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
飞不过沧海
Paddle
提交
4a93db92
P
Paddle
项目概览
飞不过沧海
/
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
4a93db92
编写于
12月 05, 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 jit namespace
test=develop
上级
8cda28f3
变更
18
隐藏空白更改
内联
并排
Showing
18 changed file
with
167 addition
and
179 deletion
+167
-179
paddle/fluid/operators/attention_lstm_op.cc
paddle/fluid/operators/attention_lstm_op.cc
+8
-8
paddle/fluid/operators/fused/fused_embedding_fc_lstm_op.cc
paddle/fluid/operators/fused/fused_embedding_fc_lstm_op.cc
+3
-3
paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc
...le/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc
+3
-3
paddle/fluid/operators/math/cpu_vec.h
paddle/fluid/operators/math/cpu_vec.h
+72
-76
paddle/fluid/operators/math/cpu_vec_test.cc
paddle/fluid/operators/math/cpu_vec_test.cc
+30
-24
paddle/fluid/operators/math/jit_code.cc
paddle/fluid/operators/math/jit_code.cc
+1
-1
paddle/fluid/operators/math/jit_code.h
paddle/fluid/operators/math/jit_code.h
+1
-1
paddle/fluid/operators/math/jit_gen.cc
paddle/fluid/operators/math/jit_gen.cc
+1
-1
paddle/fluid/operators/math/jit_kernel.cc
paddle/fluid/operators/math/jit_kernel.cc
+0
-2
paddle/fluid/operators/math/jit_kernel_blas.cc
paddle/fluid/operators/math/jit_kernel_blas.cc
+1
-2
paddle/fluid/operators/math/jit_kernel_crf_decode.cc
paddle/fluid/operators/math/jit_kernel_crf_decode.cc
+11
-13
paddle/fluid/operators/math/jit_kernel_exp.cc
paddle/fluid/operators/math/jit_kernel_exp.cc
+0
-1
paddle/fluid/operators/math/jit_kernel_layer_norm.cc
paddle/fluid/operators/math/jit_kernel_layer_norm.cc
+10
-12
paddle/fluid/operators/math/jit_kernel_macro.h
paddle/fluid/operators/math/jit_kernel_macro.h
+18
-19
paddle/fluid/operators/math/jit_kernel_test.cc
paddle/fluid/operators/math/jit_kernel_test.cc
+1
-1
paddle/fluid/platform/cpu_info.cc
paddle/fluid/platform/cpu_info.cc
+0
-2
paddle/fluid/platform/cpu_info.h
paddle/fluid/platform/cpu_info.h
+0
-3
paddle/fluid/platform/init.cc
paddle/fluid/platform/init.cc
+7
-7
未找到文件。
paddle/fluid/operators/attention_lstm_op.cc
浏览文件 @
4a93db92
...
...
@@ -231,10 +231,10 @@ use lstm_x_t as input and compute as standard LSTM.
template
<
typename
T
>
inline
void
bias_relu
(
const
int
n
,
const
T
*
x
,
const
T
*
bias
,
T
*
y
)
{
if
(
bias
)
{
math
::
vec_add_bias
<
T
,
platform
::
jit
::
avx
>
(
n
,
*
bias
,
x
,
y
);
math
::
vec_relu
<
T
,
platform
::
jit
::
avx
>
(
n
,
y
,
y
);
math
::
vec_add_bias
<
T
,
platform
::
avx
>
(
n
,
*
bias
,
x
,
y
);
math
::
vec_relu
<
T
,
platform
::
avx
>
(
n
,
y
,
y
);
}
else
{
math
::
vec_relu
<
T
,
platform
::
jit
::
avx
>
(
n
,
x
,
y
);
math
::
vec_relu
<
T
,
platform
::
avx
>
(
n
,
x
,
y
);
}
}
...
...
@@ -245,8 +245,8 @@ inline void vec_softmax(const int n, const T* x, T* y) {
for
(
int
i
=
1
;
i
<
n
;
++
i
)
{
scalar
=
scalar
<
x
[
i
]
?
x
[
i
]
:
scalar
;
}
math
::
vec_add_bias
<
T
,
platform
::
jit
::
avx
>
(
n
,
-
scalar
,
x
,
y
);
// sub
math
::
vec_exp
<
T
>
(
n
,
y
,
y
);
// exp
math
::
vec_add_bias
<
T
,
platform
::
avx
>
(
n
,
-
scalar
,
x
,
y
);
// sub
math
::
vec_exp
<
T
>
(
n
,
y
,
y
);
// exp
// sum
scalar
=
T
(
0
);
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
...
...
@@ -302,13 +302,13 @@ class AttentionLSTMKernel : public framework::OpKernel<T> {
auto
&
act_gate_str
=
ctx
.
Attr
<
std
::
string
>
(
"gate_activation"
);
auto
&
act_cell_str
=
ctx
.
Attr
<
std
::
string
>
(
"cell_activation"
);
auto
&
act_cand_str
=
ctx
.
Attr
<
std
::
string
>
(
"candidate_activation"
);
if
(
platform
::
jit
::
MayIUse
(
platform
::
jit
::
avx
))
{
math
::
VecActivations
<
T
,
platform
::
jit
::
avx
>
act_functor
;
if
(
platform
::
MayIUse
(
platform
::
avx
))
{
math
::
VecActivations
<
T
,
platform
::
avx
>
act_functor
;
act_gate
=
act_functor
(
act_gate_str
);
act_cell
=
act_functor
(
act_cell_str
);
act_cand
=
act_functor
(
act_cand_str
);
}
else
{
math
::
VecActivations
<
T
,
platform
::
jit
::
isa_any
>
act_functor
;
math
::
VecActivations
<
T
,
platform
::
isa_any
>
act_functor
;
act_gate
=
act_functor
(
act_gate_str
);
act_cell
=
act_functor
(
act_cell_str
);
act_cand
=
act_functor
(
act_cand_str
);
...
...
paddle/fluid/operators/fused/fused_embedding_fc_lstm_op.cc
浏览文件 @
4a93db92
...
...
