Replace Dot22Scalar by a GpuGemm instead and add a test.

theano/sandbox/gpuarray/opt.py

@@ -753,9 +753,11 @@ def local_gpua_dot22(node, context_name):
 @register_opt('fast_compile')
 @op_lifter([tensor.blas.Dot22Scalar])
 def local_gpua_dot22scalar(node, context_name):
-    dot = gpu_dot22(as_gpuarray_variable(node.inputs[0], context_name),
-                    as_gpuarray_variable(node.inputs[1], context_name))
-    return [node.inputs[2] * dot]
+    x, y, a = node.inputs
+    x = as_gpuarray_variable(x, context_name)
+    y = as_gpuarray_variable(y, context_name)
+    z = GpuAllocEmpty(x.dtype, context_name)(x.shape[0], y.shape[1])
+    return [GpuGemm(inplace=False)(z, a, x, y, 0)]
 
 
 @register_opt('fast_compile')

theano/sandbox/gpuarray/tests/test_opt.py

@@ -11,6 +11,7 @@ from .. import basic_ops
 from ..type import GpuArrayType, gpuarray_shared_constructor, get_context
 from ..basic_ops import (
     GpuAlloc, GpuAllocEmpty, GpuReshape, GpuFromHost, host_from_gpu)
+from ..blas import GpuGemm
 from ..elemwise import GpuCAReduceCuda, GpuCAReduceCPY, GpuElemwise
 from ..subtensor import GpuSubtensor
 
@@ -254,6 +255,17 @@ def test_local_gpu_elemwise_careduce():
     assert topo[1].op.pre_scalar_op == theano.scalar.sqr
     utt.assert_allclose(f(data), (data * data).sum(axis=1))
 
+def test_local_lift_dot22scalar():
+    x = tensor.matrix()
+    y = tensor.matrix()
+    a = tensor.scalar()
+    o = tensor.blas.Dot22Scalar()(x, y, a)
+    f = theano.function([x, y, a], o, mode=mode_with_gpu)
+    assert not any(isinstance(n.op, tensor.blas.Dot22Scalar)
+                   for n in f.maker.fgraph.apply_nodes)
+    assert any(isinstance(n.op, GpuGemm)
+               for n in f.maker.fgraph.apply_nodes)
+
 
 def test_local_gpu_subtensor():
     # Test shared forced on CPU.