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
e2d6eddd
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 >>
提交
e2d6eddd
编写于
11月 16, 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
remove ComputeDeprecated
test=develop
上级
f65ddff8
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
53 addition
and
78 deletion
+53
-78
paddle/fluid/operators/math/jit_code.cc
paddle/fluid/operators/math/jit_code.cc
+1
-0
paddle/fluid/operators/math/jit_kernel.h
paddle/fluid/operators/math/jit_kernel.h
+6
-25
paddle/fluid/operators/math/jit_kernel_blas.cc
paddle/fluid/operators/math/jit_kernel_blas.cc
+14
-14
paddle/fluid/operators/math/jit_kernel_exp.cc
paddle/fluid/operators/math/jit_kernel_exp.cc
+5
-12
paddle/fluid/operators/math/jit_kernel_rnn.cc
paddle/fluid/operators/math/jit_kernel_rnn.cc
+19
-19
paddle/fluid/operators/math/jit_kernel_test.cc
paddle/fluid/operators/math/jit_kernel_test.cc
+8
-8
未找到文件。
paddle/fluid/operators/math/jit_code.cc
浏览文件 @
e2d6eddd
...
...
@@ -178,6 +178,7 @@ bool VActJitCode::init(int d, operand_type type) {
if
(
type
==
operand_type
::
relu
)
{
return
ok
;
}
else
{
// TODO(TJ): support more
return
ok
&&
d
==
8
;
// only 8 yet
}
}
...
...
paddle/fluid/operators/math/jit_kernel.h
浏览文件 @
e2d6eddd
...
...
@@ -98,42 +98,23 @@ class VAddBiasKernel : public Kernel {
template
<
typename
T
>
class
VActKernel
:
public
Kernel
{
public:
v
irtual
void
ComputeDeprecated
(
const
T
*
x
,
T
*
y
)
const
=
0
;
v
oid
(
*
Compute
)(
const
T
*
,
T
*
,
int
)
;
};
template
<
typename
T
>
class
VReluKernel
:
public
VActKernel
<
T
>
{
public:
virtual
void
ComputeDeprecated
(
const
T
*
x
,
T
*
y
)
const
=
0
;
void
(
*
Compute
)(
const
T
*
,
T
*
,
int
);
};
class
VReluKernel
:
public
VActKernel
<
T
>
{};
template
<
typename
T
>
class
VIdentityKernel
:
public
VActKernel
<
T
>
{
public:
virtual
void
ComputeDeprecated
(
const
T
*
x
,
T
*
y
)
const
=
0
;
};
class
VIdentityKernel
:
public
VActKernel
<
T
>
{};
template
<
typename
T
>
class
VExpKernel
:
public
VActKernel
<
T
>
{
public:
virtual
void
ComputeDeprecated
(
const
T
*
x
,
T
*
y
)
const
=
0
;
void
(
*
Compute
)(
const
T
*
,
T
*
,
int
);
};
class
VExpKernel
:
public
VActKernel
<
T
>
{};
template
<
typename
T
>
class
VSigmoidKernel
:
public
VActKernel
<
T
>
{
public:
virtual
void
ComputeDeprecated
(
const
T
*
x
,
T
*
y
)
const
=
0
;
void
(
*
Compute
)(
const
T
*
,
T
*
,
int
);
};
class
VSigmoidKernel
:
public
VActKernel
<
T
>
{};
template
<
typename
T
>
class
VTanhKernel
:
public
VActKernel
<
T
>
{
public:
virtual
void
ComputeDeprecated
(
const
T
*
x
,
T
*
y
)
const
=
0
;
void
(
*
Compute
)(
const
T
*
,
T
*
,
int
);
};
class
VTanhKernel
:
public
VActKernel
<
T
>
{};
template
<
typename
T
>
class
LSTMKernel
:
public
Kernel
{
...
...
paddle/fluid/operators/math/jit_kernel_blas.cc
浏览文件 @
e2d6eddd
...