@@ -217,13 +217,13 @@ class FusedEmbeddingFCLSTMKernel : public framework::OpKernel<T> {
auto& act_gate_str = ctx.Attr<std::string>("gate_activation"); \
auto& act_cell_str = ctx.Attr<std::string>("cell_activation"); \
auto& act_cand_str = ctx.Attr<std::string>("candidate_activation"); \
if (platform::
jit::MayIUse(platform::jit::avx)) {
\
math::VecActivations<T, platform::
jit::avx> act_functor;
\
if (platform::
MayIUse(platform::avx)) {
\
math::VecActivations<T, platform::
avx> act_functor;
\
act_gate = act_functor(act_gate_str); \
act_cell = act_functor(act_cell_str); \
act_cand = act_functor(act_cand_str); \
} else { \
math::VecActivations<T, platform::
jit::isa_any> act_functor;
\
math::VecActivations<T, platform::
isa_any> act_functor;
\
act_gate = act_functor(act_gate_str); \
act_cell = act_functor(act_cell_str); \
act_cand = act_functor(act_cand_str); \
...
...
paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc
浏览文件 @
4a93db92
...
...
@@ -151,11 +151,11 @@ class FusionSeqExpandConcatFCOpKernel : public framework::OpKernel<T> {
std
::
function
<
void
(
const
int
,
const
T
*
,
T
*
)
>
fc_act
;
auto
&
fc_act_str
=
ctx
.
Attr
<
std
::
string
>
(
"fc_activation"
);
if
(
platform
::
jit
::
MayIUse
(
platform
::
jit
::
avx
))
{
math
::
VecActivations
<
T
,
platform
::
jit
::
avx
>
act_functor
;
if
(
platform
::
MayIUse
(
platform
::
avx
))
{
math
::
VecActivations
<
T
,
platform
::
avx
>
act_functor
;
fc_act
=
act_functor
(
fc_act_str
);
}
else
{
math
::
VecActivations
<
T
,
platform
::
jit
::
isa_any
>
act_functor
;
math
::
VecActivations
<
T
,
platform
::
isa_any
>
act_functor
;
fc_act
=
act_functor
(
fc_act_str
);
}
...
...
paddle/fluid/operators/math/cpu_vec.h
浏览文件 @
4a93db92
...
...
@@ -77,7 +77,7 @@ inline void vec_scal<double>(const int n, const double a, double* x) {
#endif
// MKL scal only support inplace, choose this if src and dst are not equal
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
=
platform
::
jit
::
isa_any
>
template
<
typename
T
,
platform
::
cpu_isa_t
isa
=
platform
::
isa_any
>
inline
void
vec_scal
(
const
int
n
,
const
T
a
,
const
T
*
x
,
T
*
y
)
{
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
y
[
i
]
=
a
*
x
[
i
];
...
...
@@ -85,12 +85,12 @@ inline void vec_scal(const int n, const T a, const T* x, T* y) {
}
template
<
>
inline
void
vec_scal
<
float
,
platform
::
jit
::
avx
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
inline
void
vec_scal
<
float
,
platform
::
avx
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
#ifdef __AVX__
constexpr
int
block
=
YMM_FLOAT_BLOCK
;
if
(
n
<
block
)
{
vec_scal
<
float
,
platform
::
jit
::
isa_any
>
(
n
,
a
,
x
,
y
);
vec_scal
<
float
,
platform
::
isa_any
>
(
n
,
a
,
x
,
y
);
return
;
}
const
int
rest
=
n
%
block
;
...
...
@@ -114,24 +114,24 @@ inline void vec_scal<float, platform::jit::avx>(const int n, const float a,
y
[
i
]
=
a
*
x
[
i
];
}
#else
vec_scal
<
float
,
platform
::
jit
::
isa_any
>
(
n
,
a
,
x
,
y
);
vec_scal
<
float
,
platform
::
isa_any
>
(
n
,
a
,
x
,
y
);
#endif
}
template
<
>
inline
void
vec_scal
<
float
,
platform
::
jit
::
avx2
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
vec_scal
<
float
,
platform
::
jit
::
avx
>
(
n
,
a
,
x
,
y
);
inline
void
vec_scal
<
float
,
platform
::
avx2
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
vec_scal
<
float
,
platform
::
avx
>
(
n
,
a
,
x
,
y
);
}
template
<
>
inline
void
vec_scal
<
float
,
platform
::
jit
::
avx512f
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
inline
void
vec_scal
<
float
,
platform
::
avx512f
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
// TODO(TJ): enable me
vec_scal
<
float
,
platform
::
jit
::
avx2
>
(
n
,
a
,
x
,
y
);
vec_scal
<
float
,
platform
::
avx2
>
(
n
,
a
,
x
,
y
);
}
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
=
platform
::
jit
::
isa_any
>
template
<
typename
T
,
platform
::
cpu_isa_t
isa
=
platform
::
isa_any
>
inline
void
vec_bias_sub
(
const
int
n
,
const
T
a
,
const
T
*
x
,
T
*
y
)
{
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
y
[
i
]
=
a
-
x
[
i
];
...
...
@@ -139,12 +139,12 @@ inline void vec_bias_sub(const int n, const T a, const T* x, T* y) {
}
template
<
>
inline
void
vec_bias_sub
<
float
,
platform
::
jit
::
avx
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
inline
void
vec_bias_sub
<
float
,
platform
::
avx
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
#ifdef __AVX__
constexpr
int
block
=
YMM_FLOAT_BLOCK
;
if
(
n
<
block
)
{
vec_bias_sub
<
float
,
platform
::
jit
::
isa_any
>
(
n
,
a
,
x
,
y
);
vec_bias_sub
<
float
,
platform
::
isa_any
>
(
n
,
a
,
x
,
y
);
return
;
}
const
int
rest
=
n
%
block
;
...
...
@@ -168,27 +168,25 @@ inline void vec_bias_sub<float, platform::jit::avx>(const int n, const float a,
y
[
i
]
=
a
-
x
[
i
];
}
#else
vec_bias_sub
<
float
,
platform
::
jit
::
isa_any
>
(
n
,
a
,
x
,
y
);
vec_bias_sub
<
float
,
platform
::
isa_any
>
(
n
,
a
,
x
,
y
);
#endif
}
template
<
>
inline
void
vec_bias_sub
<
float
,
platform
::
jit
::
avx2
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
vec_bias_sub
<
float
,
platform
::
jit
::
avx
>
(
n
,
a
,
x
,
y
);
inline
void
vec_bias_sub
<
float
,
platform
::
avx2
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
vec_bias_sub
<
float
,
platform
::
avx
>
(
n
,
a
,
x
,
y
);
}
template
<
>
inline
void
vec_bias_sub
<
float
,
platform
::
jit
::
avx512f
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
inline
void
vec_bias_sub
<
float
,
platform
::
avx512f
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
// TODO(TJ): enable me
vec_bias_sub
<
float
,
platform
::
jit
::
avx2
>
(
n
,
a
,
x
,
y
);
vec_bias_sub
<
float
,
platform
::
avx2
>
(
n
,
a
,
x
,
y
);
}
// out = x*y + (1-x)*z
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
=
platform
::
jit
::
isa_any
>
template
<
typename
T
,
platform
::
cpu_isa_t
isa
=
platform
::
isa_any
>
inline
void
vec_cross
(
const
int
n
,
const
T
*
x
,
const
T
*
y
,
const
T
*
z
,
T
*
out
)
{
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
out
[
i
]
=
x
[
i
]
*
y
[
i
]
+
(
static_cast
<
T
>
(
1
)
-
x
[
i
])
*
z
[
i
];
...
