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
s920243400
PaddleDetection
提交
1f0291a5
P
PaddleDetection
项目概览
s920243400
/
PaddleDetection
与 Fork 源项目一致
Fork自
PaddlePaddle / PaddleDetection
通知
2
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
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
PaddleDetection
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
0
Issue
0
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
提交
1f0291a5
编写于
11月 26, 2018
作者:
T
tensor-tang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add comments and follow comments
test=develop
上级
557229bd
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
17 addition
and
12 deletion
+17
-12
paddle/fluid/operators/math/jit_kernel_refer.h
paddle/fluid/operators/math/jit_kernel_refer.h
+4
-0
paddle/fluid/operators/math/jit_kernel_test.cc
paddle/fluid/operators/math/jit_kernel_test.cc
+13
-12
未找到文件。
paddle/fluid/operators/math/jit_kernel_refer.h
浏览文件 @
1f0291a5
...
...
@@ -116,6 +116,7 @@ void (*getActFunc(const std::string& type))(const T*, T*, int) { // NOLINT
return
nullptr
;
}
// compute ct and ht
template
<
typename
T
>
void
LSTMCtHt
(
lstm_t
*
step
,
const
lstm_attr_t
*
attr
)
{
T
*
gates
=
reinterpret_cast
<
T
*>
(
step
->
gates
);
...
...
@@ -199,6 +200,7 @@ void GRUH1(gru_t* step, const gru_attr_t* attr) {
VMul
(
gates
,
gates
+
d2
,
ht
,
d
);
}
// compute the first part of GRU: ht = act_gate(r) * ht_1
template
<
typename
T
>
void
GRUHtPart1
(
gru_t
*
step
,
const
gru_attr_t
*
attr
)
{
// W: {W_update, W_reset; W_state}
...
...
@@ -210,6 +212,8 @@ void GRUHtPart1(gru_t* step, const gru_attr_t* attr) {
VMul
(
ht_1
,
gates
+
attr
->
d
,
ht
,
attr
->
d
);
}
// compute the second part of GRU:
// ht = act_gate(u) * act_cand(s) + (1-act_gate(u)) * ht_1
template
<
typename
T
>
void
GRUHtPart2
(
gru_t
*
step
,
const
gru_attr_t
*
attr
)
{
T
*
gates
=
reinterpret_cast
<
T
*>
(
step
->
gates
);
...
...
paddle/fluid/operators/math/jit_kernel_test.cc
浏览文件 @
1f0291a5
...
...
@@ -86,7 +86,7 @@ TEST(JitKernel, vrelu) {
vrelu_intri8
(
d
,
x_data
,
zref_data
);
}
auto
si1
=
GetCurrentUS
();
VLOG
(
30
)
<<
"Vec size 8 intr takes: "
<<
(
si1
-
si0
)
/
repeat
;
VLOG
(
30
)
<<
"Vec size 8 intr takes: "
<<
(
si1
-
si0
)
/
repeat
<<
" us"
;
}
#endif
auto
ttgts
=
GetCurrentUS
();
...
...
@@ -96,7 +96,7 @@ TEST(JitKernel, vrelu) {
auto
ttgte
=
GetCurrentUS
();
VLOG
(
30
)
<<
"Vec size "
<<
d
<<
": refer takes: "
<<
(
trefe
-
trefs
)
/
repeat
<<
" us, tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
;
<<
" us, tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
<<
" us"
;
for
(
int
i
=
0
;
i
<
d
;
++
i
)
{
EXPECT_NEAR
(
ztgt_data
[
i
],
zref_data
[
i
],
1e-3
);
}
...
...
@@ -129,7 +129,7 @@ TEST(JitKernel, vaddbias) {
VLOG
(
30
)
<<
"Vec size "
<<
d
<<
": refer takes: "
<<
(
trefe
-
trefs
)
/
repeat
<<
" us, tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
;
<<
" us, tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
<<
" us"
;
for
(
int
i
=
0
;
i
<
d
;
++
i
)
{
EXPECT_NEAR
(
ztgt_data
[
i
],
zref_data
[
i
],
1e-3
);
}
...
...
@@ -182,7 +182,7 @@ TEST(JitKernel, vexp) {
#else
<<
" us, "
#endif
<<
"tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
;
<<
"tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
<<
" us"
;
for
(
int
i
=
0
;
i
<
d
;
++
i
)
{
EXPECT_NEAR
(
ztgt_data
[
i
],
zref_data
[
i
],
1e-3
);
}
...
