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
124f1df4
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 >>
提交
124f1df4
编写于
3月 19, 2019
作者:
Z
zhhsplendid
浏览文件
操作
浏览文件
下载
差异文件
Add flags for init and re-alloc gpu
test=develop
上级
22715487
d3acf680
变更
12
隐藏空白更改
内联
并排
Showing
12 changed file
with
296 addition
and
72 deletion
+296
-72
paddle/fluid/memory/allocation/CMakeLists.txt
paddle/fluid/memory/allocation/CMakeLists.txt
+3
-1
paddle/fluid/memory/allocation/allocator_facade_abs_flags_test.cc
...luid/memory/allocation/allocator_facade_abs_flags_test.cc
+7
-17
paddle/fluid/memory/allocation/allocator_facade_frac_flags_test.cc
...uid/memory/allocation/allocator_facade_frac_flags_test.cc
+94
-0
paddle/fluid/memory/allocation/legacy_allocator.cc
paddle/fluid/memory/allocation/legacy_allocator.cc
+8
-8
paddle/fluid/memory/detail/CMakeLists.txt
paddle/fluid/memory/detail/CMakeLists.txt
+2
-0
paddle/fluid/memory/detail/buddy_allocator.cc
paddle/fluid/memory/detail/buddy_allocator.cc
+14
-12
paddle/fluid/memory/detail/buddy_allocator.h
paddle/fluid/memory/detail/buddy_allocator.h
+1
-1
paddle/fluid/memory/detail/buddy_allocator_test.cc
paddle/fluid/memory/detail/buddy_allocator_test.cc
+133
-0
paddle/fluid/memory/detail/system_allocator.cc
paddle/fluid/memory/detail/system_allocator.cc
+8
-8
paddle/fluid/platform/gpu_info.cc
paddle/fluid/platform/gpu_info.cc
+22
-21
paddle/testing/paddle_gtest_main.cc
paddle/testing/paddle_gtest_main.cc
+2
-2
python/paddle/fluid/__init__.py
python/paddle/fluid/__init__.py
+2
-2
未找到文件。
paddle/fluid/memory/allocation/CMakeLists.txt
浏览文件 @
124f1df4
...
...
@@ -61,4 +61,6 @@ nv_test(allocation_and_eigen_test SRCS allocation_and_eigen_test.cu DEPS allocat
cc_test
(
retry_allocator_test SRCS retry_allocator_test.cc DEPS retry_allocator best_fit_allocator locked_allocator cpu_allocator
)
cc_test
(
allocator_facade_test SRCS allocator_facade_test.cc DEPS allocator_facade
)
cc_test
(
allocator_facade_abs_flags_test SRCS allocator_facade_abs_flags_test.cc DEPS allocator_facade
)
cc_test
(
allocator_facade_frac_flags_test SRCS allocator_facade_frac_flags_test.cc DEPS allocator_facade
)
paddle/fluid/memory/allocation/allocator_facade_test.cc
→
paddle/fluid/memory/allocation/allocator_facade_
abs_flags_
test.cc
浏览文件 @
124f1df4
...
...
@@ -19,8 +19,8 @@
#ifdef PADDLE_WITH_CUDA
DECLARE_double
(
fraction_of_gpu_memory_to_use
);
DECLARE_double
(
fraction_of_cuda_pinned_memory_to_use
);
DECLARE_uint64
(
gpu_init
_memory_in_mb
);
DECLARE_uint64
(
gpu_reallocate
_memory_in_mb
);
DECLARE_uint64
(
initial_gpu
_memory_in_mb
);
DECLARE_uint64
(
reallocate_gpu
_memory_in_mb
);
DECLARE_int64
(
gpu_allocator_retry_time
);
#endif
...
...