...
@@ -196,13 +194,13 @@ inline void vec_cross(const int n, const T* x, const T* y, const T* z, T* out) {
}
template
<
>
inline
void
vec_cross
<
float
,
platform
::
jit
::
avx
>
(
const
int
n
,
const
float
*
x
,
const
float
*
y
,
const
float
*
z
,
float
*
out
)
{
inline
void
vec_cross
<
float
,
platform
::
avx
>
(
const
int
n
,
const
float
*
x
,
const
float
*
y
,
const
float
*
z
,
float
*
out
)
{
#ifdef __AVX__
constexpr
int
block
=
YMM_FLOAT_BLOCK
;
if
(
n
<
block
)
{
vec_cross
<
float
,
platform
::
jit
::
isa_any
>
(
n
,
x
,
y
,
z
,
out
);
vec_cross
<
float
,
platform
::
isa_any
>
(
n
,
x
,
y
,
z
,
out
);
return
;
}
const
int
rest
=
n
%
block
;
...
...
@@ -228,25 +226,26 @@ inline void vec_cross<float, platform::jit::avx>(const int n, const float* x,
out
[
i
]
=
x
[
i
]
*
y
[
i
]
+
(
1.
f
-
x
[
i
])
*
z
[
i
];
}
#else
vec_cross
<
float
,
platform
::
jit
::
isa_any
>
(
n
,
x
,
y
,
z
,
out
);
vec_cross
<
float
,
platform
::
isa_any
>
(
n
,
x
,
y
,
z
,
out
);
#endif
}
template
<
>
inline
void
vec_cross
<
float
,
platform
::
jit
::
avx2
>
(
const
int
n
,
const
float
*
x
,
const
float
*
y
,
const
float
*
z
,
float
*
out
)
{
vec_cross
<
float
,
platform
::
jit
::
avx
>
(
n
,
x
,
y
,
z
,
out
);
inline
void
vec_cross
<
float
,
platform
::
avx2
>
(
const
int
n
,
const
float
*
x
,
const
float
*
y
,
const
float
*
z
,
float
*
out
)
{
vec_cross
<
float
,
platform
::
avx
>
(
n
,
x
,
y
,
z
,
out
);
}
template
<
>
inline
void
vec_cross
<
float
,
platform
::
jit
::
avx512f
>
(
const
int
n
,
const
float
*
x
,
const
float
*
y
,
const
float
*
z
,
float
*
out
)
{
inline
void
vec_cross
<
float
,
platform
::
avx512f
>
(
const
int
n
,
const
float
*
x
,
const
float
*
y
,
const
float
*
z
,
float
*
out
)
{
// TODO(TJ): enable me
vec_cross
<
float
,
platform
::
jit
::
avx
>
(
n
,
x
,
y
,
z
,
out
);
vec_cross
<
float
,
platform
::
avx
>
(
n
,
x
,
y
,
z
,
out
);
}
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
=
platform
::
jit
::
isa_any
>
template
<
typename
T
,
platform
::
cpu_isa_t
isa
=
platform
::
isa_any
>
inline
void
vec_add_bias
(
const
int
n
,
const
T
a
,
const
T
*
x
,
T
*
y
)
{
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
y
[
i
]
=
x
[
i
]
+
a
;
...
...
@@ -254,12 +253,12 @@ inline void vec_add_bias(const int n, const T a, const T* x, T* y) {
}
template
<
>
inline
void
vec_add_bias
<
float
,
platform
::
jit
::
avx
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
inline
void
vec_add_bias
<
float
,
platform
::
avx
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
#ifdef __AVX__
constexpr
int
block
=
YMM_FLOAT_BLOCK
;
if
(
n
<
block
)
{
vec_add_bias
<
float
,
platform
::
jit
::
isa_any
>
(
n
,
a
,
x
,
y
);
vec_add_bias
<
float
,
platform
::
isa_any
>
(
n
,
a
,
x
,
y
);
return
;
}
const
int
rest
=
n
%
block
;
...
...
@@ -283,32 +282,30 @@ inline void vec_add_bias<float, platform::jit::avx>(const int n, const float a,
y
[
i
]
=
x
[
i
]
+
a
;
}
#else
vec_add_bias
<
float
,
platform
::
jit
::
isa_any
>
(
n
,
a
,
x
,
y
);
vec_add_bias
<
float
,
platform
::
isa_any
>
(
n
,
a
,
x
,
y
);
#endif
}
template
<
>
inline
void
vec_add_bias
<
float
,
platform
::
jit
::
avx2
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
vec_add_bias
<
float
,
platform
::
jit
::
avx
>
(
n
,
a
,
x
,
y
);
inline
void
vec_add_bias
<
float
,
platform
::
avx2
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
vec_add_bias
<
float
,
platform
::
avx
>
(
n
,
a
,
x
,
y
);
}
template
<
>
inline
void
vec_add_bias
<
float
,
platform
::
jit
::
avx512f
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
inline
void
vec_add_bias
<
float
,
platform
::
avx512f
>
(
const
int
n
,
const
float
a
,
const
float
*
x
,
float
*
y
)
{
// TODO(TJ): enable me
vec_add_bias
<
float
,
platform
::
jit
::
avx2
>
(
n
,
a
,
x
,
y
);
vec_add_bias
<
float
,
platform
::
avx2
>
(
n
,
a
,
x
,
y
);
}
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
=
platform
::
jit
::
isa_any
>
template
<
typename
T
,
platform
::
cpu_isa_t
isa
=
platform
::
isa_any
>
inline
void
vec_identity
(
const
int
n
,
const
T
*
x
,
T
*
y
)
{
// do nothing
return
;
}
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
=
platform
::
jit
::
isa_any
>
template
<
typename
T
,
platform
::
cpu_isa_t
isa
=
platform
::
isa_any
>
inline
void
vec_sigmoid
(
const
int
n
,
const
T
*
x
,
T
*
y
)
{
const
T
min
=
SIGMOID_THRESHOLD_MIN
;
const
T
max
=
SIGMOID_THRESHOLD_MAX
;
...
