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
910e1b6a
P
Paddle
项目概览
PaddlePaddle
/
Paddle
接近 2 年 前同步成功
通知
2323
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
910e1b6a
编写于
10月 13, 2022
作者:
X
xiaohemaikoo
提交者:
GitHub
10月 13, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
logsumexp support fp16 (#45817)
上级
e86dbd62
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
111 addition
and
18 deletion
+111
-18
paddle/phi/kernels/gpu/logsumexp_grad_kernel.cu
paddle/phi/kernels/gpu/logsumexp_grad_kernel.cu
+10
-2
paddle/phi/kernels/gpu/logsumexp_kernel.cu
paddle/phi/kernels/gpu/logsumexp_kernel.cu
+4
-1
paddle/phi/kernels/impl/logsumexp_grad_kernel_impl.h
paddle/phi/kernels/impl/logsumexp_grad_kernel_impl.h
+20
-7
paddle/phi/kernels/impl/logsumexp_kernel_impl.h
paddle/phi/kernels/impl/logsumexp_kernel_impl.h
+19
-8
python/paddle/fluid/tests/unittests/test_logsumexp.py
python/paddle/fluid/tests/unittests/test_logsumexp.py
+58
-0
未找到文件。
paddle/phi/kernels/gpu/logsumexp_grad_kernel.cu
浏览文件 @
910e1b6a
...
...
@@ -15,8 +15,16 @@
#include "paddle/phi/kernels/logsumexp_grad_kernel.h"
#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/common/float16.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/impl/logsumexp_grad_kernel_impl.h"
PD_REGISTER_KERNEL
(
logsumexp_grad
,
GPU
,
ALL_LAYOUT
,
phi
::
LogsumexpGradKernel
,
float
,
double
)
{}
using
float16
=
phi
::
dtype
::
float16
;
PD_REGISTER_KERNEL
(
logsumexp_grad
,
GPU
,
ALL_LAYOUT
,
phi
::
LogsumexpGradKernel
,
float
,
double
,
float16
)
{}
paddle/phi/kernels/gpu/logsumexp_kernel.cu
浏览文件 @
910e1b6a
...
...
@@ -15,8 +15,11 @@
#include "paddle/phi/kernels/logsumexp_kernel.h"
#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/common/float16.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/impl/logsumexp_kernel_impl.h"
using
float16
=
phi
::
dtype
::
float16
;
PD_REGISTER_KERNEL
(
logsumexp
,
GPU
,
ALL_LAYOUT
,
phi
::
LogsumexpKernel
,
float
,
double
)
{}
logsumexp
,
GPU
,
ALL_LAYOUT
,
phi
::
LogsumexpKernel
,
float
,
double
,
float16
)
{}
paddle/phi/kernels/impl/logsumexp_grad_kernel_impl.h
浏览文件 @
910e1b6a
...
...
@@ -16,6 +16,7 @@
#include <type_traits>
#include <vector>
#include "paddle/phi/common/amp_type_traits.h"
#include "paddle/phi/kernels/funcs/eigen/common.h"
#include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
#include "paddle/phi/kernels/funcs/reduce_grad_functions.h"
...
...
@@ -23,6 +24,7 @@
namespace
phi
{
template
<
typename
T
>
struct
LogsumexpGradFunctor
{
template
<
typename
Context
,
typename
X
,
...
...
@@ -37,7 +39,13 @@ struct LogsumexpGradFunctor {
DY
*
dy
,
const
Dim
&
dim
,
int
size
)
{
dx
->
device
(
place
)
=
dy
->
broadcast
(
dim
)
*
(
*
x
-
y
->
broadcast
(
dim
)).
exp
();
using
MT
=
typename
phi
::
dtype
::
MPTypeTrait
<
T
>::
Type
;
auto
x_mt
=
(
*
x
).
template
cast
<
MT
>();
auto
y_mt
=
(
*
y
).
template
cast
<
MT
>();
auto
dy_mt
=
(
*
dy
).
template
cast
<
MT
>();
dx
->
device
(
place
)
=
(
dy_mt
.
broadcast
(
dim
)
*
(
x_mt
-
y_mt
.
broadcast
(
dim
)).
exp
())
.
template
cast
<
T
>();
}
};
...
...
@@ -62,11 +70,11 @@ void LogsumexpGradKernel(const Context& dev_ctx,
auto
dx
=
phi
::
EigenVector
<
T
>::
Flatten
(
*
in_grad
);
auto
&
place
=
*
dev_ctx
.
eigen_device
();
auto
broadcast_dim
=
Eigen
::
array
<
int
,
1
>
({{
static_cast
<
int
>
(
in
.
numel
())}});
LogsumexpGradFunctor
()(
LogsumexpGradFunctor
<
T
>
()(
place
,
&
x
,
&
y
,
&
dx
,
&
dy
,
broadcast_dim
,
broadcast_dim
[
0
]);
}
else
{
int
rank
=
in
.
dims
().
size
();
LogsumexpGradFunctor
functor
;
LogsumexpGradFunctor
<
T
>
functor
;
std
::
vector
<
int32_t
>
axis32
;
axis32
.
reserve
(
axis
.
size
());
std
::
for_each
(
axis
.
begin
(),
axis
.
end
(),
[
&
axis32
](
const
int64_t
&
t
)
{
...
