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
64a90b2f
P
Paddle
项目概览
PaddlePaddle
/
Paddle
接近 2 年 前同步成功
通知
2323
Star
20933
Fork
5424
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
提交
64a90b2f
编写于
12月 17, 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
use vadd, vaddrelu, lstm and gru jitkernel
上级
3713d08d
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
68 addition
and
75 deletion
+68
-75
paddle/fluid/operators/fused/fusion_gru_op.cc
paddle/fluid/operators/fused/fusion_gru_op.cc
+30
-28
paddle/fluid/operators/fused/fusion_lstm_op.cc
paddle/fluid/operators/fused/fusion_lstm_op.cc
+32
-30
paddle/fluid/operators/math/CMakeLists.txt
paddle/fluid/operators/math/CMakeLists.txt
+0
-9
paddle/fluid/operators/math/fc_compute.h
paddle/fluid/operators/math/fc_compute.h
+6
-8
未找到文件。
paddle/fluid/operators/fused/fusion_gru_op.cc
浏览文件 @
64a90b2f
...
@@ -15,9 +15,9 @@ limitations under the License. */
...
@@ -15,9 +15,9 @@ limitations under the License. */
#include "paddle/fluid/operators/fused/fusion_gru_op.h"
#include "paddle/fluid/operators/fused/fusion_gru_op.h"
#include <cstring> // for memcpy
#include <cstring> // for memcpy
#include <string>
#include <string>
#include "paddle/fluid/operators/jit/kernels.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/fc_compute.h"
#include "paddle/fluid/operators/math/fc_compute.h"
#include "paddle/fluid/operators/math/jit_kernel.h"
#include "paddle/fluid/operators/math/sequence2batch.h"
#include "paddle/fluid/operators/math/sequence2batch.h"
namespace
paddle
{
namespace
paddle
{
...
@@ -183,27 +183,29 @@ class FusionGRUKernel : public framework::OpKernel<T> {
...
@@ -183,27 +183,29 @@ class FusionGRUKernel : public framework::OpKernel<T> {
const int total_T = x_dims[0]; \
const int total_T = x_dims[0]; \
const int D3 = wh_dims[1]
const int D3 = wh_dims[1]
#define INIT_OTHER_DEFINES \
#define INIT_OTHER_DEFINES \
auto* h0 = ctx.Input<Tensor>("H0"); \
auto* h0 = ctx.Input<Tensor>("H0"); \
auto* wx = ctx.Input<Tensor>("WeightX"); \
auto* wx = ctx.Input<Tensor>("WeightX"); \
auto* bias = ctx.Input<Tensor>("Bias"); \
auto* bias = ctx.Input<Tensor>("Bias"); \
auto* hidden_out = ctx.Output<LoDTensor>("Hidden"); \
auto* hidden_out = ctx.Output<LoDTensor>("Hidden"); \
bool is_reverse = ctx.Attr<bool>("is_reverse"); \
bool is_reverse = ctx.Attr<bool>("is_reverse"); \
const int M = x_dims[1]; \
const int M = x_dims[1]; \
const int D = wh_dims[0]; \
const int D = wh_dims[0]; \
const int D2 = D * 2; \
const int D2 = D * 2; \
const math::jitkernel::gru_attr_t attr( \
const jit::gru_attr_t attr( \
D, ctx.Attr<std::string>("gate_activation"), \
D, jit::to_kerneltype(ctx.Attr<std::string>("gate_activation")), \
ctx.Attr<std::string>("activation")); \
jit::to_kerneltype(ctx.Attr<std::string>("activation"))); \
math::jitkernel::gru_t one_step; \
jit::gru_t one_step; \
const auto& ker = \
auto ComputeH1 = \
math::jitkernel::KernelPool::Instance() \
jit::Get<jit::gruh1, jit::GRUTuples, platform::CPUPlace>(attr); \
.template Get<math::jitkernel::GRUKernel<T>, \
auto ComputeHtPart1 = \
const math::jitkernel::gru_attr_t&>(attr); \
jit::Get<jit::gruhtpart1, jit::GRUTuples, platform::CPUPlace>(attr); \
const T* x_data = x->data<T>(); \
auto ComputeHtPart2 = \
const T* wx_data = wx->data<T>(); \
jit::Get<jit::gruhtpart2, jit::GRUTuples, platform::CPUPlace>(attr); \
const T* wh_data = wh->data<T>(); \
const T* x_data = x->data<T>(); \
auto place = ctx.GetPlace(); \
const T* wx_data = wx->data<T>(); \
const T* wh_data = wh->data<T>(); \
auto place = ctx.GetPlace(); \
T* xx_data = xx->mutable_data<T>(place)
T* xx_data = xx->mutable_data<T>(place)
void
SeqCompute
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
void
SeqCompute
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
...
@@ -242,7 +244,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
...