@@ -79,25 +79,15 @@ void AllocateTestCases() {
#endif
}
TEST
(
allocator
,
allocator
)
{
TEST
(
Allocator
,
SpecifyGpuMemory
)
{
#ifdef PADDLE_WITH_CUDA
FLAGS_fraction_of_gpu_memory_to_use
=
0.01
;
FLAGS_gpu_allocator_retry_time
=
500
;
FLAGS_fraction_of_cuda_pinned_memory_to_use
=
0.5
;
#endif
AllocateTestCases
();
}
TEST
(
allocator
,
specify_gpu_memory
)
{
#ifdef PADDLE_WITH_CUDA
// Set to 0.0 to test FLAGS_gpu_init_memory_in_mb and
// FLAGS_gpu_reallocate_memory_in_mb
// Set to 0.0 to test FLAGS_initial_gpu_memory_in_mb and
// FLAGS_reallocate_gpu_memory_in_mb
FLAGS_fraction_of_gpu_memory_to_use
=
0.0
;
// 512 MB
FLAGS_
gpu_init
_memory_in_mb
=
512
;
FLAGS_
initial_gpu
_memory_in_mb
=
512
;
// 4 MB
FLAGS_
gpu_reallocate
_memory_in_mb
=
4
;
FLAGS_
reallocate_gpu
_memory_in_mb
=
4
;
FLAGS_gpu_allocator_retry_time
=
500
;
FLAGS_fraction_of_cuda_pinned_memory_to_use
=
0.5
;
#endif
...
...
paddle/fluid/memory/allocation/allocator_facade_frac_flags_test.cc
0 → 100644
浏览文件 @
124f1df4
// 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/memory/allocation/allocator_facade.h"
#include <gflags/gflags.h>
#include <gtest/gtest.h>
#ifdef PADDLE_WITH_CUDA
DECLARE_double
(
fraction_of_gpu_memory_to_use
);
DECLARE_double
(
fraction_of_cuda_pinned_memory_to_use
);
DECLARE_uint64
(
initial_gpu_memory_in_mb
);
DECLARE_uint64
(
reallocate_gpu_memory_in_mb
);
DECLARE_int64
(
gpu_allocator_retry_time
);
#endif
namespace
paddle
{
namespace
memory
{
namespace
allocation
{
//! Run allocate test cases for different places
void
AllocateTestCases
()
{
auto
&
instance
=
AllocatorFacade
::
Instance
();
platform
::
Place
place
;
size_t
size
=
1024
;
{
place
=
platform
::
CPUPlace
();
size
=
1024
;
auto
cpu_allocation
=
instance
.
Alloc
(
place
,
size
);
ASSERT_NE
(
cpu_allocation
,
nullptr
);
ASSERT_NE
(
cpu_allocation
->
ptr
(),
nullptr
);
ASSERT_EQ
(
cpu_allocation
->
place
(),
place
);
ASSERT_EQ
(
cpu_allocation
->
size
(),
size
);
}
#ifdef PADDLE_WITH_CUDA
{
place
=
platform
::
CUDAPlace
(
0
);
size
=
1024
;
auto
gpu_allocation
=
instance
.
Alloc
(
place
,
size
);
ASSERT_NE
(
gpu_allocation
,
nullptr
);
ASSERT_NE
(
gpu_allocation
->
ptr
(),
nullptr
);
ASSERT_EQ
(
gpu_allocation
->
place
(),
place
);
ASSERT_GE
(
gpu_allocation
->
size
(),
size
);
}
{
// Allocate 2GB gpu memory
place
=
platform
::
CUDAPlace
(
0
);
size
=
2
*
static_cast
<
size_t
>
(
1
<<
30
);
auto
gpu_allocation
=
instance
.
Alloc
(
place
,
size
);
ASSERT_NE
(
gpu_allocation
,
nullptr
);
ASSERT_NE
(
gpu_allocation
->
ptr
(),
nullptr
);
ASSERT_EQ
(
gpu_allocation
->
place
(),
place
);
ASSERT_GE
(
gpu_allocation
->
size
(),
size
);
}
{
place
=
platform
::
CUDAPinnedPlace
();
size
=
(
1
<<
20
);
auto
cuda_pinned_allocation
=
instance
.
Alloc
(
platform
::
CUDAPinnedPlace
(),
1
<<
20
);
ASSERT_NE
(
cuda_pinned_allocation
,
nullptr
);
ASSERT_NE
(
cuda_pinned_allocation
->
ptr
(),
nullptr
);
ASSERT_EQ
(
cuda_pinned_allocation
->
place
(),
place
);
ASSERT_GE
(
cuda_pinned_allocation
->
size
(),
size
);
}
#endif
}
TEST
(
Allocator
,
Allocator
)
{
#ifdef PADDLE_WITH_CUDA
FLAGS_fraction_of_gpu_memory_to_use
=
0.01
;
FLAGS_gpu_allocator_retry_time
=
500
;
FLAGS_fraction_of_cuda_pinned_memory_to_use
=
0.5
;
#endif
AllocateTestCases
();
}
}
// namespace allocation
}
// namespace memory
}
// namespace paddle
paddle/fluid/memory/allocation/legacy_allocator.cc
浏览文件 @
124f1df4
...