...
@@ -74,21 +82,26 @@ void LogsumexpGradKernel(const Context& dev_ctx,
});
switch
(
rank
)
{
case
1
:
phi
::
funcs
::
ReduceGradFunctor
<
Context
,
T
,
1
,
LogsumexpGradFunctor
>
(
phi
::
funcs
::
ReduceGradFunctor
<
Context
,
T
,
1
,
LogsumexpGradFunctor
<
T
>
>
(
dev_ctx
,
in
,
out
,
out_grad
,
in_grad
,
functor
,
axis32
);
break
;
case
2
:
phi
::
funcs
::
ReduceGradFunctor
<
Context
,
T
,
2
,
LogsumexpGradFunctor
>
(
phi
::
funcs
::
ReduceGradFunctor
<
Context
,
T
,
2
,
LogsumexpGradFunctor
<
T
>
>
(
dev_ctx
,
in
,
out
,
out_grad
,
in_grad
,
functor
,
axis32
);
break
;
case
3
:
phi
::
funcs
::
ReduceGradFunctor
<
Context
,
T
,
3
,
LogsumexpGradFunctor
>
(
phi
::
funcs
::
ReduceGradFunctor
<
Context
,
T
,
3
,
LogsumexpGradFunctor
<
T
>
>
(
dev_ctx
,
in
,
out
,
out_grad
,
in_grad
,
functor
,
axis32
);
break
;
case
4
:
phi
::
funcs
::
ReduceGradFunctor
<
Context
,
T
,
4
,
LogsumexpGradFunctor
>
(
phi
::
funcs
::
ReduceGradFunctor
<
Context
,
T
,
4
,
LogsumexpGradFunctor
<
T
>
>
(
dev_ctx
,
in
,
out
,
out_grad
,
in_grad
,
functor
,
axis32
);
break
;
default:
PADDLE_THROW
(
phi
::
errors
::
Unimplemented
(
"Unsupported dimensions, please keep maximum dimensions of input "
"data less than 4."
));
break
;
}
}
}
...
...
paddle/phi/kernels/impl/logsumexp_kernel_impl.h
浏览文件 @
910e1b6a
...
...
@@ -16,6 +16,7 @@
#include <type_traits>
#include <vector>
#include "paddle/phi/common/amp_type_traits.h"
#include "paddle/phi/kernels/funcs/eigen/common.h"
#include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
#include "paddle/phi/kernels/funcs/reduce_function.h"
...
...
@@ -23,15 +24,17 @@
namespace
phi
{
#define HANDLE_DIM(NDIM, RDIM) \
if (ndim == NDIM && rdim == RDIM) { \
funcs::ReduceFunctor<Context, T, NDIM, RDIM, LogsumexpFunctor>( \
dev_ctx, x, out, axis, keepdim); \
#define HANDLE_DIM(NDIM, RDIM)
\
if (ndim == NDIM && rdim == RDIM) {
\
funcs::ReduceFunctor<Context, T, NDIM, RDIM, LogsumexpFunctor
<T>
>( \
dev_ctx, x, out, axis, keepdim);
\
}
template
<
typename
T
>
struct
LogsumexpFunctor
{
template
<
typename
Context
,
typename
X
,
typename
Y
,
typename
Dim
>
void
operator
()(
const
Context
&
place
,
X
*
x
,
Y
*
y
,
const
Dim
&
dim
)
{
using
MT
=
typename
phi
::
dtype
::
MPTypeTrait
<
T
>::
Type
;
auto
x_dim
=
x
->
dimensions
();
auto
t_dim
=
x_dim
;
for
(
int
i
=
0
;
i
<
static_cast
<
int
>
(
dim
.
size
());
i
++
)
{
...
...
@@ -46,12 +49,14 @@ struct LogsumexpFunctor {
r_dim
[
dim
[
i
]]
=
x_dim
[
dim
[
i
]];
}
auto
x_mt
=
(
*
x
).
template
cast
<
MT
>();
auto
y_dim
=
y
->
dimensions
();
auto
x_max
=
x
->
maximum
(
dim
);
auto
x_max
=
x
_mt
.
maximum
(
dim
);
y
->
device
(
place
)
=
(
x_max
+
(
*
x
-
x_max
.
reshape
(
t_dim
).
broadcast
(
r_dim
)).
exp
().
sum
(
dim
).
log
())
.
reshape
(
y_dim
);
(
x_mt
-
x_max
.
reshape
(
t_dim
).
broadcast
(
r_dim
)).
exp
().
sum
(
dim
).
log
())
.
reshape
(
y_dim
)
.
template
cast
<
T
>();
}
};
...
