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
3c957af1
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
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 >>
未验证
提交
3c957af1
编写于
10月 30, 2018
作者:
T
tensor-tang
提交者:
GitHub
10月 30, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #14080 from tensor-tang/refine/jit/crf2
Refine/jit/crf decoding
上级
b3b32925
64d5b438
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
310 addition
and
219 deletion
+310
-219
paddle/fluid/operators/CMakeLists.txt
paddle/fluid/operators/CMakeLists.txt
+1
-0
paddle/fluid/operators/crf_decoding_op.h
paddle/fluid/operators/crf_decoding_op.h
+5
-218
paddle/fluid/operators/math/CMakeLists.txt
paddle/fluid/operators/math/CMakeLists.txt
+1
-1
paddle/fluid/operators/math/jit_kernel.h
paddle/fluid/operators/math/jit_kernel.h
+7
-0
paddle/fluid/operators/math/jit_kernel_crf_decode.cc
paddle/fluid/operators/math/jit_kernel_crf_decode.cc
+296
-0
未找到文件。
paddle/fluid/operators/CMakeLists.txt
浏览文件 @
3c957af1
...
...
@@ -301,6 +301,7 @@ op_library(flatten_op DEPS reshape_op)
op_library
(
sequence_pad_op DEPS sequence_padding
)
op_library
(
unstack_op DEPS stack_op
)
op_library
(
fake_quantize_op DEPS memory
)
op_library
(
crf_decoding_op DEPS jit_kernel
)
op_library
(
fusion_lstm_op DEPS jit_kernel
)
if
(
WITH_GPU
)
op_library
(
conv_op DEPS vol2col depthwise_conv im2col
)
...
...
paddle/fluid/operators/crf_decoding_op.h
浏览文件 @
3c957af1
...
...
@@ -16,6 +16,7 @@ limitations under the License. */
#include <limits>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/jit_kernel.h"
#include "paddle/fluid/operators/math/math_function.h"
namespace
paddle
{
...
...
@@ -69,9 +70,6 @@ class CRFDecodingOpKernel : public framework::OpKernel<T> {
auto
emission_dims
=
emission_weights
.
dims
();
const
size_t
seq_len
=
emission_dims
[
0
];
const
size_t
tag_num
=
emission_dims
[
1
];
const
size_t
state_trans_base_idx
=
2
;
const
T
*
x
=
emission_weights
.
data
<
T
>
();
const
T
*
w
=
transition_weights
.
data
<
T
>
();
int64_t
*
path
=
decoded_path
->
data
<
int64_t
>
();
...
...
@@ -84,221 +82,10 @@ class CRFDecodingOpKernel : public framework::OpKernel<T> {
Tensor
track
;
int
*
track_value
=
track
.
mutable_data
<
int
>
(
emission_dims
,
platform
::
CPUPlace
());
#ifdef __AVX__
// It use the AVX or AVX512 instruction to deal the data as the vector of 8 or
// 16 elements per iteration. Then it can implement the parallel processing.
// Only optimize for float type.
#ifdef __AVX512F__
size_t
step_size
=
16
;
#else
size_t
step_size
=
8
;
#endif
if
(
std
::
is_same
<
T
,
float
>::
value
&&
(
tag_num
>=
step_size
))
{
size_t
steps
=
tag_num
/
step_size
;
size_t
remain
=
tag_num
%
step_size
;
int
last_offset
=
static_cast
<
int
>
(
remain
)
-
static_cast
<
int
>
(
step_size
);
// Setup the alpha initial value.
size_t
i_offset
=
0
;
for
(
size_t
i
=
0
;
i
<=
steps
;
++
i
)
{
#ifdef __AVX512F__
// Declare the variable for the content of weights, input and alpha
// values.
__m512
w_content
,
x_content
,
alpha_content
;
// Load the relevant data into the variables from un-aligned address.
w_content
=
_mm512_loadu_ps
((
const
float
*
)(
w
+
i_offset
));
x_content
=
_mm512_loadu_ps
((
const
float
*
)(
x
+
i_offset
));
alpha_content
=
_mm512_add_ps
(
w_content
,
x_content
);
// Save the alpha value.
