numpy.greater_equal.outer uses stub from pypy's micronumpy and not numpy ufunc...

[M]

EDIT: Simpler bug reproduction on the bottom of report. EDIT2: Found "real" problem, probably, see last comments.

Not works in: pypy 5.1.1 with numpy from https://bitbucket.org/pypy/numpy/ ( 2de5fffe8210079d591a132b6d9a82eb619d5967 )

Works in: normal numpy in python2 and python3

Error:

  File "lab4.py", line 12, in my_rq
    Q,R = numpy.linalg.qr(numpy.flipud(A).T)
  File "/usr/local/Cellar/pypy/5.1.1/libexec/site-packages/numpy/linalg/linalg.py", line 831, in qr
    return wrap(q), wrap(triu(r))
  File "/usr/local/Cellar/pypy/5.1.1/libexec/site-packages/numpy/lib/twodim_base.py", line 484, in triu
    mask = tri(*m.shape[-2:], k=k-1, dtype=bool)
  File "/usr/local/Cellar/pypy/5.1.1/libexec/site-packages/numpy/lib/twodim_base.py", line 415, in tri
    arange(-k, M-k, dtype=_min_int(-k, M - k)))
ValueError: outer product only supported for binary functions

Code:

def my_rq(A):
    Q,R = numpy.linalg.qr(numpy.flipud(A).T)
    R = numpy.flipud(R.T)
    Q = Q.T
    return R[:,::-1], Q[::-1,:]

This code generates RQ decomposition from numpy.linalg.qr decomposition.

A matrix:

[[ 0.45827554 -0.29474237 -0.01395746]
 [-0.05085589 -0.0545847  -0.54105993]
 [ 0.10900958  0.17834548 -0.04426782]]

---

This simple code works with normal python numpy, but not with pypy's numpy:

import numpy

a = numpy.arange(3)
b = numpy.arange(1,4)
print(a)
print(b)
print(numpy.greater_equal.outer(a,b))

---

[M]

I've upgraded to new 5.1.1 tarball after https://bitbucket.org/pypy/pypy/issues/2292/511-source-tarball-is-really-510 but bug still persists.

[M]

numpy.flipud(A).T is the same in both versions, and equals:

[[ 0.10900958 -0.05085589  0.45827554]
 [ 0.17834548 -0.0545847  -0.29474237]
 [-0.04426782 -0.54105993 -0.01395746]]

So it has to be something wrong with numpy.linalg.qr, triu or tri.

[M]

I've injected some prints into both numpies, and it seems r is the same in pypy version:

[[-0.21365823 -0.04059207  0.00932179]
 [ 0.55272163  0.54466866  0.00130836]
 [-0.13719317  0.91389158  0.54497316]]

Later I've injected into tri and this:

    print(N)
    print(-k)
    print(M-k)
    print(_min_int(0, N))
    print(_min_int(-k, M - k))

gave the same output in both.

[M]

print(arange(N, dtype=_min_int(0, N)),
                            arange(-k, M-k, dtype=_min_int(-k, M - k)))

Injected into tri gives in both versions: [0 1 2] [1 2 3]

so it must be a bug in greater_equal.outer

[M]

It looks that numpy.ufunc outer is not defined in this numpy. And in pypy there is some stub ( why it's in pypy code and not here? ) that does only:

    def _outer(self, space, __args__):
        raise OperationError(space.w_ValueError, space.wrap(
            "outer product only supported for binary functions"))

( /module/micronumpy/ufuncs.py in pypy repo )

[M]

There is outer() implemetation in core/src/umath/ufunc_object.c... Why then stub version from micronumpy is used and not this one?

[mattip]

fixed in pypy/pypy changeset 0c3ac8d64955, please reopen if I missed an edge case