InputVar
(
"Grad"
);
PADDLE_ENFORCE_EQ
(
grad_var
->
IsType
<
LoDTensor
>
(),
true
,
platform
::
errors
::
InvalidArgument
(
"Tensor holds the wrong type,Expected Var(%s)'s "
"type is LoDTensor, "
"but the received is %s"
,
ctx
.
InputNames
(
"Grad"
).
front
(),
framework
::
ToTypeName
(
grad_var
->
Type
())));
// inputs
auto
&
param
=
GET_DATA_SAFELY
(
ctx
.
Input
<
LoDTensor
>
(
"Param"
),
"Input"
,
"Param"
,
"Rmsprop"
);
auto
&
meanSquare
=
GET_DATA_SAFELY
(
ctx
.
Input
<
LoDTensor
>
(
"MeanSquare"
),
"Input"
,
"MeanSquare"
,
"Rmsprop"
);
auto
&
grad
=
GET_DATA_SAFELY
(
ctx
.
Input
<
LoDTensor
>
(
"Grad"
),
"Input"
,
"Grad"
,
"Rmsprop"
);
auto
&
mom
=
GET_DATA_SAFELY
(
ctx
.
Input
<
LoDTensor
>
(
"Moment"
),
"Input"
,
"Moment"
,
"Rmsprop"
);
auto
*
learning_rate
=
ctx
.
Input
<
Tensor
>
(
"LearningRate"
);
PADDLE_ENFORCE_EQ
(
learning_rate
->
dims
().
size
(),
1
,
platform
::
errors
::
InvalidArgument
(
"learining rate should have dimension = 1."
" But received learning rate dim [%s] "
,
learning_rate
->
dims
().
size
()));
T
lr
=
static_cast
<
T
>
(
GetAttrFromTensor
(
learning_rate
));
// constants
T
epsilon
=
static_cast
<
T
>
(
ctx
.
Attr
<
float
>
(
"epsilon"
));
T
decay
=
static_cast
<
T
>
(
ctx
.
Attr
<
float
>
(
"decay"
));
T
momentum
=
static_cast
<
T
>
(
ctx
.
Attr
<
float
>
(
"momentum"
));
bool
centered
=
ctx
.
Attr
<
bool
>
(
"centered"
);
PADDLE_ENFORCE_EQ
(
centered
,
false
,
platform
::
errors
::
Unimplemented
(
"centered=True is not supported in the xpu kernel of "
"rmsprop. use XPU_BLACK_LIST to disable this op."
));
/*
TODO(houj04): when XDNN api supports 'center', add input of
mean_grad_input and output of mean_grad_output. auto *mean_grad_input =
ctx.Input<Tensor>("MeanGrad"); auto *mean_grad_output =
ctx.Output<Tensor>("MeanGradOut");
*/
// outputs
auto
&
param_out
=
GET_DATA_SAFELY
(
ctx
.
Output
<
LoDTensor
>
(
"ParamOut"
),
"Output"
,
"ParamOut"
,
"Rmsprop"
);
auto
&
mom_out
=
GET_DATA_SAFELY
(
ctx
.
Output
<
LoDTensor
>
(
"MomentOut"
),
"Output"
,
"MomentOut"
,
"Rmsprop"
);
auto
&
mom_sqrt_out
=
GET_DATA_SAFELY
(
ctx
.
Output
<
LoDTensor
>
(
"MeanSquareOut"
),
"Output"
,
"MeanSquareOut"
,
"Rmsprop"
);
auto
&
dev_ctx
=
ctx
.
template
device_context
<
DeviceContext
>();
// int rmsprop(Context* ctx, const T* g, const T* p, const float* ms, const
// float* mom, T* p_out, float* ms_out, float* mom_out, float epsilon, float
// rho, float momentum, float lr, int n);
int
r
=
xpu
::
rmsprop
(
dev_ctx
.
x_context
(),
grad
.
template
data
<
T
>(),
param
.
template
data
<
T
>(),
meanSquare
.
template
data
<
T
>(),
mom
.
template
data
<
T
>(),
param_out
.
template
mutable_data
<
T
>(
ctx
.
GetPlace
()),
mom_sqrt_out
.
template
mutable_data
<
T
>(
ctx
.
GetPlace
()),
mom_out
.
template
mutable_data
<
T
>(
ctx
.
GetPlace
()),
epsilon
,
decay
,
momentum
,
lr
,
param
.
numel
());
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"rmsprop"
);
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_XPU_KERNEL
(
rmsprop
,
ops
::
RmspropOpXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
);
#endif
paddle/phi/kernels/xpu/rmsprop_kernel.cc
0 → 100644
浏览文件 @
1137677a
// 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/rmsprop_kernel.h"
#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/fluid/memory/memcpy.h"
namespace
phi
{
template
<
typename
T
,
typename
Context
>
void
RmspropDenseKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
param
,
const
DenseTensor
&
mean_square
,
const
DenseTensor
&
grad
,
const
DenseTensor
&
moment
,
const
DenseTensor
&
learning_rate
,
const
paddle
::
optional
<
DenseTensor
>&
mean_grad
,
float
epsilon
,
float
decay
,
float
momentum
,
bool
centered
,
DenseTensor
*
param_out
,
DenseTensor
*
moment_out
,
DenseTensor
*
mean_square_out
,
DenseTensor
*
mean_grad_out
)
{
// check input
PADDLE_ENFORCE_EQ
(
centered
,
false
,
errors
::
Unimplemented
(
"centered=True is not supported in the xpu kernel of "
"rmsprop. use XPU_BLACK_LIST to disable this op."
));
// copy learning_rate to cpu
PADDLE_ENFORCE_EQ
(
learning_rate
.
dims
().
size
(),
1
,
errors
::
InvalidArgument
(
"learining rate should have dimension = 1."
" But received learning rate dim [%s] "
,
learning_rate
.
dims
().
size
()));
T
learning_rate_cpu
=
0.0
f
;
paddle
::
memory
::
Copy
(
CPUPlace
(),
static_cast
<
void
*>
(
&
learning_rate_cpu
),
dev_ctx
.
GetPlace
(),
static_cast
<
const
void
*>
(
learning_rate
.
data
()),
sizeof
(
T
));
// alloc output
dev_ctx
.
template
Alloc
<
T
>(
param_out
);
dev_ctx
.
template
Alloc
<
T
>(
moment_out
);
dev_ctx
.
template
Alloc
<
T
>(
mean_square_out
);
// int rmsprop(Context* ctx, const T* g, const T* p, const float* ms, const
// float* mom, T* p_out, float* ms_out, float* mom_out, float epsilon, float
// rho, float momentum, float lr, int n);
int
r
=
xpu
::
rmsprop
(
dev_ctx
.
x_context
(),
grad
.
data
<
T
>
(),
param
.
data
<
T
>
(),
mean_square
.
data
<
T
>
(),
moment
.
data
<
T
>
(),
param_out
->
data
<
T
>
(),
mean_square_out
->
data
<
T
>
(),
moment_out
->
data
<
T
>
(),
epsilon
,
decay
,
momentum
,
learning_rate_cpu
,
param
.
numel
());
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"rmsprop"
);
}
}
// namespace phi
PD_REGISTER_KERNEL
(
rmsprop
,
XPU
,
ALL_LAYOUT
,
phi
::
RmspropDenseKernel
,
float
)
{}
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录