...
@@ -323,12 +320,12 @@ inline void vec_sigmoid(const int n, const T* x, T* y) {
}
template
<
>
inline
void
vec_sigmoid
<
float
,
platform
::
jit
::
avx
>
(
const
int
n
,
const
float
*
x
,
float
*
y
)
{
inline
void
vec_sigmoid
<
float
,
platform
::
avx
>
(
const
int
n
,
const
float
*
x
,
float
*
y
)
{
#ifdef __AVX__
constexpr
int
block
=
YMM_FLOAT_BLOCK
;
if
(
n
<
block
)
{
vec_sigmoid
<
float
,
platform
::
jit
::
isa_any
>
(
n
,
x
,
y
);
vec_sigmoid
<
float
,
platform
::
isa_any
>
(
n
,
x
,
y
);
return
;
}
const
int
rest
=
n
%
block
;
...
...
@@ -377,25 +374,24 @@ inline void vec_sigmoid<float, platform::jit::avx>(const int n, const float* x,
y
[
i
]
=
1.
f
/
(
1.
f
+
y
[
i
]);
}
#else
vec_sigmoid
<
float
,
platform
::
jit
::
isa_any
>
(
n
,
x
,
y
);
vec_sigmoid
<
float
,
platform
::
isa_any
>
(
n
,
x
,
y
);
#endif
}
template
<
>
inline
void
vec_sigmoid
<
float
,
platform
::
jit
::
avx2
>
(
const
int
n
,
const
float
*
x
,
float
*
y
)
{
vec_sigmoid
<
float
,
platform
::
jit
::
avx
>
(
n
,
x
,
y
);
inline
void
vec_sigmoid
<
float
,
platform
::
avx2
>
(
const
int
n
,
const
float
*
x
,
float
*
y
)
{
vec_sigmoid
<
float
,
platform
::
avx
>
(
n
,
x
,
y
);
}
template
<
>
inline
void
vec_sigmoid
<
float
,
platform
::
jit
::
avx512f
>
(
const
int
n
,
const
float
*
x
,
float
*
y
)
{
inline
void
vec_sigmoid
<
float
,
platform
::
avx512f
>
(
const
int
n
,
const
float
*
x
,
float
*
y
)
{
// TODO(TJ): enable me
vec_sigmoid
<
float
,
platform
::
jit
::
avx2
>
(
n
,
x
,
y
);
vec_sigmoid
<
float
,
platform
::
avx2
>
(
n
,
x
,
y
);
}
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
=
platform
::
jit
::
isa_any
>
template
<
typename
T
,
platform
::
cpu_isa_t
isa
=
platform
::
isa_any
>
inline
void
vec_tanh
(
const
int
n
,
const
T
*
x
,
T
*
y
)
{
vec_scal
<
T
,
isa
>
(
n
,
static_cast
<
T
>
(
2
),
x
,
y
);
vec_sigmoid
<
T
,
isa
>
(
n
,
y
,
y
);
...
...
@@ -404,7 +400,7 @@ inline void vec_tanh(const int n, const T* x, T* y) {
}
// TODO(TJ): make relu clip
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
=
platform
::
jit
::
isa_any
>
template
<
typename
T
,
platform
::
cpu_isa_t
isa
=
platform
::
isa_any
>
inline
void
vec_relu
(
const
int
n
,
const
T
*
x
,
T
*
y
)
{
for
(
int
i
=
0
;
i
<
n
;
++
i
)
{
y
[
i
]
=
x
[
i
]
>
0
?
x
[
i
]
:
0
;
...
...
@@ -412,12 +408,12 @@ inline void vec_relu(const int n, const T* x, T* y) {
}
template
<
>
inline
void
vec_relu
<
float
,
platform
::
jit
::
avx
>
(
const
int
n
,
const
float
*
x
,
float
*
y
)
{
inline
void
vec_relu
<
float
,
platform
::
avx
>
(
const
int
n
,
const
float
*
x
,
float
*
y
)
{
#ifdef __AVX__
constexpr
int
block
=
YMM_FLOAT_BLOCK
;
if
(
n
<
block
*
4
)
{
vec_relu
<
float
,
platform
::
jit
::
isa_any
>
(
n
,
x
,
y
);
vec_relu
<
float
,
platform
::
isa_any
>
(
n
,
x
,
y
);
return
;
}
...
...
@@ -441,26 +437,26 @@ inline void vec_relu<float, platform::jit::avx>(const int n, const float* x,
#undef MOVE_ONE_STEP
#else
vec_relu
<
float
,
platform
::
jit
::
isa_any
>
(
n
,
x
,
y
);
vec_relu
<
float
,
platform
::
isa_any
>
(
n
,
x
,
y
);
#endif
}
template
<
>
inline
void
vec_relu
<
float
,
platform
::
jit
::
avx2
>
(
const
int
n
,
const
float
*
x
,
float
*
y
)
{
vec_relu
<
float
,
platform
::
jit
::
avx
>
(
n
,
x
,
y
);
inline
void
vec_relu
<
float
,
platform
::
avx2
>
(
const
int
n
,
const
float
*
x
,
float
*
y
)
{
vec_relu
<
float
,
platform
::
avx
>
(
n
,
x
,
y
);
}
template
<
>
inline
void
vec_relu
<
float
,
platform
::
jit
::
avx512f
>
(
const
int
n
,
const
float
*
x
,
float
*
y
)
{
inline
void
vec_relu
<
float
,
platform
::
avx512f
>
(
const
int
n
,
const
float
*
x
,
float
*
y
)
{
// TODO(TJ): enable me
vec_relu
<
float
,
platform
::
jit
::
avx2
>
(
n
,
x
,
y
);
vec_relu
<
float
,
platform
::
avx2
>
(
n
,
x
,
y
);
}
// TODO(TJ): optimize double of sigmoid, tanh and relu if necessary
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
=
platform
::
jit
::
isa_any
>
template
<
typename
T
,
platform
::
cpu_isa_t
isa
=
platform
::
isa_any
>
class
VecActivations
{
public:
std
::
function
<
void
(
const
int
,
const
T
*
,
T
*
)
>
operator
()(
...
