Merge pull request #802 from jaberg/reshape_c_code

theano/tensor/basic.py

@@ -5256,6 +5256,7 @@ class Reshape(Op):
             raise ValueError('Cannot reshape input of shape %s to shape %s' %
                              (x.shape, shp))
 
+
     def grad(self, inp, grads):
         x, shp = inp
         g_out, = grads
@@ -5298,6 +5299,35 @@ class Reshape(Op):
             oshape.append(os_i)
         return [tuple(oshape)]
 
+    def c_code_cache_version(self):
+        return (1,)
+
+    def c_code(self, node, name, inputs, outputs, sub):
+        if isinstance(node.inputs[0], TensorVariable):
+            x, shp = inputs
+            z, = outputs
+            new_ndim = self.ndim
+            fail = sub['fail']
+            return """
+            assert (%(shp)s->nd == 1);
+            npy_intp new_dims[%(new_ndim)s];
+            PyArray_Dims newshape;
+            newshape.ptr = new_dims;
+            newshape.len = %(new_ndim)s;
+            for (int ii = 0; ii < %(new_ndim)s; ++ii)
+            {
+                // -- We do not want an explicit cast here. the shp can be any
+                // -- int* dtype. The compiler will explicitly upcast it, but
+                // -- will err if this will downcast. This could happen if the
+                // -- user pass an int64 dtype, but npy_intp endup being int32.
+                new_dims[ii] = ((dtype_%(shp)s*)(
+                        %(shp)s->data + ii * %(shp)s->strides[0]))[0];
+            }
+            Py_XDECREF(%(z)s);
+            %(z)s = (PyArrayObject *) PyArray_Newshape(%(x)s, &newshape, NPY_ANYORDER);
+            """ % locals()
+        else:
+            return Op.c_code(self, node, name, inputs, outputs, sub)
 
 def reshape(x, newshape, ndim=None, name=None):
     if ndim is None: