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
c8fac5ee
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
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
c8fac5ee
编写于
2月 26, 2021
作者:
Q
Qi Li
提交者:
GitHub
2月 26, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[ROCM] update fluid framework for rocm (part5), test=develop (#31014)
上级
580447d0
变更
19
隐藏空白更改
内联
并排
Showing
19 changed file
with
135 addition
and
48 deletion
+135
-48
paddle/fluid/framework/garbage_collector.cc
paddle/fluid/framework/garbage_collector.cc
+12
-3
paddle/fluid/framework/garbage_collector.h
paddle/fluid/framework/garbage_collector.h
+3
-3
paddle/fluid/framework/heter_service.h
paddle/fluid/framework/heter_service.h
+1
-1
paddle/fluid/framework/heterbox_trainer.cc
paddle/fluid/framework/heterbox_trainer.cc
+22
-7
paddle/fluid/framework/lod_tensor.h
paddle/fluid/framework/lod_tensor.h
+1
-1
paddle/fluid/framework/lod_tensor_test.cu
paddle/fluid/framework/lod_tensor_test.cu
+12
-2
paddle/fluid/framework/mixed_vector.h
paddle/fluid/framework/mixed_vector.h
+1
-1
paddle/fluid/framework/mixed_vector_test.cu
paddle/fluid/framework/mixed_vector_test.cu
+25
-1
paddle/fluid/framework/op_registry.h
paddle/fluid/framework/op_registry.h
+1
-1
paddle/fluid/framework/operator.cc
paddle/fluid/framework/operator.cc
+5
-1
paddle/fluid/framework/operator.h
paddle/fluid/framework/operator.h
+1
-1
paddle/fluid/framework/parallel_executor.cc
paddle/fluid/framework/parallel_executor.cc
+23
-12
paddle/fluid/framework/parallel_executor.h
paddle/fluid/framework/parallel_executor.h
+1
-1
paddle/fluid/framework/pipeline_trainer.cc
paddle/fluid/framework/pipeline_trainer.cc
+1
-1
paddle/fluid/framework/ps_gpu_trainer.cc
paddle/fluid/framework/ps_gpu_trainer.cc
+2
-1
paddle/fluid/framework/ps_gpu_worker.cc
paddle/fluid/framework/ps_gpu_worker.cc
+2
-1
paddle/fluid/framework/pull_dense_worker.cc
paddle/fluid/framework/pull_dense_worker.cc
+11
-7
paddle/fluid/framework/save_load_util.cc
paddle/fluid/framework/save_load_util.cc
+1
-1
paddle/fluid/framework/section_worker.cc
paddle/fluid/framework/section_worker.cc
+10
-2
未找到文件。
paddle/fluid/framework/garbage_collector.cc
浏览文件 @
c8fac5ee
...
...
@@ -13,7 +13,7 @@
// limitations under the License.
#include <functional>
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif
#include "gflags/gflags.h"
...
...
@@ -53,7 +53,7 @@ void XPUGarbageCollector::ClearCallback(const std::function<void()> &callback) {
}
#endif
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
UnsafeFastGPUGarbageCollector
::
UnsafeFastGPUGarbageCollector
(
const
platform
::
CUDAPlace
&
place
,
size_t
max_memory_size
)
:
GarbageCollector
(
place
,
max_memory_size
)
{}
...
...
@@ -82,18 +82,27 @@ StreamGarbageCollector::StreamGarbageCollector(const platform::CUDAPlace &place,
size_t
max_memory_size
)
:
GarbageCollector
(
place
,
max_memory_size
)
{
platform
::
CUDADeviceGuard
guard
(
place
.
device
);
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_CUDA_SUCCESS
(
hipStreamCreate
(
&
stream_
));
#else
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaStreamCreate
(
&
stream_
));
#endif
callback_manager_
.
reset
(
new
platform
::
StreamCallbackManager
(
stream_
));
}
StreamGarbageCollector
::~
StreamGarbageCollector
()
{
auto
place
=
BOOST_GET_CONST
(
platform
::
CUDAPlace
,
this
->
dev_ctx_
->
GetPlace
());
platform
::
CUDADeviceGuard
guard
(
place
.
