Skip to content

P
pytensor
提交 d3f52989 authored 作者: Pascal Lamblin's avatar Pascal Lamblin
浏览文件
操作

Refine mechanism to test different kinds of pre-allocated output storage

上级 48a14134
隐藏空白字符变更
内嵌 并排
正在显示 包含 143 行增加 和 131 行删除
theano/compile/debugmode.py
浏览文件 @ d3f52989
......@@ -11,7 +11,8 @@ from theano import gof
from theano.gof import Env, graph, utils, link
from theano.gof.link import raise_with_op
from theano.gof.cc import CLinker
from theano.configparser import config, AddConfigVar, IntParam, BoolParam
from theano.configparser import (config, AddConfigVar, BoolParam, IntParam,
StrParam)
from theano.compile.function_module import (FunctionMaker,
Function,
infer_reuse_pattern,
......@@ -47,8 +48,13 @@ AddConfigVar('DebugMode.warn_input_not_reused',
BoolParam(True))
AddConfigVar('DebugMode.check_preallocated_output',
'Test thunks with pre-allocated memory as output storage.',
BoolParam(False))
('Test thunks with pre-allocated memory as output storage. '
'This is a list of strings separated by ":". Valid values are: '
'"previous" (previously-returned memory), '
'"c_contiguous", "f_contiguous", '
'"neg_strides" (negative strides), and '
'"ALL" (all of the above).'),
StrParam(''))
import logging
_logger=logging.getLogger("theano.compile.debugmode")
......@@ -819,6 +825,109 @@ def _find_bad_optimizations2(order, reasons, r_vals):
_find_bad_optimizations = _find_bad_optimizations0
def _check_preallocated_output(node, thunk, prealloc_modes, def_val,
storage_map, r_vals, dr_vals, perform, active_order_set):
'''Try to apply thunk() on different output storages'''
# To avoid circular imports
from theano.tensor import TensorType
from theano.sandbox.cuda import cuda_available, CudaNdarrayType
if cuda_available:
from theano.sandbox.cuda import CudaNdarray
# List of (name, map) pairs of the settings to test
prealloc_maps = []
# TODO: Sparse, Scalar
# TODO: wrong shape, more stride patterns
# reuse_output: use a copy of the same storage returned the first time
# TODO: optimization warning if the storage in reuse_outputs
# is not reused
if 'previous' in prealloc_modes or 'ALL' in prealloc_modes:
reuse_outputs = {}
for r in node.outputs:
# We want to reuse the exact same memory buffer,
# so we keep the copy in r_vals
new_r = _lessbroken_deepcopy(r_vals[r])
reuse_outputs[r] = r_vals[r]
r_vals[r] = new_r
prealloc_maps.append(('previous', reuse_outputs))
# c_cont_output: use a c-continuous array
# (for TensorType and CudaNdarray, else None)
if 'c_contiguous' in prealloc_modes or 'ALL' in prealloc_modes:
c_cont_outputs = {}
for r in node.outputs:
if isinstance(r.type, (TensorType, CudaNdarrayType)):
# Build a C-contiguous buffer
new_buf = numpy.zeros(
shape=r_vals[r].shape,
dtype=r_vals[r].dtype,
order='C')
new_buf += def_val
if isinstance(r.type, CudaNdarrayType):
new_buf = CudaNdarray(new_buf)
c_cont_outputs[r] = new_buf
if len(c_cont_outputs):
prealloc_maps.append(('c_contiguous', c_cont_outputs))
# f_cont_output: use a fortran-continuous ndarray
# (for TensorType, only)
if 'f_contiguous' in prealloc_modes or 'ALL' in prealloc_modes:
f_cont_outputs = {}
for r in node.outputs:
if isinstance(r.type, TensorType):
new_buf = numpy.zeros(
shape=r_vals[r].shape,
dtype=r_vals[r].dtype,
order='F')
new_buf += def_val
f_cont_outputs[r] = new_buf
if len(f_cont_outputs):
prealloc_maps.append(('f_contiguous', f_cont_outputs))
if 'neg_strides' in prealloc_maps:
raise NotImplementedError('Negative strides in check_preallocated_output')
for (name, out_map) in prealloc_maps:
# Copy the inputs over again
for r in node.inputs:
storage_map[r][0] = _lessbroken_deepcopy(r_vals[r])
# Get the appropriate output storages
# (no copy)
for r in node.outputs:
storage_map[r][0] = out_map.get(r, None)
thunk()
# Check outputs
for r in node.outputs:
if not r.type.is_valid_value(storage_map[r][0]):
raise InvalidValueError(r, storage_map[r][0],
hint='%s with %s output' % (perform, name),
specific_hint=r.type.value_validity_msg(storage_map[r][0]))
_check_inputs(node, storage_map, r_vals, dr_vals, active_order_set,
clobber_dr_vals=False,
perform='%s with output %s' % (perform, name),
warn_input_not_reused=False)
_check_viewmap(node, storage_map)
for r in node.outputs:
if not r.type.values_eq_approx(r_vals[r], storage_map[r][0]):
# TODO: indicate it is not a C/Py problem
raise BadCLinkerOutput(r, val_py=r_vals[r], val_c=storage_map[r][0])
# Clear storage_map
for r in node.outputs:
storage_map[r][0] = None
class _EnvEvent(object):
"""A record of an event in the life of an Env.