_mm512_storeu_ps
(
reinterpret_cast
<
float
*>
(
alpha_value
+
i_offset
),
alpha_content
);
#else
// Declare the variable for the content of weights, input and alpha
// values.
__m256
w_content
,
x_content
,
alpha_content
;
// Load the relevant data into the variables from un-aligned address.
w_content
=
_mm256_loadu_ps
((
const
float
*
)(
w
+
i_offset
));
x_content
=
_mm256_loadu_ps
((
const
float
*
)(
x
+
i_offset
));
alpha_content
=
_mm256_add_ps
(
w_content
,
x_content
);
// Save the alpha value.
_mm256_storeu_ps
(
reinterpret_cast
<
float
*>
(
alpha_value
+
i_offset
),
alpha_content
);
#endif
i_offset
+=
step_size
;
if
(
i
==
steps
-
1
)
{
if
(
remain
>
0
)
{
i_offset
+=
last_offset
;
}
else
{
break
;
}
}
}
// Use the column-major strategy to get the location of maximum score.
size_t
seq_offset
=
0
;
for
(
size_t
k
=
1
;
k
<
seq_len
;
++
k
)
{
size_t
j_offset
=
0
;
for
(
size_t
j
=
0
;
j
<=
steps
;
++
j
)
{
#ifdef __AVX512F__
// Initialize the variables of maximum score and location.
__m512
max_score
=
_mm512_set1_ps
(
-
std
::
numeric_limits
<
T
>::
max
());
__m512i
max_j
=
_mm512_setzero_si512
();
#else
// Initialize the variables of maximum score and location.
__m256
max_score
=
_mm256_set1_ps
(
-
std
::
numeric_limits
<
T
>::
max
());
__m256i
max_j
=
_mm256_set1_epi32
(
0
);
#endif
// Calculate the offset of transition_weights.
size_t
trans_offset
=
state_trans_base_idx
*
tag_num
+
j_offset
;
for
(
size_t
i
=
0
;
i
<
tag_num
;
++
i
)
{
#ifdef __AVX512F__
// Initalize the content of alpha variable with related offset.
__m512
alpha_content
=
_mm512_set1_ps
(
*
(
const
float
*
)(
alpha_value
+
seq_offset
+
i
));
// Obtain the content of weights from un-aligned address.
__m512
w_content
=
_mm512_loadu_ps
((
const
float
*
)(
w
+
trans_offset
));
__m512
score_v
=
_mm512_add_ps
(
alpha_content
,
w_content
);
__mmask16
mask
=
_mm512_cmp_ps_mask
(
score_v
,
max_score
,
_CMP_GT_OS
);
// According to the mask value, it update the index of the max_score
// location.
max_j
=
_mm512_mask_set1_epi32
(
max_j
,
mask
,
i
);
// Update the max_score value.
max_score
=
_mm512_max_ps
(
max_score
,
score_v
);
#else
// Initalize the content of alpha variable with related offset.
__m256
alpha_content
=
_mm256_broadcast_ss
(
(
const
float
*
)(
alpha_value
+
seq_offset
+
i
));
// Obtain the content of weights from un-aligned address.
__m256
w_content
=
_mm256_loadu_ps
((
const
float
*
)(
w
+
trans_offset
));
__m256
score_v
=
_mm256_add_ps
(
alpha_content
,
w_content
);
__m256
mask
=
_mm256_cmp_ps
(
score_v
,
max_score
,
_CMP_GT_OS
);
#ifdef __AVX2__
// According to the mask value, it update the index of the max_score
// location.
