Add test for noncontiguous strides.

theano/gpuarray/tests/test_dnn.py

@@ -18,7 +18,7 @@ from theano.tensor.nnet import bn
 
 from .. import dnn
 from ..basic_ops import GpuAllocEmpty
-from ..type import gpuarray_shared_constructor
+from ..type import gpuarray_shared_constructor, GpuArrayType
 
 from .config import mode_with_gpu, mode_without_gpu, test_ctx_name, ref_cast
 from . import test_nnet
@@ -26,6 +26,11 @@ from .rnn_support import Model, GRU, LSTM, WrapperLayer
 
 from theano.configdefaults import SUPPORTED_DNN_CONV_ALGO_FWD
 
+try:
+    import pygpu
+except ImportError:
+    pass
+
 mode_with_gpu = mode_with_gpu.including()
 # Globally disabled for mode_without_gpu
 mode_with_gpu.check_py_code = False
@@ -1506,6 +1511,22 @@ def test_dnn_reduction_opt():
         yield dnn_reduction, 2, idtype, adtype, odtype
 
 
+def dnn_reduction_strides(shp, shuffle, slice):
+    inp = GpuArrayType('float32', (False,) * len(shp),
+                       context_name=test_ctx_name)()
+    tmp = inp.dimshuffle(shuffle)[slice]
+    res = tmp.sum(acc_dtype='float32', dtype='float32')
+    f = theano.function([inp], res, mode=mode_with_gpu)
+    assert any(isinstance(n.op, dnn.GpuDnnReduction)
+               for n in f.maker.fgraph.apply_nodes)
+    f(pygpu.zeros(shp, dtype='float32', context=inp.type.context))
+
+
+def test_dnn_reduction_strides():
+    yield dnn_reduction_strides, (2, 3, 2), (1, 0, 2), ...
+    yield dnn_reduction_strides, (2, 3, 2), (0, 1, 2), slice(None, None, -1)
+
+
 def dnn_maxargmax(nd, idtype, axis):
     inp = T.TensorType(idtype, (False,) * nd)()
     res = T.max_and_argmax(inp, axis=axis)