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
马赛克医院
Paddle
提交
b68ececb
P
Paddle
项目概览
马赛克医院
/
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
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
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
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
体验新版 GitCode,发现更多精彩内容 >>
提交
b68ececb
编写于
11月 06, 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 vaddrelu jitcode
test=develop
上级
bb09e310
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
66 addition
and
102 deletion
+66
-102
paddle/fluid/operators/math/jit_code.cc
paddle/fluid/operators/math/jit_code.cc
+15
-0
paddle/fluid/operators/math/jit_code.h
paddle/fluid/operators/math/jit_code.h
+9
-6
paddle/fluid/operators/math/jit_kernel.h
paddle/fluid/operators/math/jit_kernel.h
+6
-6
paddle/fluid/operators/math/jit_kernel_blas.cc
paddle/fluid/operators/math/jit_kernel_blas.cc
+35
-89
paddle/fluid/operators/math/jit_kernel_test.cc
paddle/fluid/operators/math/jit_kernel_test.cc
+1
-1
未找到文件。
paddle/fluid/operators/math/jit_code.cc
浏览文件 @
b68ececb
...
...
@@ -70,10 +70,16 @@ bool VAddJitCode::init(int d) { return MayIUse(avx); }
void
VAddJitCode
::
generate
()
{
int
offset
=
0
;
if
(
with_relu_
)
{
vxorps
(
ymm_zero
,
ymm_zero
,
ymm_zero
);
}
for
(
int
i
=
0
;
i
<
num_
/
AVX_FLOAT_BLOCK
;
++
i
)
{
vmovups
(
ymm_src1
,
ptr
[
param1
+
offset
]);
vmovups
(
ymm_src2
,
ptr
[
param2
+
offset
]);
vaddps
(
ymm_dst
,
ymm_src1
,
ymm_src2
);
if
(
with_relu_
)
{
vmaxps
(
ymm_dst
,
ymm_zero
,
ymm_dst
);
}
vmovups
(
ptr
[
param3
+
offset
],
ymm_dst
);
offset
+=
sizeof
(
float
)
*
AVX_FLOAT_BLOCK
;
}
...
...
@@ -82,6 +88,9 @@ void VAddJitCode::generate() {
vmovups
(
xmm_src1
,
ptr
[
param1
+
offset
]);
vmovups
(
xmm_src2
,
ptr
[
param2
+
offset
]);
vaddps
(
xmm_dst
,
xmm_src1
,
xmm_src2
);
if
(
with_relu_
)
{
vmaxps
(
xmm_dst
,
xmm_zero
,
xmm_dst
);
}
vmovups
(
ptr
[
param3
+
offset
],
xmm_dst
);
offset
+=
sizeof
(
float
)
*
4
;
rest
-=
4
;
...
...
@@ -90,6 +99,9 @@ void VAddJitCode::generate() {
vmovq
(
xmm_src1
,
ptr
[
param1
+
offset
]);
vmovq
(
xmm_src2
,
ptr
[
param2
+
offset
]);
vaddps
(
xmm_dst
,
xmm_src1
,
xmm_src2
);
if
(
with_relu_
)
{
vmaxps
(
xmm_dst
,
xmm_zero
,
xmm_dst
);
}
vmovq
(
ptr
[
param3
+
offset
],
xmm_dst
);
offset
+=
sizeof
(
float
)
*
2
;
rest
-=
2
;
...
...
@@ -98,6 +110,9 @@ void VAddJitCode::generate() {
vmovss
(
xmm_src1
,
ptr
[
param1
+
offset
]);
vmovss
(
xmm_src2
,
ptr
[
param2
+
offset
]);
vaddss
(
xmm_dst
,
xmm_src1
,
xmm_src2
);
if
(
with_relu_
)
{
vmaxps
(
xmm_dst
,
xmm_zero
,
xmm_dst
);
}
vmovss
(
ptr
[
param3
+
offset
],
xmm_dst
);
}
ret
();
...
...
paddle/fluid/operators/math/jit_code.h
浏览文件 @
b68ececb
...
...
