Reuse allocated memory when possible.

theano/tensor/__init__.py

@@ -62,4 +62,4 @@ from theano.gradient import Rop, Lop, grad, numeric_grad, verify_grad, \
 
 from theano.tensor.sort import sort, argsort
 from theano.tensor.extra_ops import (DiffOp, bincount, squeeze,
-                       repeat, bartlett, fill_diagonal)
+                       repeat, bartlett, fill_diagonal, cumsum)

theano/tensor/extra_ops.py

@@ -49,7 +49,7 @@ class CumsumOp(theano.Op):
 
     def infer_shape(self, node, shapes):
         if self.axis is None:
-            return [(np.prod(shapes[0]),)]  # Flatten
+            return [(tensor.prod(shapes[0]),)]  # Flatten
 
         return shapes
 
@@ -59,11 +59,14 @@ class CumsumOp(theano.Op):
         axis = self.axis
         fail = sub['fail']
 
-        if self.axis is None:
+        if self.axis is None or (self.axis == 0 and node.inputs[0].ndim == 1):
             code = """
                 npy_intp shape[1] = { PyArray_SIZE(%(x)s) };
+                if(!(%(z)s && PyArray_DIMS(%(z)s)[0] == shape[0]))
+                {
                     Py_XDECREF(%(z)s);
                     %(z)s = (PyArrayObject*) PyArray_SimpleNew(1, shape, type_num_%(x)s);
+                }
 
                 if (!%(z)s)
                     %(fail)s;
@@ -73,8 +76,11 @@ class CumsumOp(theano.Op):
             """ % locals()
         else:
             code = """
+                if(!(%(z)s && PyArray_CompareLists(PyArray_DIMS(%(z)s), PyArray_DIMS(%(x)s), PyArray_NDIM(%(x)s)) ))
+                {
                     Py_XDECREF(%(z)s);
                     %(z)s = (PyArrayObject*) PyArray_SimpleNew(PyArray_NDIM(%(x)s), PyArray_DIMS(%(x)s), type_num_%(x)s);
+                }
 
                 if (!%(z)s)
                     %(fail)s;
@@ -86,10 +92,10 @@ class CumsumOp(theano.Op):
         return code
 
     def c_code_cache_version(self):
-        return (1,)
+        return (2,)
 
     def __str__(self):
-        return self.__class__.__name__
+        return "%s{%s}" % (self.__class__.__name__, self.axis)
 
 
 def cumsum(x, axis=None):

theano/tensor/tests/test_extra_ops.py

@@ -22,15 +22,26 @@ class TestCumsumOp(utt.InferShapeTester):
 
     def test_cumsumOp(self):
         x = T.tensor3('x')
-        a = np.random.random((30, 50, 20)).astype(config.floatX)
+        a = np.random.random((3, 5, 2)).astype(config.floatX)
+        b = np.random.random((30, 5, 2)).astype(config.floatX)
 
         f = theano.function([x], cumsum(x), mode="DebugMode")
         assert np.allclose(np.cumsum(a), f(a))  # Test axis=None
 
+        # Test without garbage collector
+        f = theano.function([x], cumsum(x).sum(), mode=theano.compile.Mode(linker="cvm_nogc", optimizer="fast_run") )
+        assert np.allclose(np.cumsum(a).sum(), f(a))  # Test axis=None
+        assert np.allclose(np.cumsum(b).sum(), f(b))  # Would fail without re-allocation
+
         for axis in range(len(a.shape)):
             f = theano.function([x], cumsum(x, axis=axis), mode="DebugMode")
             assert np.allclose(np.cumsum(a, axis=axis), f(a))
 
+            # Test without garbage collector
+            f = theano.function([x], cumsum(x, axis=axis).sum(), mode=theano.compile.Mode(linker="cvm_nogc", optimizer="fast_run"))
+            assert np.allclose(np.cumsum(a, axis=axis).sum(), f(a))
+            assert np.allclose(np.cumsum(b, axis=axis).sum(), f(b))  # Would fail without re-allocation
+
     def test_infer_shape(self):
         x = T.tensor3('x')
         a = np.random.random((30, 50, 20)).astype(config.floatX)
@@ -53,7 +64,7 @@ class TestCumsumOp(utt.InferShapeTester):
         utt.verify_grad(self.op, [a])  # Test axis=None
 
         for axis in range(len(a.shape)):
-            utt.verify_grad(CumsumOp(axis=axis), [a])
+            utt.verify_grad(self.op_class(axis=axis), [a])
 
 
 class TestBinCountOp(utt.InferShapeTester):