invalid result on master when aggregating flat array with size=0

[shoyer]

import numpy as np
from numba import guvectorize

@guvectorize(['void(float64[:, :], float64[:])'],
              '(x, y)->()')
def my_sum(a, out):
    asum = 0.0
    for ai in a.flat:
        asum += ai
    out[0] = asum

x = np.arange(0.0).reshape(2, 0)
print repr(x)
print repr(my_sum(x)) 

array([], shape=(2, 0), dtype=float64)
1.7272338168659652e-77

I'm running the master version of numba (actually 0.15.1-28-g2f5d025) and numpy 1.9.0 on Python 2.7/OSX (via conda).

This is indeed pathological data, but this part of the bottleneck test suite failed when I tested it for numbagg.

[pitrou]

Is it specific to the use of guvectorize? Or does it also apply to normal jitted functions?

[shoyer]

@pitrou Good question. Here is my jit version, which for which I cannot reproduce the issue:

import numpy as np
from numba import jit

@jit
def my_sum(a):
    asum = 0.0
    for ai in a.flat:
        asum += ai
    return asum

x = np.arange(0.0).reshape(2, 0)
print repr(x)
print repr(my_sum(x)) 

[sklam]

This maybe a numpy bug. Consider:

import numpy as np
import numpy.core.umath_tests as ut

x = np.arange(0.0).reshape(2, 0)
y = np.arange(0.0).reshape(0, 2)
print(x, x.shape, x.data.nbytes)
print(y, y.shape, y.data.nbytes)
print(ut.matrix_multiply(x, y))

which prints:

[] (2, 0) 0
[] (0, 2) 0
[[  1.72723371e-077   1.72723371e-077]
 [  2.12435947e-314   2.78134365e-309]]

[shoyer]

Interesting. I should note that this does work on the numba 0.15.1 (I accidentally omitted that in my original report), so it is likely related to work refactoring .flat.

Also might be related to this numpy issue: numpy/numpy#5195

[pitrou]

The fact that it works with @jit is a bit fishy. It would be nice to know what the a array looks like inside the @guvectorize version.

[shoyer]

@pitrou I tried inserting some print statements, but of course that causes the function to be executed in python mode, which removes the bug.

[pitrou]

If you use the print() function and only print integers and floats, the function should be able to compile in nopython mode (add nopython=True to the jit() call to ensure this).

[shoyer]

OK, perhaps this output may be helpful:

import numpy as np
from numba import guvectorize

@guvectorize(['void(float64[:, :], float64[:])'],
              '(x, y)->()', nopython=True)
def my_sum(a, out):
    asum = 0.0
    print(asum)
    adding = 99999999 # some sentinel value
    for n, ai in enumerate(a.flat):
        print(adding)
        print(ai)
        asum += ai
    print(asum)
    out[0] = asum

for shape in [(0, 0), (0, 2), (2, 0)]:
    print('\nshape {}:'.format(shape))
    x = np.arange(0.0).reshape(shape)
    print repr(my_sum(x)) 

shape (0, 0):
0.0
0.0
0.0

shape (0, 2):
0.0
99999999
0.0
99999999
3.10503636984e+231
3.10503636984e+231
3.1050363698376418e+231

shape (2, 0):
0.0
99999999
1.32768739285e-315
1.32768739285e-315
1.3276873928472899e-315

It looks like .flat is sometimes reading from unallocated memory.

[pitrou]

Thank you. Yes, it's an issue with .flat. It only manifests on the code path for non-contigouous arrays (you can also trigger it with explicit typing in the jit() call, e.g. @jit('void(float64[:, :])', nopython=True). The problem is that if one dimension is zero-sized (but not all), the iteration logic becomes incorrect. IMO the best fix will be to add a special path if there is a zero-sized dimension, since the iterator is empty then.