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
csdn_franckjun
Paddle
提交
f45aced5
P
Paddle
项目概览
csdn_franckjun
/
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 >>
提交
f45aced5
编写于
3月 24, 2019
作者:
D
dengkaipeng
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add jit test. develop=test
上级
51536f7f
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
112 addition
and
26 deletion
+112
-26
paddle/fluid/operators/jit/more/mix/mix.cc
paddle/fluid/operators/jit/more/mix/mix.cc
+5
-5
paddle/fluid/operators/jit/more/mix/mix.h
paddle/fluid/operators/jit/more/mix/mix.h
+1
-1
paddle/fluid/operators/jit/more/mkl/mkl.cc
paddle/fluid/operators/jit/more/mkl/mkl.cc
+4
-4
paddle/fluid/operators/jit/more/mkl/mkl.h
paddle/fluid/operators/jit/more/mkl/mkl.h
+5
-5
paddle/fluid/operators/jit/refer/refer.h
paddle/fluid/operators/jit/refer/refer.h
+11
-7
paddle/fluid/operators/jit/test.cc
paddle/fluid/operators/jit/test.cc
+86
-4
未找到文件。
paddle/fluid/operators/jit/more/mix/mix.cc
浏览文件 @
f45aced5
...
...
@@ -50,7 +50,7 @@ void VTanh(const T* x, T* y, int n) {
compute_addbias
(
&
b
,
y
,
y
,
n
);
}
void
Softmax
(
const
T
*
x
,
T
*
y
,
int
n
,
int
bs
,
int
m
)
{
void
Softmax
(
const
T
*
x
,
T
*
y
,
int
n
,
int
bs
,
int
remain
)
{
auto
compute_hmax
=
KernelFuncs
<
HMaxTuple
<
T
>
,
CPUPlace
>::
Cache
().
At
(
n
);
auto
compute_hsum
=
KernelFuncs
<
HSumTuple
<
T
>
,
CPUPlace
>::
Cache
().
At
(
n
);
auto
compute_vscal
=
KernelFuncs
<
VScalTuple
<
T
>
,
CPUPlace
>::
Cache
().
At
(
n
);
...
...
@@ -66,15 +66,15 @@ void Softmax(const T* x, T* y, int n, int bs, int m) {
scalar
=
static_cast
<
T
>
(
0
)
-
scalar
;
compute_vaddbias
(
&
scalar
,
x
,
y
,
n
);
// x - max
compute_vexp
(
y
,
y
,
n
);
if
(
m
==
1
)
{
if
(
remain
==
1
)
{
compute_hsum
(
y
,
&
scalar
,
n
);
scalar
=
static_cast
<
T
>
(
1
)
/
scalar
;
compute_vscal
(
&
scalar
,
y
,
y
,
n
);
}
else
{
for
(
int
j
=
0
;
j
<
m
;
++
j
)
{
compute_stridesum
(
&
y
[
j
],
&
scalar
,
n
,
m
);
for
(
int
j
=
0
;
j
<
remain
;
++
j
)
{
compute_stridesum
(
&
y
[
j
],
&
scalar
,
n
,
remain
);
scalar
=
static_cast
<
T
>
(
1
)
/
scalar
;
compute_stridescal
(
&
scalar
,
&
y
[
j
],
&
y
[
j
],
n
,
m
);
compute_stridescal
(
&
scalar
,
&
y
[
j
],
&
y
[
j
],
n
,
remain
);
}
}
x
+=
n
;
...
...
paddle/fluid/operators/jit/more/mix/mix.h
浏览文件 @
f45aced5
...
...
@@ -26,7 +26,7 @@ using T = float;
void
VSigmoid
(
const
T
*
x
,
T
*
y
,
int
n
);
void
VTanh
(
const
T
*
x
,
T
*
y
,
int
n
);
void
Softmax
(
const
T
*
x
,
T
*
y
,
int
n
,
int
bs
,
int
m
);
void
Softmax
(
const
T
*
x
,
T
*
y
,
int
n
,
int
bs
,
int
remain
);
void
LSTMCtHt
(
lstm_t
*
step
,
const
lstm_attr_t
*
attr
);
void
LSTMC1H1
(
lstm_t
*
step
,
const
lstm_attr_t
*
attr
);
...
...
paddle/fluid/operators/jit/more/mkl/mkl.cc
浏览文件 @
f45aced5
...
...