@@ -46,35 +46,38 @@ class VMulJitCode : public JitCode {
xmm_t
xmm_src1
=
xmm_t
(
0
);
xmm_t
xmm_src2
=
xmm_t
(
1
);
xmm_t
xmm_dst
=
xmm_t
(
2
);
xmm_t
xmm_dst
=
xmm_t
(
1
);
ymm_t
ymm_src1
=
ymm_t
(
0
);
ymm_t
ymm_src2
=
ymm_t
(
1
);
ymm_t
ymm_dst
=
ymm_t
(
2
);
ymm_t
ymm_dst
=
ymm_t
(
1
);
};
class
VAddJitCode
:
public
JitCode
{
public:
DECLARE_JIT_CODE
(
VAddJitCode
);
explicit
VAddJitCode
(
int
d
,
size_t
code_size
=
256
*
1024
,
explicit
VAddJitCode
(
int
d
,
bool
with_relu
,
size_t
code_size
=
256
*
1024
,
void
*
code_ptr
=
nullptr
)
:
JitCode
(
code_size
,
code_ptr
),
num_
(
d
)
{}
:
JitCode
(
code_size
,
code_ptr
),
num_
(
d
)
,
with_relu_
(
with_relu
)
{}
static
bool
init
(
int
d
);
void
generate
()
override
;
private:
int
num_
;
bool
with_relu_
;
reg64_t
param1
{
abi_param1
};
reg64_t
param2
{
abi_param2
};
reg64_t
param3
{
abi_param3
};
xmm_t
xmm_src1
=
xmm_t
(
0
);
xmm_t
xmm_src2
=
xmm_t
(
1
);
xmm_t
xmm_dst
=
xmm_t
(
2
);
xmm_t
xmm_dst
=
xmm_t
(
1
);
xmm_t
xmm_zero
=
xmm_t
(
2
);
ymm_t
ymm_src1
=
ymm_t
(
0
);
ymm_t
ymm_src2
=
ymm_t
(
1
);
ymm_t
ymm_dst
=
ymm_t
(
2
);
ymm_t
ymm_dst
=
ymm_t
(
1
);
ymm_t
ymm_zero
=
ymm_t
(
2
);
};
}
// namespace gen
...
...
paddle/fluid/operators/math/jit_kernel.h
浏览文件 @
b68ececb
...
...
@@ -75,22 +75,22 @@ class VAddKernel : public Kernel {
};
template
<
typename
T
>
class
V
Scal
Kernel
:
public
Kernel
{
class
V
AddRelu
Kernel
:
public
Kernel
{
public:
virtual
void
Compute
(
const
T
a
,
const
T
*
x
,
T
*
y
)
const
=
0
;
virtual
void
Compute
(
const
T
a
,
T
*
x
)
const
=
0
;
void
(
*
Compute
)(
const
T
*
,
const
T
*
,
T
*
,
int
);
};
template
<
typename
T
>
class
V
AddBias
Kernel
:
public
Kernel
{
class
V
Scal
Kernel
:
public
Kernel
{
public:
virtual
void
Compute
(
const
T
a
,
const
T
*
x
,
T
*
y
)
const
=
0
;
virtual
void
Compute
(
const
T
a
,
T
*
x
)
const
=
0
;
};
template
<
typename
T
>
class
VAdd
Relu
Kernel
:
public
Kernel
{
class
VAdd
Bias
Kernel
:
public
Kernel
{
public:
virtual
void
Compute
(
const
T
*
x
,
const
T
*
y
,
T
*
z
)
const
=
0
;
virtual
void
Compute
(
const
T
a
,
const
T
*
x
,
T
*
y
)
const
=
0
;
};
template
<
typename
T
>
...
...
paddle/fluid/operators/math/jit_kernel_blas.cc
浏览文件 @
b68ececb
...
...
@@ -46,6 +46,14 @@ void VAddRefer(const T* x, const T* y, T* z, int n) {
}
}
template
<
typename
T
>
void
VAddReluRefer
(
const
T
*
x
,
const
T
*
y
,
T
*
z
,
int
n
)
{
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
z
[
i
]
=
x
[
i
]
+
y
[
i
];
z
[
i
]
=
z
[
i
]
>
0
?
z
[
i
]
:
0
;
}
}
#ifdef PADDLE_WITH_MKLML
template
<
typename
T
>
void
VMulMKL
(
const
T
*
x
,
const
T
*
y
,
T
*
z
,
int
n
);
...