max_j
=
_mm256_or_si256
(
_mm256_andnot_si256
((
__m256i
)
mask
,
max_j
),
_mm256_and_si256
((
__m256i
)
mask
,
_mm256_set1_epi32
(
i
)));
#else
__m128i
lo_max_j
=
_mm256_extractf128_si256
(
max_j
,
0
);
__m128i
hi_max_j
=
_mm256_extractf128_si256
(
max_j
,
1
);
__m128i
lo_mask
=
_mm256_extractf128_si256
((
__m256i
)
mask
,
0
);
__m128i
hi_mask
=
_mm256_extractf128_si256
((
__m256i
)
mask
,
1
);
lo_max_j
=
_mm_andnot_si128
(
lo_mask
,
lo_max_j
);
hi_max_j
=
_mm_andnot_si128
(
hi_mask
,
hi_max_j
);
lo_mask
=
_mm_and_si128
(
lo_mask
,
_mm_set1_epi32
(
i
));
hi_mask
=
_mm_and_si128
(
hi_mask
,
_mm_set1_epi32
(
i
));
lo_max_j
=
_mm_or_si128
(
lo_mask
,
lo_max_j
);
hi_max_j
=
_mm_or_si128
(
hi_mask
,
hi_max_j
);
// According to the mask value, it update the index of the max_score
// location.
max_j
=
_mm256_insertf128_si256
(
max_j
,
lo_max_j
,
0
);
max_j
=
_mm256_insertf128_si256
(
max_j
,
hi_max_j
,
1
);
#endif
// Update the max_score value.
max_score
=
_mm256_max_ps
(
max_score
,
score_v
);
#endif
trans_offset
+=
tag_num
;
}
#ifdef __AVX512F__
// Update the alpha and track values.
__m512
x_content
=
_mm512_loadu_ps
(
(
const
float
*
)(
x
+
seq_offset
+
tag_num
+
j_offset
));
max_score
=
_mm512_add_ps
(
max_score
,
x_content
);
_mm512_storeu_ps
(
reinterpret_cast
<
float
*>
(
alpha_value
+
seq_offset
+
tag_num
+
j_offset
),
max_score
);
_mm512_storeu_si512
(
reinterpret_cast
<
__m512i
*>
(
track_value
+
seq_offset
+
tag_num
+
j_offset
),
max_j
);
#else
// Update the alpha and track values.
__m256
x_content
=
_mm256_loadu_ps
(
(
const
float
*
)(
x
+
seq_offset
+
tag_num
+
j_offset
));
max_score
=
_mm256_add_ps
(
max_score
,
x_content
);
_mm256_storeu_ps
(
reinterpret_cast
<
float
*>
(
alpha_value
+
seq_offset
+
tag_num
+
j_offset
),
max_score
);
_mm256_storeu_si256
(
reinterpret_cast
<
__m256i
*>
(
track_value
+
seq_offset
+
tag_num
+
j_offset
),
max_j
);
#endif
// Calculate the offset of next step
j_offset
+=
step_size
;
if
(
j
==
steps
-
1
)
{
if
(
remain
>
0
)
{
j_offset
+=
last_offset
;
}
else
{
break
;
}
}
}
seq_offset
+=
tag_num
;
}
}
else
{
for
(
size_t
i
=
0
;
i
<
tag_num
;
++
i
)
alpha_value
[
i
]
=
w
[
i
]
+
x
[
i
];
for
(
size_t
k
=
1
;
k
<
seq_len
;
++
k
)
{
for
(
size_t
i
=
0
;
i
<
tag_num
;
++
i
)
{
T
max_score
=
-
std
::
numeric_limits
<
T
>::
max
();
int
max_j
=
0
;
for
(
size_t
j
=
0
;
j
<
tag_num
;
++
j
)
{
T
score
=
alpha_value
[(
k
-
1
)
*
tag_num
+
j
]
+
w
[(
j
+
state_trans_base_idx
)
*
tag_num
+
i
];
if
(
score
>
max_score
)
{
max_score
=
score
;
max_j
=
j
;
}
}
alpha_value
[
k
*
tag_num
+
i
]
=
max_score
+
x
[
k
*
tag_num
+
i
];
track_value
[
k
*
tag_num
+
i
]
=
max_j
;
}
}
}
#else
for
(
size_t
i
=
0
;
i
<
tag_num
;
++
i
)
alpha_value
[
i
]
=
w
[
i
]
+
x
[
i
];
for
(
size_t
k
=
1
;
k
<
seq_len
;
++
k
)
{
for
(
size_t
i
=
0
;
i
<
tag_num
;
++
i
)
{
T
max_score
=
-
std
::
numeric_limits
<
T
>::
max
();
int
max_j
=
0
;
for
(
size_t
j
=
0
;
j
<
tag_num
;
++
j
)
{
T
score
=
alpha_value
[(
k
-
1
)
*
tag_num
+
j
]
+
w
[(
j
+
state_trans_base_idx
)
*
tag_num
+
i
];
if
(
score
>
max_score
)
{
max_score
=
score
;
max_j
=
j
;
}
}
alpha_value
[
k
*
tag_num
+
i
]
=
max_score
+
x
[
k
*
tag_num
+
i
];
track_value
[
k
*
tag_num
+
i
]
=
max_j
;
}
}
#endif
const
auto
&
ker
=
math
::
jitkernel
::
KernelPool
::
Instance
()
.