...
@@ -346,7 +346,6 @@ class VReluKernelImpl : public VReluKernel<T> {
public:
JITKERNEL_DECLARE_STATIC_FUNC
;
explicit
VReluKernelImpl
(
int
d
)
:
VReluKernel
<
T
>
()
{
this
->
num_
=
d
;
// TODO(TJ): remove me when ComputeDeprecated done
#ifdef PADDLE_WITH_XBYAK
if
(
useJIT
(
d
))
{
size_t
sz
=
96
/* init size */
+
...
...
@@ -361,9 +360,6 @@ class VReluKernelImpl : public VReluKernel<T> {
this
->
Compute
=
VReluRefer
<
T
>
;
}
void
ComputeDeprecated
(
const
T
*
x
,
T
*
y
)
const
override
{
VReluRefer
(
x
,
y
,
this
->
num_
);
}
#ifdef PADDLE_WITH_XBYAK
private:
...
...
@@ -378,22 +374,26 @@ bool VReluKernelImpl<float>::useJIT(int d) {
}
#endif
REGISTER_JITKERNEL
(
vmul
,
VMulKernel
);
REGISTER_JITKERNEL
(
vadd
,
VAddKernel
);
REGISTER_JITKERNEL
(
vaddrelu
,
VAddReluKernel
);
REGISTER_JITKERNEL
(
vscal
,
VScalKernel
);
REGISTER_JITKERNEL
(
vaddbias
,
VAddBiasKernel
);
REGISTER_JITKERNEL
(
vrelu
,
VReluKernel
);
template
<
typename
T
>
inline
void
VIdentityRefer
(
const
T
*
x
,
T
*
y
,
int
n
)
{}
/* An empty JitKernel */
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
,
jit_block
>
template
<
typename
T
>
class
VIdentityKernelImpl
:
public
VIdentityKernel
<
T
>
{
public:
explicit
VIdentityKernelImpl
(
int
d
)
:
VIdentityKernel
<
T
>
()
{
this
->
num_
=
d
;
}
void
ComputeDeprecated
(
const
T
*
x
,
T
*
y
)
const
override
{}
JITKERNEL_DECLARE_STATIC_FUNC
;
explicit
VIdentityKernelImpl
(
int
d
)
:
VIdentityKernel
<
T
>
()
{
this
->
Compute
=
VIdentityRefer
<
T
>
;
}
};
REGISTER_JITKERNEL_DEPRECATED
(
videntity
,
VIdentityKernel
);
REGISTER_JITKERNEL
(
vmul
,
VMulKernel
);
REGISTER_JITKERNEL
(
vadd
,
VAddKernel
);
REGISTER_JITKERNEL
(
vaddrelu
,
VAddReluKernel
);
REGISTER_JITKERNEL
(
vscal
,
VScalKernel
);
REGISTER_JITKERNEL
(
vaddbias
,
VAddBiasKernel
);
REGISTER_JITKERNEL
(
vrelu
,
VReluKernel
);
REGISTER_JITKERNEL
(
videntity
,
VIdentityKernel
);
}
// namespace jitkernel
}
// namespace math
...
...
paddle/fluid/operators/math/jit_kernel_exp.cc
浏览文件 @
e2d6eddd
...
...
@@ -36,6 +36,7 @@ namespace jitkernel {
namespace
jit
=
platform
::
jit
;
// TODO(TJ): move refer codes to one file
// Refer code only focus on correctness
template
<
typename
T
>
void
VExpRefer
(
const
T
*
x
,
T
*
y
,
int
n
)
{
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
...
...
@@ -67,6 +68,7 @@ void VTanhRefer(const T* x, T* y, int n) {
}
#ifdef PADDLE_WITH_MKLML
// try to use MKL to speedup
template
<
typename
T
>
void
VExpMKL
(
const
T
*
x
,
T
*
y
,
int
n
);
...
...