@@ -81,7 +81,7 @@ void VScal<double>(const double* a, const double* x, double* y, int n) {
template
<
>
void
StrideScal
<
float
>
(
const
float
*
a
,
const
float
*
x
,
float
*
y
,
int
n
,
int
stride
)
{
if
(
x
==
y
)
{
platform
::
dynload
::
cblas_sscal
(
n
,
*
a
,
y
,
stride
);
platform
::
dynload
::
cblas_sscal
(
n
/
stride
,
*
a
,
y
,
stride
);
}
else
{
refer
::
StrideScal
<
float
>
(
a
,
x
,
y
,
n
,
stride
);
}
...
...
@@ -90,7 +90,7 @@ void StrideScal<float>(const float* a, const float* x, float* y, int n, int stri
template
<
>
void
StrideScal
<
double
>
(
const
double
*
a
,
const
double
*
x
,
double
*
y
,
int
n
,
int
stride
)
{
if
(
x
==
y
)
{
platform
::
dynload
::
cblas_dscal
(
n
,
*
a
,
y
,
stride
);
platform
::
dynload
::
cblas_dscal
(
n
/
stride
,
*
a
,
y
,
stride
);
}
else
{
refer
::
StrideScal
<
double
>
(
a
,
x
,
y
,
n
,
stride
);
}
...
...
@@ -148,12 +148,12 @@ void ASum<double>(const double* x, double* res, int n) {
template
<
>
void
StrideASum
<
float
>
(
const
float
*
x
,
float
*
res
,
int
n
,
int
stride
)
{
res
[
0
]
=
platform
::
dynload
::
cblas_sasum
(
n
,
x
,
stride
);
res
[
0
]
=
platform
::
dynload
::
cblas_sasum
(
n
/
stride
,
x
,
stride
);
}
template
<
>
void
StrideASum
<
double
>
(
const
double
*
x
,
double
*
res
,
int
n
,
int
stride
)
{
res
[
0
]
=
platform
::
dynload
::
cblas_dasum
(
n
,
x
,
stride
);
res
[
0
]
=
platform
::
dynload
::
cblas_dasum
(
n
/
stride
,
x
,
stride
);
}
// TODO(TJ): tuning me carefully on AVX, AVX2 and AVX512
...
...
paddle/fluid/operators/jit/more/mkl/mkl.h
浏览文件 @
f45aced5
...
...
@@ -135,7 +135,7 @@ template <typename T>
void
StrideScal
(
const
T
*
a
,
const
T
*
x
,
T
*
y
,
int
n
,
int
stride
);
template
<
typename
T
>
void
Softmax
(
const
T
*
x
,
T
*
y
,
int
n
,
int
bs
,
int
m
=
1
)
{
void
Softmax
(
const
T
*
x
,
T
*
y
,
int
n
,
int
bs
,
int
remain
=
1
)
{
std
::
vector
<
T
>
entities
(
bs
);
for
(
int
i
=
0
;
i
<
bs
;
++
i
)
{
entities
[
i
]
=
x
[
i
*
n
];
...
...
@@ -149,15 +149,15 @@ void Softmax(const T* x, T* y, int n, int bs, int m=1) {
VExp
(
y
,
y
,
n
*
bs
);
for
(
int
i
=
0
;
i
<
bs
;
++
i
)
{
T
sum
;
if
(
m
==
1
)
{
if
(
remain
==
1
)
{
ASum
(
&
y
[
i
*
n
],
&
sum
,
n
);
sum
=
static_cast
<
T
>
(
1
)
/
sum
;
VScal
(
&
sum
,
&
y
[
i
*
n
],
&
y
[
i
*
n
],
n
);
}
else
{
for
(
int
j
=
0
;
j
<
m
;
++
j
)
{
StrideASum
(
&
y
[
i
*
n
+
j
],
&
sum
,
n
/
m
,
m
);
for
(
int
j
=
0
;
j
<
remain
;
++
j
)
{
StrideASum
(
&
y
[
i
*
n
+
j
],
&
sum
,
n
,
remain
);
sum
=
static_cast
<
T
>
(
1
)
/
sum
;
StrideScal
(
&
sum
,
&
y
[
i
*
n
+
j
],
&
y
[
i
*
n
+
j
],
n
/
m
,
m
);
StrideScal
(
&
sum
,
&
y
[
i
*
n
+
j
],
&
y
[
i
*
n
+
j
],
n
,
remain
);
}
}
}
...
...
paddle/fluid/operators/jit/refer/refer.h
浏览文件 @
f45aced5
...
...
