Merge pull request #599 from lamblin/cudandarray_0_stride

theano/sandbox/cuda/cuda_ndarray.cu

@@ -1542,7 +1542,8 @@ CudaNdarray_Subscript(PyObject * py_self, PyObject * key)
             return NULL;
         }
         //initialize dimension 0 of rval
-        CudaNdarray_set_stride(rval, 0, step * CudaNdarray_HOST_STRIDES(self)[0]);
+        CudaNdarray_set_stride(rval, 0,
+                (slen == 1) ? 0 : step * CudaNdarray_HOST_STRIDES(self)[0]);
         CudaNdarray_set_dim(rval, 0, slen);
         if (verbose) std::cerr << "rval stride " << CudaNdarray_HOST_STRIDES(rval)[0] << "\n";
         // initialize dimensions > 0 of rval
@@ -1614,7 +1615,8 @@ CudaNdarray_Subscript(PyObject * py_self, PyObject * key)
                         return NULL;
                     }
                     rval->devdata += start * CudaNdarray_HOST_STRIDES(self)[d];
-                    CudaNdarray_set_stride(rval, rval_d, step * CudaNdarray_HOST_STRIDES(self)[d]);
+                    CudaNdarray_set_stride(rval, rval_d,
+                            (slen == 1) ? 0 : step * CudaNdarray_HOST_STRIDES(self)[d]);
                     CudaNdarray_set_dim(rval, rval_d, slen);
                     if (0)
                     {

theano/sandbox/cuda/tests/test_cuda_ndarray.py

@@ -463,6 +463,15 @@ def test_stride_manipulation():
     assert numpy.all(c == [[5, 4, 3], [2, 1, 0]])
 
 
+def test_subtensor_broadcastable():
+    a = numpy.zeros((2, 7), dtype='float32')
+    cuda_a = cuda_ndarray.CudaNdarray(a)
+    # Will have shape (1, 7), so the stride in the first dim should be 0
+    sub_a = cuda_a[1:]
+    assert sub_a.shape == (1, 7)
+    assert sub_a._strides[0] == 0
+
+
 def test_copy_subtensor0():
     sizeof_float=4
     a = theano._asarray(numpy.random.rand(30,20,5,5), dtype='float32')