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
7aa4d879
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 2 年 前同步成功
通知
2325
Star
20933
Fork
5424
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
未验证
提交
7aa4d879
编写于
9月 08, 2021
作者:
L
Leo Chen
提交者:
GitHub
9月 08, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add clip_by_norm fp16 kernel (#35446)
* add clip_by_norm fp16 kernel * add ut
上级
28abd5d8
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
150 addition
and
2 deletion
+150
-2
paddle/fluid/operators/clip_by_norm_op.cu
paddle/fluid/operators/clip_by_norm_op.cu
+116
-1
python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py
python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py
+34
-1
未找到文件。
paddle/fluid/operators/clip_by_norm_op.cu
浏览文件 @
7aa4d879
...
@@ -13,8 +13,123 @@ See the License for the specific language governing permissions and
...
@@ -13,8 +13,123 @@ See the License for the specific language governing permissions and
limitations under the License. */
limitations under the License. */
#include "paddle/fluid/operators/clip_by_norm_op.h"
#include "paddle/fluid/operators/clip_by_norm_op.h"
#include "paddle/fluid/operators/reduce_ops/reduce_op.cu.h"
namespace
paddle
{
namespace
operators
{
using
Tensor
=
framework
::
Tensor
;
template
<
typename
Tx
,
typename
Ty
=
Tx
>
struct
SquareTransformer
{
HOSTDEVICE
explicit
inline
SquareTransformer
(
int
n
)
{}
HOSTDEVICE
inline
Ty
operator
()(
const
Tx
&
x
)
const
{
return
static_cast
<
Ty
>
(
x
)
*
static_cast
<
Ty
>
(
x
);
}
HOSTDEVICE
inline
Ty
operator
()(
const
Tx
*
x
)
const
{
return
static_cast
<
Ty
>
(
x
[
0
])
*
static_cast
<
Ty
>
(
x
[
0
]);
}
};
template
<
typename
Tx
,
typename
Ty
=
Tx
>
struct
SquareSum
{
using
Transformer
=
SquareTransformer
<
Tx
,
Ty
>
;
inline
Ty
initial
()
{
return
static_cast
<
Ty
>
(
0.0
f
);
}
__device__
__forceinline__
Ty
operator
()(
const
Ty
&
a
,
const
Ty
&
b
)
const
{
return
b
+
a
;
}
};
template
<
>
class
ClipByNormKernel
<
platform
::
CUDADeviceContext
,
platform
::
float16
>
:
public
framework
::
OpKernel
<
platform
::
float16
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
max_norm
=
context
.
Attr
<
float
>
(
"max_norm"
);
auto
in_var
=
context
.
InputVar
(
"X"
);
auto
&
dev_ctx
=
context
.
template
device_context
<
platform
::
CUDADeviceContext
>();
Tensor
*
output
=
nullptr
;
const
Tensor
*
input
=
nullptr
;
if
(
in_var
->
IsType
<
framework
::
LoDTensor
>
())
{
input
=
context
.
Input
<
Tensor
>
(
"X"
);
output
=
context
.
Output
<
Tensor
>
(
"Out"
);
output
->
mutable_data
<
platform
::
float16
>
(
context
.
GetPlace
());
}
else
if
(
in_var
->
IsType
<
SelectedRows
>
())
{
auto
*
x
=
context
.
Input
<
SelectedRows
>
(
"X"
);
// merge ids in selected rows first
math
::
scatter
::
MergeAdd
<
platform
::
CUDADeviceContext
,
platform
::
float16
>
merge_func
;
SelectedRows
*
merged_input
=
const_cast
<
framework
::
Scope
&>
(
context
.
scope
())
.
Var
()
->
GetMutable
<
SelectedRows
>
();
merge_func
(
context
.
template
device_context
<
platform
::
CUDADeviceContext
>(),
*
x
,
merged_input
);
input
=
&
(
merged_input
->
value
());
SelectedRows
*
output_selected_rows
=
context
.
Output
<
SelectedRows
>
(
"Out"
);
output_selected_rows
->
set_rows
(
merged_input
->
rows
());
output_selected_rows
->
set_height
(
merged_input
->
height
());
output
=
output_selected_rows
->
mutable_value
();
output
->
Resize
(
merged_input
->
value
().
dims
());
output
->
mutable_data
<
platform
::
float16
>
(
context
.
GetPlace
());
}
else
{
PADDLE_THROW
(
platform
::
errors
::
InvalidArgument
(
"Invalid input variable type, only support LodTensor and "
"SelectedRows types, but got type is %s."
,
framework
::
ToTypeName
(
in_var
->
Type
())));
}
PADDLE_ENFORCE_NOT_NULL
(
input
,
platform
::
errors
::
InvalidArgument
(
"Input(X) of ClipByNormOp should not be null. "
"Please check if it is created correctly."
));
std
::
vector
<
int
>
reduce_dims
;
reduce_dims
.
resize
(
input
->
dims
().
size
());
for
(
int
i
=
0
;
i
<
reduce_dims
.
size
();
++
i
)
{
reduce_dims
[
i
]
=
i
;
}
Tensor
tmp
=
context
.