@@ -421,30 +421,34 @@ void StrideASum(const T* x, T* res, int n, int stride) {
template
<
typename
T
>
void
StrideScal
(
const
T
*
a
,
const
T
*
x
,
T
*
y
,
int
n
,
int
stride
)
{
for
(
int
i
=
0
;
i
<
n
;
i
+=
stride
)
{
y
[
i
]
=
x
[
i
]
*
a
[
0
];
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
if
(
i
%
stride
==
0
)
{
y
[
i
]
=
x
[
i
]
*
a
[
0
];
}
else
{
y
[
i
]
=
x
[
i
];
}
}
}
// y = e^(x - max(x))
// y = y / sum(y)
template
<
typename
T
>
void
Softmax
(
const
T
*
x
,
T
*
y
,
int
n
,
int
bs
=
1
,
int
m
=
1
)
{
void
Softmax
(
const
T
*
x
,
T
*
y
,
int
n
,
int
bs
=
1
,
int
remain
=
1
)
{
for
(
int
i
=
0
;
i
<
bs
;
++
i
)
{
T
scalar
;
HMax
(
x
,
&
scalar
,
n
);
scalar
=
static_cast
<
T
>
(
0
)
-
scalar
;
VAddBias
(
&
scalar
,
x
,
y
,
n
);
// x - max
VExp
(
y
,
y
,
n
);
if
(
m
==
1
)
{
if
(
remain
==
1
)
{
HSum
(
y
,
&
scalar
,
n
);
scalar
=
static_cast
<
T
>
(
1
)
/
scalar
;
VScal
(
&
scalar
,
y
,
y
,
n
);
}
else
{
for
(
int
j
=
0
;
j
<
m
;
j
++
)
{
StrideASum
(
&
y
[
j
],
&
scalar
,
n
,
m
);
for
(
int
j
=
0
;
j
<
remain
;
j
++
)
{
StrideASum
(
&
y
[
j
],
&
scalar
,
n
,
remain
);
scalar
=
static_cast
<
T
>
(
1
)
/
scalar
;
StrideScal
(
&
scalar
,
&
y
[
j
],
&
y
[
j
],
n
,
m
);
StrideScal
(
&
scalar
,
&
y
[
j
],
&
y
[
j
],
n
,
remain
);
}
}
x
+=
n
;
...
...
paddle/fluid/operators/jit/test.cc
浏览文件 @
f45aced5
...
...
@@ -723,11 +723,10 @@ void TestKernelSoftmax() {
VLOG
(
10
)
<<
"Test JITKernel: "
<<
jit
::
to_string
(
KernelTuple
::
kernel_type
);
for
(
int
bs
:
{
1
,
2
,
10
})
{
for
(
int
n
:
TestSizes
())
{
for
(
int
m
:
{
1
,
2
})
{
for
(
int
m
:
{
1
,
2
,
3
})
{
// remain
if
(
m
>
n
||
n
%
m
!=
0
)
{
continue
;
}
VLOG
(
10
)
<<
"Softmax: "
<<
bs
<<
", "
<<
n
<<
", "
<<
m
;
auto
ref
=
jit
::
GetReferFunc
<
KernelTuple
>
();
EXPECT_TRUE
(
ref
!=
nullptr
);
std
::
vector
<
T
>
x
(
bs
*
n
),
y
(
bs
*
n
);
...
...
@@ -766,6 +765,86 @@ void TestKernelSoftmax() {
}
}
template
<
typename
KernelTuple
,
typename
PlaceType
>
void
TestKernelStrideASum
()
{
using
T
=
typename
KernelTuple
::
data_type
;
VLOG
(
10
)
<<
"Test JITKernel: "
<<
jit
::
to_string
(
KernelTuple
::
kernel_type
);
for
(
int
d
:
TestSizes
())
{
for
(
int
m
:
{
1
,
2
,
3
})
{
// stride
if
(
m
>
d
||
d
%
m
!=
0
)
{
continue
;
}
auto
ref
=
jit
::
GetReferFunc
<
KernelTuple
>
();
EXPECT_TRUE
(
ref
!=
nullptr
);
std
::
vector
<
T
>
x
(
d
);
RandomVec
<
T
>
(
d
,
x
.
data
());
T
ref_res
;
ref
(
x
.
data
(),
&
ref_res
,
d
,
m
);
auto
verifier
=
[](
const
typename
KernelTuple
::
func_type
tgt
,
const
std
::
vector
<
T
>&
x
,
const
T
ref_res
,
const
int
m
)
{
EXPECT_TRUE
(
tgt
!=
nullptr
);
T
tgt_res
;
tgt
(
x
.
data
(),
&
tgt_res
,
x
.
