Don't move Elemwise to gpu when we can't generate c code for it! Test it.

theano/sandbox/cuda/opt.py

@@ -90,6 +90,8 @@ def local_gpu_elemwise_0(node):
     if isinstance(node.op, tensor.Elemwise):
         if numpy.any([i.owner and isinstance(i.owner.op, HostFromGpu) for i in node.inputs]):
             if numpy.all([o.type.dtype == 'float32' for o in node.outputs]):
+                if max_inputs_to_GpuElemwise(node)<len(node.inputs):
+                    return False
                 #don't set any inplace pattern. gpu_insert_inplace_optimizer will do it later
                 new_op = GpuElemwise(node.op.scalar_op)
 
@@ -113,9 +115,10 @@ def local_gpu_elemwise_0(node):
                 else:
                     return False
 
-                gpu_elemwise = split_huge_add_or_mul(gpu_elemwise.owner).outputs[0]
-
-                return [host_from_gpu(gpu_elemwise)]
+                gpu_elemwise = split_huge_add_or_mul(gpu_elemwise.owner)
+                if not gpu_elemwise:
+                    return False
+                return [host_from_gpu(gpu_elemwise.outputs[0])]
 @register_opt()
 @local_optimizer([])
 def local_gpu_elemwise_1(node):
@@ -130,8 +133,10 @@ def local_gpu_elemwise_1(node):
             new_op = GpuElemwise(elemwise_node.op.scalar_op)
             if all([i.dtype=='float32' for i in elemwise_node.inputs]):
                 gpu_elemwise = new_op(*[gpu_from_host(i) for i in elemwise_node.inputs])
-                gpu_elemwise = split_huge_add_or_mul(gpu_elemwise.owner).outputs[0]
-                return [gpu_elemwise]
+                gpu_elemwise = split_huge_add_or_mul(gpu_elemwise.owner)
+                if not gpu_elemwise:
+                    return False
+                return [gpu_elemwise.outputs[0]]
     return False
 
 @register_opt()
@@ -762,6 +767,8 @@ def split_huge_add_or_mul(node):
     """
     if node.op.scalar_op in (scal.add, scal.mul):
         max_nb_inputs = max_inputs_to_GpuElemwise(node)
+        if max_nb_inputs<=1 and len(node.inputs)>1:
+            return False
         while len(node.inputs)>max_nb_inputs:
             inner_op = []
             for i in range(0,len(node.inputs),max_nb_inputs):

theano/sandbox/cuda/tests/test_opt.py

@@ -176,6 +176,23 @@ def test_huge_elemwise_fusion():
     gen = lambda : theano._asarray(numpy.random.rand(*shape), dtype='float32')
     f(gen(),gen(),gen(),gen(),gen(),gen(),gen(),gen(),gen(),gen())
 
+    # Test the case where we can't put the computation on the gpu! their is too many
+    # dimensions to the input to have 2 inputs to the op!
+
+    shape = (1,2,3,4,5,6,7,2,2,3,2,1,2,2,2,)
+    ttype = tensor.tensor(dtype='float32',broadcastable=(False,)*len(shape))
+    vars = [tensor.tanh(ttype) for x in range(10)]
+    f = pfunc(vars, [vars[0]-vars[1]-vars[2]-vars[3]-vars[4]-vars[5]-vars[6]], mode=mode_with_gpu)
+    topo = f.maker.env.toposort()
+    #theano.printing.debugprint(f)
+    assert len(topo)==1
+    assert sum([isinstance(node.op, cuda.GpuElemwise) for node in topo])==0
+    assert sum([isinstance(node.op, tensor.Elemwise) for node in topo])==1
+    #let debugmode catch errors
+    gen = lambda : theano._asarray(numpy.random.rand(*shape), dtype='float32')
+    f(gen(),gen(),gen(),gen(),gen(),gen(),gen(),gen(),gen(),gen())
+
+
 def test_elemwise_fusion():
     """ Test the the GpuElemwise fusion work correctly"""
     shape = (3,4)