...
@@ -37,8 +37,8 @@ DEFINE_bool(init_allocated_mem, false,
"that initializing the allocated memory with a small value "
"during unit testing."
);
DECLARE_double
(
fraction_of_gpu_memory_to_use
);
DECLARE_uint64
(
gpu_init
_memory_in_mb
);
DECLARE_uint64
(
gpu_reallocate
_memory_in_mb
);
DECLARE_uint64
(
initial_gpu
_memory_in_mb
);
DECLARE_uint64
(
reallocate_gpu
_memory_in_mb
);
DECLARE_bool
(
benchmark
);
namespace
paddle
{
...
...
@@ -158,15 +158,15 @@ BuddyAllocator *GetGPUBuddyAllocator(int gpu_id) {
VLOG
(
10
)
<<
"
\n\n
NOTE:
\n
"
<<
"You can set GFlags environment variable "
<<
"'FLAGS_fraction_of_gpu_memory_to_use' "
<<
"or 'FLAGS_
gpu_init
_memory_in_mb' "
<<
"or 'FLAGS_
gpu_reallocate
_memory_in_mb' "
<<
"or 'FLAGS_
initial_gpu
_memory_in_mb' "
<<
"or 'FLAGS_
reallocate_gpu
_memory_in_mb' "
<<
"to change the memory size for GPU usage.
\n
"
<<
"Current 'FLAGS_fraction_of_gpu_memory_to_use' value is "
<<
FLAGS_fraction_of_gpu_memory_to_use
<<
". Current 'FLAGS_
gpu_init
_memory_in_mb' value is "
<<
FLAGS_
gpu_init
_memory_in_mb
<<
". Current 'FLAGS_
gpu_reallocate
_memory_in_mb' value is "
<<
FLAGS_
gpu_reallocate
_memory_in_mb
<<
"
\n\n
"
;
<<
". Current 'FLAGS_
initial_gpu
_memory_in_mb' value is "
<<
FLAGS_
initial_gpu
_memory_in_mb
<<
". Current 'FLAGS_
reallocate_gpu
_memory_in_mb' value is "
<<
FLAGS_
reallocate_gpu
_memory_in_mb
<<
"
\n\n
"
;
}
});
...
...
paddle/fluid/memory/detail/CMakeLists.txt
浏览文件 @
124f1df4
...
...
@@ -9,3 +9,5 @@ endif(${WITH_GPU})
cc_test
(
system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator
)
cc_library
(
buddy_allocator SRCS buddy_allocator.cc DEPS memory_block system_allocator glog
)
cc_test
(
buddy_allocator_test SRCS buddy_allocator_test.cc DEPS buddy_allocator
)
paddle/fluid/memory/detail/buddy_allocator.cc
浏览文件 @
124f1df4
...
...
@@ -14,7 +14,8 @@ limitations under the License. */
#include "paddle/fluid/memory/detail/buddy_allocator.h"
#include <utility> // for std::move
#include <algorithm>
#include <utility>
#include "glog/logging.h"
...
...
@@ -39,9 +40,10 @@ BuddyAllocator::~BuddyAllocator() {
"have actually been freed"
;
while
(
!
pool_
.
empty
())
{
auto
block
=
static_cast
<
MemoryBlock
*>
(
std
::
get
<
2
>
(
*
pool_
.
begin
()));
VLOG
(
10
)
<<
"Free from block ("
<<
block
<<
", "
<<
max_chunk_size_
<<
")"
;
VLOG
(
10
)
<<
"Free from block ("
<<
block
<<
", "
<<
block
->
size
(
cache_
)
<<
")"
;
system_allocator_
->
Free
(
block
,
max_chunk_size_
,
block
->
index
(
cache_
));
system_allocator_
->
Free
(
block
,
block
->
size
(
cache_
)
,
block
->
index
(
cache_
));
cache_
.
invalidate
(
block
);
pool_
.
erase
(
pool_
.
begin
());
}
...
...
@@ -74,7 +76,7 @@ void* BuddyAllocator::Alloc(size_t unaligned_size) {
// refill the pool if failure
if
(
it
==
pool_
.
end
())
{
it
=
RefillPool
();
it
=
RefillPool
(
size
);
// if still failure, fail fatally
if
(
it
==
pool_
.
end
())
{
return
nullptr
;
...
