Merge pull request #6169 from lamblin/fix_gemv_dot

theano/gpuarray/blas.py

@@ -120,16 +120,12 @@ class GpuGemv(BlasOp):
             %(out)s->ga.nd = 0;
             %(A)s->ga.nd = 1;
             %(A)s->ga.dimensions[0] = %(A)s->ga.dimensions[1];
-            if (%(A)s->ga.flags & GA_C_CONTIGUOUS) {
             ssize_t a_stride0 = %(A)s->ga.strides[0];
             %(A)s->ga.strides[0] = %(A)s->ga.strides[1];
             if (pygpu_blas_rdot(%(x)s, %(A)s, %(out)s, 0) == -1) {
                 %(fail)s
             }
             %(A)s->ga.strides[0] = a_stride0;
-            } else if (pygpu_blas_rdot(%(x)s, %(A)s, %(out)s, 0) == -1) {
-                %(fail)s
-            }
             %(out)s->ga.nd = 1;
             %(A)s->ga.nd = 2;
             %(A)s->ga.dimensions[0] = 1;
@@ -149,7 +145,7 @@ class GpuGemv(BlasOp):
         return code
 
     def c_code_cache_version(self):
-        return (7,)
+        return (8,)
 
 gpugemv_no_inplace = GpuGemv(inplace=False)
 gpugemv_inplace = GpuGemv(inplace=True)

theano/gpuarray/tests/test_blas.py

@@ -234,3 +234,14 @@ def test_gemv_zeros():
     tmp = f(A, b)
     assert np.allclose(tmp,
                        np.zeros((dim,)))
+
+
+def test_gemv_dot_strides():
+    # Reported in https://github.com/Theano/Theano/issues/6142
+    xv = rand(5)
+    yv = rand(5, 1)
+    x = gpuarray_shared_constructor(xv)
+    y = gpuarray_shared_constructor(yv, broadcastable=(False, True))
+    f = theano.function([], tensor.dot(x, y[::-1]), mode=mode_with_gpu)
+    out = f()
+    utt.assert_allclose(out, np.dot(xv, yv[::-1]))