@@ -242,7 +244,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
}
else
{
}
else
{
one_step
.
gates
=
xx_data
;
one_step
.
gates
=
xx_data
;
one_step
.
ht
=
hidden_out_data
;
one_step
.
ht
=
hidden_out_data
;
ker
->
ComputeH1
(
&
one_step
,
&
attr
);
ComputeH1
(
&
one_step
,
&
attr
);
prev_hidden_data
=
hidden_out_data
;
prev_hidden_data
=
hidden_out_data
;
tstart
=
1
;
tstart
=
1
;
move_step
();
move_step
();
...
@@ -255,12 +257,12 @@ class FusionGRUKernel : public framework::OpKernel<T> {
...
@@ -255,12 +257,12 @@ class FusionGRUKernel : public framework::OpKernel<T> {
one_step
.
gates
=
xx_data
;
one_step
.
gates
=
xx_data
;
one_step
.
ht_1
=
prev_hidden_data
;
one_step
.
ht_1
=
prev_hidden_data
;
one_step
.
ht
=
hidden_out_data
;
one_step
.
ht
=
hidden_out_data
;
ker
->
ComputeHtPart1
(
&
one_step
,
&
attr
);
ComputeHtPart1
(
&
one_step
,
&
attr
);
// gemm rt * Ws
// gemm rt * Ws
blas
.
GEMM
(
CblasNoTrans
,
CblasNoTrans
,
1
,
D
,
D
,
static_cast
<
T
>
(
1
),
blas
.
GEMM
(
CblasNoTrans
,
CblasNoTrans
,
1
,
D
,
D
,
static_cast
<
T
>
(
1
),
hidden_out_data
,
D
,
wh_state_data
,
D
,
static_cast
<
T
>
(
1
),
hidden_out_data
,
D
,
wh_state_data
,
D
,
static_cast
<
T
>
(
1
),
xx_data
+
D2
,
D3
);
xx_data
+
D2
,
D3
);
ker
->
ComputeHtPart2
(
&
one_step
,
&
attr
);
ComputeHtPart2
(
&
one_step
,
&
attr
);
// save prev
// save prev
prev_hidden_data
=
hidden_out_data
;
prev_hidden_data
=
hidden_out_data
;
move_step
();
move_step
();
...
@@ -324,7 +326,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
...
@@ -324,7 +326,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
for
(
int
i
=
0
;
i
<
max_bs
;
++
i
)
{
for
(
int
i
=
0
;
i
<
max_bs
;
++
i
)
{
one_step
.
gates
=
cur_in_data
;
one_step
.
gates
=
cur_in_data
;
one_step
.
ht
=
cur_out_data
;
one_step
.
ht
=
cur_out_data
;
ker
->
ComputeH1
(
&
one_step
,
&
attr
);
ComputeH1
(
&
one_step
,
&
attr
);
// add offset
// add offset
cur_in_data
+=
D3
;
cur_in_data
+=
D3
;
cur_out_data
+=
D
;
cur_out_data
+=
D
;
...
@@ -352,7 +354,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
...
@@ -352,7 +354,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
one_step
.
gates
=
cur_batched_data
;
one_step
.
gates
=
cur_batched_data
;
one_step
.
ht_1
=
cur_prev_hidden_data
;
one_step
.
ht_1
=
cur_prev_hidden_data
;
one_step
.
ht
=
cur_out_data
;
one_step
.
ht
=
cur_out_data
;
ker
->
ComputeHtPart1
(
&
one_step
,
&
attr
);
ComputeHtPart1
(
&
one_step
,
&
attr
);
cur_batched_data
+=
D3
;
cur_batched_data
+=
D3
;
cur_prev_hidden_data
+=
D
;
cur_prev_hidden_data
+=
D
;
...
@@ -370,7 +372,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
...
@@ -370,7 +372,7 @@ class FusionGRUKernel : public framework::OpKernel<T> {
one_step
.
gates
=
cur_batched_data
;
one_step
.
gates
=
cur_batched_data
;
one_step
.
ht_1
=
cur_prev_hidden_data
;
one_step
.
ht_1
=
cur_prev_hidden_data
;
one_step
.
ht
=
cur_out_data
;
one_step
.
ht
=
cur_out_data
;
ker
->
ComputeHtPart2
(
&
one_step
,
&
attr
);
ComputeHtPart2
(
&
one_step
,
&
attr
);
cur_batched_data
+=
D3
;
cur_batched_data
+=
D3
;
cur_prev_hidden_data
+=
D
;
cur_prev_hidden_data
+=
D
;
cur_out_data
+=
D
;
cur_out_data
+=
D
;
...
...
paddle/fluid/operators/fused/fusion_lstm_op.cc
浏览文件 @
64a90b2f
...
@@ -14,9 +14,9 @@ limitations under the License. */
...