device
);
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_CUDA_SUCCESS
(
hipStreamSynchronize
(
stream_
));
PADDLE_ENFORCE_CUDA_SUCCESS
(
hipStreamDestroy
(
stream_
));
#else
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaStreamSynchronize
(
stream_
));
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaStreamDestroy
(
stream_
));
#endif
}
cuda
Stream_t
StreamGarbageCollector
::
stream
()
const
{
return
stream_
;
}
gpu
Stream_t
StreamGarbageCollector
::
stream
()
const
{
return
stream_
;
}
void
StreamGarbageCollector
::
Wait
()
const
{
callback_manager_
->
Wait
();
}
...
...
paddle/fluid/framework/garbage_collector.h
浏览文件 @
c8fac5ee
...
...
@@ -80,7 +80,7 @@ class XPUGarbageCollector : public GarbageCollector {
};
#endif
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
class
UnsafeFastGPUGarbageCollector
:
public
GarbageCollector
{
public:
UnsafeFastGPUGarbageCollector
(
const
platform
::
CUDAPlace
&
place
,
...
...
@@ -110,13 +110,13 @@ class StreamGarbageCollector : public GarbageCollector {
void
Wait
()
const
override
;
cuda
Stream_t
stream
()
const
;
gpu
Stream_t
stream
()
const
;
protected:
void
ClearCallback
(
const
std
::
function
<
void
()
>
&
callback
)
override
;
private:
cuda
Stream_t
stream_
;
gpu
Stream_t
stream_
;
std
::
unique_ptr
<
platform
::
StreamCallbackManager
>
callback_manager_
;
};
...
...
paddle/fluid/framework/heter_service.h
浏览文件 @
c8fac5ee
...
...
@@ -152,7 +152,7 @@ class HeterObjectPool {
std
::
lock_guard
<
std
::
mutex
>
lock
(
mutex_
);
if
(
pool_
.
empty
())
{
num_
+=
1
;
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
VLOG
(
0
)
<<
"pool construct size: "
<<
num_
;
#endif
return
std
::
make_shared
<
T
>
();
...
...
paddle/fluid/framework/heterbox_trainer.cc
浏览文件 @
c8fac5ee
...
...
@@ -21,9 +21,10 @@ limitations under the License. */
#include "paddle/fluid/framework/device_worker_factory.h"
#include "paddle/fluid/framework/fleet/fleet_wrapper.h"
#include "paddle/fluid/framework/trainer.h"
#if (defined PADDLE_WITH_CUDA || defined PADDLE_WITH_XPU) && \
#if (defined PADDLE_WITH_CUDA || defined PADDLE_WITH_HIP || \
defined PADDLE_WITH_XPU) && \
(defined PADDLE_WITH_PSLIB)
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif
namespace
paddle
{
...
...
@@ -48,16 +49,25 @@ void HeterBoxTrainer::Initialize(const TrainerDesc& trainer_desc,
dataset
->
GetReaders
();
for
(
int
i
=
0
;
i
<
place_num
;
++
i
)
{
int
num
=
trainer_desc
.
worker_places
(
i
);
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
platform
::
CUDAPlace
place
=
platform
::
CUDAPlace
(
num
);
platform
::
CUDADeviceGuard
guard
(
place
.
device
);
cudaStream_t
stream
;
gpuStream_t
stream
;
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_CUDA_SUCCESS
(
hipStreamCreate
(
&
stream
));
#else
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaStreamCreate
(
&
stream
));
#endif
copy_streams_
.
push_back
(
stream
);
places_
.
push_back
(
place
);
cudaEvent_t
event
;
gpuEvent_t
event
;
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_CUDA_SUCCESS
(
hipEventCreateWithFlags
(
&
event
,
hipEventDisableTiming
));
#else
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaEventCreateWithFlags
(
&
event
,
cudaEventDisableTiming
));
#endif
events_
.
push_back
(
event
);
#endif
#ifdef PADDLE_WITH_XPU
...
...
