Add an optimization to move mrg_uniform to the GPU.

theano/sandbox/rng_mrg.py

@@ -992,7 +992,7 @@ class GPUA_mrg_uniform(GpuKernelBase, mrg_uniform_base):
         """ % locals()
 
     def c_code_cache_version(self):
-        return (2, self.GpuKernelBase_version)
+        return (3, self.GpuKernelBase_version)
 
 
 def guess_n_streams(size, warn=True):
@@ -1350,19 +1350,25 @@ class MRG_RandomStreams(object):
         return final_samples
 
 from theano.sandbox.gpuarray.opt import (register_opt as register_gpua,
-                                         op_lifter as gpua_lifter)
+                                         host_from_gpu as host_from_gpua)
 
 @register_gpua()
-@gpua_lifter([mrg_uniform])
+@local_optimizer([mrg_uniform])
 def local_gpua_mrg(node):
-    return GPUA_mrg_uniform.new(node.inputs[0], node.op.output_type.ndim,
-                                node.op.output_type.dtype, node.inputs[1])
+    if (type(node.op) == mrg_uniform and
+        isinstance(node.inputs[0].type, GpuArrayType)):
+        outs = GPUA_mrg_uniform.new(node.inputs[0],
+                                    node.op.output_type.ndim,
+                                    node.op.output_type.dtype,
+                                    node.inputs[1])
+        return [outs[0], host_from_gpua(outs[1])]
 
 
-@local_optimizer([mrg_uniform])
+MRG_RNGs = (mrg_uniform, GPU_mrg_uniform, GPUA_mrg_uniform)
+@local_optimizer(MRG_RNGs)
 def mrg_random_make_inplace(node):
     op = node.op
-    if isinstance(op, mrg_uniform) and not op.inplace:
+    if isinstance(op, MRG_RNGs) and not op.inplace:
         # op might be gpu version
         new_op = op.__class__(op.output_type, inplace=True)
         return new_op.make_node(*node.inputs).outputs