Skip to content

P
pytensor
提交 5b783de4 authored 作者: Pascal Lamblin's avatar Pascal Lamblin
浏览文件
操作

In DebugMode, call the thunks on different preallocated output storages

上级 8bd84e5e
隐藏空白字符变更
内嵌 并排
正在显示 包含 136 行增加 和 13 行删除
theano/compile/debugmode.py
浏览文件 @ 5b783de4
......@@ -1163,18 +1163,19 @@ class _Linker(gof.link.LocalLinker):
if not r.type.is_valid_value(storage_map[r][0]):
raise InvalidValueError(r, storage_map[r][0], client_node=node)
## On the first call to thunk_py(), its output storage will be None
if thunk_py:
debug(i, "DEBUGMODE running thunk_py")
debug(i, "DEBUGMODE running thunk_py with None as output storage")
try:
thunk_py()
except utils.MethodNotDefined:
thunk_py = None #shouldn't have put it into the list in the first place
if thunk_py:
# check output values for type-correctness
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]))
#if r in r_vals:
_check_inputs(node, storage_map, r_vals, dr_vals, active_order_set,
clobber_dr_vals=True, perform='py',
......@@ -1182,18 +1183,78 @@ class _Linker(gof.link.LocalLinker):
_check_viewmap(node, storage_map)
# 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],
#[(id(o), numpy.asarray(storage_map[o][0])[0,0]) for o in node.outputs])
sys.stdout.flush()
#retrieve each output from the storage_map
# Retrieve each output from the storage_map
# The return values of this first run will be the reference ones
for r in node.outputs:
assert r not in r_vals
# print >> sys.stderr, i, "DEBUGMODE storing reference output %x" % id(storage_map[r][0])
r_vals[r] = storage_map[r][0]
storage_map[r][0] = None #clear the storage_map of outputs for the thunk_c
## 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')
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
# 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],
#[(id(o), numpy.asarray(storage_map[o][0])[0,0]) for o in node.outputs])
sys.stdout.flush()
if thunk_c:
clobber = True
......@@ -1219,6 +1280,7 @@ class _Linker(gof.link.LocalLinker):
clobber = False
debug(i, "DEBUGMODE running thunk_c")
## First time, with None in output_storage
try:
thunk_c()
except:
......@@ -1241,11 +1303,7 @@ class _Linker(gof.link.LocalLinker):
_check_viewmap(node, storage_map)
# 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],
#[(id(o), numpy.asarray(storage_map[o][0])[0,0]) for o in node.outputs])
sys.stdout.flush()
# Check with Python result
for r in node.outputs:
if r in r_vals:
#print >> sys.stderr, i, "DEBUGMODE clearing output", r
......@@ -1262,6 +1320,71 @@ class _Linker(gof.link.LocalLinker):
storage_map[r][0] = None #clear the storage_map for the thunk_c
## 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
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
# 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],
#[(id(o), numpy.asarray(storage_map[o][0])[0,0]) for o in node.outputs])
sys.stdout.flush()
# we're done with this thunk
# clear everything out of the storage_map
......
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