......@@ -1029,9 +1138,7 @@ class _Linker(gof.link.LocalLinker):
#can't import at toplevel because of circular import
# TODO: don't do this ugly hacky way of setting the filter_checks_isfinite
from theano.tensor import TensorType #to set filter_check_isfinite
from theano.sandbox.cuda import cuda_available, CudaNdarrayType
if cuda_available:
from theano.sandbox.cuda import CudaNdarray
from theano import tests # for config.unittests.rseed
env = self.env
input_storage_ = input_storage
output_storage_ = output_storage
......@@ -1092,6 +1199,13 @@ class _Linker(gof.link.LocalLinker):
else:
no_recycling = [storage_map[r] for r in no_recycling if r not in env.inputs]
# Precompute some things for storage pre-allocation
prealloc_modes = config.DebugMode.check_preallocated_output.split(':')
try:
def_val = int(config.unittests.rseed)
except ValueError:
def_val = 666
#####
# This is the function that runs when you evaluate the graph
#####
......@@ -1133,14 +1247,8 @@ class _Linker(gof.link.LocalLinker):
storage_map[r][0] = None
r_vals_initialized.append(r)
# Debug Mode complains if someone provides memory buffers
# for the outputs (where the linker can choose to store the
# outputs). Since this is what scan does by default, we will
# delete the output_storage for now. This code is going to
# change when someone decides to go over the debug code
# again, and try to include checks for such behaviour as
# well.
# TODO: store them in another map, and test the thunks on
# them as output storages.
for r in storage_map:
if r in env.outputs:
storage_map[r][0] = None
......@@ -1199,68 +1307,16 @@ class _Linker(gof.link.LocalLinker):
storage_map[r][0] = None #clear the storage_map of outputs for the thunk_c
if config.DebugMode.check_preallocated_output:
## Then, try to use different output storages
# reuse_output: use a copy of the same storage returned the first time
# TODO: optimization warning if the storage in reuse_outputs
# is not reused
# c_cont_output: use a c-continuous ndarray (for TensorType, else None)
# f_cont_output: use a fortran-continuous ndarray (for TensorType, else None)
# TODO: Sparse, Scalar
# TODO: wrong shape, more stride patterns
reuse_outputs = {}
c_cont_outputs = {}
f_cont_outputs = {}
for r in node.outputs:
r_val = r_vals[r]
reuse_outputs[r] = _lessbroken_deepcopy(r_val)
if isinstance(r.type, TensorType):
c_cont_outputs[r] = numpy.empty(
shape=r_val.shape,
dtype=r_val.dtype,
order='C')
f_cont_outputs[r] = numpy.empty(
shape=r_val.shape,
dtype=r_val.dtype,
order='F')
elif isinstance(r.type, CudaNdarrayType):
# CudaNdarray supports only C-contiguous
c_cont_outputs[r] = CudaNdarray.zeros(
r_val.shape)
for out_map in (reuse_outputs, c_cont_outputs, f_cont_outputs):
if len(out_map) == 0:
# All storages are None, no need to test that again
continue
# Copy the inputs over again
for r in node.inputs:
storage_map[r][0] = _lessbroken_deepcopy(r_vals[r])
# Copy the appropriate output storages
for r in node.outputs:
storage_map[r][0] = out_map.get(r, None)
thunk_py()
# Check outputs
for r in node.outputs:
if not r.type.is_valid_value(storage_map[r][0]):
