Pack/unpack external broken for long double real and complex datatypes

[dalcinl]

Looks like the pack/unpack external operations do not roundtrip for long double (real and complex) datatypes with current ompi@master.

This is my trivial Python reproducer

from mpi4py import MPI
import numpy as np

datatype = MPI.LONG_DOUBLE
typecode = MPI._typecode(datatype)
iarray = np.array([1,2,3], dtype=typecode)
oarray = np.zeros(3, dtype=typecode)

n = datatype.Pack_external_size("external32", len(iarray))
work = bytearray(n)
datatype.Pack_external("external32", iarray, work, 0)
datatype.Unpack_external("external32", work, 0, oarray)

print(iarray)
print(oarray)

Running on my Linux-64 workstation I get the following output (both lines are expected to be the same):

[1. 2. 3.]
[8.40525786e-4933 8.40525786e-4933 1.26078868e-4932]

@jsquyres I understand that supporting external32 for long double [complex] is a bit tricky. You probably have to rely on a compiler quad-precision type like GCC __float128 for a proper implementation, see MPI 3.1 Table 13.2 pp. 580 RFC 1832, Sec. 3.8, but if it cannot be supported, then perhaps you should error rather that let things success silently and end up with broken data? Or maybe this is just a trivial bug?

[gpaulsen]

@markalle You posted a fix for external32, that now needs rebasing. Might PR #7919 fix this issue?

[markalle]

Offhand I doubt my 7919 would be related, although if there is a difference between the system sizeof(long double) and the MPI_Type_size(MPI_LONG_DOUBLE,) then it could be.

I just now ran a similar C program:
https://gist.github.com/markalle/ad7e69f026471e2baa8e842c938d8048
and it was successful, at least on my system where sizeof(long double) and MPI_Type_size(MPI_LONG_DOUBLE,) were both 16.

[bosilca]

With the current master plus PR #8769 both the C and python code give the correct result on x86_64 and ARM (M1).