Merge pull request numpy#8042 from pv/fix-nditer-contig

BUG: core: fix bug in NpyIter buffering with discontinuous arrays

numpy/core/src/multiarray/nditer_api.c

@@ -2334,7 +2334,7 @@ npyiter_copy_to_buffers(NpyIter *iter, char **prev_dataptrs)
                 }
                 else {
                     /* It's all in one stride in the reduce outer loop */
-                    if ((reduce_outerdim > 0) &&
+                    if ((reduce_outerdim == 1) &&
                                     (transfersize/reduce_innersize <=
                                             NAD_SHAPE(reduce_outeraxisdata) -
                                             NAD_INDEX(reduce_outeraxisdata))) {

numpy/core/tests/test_nditer.py

@@ -2374,6 +2374,22 @@ def test_iter_buffering_reduction():
     it.reset()
     assert_equal(it[0], [1, 2, 1, 2])
 
+    # Iterator inner loop should take argument contiguity into account
+    x = np.ones((7, 13, 8), np.int8)[4:6,1:11:6,1:5].transpose(1, 2, 0)
+    x[...] = np.arange(x.size).reshape(x.shape)
+    y_base = np.arange(4*4, dtype=np.int8).reshape(4, 4)
+    y_base_copy = y_base.copy()
+    y = y_base[::2,:,None]
+
+    it = np.nditer([y, x],
+                   ['buffered', 'external_loop', 'reduce_ok'],
+                   [['readwrite'], ['readonly']])
+    for a, b in it:
+        a.fill(2)
+
+    assert_equal(y_base[1::2], y_base_copy[1::2])
+    assert_equal(y_base[::2], 2)
+
 def test_iter_buffering_reduction_reuse_reduce_loops():
     # There was a bug triggering reuse of the reduce loop inappropriately,
     # which caused processing to happen in unnecessarily small chunks

numpy/core/tests/test_ufunc.py

@@ -1257,6 +1257,27 @@ def test_NotImplemented_not_returned(self):
         for f in binary_funcs:
             assert_raises(TypeError, f, a, b)
 
+    def test_reduce_noncontig_output(self):
+        # Check that reduction deals with non-contiguous output arrays
+        # appropriately.
+        #
+        # gh-8036
+
+        x = np.arange(7*13*8, dtype=np.int16).reshape(7, 13, 8)
+        x = x[4:6,1:11:6,1:5].transpose(1, 2, 0)
+        y_base = np.arange(4*4, dtype=np.int16).reshape(4, 4)
+        y = y_base[::2,:]
+
+        y_base_copy = y_base.copy()
+
+        r0 = np.add.reduce(x, out=y.copy(), axis=2)
+        r1 = np.add.reduce(x, out=y, axis=2)
+
+        # The results should match, and y_base shouldn't get clobbered
+        assert_equal(r0, r1)
+        assert_equal(y_base[1,:], y_base_copy[1,:])
+        assert_equal(y_base[3,:], y_base_copy[3,:])
+
 
 if __name__ == "__main__":
     run_module_suite()