gcd, lcm, subtract, sign, and negative don't match numpy behavior with bool inputs

[eriknw]

Reporting a bug

• I have tried using the latest released version of Numba (most recent is
  visible in the change log (https://github.com/numba/numba/blob/main/CHANGE_LOG).
• I have included a self contained code sample to reproduce the problem.
  i.e. it's possible to run as 'python bug.py'.

I noticed that numba doesn't match NumPy's behavior for np.gcd, np.lcm, and np.sign with bool inputs. Notably, NumPy raises, but numba treats bools as ints:

>>> import numpy as np
>>> import numba
>>> @numba.njit
... def gcd(x, y):
...     return np.gcd(x, y)
...
>>> gcd(True, True)
1
>>> np.gcd(True, True)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
numpy.core._exceptions._UFuncNoLoopError: ufunc 'gcd' did not contain a loop with signature matching types (dtype('bool'), dtype('bool')) -> dtype('bool')
>>> @numba.njit
... def sign(x):
...     return np.sign(x)
...
>>> sign(True)
1
>>> np.sign(True)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
numpy.core._exceptions._UFuncNoLoopError: ufunc 'sign' did not contain a loop with signature matching types dtype('bool') -> dtype('bool')
>>> @numba.njit
... def lcm(x, y):
...     return np.lcm(x, y)
...
>>> lcm(True, True)
1
>>> np.lcm(True, True)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
numpy.core._exceptions._UFuncNoLoopError: ufunc 'lcm' did not contain a loop with signature matching types (dtype('bool'), dtype('bool')) -> dtype('bool')

I can get by with the current behavior, but thought you would want to know.

[eriknw]

Two more: subtract and negative

>>> @numba.njit
... def subtract(x, y):
...     return np.subtract(x, y)
...
>>> subtract(True, True)
0
>>> np.subtract(True, True)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: numpy boolean subtract, the `-` operator, is not supported, use the bitwise_xor, the `^` operator, or the logical_xor function instead.

[stuartarchibald]

Thanks for the report, I can reproduce.

[stuartarchibald]

Think this might need some logic like:

            if not set([x.char for x in np_input_types]).intersection(set([x for x in ufunc_inputs])):
                continue

around here:

numba/numba/np/numpy_support.py

Lines 521 to 523 in fa3630e

	if found:
	# Found: determine the Numba types for the loop's inputs and
	# outputs.

i.e. once a loop has been found, require that at least one of the input types is present in the selected loop version. This may however be too restrictive.

[czgdp1807]

I am able to reproduce the incorrect behaviour of Numba. In addition, I added your check as follows and I get the following output with sign example in the code provided by @eriknw.

diff

diff --git a/numba/np/numpy_support.py b/numba/np/numpy_support.py
index eea1b612d..27b1558cd 100644
--- a/numba/np/numpy_support.py
+++ b/numba/np/numpy_support.py
@@ -521,6 +521,13 @@ def ufunc_find_matching_loop(ufunc, arg_types):
         if found:
             # Found: determine the Numba types for the loop's inputs and
             # outputs.
+            np_input_types_set = set([x.char for x in np_input_types])
+            output_types_set = set([x for x in ufunc_inputs])
+            if not np_input_types_set.intersection(output_types_set):
+                print("found but not correct: ", [x.char for x in np_input_types], [x for x in ufu
nc_inputs])
+                continue
+            else:
+                print("found and correct: ", [x.char for x in np_input_types], [x for x in ufunc_i
nputs])
             try:
                 inputs = choose_types(input_types, ufunc_inputs)
                 outputs = choose_types(output_types, ufunc_outputs)
@@ -534,8 +541,10 @@ def ufunc_find_matching_loop(ufunc, arg_types):
             except errors.NumbaNotImplementedError:
                 # One of the selected dtypes isn't supported by Numba
                 # (e.g. float16), try other candidates
+                print("errored")
                 continue
             else:
+                print("returning, :", inputs, outputs, candidate)
                 return UFuncLoopSpec(inputs, outputs, candidate)
 