...
@@ -131,7 +139,7 @@ class VAddKernelImpl : public VAddKernel<T> {
explicit
VAddKernelImpl
(
int
d
)
:
VAddKernel
<
T
>
()
{
if
(
useJIT
(
d
))
{
size_t
sz
=
96
+
d
/
AVX_FLOAT_BLOCK
*
4
*
8
;
jitcode_
.
reset
(
new
gen
::
VAddJitCode
(
d
,
sz
>
4096
?
sz
:
4096
));
jitcode_
.
reset
(
new
gen
::
VAddJitCode
(
d
,
false
,
sz
>
4096
?
sz
:
4096
));
this
->
Compute
=
jitcode_
->
getCode
<
void
(
*
)(
const
T
*
,
const
T
*
,
T
*
,
int
)
>
();
return
;
...
...
@@ -164,10 +172,36 @@ bool VAddKernelImpl<double>::useMKL(int d) {
return
true
;
}
/* VAddRelu JitKernel */
template
<
typename
T
>
class
VAddReluKernelImpl
:
public
VAddReluKernel
<
T
>
{
public:
DECLARE_STATIC_FUNC
;
explicit
VAddReluKernelImpl
(
int
d
)
:
VAddReluKernel
<
T
>
()
{
if
(
useJIT
(
d
))
{
size_t
sz
=
96
+
d
/
AVX_FLOAT_BLOCK
*
4
*
8
;
jitcode_
.
reset
(
new
gen
::
VAddJitCode
(
d
,
true
,
sz
>
4096
?
sz
:
4096
));
this
->
Compute
=
jitcode_
->
getCode
<
void
(
*
)(
const
T
*
,
const
T
*
,
T
*
,
int
)
>
();
return
;
}
this
->
Compute
=
VAddReluRefer
<
T
>
;
}
private:
std
::
unique_ptr
<
gen
::
VAddJitCode
>
jitcode_
{
nullptr
};
};
template
<
>
bool
VAddReluKernelImpl
<
float
>::
useJIT
(
int
d
)
{
return
gen
::
VAddJitCode
::
init
(
d
);
}
#undef DECLARE_STATIC_FUNC
REGISTER_JITKERNEL
(
vmul
,
VMulKernel
);
REGISTER_JITKERNEL
(
vadd
,
VAddKernel
);
REGISTER_JITKERNEL
(
vaddrelu
,
VAddReluKernel
);
/* VSCAL JitKernel */
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
,
jit_block
>
...
...
@@ -404,97 +438,9 @@ 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
,
kGT16
);
#endif
#ifdef __AVX2__
INTRI8_FLOAT
(
jit
::
avx2
);
INTRI16_FLOAT
(
jit
::
avx2
);
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
,
kGT16
);
#endif
#undef INTRI8_FLOAT
#undef INTRI16_FLOAT
#undef INTRI_COMMON_FLOAT
REGISTER_JITKERNEL_DEPRECATED
(
vscal
,
VScalKernel
);
REGISTER_JITKERNEL_DEPRECATED
(
vaddb
,
VAddBiasKernel
);
REGISTER_JITKERNEL_DEPRECATED
(
vrelu
,
VReluKernel
);
REGISTER_JITKERNEL_DEPRECATED
(
vaddrelu
,
VAddReluKernel
);
REGISTER_JITKERNEL_DEPRECATED
(
videntity
,
VIdentityKernel
);
}
// namespace jitkernel
...
...
paddle/fluid/operators/math/jit_kernel_test.cc
浏览文件 @
b68ececb
...
...
@@ -757,7 +757,7 @@ TEST(JitKernel, vaddrelu) {
auto
tmkle
=
GetCurrentUS
();
auto
ttgts
=
GetCurrentUS
();
for
(
int
i
=
0
;
i
<
repeat
;
++
i
)
{
ker
->
Compute
(
x_data
,
y_data
,
ztgt_data
);
ker
->
Compute
(
x_data
,
y_data
,
ztgt_data
,
d
);
}
auto
ttgte
=
GetCurrentUS
();
VLOG
(
3
)
<<
"Vec size "
<<
d
<<
": refer takes: "
<<
(
trefe
-
trefs
)
/
repeat
...
...
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