...
paddle/fluid/operators/math/cpu_vec_test.cc
浏览文件 @
4a93db92
...
...
@@ -104,38 +104,42 @@ void TestAndBench(const int n, std::function<void(const int, const T*, T*)> tgt,
}
TEST
(
CpuVecTest
,
sigmoid
)
{
namespace
jit
=
paddle
::
platform
::
jit
;
namespace
platform
=
paddle
::
platform
;
using
namespace
paddle
::
operators
::
math
;
// NOLINT
for
(
auto
sz
:
{
1
,
2
,
15
,
16
,
30
,
32
,
128
,
200
,
512
})
{
TestAndBench
<
float
>
(
sz
,
vec_sigmoid
<
float
>
,
ref_sigmoid
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_sigmoid
<
float
,
jit
::
avx
>
,
ref_sigmoid
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_sigmoid
<
float
,
jit
::
avx2
>
,
ref_sigmoid
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_sigmoid
<
float
,
jit
::
avx512f
>
,
TestAndBench
<
float
>
(
sz
,
vec_sigmoid
<
float
,
platform
::
avx
>
,
ref_sigmoid
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_sigmoid
<
float
,
platform
::
avx2
>
,
ref_sigmoid
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_sigmoid
<
float
,
platform
::
avx512f
>
,
ref_sigmoid
<
float
>
);
}
TestAndBench
<
double
>
(
30
,
vec_sigmoid
<
double
>
,
ref_sigmoid
<
double
>
);
}
TEST
(
CpuVecTest
,
tanh
)
{
namespace
jit
=
paddle
::
platform
::
jit
;
namespace
platform
=
paddle
::
platform
;
using
namespace
paddle
::
operators
::
math
;
// NOLINT
for
(
auto
sz
:
{
1
,
2
,
15
,
16
,
30
,
32
,
128
,
200
,
512
})
{
TestAndBench
<
float
>
(
sz
,
vec_tanh
<
float
>
,
ref_tanh
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_tanh
<
float
,
jit
::
avx
>
,
ref_tanh
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_tanh
<
float
,
jit
::
avx2
>
,
ref_tanh
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_tanh
<
float
,
jit
::
avx512f
>
,
ref_tanh
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_tanh
<
float
,
platform
::
avx
>
,
ref_tanh
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_tanh
<
float
,
platform
::
avx2
>
,
ref_tanh
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_tanh
<
float
,
platform
::
avx512f
>
,
ref_tanh
<
float
>
);
}
TestAndBench
<
double
>
(
30
,
vec_tanh
<
double
>
,
ref_tanh
<
double
>
);
}
TEST
(
CpuVecTest
,
relu
)
{
namespace
jit
=
paddle
::
platform
::
jit
;
namespace
platform
=
paddle
::
platform
;
using
namespace
paddle
::
operators
::
math
;
// NOLINT
for
(
auto
sz
:
{
1
,
2
,
15
,
16
,
30
,
32
,
128
,
200
,
512
})
{
TestAndBench
<
float
>
(
sz
,
vec_relu
<
float
>
,
ref_relu
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_relu
<
float
,
jit
::
avx
>
,
ref_relu
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_relu
<
float
,
jit
::
avx2
>
,
ref_relu
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_relu
<
float
,
jit
::
avx512f
>
,
ref_relu
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_relu
<
float
,
platform
::
avx
>
,
ref_relu
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_relu
<
float
,
platform
::
avx2
>
,
ref_relu
<
float
>
);
TestAndBench
<
float
>
(
sz
,
vec_relu
<
float
,
platform
::
avx512f
>
,
ref_relu
<
float
>
);
}
TestAndBench
<
double
>
(
30
,
vec_relu
<
double
>
,
ref_relu
<
double
>
);
}
...
...
@@ -162,38 +166,40 @@ void TestInplace(const int n, std::function<void(const int, const T*, T*)> tgt,
}
TEST
(
CpuVecTest
,
inplace_sigmoid
)
{
namespace
jit
=
paddle
::
platform
::
jit
;
namespace
platform
=
paddle
::
platform
;
using
namespace
paddle
::
operators
::
math
;
// NOLINT
for
(
auto
sz
:
{
1
,
2
,
15
,
16
,
30
,
32
,
128
,
200
,
512
})
{
TestInplace
<
float
>
(
sz
,
vec_sigmoid
<
float
>
,
ref_sigmoid
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_sigmoid
<
float
,
jit
::
avx
>
,
ref_sigmoid
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_sigmoid
<
float
,
jit
::
avx2
>
,
ref_sigmoid
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_sigmoid
<
float
,
jit
::
avx512f
>
,
TestInplace
<
float
>
(
sz
,
vec_sigmoid
<
float
,
platform
::
avx
>
,
ref_sigmoid
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_sigmoid
<
float
,
platform
::
avx2
>
,
ref_sigmoid
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_sigmoid
<
float
,
platform
::
avx512f
>
,
ref_sigmoid
<
float
>
);
}
TestInplace
<
double
>
(
30
,
vec_sigmoid
<
double
>
,
ref_sigmoid
<
double
>
);
}
TEST
(
CpuVecTest
,
inplace_tanh
)
{
namespace
jit
=
paddle
::
platform
::
jit
;
namespace
platform
=
paddle
::
platform
;
using
namespace
paddle
::
operators
::
math
;
// NOLINT
for
(
auto
sz
:
{
1
,
2
,
15
,
16
,
30
,
32
,
128
,
200
,
512
})
{
TestInplace
<
float
>
(
sz
,
vec_tanh
<
float
>
,
ref_tanh
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_tanh
<
float
,
jit
::
avx
>
,
ref_tanh
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_tanh
<
float
,
jit
::
avx2
>
,
ref_tanh
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_tanh
<
float
,
jit
::
avx512f
>
,
ref_tanh
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_tanh
<
float
,
platform
::
avx
>
,
ref_tanh
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_tanh
<
float
,
platform
::
avx2
>
,
ref_tanh
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_tanh
<
float
,
platform
::
avx512f
>
,
ref_tanh
<
float
>
);
}
TestInplace
<
double
>
(
30
,
vec_tanh
<
double
>
,
ref_tanh
<
double
>
);
}
TEST
(
CpuVecTest
,
inplace_relu
)
{
namespace
jit
=
paddle
::
platform
::
jit
;
namespace
platform
=
paddle
::
platform
;
using
namespace
paddle
::
operators
::
math
;
// NOLINT
for
(
auto
sz
:
{
1
,
2
,
15
,
16
,
30
,
32
,
128
,
200
,
512
})
{
TestInplace
<
float
>
(
sz
,
vec_relu
<
float
>
,
ref_relu
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_relu
<
float
,
jit
::
avx
>
,
ref_relu
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_relu
<
float
,
jit
::
avx2
>
,
ref_relu
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_relu
<
float
,
jit
::
avx512f
>
,
ref_relu
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_relu
<
float
,
platform
::
avx
>
,
ref_relu
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_relu
<
float
,
platform
::
avx2
>
,
ref_relu
<
float
>
);
TestInplace
<
float
>
(
sz
,
vec_relu
<
float
,
platform
::
avx512f
>
,
ref_relu
<
float
>
);
}
TestInplace
<
double
>
(
30
,
vec_relu
<
double
>
,
ref_relu
<
double
>
);
}
paddle/fluid/operators/math/jit_code.cc
浏览文件 @
4a93db92
...
