Fails to Compile with IBM XLF 16.1.1

I noticed that some XLF support was mentioned in an older change log. After examining Makefile.power, I noticed that the version being called out is XLF 11.1 (paths reference /opt/ibmcmp/xlf/11.1).

I first compiled with the default settings and ran into linking errors. I then adjusted the paths in Makefile.power to /opt/ibm/xlf/16.1.1 and ran into the same linking errors.

I'm attaching the compressed output for make VERBOSE=1 2>&1 | tee ~/oblas.log as the original was 4 MB.

oblas.log.gz

[martin-frbg]

The references to xlf 11 are only in optional sections for linking to lIBM's ESSL library. What appears to happen is that the gensymbol script hits some kind of incompatibility with your shell or basic os tools, and generates names with a trailing 1 for all Lapack function. This results in an incomplete shared library, while the static libopenblas.a is probably usable.

[martin-frbg]

which operating system version are you seeing this with ?

The OS is Fedora 38. I didn't realise linking to ESSL was an option. Is there a flag for it? The modern essl library installs to /opt/ibmmath/essl/<Version>/lib64

To test your idea of the shared library being the issue, I compiled again with the no shared library flag. I get the following error:

xlf -O2 -q64 -openmp  -O2 -qrecur -qnosave -mcpu=power9 -mtune=power9 -DUSE_OPENMP -fopenmp  -o test_sbgemm compare_sgemm_sbgemm.c ../libopenblas_power9p-r0.3.24.a -lm -lpthread -lm -lpthread -L/usr/lib/gcc/ppc64le-redhat-linux/13 -L/usr/lib/gcc/ppc64le-redhat-linux/13/../../../../lib64 -L/lib/../lib64 -L/usr/lib/../lib64 -L/opt/OpenBLAS-0.3.24-MT/GCC/lib -L. -L/usr/lib/gcc/ppc64le-redhat-linux/13/../../..  -lc 
xlf -O2 -q64 -openmp -fPIC -O2 -qrecur -qnosave -mcpu=power9 -mtune=power9  -DUSE_OPENMP -fopenmp -c sblat3.f  -o sblat3.o
xlf -O2 -q64 -openmp -fPIC -O2 -qrecur -qnosave -mcpu=power9 -mtune=power9  -DUSE_OPENMP -fopenmp -c dblat3.f  -o dblat3.o
xlf -O2 -q64 -openmp -fPIC -O2 -qrecur -qnosave -mcpu=power9 -mtune=power9  -DUSE_OPENMP -fopenmp -c cblat3.f  -o cblat3.o
xlf -O2 -q64 -openmp -fPIC -O2 -qrecur -qnosave -mcpu=power9 -mtune=power9  -DUSE_OPENMP -fopenmp -c zblat3.f  -o zblat3.o
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mcpu=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mtune=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mcpu=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mtune=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mcpu=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mtune=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mcpu=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mtune=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mcpu=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mtune=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mcpu=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mtune=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mcpu=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mtune=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mcpu=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mtune=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mcpu=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mtune=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-218 (W) file compare_sgemm_sbgemm.c contains an incorrect file suffix
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mcpu=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mtune=power9 is not recognized - passed to ld
/usr/bin/ld: unrecognised emulation mode: tune=power9
Supported emulations: elf64lppc elf32lppc elf32lppclinux elf32lppcsim elf64ppc elf32ppc elf32ppclinux elf32ppcsim elf32_spu i386pep i386pe elf64bpf
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mcpu=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mtune=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mcpu=power9 is not recognized - passed to ld
/opt/ibm/xlf/16.1.1/bin/.orig/xlf: 1501-216 (W) command option -mtune=power9 is not recognized - passed to ld
make[1]: *** [Makefile:329: test_sbgemm] Error 1

[martin-frbg]

huh, why is it trying to compile a c source with the fortran compiler here ?

[martin-frbg]

found me a machine with xlf16, testing some fixes...

[martin-frbg]

there is still a problem in LAPACKE even with PR #4238 - gcc expects the symbols to have a trailing underscore, but xlf does not add one (maybe adding -qextname to FC will help)