template
Get
<
math
::
jitkernel
::
CRFDecodeKernel
<
T
>
>
(
static_cast
<
int
>
(
tag_num
));
ker
->
Compute
(
static_cast
<
int
>
(
seq_len
),
x
,
w
,
alpha_value
,
track_value
);
T
max_score
=
-
std
::
numeric_limits
<
T
>::
max
();
int
max_i
=
0
;
for
(
size_t
i
=
0
;
i
<
tag_num
;
++
i
)
{
...
...
paddle/fluid/operators/math/CMakeLists.txt
浏览文件 @
3c957af1
...
...
@@ -76,6 +76,6 @@ endif()
cc_test
(
concat_test SRCS concat_test.cc DEPS concat_and_split
)
cc_test
(
cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info
)
cc_library
(
jit_kernel
SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_rnn.cc
SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_rnn.cc
jit_kernel_crf_decode.cc
DEPS cpu_info cblas
)
cc_test
(
jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel
)
paddle/fluid/operators/math/jit_kernel.h
浏览文件 @
3c957af1
...
...
@@ -151,6 +151,13 @@ class GRUKernel : public Kernel {
virtual
void
ComputeHtPart2
(
T
*
gates
,
const
T
*
ht_1
,
T
*
ht
)
const
=
0
;
};
template
<
typename
T
>
class
CRFDecodeKernel
:
public
Kernel
{
public:
virtual
void
Compute
(
const
int
seq_len
,
const
T
*
x
,
const
T
*
w
,
T
*
alpha
,
int
*
track
)
const
=
0
;
};
}
// namespace jitkernel
}
// namespace math
}
// namespace operators
...
...