@@ -140,8 +150,13 @@ void HeterBoxTrainer::InitTrainerEnv(const ProgramDesc& main_program,
_ForEachDataType_
(
HeterMemcpyFunc
);
}
}
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_CUDA_SUCCESS
(
hipEventRecord
(
event
,
stream
));
hipEventSynchronize
(
event
);
#else
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaEventRecord
(
event
,
stream
));
cudaEventSynchronize
(
event
);
#endif
}
place_
=
place
;
}
...
...
@@ -150,7 +165,7 @@ template <typename T>
void
HeterBoxTrainer
::
HeterMemCpy
(
LoDTensor
*
thread_tensor
,
LoDTensor
*
root_tensor
,
const
paddle
::
platform
::
Place
&
thread_place
,
cuda
Stream_t
stream
)
{
gpu
Stream_t
stream
)
{
T
*
thread_ptr
=
thread_tensor
->
mutable_data
<
T
>
(
root_tensor
->
dims
(),
thread_place
);
T
*
root_ptr
=
root_tensor
->
data
<
T
>
();
...
...
@@ -171,7 +186,7 @@ void HeterBoxTrainer::InitOtherEnv(const ProgramDesc& main_program) {
for
(
size_t
i
=
0
;
i
<
places_
.
size
();
++
i
)
{
pull_dense_worker_
->
AddThreadScope
(
workers_
[
i
]
->
GetThreadScope
());
pull_dense_worker_
->
AddPlace
(
places_
[
i
]);
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
pull_dense_worker_
->
AddStream
(
copy_streams_
[
i
]);
#endif
}
...
...
paddle/fluid/framework/lod_tensor.h
浏览文件 @
c8fac5ee
...
...
@@ -18,7 +18,7 @@ limitations under the License. */
#include <string>
#include <utility>
#include <vector>
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
#endif
...
...
paddle/fluid/framework/lod_tensor_test.cu
浏览文件 @
c8fac5ee
...
...
@@ -12,8 +12,6 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include <cuda.h>
#include <cuda_runtime.h>
#include <stdio.h>
#include "gtest/gtest.h"
...
...
@@ -34,8 +32,14 @@ TEST(LoD, data) {
auto
&
v
=
lod
[
0
];
paddle
::
platform
::
CUDAPlace
gpu
(
0
);
#ifdef PADDLE_WITH_HIP
hipLaunchKernelGGL
(
test
,
dim3
(
1
),
dim3
(
1
),
0
,
0
,
v
.
CUDAMutableData
(
gpu
),
v
.
size
());
hipDeviceSynchronize
();
#else
test
<<<
1
,
1
>>>
(
v
.
CUDAMutableData
(
gpu
),
v
.
size
());
cudaDeviceSynchronize
();
#endif
for
(
size_t
i
=
0
;
i
<
v
.
size
();
++
i
)
{
EXPECT_EQ
(
v
[
i
],
i
*
2
);
}
...
...
@@ -59,8 +63,14 @@ TEST(LoDTensor, LoDInGPU) {
auto
lod
=
lod_tensor
.
lod
();
#ifdef PADDLE_WITH_HIP
hipLaunchKernelGGL
(
test
,
dim3
(
1
),
dim3
(
8
),
0
,
0
,
lod
[
0
].
CUDAMutableData
(
place
),
lod
[
0
].
size
());
hipDeviceSynchronize
();
#else
test
<<<
1
,
8
>>>
(
lod
[
0
].
CUDAMutableData
(
place
),
lod
[
0
].
size
());
cudaDeviceSynchronize
();
#endif
for
(
size_t
i
=
0
;
i
<
src_lod
[
0
].
size
();
++
i
)
{
EXPECT_EQ
(
lod
[
0
].
data
()[
i
],
src_lod
[
0
].
data
()[
i
]
*
2
);
...
...
paddle/fluid/framework/mixed_vector.h
浏览文件 @
c8fac5ee
...
...
@@ -31,7 +31,7 @@ limitations under the License. */
namespace
paddle
{
namespace
framework
{
#if defined(PADDLE_WITH_CUDA)
#if defined(PADDLE_WITH_CUDA)
|| defined(PADDLE_WITH_HIP)
// Vector<T> implements the std::vector interface, and can get Data or
// MutableData from any place. The data will be synced implicitly inside.
template
<
typename
T
>
...
...
paddle/fluid/framework/mixed_vector_test.cu
浏览文件 @
c8fac5ee
...
...
@@ -12,7 +12,13 @@
See the License for the specific language governing permissions and
limitations under the License. */
#ifdef PADDLE_WITH_CUDA
#include <cuda_runtime.h>
#endif
#ifdef PADDLE_WITH_HIP
#include <hip/hip_runtime.h>
#endif
#include <memory>
#include "glog/logging.h"
...
...
@@ -22,6 +28,7 @@
template
<
typename
T
>
using
vec
=
paddle
::
framework
::
Vector
<
T
>
;
using
gpuStream_t
=
paddle
::
gpuStream_t
;
static
__global__
void
multiply_10
(
int
*
ptr
)
{
for
(
int
i
=
0
;
i
<
10
;
++
i
)
{
...
...
@@ -29,7 +36,7 @@ static __global__ void multiply_10(int* ptr) {
}
}
cuda
Stream_t
GetCUDAStream
(
paddle
::
platform
::
CUDAPlace
place
)
{
gpu
Stream_t
GetCUDAStream
(
paddle
::
platform
::
CUDAPlace
place
)
{
return
reinterpret_cast
<
const
paddle
::
platform
::
CUDADeviceContext
*>
(
paddle
::
platform
::
DeviceContextPool
::
Instance
().
Get
(
place
))
->
stream
();
...
...
@@ -43,7 +50,12 @@ TEST(mixed_vector, GPU_VECTOR) {
ASSERT_EQ
(
tmp
.
size
(),
10UL
);
paddle
::
platform
::
CUDAPlace
gpu
(
0
);
#ifdef PADDLE_WITH_HIP
hipLaunchKernelGGL
(
multiply_10
,
dim3
(
1
),
dim3
(
1
),
0
,
GetCUDAStream
(
gpu
),
tmp
.
MutableData
(
gpu
));
#else
multiply_10
<<<
1
,
1
,
0
,
GetCUDAStream
(
gpu
)
>>>
(
tmp
.
MutableData
(
gpu
));
#endif
for
(
int
i
=
0
;
i
<
10
;
++
i
)
{
ASSERT_EQ
(
tmp
[
i
],
i
*
10
);
...
...
@@ -64,11 +76,23 @@ TEST(mixed_vector, MultiGPU) {
ASSERT_EQ
(
tmp
.
size
(),
10UL
);
paddle
::
platform
::
CUDAPlace
gpu0
(
0
);
paddle
::
platform
::
SetDeviceId
(
0
);
#ifdef PADDLE_WITH_HIP
hipLaunchKernelGGL
(
multiply_10
,
dim3
(
1
),
dim3
(
1
),
0
,
GetCUDAStream
(
gpu0
),
tmp
.
MutableData
(
gpu0
));
#else
multiply_10
<<<
1
,
1
,
0
,
GetCUDAStream
(
gpu0
)
>>>
(
tmp
.
MutableData
(
gpu0
));
#endif
paddle
::
platform
::
CUDAPlace
gpu1
(
1
);
auto
*
gpu1_ptr
=
tmp
.
MutableData
(
gpu1
);
paddle
::
platform
::
SetDeviceId
(
1
);
#ifdef PADDLE_WITH_HIP
hipLaunchKernelGGL
(
multiply_10
,
dim3
(
1
),
dim3
(
1
),
0
,
GetCUDAStream
(
gpu1
),
gpu1_ptr
);
#else
multiply_10
<<<
1
,
1
,
0
,
GetCUDAStream
(
gpu1
)
>>>
(
gpu1_ptr
);
#endif
for
(
int
i
=
0
;
i
<
10
;
++
i
)
{
ASSERT_EQ
(
tmp
[
i
],
i
*
100
);
}
...
...
paddle/fluid/framework/op_registry.h
浏览文件 @
c8fac5ee
...
...
@@ -369,7 +369,7 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
// TODO(fengjiayi): The following macros
// seems ugly, do we have better method?
#if
ndef PADDLE_WITH_CUDA
#if
!defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
#define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU)
#else
#define USE_OP_KERNEL(op_type) \
...