...
@@ -22,7 +22,7 @@ namespace math {
namespace
jitkernel
{
namespace
gen
{
using
namespace
platform
::
jit
;
// NOLINT
using
namespace
platform
;
// NOLINT
bool
VXXJitCode
::
init
(
int
d
,
int
scalar_index
)
{
// It's not necessary to use avx512 since it would slow down the frequency
...
...
paddle/fluid/operators/math/jit_code.h
浏览文件 @
4a93db92
...
...
@@ -179,7 +179,7 @@ class VActJitCode : public JitCode {
template
<
typename
JMM
>
void
exp_jmm
(
JMM
&
dst
,
JMM
&
src
,
int
src_idx
=
11
,
int
fx_idx
=
12
,
// NOLINT
int
fy_idx
=
13
,
int
mask_idx
=
14
,
int
tmp_idx
=
15
)
{
using
namespace
platform
::
jit
;
// NOLINT
using
namespace
platform
;
// NOLINT
// check all idx can not equal
JMM
jmm_src
=
JMM
(
src_idx
);
JMM
jmm_fx
=
JMM
(
fx_idx
);
...
...
paddle/fluid/operators/math/jit_gen.cc
浏览文件 @
4a93db92
...
...
@@ -36,7 +36,7 @@ void JitCode::preCode() {
for
(
int
i
=
0
;
i
<
num_g_abi_regs
;
++
i
)
{
push
(
Xbyak
::
Reg64
(
g_abi_regs
[
i
]));
}
if
(
platform
::
jit
::
MayIUse
(
platform
::
jit
::
avx512f
))
{
if
(
platform
::
MayIUse
(
platform
::
avx512f
))
{
mov
(
reg_EVEX_max_8b_offt
,
2
*
EVEX_max_8b_offt
);
}
}
...
...
paddle/fluid/operators/math/jit_kernel.cc
浏览文件 @
4a93db92
...
...
@@ -21,8 +21,6 @@ namespace operators {
namespace
math
{
namespace
jitkernel
{
namespace
jit
=
platform
::
jit
;
KernelPool
&
KernelPool
::
Instance
()
{
static
thread_local
KernelPool
g_jit_kernels
;
return
g_jit_kernels
;
...
...
paddle/fluid/operators/math/jit_kernel_blas.cc
浏览文件 @
4a93db92
...
...
@@ -30,7 +30,6 @@ namespace paddle {
namespace
operators
{
namespace
math
{
namespace
jitkernel
{
namespace
jit
=
platform
::
jit
;
#ifdef PADDLE_WITH_MKLML
template
<
typename
T
>
...
...
@@ -125,7 +124,7 @@ bool VMulKernelImpl<float>::useJIT(int d) {
#ifdef PADDLE_WITH_MKLML
template
<
>
bool
VMulKernelImpl
<
float
>::
useMKL
(
int
d
)
{
return
jit
::
MayIUse
(
jit
::
avx512f
)
&&
d
>
512
;
return
platform
::
MayIUse
(
platform
::
avx512f
)
&&
d
>
512
;
}
template
<
>
...
...
paddle/fluid/operators/math/jit_kernel_crf_decode.cc
浏览文件 @
4a93db92
...
...
@@ -25,10 +25,8 @@ namespace operators {
namespace
math
{
namespace
jitkernel
{
namespace
jit
=
platform
::
jit
;
/* CRF Decode JitKernel */
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
,
jit_block
>
template
<
typename
T
,
platform
::
cpu_isa_t
isa
,
jit_block
>
class
CRFDecodeKernelImpl
:
public
CRFDecodeKernel
<
T
>
{
public:
explicit
CRFDecodeKernelImpl
(
int
tag_num
)
:
CRFDecodeKernel
<
T
>
()
{
...
...
@@ -101,7 +99,7 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
#define INTRIAVX_FLOAT(block) \
template <> \
CRFDecodeKernelImpl<float,
jit::avx, block>::CRFDecodeKernelImpl(
\
CRFDecodeKernelImpl<float,
platform::avx, block>::CRFDecodeKernelImpl(
\
int tag_num) \
: CRFDecodeKernel<float>() { \
this->num_ = tag_num; \
...
...
@@ -109,7 +107,7 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
this->rest_ = this->num_ % YMM_FLOAT_BLOCK; \
} \
template <> \
void CRFDecodeKernelImpl<float,
jit::avx, block>::Compute(
\
void CRFDecodeKernelImpl<float,
platform::avx, block>::Compute(
\
const int seq_len, const float* x, const float* w, float* alpha, \
int* track) const { \
INIT_ALPHA(YMM_FLOAT_BLOCK) \
...
...
@@ -204,7 +202,7 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
#define INTRIAVX512_FLOAT(block) \
template <> \
CRFDecodeKernelImpl<float,
jit::avx512f, block>::CRFDecodeKernelImpl(
\
CRFDecodeKernelImpl<float,
platform::avx512f, block>::CRFDecodeKernelImpl(
\
int tag_num) \
: CRFDecodeKernel<float>() { \
this->num_ = tag_num; \
...
...
