diehlpk / octotiger-spack

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Octotiger-Spack

This repository contains the Spack package for the astrophysics code Octo-Tiger. Additionally, the repository also includes new/modified/updated spack packages of software Octo-Tiger depends on. The repo is meant to replace the old collection of Octo-Tiger buildscripts as these become increasingly unwieldy and uncomfortable to use.

Package features:

  • Supports builds all CPU/GPU variants of Octotiger (Kokkos, CUDA, HIP, SYCL and the older Vc kernels) and all available SIMD backends
  • Supports distributed builds (with the HPX networking backends tcp, mpi or lci)
  • Supports older versions of Octo-Tiger as well (0.8.0 / Benchmark Version and 0.9.0)

List of added/modified packages:

  • Added Octo-Tiger package
  • Added CPPuddle package
  • Modified HPX package by adding a SYCL variant (depends on dpcpp; tested on NVIDIA/AMD GPUs)
  • Modified HPX package by adding a LCI parcelport variant to have an additional networking backend available
  • Modified HPX-Kokkos package by adding SYCL variant, adding variants for the future types and exposing the hpx-kokkos tests to spack
  • Modified Kokkos package by adding an use_unsupported_sycl_arch variant (which allows using the SYCL execution space on non-Intel GPU - which in turn allowed me to test octotiger with SYCL on my available hardware). Also includes a patch for the kokkos CMakeLists which allows using this on AMDGPUs (not just NVIDIA)
  • Added some patches for kokkos/kokkos-nvcc wrapper for NIXOs usage
  • Modified dpcpp package by using the package.py from the this spack PR that's updating (not merged as of yet) and by attempting to fix the library/linking issue in dependent packages

Note: Additions and changes might eventually be submitted to the main spack repository! For now, this repo allows me to rapidly experiment with different builds on different machines until I get to the point where all Octo-Tiger features and relevant versions are working on the intended target platforms.

Repo installation:

# spack install
git clone --depth=100 --branch=releases/v0.20 https://github.com/spack/spack.git /path/to/spack
cd /path/to/spack
. share/spack/setup-env.sh
# spack repo install
git clone https://github.com/G-071/octotiger-spack.git /path/to/octotiger-spack
spack repo add /path/to/octotiger-spack
# use system packages
spack compiler find # search currently loaded compilers
spack external find cuda # replace cuda by desired packages or leave blank
# Check package availability and its variants:
spack info octotiger

Usage examples

  • Basic Octo-Tiger installation:
# Basic octotiger installation
spack install --fresh --test=root octotiger~cuda~kokkos
# octotiger install with specific compiler/cuda versions and run octotiger tests
spack install --fresh --test=root octotiger+cuda+kokkos%gcc@11^cuda@11.8.89^cmake@3.26.4^kokkos@3.7.00 cuda_arch=75
# Use one of the installs
spack load octotiger~cuda~kokkos
octotiger --help
  • Octo-Tiger CUDA/Kokkos dev build (RTX 2060):
module load cuda/11.8
spack external find cuda # finds cuda@11.8.89 in this example...
# Get fresh src dir
git clone https://github.com/STEllAR-GROUP/octotiger
cd octotiger
# Setup development shell with all dependencies (drops user into the build process right after cmake)
spack dev-build --fresh --drop-in bash --until cmake --test=root octotiger+cuda+kokkos cuda_arch=75 @master%gcc@11^cuda@11.8.89
cd spack_build_id #exact dir name ist printed by the last command
# Use with usual edit-make-test cycle after editing the src directory...
make -j16
ctest --output-on-failure
./octotiger --help
  • Octo-Tiger HIP/Kokkos dev build (MI100):
module load rocm/5.4.6
spack compiler find # should find rocmcc@5.4.6
spack external find hip llvm-amdgpu hsa-rocr-dev # other erquired 5.4.6 packges from the rocm module
git clone https://github.com/STEllAR-GROUP/octotiger
cd octotiger
# long, manual spec for dev build:
spack dev-build --drop-in bash --until cmake --test=root octotiger+rocm+kokkos amdgpu_target=gfx908@master%rocmcc@5.4.6 ^asio@1.16.0^hpx max_cpu_count=128 amdgpu_target=gfx908 ^hip@5.4.6 ^llvm-amdgpu@5.4.6^kokkos amdg
pu_target=gfx908 ^hpx-kokkos amdgpu_target=gfx908
# new alternative (octotiger package now internally propgates the amdgpu_target to hpx, kokkos and hpx-kokkos, enabling this shorter version):
spack dev-build --drop-in bash --until cmake --test=root octotiger+rocm+kokkos amdgpu_target=gfx908@master%rocmcc@5.4.6 ^asio@1.16.0^hpx max_cpu_count=128 ^hip@5.4.6 ^llvm-amdgpu@5.4.6
cd spack_build_id #exact dir name ist printed by the last command
# Use with usual edit-make-test cycle after editing the src directory...
make -j32
ctest --output-on-failure
  • Octo-Tiger SYCL/Kokkos dev build (V100):
module load cuda/12
git clone https://github.com/STEllAR-GROUP/octotiger
cd octotiger
spack dev-build --fresh --drop-in bash --until cmake --test=root octotiger@master -cuda -rocm +sycl %gcc@10 ^kokkos use_unsupported_sycl_arch=70 ^hpx sycl_target_arch=70 ^cppuddle
cd spack_build_id #exact dir name ist printed by the last command
# Needed without the libs fix in the dpcpp package
#export LD_LIBRARY_PATH="$(spack find -l --paths dpcpp | tail -n1 | awk '{ print $3 }')/lib":${LD_LIBRARY_PATH}
# Use with usual edit-make-test cycle after editing the src directory...
make -j32
ctest --output-on-failure
# 0.9.0 is the first version with Kokkos, a GPU hydro solver and ROCm support. 
# Based on HPX 1.6.0 - 1.7.1
spack install octotiger@0.9.0 +cuda +kokkos cuda_arch=70 %clang@12
spack load octotiger@0.9.0 +cuda +kokkos cuda_arch=70 %clang@12


# 0.8.0 (or benchmark version) is the legacy version with basic GPU support in the gravity solver but CPU-only hydro kernels.
# Based on HPX 1.4.1 
spack install octotiger@0.8.0 +cuda cuda_arch=70 %gcc@11
spack load octotiger@0.8.0 +cuda cuda_arch=70 %gcc@11

Notes for usage on nixos / with nix:

For now, use this spack fork when using octotiger-spack on nixos. It contains multiple small workarounds and a nix-shell (based on the suggestion in this spack PR).

To use it, go to the spack root directory, create the new tmp dir spack_tmp (avoids issues to to small tmp dirs on some of my machines) and load the shell via nix-shell share/spack/shell.nix. Afterwards, use spack as usual! I tested it with the Kokkos/CUDA usage example posted above and it seems to work well.

About


Languages

Language:Python 100.0%