Merge pull request #5321 from ReyhaneAskari/fix_5253

theano/gpuarray/basic_ops.py

@@ -1245,8 +1245,19 @@ class GpuJoin(HideC, Join):
 
     """
     _f16_ok = True
+    __props__ = ("view",)
     params_type = gpu_context_type
 
+    def __init__(self, view=-1):
+        self.view = view
+        if view != -1:
+            # since the first input is always the axis, the tensors
+            # start from index 1.
+            self.view_map = {0: [1 + view]}
+
+    def __str__(self):
+        return Join.__str__(self)
+
     def make_node(self, axis, *tensors):
         node = Join.make_node(self, axis, *tensors)
 
@@ -1265,61 +1276,98 @@ class GpuJoin(HideC, Join):
 
     def perform(self, node, axis_and_tensors, out_, ctx):
         out, = out_
+        view = self.view
         axis = int(axis_and_tensors[0])
+        tensors = axis_and_tensors[1:]
+
         if axis < -axis_and_tensors[1].ndim:
             raise IndexError
         if axis < 0:
             axis += axis_and_tensors[1].ndim
-        tensors = axis_and_tensors[1:]
-        out[0] = pygpu.concatenate(tensors, axis=axis, context=ctx).astype(
-            node.outputs[0].dtype)
+        # we check these tensors for being empty.
+        if (view != -1) and numpy.all(
+                [tensor.shape[axis] == 0 for tensor in
+                 tensors[0:view] + tensors[view + 1:]]):
+            out[0] = tensors[view]
+        else:
+            out[0] = pygpu.concatenate(tensors, axis=axis, context=ctx).astype(
+                node.outputs[0].dtype)
 
     def c_code_cache_version(self):
-        return (2,)
+        return (3,)
 
     def c_support_code(self):
         return """
-#if PY_MAJOR_VERSION >= 3
-#define PyInt_AsLong PyLong_AsLong
-#endif
-"""
+        #if PY_MAJOR_VERSION >= 3
+        #define PyInt_AsLong PyLong_AsLong
+        #endif
+        """
+
+    def c_headers(self):
+        return ['<numpy_compat.h>']
 
     def c_code(self, node, name, inputs, out_, sub):
+        axis, tensors = inputs[0], inputs[1:]
         copy_to_list = []
         restype = pygpu.gpuarray.dtype_to_typecode(node.outputs[0].dtype)
-        for i, inp in enumerate(inputs[1:]):
+        view = self.view
+        non_empty_tensor = tensors[view]
+        for i, inp in enumerate(tensors):
             copy_to_list.append("als[%s] = &%s->ga;" % (i, inp))
-        return """
-const GpuArray **als = (const GpuArray **)PyMem_Malloc(sizeof(GpuArray *) *
+
+        n = len(tensors)
+        fail = sub['fail']
+        out = out_[0]
+        copy_inputs_to_list = '\n'.join(copy_to_list)
+        restype = restype
+        ctx = sub['params']
+
+        code = """
+        const GpuArray **als = (const GpuArray **)PyMem_Malloc(sizeof(GpuArray *) *
                                                        %(n)s);
-if (als == NULL) {
-  PyErr_NoMemory();
-  %(fail)s
-}
-%(copy_inputs_to_list)s
-Py_XDECREF(%(out)s);
-{
-int axis = PyInt_AsLong((PyObject *)%(axis)s);
-if (axis < 0) {
-  if (axis == -1 && PyErr_Occurred()) {
-    %(fail)s
-  }
-  axis += als[0]->nd;
-  if (axis < 0) {
-    PyErr_SetString(PyExc_IndexError, "invalid axis");
-    %(fail)s
-  }
-}
-%(out)s = pygpu_concatenate(als, %(n)s, axis,
-                            %(restype)s, (PyObject *)&PyGpuArrayType,
-                            %(ctx)s);
-}
-PyMem_Free(als);
-if (%(out)s == NULL)
-  %(fail)s
-        """ % dict(n=len(inputs[1:]), fail=sub['fail'], out=out_[0],
-                   axis=inputs[0], copy_inputs_to_list='\n'.join(copy_to_list),
-                   restype=restype, ctx=sub['params'])
+        if (als == NULL) {
+            PyErr_NoMemory();
+            %(fail)s
+        }
+        %(copy_inputs_to_list)s
+        Py_XDECREF(%(out)s);
+        {
+            int axis = PyInt_AsLong((PyObject *)%(axis)s);
+            if (axis < 0) {
+                if (axis == -1 && PyErr_Occurred()) {
+                    %(fail)s
+                }
+                axis += als[0]->nd;
+                if (axis < 0) {
+                    PyErr_SetString(PyExc_IndexError, "invalid axis");
+                    %(fail)s
+                }
+            }
+
+            int tensors_lens_sum;
+            if(%(view)s != -1) {
+                tensors_lens_sum = 0;
+                for(int i=0; i < %(n)s; i++){
+                    tensors_lens_sum += als[i]->dimensions[axis];
+                }
+                tensors_lens_sum -= PyGpuArray_DIM(%(non_empty_tensor)s, axis);
+            }
+
+            if(%(view)s != -1 && tensors_lens_sum == 0) {
+                Py_INCREF(%(non_empty_tensor)s);
+                %(out)s = %(non_empty_tensor)s;
+            }else{
+                %(out)s = pygpu_concatenate(als, %(n)s, axis,
+                                            %(restype)s, (PyObject *)&PyGpuArrayType,
+                                            %(ctx)s);
+            }
+        }
+        PyMem_Free(als);
+        if (%(out)s == NULL)
+            %(fail)s
+
+        """ % locals()
+        return code
 