paddle/fluid/operators/math/jit_kernel_crf_decode.cc
0 → 100644
浏览文件 @
3c957af1
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/math/jit_kernel.h"
#include <limits>
#include <string>
#include "paddle/fluid/operators/math/jit_kernel_macro.h"
#ifdef __AVX__
#include <immintrin.h>
#endif
namespace
paddle
{
namespace
operators
{
namespace
math
{
namespace
jitkernel
{
namespace
jit
=
platform
::
jit
;
/* CRF Decode JitKernel */
template
<
typename
T
,
platform
::
jit
::
cpu_isa_t
isa
,
jit_block
>
class
CRFDecodeKernelImpl
:
public
CRFDecodeKernel
<
T
>
{
public:
explicit
CRFDecodeKernelImpl
(
int
tag_num
)
:
CRFDecodeKernel
<
T
>
()
{
this
->
num_
=
tag_num
;
}
void
Compute
(
const
int
seq_len
,
const
T
*
x
,
const
T
*
w
,
T
*
alpha
,
int
*
track
)
const
override
{
constexpr
int
state_trans_base_idx
=
2
;
for
(
int
i
=
0
;
i
<
this
->
num_
;
++
i
)
{
alpha
[
i
]
=
w
[
i
]
+
x
[
i
];
}
for
(
int
k
=
1
;
k
<
seq_len
;
++
k
)
{
for
(
int
i
=
0
;
i
<
this
->
num_
;
++
i
)
{
T
max_score
=
-
std
::
numeric_limits
<
T
>::
max
();
int
max_j
=
0
;
for
(
int
j
=
0
;
j
<
this
->
num_
;
++
j
)
{
T
score
=
alpha
[(
k
-
1
)
*
this
->
num_
+
j
]
+
w
[(
j
+
state_trans_base_idx
)
*
this
->
num_
+
i
];
if
(
score
>
max_score
)
{
max_score
=
score
;
max_j
=
j
;
}
}
alpha
[
k
*
this
->
num_
+
i
]
=
max_score
+
x
[
k
*
this
->
num_
+
i
];
track
[
k
*
this
->
num_
+
i
]
=
max_j
;
}
}
}
};
#define INIT_ALPHA(step_size) \
/* Setup the alpha initial value.*/
\
int i_offset = 0; \
int last_offset = this->rest_ - step_size; \
for (int i = 0; i <= this->end_; ++i) { \
/* weights, input and alpha values. */
\
__m256 w_content, x_content, alpha_content; \
/* Load the relevant data into the variables from un-aligned address.*/
\
w_content = _mm256_loadu_ps(w + i_offset); \
x_content = _mm256_loadu_ps(x + i_offset); \
alpha_content = _mm256_add_ps(w_content, x_content); \
_mm256_storeu_ps(alpha + i_offset, alpha_content); \
i_offset += step_size; \
if (i == this->end_ - 1) { \
if (this->rest_ > 0) { \
i_offset += last_offset; \
} else { \
break; \
} \
} \
}
#define UPDATE_ALPHA(step_size) \
/* Update the alpha and track values. */
\
__m256 x_content = _mm256_loadu_ps(x + seq_offset + this->num_ + j_offset); \
max_score = _mm256_add_ps(max_score, x_content); \
_mm256_storeu_ps(alpha + seq_offset + this->num_ + j_offset, max_score); \
_mm256_storeu_si256( \
reinterpret_cast<__m256i*>(track + seq_offset + this->num_ + j_offset), \
max_j); \
/* Calculate the offset of next step*/
\
j_offset += step_size; \
if (j == this->end_ - 1) { \
if (this->rest_ > 0) { \
j_offset += last_offset; \
} else { \
break; \
} \
}
#define INTRIAVX_FLOAT(block) \
template <> \
CRFDecodeKernelImpl<float, jit::avx, block>::CRFDecodeKernelImpl( \
int tag_num) \
: CRFDecodeKernel<float>() { \
this->num_ = tag_num; \
this->end_ = this->num_ / AVX_FLOAT_BLOCK; \
this->rest_ = this->num_ % AVX_FLOAT_BLOCK; \
} \
template <> \
void CRFDecodeKernelImpl<float, jit::avx, block>::Compute( \
const int seq_len, const float* x, const float* w, float* alpha, \
int* track) const { \
INIT_ALPHA(AVX_FLOAT_BLOCK) \
/* Use the column-major strategy to get the location of maximum score.*/
\
int seq_offset = 0; \
constexpr int state_trans_base_idx = 2; \