size
(),
m
);
ExpectEQ
<
T
>
(
&
tgt_res
,
&
ref_res
,
1
);
};
TestAllImpls
<
KernelTuple
,
PlaceType
>
(
d
,
verifier
,
x
,
ref_res
,
m
);
}
}
}
template
<
typename
KernelTuple
,
typename
PlaceType
>
void
TestKernelStrideScal
()
{
using
T
=
typename
KernelTuple
::
data_type
;
VLOG
(
10
)
<<
"Test JITKernel: "
<<
jit
::
to_string
(
KernelTuple
::
kernel_type
);
// for (int d : TestSizes()) {
// for (int m : {1, 2, 3}) { // stride
for
(
int
d
:
{
4
})
{
for
(
int
m
:
{
2
})
{
// stride
if
(
m
>
d
||
d
%
m
!=
0
)
{
continue
;
}
auto
ref
=
jit
::
GetReferFunc
<
KernelTuple
>
();
EXPECT_TRUE
(
ref
!=
nullptr
);
const
T
a
=
static_cast
<
T
>
(
3
);
std
::
vector
<
T
>
x
(
d
),
yref
(
d
);
std
::
vector
<
T
>
xinp
(
d
);
// inplace test
RandomVec
<
T
>
(
d
,
x
.
data
());
std
::
copy
(
x
.
begin
(),
x
.
end
(),
xinp
.
begin
());
const
T
*
x_data
=
x
.
data
();
T
*
yref_data
=
yref
.
data
();
T
*
xinp_data
=
xinp
.
data
();
// test refer code inplace
ref
(
&
a
,
x_data
,
yref_data
,
d
,
m
);
ref
(
&
a
,
xinp_data
,
xinp_data
,
d
,
m
);
ExpectEQ
<
T
>
(
xinp_data
,
yref_data
,
d
);
auto
verifier
=
[](
const
typename
KernelTuple
::
func_type
tgt
,
const
T
a
,
const
std
::
vector
<
T
>&
x
,
const
std
::
vector
<
T
>&
yref
,
const
int
m
)
{
EXPECT_TRUE
(
tgt
!=
nullptr
);
EXPECT_EQ
(
yref
.
size
(),
x
.
size
());
const
T
*
x_data
=
x
.
data
();
const
T
*
yref_data
=
yref
.
data
();
const
int
d
=
yref
.
size
();
std
::
vector
<
T
>
ytgt
(
d
);
T
*
ytgt_data
=
ytgt
.
data
();
// test normal
tgt
(
&
a
,
x_data
,
ytgt_data
,
d
,
m
);
ExpectEQ
<
T
>
(
ytgt_data
,
yref_data
,
d
);
// test inplace x
std
::
copy
(
x
.
begin
(),
x
.
end
(),
ytgt
.
begin
());
tgt
(
&
a
,
ytgt_data
,
ytgt_data
,
d
,
m
);
ExpectEQ
<
T
>
(
ytgt_data
,
yref_data
,
d
);
};
TestAllImpls
<
KernelTuple
,
PlaceType
>
(
d
,
verifier
,
a
,
x
,
yref
,
m
);
}
}
}
template
<
typename
KernelTuple
,
typename
PlaceType
>
void
TestKernelSgd
()
{
using
T
=
typename
KernelTuple
::
data_type
;
...
...
@@ -918,7 +997,7 @@ TEST(JITKernel_pool, more) {
EXPECT_EQ
(
kers
.
size
(),
10UL
);
#else
#ifdef PADDLE_WITH_MKLML
EXPECT_EQ
(
kers
.
size
(),
2
1
UL
);
EXPECT_EQ
(
kers
.
size
(),
2
2
UL
);
#else
EXPECT_EQ
(
kers
.
size
(),
8UL
);
#endif
...
...
@@ -927,7 +1006,7 @@ TEST(JITKernel_pool, more) {
TEST
(
JITKernel_pool
,
refer
)
{
const
auto
&
kers
=
jit
::
ReferKernelPool
::
Instance
().
AllKernels
();
EXPECT_EQ
(
kers
.
size
(),
29
UL
);
EXPECT_EQ
(
kers
.
size
(),
31
UL
);
}
// test helper
...
...
@@ -1298,3 +1377,6 @@ TEST_CPU_KERNEL(MatMul);
TEST_CPU_KERNEL
(
Softmax
);
TEST_CPU_KERNEL
(
Sgd
);
TEST_CPU_KERNEL
(
VBroadcast
);
TEST_CPU_KERNEL
(
StrideASum
);
TEST_CPU_KERNEL
(
StrideScal
);
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