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
Xiaomi
Mace
提交
9ba7bce9
Mace
项目概览
Xiaomi
/
Mace
通知
106
Star
40
Fork
27
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
列表
看板
标记
里程碑
合并请求
0
DevOps
流水线
流水线任务
计划
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
Mace
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
0
Issue
0
列表
看板
标记
里程碑
合并请求
0
合并请求
0
Pages
DevOps
DevOps
流水线
流水线任务
计划
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
流水线任务
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
提交
9ba7bce9
编写于
3月 16, 2018
作者:
刘
刘琦
浏览文件
操作
浏览文件
下载
差异文件
Merge branch 'gpu-mem' into 'master'
Add more sophisticated mem optimization See merge request !288
上级
f0b55c29
75bca065
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
26 addition
and
11 deletion
+26
-11
mace/python/tools/memory_optimizer.py
mace/python/tools/memory_optimizer.py
+26
-11
未找到文件。
mace/python/tools/memory_optimizer.py
浏览文件 @
9ba7bce9
...
...
@@ -43,6 +43,9 @@ class MemoryOptimizer(object):
mem_size
[
1
]
=
output_shape
[
0
]
*
output_shape
[
1
]
return
mem_size
def
mem_area
(
self
,
memory_size
):
return
memory_size
[
0
]
*
memory_size
[
1
]
def
optimize
(
self
):
for
op
in
self
.
net_def
.
op
:
if
self
.
is_buffer_image_op
(
op
):
...
...
@@ -54,22 +57,34 @@ class MemoryOptimizer(object):
print
(
'WARNING: the number of output shape is not equal to the number of output.'
)
return
for
i
in
range
(
len
(
op
.
output
)):
if
len
(
self
.
idle_mem
)
==
0
:
# allocate new mem
op_mem_size
=
self
.
get_mem_size
(
op
.
type
,
op
.
output_shape
[
i
].
dims
)
mem_id
=
-
1
if
len
(
self
.
idle_mem
)
>
0
:
best_mem_candidate_id
=
-
1
best_mem_candidate_delta_area
=
sys
.
maxint
best_mem_candidate_shape
=
[]
for
mid
in
self
.
idle_mem
:
reuse_mem_size
=
self
.
mem_block
[
mid
]
resize_mem_size
=
[
max
(
reuse_mem_size
[
0
],
op_mem_size
[
0
]),
max
(
reuse_mem_size
[
1
],
op_mem_size
[
1
])]
delta_mem_area
=
self
.
mem_area
(
resize_mem_size
)
-
self
.
mem_area
(
reuse_mem_size
)
if
delta_mem_area
<
best_mem_candidate_delta_area
:
best_mem_candidate_id
=
mid
best_mem_candidate_delta_area
=
delta_mem_area
best_mem_candidate_shape
=
resize_mem_size
if
best_mem_candidate_delta_area
<=
self
.
mem_area
(
op_mem_size
):
# reuse
self
.
mem_block
[
best_mem_candidate_id
]
=
best_mem_candidate_shape
mem_id
=
best_mem_candidate_id
self
.
idle_mem
.
remove
(
mem_id
)
if
mem_id
==
-
1
:
mem_id
=
self
.
total_mem_count
self
.
total_mem_count
+=
1
else
:
# reuse mem
mem_id
=
self
.
idle_mem
.
pop
()
self
.
mem_block
[
mem_id
]
=
op_mem_size
op
.
mem_id
.
extend
([
mem_id
])
self
.
op_mem
[
op
.
output
[
i
]]
=
mem_id
if
mem_id
not
in
self
.
mem_block
:
self
.
mem_block
[
mem_id
]
=
[
0
,
0
]
mem_size
=
self
.
mem_block
[
mem_id
]
op_mem_size
=
self
.
get_mem_size
(
op
.
type
,
op
.
output_shape
[
i
].
dims
)
mem_size
[
0
]
=
max
(
mem_size
[
0
],
op_mem_size
[
0
])
mem_size
[
1
]
=
max
(
mem_size
[
1
],
op_mem_size
[
1
])
# de-ref input tensor mem
for
ipt
in
op
.
input
:
...
...
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