@@ -112,7 +114,6 @@ class VExpKernelImpl : public VExpKernel<T> {
public:
JITKERNEL_DECLARE_STATIC_FUNC
;
explicit
VExpKernelImpl
(
int
d
)
:
VExpKernel
<
T
>
()
{
this
->
num_
=
d
;
// TODO(TJ): remove me when ComputeDeprecated done
#ifdef PADDLE_WITH_XBYAK
if
(
useJIT
(
d
))
{
size_t
sz
=
96
+
d
/
AVX_FLOAT_BLOCK
*
4
*
8
;
// should change
...
...
@@ -130,9 +131,7 @@ class VExpKernelImpl : public VExpKernel<T> {
#endif
this
->
Compute
=
VExpRefer
<
T
>
;
}
void
ComputeDeprecated
(
const
T
*
x
,
T
*
y
)
const
override
{
VExpRefer
(
x
,
y
,
this
->
num_
);
}
#ifdef PADDLE_WITH_XBYAK
private:
...
...
@@ -166,7 +165,6 @@ class VSigmoidKernelImpl : public VSigmoidKernel<T> {
public:
JITKERNEL_DECLARE_STATIC_FUNC
;
explicit
VSigmoidKernelImpl
(
int
d
)
:
VSigmoidKernel
<
T
>
()
{
this
->
num_
=
d
;
// TODO(TJ): remove me when ComputeDeprecated done
#ifdef PADDLE_WITH_XBYAK
if
(
useJIT
(
d
))
{
size_t
sz
=
96
+
d
/
AVX_FLOAT_BLOCK
*
4
*
8
;
// should change
...
...
@@ -186,9 +184,7 @@ class VSigmoidKernelImpl : public VSigmoidKernel<T> {
#endif
this
->
Compute
=
VSigmoidRefer
<
T
>
;
}
void
ComputeDeprecated
(
const
T
*
x
,
T
*
y
)
const
override
{
VSigmoidRefer
(
x
,
y
,
this
->
num_
);
}
#ifdef PADDLE_WITH_XBYAK
private:
...
...
@@ -221,7 +217,6 @@ class VTanhKernelImpl : public VTanhKernel<T> {
public:
JITKERNEL_DECLARE_STATIC_FUNC
;
explicit
VTanhKernelImpl
(
int
d
)
:
VTanhKernel
<
T
>
()
{
this
->
num_
=
d
;
// TODO(TJ): remove me when ComputeDeprecated done
#ifdef PADDLE_WITH_XBYAK
if
(
useJIT
(
d
))
{
size_t
sz
=
96
+
d
/
AVX_FLOAT_BLOCK
*
4
*
8
;
// should change
...
...
@@ -241,9 +236,7 @@ class VTanhKernelImpl : public VTanhKernel<T> {
#endif
this
->
Compute
=
VTanhRefer
<
T
>
;
}
void
ComputeDeprecated
(
const
T
*
x
,
T
*
y
)
const
override
{
VTanhRefer
(
x
,
y
,
this
->
num_
);
}
#ifdef PADDLE_WITH_XBYAK
private:
...
...
paddle/fluid/operators/math/jit_kernel_rnn.cc
浏览文件 @
e2d6eddd
...
...