@@ -212,7 +210,7 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
this->rest_ = this->num_ % ZMM_FLOAT_BLOCK; \
} \
template <> \
void CRFDecodeKernelImpl<float,
jit::avx512f, block>::Compute(
\
void CRFDecodeKernelImpl<float,
platform::avx512f, block>::Compute(
\
const int seq_len, const float* x, const float* w, float* alpha, \
int* track) const { \
INIT_ALPHA(ZMM_FLOAT_BLOCK) \
...
...
@@ -270,14 +268,14 @@ INTRIAVX_FLOAT(kEQ16);
INTRIAVX_FLOAT
(
kGT16
);
#endif
#ifdef __AVX2__
INTRIAVX2_FLOAT
(
jit
::
avx2
,
kEQ8
);
INTRIAVX2_FLOAT
(
jit
::
avx2
,
kGT8LT16
);
INTRIAVX2_FLOAT
(
jit
::
avx2
,
kEQ16
);
INTRIAVX2_FLOAT
(
jit
::
avx2
,
kGT16
);
INTRIAVX2_FLOAT
(
platform
::
avx2
,
kEQ8
);
INTRIAVX2_FLOAT
(
platform
::
avx2
,
kGT8LT16
);
INTRIAVX2_FLOAT
(
platform
::
avx2
,
kEQ16
);
INTRIAVX2_FLOAT
(
platform
::
avx2
,
kGT16
);
#endif
#ifdef __AVX512F__
INTRIAVX2_FLOAT
(
jit
::
avx512f
,
kEQ8
);
INTRIAVX2_FLOAT
(
jit
::
avx512f
,
kGT8LT16
);
INTRIAVX2_FLOAT
(
platform
::
avx512f
,
kEQ8
);
INTRIAVX2_FLOAT
(
platform
::
avx512f
,
kGT8LT16
);
INTRIAVX512_FLOAT
(
kEQ16
);
INTRIAVX512_FLOAT
(
kGT16
);
#endif
...
...
paddle/fluid/operators/math/jit_kernel_exp.cc
浏览文件 @
4a93db92
...
...
@@ -29,7 +29,6 @@ namespace paddle {
namespace
operators
{
namespace
math
{
namespace
jitkernel
{
namespace
jit
=
platform
::
jit
;
#ifdef PADDLE_WITH_MKLML
// try to use MKL to speedup
...
...
paddle/fluid/operators/math/jit_kernel_layer_norm.cc
浏览文件 @
4a93db92
...
...
@@ -22,10 +22,8 @@ namespace operators {
namespace
math
{
namespace
jitkernel
{
namespace
jit
=
platform
::
jit
;
/* Layer Norm JitKernel */
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
,
jit_block
>
template
<
typename
T
,
platform
::
cpu_isa_t
isa
,
jit_block
>
class
LayerNormKernelImpl
:
public
LayerNormKernel
<
T
>
{
public:
explicit
LayerNormKernelImpl
(
int
right
)
:
LayerNormKernel
<
T
>
()
{
...
...
@@ -90,7 +88,7 @@ class LayerNormKernelImpl : public LayerNormKernel<T> {
this->end_ = this->num_ - this->rest_; \
} \
template <> \
void LayerNormKernelImpl<float,
jit::avx, block>::Compute(
\
void LayerNormKernelImpl<float,
platform::avx, block>::Compute(
\
float* x, float* out, float* mean, float* var, const float* scale, \
const float* bias, int height, const float epsilon) const { \
__m256 sum; \
...
...
@@ -219,16 +217,16 @@ class LayerNormKernelImpl : public LayerNormKernel<T> {
}
#ifdef __AVX__
INTRIAVX_FLOAT
(
jit
::
avx
,
kEQ8
);
INTRIAVX_FLOAT
(
jit
::
avx
,
kGT8LT16
);
INTRIAVX_FLOAT
(
jit
::
avx
,
kEQ16
);
INTRIAVX_FLOAT
(
jit
::
avx
,
kGT16
);
INTRIAVX_FLOAT
(
platform
::
avx
,
kEQ8
);
INTRIAVX_FLOAT
(
platform
::
avx
,
kGT8LT16
);
INTRIAVX_FLOAT
(
platform
::
avx
,
kEQ16
);
INTRIAVX_FLOAT
(
platform
::
avx
,
kGT16
);
#endif
#ifdef __AVX2__
INTRIAVX_FLOAT
(
jit
::
avx2
,
kEQ8
);
INTRIAVX_FLOAT
(
jit
::
avx2
,
kGT8LT16
);
INTRIAVX_FLOAT
(
jit
::
avx2
,
kEQ16
);
INTRIAVX_FLOAT
(
jit
::
avx2
,
kGT16
);
INTRIAVX_FLOAT
(
platform
::
avx2
,
kEQ8
);
INTRIAVX_FLOAT
(
platform
::
avx2
,
kGT8LT16
);
INTRIAVX_FLOAT
(
platform
::
avx2
,
kEQ16
);
INTRIAVX_FLOAT
(
platform
::
avx2
,
kGT16
);
#endif
#undef INTRIAVX_FLOAT
...
...
paddle/fluid/operators/math/jit_kernel_macro.h
浏览文件 @
4a93db92
...
...
@@ -92,7 +92,6 @@ namespace jitkernel {
JITKERNEL_DECLARE, JITKERNEL_FIND_KEY, \
JITKERNEL_IMPL)
namespace
jit
=
platform
::
jit
;
// TODO(TJ): below defines are deprecated, would be remove recently
#define SEARCH_BLOCK(macro_, ker, dtype, isa) \
if (d < YMM_FLOAT_BLOCK) { \
...
...
@@ -107,15 +106,15 @@ namespace jit = platform::jit;
macro_(ker, dtype, isa, kGT16); \
}
#define SEARCH_ISA_BLOCK(macro_, ker, dtype) \
if (
jit::MayIUse(jit::avx512f)) {
\
SEARCH_BLOCK(macro_, ker, dtype,
jit
::avx512f); \
} else if (
jit::MayIUse(jit::avx2)) {
\
SEARCH_BLOCK(macro_, ker, dtype,
jit
::avx2); \
} else if (
jit::MayIUse(jit::avx)) {
\
SEARCH_BLOCK(macro_, ker, dtype,
jit
::avx); \
} else { \
SEARCH_BLOCK(macro_, ker, dtype,
jit
::isa_any); \
#define SEARCH_ISA_BLOCK(macro_, ker, dtype)
\
if (
platform::MayIUse(platform::avx512f)) {
\
SEARCH_BLOCK(macro_, ker, dtype,
platform
::avx512f); \
} else if (
platform::MayIUse(platform::avx2)) {
\
SEARCH_BLOCK(macro_, ker, dtype,
platform
::avx2); \
} else if (
platform::MayIUse(platform::avx)) {
\
SEARCH_BLOCK(macro_, ker, dtype,
platform
::avx); \
} else {
\
SEARCH_BLOCK(macro_, ker, dtype,
platform
::isa_any); \
}
#define JITKERNEL_KEY(ker_key, dtype_key) \
...