raise InvalidValueError(r, storage_map[r][0], hint='perform output', specific_hint = r.type.value_validity_msg(storage_map[r][0]))
_check_inputs(node, storage_map, r_vals, dr_vals, active_order_set,
clobber_dr_vals=False, perform='py',
warn_input_not_reused=False)
_check_viewmap(node, storage_map)
for r in node.outputs:
if not r.type.values_eq_approx(r_vals[r], storage_map[r][0]):
# TODO: indicate it is not a C/Py problem
raise BadCLinkerOutput(r, val_py=r_vals[r], val_c=storage_map[r][0])
# Clear storage_map
for r in node.outputs:
storage_map[r][0] = None
_check_preallocated_output(
node=node,
thunk=thunk_py,
prealloc_modes=prealloc_modes,
def_val=def_val,
storage_map=storage_map,
r_vals=r_vals,
dr_vals=dr_vals,
perform='py',
active_order_set=active_order_set)
# print >> sys.stderr, i, "DEBUGMODE thunk_py %100s %50s %30s" % (node,
#[(id(o), numpy.asarray(storage_map[o][0])[0,0]) for o in node.inputs],
......@@ -1332,65 +1388,21 @@ class _Linker(gof.link.LocalLinker):
storage_map[r][0] = None #clear the storage_map for the thunk_c
if config.DebugMode.check_preallocated_output:
## Then, try to use different output storages
# TODO: factorize that code with the one for Python above
reuse_outputs = {}
c_cont_outputs = {}
f_cont_outputs = {}
for r in node.outputs:
r_val = r_vals[r]
reuse_outputs[r] = _lessbroken_deepcopy(r_val)
if isinstance(r.type, TensorType):
c_cont_outputs[r] = numpy.empty(
shape=r_val.shape,
dtype=r_val.dtype,
order='C')
f_cont_outputs[r] = numpy.empty(
shape=r_val.shape,
dtype=r_val.dtype,
order='F')
for out_map in (reuse_outputs, c_cont_outputs, f_cont_outputs):
if len(out_map) == 0:
# All storages are None, no need to test that again
continue
# Copy the inputs over again
for r in node.inputs:
storage_map[r][0] = _lessbroken_deepcopy(r_vals[r])
# Copy the appropriate output storages
for r in node.outputs:
#storage_map[r][0] = out_map.get(r, None)
if r in out_map:
storage_map[r][0] = out_map[r]
else:
print 'not tensor?', r
def thunk():
try:
thunk_c()
except:
raise_with_op(node)
# Check outputs
for r in node.outputs:
if not r.type.is_valid_value(storage_map[r][0]):
raise InvalidValueError(r, storage_map[r][0], hint='perform output', specific_hint = r.type.value_validity_msg(storage_map[r][0]))
_check_inputs(node, storage_map, r_vals, dr_vals, active_order_set,
clobber_dr_vals=False, perform='c',
warn_input_not_reused=False)
_check_viewmap(node, storage_map)
for r in node.outputs:
if not r.type.values_eq_approx(r_vals[r], storage_map[r][0]):
# TODO: indicate it is not a C/Py problem
raise BadCLinkerOutput(r, val_py=r_vals[r], val_c=storage_map[r][0])
# Clear storage map
for r in node.outputs:
storage_map[r][0] = None
_check_preallocated_output(
node=node,
thunk=thunk,
prealloc_modes=prealloc_modes,
def_val=def_val,
storage_map=storage_map,
r_vals=r_vals,
dr_vals=dr_vals,
perform='c code',
active_order_set=active_order_set)
# print >> sys.stderr, i, "DEBUGMODE thunk_c %100s %50s %30s" % (node,
#[(id(o), numpy.asarray(storage_map[o][0])[0,0]) for o in node.inputs],
......
Markdown 格式
0% 或
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！
请 注册 或者 后发表评论