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
7ef2699e
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2297
Star
20931
Fork
5422
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
体验新版 GitCode,发现更多精彩内容 >>
提交
7ef2699e
编写于
10月 11, 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
init peephole runtime kernel
上级
3ee8f2c6
变更
4
显示空白变更内容
内联
并排
Showing
4 changed file
with
104 addition
and
34 deletion
+104
-34
paddle/fluid/operators/fusion_lstm_op.cc
paddle/fluid/operators/fusion_lstm_op.cc
+6
-8
paddle/fluid/operators/math/jit_kernel.h
paddle/fluid/operators/math/jit_kernel.h
+2
-1
paddle/fluid/operators/math/jit_kernel_lstm.cc
paddle/fluid/operators/math/jit_kernel_lstm.cc
+84
-18
paddle/fluid/operators/math/jit_kernel_test.cc
paddle/fluid/operators/math/jit_kernel_test.cc
+12
-7
未找到文件。
paddle/fluid/operators/fusion_lstm_op.cc
浏览文件 @
7ef2699e
...
...
@@ -400,10 +400,9 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
}
else
{
const
auto
&
ker
=
math
::
jitkernel
::
KernelPool
::
Instance
()
.
template
Get
<
math
::
jitkernel
::
LSTMKernel
<
T
>,
int
,
const
std
::
string
&
,
const
std
::
string
&
,
const
std
::
string
&>
(
D
,
act_gate_str
,
act_cand_str
,
act_cell_str
);
.
template
Get
<
math
::
jitkernel
::
LSTMKernel
<
T
>,
const
std
::
string
&
,
const
std
::
string
&
,
const
std
::
string
&>
(
act_gate_str
,
act_cand_str
,
act_cell_str
,
D
,
false
);
for
(
int
i
=
0
;
i
<
N
;
++
i
)
{
PROCESS_H0C0
...
...
@@ -545,10 +544,9 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
}
else
{
const
auto
&
ker
=
math
::
jitkernel
::
KernelPool
::
Instance
()
.
template
Get
<
math
::
jitkernel
::
LSTMKernel
<
T
>,
int
,
const
std
::
string
&
,
const
std
::
string
&
,
const
std
::
string
&>
(
D
,
act_gate_str
,
act_cand_str
,
act_cell_str
);
.
template
Get
<
math
::
jitkernel
::
LSTMKernel
<
T
>,
const
std
::
string
&
,
const
std
::
string
&
,
const
std
::
string
&>
(
act_gate_str
,
act_cand_str
,
act_cell_str
,
D
,
false
);
for
(
int
step
=
tstart
;
step
<
max_seq_len
;
++
step
)
{
const
int
cur_bs
=
batch_starts
[
step
+
1
]
-
batch_starts
[
step
];
...
...
paddle/fluid/operators/math/jit_kernel.h
浏览文件 @
7ef2699e
...
...
@@ -125,7 +125,8 @@ class VTanhKernel : public VActKernel<T> {
template
<
typename
T
>
class
LSTMKernel
:
public
Kernel
{
public:
virtual
void
ComputeCtHt
(
T
*
gates
,
const
T
*
ct_1
,
T
*
ct
,
T
*
ht
)
const
=
0
;
virtual
void
ComputeCtHt
(
T
*
gates
,
const
T
*
ct_1
,
T
*
ct
,
T
*
ht
,
T
*
checked
=
nullptr
)
const
=
0
;
};
}
// namespace jitkernel
...
...
paddle/fluid/operators/math/jit_kernel_lstm.cc
浏览文件 @
7ef2699e
...
...
@@ -86,9 +86,9 @@ __m256 AVXActImpl<kIdentity>::Compute(__m256 x) const {
template
<
typename
T
,
jit
::
cpu_isa_t
isa
,
jit_block
>
class
LSTMKernelImpl
:
public
LSTMKernel
<
T
>
{
public:
explicit
LSTMKernelImpl
(
int
d
,
const
std
::
string
&
act_gate
,
explicit
LSTMKernelImpl
(
const
std
::
string
&
act_gate
,
const
std
::
string
&
act_cand
,
const
std
::
string
&
act_cell
)
const
std
::
string
&
act_cell
,
int
d
)
:
LSTMKernel
<
T
>
()
{
d_
=
d
;
d2_
=
d
*
2
;
...
...
@@ -134,7 +134,8 @@ class LSTMKernelImpl : public LSTMKernel<T> {
#endif
}
void
ComputeCtHt
(
T
*
gates
,
const
T
*
ct_1
,
T
*
ct
,
T
*
ht
)
const
override
{
void
ComputeCtHt
(
T
*
gates
,
const
T
*
ct_1
,
T
*
ct
,
T
*
ht
,
T
*
checked
)
const
override
{
// gates: W_ch, W_ih, W_fh, W_oh
act_gate_3d_
->
Compute
(
gates
+
d_
,
gates
+
d_
);
...