...
@@ -156,10 +155,10 @@ namespace jit = platform::jit;
marco_declare, macro_key, macro_impl)
#define FOR_EACH_ISA(macro_, block) \
macro_(
jit::avx512f, block);
\
macro_(
jit::avx2, block);
\
macro_(
jit::avx, block);
\
macro_(
jit
::isa_any, block)
macro_(
platform::avx512f, block);
\
macro_(
platform::avx2, block);
\
macro_(
platform::avx, block);
\
macro_(
platform
::isa_any, block)
#define FOR_EACH_BLOCK(macro_, isa) \
macro_(isa, kLT8); \
...
...
@@ -168,11 +167,11 @@ namespace jit = platform::jit;
macro_(isa, kEQ16); \
macro_(isa, kGT16)
#define FOR_EACH_ISA_BLOCK(macro_) \
FOR_EACH_BLOCK(macro_,
jit
::avx512f); \
FOR_EACH_BLOCK(macro_,
jit
::avx2); \
FOR_EACH_BLOCK(macro_,
jit
::avx); \
FOR_EACH_BLOCK(macro_,
jit
::isa_any)
#define FOR_EACH_ISA_BLOCK(macro_)
\
FOR_EACH_BLOCK(macro_,
platform
::avx512f); \
FOR_EACH_BLOCK(macro_,
platform
::avx2); \
FOR_EACH_BLOCK(macro_,
platform
::avx); \
FOR_EACH_BLOCK(macro_,
platform
::isa_any)
}
// namespace jitkernel
}
// namespace math
...
...
paddle/fluid/operators/math/jit_kernel_test.cc
浏览文件 @
4a93db92
...
...
@@ -705,7 +705,7 @@ TEST(JitKernel, pool) {
jit
::
lstm_attr_t
attr
(
frame_size
,
act_gate
,
act_cand
,
act_cell
,
false
);
// empty call it to avoid unknown flag 'use_pinned_memory' on Mac
paddle
::
platform
::
jit
::
MayIUse
(
paddle
::
platform
::
jit
::
avx
);
paddle
::
platform
::
MayIUse
(
paddle
::
platform
::
avx
);
const
auto
&
plstm1
=
jit
::
KernelPool
::
Instance
()
.
template
Get
<
jit
::
LSTMKernel
<
float
>,
const
jit
::
lstm_attr_t
&>
(
attr
);
...
...
paddle/fluid/platform/cpu_info.cc
浏览文件 @
4a93db92
...
...
@@ -123,7 +123,6 @@ size_t CUDAPinnedMaxChunkSize() {
return
CUDAPinnedMaxAllocSize
()
/
256
;
}
namespace
jit
{
#ifdef PADDLE_WITH_XBYAK
static
Xbyak
::
util
::
Cpu
cpu
;
bool
MayIUse
(
const
cpu_isa_t
cpu_isa
)
{
...
...
@@ -165,6 +164,5 @@ bool MayIUse(const cpu_isa_t cpu_isa) {
}
#endif
}
// namespace jit
}
// namespace platform
}
// namespace paddle
paddle/fluid/platform/cpu_info.h
浏览文件 @
4a93db92
...
...
@@ -39,7 +39,6 @@ size_t CUDAPinnedMinChunkSize();
//! Get the maximum chunk size for buddy allocator.
size_t
CUDAPinnedMaxChunkSize
();
namespace
jit
{
typedef
enum
{
isa_any
,
sse42
,
...
...
@@ -55,7 +54,5 @@ typedef enum {
// May I use some instruction
bool
MayIUse
(
const
cpu_isa_t
cpu_isa
);
}
// namespace jit
}
// namespace platform
}
// namespace paddle
paddle/fluid/platform/init.cc
浏览文件 @
4a93db92
...
...
@@ -116,7 +116,7 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
#endif
#if !defined(_WIN32) && !defined(__APPLE__) && !defined(__OSX__)
if
(
platform
::
jit
::
MayIUse
(
platform
::
jit
::
avx
))
{
if
(
platform
::
MayIUse
(
platform
::
avx
))
{
#ifndef __AVX__
LOG
(
WARNING
)
<<
"AVX is available, Please re-compile on local machine"
;
#endif
...
...
@@ -131,10 +131,10 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
" version or compile from source code."
#ifdef __AVX512F__
if
(
!
platform
::
jit
::
MayIUse
(
platform
::
jit
::
avx512f
))
{
if
(
platform
::
jit
::
MayIUse
(
platform
::
jit
::
avx2
))
{
if
(
!
platform
::
MayIUse
(
platform
::
avx512f
))
{
if
(
platform
::
MayIUse
(
platform
::
avx2
))
{
AVX_GUIDE
(
AVX512
,
AVX2
);
}
else
if
(
platform
::
jit
::
MayIUse
(
platform
::
jit
::
avx
))
{
}
else
if
(
platform
::
MayIUse
(
platform
::
avx
))
{
AVX_GUIDE
(
AVX512
,
AVX
);
}
else
{
AVX_GUIDE
(
AVX512
,
NonAVX
);
...
...
@@ -143,8 +143,8 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
#endif
#ifdef __AVX2__
if
(
!
platform
::
jit
::
MayIUse
(
platform
::
jit
::
avx2
))
{
if
(
platform
::
jit
::
MayIUse
(
platform
::
jit
::
avx
))
{
if
(
!
platform
::
MayIUse
(
platform
::
avx2
))
{
if
(
platform
::
MayIUse
(
platform
::
avx
))
{
AVX_GUIDE
(
AVX2
,
AVX
);
}
else
{
AVX_GUIDE
(
AVX2
,
NonAVX
);
...
...
@@ -153,7 +153,7 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
#endif
#ifdef __AVX__
if
(
!
platform
::
jit
::
MayIUse
(
platform
::
jit
::
avx
))
{
if
(
!
platform
::
MayIUse
(
platform
::
avx
))
{
AVX_GUIDE
(
AVX
,
NonAVX
);
}
#endif
...
...
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