Elementwise equality comparison with zero-length void dtype fails on NumPy 1.21

@seberg

Reproducing code example:

With NumPy 1.21:

>>> import numpy as np
>>> x = np.zeros((3,), np.void)
>>> x == x
/usr/local/lib/python3.7/dist-packages/ipykernel_launcher.py:4: FutureWarning: elementwise == comparison failed and returning scalar instead; this will raise an error or perform elementwise comparison in the future.
False

I believe the error comes from these lines:

numpy/numpy/core/src/multiarray/arrayobject.c

Lines 1361 to 1375 in 1581032

    
           _res = PyArray_CanCastTypeTo(PyArray_DESCR(self), 
        
                                        PyArray_DESCR(array_other), 
        
                                        NPY_EQUIV_CASTING); 
        
           if (_res == 0) { 
        
               /* 2015-05-07, 1.10 */ 
        
               Py_DECREF(array_other); 
        
               if (DEPRECATE_FUTUREWARNING( 
        
                       "elementwise == comparison failed and returning scalar " 
        
                       "instead; this will raise an error or perform " 
        
                       "elementwise comparison in the future.") < 0) { 
        
                   return NULL; 
        
               } 
        
               Py_INCREF(Py_False); 
        
               return Py_False; 
        
           }

On NumPy 1.20, the output is correct:

array([ True,  True,  True])

Or conversely, if we use a non-zero length void dtype, e.g.,

>>> x = np.array([b'foo', b'bar'], np.void)
>>> x == x
array([ True,  True])

CC @seberg since this smells similar to #19085

Probably should just fix this local to comparisons, unless you found the issue also in another context? It is not like you can do a whole lot with voids, especially empty ones...

We have only encountered this in comparisons, so that would work for us.

The empty void dtype is convenient for JAX as a logical "placeholder" for arrays with a particular shape that doesn't require allocating memory (the arrays are always zero).

Would be nice if you can test gh-19322, just to be sure.

@seberg

@seberg I can confirm that NumPy built at your PR gh-19322 fixes the test failures we saw in the JAX test suite with NumPy 1.21.

Any chance we can get this fix into 1.21.1 if/when that is a thing?

Yes, that PR is practically backport only, the whole branch should use different logic in 1.22/2.0

Disable float0 tests that fail under NumPy 1.21.

Navigation Menu

Repository navigation

Elementwise equality comparison with zero-length void dtype fails on NumPy 1.21 #19305

Reproducing code example:

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	_res = PyArray_CanCastTypeTo(PyArray_DESCR(self),
	PyArray_DESCR(array_other),
	NPY_EQUIV_CASTING);
	if (_res == 0) {
	/* 2015-05-07, 1.10 */
	Py_DECREF(array_other);
	if (DEPRECATE_FUTUREWARNING(
	"elementwise == comparison failed and returning scalar "
	"instead; this will raise an error or perform "
	"elementwise comparison in the future.") < 0) {
	return NULL;
	}
	Py_INCREF(Py_False);
	return Py_False;
	}

Elementwise equality comparison with zero-length void dtype fails on NumPy 1.21 #19305

Description

Reproducing code example:

Activity

seberg commented on Jun 23, 2021

shoyer commented on Jun 23, 2021

seberg commented on Jun 24, 2021

hawkinsp commented on Jun 24, 2021

seberg commented on Jun 24, 2021

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions