Skip to content

P
pytensor
提交 c4f1ede2 authored 作者: Frédéric Bastien's avatar Frédéric Bastien
浏览文件
操作

Merge pull request #3475 from carriepl/scan_backend_speedup

Scan backend speedup
上级 60026783 9a70e813
隐藏空白字符变更
内嵌 并排
正在显示 包含 170 行增加 和 156 行删除
theano/scan_module/scan_op.py
浏览文件 @ c4f1ede2
...@@ -1175,13 +1175,12 @@ class Scan(PureOp): ...@@ -1175,13 +1175,12 @@ class Scan(PureOp):
offset = self.nit_sot_arg_offset + self.n_nit_sot offset = self.nit_sot_arg_offset + self.n_nit_sot
other_args = args[offset:] other_args = args[offset:]
input_storage = self.fn.input_storage input_storage = self.fn.input_storage
old_input_storage = [None] * len(input_storage) nb_mitmot_in = sum(map(len, self.tap_array[:self.n_mit_mot]))
old_input_data = [None] * len(input_storage) old_mitmot_input_storage = [None] * nb_mitmot_in
input_reused = [None] * len(input_storage) old_mitmot_input_data = [None] * nb_mitmot_in
output_storage = self.fn.output_storage output_storage = self.fn.output_storage
old_output_storage = [None] * len(output_storage) old_output_storage = [None] * len(output_storage)
old_output_data = [None] * len(output_storage) old_output_data = [None] * len(output_storage)
output_reused = [None] * len(output_storage)
fn = self.fn.fn fn = self.fn.fn
offset = (self.n_seqs + sum(map(len, self.tap_array[:self.n_outs])) + offset = (self.n_seqs + sum(map(len, self.tap_array[:self.n_outs])) +
self.n_shared_outs) self.n_shared_outs)
...@@ -1281,21 +1280,22 @@ class Scan(PureOp): ...@@ -1281,21 +1280,22 @@ class Scan(PureOp):
old_output_data[idx] = None old_output_data[idx] = None
# 4.6. Keep a reference to the variables (ndarrays, CudaNdarrays, # 4.6. Keep a reference to the variables (ndarrays, CudaNdarrays,
# etc) currently in the input_storage to be able to compare them # etc) associated with mitmot inputs currently in the
# with the content of the input_storage after the execution of the # input_storage to be able to compare them with the content of the
# function. Also keep pointers to their data to be able to detect # input_storage after the execution of the function. Also keep
# cases where outputs reused the allocated object but alter the # pointers to their data to be able to detect cases where outputs
# memory region they refer to. # reused the allocated object but alter the memory region they
for idx in xrange(len(input_storage)): # refer to.
var = input_storage[idx].storage[0] for idx in xrange(nb_mitmot_in):
old_input_storage[idx] = var var = input_storage[idx + self.n_seqs].storage[0]
old_mitmot_input_storage[idx] = var
if hasattr(var, 'gpudata'): if hasattr(var, 'gpudata'):
old_input_data[idx] = var.gpudata old_mitmot_input_data[idx] = var.gpudata
elif hasattr(var, 'data'): elif hasattr(var, 'data'):
old_input_data[idx] = var.data old_mitmot_input_data[idx] = var.data
else: else:
old_input_data[idx] = None old_mitmot_input_data[idx] = None
# 5.1 compute outputs # 5.1 compute outputs
t0_fn = time.time() t0_fn = time.time()
...@@ -1338,71 +1338,39 @@ class Scan(PureOp): ...@@ -1338,71 +1338,39 @@ class Scan(PureOp):
storage.data = output_storage[offset_out].data storage.data = output_storage[offset_out].data
offset_out -= 1 offset_out -= 1
# 5.3. Check which of the pre-allocated outputs (if applicable)
# have been reused by the inner function
for idx in xrange(len(output_storage)):
# If the storage map does not contain the same object, then
# the pre-allocated output has not been reused
new_var = output_storage[idx].storage[0]
if old_output_storage[idx] is new_var:
# The pre-allocated output is only considered as having
# been reused if it still points to the same data as it
# did before the execution of the inner function
if old_output_data[idx] is None:
output_reused[idx] = False
else:
if hasattr(new_var, 'gpudata'):
output_reused[idx] = (new_var.gpudata ==
old_output_data[idx])
elif hasattr(new_var, 'data'):
output_reused[idx] = (new_var.data ==
old_output_data[idx])
else:
output_reused[idx] = False
# 5.4 Check which of the input storage have been modified by the
# inner function
for idx in xrange(len(input_storage)):
# If the storage map does not contain the same object, then
# the pre-allocated output has not been reused
new_var = input_storage[idx].storage[0]
if old_input_storage[idx] is new_var:
# The pre-allocated output is only considered as having
# been reused if it still points to the same data as it
# did before the execution of the inner function
if old_input_data[idx] is None:
input_reused[idx] = False
else:
if hasattr(new_var, 'gpudata'):
input_reused[idx] = (new_var.gpudata ==
old_input_data[idx])
elif hasattr(new_var, 'data'):
input_reused[idx] = (new_var.data ==
old_input_data[idx])
else:
input_reused[idx] = False
t_fn += dt_fn t_fn += dt_fn
offset_out = 0 offset_out = 0
# 5.5 Copy over the values for mit_mot outputs # 5.3 Copy over the values for mit_mot outputs
mitmot_inp_offset = self.n_seqs mitmot_inp_offset = 0
mitmot_out_idx = 0 mitmot_out_idx = 0
for j in xrange(self.n_mit_mot): for j in xrange(self.n_mit_mot):
for k in self.mit_mot_out_slices[j]: for k in self.mit_mot_out_slices[j]:
if self.mitmots_preallocated[mitmot_out_idx]: if self.mitmots_preallocated[mitmot_out_idx]:
# This output tap has been preallocated. If the # This output tap has been preallocated.
# corresponding input storage has been replaced,
# recover the value as usual. Otherwise, the input was
# modified inplace and nothing needs to be done.
inp_idx = (mitmot_inp_offset + inp_idx = (mitmot_inp_offset +
self.tap_array[j].index(k)) self.tap_array[j].index(k))
if not input_reused[inp_idx]:
# Verify whether the input points to the same data as
# it did before the execution of the inner function.
old_var = old_mitmot_input_storage[inp_idx]
new_var = input_storage[self.n_seqs + inp_idx].storage[0]
if old_var is new_var:
old_data = old_mitmot_input_data[inp_idx]
if hasattr(new_var, 'gpudata'):
same_data = (new_var.gpudata == old_data)
elif hasattr(new_var, 'data'):
same_data = (new_var.data == old_data)
else:
same_data = False
# If the corresponding input storage still points to
# the same data, it has been modified inplace and
# nothing needs to be done. Otherwise, recover the
# and store it in `outs` as usual
if not same_data:
outs[j][0][k + pos[j]] = \ outs[j][0][k + pos[j]] = \
input_storage[inp_idx].storage[0] input_storage[self.n_seqs + inp_idx].storage[0]
else: else:
# This output tap has not been preallocated, recover # This output tap has not been preallocated, recover
...@@ -1415,21 +1383,42 @@ class Scan(PureOp): ...@@ -1415,21 +1383,42 @@ class Scan(PureOp):
mitmot_inp_offset += len(self.tap_array[j]) mitmot_inp_offset += len(self.tap_array[j])
# 5.6 Copy over the values for mit_sot/sit_sot outputs # 5.4 Copy over the values for mit_sot/sit_sot outputs
begin = self.n_mit_mot begin = self.n_mit_mot
end = self.n_outs end = self.n_outs
offset_out -= self.n_mit_mot offset_out -= self.n_mit_mot
for j in xrange(begin, end): for j in xrange(begin, end):
if (store_steps[j] == 1 or self.vector_outs[j] or
not output_reused[offset_out + j]): # Copy the output value to `outs`, if necessary
if store_steps[j] == 1 or self.vector_outs[j]:
outs[j][0][pos[j]] = \ outs[j][0][pos[j]] = \
output_storage[offset_out + j].storage[0] output_storage[offset_out + j].storage[0]
else:
# Check whether the initialization of the output storage
# map for this output has been reused.
old_var = old_output_storage[offset_out + j]
new_var = output_storage[offset_out + j].storage[0]
if old_var is new_var:
old_data = old_output_data[offset_out + j]
if old_data is None:
output_reused = False
elif hasattr(new_var, 'gpudata'):
output_reused = (new_var.gpudata == old_data)
elif hasattr(new_var, 'data'):
output_reused = (new_var.data == old_data)
else:
output_reused = False
if not output_reused:
outs[j][0][pos[j]] = \
output_storage[offset_out + j].storage[0]
# 5.7 Copy over the values for nit_sot outputs # 5.5 Copy over the values for nit_sot outputs
begin = end begin = end
end += self.n_nit_sot end += self.n_nit_sot
for j in xrange(begin, end): for j in xrange(begin, end):
if i == 0: if i == 0:
jout = j + offset_out jout = j + offset_out
shape = (store_steps[j],) + \ shape = (store_steps[j],) + \
...@@ -1445,12 +1434,30 @@ class Scan(PureOp): ...@@ -1445,12 +1434,30 @@ class Scan(PureOp):
elif outs[j][0].shape[0] != store_steps[j]: elif outs[j][0].shape[0] != store_steps[j]:
outs[j][0] = outs[j][0][:store_steps[j]] outs[j][0] = outs[j][0][:store_steps[j]]
outs[j][0][pos[j]] = output_storage[jout].storage[0] outs[j][0][pos[j]] = output_storage[jout].storage[0]
elif (store_steps[j] == 1 or self.vector_outs[j] or elif store_steps[j] == 1 or self.vector_outs[j]:
not output_reused[offset_out + j]):
outs[j][0][pos[j]] = \ outs[j][0][pos[j]] = \
output_storage[j + offset_out].storage[0]
else:
# Check whether the initialization of the output storage map
# for this output has been reused.
old_var = old_output_storage[offset_out + j]
old_data = old_output_data[offset_out + j]
new_var = output_storage[offset_out + j].storage[0]
if old_var is new_var:
if old_data is None:
output_reused = False
elif hasattr(new_var, 'gpudata'):
output_reused = (new_var.gpudata == old_data)
elif hasattr(new_var, 'data'):
output_reused = (new_var.data == old_data)
else:
output_reused = False
if not output_reused:
outs[j][0][pos[j]] = \
output_storage[j + offset_out].storage[0] output_storage[j + offset_out].storage[0]
# 5.8 Copy over the values for outputs corresponding to shared # 5.6 Copy over the values for outputs corresponding to shared
# variables # variables
begin = end begin = end
end += self.n_shared_outs end += self.n_shared_outs
......
theano/scan_module/scan_perform.c
浏览文件 @ c4f1ede2
This source diff could not be displayed because it is too large. You can view the blob instead.
theano/scan_module/scan_perform.pyx
浏览文件 @ c4f1ede2
...@@ -62,7 +62,7 @@ import copy ...@@ -62,7 +62,7 @@ import copy
def get_version(): def get_version():
return 0.287 return 0.289
@cython.boundscheck(False) @cython.boundscheck(False)
def perform( def perform(
...@@ -195,8 +195,6 @@ def perform( ...@@ -195,8 +195,6 @@ def perform(
cdef unsigned int len_output_storage = (n_mit_mot_outs + n_mit_sot + cdef unsigned int len_output_storage = (n_mit_mot_outs + n_mit_sot +
n_sit_sot + n_nit_sot + n_sit_sot + n_nit_sot +
n_shared_outs) n_shared_outs)
cdef int input_reused[500] # max 500 inputs
cdef int output_reused[500] # max 500 outputs
if n_steps < 0: if n_steps < 0:
...@@ -256,9 +254,11 @@ def perform( ...@@ -256,9 +254,11 @@ def perform(
offset = nit_sot_arg_offset + n_nit_sot offset = nit_sot_arg_offset + n_nit_sot
other_args = args[offset:] other_args = args[offset:]
input_storage = fnct.input_storage input_storage = fnct.input_storage
len_input_storage = len(input_storage) nb_mitmot_in = 0
old_input_storage = [None] * len_input_storage for idx in range(n_mit_mot):
old_input_data = [None] * len_input_storage nb_mitmot_in += tap_array_len[idx]
old_mitmot_input_storage = [None] * nb_mitmot_in
old_mitmot_input_data = [None] * nb_mitmot_in
output_storage = fnct.output_storage output_storage = fnct.output_storage
old_output_storage = [None] * len_output_storage old_output_storage = [None] * len_output_storage
old_output_data = [None] * len_output_storage old_output_data = [None] * len_output_storage
...@@ -366,21 +366,21 @@ def perform( ...@@ -366,21 +366,21 @@ def perform(
old_output_data[idx] = None old_output_data[idx] = None
# 4.6. Keep a reference to the variables (ndarrays, CudaNdarrays, # 4.6. Keep a reference to the variables (ndarrays, CudaNdarrays,
# etc) currently in the input_storage to be able to compare them # etc) associated with mitmot inputs currently in the input_storage to
# with the content of the input_storage after the execution of the # be able to compare them with the content of the input_storage after
# function. Also keep pointers to their data to be able to detect # the execution of the function. Also keep pointers to their data to
# cases where outputs reused the allocated object but alter the # be able to detect cases where outputs reused the allocated object