@@ -175,26 +175,26 @@ class LSTMKernelImpl : public LSTMKernel<T> {
void
ComputeCtHt
(
T
*
gates
,
const
T
*
ct_1
,
T
*
ct
,
T
*
ht
,
const
T
*
wp_data
,
T
*
checked
)
const
override
{
// gates: W_ch, W_ih, W_fh, W_oh
act_gate_d3_
->
Compute
Deprecated
(
gates
+
d_
,
gates
+
d
_
);
act_gate_d3_
->
Compute
(
gates
+
d_
,
gates
+
d_
,
d3
_
);
/* C_t = C_t-1 * fgated + cand_gated * igated */
act_cand_d_
->
Compute
Deprecated
(
gates
,
gates
);
act_cand_d_
->
Compute
(
gates
,
gates
,
d_
);
vmul_d_
->
Compute
(
gates
,
gates
+
d_
,
gates
+
d_
,
d_
);
vmul_d_
->
Compute
(
ct_1
,
gates
+
d2_
,
gates
+
d2_
,
d_
);
vadd_d_
->
Compute
(
gates
+
d_
,
gates
+
d2_
,
ct
,
d_
);
/* H_t = act_cell(C_t) * ogated */
act_cell_d_
->
Compute
Deprecated
(
ct
,
gates
+
d2
_
);
act_cell_d_
->
Compute
(
ct
,
gates
+
d2_
,
d
_
);
vmul_d_
->
Compute
(
gates
+
d2_
,
gates
+
d3_
,
ht
,
d_
);
}
void
ComputeC1H1
(
T
*
gates
,
T
*
ct
,
T
*
ht
,
const
T
*
wp_data
)
const
override
{
/* C_t = igated * cgated*/
act_gate_d_
->
Compute
Deprecated
(
gates
+
d_
,
gates
+
d_
);
act_cand_d_
->
Compute
Deprecated
(
gates
,
gates
);
act_gate_d_
->
Compute
(
gates
+
d_
,
gates
+
d_
,
d_
);
act_cand_d_
->
Compute
(
gates
,
gates
,
d_
);
vmul_d_
->
Compute
(
gates
,
gates
+
d_
,
ct
,
d_
);
/* H_t = act_cell(C_t) * ogated */
act_gate_d_
->
Compute
Deprecated
(
gates
+
d3_
,
gates
+
d3
_
);
act_cell_d_
->
Compute
Deprecated
(
ct
,
gates
+
d2
_
);
act_gate_d_
->
Compute
(
gates
+
d3_
,
gates
+
d3_
,
d
_
);
act_cell_d_
->
Compute
(
ct
,
gates
+
d2_
,
d
_
);
vmul_d_
->
Compute
(
gates
+
d2_
,
gates
+
d3_
,
ht
,
d_
);
}
...
...
@@ -292,32 +292,32 @@ class PeepholeKernelImpl : public LSTMKernel<T> {
vmul_d_
->
Compute
(
wp_data
,
ct_1
,
checked
,
d_
);
vmul_d_
->
Compute
(
wp_data
+
d_
,
ct_1
,
checked
+
d_
,
d_
);
vadd_d2_
->
Compute
(
checked
,
gates
+
d_
,
gates
+
d_
,
d2_
);
act_gate_d2_
->
Compute
Deprecated
(
gates
+
d_
,
gates
+
d
_
);
act_gate_d2_
->
Compute
(
gates
+
d_
,
gates
+
d_
,
d2
_
);
/* C_t = C_t-1 * fgated + cand_gated * igated*/
act_cand_d_
->
Compute
Deprecated
(
gates
,
gates
);
act_cand_d_
->
Compute
(
gates
,
gates
,
d_
);
vmul_d_
->
Compute
(
gates
,
gates
+
d_
,
gates
+
d_
,
d_
);
vmul_d_
->
Compute
(
ct_1
,
gates
+
d2_
,
gates
+
d2_
,
d_
);
vadd_d_
->
Compute
(
gates
+
d_
,
gates
+
d2_
,
ct
,
d_
);
/* get ogated*/
vmul_d_
->
Compute
(
wp_data
+
d2_
,
ct
,
gates
+
d_
,
d_
);