...
@@ -187,7 +189,8 @@ void* BuddyAllocator::SystemAlloc(size_t size) {
return
static_cast
<
MemoryBlock
*>
(
p
)
->
data
();
}
BuddyAllocator
::
PoolSet
::
iterator
BuddyAllocator
::
RefillPool
()
{
BuddyAllocator
::
PoolSet
::
iterator
BuddyAllocator
::
RefillPool
(
size_t
request_bytes
)
{
size_t
allocate_bytes
=
max_chunk_size_
;
size_t
index
=
0
;
...
...
@@ -195,14 +198,13 @@ BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() {
if
(
system_allocator_
->
UseGpu
())
{
if
((
total_used_
+
total_free_
)
==
0
)
{
// Compute the allocation size for gpu for the first allocation.
max_chunk_size_
=
platform
::
GpuMaxChunkSize
();
allocate_bytes
=
platform
::
GpuInitAllocSize
();
allocate_bytes
=
std
::
max
(
platform
::
GpuInitAllocSize
(),
request_bytes
);
}
else
{
// Reallocation size
if
(
realloc_size_
==
0
)
{
realloc_size_
=
platform
::
GpuReallocSize
();
}
allocate_bytes
=
realloc_size_
;
allocate_bytes
=
std
::
max
(
realloc_size_
,
request_bytes
)
;
}
}
#endif
...
...
@@ -298,12 +300,12 @@ void BuddyAllocator::CleanIdleFallBackAlloc() {
VLOG
(
10
)
<<
"Return block "
<<
block
<<
" to fallback allocator."
;
system_allocator_
->
Free
(
block
,
max_chunk_size_
,
block
->
index
(
cache_
));
system_allocator_
->
Free
(
block
,
block
->
size
(
cache_
)
,
block
->
index
(
cache_
));
cache_
.
invalidate
(
block
);
pool
=
PoolSet
::
reverse_iterator
(
pool_
.
erase
(
std
::
next
(
pool
).
base
()));
total_free_
-=
max_chunk_size_
;
total_free_
-=
block
->
size
(
cache_
)
;
fallback_alloc_count_
--
;
// If no fall allocation exists, return directly
...
...
@@ -334,12 +336,12 @@ void BuddyAllocator::CleanIdleNormalAlloc() {
VLOG
(
10
)
<<
"Return block "
<<
block
<<
" to base allocator."
;
system_allocator_
->
Free
(
block
,
max_chunk_size_
,
block
->
index
(
cache_
));
system_allocator_
->
Free
(
block
,
block
->
size
(
cache_
)
,
block
->
index
(
cache_
));
cache_
.
invalidate
(
block
);
pool
=
PoolSet
::
reverse_iterator
(
pool_
.
erase
(
std
::
next
(
pool
).
base
()));
total_free_
-=
max_chunk_size_
;
total_free_
-=
block
->
size
(
cache_
)
;
if
(
!
shall_free_alloc
())
return
;
}
...
...
paddle/fluid/memory/detail/buddy_allocator.h
浏览文件 @
124f1df4
...
...
@@ -60,7 +60,7 @@ class BuddyAllocator {
void
*
SystemAlloc
(
size_t
size
);
/*! \brief If existing chunks are not suitable, refill pool */
PoolSet
::
iterator
RefillPool
();
PoolSet
::
iterator
RefillPool
(
size_t
request_bytes
);
/**
* \brief Find the suitable chunk from existing pool and split
...
...
