The latest PGI compiler (v 20.4) is failing to properly compile OpenMPI (v
4.0.1 or 4.0.4) - everything appears to build fine, but fails to run on
./hello: symbol lookup error:
/share/apps/pgi/20.4/openmpi-4.0.4/lib/pmix/mca_gds_ds21.so: undefined
symbol: __atomic_compare_exchange
closer scrutiny shows:
1- mca_gds_ds21.so is much smaller w/ PGI 20.4 than PGI 19.x versions:
-rwxr-xr-x 1 root root 87K Oct 1 2019
/share/apps/pgi/19.5/openmpi-4.0.1/lib/pmix/mca_gds_ds21.so
-rwxr-xr-x 1 root root 87K Oct 8 2019
/share/apps/pgi/19.9/openmpi-4.0.1/lib/pmix/mca_gds_ds21.so
-rwxr-xr-x 1 root root 38K Jul 8 14:08
/share/apps/pgi/20.4/openmpi-4.0.4/lib/pmix/mca_gds_ds21.so
2- PGI 20.4 fails to properly parse atomic_stdc.h in
opal/mca/pmix/pmix3x/pmix/src/mca/base
I've reported this to PGI, but if someone is aware if a fix/workaround I'd
appreciate it. I have not digged further down to understand the root cause
(too many layers)
FYI, when using PGI 19.9, I see in the output of make:
make[5]: Entering directory
`/local/hpc/pgi/19.9/openmpi-4.0.1/opal/mca/pmix/pmix3x/pmix/src/mca/base'
CC pmix_mca_base_close.lo
CC pmix_mca_base_cmd_line.lo
CC pmix_mca_base_component_compare.lo
ie, no compilation warnings
The same output, when using PGI 20.4, shows
make[5]: Entering directory
`/local/hpc/pgi/20.4/openmpi-4.0.4/opal/mca/pmix/pmix3x/pmix/src/mca/base'
CC pmix_mca_base_close.lo
"/local/hpc/pgi/20.4/openmpi-4.0.4/opal/mca/pmix/pmix3x/pmix/src/atomics/sys/atomic_stdc.h",
line 115: warning: argument of type
"pmix_atomic_int32_t *" is incompatible with parameter of type
"volatile void *"
PMIX_ATOMIC_STDC_DEFINE_FETCH_OP(add, 32, int32_t, +)
^
and a slew of such warnings as of that point. The compilation proceeds,
tho, and make completes.
Clearly PGI 20.4 is barfing on the macros defined in atomic_stdc.h (hence
the smaller .sl and the missing symbols) - any idea why?
[I've successfully built OpenMPI w/ Intel's, GCC and PGI compilers w/
numerous versions of ea compiler]
Stay sane, safe and healthy, 6+ft away! Cheers,
Sylvain
--
