Compiliation error with NVHPC toolkit 22.1

[RonRahaman]

Hi all,

I'm encountering some compilation errors when building UCX 1.12.1 with our site's installation of NVHPC Toolkit 22.1 on RHEL 7.9. I'm using NVHPC for both CUDA and MPI. My config.log and our compilation error are below. Thanks!

config.log.txt

  CC       ucx_perftest-perftest.o
"/storage/home/hcodaman1/rrahaman6/manual-src/ucx-1.12.1/contrib/../src/tools/perf/perftest.c", line 640: error #1144: arithmetic on pointer to void or function type
          MPI_Recv(buffer + offset, max - offset, MPI_BYTE, src, MPI_ANY_TAG,
                          ^

"/storage/home/hcodaman1/rrahaman6/manual-src/ucx-1.12.1/contrib/../src/tools/perf/perftest.c", line 790: error #144: a value of type "void *" cannot be used to initialize an entity of type "ucx_perf_rte_exchange_vec_func_t"
          .exchange_vec  = (void*)ucs_empty_function,
                           ^

2 errors detected in the compilation of "/storage/home/hcodaman1/rrahaman6/manual-src/ucx-1.12.1/contrib/../src/tools/perf/perftest.c".
make[3]: *** [ucx_perftest-perftest.o] Error 2

---

Ron Rahaman
Research Scientist II
Partnership for an Advanced Computing Environment (PACE)
Georgia Institute of Technology

[RonRahaman]

In the NVHPC build, nvc is used for both CC and MPICC (the mpicc command uses nvc plus the MPI header/library flags). I'm wondering if nvc is pretty strict about pointer arithmetic and casting.

I can build by combining our toolchains for GCC 8.3 + OpenMPI 4.1.2 + CUDA 11.2 just fine.

[RonRahaman]

I was able to get this working with "--diag_suppress=1144 --diag_suppress=144 --diag_suppress=513" in CFLAGS