paddle/fluid/memory/detail/buddy_allocator_test.cc
0 → 100644
浏览文件 @
124f1df4
/* Copyright (c) 2016 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/memory/detail/buddy_allocator.h"
#include <memory>
#include "gflags/gflags.h"
#include "gtest/gtest.h"
#include "paddle/fluid/memory/detail/system_allocator.h"
#include "paddle/fluid/platform/gpu_info.h"
#ifdef PADDLE_WITH_CUDA
DECLARE_double
(
fraction_of_gpu_memory_to_use
);
DECLARE_uint64
(
initial_gpu_memory_in_mb
);
DECLARE_uint64
(
reallocate_gpu_memory_in_mb
);
#endif
namespace
paddle
{
namespace
memory
{
namespace
detail
{
constexpr
static
int
test_gpu_id
=
0
;
void
TestBuddyAllocator
(
BuddyAllocator
*
allocator
,
size_t
size_bytes
)
{
bool
freed
=
false
;
size_t
used_bytes
=
allocator
->
Used
();
if
(
size_bytes
>
0
)
{
void
*
p
=
allocator
->
Alloc
(
size_bytes
);
EXPECT_NE
(
p
,
nullptr
);
#ifdef PADDLE_WITH_CUDA
if
(
size_bytes
<
platform
::
GpuMaxChunkSize
())
{
#else
if
(
size_bytes
<
platform
::
CpuMaxChunkSize
())
{
#endif
// Not allocate from SystemAllocator
EXPECT_GE
(
allocator
->
Used
(),
used_bytes
+
size_bytes
);
}
else
{
// Allocate from SystemAllocator doesn't count in Used()
EXPECT_EQ
(
allocator
->
Used
(),
used_bytes
);
}
int
*
intp
=
static_cast
<
int
*>
(
p
);
std
::
shared_ptr
<
int
>
ptr
(
intp
,
[
&
](
void
*
p
)
{
allocator
->
Free
(
intp
);
freed
=
true
;
});
}
else
{
freed
=
true
;
}
EXPECT_EQ
(
used_bytes
,
allocator
->
Used
());
EXPECT_TRUE
(
freed
);
}
#ifdef PADDLE_WITH_CUDA
TEST
(
BuddyAllocator
,
GpuFraction
)
{
FLAGS_fraction_of_gpu_memory_to_use
=
0.01
;
BuddyAllocator
buddy_allocator
(
std
::
unique_ptr
<
SystemAllocator
>
(
new
GPUAllocator
(
test_gpu_id
)),
platform
::
GpuMinChunkSize
(),
platform
::
GpuMaxChunkSize
());
TestBuddyAllocator
(
&
buddy_allocator
,
10
);
TestBuddyAllocator
(
&
buddy_allocator
,
10
<<
10
);
TestBuddyAllocator
(
&
buddy_allocator
,
10
<<
20
);
TestBuddyAllocator
(
&
buddy_allocator
,
2
*
static_cast
<
size_t
>
(
1
<<
30
));
}
TEST
(
BuddyAllocator
,
InitRealloc
)
{
FLAGS_initial_gpu_memory_in_mb
=
100
;
FLAGS_reallocate_gpu_memory_in_mb
=
50
;
EXPECT_EQ
(
platform
::
GpuMaxChunkSize
(),
static_cast
<
size_t
>
(
100
<<
20
));
BuddyAllocator
buddy_allocator
(
std
::
unique_ptr
<
SystemAllocator
>
(
new
GPUAllocator
(
test_gpu_id
)),
platform
::
GpuMinChunkSize
(),
platform
::
GpuMaxChunkSize
());
// Less then initial size and reallocate size
TestBuddyAllocator
(
&
buddy_allocator
,
10
<<
20
);
// Between initial size and reallocate size and not exceed pool
TestBuddyAllocator
(
&
buddy_allocator
,
80
<<
20
);
// Less then reallocate size and exceed pool
TestBuddyAllocator
(
&
buddy_allocator
,
40
<<
20
);
// Greater then reallocate size and exceed pool
TestBuddyAllocator
(
&
buddy_allocator
,
80
<<
20
);
// Greater then initial size and reallocate size
TestBuddyAllocator
(
&
buddy_allocator
,
2
*
static_cast
<
size_t
>
(
1
<<
30
));
}
TEST
(
BuddyAllocator
,
ReallocSizeGreaterThanInit
)
{
FLAGS_initial_gpu_memory_in_mb
=
5
;
FLAGS_reallocate_gpu_memory_in_mb
=
10
;
EXPECT_EQ
(
platform
::
GpuMaxChunkSize
(),
static_cast
<
size_t
>
(
10
<<
20
));
BuddyAllocator
buddy_allocator
(
std
::
unique_ptr
<
SystemAllocator
>
(
new
GPUAllocator
(
test_gpu_id
)),
platform
::
GpuMinChunkSize
(),
platform
::
GpuMaxChunkSize
());
// Less then initial size and reallocate size
TestBuddyAllocator
(
&
buddy_allocator
,
1
<<
20
);
// Between initial size and reallocate size and not exceed pool
TestBuddyAllocator
(
&
buddy_allocator
,
3
<<
20
);
// Less then initial size and exceed pool
TestBuddyAllocator
(
&
buddy_allocator
,
3
<<
20
);
// Less then reallocate size and not exceed pool (now pool is 15 MB, used 7
// MB)
TestBuddyAllocator
(
&
buddy_allocator
,
7
<<
20
);
// Less then reallocate size and exceed pool
TestBuddyAllocator
(
&
buddy_allocator
,
8
<<
20
);
// Greater then initial size and reallocate size
TestBuddyAllocator
(
&
buddy_allocator
,
2
*
static_cast
<
size_t
>
(
1
<<
30
));
}
#endif
}
// namespace detail
}
// namespace memory
}
// namespace paddle
paddle/fluid/memory/detail/system_allocator.cc
浏览文件 @
124f1df4
...
