np.divide.accumulate gives incorrect results for float64 on Numpy 1.15.0 using Intel compiler

I have a unit test failing on np.divide.accumulate with numpy 1.15.0 compiled with the Intel compiler. The same code is fine with a local build of GCC.

Perhaps this is related to ENH: umath: don’t make temporary copies for in-place accumulation (https://github.com/numpy/numpy/pull/10665) where @juliantaylor fixed numpy/core/src/umath/loops.c.src for GCC’s ivdep pragma. That was reviewed by @eric-wieser .

Reproducing code example

Source code was 1.15.0 compiled with the Intel compiler 2018.1.163. where np.divide.accumulate fails in the numpy unit tests.

I can reproduce with this Python 2.7 test case for float64 dtype. It runs correctly for float32 and float128 on the Intel compiler. It also runs correctly for float64 using GCC instead of Intel.

$ python -c "import numpy as np; acc = np.divide.accumulate;  
  a = np.ones(8, dtype=np.float64);  
  print acc(a, out=2*np.ones_like(a)); print acc(a); print acc(a); print acc(a); 
  print acc(a, out=2*np.ones_like(a))"
[1. 1. 2. 2. 2. 2. 2. 2.]
[1. 1. 1. 1. 2. 2. 2. 2.]
[1. 1. 1. 1. 1. 1. 2. 2.]
[1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 2. 2. 2. 2. 2. 2.]

Notice how only the next two elements are set each time np.divide.accumulate runs. That suggests the float64s are being processed in pairs. Then in the final line, specifying the output array again returns to the first line’s incorrect behavior.

Numpy/Python version information:

numpy 1.15.0 build with ICC 2018.1.163.
python 2.7.11 built with GCC 4.9.3.

[seberg]

Thanks @svenevs just in case, one thing that would be nice is to ensure that our test suits finds these errors (if possible). Probably that is already the case, but if not it would be good to add them. After that anyone using these compilers will know that something is off.

[seberg]

@svenevs you should probably start with doing modifications here:

https://github.com/numpy/numpy/blob/1dfb0ab7358b0cae068b41b3d07452713594ef6b/numpy/core/src/umath/loops.c.src#L1820

which are the loops that get executed for the typical floating point types. Note that “kind” in that code bunch can be divide