...
paddle/fluid/framework/operator.cc
浏览文件 @
c8fac5ee
...
...
@@ -193,7 +193,7 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
try
{
VLOG
(
4
)
<<
place
<<
" "
<<
DebugStringEx
(
&
scope
);
if
(
platform
::
is_gpu_place
(
place
))
{
#if
ndef PADDLE_WITH_CUDA
#if
!defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
PADDLE_THROW
(
platform
::
errors
::
Unavailable
(
"Cannot run operator on place %s, please recompile paddle or "
"reinstall Paddle with CUDA support."
,
...
...
@@ -1166,6 +1166,10 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
#if defined(PADDLE_WITH_CUDA)
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaGetLastError
());
VLOG
(
4
)
<<
"Operator("
<<
Type
()
<<
"): context wait and get last error"
;
#endif
#if defined(PADDLE_WITH_HIP)
PADDLE_ENFORCE_CUDA_SUCCESS
(
hipGetLastError
());
VLOG
(
4
)
<<
"Operator("
<<
Type
()
<<
"): context wait and get last error"
;
#endif
}
...
...
paddle/fluid/framework/operator.h
浏览文件 @
c8fac5ee
...
...
@@ -384,7 +384,7 @@ class ExecutionContext {
return
device_context_
;
}
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
const
inline
platform
::
CUDADeviceContext
&
cuda_device_context
()
const
{
PADDLE_ENFORCE_EQ
(
platform
::
is_gpu_place
(
device_context_
.
GetPlace
()),
true
,
platform
::
errors
::
PreconditionNotMet
(
...
...
paddle/fluid/framework/parallel_executor.cc
浏览文件 @
c8fac5ee
...
...
@@ -37,7 +37,7 @@ limitations under the License. */
#include "paddle/fluid/platform/event.h"
#include "paddle/fluid/platform/profiler.h"
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif
...
...
@@ -60,7 +60,7 @@ static std::once_flag gProfileOnce;
static
bool
gProfileStarted
=
false
;
#endif
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
std
::
once_flag
p2p_init_flag
;
#endif
...
...
@@ -132,7 +132,7 @@ class ParallelExecutorPrivate {
}
}
#if defined(PADDLE_WITH_NCCL)
#if defined(PADDLE_WITH_NCCL)
|| defined(PADDLE_WITH_RCCL)
void
InitNCCLCtxs
(
framework
::
Scope
*
scope
,
const
BuildStrategy
&
bst
)
{
VLOG
(
1
)
<<
"nccl comm num:"
<<
bst
.
nccl_comm_num_
<<
", nranks:"
<<
nranks_
<<
", num_trainers:"
<<
bst
.
num_trainers_
...
...
@@ -371,7 +371,7 @@ class ParallelExecutorPrivate {
std
::
unordered_map
<
std
::
string
,
bool
>
is_persistable_
;
#if defined(PADDLE_WITH_NCCL)
#if defined(PADDLE_WITH_NCCL)
|| defined(PADDLE_WITH_RCCL)
platform
::
NCCLCommunicator
*
nccl_ctxs_
{
nullptr
};
#elif defined(PADDLE_WITH_XPU_BKCL)
platform
::
BKCLCommunicator
*
bkcl_ctxs_
{
nullptr
};
...
...
@@ -483,7 +483,7 @@ ir::Graph *ParallelExecutorPrivate::ApplyMemoryOptimizePass(ir::Graph *graph) {
}
std
::
unique_ptr
<
GarbageCollector
>
gc
;
if
(
platform
::
is_gpu_place
(
place
))
{
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
if
(
IsFastEagerDeletionModeEnabled
())
{
gc
.
reset
(
new
UnsafeFastGPUGarbageCollector
(
BOOST_GET_CONST
(
platform
::
CUDAPlace
,
place
),
max_memory_size
));
...
...
@@ -572,7 +572,7 @@ bool ParallelExecutor::NeedCreateLocalExeScope() {
}
void
InitP2P
(
const
std
::
vector
<
platform
::
Place
>
&
places
)
{
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
std
::
call_once
(
p2p_init_flag
,
[
&
]()
{
int
count
=
places
.
size
();
if
(
count
<=
1
)
return
;
...