for (int k = 1; k < seq_len; ++k) { \
int j_offset = 0; \
for (int j = 0; j <= this->end_; ++j) { \
/* Initialize the variables of maximum score and location.*/
\
__m256 max_score = _mm256_set1_ps(-std::numeric_limits<float>::max()); \
__m256i max_j = _mm256_set1_epi32(0); \
/* Calculate the offset of transition_weights.*/
\
int trans_offset = state_trans_base_idx * this->num_ + j_offset; \
for (int i = 0; i < this->num_; ++i) { \
/* Initalize the content of alpha variable with related offset.*/
\
__m256 alpha_content = _mm256_broadcast_ss(alpha + seq_offset + i); \
/* Obtain the content of weights from un-aligned address.*/
\
__m256 w_content = _mm256_loadu_ps(w + trans_offset); \
__m256 score_v = _mm256_add_ps(alpha_content, w_content); \
__m256 mask = _mm256_cmp_ps(score_v, max_score, _CMP_GT_OS); \
/* According to the mask value, update the index of the max_score.*/
\
/* AVX instructions.*/
\
__m128i lo_max_j = _mm256_extractf128_si256(max_j, 0); \
__m128i hi_max_j = _mm256_extractf128_si256(max_j, 1); \
__m128i lo_mask = _mm256_extractf128_si256((__m256i)mask, 0); \
__m128i hi_mask = _mm256_extractf128_si256((__m256i)mask, 1); \
lo_max_j = _mm_andnot_si128(lo_mask, lo_max_j); \
hi_max_j = _mm_andnot_si128(hi_mask, hi_max_j); \
lo_mask = _mm_and_si128(lo_mask, _mm_set1_epi32(i)); \
hi_mask = _mm_and_si128(hi_mask, _mm_set1_epi32(i)); \
lo_max_j = _mm_or_si128(lo_mask, lo_max_j); \
hi_max_j = _mm_or_si128(hi_mask, hi_max_j); \
max_j = _mm256_insertf128_si256(max_j, lo_max_j, 0); \
max_j = _mm256_insertf128_si256(max_j, hi_max_j, 1); \
/* AVX done*/
\
/* Update the max_score value.*/
\
max_score = _mm256_max_ps(max_score, score_v); \
trans_offset += this->num_; \
} \
UPDATE_ALPHA(AVX_FLOAT_BLOCK) \
} \
seq_offset += this->num_; \
} \
}
#define INTRIAVX2_FLOAT(isa, block) \
template <> \
CRFDecodeKernelImpl<float, isa, block>::CRFDecodeKernelImpl(int tag_num) \
: CRFDecodeKernel<float>() { \
this->num_ = tag_num; \
this->end_ = this->num_ / AVX2_FLOAT_BLOCK; \
this->rest_ = this->num_ % AVX2_FLOAT_BLOCK; \
} \
template <> \
void CRFDecodeKernelImpl<float, isa, block>::Compute( \
const int seq_len, const float* x, const float* w, float* alpha, \
int* track) const { \
INIT_ALPHA(AVX2_FLOAT_BLOCK) \
/* Use the column-major strategy to get the location of maximum score.*/
\
int seq_offset = 0; \
constexpr int state_trans_base_idx = 2; \
for (int k = 1; k < seq_len; ++k) { \
int j_offset = 0; \
for (int j = 0; j <= this->end_; ++j) { \
/* Initialize the variables of maximum score and location.*/
\
__m256 max_score = _mm256_set1_ps(-std::numeric_limits<float>::max()); \
__m256i max_j = _mm256_set1_epi32(0); \
/* Calculate the offset of transition_weights.*/
\
int trans_offset = state_trans_base_idx * this->num_ + j_offset; \
for (int i = 0; i < this->num_; ++i) { \
/* Initalize the content of alpha variable with related offset.*/
\
__m256 alpha_content = _mm256_broadcast_ss(alpha + seq_offset + i); \
/* Obtain the content of weights from un-aligned address.*/
\
__m256 w_content = _mm256_loadu_ps(w + trans_offset); \
__m256 score_v = _mm256_add_ps(alpha_content, w_content); \
__m256 mask = _mm256_cmp_ps(score_v, max_score, _CMP_GT_OS); \
/* According to the mask value, update the index of the max_score.*/
