Add a config variable to force ops to work synchronously.

theano/sandbox/gpuarray/__init__.py

 import logging
 
 import theano
-from theano.configparser import config
+from theano.configparser import config, AddConfigVar, BoolParam
 from theano.compile import optdb
 
 _logger_name = 'theano.sandbox.gpuarray'
@@ -18,12 +18,18 @@ try:
 except ImportError:
     pygpu = None
 
+AddConfigVar('gpuarray.sync',
+             """If True, every op will make sure its work is done before
+                returning.  Setting this to True will slow down execution,
+                but give much more accurate results in profiling.""",
+             BoolParam(False),
+             in_c_key=True)
+
 # This is for documentation not to depend on the availability of pygpu
 from type import (GpuArrayType, GpuArrayVariable, GpuArrayConstant,
                   GpuArraySharedVariable, gpuarray_shared_constructor)
 import opt
 
-
 def init_dev(dev):
     global pygpu_activated
     context = pygpu.init(dev)

theano/sandbox/gpuarray/elemwise.py

@@ -2,7 +2,7 @@ import copy
 from itertools import izip
 
 import numpy
-from theano import Op, Apply, scalar
+from theano import Op, Apply, scalar, config
 from theano.tensor.elemwise import Elemwise, DimShuffle
 
 try:
@@ -163,6 +163,8 @@ class GpuElemwise(HideC, Elemwise):
 
         # the dict call is there to avoid a syntax error in python < 2.6
         node._cache_elemwise_k(*args, **dict(broadcast=True))
+        if config.gpuarray.sync:
+            output_storage[0][0].sync()
 
 
 class SupportCodeError(Exception):