Add the theano flags gpuarray.sched

doc/library/config.txt

@@ -487,6 +487,21 @@ import theano and print the config variable, as in:
         automatically to get more memory. But this can cause
         fragmentation, see note above.
 
+.. attribute:: config.gpuarray.sched
+
+    String value: ``'default'``, ``'multi'``, ``'single'``
+
+    Default: ``'default'``
+
+    Control the stream mode of contexts.
+
+    The sched parameter passed for context creation to pygpu.  With
+    CUDA, using "multi" mean using the parameter
+    cudaDeviceScheduleYield. This is useful to lower the CPU overhead
+    when waiting for GPU. One user found that it speed up its other
+    process that was doing data augmentation.
+
+
 .. attribute:: config.gpuarray.single_stream
 
     Boolean value

theano/configdefaults.py

@@ -242,6 +242,16 @@ AddConfigVar('gpuarray.preallocate',
              FloatParam(0),
              in_c_key=False)
 
+AddConfigVar('gpuarray.sched',
+             """The sched parameter passed for context creation to pygpu.
+                With CUDA, using "multi" mean using the parameter
+                cudaDeviceScheduleYield. This is useful to lower the
+                CPU overhead when waiting for GPU. One user found that it
+                speed up its other process that was doing data augmentation.
+             """,
+             EnumStr("default", "multi", "single"),
+             in_c_key=True)
+
 AddConfigVar('gpuarray.single_stream',
              """
              If your computations are mostly lots of small elements,

theano/gpuarray/__init__.py

@@ -63,7 +63,8 @@ def init_dev(dev, name=None):
     if dev not in init_dev.devmap:
         ctx = pygpu.init(dev,
                          disable_alloc_cache=config.gpuarray.preallocate < 0,
-                         single_stream=config.gpuarray.single_stream)
+                         single_stream=config.gpuarray.single_stream,
+                         sched=config.gpuarray.sched)
         init_dev.devmap[dev] = ctx
         if config.gpuarray.preallocate > 0:
             MB = (1024 * 1024)