np.logical_xor.accumulate fails on 1.24 on mac

[knutdrand]

Describe the issue:

.Accumulate on logical_xor gives wrong result on newest version. I suspect the same for bitwise_xor.

Reproduce the code example:

import numpy as np

a = np.array([False, False,  True, False, False, False, False,  True, False, False,  True, False, False,  True, False, False, False, False, False, False])
np.logical_xor.accumulate(a)
array([False, False,  True,  True, False, False, False,  True,  True,
       False,  True,  True, False,  True,  True, False, False, False,
       False, False])
# Should be [False, False,  True,  True, True, True, True, False,,,,]

Error message:

No response

NumPy/Python version information:

numpy=1.24
python 3.8,
mac version 12.5.1

Context for the issue:

No response

[seberg]

Again ping @seiko2plus and maybe @Developer-Ecosystem-Engineering. This looks like probably the identical issue as gh-22840 (just now slightly different with the reduction). Presumably introduced by gh-22167.

With a view (although I am not sure how important it is directly without digging in):

>>> a = np.array([False, False,  True, False, False, False, False,  True, False, False,  True, False, False,  True, False, False, False, False, False, False])
>>> np.logical_xor.accumulate(a).view(np.uint8)
array([  0,   0, 254, 254,   0,   0,   0, 254, 254,   0, 254, 254,   0,
       254, 254,   0,   0,   0,   0,   0], dtype=uint8)

[Developer-Ecosystem-Engineering]

Hi @seberg,

Appears related to #21483. The aliases for BOOL_logical_xor was BOOL_not_equal, which was subsequently changed

[seiko2plus]

Memory overlaps! The comparison loops of all data types are free of overlap checking:

numpy/numpy/core/src/umath/fast_loop_macros.h

Lines 381 to 392 in 0d1bb8e

	#define IS_BLOCKABLE_BINARY_BOOL(esize, vsize) \
	(steps[0] == (esize) && steps[0] == steps[1] && steps[2] == (1) && \
	npy_is_aligned(args[1], (esize)) && \
	npy_is_aligned(args[0], (esize)))
	#define IS_BLOCKABLE_BINARY_SCALAR1_BOOL(esize, vsize) \
	(steps[0] == 0 && steps[1] == (esize) && steps[2] == (1) && \
	npy_is_aligned(args[1], (esize)))
	#define IS_BLOCKABLE_BINARY_SCALAR2_BOOL(esize, vsize) \
	(steps[0] == (esize) && steps[1] == 0 && steps[2] == (1) && \
	npy_is_aligned(args[0], (esize)))

numpy/numpy/core/src/umath/loops_comparison.dispatch.c.src

Lines 315 to 328 in 0d1bb8e

	/* argument one scalar */
	if (IS_BLOCKABLE_BINARY_SCALAR1_BOOL(sizeof(@type@), NPY_SIMD_WIDTH)) {
	simd_binary_scalar1_@kind@_@sfx@(args, dimensions[0]);
	return;
	}
	/* argument two scalar */
	else if (IS_BLOCKABLE_BINARY_SCALAR2_BOOL(sizeof(@type@), NPY_SIMD_WIDTH)) {
	simd_binary_scalar2_@kind@_@sfx@(args, dimensions[0]);
	return;
	}
	else if (IS_BLOCKABLE_BINARY_BOOL(sizeof(@type@), NPY_SIMD_WIDTH)) {
	simd_binary_@kind@_@sfx@(args, dimensions[0]);
	return;
	}

+    // multiply by sizeof(@type@) due to SIMD unroll
     /* argument one scalar */
-    if (IS_BLOCKABLE_BINARY_SCALAR1_BOOL(sizeof(@type@), NPY_SIMD_WIDTH)) {
+    if (IS_BLOCKABLE_BINARY_SCALAR1(sizeof(@type@), NPY_SIMD_WIDTH*sizeof(@type@))) {
         simd_binary_scalar1_@kind@_@sfx@(args, dimensions[0]);
         return;
     }
     /* argument two scalar */
-    else if (IS_BLOCKABLE_BINARY_SCALAR2_BOOL(sizeof(@type@), NPY_SIMD_WIDTH)) {
+    else if (IS_BLOCKABLE_BINARY_SCALAR2(sizeof(@type@), NPY_SIMD_WIDTH*sizeof(@type@))) {
         simd_binary_scalar2_@kind@_@sfx@(args, dimensions[0]);
         return;
     }
-    else if (IS_BLOCKABLE_BINARY_BOOL(sizeof(@type@), NPY_SIMD_WIDTH)) {
+    else if (IS_BLOCKABLE_BINARY(sizeof(@type@), NPY_SIMD_WIDTH*sizeof(@type@))) {
         simd_binary_@kind@_@sfx@(args, dimensions[0]);
         return;
     }