     return None

Output

found and correct:  ['d', 'd'] ['d', 'd']
returning, : [float64, float64] [float64] dd->d
found and correct:  ['d', 'd'] ['d', 'd']
returning, : [float64, float64] [float64] dd->d
found but not correct:  ['?'] ['b']
found but not correct:  ['?'] ['B']
found but not correct:  ['?'] ['h']
found but not correct:  ['?'] ['H']
found but not correct:  ['?'] ['i']
found but not correct:  ['?'] ['I']
found but not correct:  ['?'] ['l']
found but not correct:  ['?'] ['L']
found but not correct:  ['?'] ['q']
found but not correct:  ['?'] ['Q']
found but not correct:  ['?'] ['f']
found but not correct:  ['?'] ['d']
found but not correct:  ['?'] ['g']
found but not correct:  ['?'] ['F']
found but not correct:  ['?'] ['D']
found but not correct:  ['?'] ['G']
found but not correct:  ['?'] ['b']
found but not correct:  ['?'] ['B']
found but not correct:  ['?'] ['h']
found but not correct:  ['?'] ['H']
found but not correct:  ['?'] ['i']
found but not correct:  ['?'] ['I']
found but not correct:  ['?'] ['l']
found but not correct:  ['?'] ['L']
found but not correct:  ['?'] ['q']
found but not correct:  ['?'] ['Q']
found but not correct:  ['?'] ['f']
found but not correct:  ['?'] ['d']
found but not correct:  ['?'] ['g']
found but not correct:  ['?'] ['F']
found but not correct:  ['?'] ['D']
found but not correct:  ['?'] ['G']
found but not correct:  ['?'] ['b']
found but not correct:  ['?'] ['B']
found but not correct:  ['?'] ['h']
found but not correct:  ['?'] ['H']
found but not correct:  ['?'] ['i']
found but not correct:  ['?'] ['I']
found but not correct:  ['?'] ['l']
found but not correct:  ['?'] ['L']
found but not correct:  ['?'] ['q']
found but not correct:  ['?'] ['Q']
found but not correct:  ['?'] ['f']
found but not correct:  ['?'] ['d']
found but not correct:  ['?'] ['g']
found but not correct:  ['?'] ['F']
found but not correct:  ['?'] ['D']
found but not correct:  ['?'] ['G']
found but not correct:  ['?'] ['b']
found but not correct:  ['?'] ['B']
found but not correct:  ['?'] ['h']
found but not correct:  ['?'] ['H']
found but not correct:  ['?'] ['i']
found but not correct:  ['?'] ['I']
found but not correct:  ['?'] ['l']
found but not correct:  ['?'] ['L']
found but not correct:  ['?'] ['q']
found but not correct:  ['?'] ['Q']
found but not correct:  ['?'] ['f']
found but not correct:  ['?'] ['d']
found but not correct:  ['?'] ['g']
found but not correct:  ['?'] ['F']
found but not correct:  ['?'] ['D']
found but not correct:  ['?'] ['G']
Traceback (most recent call last):
  File "../debug.py", line 320, in <module>
    test_np_gcd_lcm_sub()
  File "../debug.py", line 310, in test_np_gcd_lcm_sub
    print(sign(True))
  File "/Users/czgdp1807/numba_project/numba/numba/core/dispatcher.py", line 468, in _compile_for_args
    error_rewrite(e, 'typing')
  File "/Users/czgdp1807/numba_project/numba/numba/core/dispatcher.py", line 409, in error_rewrite
    raise e.with_traceback(None)
numba.core.errors.TypingError: Failed in nopython mode pipeline (step: nopython frontend)
No implementation of function Function(<ufunc 'sign'>) found for signature:
 
 >>> sign(bool)
 
There are 2 candidate implementations:
  - Of which 2 did not match due to:
  Overload in function 'Numpy_rules_ufunc.generic': File: numba/core/typing/npydecl.py: Line 98.
    With argument(s): '(bool)':
   Rejected as the implementation raised a specific error:
     TypingError: can't resolve ufunc sign for types [bool]
  raised from /Users/czgdp1807/numba_project/numba/numba/core/typing/npydecl.py:107

During: resolving callee type: Function(<ufunc 'sign'>)
During: typing of call at ../debug.py (308)


File "../debug.py", line 308:
    def sign(x):
        return np.sign(x)
        ^

I have a couple of questions from the output,

1. What is meant by np_input_types and ufunc_types? As far as I can figure out, np_input_types is the actual type of the input provided by the user, here True and ufunc_types is made available from keys of ufunc_db. Am I right?
2. Why returning:, happens multiple times. I noticed that the return UFuncLoopSpec(inputs, outputs, candidate) gets executed multiple times. I am unable to figure out, why?

Thanks.