# memory region they refer to. # but alter the memory region they refer to.
for idx in xrange(len(input_storage)): for idx in xrange(nb_mitmot_in):
var = input_storage[idx].storage[0] var = input_storage[idx + n_seqs].storage[0]
old_input_storage[idx] = var old_mitmot_input_storage[idx] = var
if hasattr(var, 'gpudata'): if hasattr(var, 'gpudata'):
old_input_data[idx] = var.gpudata old_mitmot_input_data[idx] = var.gpudata
elif hasattr(var, 'data'): elif hasattr(var, 'data'):
old_input_data[idx] = var.data old_mitmot_input_data[idx] = var.data
else: else:
old_input_data[idx] = None old_mitmot_input_data[idx] = None
# 5.1 compute outputs # 5.1 compute outputs
t0_fn = time.time() t0_fn = time.time()
...@@ -415,68 +415,37 @@ def perform( ...@@ -415,68 +415,37 @@ def perform(
storage.data = output_storage[offset_out].data storage.data = output_storage[offset_out].data
offset_out -= 1 offset_out -= 1
# 5.3. Check which of the pre-allocated outputs (if applicable)
# have been reused by the inner function
for idx in range(len_output_storage):
# If the storage map does not contain the same object, then
# the pre-allocated output has not been reused
new_var = output_storage[idx].storage[0]
if old_output_storage[idx] is new_var:
# The pre-allocated output is only considered as having
# been reused if it still points to the same data as it
# did before the execution of the inner function
if old_output_data[idx] is None:
output_reused[idx] = False
else:
if hasattr(new_var, 'gpudata'):
output_reused[idx] = (new_var.gpudata ==
old_output_data[idx])
elif hasattr(new_var, 'data'):
output_reused[idx] = (new_var.data ==
old_output_data[idx])
else:
output_reused[idx] = False
# 5.4. Check which of the input storage have been modified by the
# inner function
for idx in xrange(len(input_storage)):
# If the storage map does not contain the same object, then
# the pre-allocated output has not been reused
new_var = input_storage[idx].storage[0]
if old_input_storage[idx] is new_var:
# The pre-allocated output is only considered as having
# been reused if it still points to the same data as it
# did before the execution of the inner function
if old_input_data[idx] is None:
input_reused[idx] = False
else:
if hasattr(new_var, 'gpudata'):
input_reused[idx] = (new_var.gpudata ==
old_input_data[idx])
elif hasattr(new_var, 'data'):
input_reused[idx] = (new_var.data ==
old_input_data[idx])
else:
input_reused[idx] = False
offset_out = 0 offset_out = 0
# 5.5 Copy over the values for mit_mot outputs
mitmot_inp_offset = self.n_seqs # 5.3 Copy over the values for mit_mot outputs
mitmot_inp_offset = 0
mitmot_out_idx = 0 mitmot_out_idx = 0
for j in xrange(self.n_mit_mot): for j in xrange(self.n_mit_mot):
for k in self.mit_mot_out_slices[j]: for k in self.mit_mot_out_slices[j]:
if mitmots_preallocated[<unsigned int>mitmot_out_idx]: if mitmots_preallocated[<unsigned int>mitmot_out_idx]:
# This output tap has been preallocated. If the # This output tap has been preallocated.
# corresponding input storage has been replaced,
# recover the value as usual. Otherwise, the input was
# modified inplace and nothing needs to be done.
inp_idx = (mitmot_inp_offset + inp_idx = (mitmot_inp_offset +
self.tap_array[j].index(k)) self.tap_array[j].index(k))
if not input_reused[inp_idx]:
# Verify whether the input points to the same data as
# it did before the execution of the inner function.
old_var = old_mitmot_input_storage[inp_idx]
new_var = input_storage[n_seqs + inp_idx].storage[0]
if old_var is new_var:
old_data = old_mitmot_input_data[inp_idx]
if hasattr(new_var, 'gpudata'):
same_data = (new_var.gpudata == old_data)
elif hasattr(new_var, 'data'):
same_data = (new_var.data == old_data)
else:
same_data = False
# If the corresponding input storage has been replaced,
# recover the value as usual. Otherwise, the input was
# modified inplace and nothing needs to be done.
if not same_data:
outs[j][0][<unsigned int>(k + pos[j])] = \ outs[j][0][<unsigned int>(k + pos[j])] = \
input_storage[<unsigned int>inp_idx].storage[0] input_storage[<unsigned int>(n_seqs + inp_idx)].storage[0]
else: else:
# This output tap has not been preallocated, recover # This output tap has not been preallocated, recover
...@@ -489,21 +458,42 @@ def perform( ...@@ -489,21 +458,42 @@ def perform(
mitmot_inp_offset += len(self.tap_array[j]) mitmot_inp_offset += len(self.tap_array[j])
# 5.6 Copy over the values for mit_sot/sit_sot outputs # 5.4 Copy over the values for mit_sot/sit_sot outputs
begin = n_mit_mot begin = n_mit_mot
end = n_outs end = n_outs
offset_out -= n_mit_mot offset_out -= n_mit_mot
for j in range(begin, end): for j in range(begin, end):
if (store_steps[j] == 1 or vector_outs[j] == 1 or
not output_reused[<unsigned int>(offset_out+j)]):
# Copy the output value to `outs`, if necessary
if store_steps[j] == 1 or vector_outs[j] == 1:
outs[j][0][pos[j]] = output_storage[<unsigned int>(offset_out+j)].storage[0] outs[j][0][pos[j]] = output_storage[<unsigned int>(offset_out+j)].storage[0]
else:
# Check whether the initialization of the output storage map
# for this output has been reused.
old_var = old_output_storage[offset_out + j]
old_data = old_output_data[offset_out + j]
new_var = output_storage[offset_out + j].storage[0]
if old_var is new_var:
if old_data is None:
output_reused = False
elif hasattr(new_var, 'gpudata'):
output_reused = (new_var.gpudata == old_data)
elif hasattr(new_var, 'data'):
output_reused = (new_var.data == old_data)
else:
output_reused = False
if not output_reused:
outs[j][0][pos[j]] = \
output_storage[<unsigned int>(offset_out+j)].storage[0]
# 5.7 Copy over the values for nit_sot outputs
# 5.5 Copy over the values for nit_sot outputs
begin = end begin = end
end += n_nit_sot end += n_nit_sot
for j in range(begin,end): for j in range(begin,end):
if i == 0: if i == 0:
jout = j+offset_out jout = j+offset_out
shape = (store_steps[j],) + output_storage[jout].storage[0].shape shape = (store_steps[j],) + output_storage[jout].storage[0].shape
...@@ -518,11 +508,28 @@ def perform( ...@@ -518,11 +508,28 @@ def perform(
elif outs[j][0].shape[0] != store_steps[j]: elif outs[j][0].shape[0] != store_steps[j]:
outs[j][0] = outs[j][0][:store_steps[j]] outs[j][0] = outs[j][0][:store_steps[j]]
outs[j][0][pos[j]] = output_storage[jout].storage[0] outs[j][0][pos[j]] = output_storage[jout].storage[0]
elif (store_steps[j] == 1 or vector_outs[j] == 1 or elif store_steps[j] == 1 or vector_outs[j] == 1:
not output_reused[<unsigned int>(offset_out+j)]):
outs[j][0][pos[j]] = output_storage[j+offset_out].storage[0] outs[j][0][pos[j]] = output_storage[j+offset_out].storage[0]
else:
# Check whether the initialization of the output storage map
# for this output has been reused.
old_var = old_output_storage[offset_out + j]
old_data = old_output_data[offset_out + j]
new_var = output_storage[offset_out + j].storage[0]
if old_var is new_var:
if old_data is None:
output_reused = False
elif hasattr(new_var, 'gpudata'):
output_reused = (new_var.gpudata == old_data)
elif hasattr(new_var, 'data'):
output_reused = (new_var.data == old_data)
else:
output_reused = False
if not output_reused:
outs[j][0][pos[j]] = output_storage[j+offset_out].storage[0]
# 5.8 Copy over the values for outputs corresponding to shared # 5.6 Copy over the values for outputs corresponding to shared
# variables # variables
begin = end begin = end
end += n_shared_outs end += n_shared_outs
......
theano/scan_module/scan_perform_ext.py
浏览文件 @ c4f1ede2
...@@ -17,7 +17,7 @@ from theano.gof import cmodule ...@@ -17,7 +17,7 @@ from theano.gof import cmodule
_logger = logging.getLogger('theano.scan_module.scan_perform') _logger = logging.getLogger('theano.scan_module.scan_perform')
version = 0.287 # must match constant returned in function get_version() version = 0.289 # must match constant returned in function get_version()
need_reload = False need_reload = False
......
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