...
@@ -162,7 +163,8 @@ class LSTMKernelImpl : public LSTMKernel<T> {
#define INTRI8_FLOAT(isa) \
template <> \
void LSTMKernelImpl<float, isa, kEQ8>::ComputeCtHt( \
float* gates, const float* ct_1, float* ct, float* ht) const { \
float* gates, const float* ct_1, float* ct, float* ht, float* checked) \
const { \
/* gates: W_ch, W_ih, W_fh, W_oh */
\
__m256 c, i, f, o; \
c = _mm256_loadu_ps(gates); \
...
...
@@ -192,21 +194,86 @@ INTRI8_FLOAT(jit::avx2);
INTRI8_FLOAT
(
jit
::
avx512f
);
#endif
/* Peephole JitKernel */
template
<
typename
T
,
jit
::
cpu_isa_t
isa
,
jit_block
>
class
PeepholeKernelImpl
:
public
LSTMKernel
<
T
>
{
public:
explicit
PeepholeKernelImpl
(
const
std
::
string
&
act_gate
,
const
std
::
string
&
act_cand
,
const
std
::
string
&
act_cell
,
int
d
)
:
LSTMKernel
<
T
>
()
{
d_
=
d
;
d2_
=
d
*
2
;
d3_
=
d
*
3
;
auto
GetActKernel
=
[
&
](
const
std
::
string
&
type
,
int
n
)
->
std
::
shared_ptr
<
const
VActKernel
<
T
>>
{
if
(
type
==
"sigmoid"
)
{
return
std
::
dynamic_pointer_cast
<
const
VActKernel
<
T
>>
(
KernelPool
::
Instance
().
template
Get
<
VSigmoidKernel
<
T
>
>
(
n
));
}
else
if
(
type
==
"relu"
)
{
return
std
::
dynamic_pointer_cast
<
const
VActKernel
<
T
>>
(
KernelPool
::
Instance
().
template
Get
<
VReluKernel
<
T
>
>
(
n
));
}
else
if
(
type
==
"tanh"
)
{
return
std
::
dynamic_pointer_cast
<
const
VActKernel
<
T
>>
(
KernelPool
::
Instance
().
template
Get
<
VTanhKernel
<
T
>
>
(
n
));
}
else
if
(
type
==
"identity"
||
type
==
""
)
{
return
std
::
dynamic_pointer_cast
<
const
VActKernel
<
T
>>
(
KernelPool
::
Instance
().
template
Get
<
VIdentityKernel
<
T
>
>
(
n
));
}
PADDLE_THROW
(
"Not support type: %s"
,
type
);
};
act_gate_3d_
=
GetActKernel
(
act_gate
,
d
*
3
);
act_cand_d_
=
GetActKernel
(
act_cand
,
d
);
act_cell_d_
=
GetActKernel
(
act_cell
,
d
);
vmul_d_
=
KernelPool
::
Instance
().
template
Get
<
VMulKernel
<
T
>
>
(
d
);
vadd_d_
=
KernelPool
::
Instance
().
template
Get
<
VAddKernel
<
T
>
>
(
d
);
}
void
ComputeCtHt
(
T
*
gates
,
const
T
*
ct_1
,
T
*
ct
,
T
*
ht
,
T
*
checked
)
const
override
{
// gates: W_ch, W_ih, W_fh, W_oh
act_gate_3d_
->
Compute
(
gates
+
d_
,
gates
+
d_
);
/* C_t = C_t-1 * fgated + cand_gated * igated */
act_cand_d_
->
Compute
(
gates
,
gates
);
vmul_d_
->
Compute
(
gates
,
gates
+
d_
,
gates
+
d_
);
vmul_d_
->
Compute
(
ct_1
,
gates
+
d2_
,
gates
+
d2_
);
vadd_d_
->
Compute
(
gates
+
d_
,
gates
+
d2_
,
ct
);
/* H_t = act_cell(C_t) * ogated */
act_cell_d_
->
Compute
(
ct
,
gates
+
d2_
);
vmul_d_
->
Compute
(
gates
+
d2_
,
gates
+
d3_
,
ht
);
}
private:
int
d_
,
d2_
,
d3_
;
std
::
shared_ptr
<
const
VActKernel
<
T
>>
act_gate_3d_
,
act_cand_d_
,
act_cell_d_
;
std
::
shared_ptr
<
const
VMulKernel
<
T
>>
vmul_d_
;
std
::
shared_ptr
<
const
VAddKernel
<
T
>>
vadd_d_
;
};
#define JITKERNEL_DECLARE_LSTM(ker_class, ker_dtype) \
template <> \
std::shared_ptr<const
ker_class<ker_dtype>>
\
KernelPool::Get<
ker_class<ker_dtype>, int, const std::string&,
\
const std::string&, const std::string&
>(
\
int d, const std::string& act_gate, const std::string& act_cand,
\
const std::string& act_cell)
std::shared_ptr<const
LSTMKernel<ker_dtype>>
\
KernelPool::Get<
LSTMKernel<ker_dtype>, const std::string&,