AllocateTmpTensor
<
float
,
platform
::
CUDADeviceContext
>
(
{
1
},
dev_ctx
);
TensorReduceFunctorImpl
<
platform
::
float16
,
float
,
SquareSum
>
(
*
input
,
&
tmp
,
reduce_dims
,
dev_ctx
.
stream
());
auto
tmp_eigen
=
EigenVector
<
float
>::
Flatten
(
tmp
);
auto
x_norm
=
tmp_eigen
.
sqrt
();
auto
x
=
EigenVector
<
platform
::
float16
>::
Flatten
(
*
input
);
auto
out
=
EigenVector
<
platform
::
float16
>::
Flatten
(
*
output
);
auto
&
place
=
*
context
.
template
device_context
<
platform
::
CUDADeviceContext
>()
.
eigen_device
();
auto
temp
=
(
x_norm
<=
max_norm
).
template
cast
<
float
>();
auto
epsilon
=
((
x_norm
<=
static_cast
<
float
>
(
1e-30
)).
all
().
template
cast
<
float
>())
*
static_cast
<
float
>
(
1e-6
);
auto
scaling
=
(
temp
+
(
static_cast
<
float
>
(
1
)
-
temp
)
*
max_norm
/
(
x_norm
+
epsilon
))
.
template
cast
<
platform
::
float16
>();
Eigen
::
array
<
int
,
1
>
one_dim
{{
1
}};
Eigen
::
DSizes
<
int
,
1
>
m_dsize
(
input
->
numel
());
out
.
device
(
place
)
=
x
*
scaling
.
reshape
(
one_dim
).
broadcast
(
m_dsize
);
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
namespace
ops
=
paddle
::
operators
;
namespace
plat
=
paddle
::
platform
;
REGISTER_OP_CUDA_KERNEL
(
REGISTER_OP_CUDA_KERNEL
(
clip_by_norm
,
clip_by_norm
,
ops
::
ClipByNormKernel
<
paddle
::
platform
::
CUDADeviceContext
,
float
>
);
ops
::
ClipByNormKernel
<
paddle
::
platform
::
CUDADeviceContext
,
float
>
,
ops
::
ClipByNormKernel
<
paddle
::
platform
::
CUDADeviceContext
,
plat
::
float16
>
);
python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py
浏览文件 @
7aa4d879
...
@@ -25,8 +25,9 @@ import paddle.fluid.core as core
...
@@ -25,8 +25,9 @@ import paddle.fluid.core as core
class
TestClipByNormOp
(
OpTest
):
class
TestClipByNormOp
(
OpTest
):
def
setUp
(
self
):
def
setUp
(
self
):
self
.
max_relative_error
=
0.006
self
.
max_relative_error
=
0.006
self
.
init_dtype
()
self
.
initTestCase
()
self
.
initTestCase
()
input
=
np
.
random
.
random
(
self
.
shape
).
astype
(
"float32"
)
input
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
input
[
np
.
abs
(
input
)
<
self
.
max_relative_error
]
=
0.5
input
[
np
.
abs
(
input
)
<
self
.
max_relative_error
]
=
0.5
self
.
op_type
=
"clip_by_norm"
self
.
op_type
=
"clip_by_norm"
self
.
inputs
=
{
'X'
:
input
,
}
self
.
inputs
=
{
'X'
:
input
,
}
...
@@ -46,6 +47,9 @@ class TestClipByNormOp(OpTest):
...
@@ -46,6 +47,9 @@ class TestClipByNormOp(OpTest):
self
.
shape
=
(
100
,
)
self
.
shape
=
(
100
,
)
self
.
max_norm
=
1.0
self
.
max_norm
=
1.0
def
init_dtype
(
self
):
self
.
dtype
=
np
.
float32
class
TestCase1
(
TestClipByNormOp
):
class
TestCase1
(
TestClipByNormOp
):
def
initTestCase
(
self
):
def
initTestCase
(
self
):
...
@@ -65,6 +69,35 @@ class TestCase3(TestClipByNormOp):
...
@@ -65,6 +69,35 @@ class TestCase3(TestClipByNormOp):
self
.
max_norm
=
1.0
self
.
max_norm
=
1.0
class
TestClipByNormOpFp16
(
TestClipByNormOp
):
def
init_dtype
(
self
):
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
if
core
.
is_compiled_with_cuda
():
place
=
core
.
CUDAPlace
(
0
)
if
core
.
is_float16_supported
(
place
):
self
.
check_output_with_place
(
place
,
atol
=
0.001
)
class
TestClipByNormOpFp16Case1
(
TestClipByNormOpFp16
):
def
initTestCase
(
self
):
self
.
shape
=
(
100
,
)
self
.
max_norm
=
1e20
class
TestClipByNormOpFp16Case2
(
TestClipByNormOpFp16
):
def
initTestCase
(
self
):
self
.
shape
=
(
16
,
16
)
self
.
max_norm
=
0.1
class
TestClipByNormOpFp16Case3
(
TestClipByNormOpFp16
):
def
initTestCase
(
self
):
self
.
shape
=
(
4
,
8
,
16
)
self
.
max_norm
=
1.0
class
TestClipByNormOpWithSelectedRows
(
unittest
.
TestCase
):
class
TestClipByNormOpWithSelectedRows
(
unittest
.
TestCase
):
def
check_with_place
(
self
,
place
):
def
check_with_place
(
self
,
place
):
self
.
config_test_case
()
self
.
config_test_case
()
...
...
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