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
sven357
Paddle
提交
7cb19a59
P
Paddle
项目概览
sven357
/
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
7cb19a59
编写于
10月 19, 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
fuse elementwise_add and relu
上级
3c249283
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
112 addition
and
9 deletion
+112
-9
paddle/fluid/operators/math/fc_compute.h
paddle/fluid/operators/math/fc_compute.h
+15
-9
paddle/fluid/operators/math/jit_kernel.h
paddle/fluid/operators/math/jit_kernel.h
+6
-0
paddle/fluid/operators/math/jit_kernel_blas.cc
paddle/fluid/operators/math/jit_kernel_blas.cc
+91
-0
未找到文件。
paddle/fluid/operators/math/fc_compute.h
浏览文件 @
7cb19a59
...
...
@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/jit_kernel.h" // TODO(TJ): add deps
DECLARE_int32
(
paddle_num_threads
);
...
...
@@ -30,20 +31,25 @@ inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
if
(
B
==
NULL
)
{
return
;
}
if
(
relu
)
{
const
auto
&
vaddrelu
=
jitkernel
::
KernelPool
::
Instance
()
.
template
Get
<
jitkernel
::
VAddReluKernel
<
T
>
>
(
N
);
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
T
*
dst
=
Y
+
i
*
N
;
vaddrelu
->
Compute
(
B
,
dst
,
dst
);
}
}
else
{
const
auto
&
vadd
=
jitkernel
::
KernelPool
::
Instance
()
.
template
Get
<
jitkernel
::
VAddKernel
<
T
>
>
(
N
);
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for if (FLAGS_paddle_num_threads > 1)
#endif
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
blas
.
AXPY
(
N
,
static_cast
<
T
>
(
1
),
B
,
Y
+
i
*
N
);
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
T
*
dst
=
Y
+
i
*
N
;
vadd
->
Compute
(
B
,
dst
,
dst
);
}
}
if
(
!
relu
)
{
return
;
}
// TODO(TJ): fuse relu
LOG
(
FATAL
)
<<
"Not implemented!"
;
}
}
// namespace math
...
...
paddle/fluid/operators/math/jit_kernel.h
浏览文件 @
7cb19a59
...
...
@@ -86,6 +86,12 @@ class VAddBiasKernel : public Kernel {
virtual
void
Compute
(
const
T
a
,
const
T
*
x
,
T
*
y
)
const
=
0
;
};
template
<
typename
T
>
class
VAddReluKernel
:
public
Kernel
{
public:
virtual
void
Compute
(
const
T
*
x
,
const
T
*
y
,
T
*
z
)
const
=
0
;
};
template
<
typename
T
>
class
VActKernel
:
public
Kernel
{
public:
...
...
paddle/fluid/operators/math/jit_kernel_blas.cc
浏览文件 @
7cb19a59
...
...
@@ -378,11 +378,102 @@ class VIdentityKernelImpl : public VIdentityKernel<T> {
void
Compute
(
const
T
*
x
,
T
*
y
)
const
override
{}
};
/* VAddRelu JitKernel */
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
,
jit_block
>
class
VAddReluKernelImpl
:
public
VAddReluKernel
<
T
>
{
public:
explicit
VAddReluKernelImpl
(
int
d
)
:
VAddReluKernel
<
T
>
()
{
this
->
num_
=
d
;
}
void
Compute
(
const
T
*
x
,
const
T
*
y
,
T
*
z
)
const
override
{
for
(
int
i
=
0
;
i
<
this
->
num_
;
++
i
)
{
z
[
i
]
=
x
[
i
]
+
y
[
i
];
z
[
i
]
=
z
[
i
]
>
0
?