...
@@ -590,14 +590,24 @@ void InitP2P(const std::vector<platform::Place> &places) {
for
(
int
j
=
0
;
j
<
count
;
++
j
)
{
if
(
devices
[
i
]
==
devices
[
j
])
continue
;
int
can_acess
=
-
1
;
#ifdef PADDLE_WITH_HIP
hipError_t
ret
=
hipDeviceCanAccessPeer
(
&
can_acess
,
devices
[
i
],
devices
[
j
]);
if
(
ret
!=
hipSuccess
||
can_acess
!=
1
)
{
#else
cudaError_t
ret
=
cudaDeviceCanAccessPeer
(
&
can_acess
,
devices
[
i
],
devices
[
j
]);
if
(
ret
!=
cudaSuccess
||
can_acess
!=
1
)
{
#endif
LOG
(
WARNING
)
<<
"Cannot enable P2P access from "
<<
devices
[
i
]
<<
" to "
<<
devices
[
j
];
}
else
{
platform
::
CUDADeviceGuard
guard
(
devices
[
i
]);
#ifdef PADDLE_WITH_HIP
hipDeviceEnablePeerAccess
(
devices
[
j
],
0
);
#else
cudaDeviceEnablePeerAccess
(
devices
[
j
],
0
);
#endif
}
}
}
...
...
@@ -630,7 +640,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
BuildStrategy
::
ReduceStrategy
::
kAllReduce
;
member_
->
use_all_reduce_
=
true
;
}
#if
defined(PADDLE_WITH_CUDA
) && defined(_WIN32)
#if
(defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
) && defined(_WIN32)
if
(
member_
->
IsUseCUDA
(
member_
->
use_device_
))
{
PADDLE_ENFORCE_EQ
(
places
.
size
(),
1
,
...
...
@@ -638,7 +648,8 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
}
#endif
#if defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_NCCL)
#if (defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)) && \
(!defined(PADDLE_WITH_NCCL) && !defined(PADDLE_WITH_RCCL))
if
(
member_
->
IsUseCUDA
(
member_
->
use_device_
))
{
PADDLE_ENFORCE_EQ
(
places
.
size
(),
1
,
...
...
@@ -710,7 +721,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
}
if
(
member_
->
IsUseCUDA
(
member_
->
use_device_
)
&&
member_
->
nranks_
>
1
)
{
#if defined(PADDLE_WITH_NCCL)
#if defined(PADDLE_WITH_NCCL)
|| defined(PADDLE_WITH_RCCL)
member_
->
InitOrGetNCCLCommunicator
(
scope
,
&
member_
->
build_strategy_
);
// Initialize device context's nccl comm, will be used by normal
...
...
@@ -774,7 +785,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
// ncclOp
std
::
vector
<
ir
::
Graph
*>
async_graphs
(
places
.
size
());
#if defined(PADDLE_WITH_NCCL)
#if defined(PADDLE_WITH_NCCL)
|| defined(PADDLE_WITH_RCCL)
if
(
member_
->
build_strategy_
.
async_mode_
)
{
VLOG
(
3
)
<<
"use local async mode"
;
graph
=
member_
->
build_strategy_
.
Apply
(
...
...
@@ -885,7 +896,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
final_graphs
=
async_graphs
;
}
else
if
(
member_
->
build_strategy_
.
enable_parallel_graph_
)
{
VLOG
(
3
)
<<
"use ParallelSSAGraphExecutor"
;
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
// TODO(Yancey1989): Remove passing in the main_program when
// allreduce_seq_pass doesn't need it as the attr.
bool
is_inference
=
details
::
IsDataParallelInferenceGraph
(
*
graph
);
...
...
@@ -996,7 +1007,7 @@ void ParallelExecutor::BCastParamsToDevices(
}
auto
&
dims
=
main_tensor
.
dims
();
if
(
paddle
::
platform
::
is_gpu_place
(
main_tensor
.
place
()))
{
#if defined(PADDLE_WITH_NCCL)
#if defined(PADDLE_WITH_NCCL)
|| defined(PADDLE_WITH_RCCL)
std
::
vector
<
void
*>
buffers
;
buffers
.
reserve
(
member_
->
places_
.
size
());
size_t
numel
=
main_tensor
.
numel
();
...
