Merge pull request #3712 from nouiz/crash_gpu_from_host

theano/sandbox/cuda/tests/test_opt.py

@@ -504,6 +504,24 @@ def test_pdbbreakpoint_op():
     assert topo[-1].op == cuda.host_from_gpu
 
 
+def test_local_gpu_elemwise_careduce():
+    x = theano.tensor.fmatrix()
+    o = (x * x).sum()
+    f = theano.function([x], o, mode=mode_with_gpu)
+    topo = f.maker.fgraph.toposort()
+    assert len(topo) == 3
+    assert topo[1].op.pre_scalar_op == theano.scalar.sqr
+    data = numpy.random.rand(3, 4).astype('float32')
+    utt.assert_allclose(f(data), (data * data).sum())
+
+    o = (x * x).sum(axis=1)
+    f = theano.function([x], o, mode=mode_with_gpu)
+    topo = f.maker.fgraph.toposort()
+    assert len(topo) == 3
+    assert topo[1].op.pre_scalar_op == theano.scalar.sqr
+    utt.assert_allclose(f(data), (data * data).sum(axis=1))
+
+
 def test_huge_elemwise_fusion():
     """ Test the the GpuElemwise fusion work correctly
         We check that we fuse one node with part of its input

theano/sandbox/gpuarray/basic_ops.py

@@ -403,6 +403,7 @@ class GpuFromHost(Op):
         return """
         PyArrayObject *%(name)s_tmp;
         %(name)s_tmp = PyArray_GETCONTIGUOUS(%(inp)s);
+        int err;
         if (%(name)s_tmp == NULL)
           %(fail)s
 
@@ -411,8 +412,8 @@ class GpuFromHost(Op):
                               (size_t *)PyArray_DIMS(%(name)s_tmp),
                               get_typecode((PyObject *)PyArray_DESCR(%(name)s_tmp)))) {
           Py_BEGIN_ALLOW_THREADS
-          int err = GpuArray_write(&%(out)s->ga, PyArray_DATA(%(name)s_tmp),
-                                   PyArray_NBYTES(%(name)s_tmp));
+          err = GpuArray_write(&%(out)s->ga, PyArray_DATA(%(name)s_tmp),
+                               PyArray_NBYTES(%(name)s_tmp));
           Py_END_ALLOW_THREADS
           Py_DECREF(%(name)s_tmp);
           if (err != GA_NO_ERROR) {

theano/sandbox/gpuarray/opt.py

@@ -867,12 +867,13 @@ def local_gpu_elemwise_careduce(node):
             isinstance(node.inputs[0].owner.op, GpuElemwise) and
             # The Op support all scalar with 1 inputs.  We don't
             # automatically add more case, as some like trigonometic
-            # operation with some reduction pattern will probably result
-            # to slow down.
+            # operation with some reduction pattern will probably results
+            # in slow down.
             isinstance(node.inputs[0].owner.op.scalar_op, scalar.basic.Sqr)):
         op = node.op
         inp = node.inputs[0].owner.inputs[0]
         return [GpuCAReduceCuda(scalar_op=op.scalar_op,
+                                axis=op.axis,
                                 reduce_mask=op.reduce_mask,
                                 pre_scalar_op=scalar.basic.sqr)(inp)]
 

theano/sandbox/gpuarray/tests/test_opt.py

@@ -212,7 +212,15 @@ def test_local_gpu_elemwise_careduce():
     topo = f.maker.fgraph.toposort()
     assert len(topo) == 3
     assert topo[1].op.pre_scalar_op == theano.scalar.sqr
-    f(numpy.random.rand(3, 4).astype(theano.config.floatX))
+    data = numpy.random.rand(3, 4).astype(theano.config.floatX)
+    utt.assert_allclose(f(data), (data * data).sum())
+
+    o = (x * x).sum(axis=1)
+    f = theano.function([x], o, mode=mode_with_gpu)
+    topo = f.maker.fgraph.toposort()
+    assert len(topo) == 3
+    assert topo[1].op.pre_scalar_op == theano.scalar.sqr
+    utt.assert_allclose(f(data), (data * data).sum(axis=1))
 
 
 def test_local_gpu_subtensor():