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
OPTHREE
Paddle
提交
ea47685f
P
Paddle
项目概览
OPTHREE
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
列表
看板
标记
里程碑
合并请求
0
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
Issue
1
列表
看板
标记
里程碑
合并请求
0
合并请求
0
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
ea47685f
编写于
11月 29, 2018
作者:
T
Tao Luo
提交者:
GitHub
11月 29, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #14646 from jczaja/prv-softmax-mkl-sasum
Softmax for inference MKL further changes
上级
2238b965
48e1b97e
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
39 addition
and
8 deletion
+39
-8
paddle/fluid/operators/math/blas.h
paddle/fluid/operators/math/blas.h
+8
-0
paddle/fluid/operators/math/blas_impl.h
paddle/fluid/operators/math/blas_impl.h
+26
-0
paddle/fluid/operators/math/softmax_impl.h
paddle/fluid/operators/math/softmax_impl.h
+2
-5
paddle/fluid/operators/softmax_op.h
paddle/fluid/operators/softmax_op.h
+1
-3
paddle/fluid/platform/dynload/mklml.h
paddle/fluid/platform/dynload/mklml.h
+2
-0
未找到文件。
paddle/fluid/operators/math/blas.h
浏览文件 @
ea47685f
...
...
@@ -168,6 +168,9 @@ class Blas {
template
<
typename
T
>
void
SCAL
(
int
n
,
const
T
a
,
T
*
x
)
const
;
template
<
typename
T
>
T
ASUM
(
int
n
,
T
*
x
,
int
inc
)
const
;
template
<
typename
T
>
void
BatchedGEMM
(
CBLAS_TRANSPOSE
transA
,
CBLAS_TRANSPOSE
transB
,
int
M
,
int
N
,
int
K
,
T
alpha
,
const
T
*
A
,
const
T
*
B
,
T
beta
,
T
*
C
,
...
...
@@ -269,6 +272,11 @@ class BlasT : private Blas<DeviceContext> {
Base
()
->
template
SCAL
<
T
>(
args
...);
}
template
<
typename
...
ARGS
>
T
ASUM
(
ARGS
...
args
)
const
{
return
Base
()
->
template
ASUM
<
T
>(
args
...);
}
template
<
typename
...
ARGS
>
void
BatchedGEMM
(
ARGS
...
args
)
const
{
Base
()
->
template
BatchedGEMM
<
T
>(
args
...);
...
...
paddle/fluid/operators/math/blas_impl.h
浏览文件 @
ea47685f
...
...
@@ -84,6 +84,11 @@ struct CBlas<float> {
platform
::
dynload
::
cblas_sscal
(
args
...);
}
template
<
typename
...
ARGS
>
static
float
ASUM
(
ARGS
...
args
)
{
return
platform
::
dynload
::
cblas_sasum
(
args
...);
}
template
<
typename
...
ARGS
>
static
void
GEMM_BATCH
(
ARGS
...
args
)
{
platform
::
dynload
::
cblas_sgemm_batch
(
args
...);
...
...
@@ -174,6 +179,11 @@ struct CBlas<double> {
platform
::
dynload
::
cblas_dscal
(
args
...);
}
template
<
typename
...
ARGS
>
static
double
ASUM
(
ARGS
...
args
)
{
return
platform
::
dynload
::
cblas_dasum
(
args
...);
}
template
<
typename
...
ARGS
>
static
void
GEMM_BATCH
(
ARGS
...
args
)
{
platform
::
dynload
::
cblas_dgemm_batch
(
args
...);
...
...
@@ -268,6 +278,7 @@ struct CBlas<platform::float16> {
static
void
VPOW
(...)
{
PADDLE_THROW
(
"float16 VPOW not supported on CPU"
);
}
static
void
DOT
(...)
{
PADDLE_THROW
(
"float16 DOT not supported on CPU"
);
};
static
void
SCAL
(...)
{
PADDLE_THROW
(
"float16 SCAL not supported on CPU"
);
};
static
void
ASUM
(...)
{
PADDLE_THROW
(
"float16 ASUM not supported on CPU"
);
};
#ifdef PADDLE_WITH_MKLML
static
void
GEMM_BATCH
(...)
{
PADDLE_THROW
(
"float16 GEMM_BATCH not supported on CPU"
);
...
...
@@ -476,6 +487,21 @@ void Blas<platform::CPUDeviceContext>::SCAL(int n, const T a, T *x) const {
#endif
}
template
<
>
template
<
typename
T
>
T
Blas
<
platform
::
CPUDeviceContext
>::
ASUM
(
int
n
,
T
*
x
,
int
inc
)
const
{
auto
sum
=
static_cast
<
T
>
(
0.0
);
#ifdef PADDLE_WITH_MKLML
sum
=
CBlas
<
T
>::
ASUM
(
n
,
x
,
inc
);
#else
// TODO(jczaja): check if openblas does provide cblas_sasum/cblas_dasum
for
(
int
c
=
0
;
c
<
n
;
++
c
)
{
sum
+=
x
[
c
];
}
#endif
return
sum
;
}
template
<
>
template
<
typename
T
>
void
Blas
<
platform
::
CPUDeviceContext
>::
GEMV
(
bool
trans_a
,
int
M
,
int
N
,
T
alpha
,
...
...
paddle/fluid/operators/math/softmax_impl.h
浏览文件 @
ea47685f
...
...
@@ -100,11 +100,8 @@ class SoftmaxFunctor<DeviceContext, float, true, enable_if_CPU<DeviceContext>> {
blas
.
VEXP
(
num_classes
*
batch_size
,
out_data
,
out_data
);
for
(
int
n
=
0
;
n
<
batch_size
;
++
n
)
{
entities
[
n
]
=
out_data
[
n
*
num_classes
];
for
(
int
c
=
1
;
c
<
num_classes
;
++
c
)
{
entities
[
n
]
+=
out_data
[
n
*
num_classes
+
c
];
}
blas
.
SCAL
(
num_classes
,
1.0
f
/
entities
[
n
],
&
out_data
[
n
*
num_classes
]);
auto
sum
=
blas
.
ASUM
(
num_classes
,
&
out_data
[
n
*
num_classes
],
1
);
blas
.
SCAL
(
num_classes
,
1.0
f
/
sum
,
&
out_data
[
n
*
num_classes
]);
}
}
};
...
...
paddle/fluid/operators/softmax_op.h
浏览文件 @
ea47685f
...
...
@@ -36,9 +36,7 @@ class SoftmaxKernel : public framework::OpKernel<T> {
Tensor
Out_2d
=
framework
::
ReshapeToMatrix
(
*
Out
,
rank
-
1
);
#ifdef PADDLE_ON_INFERENCE
math
::
SoftmaxFunctor
<
DeviceContext
,
T
,
std
::
is_same
<
DeviceContext
,
platform
::
CPUDeviceContext
>::
value
>
()(
math
::
SoftmaxFunctor
<
DeviceContext
,
T
,
true
>
()(
context
.
template
device_context
<
DeviceContext
>(),
&
X_2d
,
&
Out_2d
);
#else
math
::
SoftmaxFunctor
<
DeviceContext
,
T
,
false
>
()(
...
...
paddle/fluid/platform/dynload/mklml.h
浏览文件 @
ea47685f
...
...
@@ -68,6 +68,8 @@ extern void* mklml_dso_handle;
__macro(cblas_dgemm_batch); \
__macro(cblas_sdot); \
__macro(cblas_ddot); \
__macro(cblas_sasum); \
__macro(cblas_dasum); \
__macro(cblas_sscal); \
__macro(cblas_dscal); \
__macro(vsAdd); \
...
...
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