提交 6005ac2b authored 作者: nouiz's avatar nouiz

Merge pull request #711 from jaberg/VM_allow_gc_and_conditionals

Vm allow gc of all intermediate results at the end of function call. Only the lazy op cause intermediate result to don't be gc during compilation.
......@@ -208,6 +208,20 @@ class Stack(VM):
node_idx[node] = i
self.apply_time[node] = 0
self.outputs_size[node] = []
# XXX: inconsistent style - why modify node here rather
# than track destroy_dependencies with dictionary like
# storage_map?
#
# destroy_dependencies
# --------------------
# The destroy_dependencies is a list of variables that are implicit
# dependencies induced by a destroy_map (compare node.inputs which
# are *explicit* dependencies). The variables in
# destroy_dependencies would be impossible to compute after the
# current `node` runs, because node.thunk() is going to destroy a
# common input variable needed by whatever node owns each variable
# in destroy_depenencies.
node.destroy_dependencies = []
if node in ords:
for prereq in ords[node]:
......@@ -235,6 +249,7 @@ class Stack(VM):
# Some thunks on some computers run faster than the granularity
# of the time.time clock.
# Profile output looks buggy if a node has run but takes 0 time.
# (and profile code might hide real bugs if it rounds up 0)
dt = max(time.time() - t0, 1e-10)
if self.callback is not None:
self.callback(
......@@ -270,32 +285,25 @@ class Stack(VM):
last_apply_stack_len = apply_stack_len
current_apply = apply_stack.pop()
current_inputs = current_apply.inputs
current_outputs = current_apply.outputs
current_deps = current_inputs + current_apply.destroy_dependencies
# Use these for loops + breaks to short circuit evaluation
# This is a significant performance point
computed_ins = True
for i in current_apply.inputs:
if not compute_map[i][0]:
computed_ins = False
break
computed_outs = True
for o in current_apply.outputs:
if not compute_map[o][0]:
computed_outs = False
break
if computed_ins:
for d in current_apply.destroy_dependencies:
if not compute_map[d][0]:
computed_ins = False
break
computed_ins = all(compute_map[v][0] for v in current_deps)
computed_outs = all(compute_map[v][0] for v in current_outputs)
if not thunks[self.node_idx[current_apply]].lazy:
#
# stack loop: Normal Non-Lazy Case
# ================================
#
# Check if all inputs are in place
# If so compute thunk and remove it from the apply_stack
# If not leave it in, and add to the apply_stack those
# that will produce you those inputs
if computed_ins and not computed_outs:
# -- Non-lazy case: have inputs, time to compute outputs
try:
_, dt = self.run_thunk_of_node(current_apply)
del _
......@@ -323,30 +331,34 @@ class Stack(VM):
raise_with_op(current_apply)
for o in current_apply.outputs:
compute_map[o][0] = 1
# Garbage Collection -> check if anybody else uses
# this input
if self.allow_gc:
for i in current_apply.inputs:
if (dependencies[i] and i.owner
and i not in self.outputs):
empty_storage_map = True
for x in dependencies[i]:
if not compute_map[x][0]:
empty_storage_map = False
break
if empty_storage_map:
# Garbage Collection -> check if anybody else uses
# this input
if (dependencies[i]
and i.owner
and i not in self.outputs):
if all(compute_map[v][0]
for v in dependencies[i]):
storage_map[i][0] = None
compute_map[i][0] = 0
elif not computed_ins:
# -- Non-lazy case, need inputs
apply_stack.append(current_apply)
apply_stack.extend(inp.owner for inp
in current_apply.inputs if inp.owner)
apply_stack.extend(inp.owner for inp
in current_apply.destroy_dependencies
if inp.owner)
apply_stack.extend(inp.owner
for inp in current_deps
if inp.owner)
elif not computed_outs:
# Try and run it to see if it works
#
# stack loop: Lazy Evaluation Case
# ================================
#
# Lazy evaluation protocol is to run the thunk with the
# current storage_map and compute_map accessed via closure,
# and the thunk will return a list of variables from its input
# list that it requires.
try:
requires, dt = self.run_thunk_of_node(current_apply)
self.apply_time[current_apply] += dt
......@@ -389,6 +401,16 @@ class Stack(VM):
break
if empty_storage_map:
storage_map[i][0] = None
compute_map[i][0] = None
# Hacky coarse gc final pass
# This is required until we have a proper gc algorithm for graphs with
# lazy evaluation. See discussion on theano-dev June 19 2012.
if self.allow_gc:
for v in storage_map:
if v.owner and 'output' not in zip(*v.clients)[0]:
storage_map[v][0] = None
compute_map[v][0] = 0
try:
......@@ -451,10 +473,35 @@ class VM_Linker(link.LocalLinker):
# admittedly confusing, and it could use some cleaning up. The base
# Linker object should probably go away completely.
def compute_gc_dependencies(self, smap):
def compute_gc_dependencies(self, variables):
"""
Returns dict: variable K -> list of variables [v1, v2, v3, ...]
for each K in variables.
The variables v1, v2, ... are the full set of variables that depend
directly on K. When we know that none of them will need to be
computed, we know that:
* K will not need to be computed
* if K is already computed, it can be released for garbage collection
Parameters
----------
variables - iterable over the variables used in a graph computation.
N.B. gc means garbage collection
"""
dependencies = {}
for k in smap:
for k in variables:
dependencies[k] = []
# If k has no owner, it is an input / constant and its value
# should not be removed from the storage_map because we have no
# way of getting it back.
#
# XXX if k has no clients... what is it doing in the computation?
if k.owner and k.clients:
ls = []
is_output = 0
......
Markdown 格式
0%
您添加了 0 到此讨论。请谨慎行事。
请先完成此评论的编辑!
注册 或者 后发表评论