z
[
i
]
:
0
;
}
}
};
#define INTRI8_FLOAT(isa) \
template <> \
void VAddReluKernelImpl<float, isa, kEQ8>::Compute( \
const float* x, const float* y, float* z) const { \
__m256 tmpx = _mm256_loadu_ps(x); \
__m256 tmpy = _mm256_loadu_ps(y); \
tmpy = _mm256_add_ps(tmpx, tmpy); \
tmpy = _mm256_max_ps(tmpy, _mm256_setzero_ps()); \
_mm256_storeu_ps(z, tmpy); \
}
#define INTRI16_FLOAT(isa) \
template <> \
void VAddReluKernelImpl<float, isa, kEQ16>::Compute( \
const float* x, const float* y, float* z) const { \
__m256 zeros = _mm256_setzero_ps(); \
__m256 tmp0 = _mm256_loadu_ps(x); \
__m256 tmp1 = _mm256_loadu_ps(y); \
tmp0 = _mm256_add_ps(tmp0, tmp1); \
tmp0 = _mm256_max_ps(tmp0, zeros); \
tmp1 = _mm256_loadu_ps(x + 8); \
__m256 tmp2 = _mm256_loadu_ps(y + 8); \
tmp1 = _mm256_add_ps(tmp1, tmp2); \
tmp1 = _mm256_max_ps(tmp1, zeros); \
_mm256_storeu_ps(z, tmp0); \
_mm256_storeu_ps(z + 8, tmp1); \
}
#define INTRI_COMMON_FLOAT(isa, block) \
template <> \
VAddReluKernelImpl<float, isa, block>::VAddReluKernelImpl(int d) \
: VAddReluKernel<float>() { \
this->num_ = d; \
this->end_ = d - d % AVX_FLOAT_BLOCK; \
this->rest_ = d - this->end_; \
} \
template <> \
void VAddReluKernelImpl<float, isa, block>::Compute( \
const float* x, const float* y, float* z) const { \
__m256 zeros = _mm256_setzero_ps(); \
for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) { \
__m256 tmpx = _mm256_loadu_ps(x + i); \
__m256 tmpy = _mm256_loadu_ps(y + i); \
tmpy = _mm256_add_ps(tmpx, tmpy); \
tmpy = _mm256_max_ps(tmpy, zeros); \
_mm256_storeu_ps(z + i, tmpy); \
} \
for (int i = this->end_; i < this->num_; ++i) { \
z[i] = x[i] + y[i]; \
z[i] = z[i] > 0 ? z[i] : 0; \
} \
}
#ifdef __AVX__
INTRI8_FLOAT
(
jit
::
avx
);
INTRI16_FLOAT
(
jit
::
avx
);
INTRI_COMMON_FLOAT
(
jit
::
avx
,
kGT8LT16
);
INTRI_COMMON_FLOAT
(
jit
::
avx
,
kGT16
);
#endif
#ifdef __AVX2__
INTRI8_FLOAT
(
jit
::
avx2
);
INTRI16_FLOAT
(
jit
::
avx2
);
INTRI_COMMON_FLOAT
(
jit
::
avx2
,
kGT8LT16
);
INTRI_COMMON_FLOAT
(
jit
::
avx2
,
kGT16
);
#endif
#ifdef __AVX512F__
// TODO(TJ): refine avx512
INTRI8_FLOAT
(
jit
::
avx512f
);
INTRI16_FLOAT
(
jit
::
avx512f
);
INTRI_COMMON_FLOAT
(
jit
::
avx512f
,
kGT8LT16
);
INTRI_COMMON_FLOAT
(
jit
::
avx512f
,
kGT16
);
#endif
#undef INTRI8_FLOAT
#undef INTRI16_FLOAT
#undef INTRI_COMMON_FLOAT
REGISTER_JITKERNEL
(
vmul
,
VMulKernel
);
REGISTER_JITKERNEL
(
vadd
,
VAddKernel
);
REGISTER_JITKERNEL
(
vscal
,
VScalKernel
);
REGISTER_JITKERNEL
(
vaddb
,
VAddBiasKernel
);
REGISTER_JITKERNEL
(
vrelu
,
VReluKernel
);
REGISTER_JITKERNEL
(
vaddrelu
,
VAddReluKernel
);
REGISTER_JITKERNEL
(
videntity
,
VIdentityKernel
);
}
// namespace jitkernel
...
...
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