...
@@ -238,7 +238,7 @@ TEST(JitKernel, vsigmoid) {
VLOG
(
30
)
<<
"Vec size "
<<
d
<<
": refer takes: "
<<
(
trefe
-
trefs
)
/
repeat
<<
" us, better(jit exp) takes: "
<<
(
tmkle
-
tmkls
)
/
repeat
<<
" us, tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
;
<<
" us, tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
<<
" us"
;
for
(
int
i
=
0
;
i
<
d
;
++
i
)
{
EXPECT_NEAR
(
ztgt_data
[
i
],
zref_data
[
i
],
1e-3
);
}
...
...
@@ -299,7 +299,7 @@ TEST(JitKernel, vtanh) {
VLOG
(
30
)
<<
"Vec size "
<<
d
<<
": refer takes: "
<<
(
trefe
-
trefs
)
/
repeat
<<
" us, better(jit exp) takes: "
<<
(
tmkle
-
tmkls
)
/
repeat
<<
" us, tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
;
<<
" us, tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
<<
" us"
;
for
(
int
i
=
0
;
i
<
d
;
++
i
)
{
EXPECT_NEAR
(
ztgt_data
[
i
],
zref_data
[
i
],
1e-3
);
}
...
...
@@ -400,7 +400,7 @@ TEST(JitKernel, lstm) {
VLOG
(
30
)
<<
"Vec size "
<<
d
<<
": refer takes: "
<<
(
trefe
-
trefs
)
/
repeat
<<
" us, better(jit) takes: "
<<
(
tmkle
-
tmkls
)
/
repeat
<<
" us, tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
;
<<
" us, tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
<<
" us"
;
}
}
...
...
@@ -474,7 +474,7 @@ TEST(JitKernel, vscal) {
}
auto
si3
=
GetCurrentUS
();
VLOG
(
30
)
<<
"Vec size 8 intr takes: "
<<
(
si1
-
si0
)
/
repeat
<<
" us, inplace: "
<<
(
si3
-
si2
)
/
repeat
;
<<
" us, inplace: "
<<
(
si3
-
si2
)
/
repeat
<<
" us"
;
}
#endif
...
...
@@ -498,7 +498,8 @@ TEST(JitKernel, vscal) {
<<
" us, "
#endif
<<
"tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
<<
"us, tgt inplace takes: "
<<
(
ttgte1
-
ttgts1
)
/
repeat
;
<<
"us, tgt inplace takes: "
<<
(
ttgte1
-
ttgts1
)
/
repeat
<<
" us"
;
for
(
int
i
=
0
;
i
<
d
;
++
i
)
{
EXPECT_NEAR
(
ztgt_data
[
i
],
zref_data
[
i
],
1e-3
);
}
...
...
@@ -573,7 +574,7 @@ TEST(JitKernel, vmul) {
#else
<<
" us, "
#endif
<<
"tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
;
<<
"tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
<<
" us"
;
for
(
int
i
=
0
;
i
<
d
;
++
i
)
{
EXPECT_NEAR
(
ztgt_data
[
i
],
zref_data
[
i
],
1e-3
);
}
...
...
@@ -648,7 +649,7 @@ TEST(JitKernel, vadd) {
#else
<<
" us, "
#endif
<<
"tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
;
<<
"tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
<<
" us"
;
for
(
int
i
=
0
;
i
<
d
;
++
i
)
{
EXPECT_NEAR
(
ztgt_data
[
i
],
zref_data
[
i
],
1e-3
);
}
...
...
@@ -701,7 +702,7 @@ TEST(JitKernel, vaddrelu) {
VLOG
(
30
)
<<
"Vec size "
<<
d
<<
": refer takes: "
<<
(
trefe
-
trefs
)
/
repeat
<<
" us, better takes: "
<<
(
tmkle
-
tmkls
)
/
repeat
<<
" us, "
<<
"tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
;
<<
"tgt takes: "
<<
(
ttgte
-
ttgts
)
/
repeat
<<
" us"
;
for
(
int
i
=
0
;
i
<
d
;
++
i
)
{
EXPECT_NEAR
(
ztgt_data
[
i
],
zref_data
[
i
],
1e-3
);
}
...
...
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