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
d7aef892
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
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
未验证
提交
d7aef892
编写于
8月 01, 2023
作者:
R
ronnywang
提交者:
GitHub
8月 01, 2023
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[ROCM] fix concat and split (#55821)
上级
a00f5bd4
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
27 addition
and
56 deletion
+27
-56
cmake/hip.cmake
cmake/hip.cmake
+7
-4
paddle/phi/common/bfloat16.h
paddle/phi/common/bfloat16.h
+4
-0
paddle/phi/common/complex.h
paddle/phi/common/complex.h
+4
-0
paddle/phi/common/float16.h
paddle/phi/common/float16.h
+4
-0
paddle/phi/kernels/funcs/concat_and_split_functor.cu
paddle/phi/kernels/funcs/concat_and_split_functor.cu
+4
-52
paddle/phi/kernels/funcs/segmented_array.h
paddle/phi/kernels/funcs/segmented_array.h
+4
-0
未找到文件。
cmake/hip.cmake
浏览文件 @
d7aef892
...
...
@@ -113,11 +113,14 @@ list(APPEND HIP_CXX_FLAGS -Wno-dangling-gsl)
list
(
APPEND HIP_CXX_FLAGS -Wno-unused-value
)
list
(
APPEND HIP_CXX_FLAGS -Wno-braced-scalar-init
)
list
(
APPEND HIP_CXX_FLAGS -Wno-return-type
)
list
(
APPEND HIP_CXX_FLAGS -Wno-pragma-once-outside-header
)
if
(
WITH_CINN
)
list
(
APPEND HIP_CXX_FLAGS -std=c++14
)
else
()
list
(
APPEND HIP_CXX_FLAGS -std=c++17
)
endif
()
list
(
APPEND HIP_CXX_FLAGS --gpu-max-threads-per-block=1024
)
if
(
CMAKE_BUILD_TYPE MATCHES Debug
)
list
(
APPEND HIP_CXX_FLAGS -g2
)
...
...
@@ -130,11 +133,11 @@ set(HIP_CLANG_FLAGS ${HIP_CXX_FLAGS})
# Ask hcc to generate device code during compilation so we can use
# host linker to link.
list
(
APPEND HIP_HCC_FLAGS -fno-gpu-rdc
)
list
(
APPEND HIP_HCC_FLAGS --
amdgpu-target
=gfx906
)
list
(
APPEND HIP_HCC_FLAGS --
amdgpu-target
=gfx908
)
list
(
APPEND HIP_HCC_FLAGS --
offload-arch
=gfx906
)
list
(
APPEND HIP_HCC_FLAGS --
offload-arch
=gfx908
)
list
(
APPEND HIP_CLANG_FLAGS -fno-gpu-rdc
)
list
(
APPEND HIP_CLANG_FLAGS --
amdgpu-target
=gfx906
)
list
(
APPEND HIP_CLANG_FLAGS --
amdgpu-target
=gfx908
)
list
(
APPEND HIP_CLANG_FLAGS --
offload-arch
=gfx906
)
list
(
APPEND HIP_CLANG_FLAGS --
offload-arch
=gfx908
)
if
(
HIP_COMPILER STREQUAL clang
)
set
(
hip_library_name amdhip64
)
...
...
paddle/phi/common/bfloat16.h
浏览文件 @
d7aef892
...
...
@@ -31,11 +31,15 @@
#include <cuda_bf16.h>
#endif
#ifndef PADDLE_WITH_HIP
#if !defined(_WIN32)
#define PADDLE_ALIGN(x) __attribute__((aligned(x)))
#else
#define PADDLE_ALIGN(x) __declspec(align(x))
#endif
#else
#define PADDLE_ALIGN(x)
#endif
namespace
phi
{
namespace
dtype
{
...
...
paddle/phi/common/complex.h
浏览文件 @
d7aef892
...
...
@@ -31,11 +31,15 @@
#include <thrust/complex.h> // NOLINT
#endif
#ifndef PADDLE_WITH_HIP
#if !defined(_WIN32)
#define PADDLE_ALIGN(x) __attribute__((aligned(x)))
#else
#define PADDLE_ALIGN(x) __declspec(align(x))
#endif
#else
#define PADDLE_ALIGN(x)
#endif
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
// todo
...
...
paddle/phi/common/float16.h
浏览文件 @
d7aef892
...
...
@@ -51,11 +51,15 @@
#include <hip/hip_fp16.h>
#endif
#ifndef PADDLE_WITH_HIP
#if !defined(_WIN32)
#define PADDLE_ALIGN(x) __attribute__((aligned(x)))
#else
#define PADDLE_ALIGN(x) __declspec(align(x))
#endif
#else
#define PADDLE_ALIGN(x)
#endif
#define CUDA_ARCH_FP16_SUPPORTED(CUDA_ARCH) (CUDA_ARCH >= 600)
...
...
paddle/phi/kernels/funcs/concat_and_split_functor.cu
浏览文件 @
d7aef892
...
...
@@ -49,11 +49,15 @@ static inline void GetBlockDims(const phi::GPUContext& context,
*
grid_dims
=
dim3
(
grid_cols
,
grid_rows
,
1
);
}
#ifndef PADDLE_WITH_HIP
#if !defined(_WIN32)
#define PADDLE_ALIGN(x) __attribute__((aligned(x)))
#else
#define PADDLE_ALIGN(x)
#endif
#else
#define PADDLE_ALIGN(x)
#endif
template
<
typename
T
,
int
Size
>
struct
PointerWrapper
{
...
