Merge pull request #470 from lamblin/fix_blas_tests

theano/sandbox/cuda/blas.py

@@ -306,7 +306,7 @@ class GpuGemv(GpuOp):
         return Apply(self, [z, a, x, y, b], [z.type()])
 
     def c_code_cache_version(self):
-        return (2,)
+        return (3,)
 
     def c_code(self, node, name, inputs, outputs, sub):
         #z_out = alpha * dot(x,y) + beta * z_in
@@ -333,6 +333,7 @@ class GpuGemv(GpuOp):
             Py_INCREF(%(z_out)s);
         }
         else if (%(z_out)s
+                && (CudaNdarray_HOST_DIMS(%(z_out)s)[0] == CudaNdarray_HOST_DIMS(%(z_in)s)[0])
                 && ((CudaNdarray_HOST_STRIDES(%(z_out)s)[0] > 0)
                     || ((CudaNdarray_HOST_STRIDES(%(z_out)s)[0] == 0)
                         && (CudaNdarray_HOST_DIMS(%(z_out)s)[0] == 1))))

theano/tensor/blas_scipy.py

@@ -45,10 +45,13 @@ class ScipyGer(Ger):
             # in-place on a, even when I tell it to.
             A = cA[0]
             if A.size == 0:
-                # We don't have to do anything, A is empty.
+                # We don't have to compute anything, A is empty.
                 # We need this special case because Numpy considers it
                 # C-contiguous, wich is confusing.
-                pass
+                if not self.destructive:
+                    # Sometimes numpy thinks empty matrices can share memory,
+                    # so here to stop DebugMode from complaining.
+                    A = A.copy()
             elif A.flags['C_CONTIGUOUS']:
                 A = local_ger(calpha[0], cy[0], cx[0], a=A.T,
                         overwrite_a=int(self.destructive)).T