\
/* AVX2 instructions.*/
\
max_j = _mm256_or_si256( \
_mm256_andnot_si256((__m256i)mask, max_j), \
_mm256_and_si256((__m256i)mask, _mm256_set1_epi32(i))); \
/* Update the max_score value.*/
\
max_score = _mm256_max_ps(max_score, score_v); \
trans_offset += this->num_; \
} \
UPDATE_ALPHA(AVX2_FLOAT_BLOCK) \
} \
seq_offset += this->num_; \
} \
}
#define INTRIAVX512_FLOAT(block) \
template <> \
CRFDecodeKernelImpl<float, jit::avx512f, block>::CRFDecodeKernelImpl( \
int tag_num) \
: CRFDecodeKernel<float>() { \
this->num_ = tag_num; \
this->end_ = this->num_ / AVX512_FLOAT_BLOCK; \
this->rest_ = this->num_ % AVX512_FLOAT_BLOCK; \
} \
template <> \
void CRFDecodeKernelImpl<float, jit::avx512f, block>::Compute( \
const int seq_len, const float* x, const float* w, float* alpha, \
int* track) const { \
INIT_ALPHA(AVX512_FLOAT_BLOCK) \
/* Use the column-major strategy to get the location of maximum score.*/
\
int seq_offset = 0; \
constexpr int state_trans_base_idx = 2; \
for (int k = 1; k < seq_len; ++k) { \
int j_offset = 0; \
for (int j = 0; j <= this->end_; ++j) { \
/* Initialize the variables of maximum score and location.*/
\
__m512 max_score = _mm512_set1_ps(-std::numeric_limits<float>::max()); \
__m512i max_j = _mm512_setzero_si512(); \
/* Calculate the offset of transition_weights.*/
\
int trans_offset = state_trans_base_idx * this->num_ + j_offset; \
for (int i = 0; i < this->num_; ++i) { \
/* Initalize the content of alpha variable with related offset.*/
\
__m512 alpha_content = _mm512_set1_ps(*(alpha + seq_offset + i)); \
/* Obtain the content of weights from un-aligned address.*/
\
__m512 w_content = _mm512_loadu_ps(w + trans_offset); \
__m512 score_v = _mm512_add_ps(alpha_content, w_content); \
__mmask16 mask = _mm512_cmp_ps_mask(score_v, max_score, _CMP_GT_OS); \
/* AVX512 instructions.*/
\
max_j = _mm512_mask_set1_epi32(max_j, mask, i); \
/* Update the max_score value.*/
\
max_score = _mm512_max_ps(max_score, score_v); \
trans_offset += this->num_; \
} \
/* Update the alpha and track values.*/
\
__m512 x_content = \
_mm512_loadu_ps(x + seq_offset + this->num_ + j_offset); \
max_score = _mm512_add_ps(max_score, x_content); \
_mm512_storeu_ps(alpha + seq_offset + this->num_ + j_offset, \
max_score); \
_mm512_storeu_si512(reinterpret_cast<__m512i*>(track + seq_offset + \
this->num_ + j_offset), \
max_j); \
/* Calculate the offset of next step*/
\
j_offset += AVX512_FLOAT_BLOCK; \
if (j == this->end_ - 1) { \
if (this->rest_ > 0) { \
j_offset += last_offset; \
} else { \
break; \
} \
} \
} \
seq_offset += this->num_; \
} \
}
#ifdef __AVX__
INTRIAVX_FLOAT
(
kEQ8
);
INTRIAVX_FLOAT
(
kGT8LT16
);
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
);
#endif
#ifdef __AVX512F__
INTRIAVX2_FLOAT
(
jit
::
avx512f
,
kEQ8
);
INTRIAVX2_FLOAT
(
jit
::
avx512f
,
kGT8LT16
);
INTRIAVX512_FLOAT
(
kEQ16
);
INTRIAVX512_FLOAT
(
kGT16
);
#endif
#undef INTRIAVX512_FLOAT
#undef INTRIAVX2_FLOAT
#undef INTRIAVX_FLOAT
#undef INIT_ALPHA
#undef UPDATE_ALPHA
REGISTER_JITKERNEL
(
crf_decode
,
CRFDecodeKernel
);
}
// namespace jitkernel
}
// namespace math
}
// namespace operators
}
// namespace paddle
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