...
@@ -572,15 +576,6 @@ void ConcatFunctorWithIndexType(const phi::GPUContext& ctx,
std
::
vector
<
IndexT
>
inputs_col_vec
(
inputs_col_num
,
0
);
const
T
**
inputs_data
=
inputs_data_vec
.
data
();
IndexT
*
inputs_col
=
inputs_col_vec
.
data
();
#ifdef PADDLE_WITH_HIP
// TODO(chentianyu03): try to find a method to remove the Alloc function
phi
::
Allocator
::
AllocationPtr
data_alloc
=
phi
::
memory_utils
::
Alloc
(
phi
::
GPUPinnedPlace
(),
in_num
*
sizeof
(
T
*
));
inputs_data
=
reinterpret_cast
<
const
T
**>
(
data_alloc
->
ptr
());
phi
::
Allocator
::
AllocationPtr
col_alloc
=
phi
::
memory_utils
::
Alloc
(
phi
::
GPUPinnedPlace
(),
inputs_col_num
*
sizeof
(
IndexT
));
inputs_col
=
reinterpret_cast
<
IndexT
*>
(
col_alloc
->
ptr
());
#endif
bool
has_same_shape
=
true
;
for
(
int
i
=
0
;
i
<
in_num
;
++
i
)
{
...
...
@@ -604,19 +599,6 @@ void ConcatFunctorWithIndexType(const phi::GPUContext& ctx,
in_num
,
limit_num
,
has_same_shape
);
#ifdef PADDLE_WITH_HIP
// Prevent pinned memory from being covered and release the memory after
// kernel launch of the stream is executed (reapply pinned memory next time)
auto
*
data_alloc_released
=
data_alloc
.
release
();
auto
*
col_alloc_released
=
col_alloc
.
release
();
ctx
.
AddStreamCallback
([
data_alloc_released
,
col_alloc_released
]
{
VLOG
(
4
)
<<
"Delete cuda pinned at "
<<
data_alloc_released
;
VLOG
(
4
)
<<
"Delete cuda pinned at "
<<
col_alloc_released
;
phi
::
memory_utils
::
AllocationDeleter
(
data_alloc_released
);
phi
::
memory_utils
::
AllocationDeleter
(
col_alloc_released
);
});
#endif
}
template
<
typename
T
>
...
...
@@ -780,25 +762,6 @@ void SplitFunctorDispatchWithIndexType(
IndexT
*
outs_cols
=
outputs_cols_vec
.
data
();
T
**
outs_data
=
nullptr
;
// There are some differences between hip runtime and NV runtime.
// In NV, when the pageable memory data less than 64K is transferred from
// hosttodevice, it will be automatically asynchronous.
// However, only pinned memory in hip can copy asynchronously
// https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#concurrent-execution-host-device
// 3.2.6.1. Concurrent Execution between Host and Device
// Memory copies from host to device of a memory block of 64 KB or less
#ifdef PADDLE_WITH_HIP
phi
::
Allocator
::
AllocationPtr
data_alloc
,
cols_alloc
;
// TODO(chentianyu03): try to find a method to remove the Alloc function
data_alloc
=
phi
::
memory_utils
::
Alloc
(
phi
::
GPUPinnedPlace
(),
out_num
*
sizeof
(
T
*
));
outs_data
=
reinterpret_cast
<
T
**>
(
data_alloc
->
ptr
());
// TODO(chentianyu03): try to find a method to remove the Alloc function
cols_alloc
=
phi
::
memory_utils
::
Alloc
(
phi
::
GPUPinnedPlace
(),
(
out_cols_num
)
*
sizeof
(
IndexT
));
outs_cols
=
reinterpret_cast
<
IndexT
*>
(
cols_alloc
->
ptr
());
#endif
outs_cols
[
0
]
=
0
;
for
(
int
i
=
0
;
i
<
out_num
;
++
i
)
{
IndexT
t_col
=
ref_ins
.
at
(
i
)
->
numel
()
/
out_row
;
...
...
@@ -835,17 +798,6 @@ void SplitFunctorDispatchWithIndexType(
outs_data
));
}
}
#ifdef PADDLE_WITH_HIP
// Prevent pinned memory from being covered and release the memory after
// kernel launch of the stream is executed (reapply pinned memory next time)
auto
*
data_alloc_released
=
data_alloc
.
release
();
auto
*
cols_alloc_released
=
cols_alloc
.
release
();
ctx
.
AddStreamCallback
([
data_alloc_released
,
cols_alloc_released
]
{
phi
::
memory_utils
::
AllocationDeleter
(
data_alloc_released
);
phi
::
memory_utils
::
AllocationDeleter
(
cols_alloc_released
);
});
#endif
}
template
<
typename
T
>
...
...
paddle/phi/kernels/funcs/segmented_array.h
浏览文件 @
d7aef892
...
...
@@ -21,11 +21,15 @@
namespace
phi
{
namespace
funcs
{
#ifndef PADDLE_WITH_HIP
#if !defined(_WIN32)
#define PADDLE_ALIGN(x) __attribute__((aligned(x)))
#else
#define PADDLE_ALIGN(x)
#endif
#else
#define PADDLE_ALIGN(x)
#endif
enum
class
SegmentedArraySize
{
kVariableLength
=
0
,
...
...
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