...
@@ -32,8 +32,8 @@ limitations under the License. */
DECLARE_bool
(
use_pinned_memory
);
DECLARE_double
(
fraction_of_gpu_memory_to_use
);
DECLARE_uint64
(
gpu_init
_memory_in_mb
);
DECLARE_uint64
(
gpu_reallocate
_memory_in_mb
);
DECLARE_uint64
(
initial_gpu
_memory_in_mb
);
DECLARE_uint64
(
reallocate_gpu
_memory_in_mb
);
namespace
paddle
{
namespace
memory
{
...
...
@@ -125,15 +125,15 @@ void* GPUAllocator::Alloc(size_t* index, size_t size) {
LOG
(
WARNING
)
<<
"Cannot malloc "
<<
size
/
1024.0
/
1024.0
<<
" MB GPU memory. Please shrink "
"FLAGS_fraction_of_gpu_memory_to_use or "
"FLAGS_
gpu_init
_memory_in_mb or "
"FLAGS_
gpu_reallocate
_memory_in_mb"
"FLAGS_
initial_gpu
_memory_in_mb or "
"FLAGS_
reallocate_gpu
_memory_in_mb"
"environment variable to a lower value. "
<<
"Current FLAGS_fraction_of_gpu_memory_to_use value is "
<<
FLAGS_fraction_of_gpu_memory_to_use
<<
". Current FLAGS_
gpu_init
_memory_in_mb value is "
<<
FLAGS_
gpu_init
_memory_in_mb
<<
". Current FLAGS_
gpu_reallocate
_memory_in_mb value is "
<<
FLAGS_
gpu_reallocate
_memory_in_mb
;
<<
". Current FLAGS_
initial_gpu
_memory_in_mb value is "
<<
FLAGS_
initial_gpu
_memory_in_mb
<<
". Current FLAGS_
reallocate_gpu
_memory_in_mb value is "
<<
FLAGS_
reallocate_gpu
_memory_in_mb
;
return
nullptr
;
}
}
...
...
paddle/fluid/platform/gpu_info.cc
浏览文件 @
124f1df4
...
...
@@ -39,19 +39,20 @@ DEFINE_double(fraction_of_gpu_memory_to_use, fraction_of_gpu_memory_to_use,
"additional trunks of the same size will be requested from gpu "
"until the gpu has no memory left for another trunk."
);
DEFINE_uint64
(
gpu_init_memory_in_mb
,
0ul
,
"Allocate a trunk of gpu memory whose byte size is specified by "
"the flag. Future memory usage will be allocated from the "
"truck. If the trunk doesn't have enough gpu memory, additional "
"trunks of the gpu memory will be requested from gpu with size "
"speified by FLAGS_gpu_reallocate_memory_in_mb until the gpu has "
"no memory left for the additional trunk. Note: if you set this "
"flag, the memory size set by "
"FLAGS_fraction_of_gpu_memory_to_use will be overrided by this "
"flag. If you don't set this flag, PaddlePaddle will use "
"FLAGS_fraction_of_gpu_memory_to_use to allocate gpu memory"
);
DEFINE_uint64
(
gpu_reallocate_memory_in_mb
,
0ul
,
DEFINE_uint64
(
initial_gpu_memory_in_mb
,
0ul
,
"Allocate a trunk of gpu memory whose byte size is specified by "
"the flag. Future memory usage will be allocated from the "
"truck. If the trunk doesn't have enough gpu memory, additional "
"trunks of the gpu memory will be requested from gpu with size "
"specified by FLAGS_reallocate_gpu_memory_in_mb until the gpu has "
"no memory left for the additional trunk. Note: if you set this "
"flag, the memory size set by "
"FLAGS_fraction_of_gpu_memory_to_use will be overrided by this "
"flag. If you don't set this flag, PaddlePaddle will use "
"FLAGS_fraction_of_gpu_memory_to_use to allocate gpu memory"
);
DEFINE_uint64
(
reallocate_gpu_memory_in_mb
,
0ul
,
"If this flag is set, Paddle will reallocate the gpu memory with "
"size specified by this flag. Else Paddle will reallocate by "
"FLAGS_fraction_of_gpu_memory_to_use"
);
...