 gpu_join = GpuJoin()
 

theano/gpuarray/tests/test_basic_ops.py

@@ -453,3 +453,24 @@ def test_hostfromgpu_shape_i():
     assert isinstance(topo[1].op, theano.compile.Shape_i)
     assert isinstance(topo[2].op, theano.tensor.opt.MakeVector)
     assert tuple(f(cv)) == (5, 4)
+
+
+def test_Gpujoin_inplace():
+    """Test Gpujoin to work inplace.
+
+    This function tests the case when several elements are passed to the
+    Gpujoin function but all except one of them are empty. In this case
+    Gpujoin should work inplace and the output should be the view of the
+    non-empty element.
+    """
+    s = T.lscalar()
+    data = numpy.array([3, 4, 5], dtype=theano.config.floatX)
+    x = gpuarray_shared_constructor(data, borrow=True)
+    z = T.zeros((s,))
+
+    join = GpuJoin(view=0)
+    c = join(0, x, z)
+
+    f = theano.function([s], theano.Out(c, borrow=True))
+    assert x.get_value(borrow=True, return_internal_type=True) is f(0)
+    assert numpy.allclose(f(0), [3, 4, 5])

theano/tensor/basic.py

@@ -3896,9 +3896,12 @@ class Join(Op):
 
     def __str__(self):
         if self.view == -1:
-            return "Join"
+            return self.__class__.__name__
         else:
-            return super(Join, self).__str__()
+            return "%s{%s}" % (
+                self.__class__.__name__,
+                ", ".join("%s=%r" % (p, getattr(self, p))
+                          for p in self.__props__))
 
     def __setstate__(self, d):
         self.__dict__.update(d)
@@ -4044,28 +4047,37 @@ class Join(Op):
         out, = outputs
         fail = sub['fail']
         adtype = node.inputs[0].type.dtype_specs()[1]
-        code = """
-        int axis = ((%(adtype)s *)PyArray_DATA(%(axis)s))[0];
-        int tensors_lens_sum = 0""" % locals()
+        copy_to_list = []
+
         for i, inp in enumerate(tensors):
-            code += """ + PyArray_DIM(%(inp)s, axis) """ % locals()
-        code += """;\n
-        tensors_lens_sum -= PyArray_DIM(%(non_empty_tensor)s, axis);
+            copy_to_list.append(
+                """Py_INCREF(%s);
+                   PyList_SetItem(list, %s, (PyObject*)%s);"""
+                % (inp, i, inp))
+
+        copy_inputs_to_list = '\n'.join(copy_to_list)
+        n = len(tensors)
+        khar = "printf(\"tensors_lens_sum: %d\", tensors_lens_sum);"
 
-        if(%(view)s != -1 && tensors_lens_sum == 0){
+        code = """
+        int axis = ((%(adtype)s *)PyArray_DATA(%(axis)s))[0];
+        PyObject* list = PyList_New(%(l)s);
+        %(copy_inputs_to_list)s
+        int tensors_lens_sum;
+        if(%(view)s != -1) {
+            tensors_lens_sum = 0;
+
+            for(int i=0; i < %(n)s; i++){
+                tensors_lens_sum += PyArray_DIM((PyArrayObject *)(PyList_GetItem(list, i)), axis);
+            }
+            %(khar)s
+            tensors_lens_sum -= PyArray_DIM(%(non_empty_tensor)s, axis);
+        }
+        if(%(view)s != -1 && tensors_lens_sum == 0) {
             Py_XDECREF(%(out)s);
             Py_INCREF(%(non_empty_tensor)s);
             %(out)s = %(non_empty_tensor)s;
-        }
-        else{
-            PyObject* list = PyList_New(%(l)s);
-        """ % locals()
-        for i, inp in enumerate(tensors):
-            code += """
-            Py_INCREF(%(inp)s);
-            PyList_SetItem(list, %(i)s, (PyObject*)%(inp)s);
-            """ % locals()
-        code += """
+        }else{
             //PyObject* PyArray_Concatenate(PyObject* obj, int axis)
             int ndim = PyArray_NDIM(%(input_1)s);
             if( axis < -ndim ){