1#!/usr/bin/python3 2 3# Use GNU compilers: 4# 5# Note cray-libsci provides BLAS etc. In summary, we have 6# module use /soft/modulefiles 7# module unload darshan 8# module load PrgEnv-gnu cray-libsci nvhpc-mixed craype-accel-nvidia80 cudatoolkit-standalone/12.4.1 9# export MPICH_GPU_SUPPORT_ENABLED=1 10# export MPICH_GPU_IPC_ENABLED=0 11# 12# $ module list 13# Currently Loaded Modules: 14# 1) libfabric/1.15.2.0 7) nghttp2/1.57.0-ciat5hu 13) cray-mpich/8.1.28 19) cray-libsci/23.12.5 15# 2) craype-network-ofi 8) curl/8.4.0-2ztev25 14) cray-pmi/6.1.13 20) nvhpc-mixed/23.9 16# 3) perftools-base/23.12.0 9) cmake/3.27.7 15) cray-pals/1.3.4 21) craype-accel-nvidia80 17# 4) darshan/3.4.4 10) cudatoolkit-standalone/12.4.1 16) cray-libpals/1.3.4 18# 5) gcc-native/12.3 11) craype/2.7.30 17) craype-x86-milan 19# 6) spack-pe-base/0.6.1 12) cray-dsmml/0.2.2 18) PrgEnv-gnu/8.5.0 20# 21# With above, Cray-MPICH GPU-aware MPI works on a node but still fail with multiple nodes. In the latter case, you can 22# add the PETSc runtime option -use_gpu_aware_mpi 0 as a workaround. 23 24if __name__ == '__main__': 25 import sys 26 import os 27 sys.path.insert(0, os.path.abspath('config')) 28 import configure 29 configure_options = [ 30 '--with-cc=cc', 31 '--with-cxx=CC', 32 '--with-fc=ftn', 33 '--with-debugging=0', 34 '--with-cuda', 35 '--with-cudac=nvcc', 36 '--with-cuda-arch=80', # Since there is no easy way to auto-detect the cuda arch on the gpu-less Polaris login nodes, we explicitly set it. 37 '--download-kokkos', 38 '--download-kokkos-kernels', 39 '--download-umpire', 40 '--download-hypre', 41 ] 42 configure.petsc_configure(configure_options) 43 44# Use NVHPC compilers 45# 46# Unset so that cray won't add -gpu to nvc even when craype-accel-nvidia80 is loaded 47# unset CRAY_ACCEL_TARGET 48# module load nvhpc/22.11 PrgEnv-nvhpc 49# 50# I met two problems with nvhpc and Kokkos (and Kokkos-Kernels) 4.2.0. 51# 1) Kokkos-Kernles failed at configuration to find TPL cublas and cusparse from NVHPC. 52# As a workaround, I just load cudatoolkit-standalone/11.8.0 to let KK use cublas and cusparse from cudatoolkit-standalone. 53# 2) KK failed at compilation 54# "/home/jczhang/petsc/arch-kokkos-dbg/externalpackages/git.kokkos-kernels/batched/dense/impl/KokkosBatched_Gemm_Serial_Internal.hpp", line 94: error: expression must have a constant value 55# constexpr int nbAlgo = Algo::Gemm::Blocked::mb(); 56# ^ 57# "/home/jczhang/petsc/arch-kokkos-dbg/externalpackages/git.kokkos-kernels/blas/impl/KokkosBlas_util.hpp", line 58: note: cannot call non-constexpr function "__builtin_is_device_code" (declared implicitly) 58# KOKKOS_IF_ON_HOST((return 4;)) 59# ^ 60# detected during: 61# 62# It is a KK problem and I have to wait for their fix. 63