\
const std::string&, const std::string&
, int, bool>(
\
const std::string& act_gate, const std::string& act_cand,
\
const std::string& act_cell
, int d, bool use_peephole
)
#define JITKERNEL_KEY_LSTM(ker_key, dtype_key) \
#ker_key #dtype_key + std::to_string(d) + act_gate + act_cand + act_cell
#ker_key #dtype_key + std::to_string(d) + act_gate + act_cand + act_cell + \
(use_peephole ? "p" : "n")
#define JITKERNEL_NEW_LSTM_IMPL(ker, dtype, isa, k) \
if (use_peephole) { \
p = std::dynamic_pointer_cast<ker<dtype>>( \
std::make_shared<ker##Impl<dtype, isa, k>>(d, act_gate, act_cand, \
act_cell))
std::make_shared<PeepholeKernelImpl<dtype, isa, k>>( \
act_gate, act_cand, act_cell, d)); \
} else { \
p = std::dynamic_pointer_cast<ker<dtype>>( \
std::make_shared<ker##Impl<dtype, isa, k>>(act_gate, act_cand, \
act_cell, d)); \
}
REGISTER_JITKERNEL_ARGS
(
lstm
,
LSTMKernel
,
JITKERNEL_DECLARE_LSTM
,
JITKERNEL_KEY_LSTM
,
JITKERNEL_NEW_LSTM_IMPL
);
...
...
@@ -215,7 +282,6 @@ REGISTER_JITKERNEL_ARGS(lstm, LSTMKernel, JITKERNEL_DECLARE_LSTM,
#undef JITKERNEL_DECLARE_LSTM
#undef JITKERNEL_KEY_LSTM
#undef JITKERNEL_NEW_LSTM_IMPL
}
// namespace jitkernel
}
// namespace math
}
// namespace operators
...
...
paddle/fluid/operators/math/jit_kernel_test.cc
浏览文件 @
7ef2699e
...
...
@@ -390,9 +390,9 @@ TEST(JitKernel, lstm) {
std
::
string
act_gate
=
"sigmoid"
,
act_cand
=
"tanh"
,
act_cell
=
"tanh"
;
const
auto
&
ker
=
jit
::
KernelPool
::
Instance
()
.
template
Get
<
jit
::
LSTMKernel
<
float
>,
int
,
const
std
::
string
&
,
.
template
Get
<
jit
::
LSTMKernel
<
float
>,
const
std
::
string
&
,
const
std
::
string
&
,
const
std
::
string
&>
(
d
,
act_gate
,
act_cand
,
act_cell
);
act_gate
,
act_cand
,
act_cell
,
d
,
false
);
// below kernels are used to compute refer
const
auto
&
vsigmoid_3d
=
jit
::
KernelPool
::
Instance
().
template
Get
<
jit
::
VSigmoidKernel
<
float
>
>
(
...
...
@@ -717,15 +717,20 @@ TEST(JitKernel, pool) {
std
::
string
act_gate
=
"sigmoid"
,
act_cand
=
"tanh"
,
act_cell
=
"tanh"
;
const
auto
&
plstm1
=
jit
::
KernelPool
::
Instance
()
.
template
Get
<
jit
::
LSTMKernel
<
float
>,
int
,
const
std
::
string
&
,
.
template
Get
<
jit
::
LSTMKernel
<
float
>,
const
std
::
string
&
,
const
std
::
string
&
,
const
std
::
string
&>
(
frame_size
,
act_gate
,
act_cand
,
act_cell
);
act_gate
,
act_cand
,
act_cell
,
frame_size
,
false
);
const
auto
&
plstm2
=
jit
::
KernelPool
::
Instance
()
.
template
Get
<
jit
::
LSTMKernel
<
float
>,
int
,
const
std
::
string
&
,
.
template
Get
<
jit
::
LSTMKernel
<
float
>,
const
std
::
string
&
,
const
std
::
string
&
,
const
std
::
string
&>
(
frame_size
,
act_gate
,
act_cand
,
act_cell
);
EXPECT_EQ
(
plstm1
,
plstm2
);
act_gate
,
act_cand
,
act_cell
,
frame_size
,
false
);
const
auto
&
peephole
=
jit
::
KernelPool
::
Instance
()
.
template
Get
<
jit
::
LSTMKernel
<
float
>,
const
std
::
string
&
,
const
std
::
string
&
,
const
std
::
string
&>
(
act_gate
,
act_cand
,
act_cell
,
frame_size
,
true
);
EXPECT_TRUE
(
plstm1
!=
peephole
);
const
auto
&
pvmul_f
=
jit
::
KernelPool
::
Instance
().
template
Get
<
jit
::
VMulKernel
<
float
>
>
(
4
);
...
...
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