...
@@ -202,12 +203,12 @@ size_t GpuMaxAllocSize() {
}
size_t
GpuInitAllocSize
()
{
if
(
FLAGS_
gpu_init
_memory_in_mb
>
0ul
)
{
// Initial memory will be allocated by FLAGS_
gpu_init
_memory_in_mb
return
static_cast
<
size_t
>
(
FLAGS_
gpu_init
_memory_in_mb
<<
20
);
if
(
FLAGS_
initial_gpu
_memory_in_mb
>
0ul
)
{
// Initial memory will be allocated by FLAGS_
initial_gpu
_memory_in_mb
return
static_cast
<
size_t
>
(
FLAGS_
initial_gpu
_memory_in_mb
<<
20
);
}
// FLAGS_
gpu_init
_memory_in_mb is 0, initial memory will be allocated by
// FLAGS_
initial_gpu
_memory_in_mb is 0, initial memory will be allocated by
// fraction
size_t
total
=
0
;
size_t
available
=
0
;
...
...
@@ -220,12 +221,12 @@ size_t GpuInitAllocSize() {
}
size_t
GpuReallocSize
()
{
if
(
FLAGS_
gpu_reallocate
_memory_in_mb
>
0ul
)
{
// Additional memory will be allocated by FLAGS_
gpu_reallocate
_memory_in_mb
return
static_cast
<
size_t
>
(
FLAGS_
gpu_reallocate
_memory_in_mb
<<
20
);
if
(
FLAGS_
reallocate_gpu
_memory_in_mb
>
0ul
)
{
// Additional memory will be allocated by FLAGS_
reallocate_gpu
_memory_in_mb
return
static_cast
<
size_t
>
(
FLAGS_
reallocate_gpu
_memory_in_mb
<<
20
);
}
// FLAGS_
gpu_reallocate
_memory_in_mb is 0, additional memory will be allocated
// FLAGS_
reallocate_gpu
_memory_in_mb is 0, additional memory will be allocated
// by fraction
size_t
total
=
0
;
size_t
available
=
0
;
...
...
paddle/testing/paddle_gtest_main.cc
浏览文件 @
124f1df4
...
...
@@ -41,8 +41,8 @@ int main(int argc, char** argv) {
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
envs
.
push_back
(
"fraction_of_gpu_memory_to_use"
);
envs
.
push_back
(
"
gpu_init
_memory_in_mb"
);
envs
.
push_back
(
"
gpu_reallocate_memory_bytes
"
);
envs
.
push_back
(
"
initial_gpu
_memory_in_mb"
);
envs
.
push_back
(
"
reallocate_gpu_memory_in_mb
"
);
envs
.
push_back
(
"allocator_strategy"
);
#elif __clang__
envs
.
push_back
(
"use_mkldnn"
);
...
...
python/paddle/fluid/__init__.py
浏览文件 @
124f1df4
...
...
@@ -162,8 +162,8 @@ def __bootstrap__():
if
core
.
is_compiled_with_cuda
():
read_env_flags
+=
[
'fraction_of_gpu_memory_to_use'
,
'
gpu_init
_memory_in_mb'
,
'
gpu_reallocate
_memory_in_mb'
,
'cudnn_deterministic'
,
'fraction_of_gpu_memory_to_use'
,
'
initial_gpu
_memory_in_mb'
,
'
reallocate_gpu
_memory_in_mb'
,
'cudnn_deterministic'
,
'enable_cublas_tensor_op_math'
,
'conv_workspace_size_limit'
,
'cudnn_exhaustive_search'
,
'memory_optimize_debug'
,
'selected_gpus'
,
'sync_nccl_allreduce'
,
'limit_of_tmp_allocation'
,
...
...
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