@@ -14,9 +14,9 @@ limitations under the License. */
#include "paddle/fluid/operators/fused/fusion_lstm_op.h"
#include "paddle/fluid/operators/fused/fusion_lstm_op.h"
#include <string>
#include <string>
#include "paddle/fluid/operators/jit/kernels.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/fc_compute.h"
#include "paddle/fluid/operators/math/fc_compute.h"
#include "paddle/fluid/operators/math/jit_kernel.h"
#include "paddle/fluid/operators/math/sequence2batch.h"
#include "paddle/fluid/operators/math/sequence2batch.h"
namespace
paddle
{
namespace
paddle
{
...
@@ -236,31 +236,33 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
...
@@ -236,31 +236,33 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
const int D = wh_dims[0]; \
const int D = wh_dims[0]; \
const int D4 = wh_dims[1]
const int D4 = wh_dims[1]
#define INIT_OTHER_DEFINES \
#define INIT_OTHER_DEFINES \
const T* x_data = x->data<T>(); \
const T* x_data = x->data<T>(); \
const T* wx_data = wx->data<T>(); \
const T* wx_data = wx->data<T>(); \
const T* wh_data = wh->data<T>(); \
const T* wh_data = wh->data<T>(); \
/* diagonal weight*/
\
/* diagonal weight*/
\
const T* wp_data = bias->data<T>() + D4; \
const T* wp_data = bias->data<T>() + D4; \
/* for peephole only*/
\
/* for peephole only*/
\
T* checked_cell_data = nullptr; \
T* checked_cell_data = nullptr; \
auto place = ctx.GetPlace(); \
auto place = ctx.GetPlace(); \
if (use_peepholes) { \
if (use_peepholes) { \
/* w_ic * Ct-1, w_fc * Ct-1 ; w_oc * Ct => ih*/
\
/* w_ic * Ct-1, w_fc * Ct-1 ; w_oc * Ct => ih*/
\
auto* checked_cell = ctx.Output<Tensor>("CheckedCell"); \
auto* checked_cell = ctx.Output<Tensor>("CheckedCell"); \
checked_cell_data = checked_cell->mutable_data<T>(place); \
checked_cell_data = checked_cell->mutable_data<T>(place); \
} \
} \
const math::jitkernel::lstm_attr_t attr( \
const jit \
D, ctx.Attr<std::string>("gate_activation"), \
: lstm_attr_t attr( \
ctx.Attr<std::string>("candidate_activation"), \
D, jit::to_kerneltype(ctx.Attr<std::string>("gate_activation")), \
ctx.Attr<std::string>("cell_activation"), use_peepholes); \
jit::to_kerneltype(ctx.Attr<std::string>("candidate_activation")), \
math::jitkernel::lstm_t one_step; \
jit::to_kerneltype(ctx.Attr<std::string>("cell_activation")), \
one_step.wp = wp_data; \
use_peepholes); \
one_step.checked = checked_cell_data; \
math::jitkernel::lstm_t one_step; \
const auto& ker = \
one_step.wp = wp_data; \
math::jitkernel::KernelPool::Instance() \
one_step.checked = checked_cell_data; \
.template Get<math::jitkernel::LSTMKernel<T>, \
auto ComputeC1H1 = \
const math::jitkernel::lstm_attr_t&>(attr)
jit::Get<jit::lstmc1h1, jit::LSTMTuples, platform::CPUPlace>(attr); \
auto ComputeCtHt = \
jit::Get<jit::lstmctht, jit::LSTMTuples, platform::CPUPlace>(attr)
// Wh GEMM
// Wh GEMM
#define GEMM_WH_ADDON(bs, prev, out) \
#define GEMM_WH_ADDON(bs, prev, out) \
...
@@ -306,7 +308,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
...
@@ -306,7 +308,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
one_step
.
gates
=
xx_data
;
one_step
.
gates
=
xx_data
;
one_step
.
ct
=
c_out_data
;
one_step
.
ct
=
c_out_data
;
one_step
.
ht
=
h_out_data
;
one_step
.
ht
=
h_out_data
;
ker
->
ComputeC1H1
(
&
one_step
,
&
attr
);
ComputeC1H1
(
&
one_step
,
&
attr
);
tstart
=
1
;
tstart
=
1
;
// move one step
// move one step
prev_h_data
=
h_out_data
;
prev_h_data
=
h_out_data
;
...
@@ -322,7 +324,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
...
@@ -322,7 +324,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
one_step
.
ct_1
=
prev_c_data
;
one_step
.
ct_1
=
prev_c_data
;
one_step
.
ct
=
c_out_data
;
one_step
.
ct
=
c_out_data
;
one_step
.
ht
=
h_out_data
;
one_step
.
ht
=
h_out_data
;
ker
->
ComputeCtHt
(
&
one_step
,
&
attr
);
ComputeCtHt
(
&
one_step
,
&
attr
);
// move one step
// move one step
prev_h_data
=
h_out_data
;
prev_h_data
=
h_out_data
;
prev_c_data
=
c_out_data
;
prev_c_data
=
c_out_data
;
...
