Add a python implementation of GpuGemv.

theano/sandbox/gpuarray/blas.py

@@ -3,6 +3,12 @@ from theano import Apply, config
 from theano.tensor.blas import Gemv
 from theano.sandbox.gpuarray.basic_ops import (HideC, as_gpuarray_variable)
 
+try:
+    from pygpu import blas
+except ImportError, e:
+    # To make sure theano is importable
+    pass
+
 class GpuGemv(HideC, Gemv):
     def make_node(self, y, alpha, A, x, beta):
         res = Gemv.make_node(self, y, alpha, A, x, beta)
@@ -11,8 +17,10 @@ class GpuGemv(HideC, Gemv):
         y = as_gpuarray_variable(y)
         return Apply(self, [y, alpha, A, x, beta], [y.type()])
 
-    def perform(*args):
-        raise NotImplementedError
+    def perform(self, node, inputs, out_storage):
+        y, alpha, A, x, beta = inputs
+        out_storage[0][0] = blas.sgemv(alpha, A, x, beta, y, trans=False,
+                                       overwrite_y=self.inplace)
 
     def c_code(self, node, name, inp, out, sub):
         vars = dict(out=out[0], y=inp[0], alpha=inp[1], A=inp[2], x=inp[3],