Skip to content
项目
群组
代码片段
帮助
当前项目
正在载入...
登录 / 注册
切换导航面板
P
pytensor
项目
项目
详情
活动
周期分析
仓库
仓库
文件
提交
分支
标签
贡献者
图表
比较
统计图
议题
0
议题
0
列表
看板
标记
里程碑
合并请求
0
合并请求
0
CI / CD
CI / CD
流水线
作业
日程
统计图
Wiki
Wiki
代码片段
代码片段
成员
成员
折叠边栏
关闭边栏
活动
图像
聊天
创建新问题
作业
提交
问题看板
Open sidebar
testgroup
pytensor
Commits
a2d05adc
提交
a2d05adc
authored
9月 21, 2022
作者:
Brandon T. Willard
提交者:
Brandon T. Willard
10月 04, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Do not overwrite arguments in Numba's Scan implementation
上级
205b7a84
隐藏空白字符变更
内嵌
并排
正在显示
1 个修改的文件
包含
30 行增加
和
26 行删除
+30
-26
scan.py
aesara/link/numba/dispatch/scan.py
+30
-26
没有找到文件。
aesara/link/numba/dispatch/scan.py
浏览文件 @
a2d05adc
from
textwrap
import
dedent
,
indent
import
numpy
as
np
import
numpy
as
np
from
numba
import
types
from
numba
import
types
from
numba.extending
import
overload
from
numba.extending
import
overload
...
@@ -72,11 +74,10 @@ def numba_funcify_Scan(op, node, **kwargs):
...
@@ -72,11 +74,10 @@ def numba_funcify_Scan(op, node, **kwargs):
allocate_mem_to_nit_sot
=
""
allocate_mem_to_nit_sot
=
""
for
_name
in
outer_in_seqs_names
:
for
_name
in
outer_in_seqs_names
:
# A sequence with multiple taps is provided as multiple modified
# A sequence with multiple taps is provided as multiple modified input
# input sequences to the Scan Op sliced appropriately
# sequences--all sliced so as to keep following the logic of a normal
# to keep following the logic of a normal sequence.
# sequence.
index
=
"[i]"
inner_in_indexed
.
append
(
f
"{_name}[i]"
)
inner_in_indexed
.
append
(
_name
+
index
)
name_to_input_map
=
dict
(
zip
(
input_names
,
node
.
inputs
[
1
:]))
name_to_input_map
=
dict
(
zip
(
input_names
,
node
.
inputs
[
1
:]))
mit_sot_name_to_taps
=
dict
(
zip
(
outer_in_mit_sot_names
,
mit_sot_in_taps
))
mit_sot_name_to_taps
=
dict
(
zip
(
outer_in_mit_sot_names
,
mit_sot_in_taps
))
...
@@ -88,31 +89,34 @@ def numba_funcify_Scan(op, node, **kwargs):
...
@@ -88,31 +89,34 @@ def numba_funcify_Scan(op, node, **kwargs):
for
_tap
in
curr_taps
:
for
_tap
in
curr_taps
:
index
=
idx_to_str
(
_tap
-
min_tap
)
index
=
idx_to_str
(
_tap
-
min_tap
)
inner_in_indexed
.
append
(
_name
+
index
)
inner_in_indexed
.
append
(
f
"{_name}{index}"
)
inner_out_name_to_index
[
_name
]
=
-
min_tap
inner_out_name_to_index
[
_name
]
=
-
min_tap
if
_name
in
outer_in_sit_sot_names
:
if
_name
in
outer_in_sit_sot_names
:
# Note that the outputs with single taps which are not
# Note that the outputs with single, non-`-1` taps are (e.g. `taps
# -1 are (for instance taps = [-2]) are classified
# = [-2]`) are classified as mit-sot, so the code for handling
# as mit-sot so the code for handling sit-sots remains
# sit-sots remains constant as follows
# constant as follows
inner_in_indexed
.
append
(
f
"{_name}[i]"
)
index
=
"[i]"
inner_in_indexed
.