@@ -402,7 +404,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
...
@@ -402,7 +404,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
one_step
.
gates
=
cur_in_data
;
one_step
.
gates
=
cur_in_data
;
one_step
.
ct
=
cur_c_out_data
;
one_step
.
ct
=
cur_c_out_data
;
one_step
.
ht
=
cur_h_out_data
;
one_step
.
ht
=
cur_h_out_data
;
ker
->
ComputeC1H1
(
&
one_step
,
&
attr
);
ComputeC1H1
(
&
one_step
,
&
attr
);
cur_in_data
+=
D4
;
cur_in_data
+=
D4
;
cur_c_out_data
+=
D
;
cur_c_out_data
+=
D
;
...
@@ -432,7 +434,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
...
@@ -432,7 +434,7 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
one_step
.
ct_1
=
cur_prev_c_data
;
one_step
.
ct_1
=
cur_prev_c_data
;
one_step
.
ct
=
cur_c_out_data
;
one_step
.
ct
=
cur_c_out_data
;
one_step
.
ht
=
cur_h_out_data
;
one_step
.
ht
=
cur_h_out_data
;
ker
->
ComputeCtHt
(
&
one_step
,
&
attr
);
ComputeC1H1
(
&
one_step
,
&
attr
);
// move one batch
// move one batch
cur_in_data
+=
D4
;
cur_in_data
+=
D4
;
...
...
paddle/fluid/operators/math/CMakeLists.txt
浏览文件 @
64a90b2f
...
@@ -73,12 +73,3 @@ if(WITH_GPU)
...
@@ -73,12 +73,3 @@ if(WITH_GPU)
endif
()
endif
()
cc_test
(
concat_test SRCS concat_test.cc DEPS concat_and_split
)
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_test
(
cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info
)
# set(JIT_KERNEL_SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_rnn.cc jit_kernel_crf_decode.cc jit_kernel_layer_norm.cc)
# set(JIT_KERNEL_DEPS cpu_info cblas gflags enforce)
# if(WITH_XBYAK)
# list(APPEND JIT_KERNEL_SRCS jit_gen.cc jit_code.cc)
# list(APPEND JIT_KERNEL_DEPS xbyak)
# endif()
# cc_library(jit_kernel SRCS ${JIT_KERNEL_SRCS} DEPS ${JIT_KERNEL_DEPS})
# cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel)
paddle/fluid/operators/math/fc_compute.h
浏览文件 @
64a90b2f
...
@@ -14,8 +14,8 @@ limitations under the License. */
...
@@ -14,8 +14,8 @@ limitations under the License. */
#pragma once
#pragma once
#include "paddle/fluid/operators/jit/kernels.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/jit_kernel.h"
namespace
paddle
{
namespace
paddle
{
namespace
operators
{
namespace
operators
{
...
@@ -30,22 +30,20 @@ inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
...
@@ -30,22 +30,20 @@ inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
return
;
return
;
}
}
if
(
relu
)
{
if
(
relu
)
{
const
auto
&
vaddrelu
=
jitkernel
::
KernelPool
::
Instance
()
auto
compute
=
.
template
Get
<
jitkernel
::
VAddReluKernel
<
T
>
>
(
N
);
jit
::
Get
<
jit
::
vaddrelu
,
jit
::
XYZNTuples
,
platform
::
CPUPlcace
>
(
N
);
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
T
*
dst
=
Y
+
i
*
N
;
T
*
dst
=
Y
+
i
*
N
;
vaddrelu
->
C
ompute
(
B
,
dst
,
dst
,
N
);
c
ompute
(
B
,
dst
,
dst
,
N
);
}
}
}
else
{
}
else
{
const
auto
&
vadd
=
jitkernel
::
KernelPool
::
Instance
()
auto
compute
=
jit
::
Get
<
jit
::
vadd
,
jit
::
XYZNTuples
,
platform
::
CPUPlcace
>
(
N
);
.
template
Get
<
jitkernel
::
VAddKernel
<
T
>
>
(
N
);
#ifdef PADDLE_WITH_MKLML
#ifdef PADDLE_WITH_MKLML
#pragma omp parallel for
#pragma omp parallel for
#endif
#endif
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
for
(
int
i
=
0
;
i
<
M
;
i
++
)
{
T
*
dst
=
Y
+
i
*
N
;
T
*
dst
=
Y
+
i
*
N
;
vadd
->
C
ompute
(
B
,
dst
,
dst
,
N
);
c
ompute
(
B
,
dst
,
dst
,
N
);
}
}
}
}
}
}
...
...
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