...
paddle/fluid/framework/parallel_executor.h
浏览文件 @
c8fac5ee
...
...
@@ -32,7 +32,7 @@ limitations under the License. */
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/device_context.h"
#if defined(PADDLE_WITH_NCCL)
#if defined(PADDLE_WITH_NCCL)
|| defined(PADDLE_WITH_RCCL)
#include "paddle/fluid/platform/nccl_helper.h"
#endif
...
...
paddle/fluid/framework/pipeline_trainer.cc
浏览文件 @
c8fac5ee
...
...
@@ -12,7 +12,7 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#if defined(PADDLE_WITH_NCCL)
#if defined(PADDLE_WITH_NCCL)
|| defined(PADDLE_WITH_RCCL)
#include "paddle/fluid/framework/data_feed_factory.h"
#include "paddle/fluid/framework/device_worker_factory.h"
#include "paddle/fluid/framework/trainer.h"
...
...
paddle/fluid/framework/ps_gpu_trainer.cc
浏览文件 @
c8fac5ee
...
...
@@ -24,7 +24,8 @@ limitations under the License. */
#include "paddle/fluid/framework/fleet/heter_ps/feature_value.h"
#include "paddle/fluid/framework/fleet/ps_gpu_wrapper.h"
#include "paddle/fluid/framework/trainer.h"
#if (defined PADDLE_WITH_NCCL) && (defined PADDLE_WITH_PSLIB)
#if (defined PADDLE_WITH_NCCL || defined PADDLE_WITH_RCCL) && \
(defined PADDLE_WITH_PSLIB)
#include "paddle/fluid/platform/cuda_device_guard.h"
namespace
paddle
{
...
...
paddle/fluid/framework/ps_gpu_worker.cc
浏览文件 @
c8fac5ee
...
...
@@ -19,7 +19,8 @@ limitations under the License. */
#include "paddle/fluid/platform/cpu_helper.h"
#include "paddle/fluid/string/string_helper.h"
#if (defined PADDLE_WITH_NCCL) && (defined PADDLE_WITH_PSLIB)
#if (defined PADDLE_WITH_NCCL || defined PADDLE_WITH_RCCL) && \
(defined PADDLE_WITH_PSLIB)
#include "paddle/fluid/platform/cuda_device_guard.h"
#if defined _WIN32 || defined __APPLE__
...
...
paddle/fluid/framework/pull_dense_worker.cc
浏览文件 @
c8fac5ee
...
...
@@ -59,17 +59,19 @@ void PullDenseWorker::Initialize(const TrainerDesc& param) {
current_version_
[
tid
]
=
0
;
}
fleet_ptr_
=
FleetWrapper
::
GetInstance
();
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
copy_streams_
.
clear
();
#endif
#if (defined PADDLE_WITH_CUDA) || (defined PADDLE_WITH_XPU)
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP) || \
defined(PADDLE_WITH_XPU)
places_
.
clear
();
thread_scopes_
.
clear
();
#endif
}
void
PullDenseWorker
::
CreatePinVar
()
{
#if (defined PADDLE_WITH_CUDA) || (defined PADDLE_WITH_XPU)
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP) || \
defined(PADDLE_WITH_XPU)
// for (auto& v : dense_value_names_) {
// for (auto& name : v.second) {
for
(
int
i
=
0
;
i
<
dwp_param_
.
program_config
(
0
).
pull_dense_table_id_size
();
...
...
@@ -84,7 +86,7 @@ void PullDenseWorker::CreatePinVar() {
auto
*
ptr
=
root_scope_
->
Var
(
name
+
"pin"
);
InitializeVariable
(
ptr
,
proto
::
VarType
::
LOD_TENSOR
);
LoDTensor
*
pin_tensor
=
ptr
->
GetMutable
<
LoDTensor
>
();
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
pin_tensor
->
mutable_data
<
float
>
(
tensor
->
dims
(),
platform
::
CUDAPinnedPlace
());
#endif
...
...