append
(
_name
+
index
)
inner_out_name_to_index
[
_name
]
=
1
inner_out_name_to_index
[
_name
]
=
1
if
_name
in
outer_in_nit_sot_names
:
if
_name
in
outer_in_nit_sot_names
:
inner_out_name_to_index
[
_name
]
=
0
output_name
=
f
"{_name}_nitsot_storage"
# In case of nit-sots we are provided shape of the array
inner_out_name_to_index
[
output_name
]
=
0
# instead of actual arrays like other cases, hence we
# In case of nit-sots we are provided the shape of the array
# allocate space for the results accordingly.
# instead of actual arrays (like other cases), hence we allocate
# space for the results accordingly.
curr_nit_sot_position
=
input_names
.
index
(
_name
)
-
n_seqs
curr_nit_sot_position
=
input_names
.
index
(
_name
)
-
n_seqs
curr_nit_sot
=
inner_fg
.
outputs
[
curr_nit_sot_position
]
curr_nit_sot
=
inner_fg
.
outputs
[
curr_nit_sot_position
]
mem_shape
=
[
"1"
]
*
curr_nit_sot
.
ndim
mem_shape
=
[
"1"
]
*
curr_nit_sot
.
ndim
curr_dtype
=
curr_nit_sot
.
type
.
numpy_dtype
.
name
curr_dtype
=
curr_nit_sot
.
type
.
numpy_dtype
.
name
allocate_mem_to_nit_sot
+=
f
"""
allocate_mem_to_nit_sot
+=
dedent
(
{_name} = [np.zeros(({create_arg_string(mem_shape)}), dtype=np.{curr_dtype})]*{_name}.item()
f
"""
"""
{output_name} = [
np.empty(({create_arg_string(mem_shape)},), dtype=np.{curr_dtype}) for i in range({_name}.item())
]"""
)
# The non_seqs are passed to inner function as-is
# The non_seqs are passed to inner function as-is
inner_in_indexed
+=
outer_in_non_seqs_names
inner_in_indexed
+=
outer_in_non_seqs_names
inner_out_indexed
=
[
inner_out_indexed
=
[
...
@@ -121,7 +125,7 @@ def numba_funcify_Scan(op, node, **kwargs):
...
@@ -121,7 +125,7 @@ def numba_funcify_Scan(op, node, **kwargs):
while_logic
=
""
while_logic
=
""
if
op
.
info
.
as_while
:
if
op
.
info
.
as_while
:
# The inner function will
be returning a boolean as last argument
# The inner function will
return a boolean as the last value
inner_out_indexed
.
append
(
"while_flag"
)
inner_out_indexed
.
append
(
"while_flag"
)
while_logic
+=
"""
while_logic
+=
"""
if while_flag:
if while_flag:
...
@@ -137,18 +141,18 @@ def numba_funcify_Scan(op, node, **kwargs):
...
@@ -137,18 +141,18 @@ def numba_funcify_Scan(op, node, **kwargs):
global_env
=
locals
()
global_env
=
locals
()
global_env
[
"np"
]
=
np
global_env
[
"np"
]
=
np
output_names
=
outer_in_mit_sot_names
+
outer_in_sit_sot_names
output_names
+=
[
f
"{n}_nitsot_storage"
for
n
in
outer_in_nit_sot_names
]
scan_op_src
=
f
"""
scan_op_src
=
f
"""
def scan(n_steps, {", ".join(input_names)}):
def scan(n_steps, {", ".join(input_names)}):
{allocate_mem_to_nit_sot}
{indent(allocate_mem_to_nit_sot, " " * 4)}
for i in range(n_steps):
for i in range(n_steps):
inner_args = {create_tuple_string(inner_in_indexed)}
inner_args = {create_tuple_string(inner_in_indexed)}
{create_tuple_string(inner_out_indexed)} = numba_at_inner_func(*inner_args)
{create_tuple_string(inner_out_indexed)} = numba_at_inner_func(*inner_args)
{while_logic}
{while_logic}
return {create_arg_string(
return {create_arg_string(output_names)}
outer_in_mit_sot_names +
outer_in_sit_sot_names +
outer_in_nit_sot_names
)}
"""
"""
scalar_op_fn
=
compile_function_src
(
scalar_op_fn
=
compile_function_src
(
scan_op_src
,
"scan"
,
{
**
globals
(),
**
global_env
}
scan_op_src
,
"scan"
,
{
**
globals
(),
**
global_env
}
...
...
编写
预览
Markdown
格式
0%
重试
或
添加新文件
添加附件
取消
您添加了
0
人
到此讨论。请谨慎行事。
请先完成此评论的编辑!
取消
请
注册
或者
登录
后发表评论