vadd_d_
->
Compute
(
gates
+
d_
,
gates
+
d3_
,
gates
+
d3_
,
d_
);
act_gate_d_
->
Compute
Deprecated
(
gates
+
d3_
,
gates
+
d3
_
);
act_gate_d_
->
Compute
(
gates
+
d3_
,
gates
+
d3_
,
d
_
);
/* H_t = act_cell(C_t) * ogated */
act_cell_d_
->
Compute
Deprecated
(
ct
,
gates
+
d2
_
);
act_cell_d_
->
Compute
(
ct
,
gates
+
d2_
,
d
_
);
vmul_d_
->
Compute
(
gates
+
d2_
,
gates
+
d3_
,
ht
,
d_
);
}
void
ComputeC1H1
(
T
*
gates
,
T
*
ct
,
T
*
ht
,
const
T
*
wp_data
)
const
override
{
/* C_t = igated * cgated*/
act_gate_d_
->
Compute
Deprecated
(
gates
+
d_
,
gates
+
d_
);
act_cand_d_
->
Compute
Deprecated
(
gates
,
gates
);
act_gate_d_
->
Compute
(
gates
+
d_
,
gates
+
d_
,
d_
);
act_cand_d_
->
Compute
(
gates
,
gates
,
d_
);
vmul_d_
->
Compute
(
gates
,
gates
+
d_
,
ct
,
d_
);
/* get outgated, put W_oc * C_t on igated */
vmul_d_
->
Compute
(
wp_data
+
d2_
,
ct
,
gates
+
d_
,
d_
);
vadd_d_
->
Compute
(
gates
+
d_
,
gates
+
d3_
,
gates
+
d3_
,
d_
);
/* H_t = act_cell(C_t) * ogated */
act_gate_d_
->
Compute
Deprecated
(
gates
+
d3_
,
gates
+
d3
_
);
act_cell_d_
->
Compute
Deprecated
(
ct
,
gates
+
d2
_
);
act_gate_d_
->
Compute
(
gates
+
d3_
,
gates
+
d3_
,
d
_
);
act_cell_d_
->
Compute
(
ct
,
gates
+
d2_
,
d
_
);
vmul_d_
->
Compute
(
gates
+
d2_
,
gates
+
d3_
,
ht
,
d_
);
}
...
...
@@ -376,20 +376,20 @@ class GRUKernelImpl : public GRUKernel<T> {
}
void
ComputeH1
(
T
*
gates
,
T
*
ht
)
const
override
{
act_gate_d_
->
Compute
Deprecated
(
gates
,
gates
);
act_state_d_
->
Compute
Deprecated
(
gates
+
d2_
,
gates
+
d2
_
);
act_gate_d_
->
Compute
(
gates
,
gates
,
d_
);
act_state_d_
->
Compute
(
gates
+
d2_
,
gates
+
d2_
,
d
_
);
vmul_d_
->
Compute
(
gates
,
gates
+
d2_
,
ht
,
d_
);
}
void
ComputeHtPart1
(
T
*
gates
,
const
T
*
ht_1
,
T
*
ht
)
const
override
{
// W: {W_update, W_reset; W_state}
act_gate_d2_
->
Compute
Deprecated
(
gates
,
gates
);
act_gate_d2_
->
Compute
(
gates
,
gates
,
d2_
);
vmul_d_
->
Compute
(
ht_1
,
gates
+
d_
,
ht
,
d_
);
}
void
ComputeHtPart2
(
T
*
gates
,
const
T
*
ht_1
,
T
*
ht
)
const
override
{
T
*
y
=
gates
+
d2_
;
act_state_d_
->
Compute
Deprecated
(
y
,
y
);
act_state_d_
->
Compute
(
y
,
y
,
d_
);
// out = zt*ht~ + (1-zt)*ht_1
for
(
int
i
=
0
;
i
<
d_
;
++
i
)
{
ht
[
i
]
=
gates
[
i
]
*
y
[
i
]
+
(
static_cast
<
T
>
(
1
)
-
gates
[
i
])
*
ht_1
[
i
];
...
...
paddle/fluid/operators/math/jit_kernel_test.cc
浏览文件 @
e2d6eddd
...
...
