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
d00bd9eb
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2299
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
提交
d00bd9eb
编写于
4月 06, 2018
作者:
K
Kexin Zhao
提交者:
Yi Wang
4月 06, 2018
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Update the cuda API and enable tensor core for GEMM (#9622)
* change from hgemm to gemmEx * fix cpplint
上级
517f6195
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
39 addition
and
12 deletion
+39
-12
paddle/fluid/operators/math/math_function.cu
paddle/fluid/operators/math/math_function.cu
+24
-9
paddle/fluid/platform/dynload/cublas.cc
paddle/fluid/platform/dynload/cublas.cc
+4
-0
paddle/fluid/platform/dynload/cublas.h
paddle/fluid/platform/dynload/cublas.h
+11
-3
未找到文件。
paddle/fluid/operators/math/math_function.cu
浏览文件 @
d00bd9eb
...
...
@@ -39,18 +39,33 @@ void gemm<platform::CUDADeviceContext, float16>(
cublasOperation_t
cuTransB
=
(
transB
==
CblasNoTrans
)
?
CUBLAS_OP_N
:
CUBLAS_OP_T
;
const
half
h_alpha
=
static_cast
<
const
half
>
(
alpha
);
const
half
h_beta
=
static_cast
<
const
half
>
(
beta
);
const
half
*
h_A
=
reinterpret_cast
<
const
half
*>
(
A
);
const
half
*
h_B
=
reinterpret_cast
<
const
half
*>
(
B
);
half
*
h_C
=
reinterpret_cast
<
half
*>
(
C
);
float
h_alpha
=
static_cast
<
float
>
(
alpha
);
float
h_beta
=
static_cast
<
float
>
(
beta
);
// TODO(kexinzhao): add processing code for compute capability < 53 case
PADDLE_ENFORCE_GE
(
context
.
GetComputeCapability
(),
53
,
"cublas Hgemm requires GPU compute capability >= 53"
);
PADDLE_ENFORCE
(
platform
::
dynload
::
cublasHgemm
(
context
.
cublas_handle
(),
cuTransB
,
cuTransA
,
N
,
M
,
K
,
&
h_alpha
,
h_B
,
ldb
,
h_A
,
lda
,
&
h_beta
,
h_C
,
N
));
"cublas fp16 gemm requires GPU compute capability >= 53"
);
cublasGemmAlgo_t
algo
=
CUBLAS_GEMM_DFALT
;
#if CUDA_VERSION >= 9000
if
(
context
.
GetComputeCapability
()
>=
70
)
{
PADDLE_ENFORCE
(
platform
::
dynload
::
cublasSetMathMode
(
context
.
cublas_handle
(),
CUBLAS_TENSOR_OP_MATH
));
algo
=
CUBLAS_GEMM_DFALT_TENSOR_OP
;
}
else
{
PADDLE_ENFORCE
(
platform
::
dynload
::
cublasSetMathMode
(
context
.
cublas_handle
(),
CUBLAS_DEFAULT_MATH
));
}
#endif
// cublasHgemm does true FP16 computation which is slow for non-Volta
// GPUs. So use cublasGemmEx instead which does pesudo FP16 computation:
// input/output in fp16, computation in fp32, which can also be accelerated
// using tensor cores in volta GPUs.
PADDLE_ENFORCE
(
platform
::
dynload
::
cublasGemmEx
(
context
.
cublas_handle
(),
cuTransB
,
cuTransA
,
N
,
M
,
K
,
&
h_alpha
,
B
,
CUDA_R_16F
,
ldb
,
A
,
CUDA_R_16F
,
lda
,
&
h_beta
,
C
,
CUDA_R_16F
,
N
,
CUDA_R_32F
,
algo
));
}
template
<
>
...
...
paddle/fluid/platform/dynload/cublas.cc
浏览文件 @
d00bd9eb
...
...
@@ -24,6 +24,10 @@ void *cublas_dso_handle = nullptr;
CUBLAS_BLAS_ROUTINE_EACH
(
DEFINE_WRAP
);
#ifdef CUBLAS_BLAS_ROUTINE_EACH_R2
CUBLAS_BLAS_ROUTINE_EACH_R2
(
DEFINE_WRAP
);
#endif
}
// namespace dynload
}
// namespace platform
}
// namespace paddle
paddle/fluid/platform/dynload/cublas.h
浏览文件 @
d00bd9eb
...
...
@@ -15,8 +15,9 @@ limitations under the License. */
#pragma once
#include <cublas_v2.h>
#include <cuda.h>
#include <dlfcn.h>
#include <mutex>
#include <mutex>
// NOLINT
#include "paddle/fluid/platform/dynload/dynamic_loader.h"
namespace
paddle
{
...
...
@@ -70,6 +71,7 @@ extern void *cublas_dso_handle;
__macro(cublasDgemm_v2); \
__macro(cublasHgemm); \
__macro(cublasSgemmEx); \
__macro(cublasGemmEx); \
__macro(cublasSgeam_v2); \
__macro(cublasDgeam_v2); \
__macro(cublasCreate_v2); \
...
...
@@ -89,9 +91,15 @@ extern void *cublas_dso_handle;
__macro(cublasSgetrfBatched); \
__macro(cublasSgetriBatched); \
__macro(cublasDgetrfBatched); \
__macro(cublasDgetriBatched)
__macro(cublasDgetriBatched)
;
CUBLAS_BLAS_ROUTINE_EACH
(
DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP
);
CUBLAS_BLAS_ROUTINE_EACH
(
DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP
)
// APIs available after CUDA 9.0
#if CUDA_VERSION >= 9000
#define CUBLAS_BLAS_ROUTINE_EACH_R2(__macro) __macro(cublasSetMathMode);
CUBLAS_BLAS_ROUTINE_EACH_R2
(
DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP
)
#endif
#undef DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP
}
// namespace dynload
...
...
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