...
@@ -74,10 +79,16 @@ void LogsumexpKernel(const Context& dev_ctx,
auto
output
=
phi
::
EigenScalar
<
T
>::
From
(
*
out
);
auto
&
place
=
*
dev_ctx
.
eigen_device
();
auto
reduce_dim
=
Eigen
::
array
<
int
,
1
>
({{
0
}});
LogsumexpFunctor
()(
place
,
&
input
,
&
output
,
reduce_dim
);
LogsumexpFunctor
<
T
>
()(
place
,
&
input
,
&
output
,
reduce_dim
);
}
else
{
int
ndim
=
input_dim_size
;
int
rdim
=
axis
.
size
();
if
(
ndim
>
4
)
{
PADDLE_THROW
(
phi
::
errors
::
Unimplemented
(
"Unsupported dimensions, please keep maximum dimensions of input "
"data less than 4."
));
}
// comments for accelerating compiling temporarily.
// HANDLE_DIM(6, 5);
// HANDLE_DIM(6, 4);
...
...
python/paddle/fluid/tests/unittests/test_logsumexp.py
浏览文件 @
910e1b6a
...
...
@@ -15,6 +15,7 @@
import
paddle
import
unittest
import
numpy
as
np
import
paddle.fluid.core
as
core
from
op_test
import
OpTest
...
...
@@ -35,6 +36,22 @@ def logsumexp_wrapper(x, axis=None, keepdim=False, allreduce=False):
return
paddle
.
logsumexp
(
x
,
axis
,
keepdim
)
def
logsumexp_op_grad
(
x
,
axis
=
None
,
keepdim
=
False
,
reduce_all
=
False
):
paddle
.
disable_static
()
tensor_x
=
paddle
.
to_tensor
(
x
)
tensor_x
.
stop_gradient
=
False
out
=
logsumexp_wrapper
(
tensor_x
,
axis
,
keepdim
,
reduce_all
)
grad
=
paddle
.
grad
(
out
,
[
tensor_x
])
x_grad
=
grad
[
0
].
numpy
()
paddle
.
enable_static
()
return
x_grad
def
logsumexp_ref_grad
(
x
):
sum
=
np
.
exp
(
x
).
sum
()
return
np
.
exp
(
x
)
/
sum
class
TestLogsumexp
(
OpTest
):
def
setUp
(
self
):
...
...
@@ -125,6 +142,47 @@ class TestLogsumexp_reduce_all(TestLogsumexp):
self
.
user_defined_grad_outputs
=
[
np
.
ones
(
1
,
dtype
=
self
.
dtype
)]
class
TestLogsumexp_FP32
(
TestLogsumexp
):
def
set_attrs
(
self
):
self
.
dtype
=
'float32'
def
test_check_grad
(
self
):
self
.
__class__
.
dtype
=
self
.
dtype
x_grad
=
logsumexp_op_grad
(
self
.
inputs
[
'X'
])
ref_x_grad
=
logsumexp_ref_grad
(
self
.
inputs
[
'X'
])
np
.
testing
.
assert_allclose
(
x_grad
,
ref_x_grad
,
rtol
=
1e-08
,
atol
=
1e-08
)
@
unittest
.
skipIf
(
not
core
.
is_compiled_with_cuda
(),
"core is not compiled with CUDA"
)
class
TestLogsumexp_FP16
(
TestLogsumexp
):
def
set_attrs
(
self
):
self
.
dtype
=
'float16'
def
test_check_output
(
self
):
ref_x
=
self
.
inputs
[
'X'
].
astype
(
np
.
float32
)
out_ref
=
ref_logsumexp
(
ref_x
)
paddle
.
disable_static
()
x
=
self
.
inputs
[
'X'
].
astype
(
np
.
float16
)
tensor_x
=
paddle
.
to_tensor
(
x
)
out_pad
=
logsumexp_wrapper
(
tensor_x
)
paddle
.
enable_static
()
np
.
testing
.
assert_allclose
(
out_pad
.
numpy
(),
out_ref
,
rtol
=
1e-03
,
atol
=
1e-08
)
def
test_check_grad
(
self
):
self
.
__class__
.
dtype
=
self
.
dtype
ref_x
=
self
.
inputs
[
'X'
].
astype
(
np
.
float32
)
ref_x_grad
=
logsumexp_ref_grad
(
ref_x
)
x
=
self
.
inputs
[
'X'
].
astype
(
np
.
float16
)
x_grad
=
logsumexp_op_grad
(
x
)
np
.
testing
.
assert_allclose
(
x_grad
,
ref_x_grad
,
rtol
=
1e-03
,
atol
=
1e-05
)
class
TestLogsumexpError
(
unittest
.
TestCase
):
def
test_errors
(
self
):
...
...
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