@@ -113,7 +115,8 @@ void PullDenseWorker::Wait(std::vector<::std::future<int32_t>>* status_vec) {
exit
(
-
1
);
}
status_vec
->
resize
(
0
);
#if (defined PADDLE_WITH_CUDA) || (defined PADDLE_WITH_XPU)
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP) || \
defined(PADDLE_WITH_XPU)
for
(
size_t
i
=
0
;
i
<
places_
.
size
();
++
i
)
{
// for (auto& v : dense_value_names_) {
...
...
@@ -131,7 +134,7 @@ void PullDenseWorker::Wait(std::vector<::std::future<int32_t>>* status_vec) {
Variable
*
var
=
thread_scopes_
[
i
]
->
FindVar
(
name
);
LoDTensor
*
tensor
=
var
->
GetMutable
<
LoDTensor
>
();
float
*
w
=
tensor
->
data
<
float
>
();
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
memory
::
Copy
(
BOOST_GET_CONST
(
platform
::
CUDAPlace
,
places_
[
i
]),
w
,
platform
::
CUDAPinnedPlace
(),
pin_w
,
sizeof
(
float
)
*
tensor
->
numel
(),
copy_streams_
[
i
]);
...
...
@@ -161,7 +164,8 @@ void PullDenseWorker::PullDense(bool force_update) {
uint64_t
tid
=
static_cast
<
uint64_t
>
(
dwp_param_
.
program_config
(
0
).
pull_dense_table_id
(
i
));
if
(
force_update
||
CheckUpdateParam
(
tid
))
{
#if (defined PADDLE_WITH_CUDA) || (defined PADDLE_WITH_XPU)
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP) || \
defined(PADDLE_WITH_XPU)
VLOG
(
3
)
<<
"pull dense "
<<
force_update
<<
" "
<<
tid
;
fleet_ptr_
->
PullDenseVarsAsync
(
*
root_scope_
,
tid
,
dense_value_names_
[
tid
],
&
pull_dense_status_
,
false
);
...
...
paddle/fluid/framework/save_load_util.cc
浏览文件 @
c8fac5ee
...
...
@@ -297,7 +297,7 @@ bool SaveTensorToDisk(const std::string& file_name,
tensor
->
numel
()
*
framework
::
SizeOfType
(
tensor
->
type
());
auto
*
data_ptr
=
tensor
->
data
<
void
>
();
if
(
platform
::
is_gpu_place
(
tensor
->
place
()))
{
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
framework
::
Tensor
temp
;
TensorCopySync
(
*
tensor
,
platform
::
CPUPlace
(),
&
temp
);
data_ptr
=
temp
.
data
<
void
>
();
...
...
paddle/fluid/framework/section_worker.cc
浏览文件 @
c8fac5ee
...
...
@@ -9,7 +9,7 @@ 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. */
#if defined(PADDLE_WITH_NCCL)
#if defined(PADDLE_WITH_NCCL)
|| defined(PADDLE_WITH_RCCL)
#include <float.h>
#include "paddle/fluid/framework/device_worker.h"
#include "paddle/fluid/framework/executor_gc_helper.h"
...
...
@@ -38,7 +38,7 @@ void SectionWorker::TrainFiles() {
std
::
unique_ptr
<
GarbageCollector
>
gc
;
auto
unused_vars_
=
GetUnusedVars
(
program_
->
Block
(
0
),
ops_
,
skip_vars_
);
if
(
max_memory_size
>=
0
)
{
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
if
(
platform
::
is_gpu_place
(
place_
))
{
if
(
IsFastEagerDeletionModeEnabled
())
{
gc
.
reset
(
new
UnsafeFastGPUGarbageCollector
(
...
...
@@ -70,7 +70,11 @@ void SectionWorker::TrainFiles() {
}
}
}
#ifdef PADDLE_WITH_RCCL
hipDeviceSynchronize
();
#else
cudaDeviceSynchronize
();
#endif
}
// backward pass
...
...
@@ -89,7 +93,11 @@ void SectionWorker::TrainFiles() {
}
}
}
#ifdef PADDLE_WITH_RCCL
hipDeviceSynchronize
();
#else
cudaDeviceSynchronize
();
#endif
}
// update pass
...
...
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