@@ -181,7 +181,7 @@ TEST(JitKernel, vexp) {
auto
ttgts
=
GetCurrentUS
();
for
(
int
i
=
0
;
i
<
repeat
;
++
i
)
{
// ker->Compute
Deprecated
(x_data, ztgt_data);
// ker->Compute(x_data, ztgt_data);
ker
->
Compute
(
x_data
,
ztgt_data
,
d
);
}
auto
ttgte
=
GetCurrentUS
();
...
...
@@ -345,8 +345,8 @@ void lstm_ctht_ref(
const
std
::
shared_ptr
<
const
paddle
::
operators
::
math
::
jitkernel
::
VExpKernel
<
float
>>&
vexp_1
,
const
int
d
,
float
*
gates
,
const
float
*
ct_1
,
float
*
ct
,
float
*
ht
)
{
vsigmoid_3d
->
Compute
Deprecated
(
gates
+
d
,
gates
+
d
);
vtanh_d
->
Compute
Deprecated
(
gates
,
gates
);
vsigmoid_3d
->
Compute
(
gates
+
d
,
gates
+
d
,
3
*
d
);
vtanh_d
->
Compute
(
gates
,
gates
,
d
);
const
float
*
i
=
gates
+
d
,
*
f
=
gates
+
d
*
2
,
*
o
=
gates
+
d
*
3
;
const
float
min
=
SIGMOID_THRESHOLD_MIN
;
const
float
max
=
SIGMOID_THRESHOLD_MAX
;
...
...
@@ -356,7 +356,7 @@ void lstm_ctht_ref(
// H_t = act_cell(C_t) * ogated
float
tmp
=
ct
[
k
]
*
2
;
tmp
=
0.
f
-
((
tmp
<
min
)
?
min
:
((
tmp
>
max
)
?
max
:
tmp
));
vexp_1
->
Compute
Deprecated
(
&
tmp
,
&
tmp
);
vexp_1
->
Compute
(
&
tmp
,
&
tmp
,
1
);
tmp
=
2.
f
/
(
1.
f
+
tmp
)
-
1.
f
;
ht
[
k
]
=
tmp
*
o
[
k
];
}
...
...
@@ -374,13 +374,13 @@ void lstm_ctht_better(
const
paddle
::
operators
::
math
::
jitkernel
::
VAddKernel
<
float
>>&
vadd_d
,
const
int
d
,
float
*
gates
,
const
float
*
ct_1
,
float
*
ct
,
float
*
ht
)
{
int
d2
=
d
*
2
;
vsigmoid_3d
->
Compute
Deprecated
(
gates
+
d
,
gates
+
d
);
vtanh_d
->
Compute
Deprecated
(
gates
,
gates
);
vsigmoid_3d
->
Compute
(
gates
+
d
,
gates
+
d
,
3
*
d
);
vtanh_d
->
Compute
(
gates
,
gates
,
d
);
vmul_d
->
Compute
(
gates
,
gates
+
d
,
gates
+
d
,
d
);
vmul_d
->
Compute
(
ct_1
,
gates
+
d2
,
gates
+
d2
,
d
);
vadd_d
->
Compute
(
gates
+
d
,
gates
+
d2
,
ct
,
d
);
/* H_t = act_cell(C_t) * ogated */
vtanh_d
->
Compute
Deprecated
(
ct
,
gates
+
d2
);
vtanh_d
->
Compute
(
ct
,
gates
+
d2
,
d
);
vmul_d
->
Compute
(
gates
+
d2
,
gates
+
d
*
3
,
ht
,
d
);
}
...
...
@@ -737,7 +737,7 @@ void vaddrelu_better(
const
paddle
::
operators
::
math
::
jitkernel
::
VReluKernel
<
float
>>&
vrelu
,
const
float
*
x
,
const
float
*
y
,
float
*
z
,
int
d
)
{
vadd
->
Compute
(
x
,
y
,
z
,
d
);
vrelu
->
Compute
Deprecated
(
z
,
z
);
vrelu
->
Compute
(
z
,
z
,
d
);
}
TEST
(
JitKernel
,
vaddrelu
)
{
...
...
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