From c42ce439e44f5a8261fdd9d5e4374a0dc26fa71e Mon Sep 17 00:00:00 2001 From: Danny Taller <66029857+dtaller@users.noreply.github.com> Date: Wed, 18 Nov 2020 11:52:21 -0800 Subject: hip support for umpire, chai, raja, camp (#19715) * create HipPackage base class and do some refactoring * comments and added conflict to raja for openmp with hip --- lib/spack/spack/build_systems/hip.py | 138 +++++++++++++++++++++++++++++++++++ lib/spack/spack/pkgkit.py | 1 + 2 files changed, 139 insertions(+) create mode 100644 lib/spack/spack/build_systems/hip.py (limited to 'lib') diff --git a/lib/spack/spack/build_systems/hip.py b/lib/spack/spack/build_systems/hip.py new file mode 100644 index 0000000000..da44f1428d --- /dev/null +++ b/lib/spack/spack/build_systems/hip.py @@ -0,0 +1,138 @@ +# Copyright 2013-2020 Lawrence Livermore National Security, LLC and other +# Spack Project Developers. See the top-level COPYRIGHT file for details. +# +# SPDX-License-Identifier: (Apache-2.0 OR MIT) + +# Troubleshooting advice for +hip builds: +# +# 1. When building with clang, go your compilers.yaml, +# add an entry for the amd version of clang, as below. +# This will ensure that your entire package is compiled/linked +# with the same compiler version. If you use a different version of +# clang which is linked against a different version of the gcc library, +# you will get errors along the lines of: +# undefined reference to +# `std::__throw_out_of_range_fmt(char const*, ...)@@GLIBCXX_3.4.20' +# which is indicative of a mismatch in standard library versions. +# +# in compilers.yaml +# - compiler: +# spec: clang@amd +# paths: +# cc: /opt/rocm/llvm/bin/clang +# cxx: /opt/rocm/llvm/bin/clang++ +# f77: +# fc: +# flags: {} +# operating_system: rhel7 +# target: x86_64 +# modules: [] +# environment: {} +# extra_rpaths: [] +# +# +# 2. hip and its dependencies are currently NOT picked up by spack +# automatically, and should therefore be added to packages.yaml by hand: +# +# in packages.yaml: +# hip: +# externals: +# - spec: hip@3.8.20371-d1886b0b +# prefix: /opt/rocm/hip +# extra_attributes: +# compilers: +# c: /opt/rocm/llvm/bin/clang++ +# c++: /opt/rocm/llvm/bin/clang++ +# hip: /opt/rocm/hip/bin/hipcc +# buildable: false +# hsa-rocr-dev: +# externals: +# - spec: hsa-rocr-dev +# prefix: /opt/rocm +# extra_attributes: +# compilers: +# c: /opt/rocm/llvm/bin/clang++ +# cxx: /opt/rocm/llvm/bin/clang++ +# buildable: false +# llvm-amdgpu: +# externals: +# - spec: llvm-amdgpu +# prefix: /opt/rocm/llvm +# extra_attributes: +# compilers: +# c: /opt/rocm/llvm/bin/clang++ +# cxx: /opt/rocm/llvm/bin/clang++ +# buildable: false +# +# 3. In part 2, DO NOT list the path to hsa as /opt/rocm/hsa ! You want spack +# to find hsa in /opt/rocm/include/hsa/hsa.h . The directory of +# /opt/rocm/hsa also has an hsa.h file, but it won't be found because spack +# does not like its directory structure. +# + +from spack.package import PackageBase +from spack.directives import depends_on, variant, conflicts + + +class HipPackage(PackageBase): + """Auxiliary class which contains HIP variant, dependencies and conflicts + and is meant to unify and facilitate its usage. Closely mimics CudaPackage. + + Maintainers: dtaller + """ + + # https://llvm.org/docs/AMDGPUUsage.html + # Possible architectures + amdgpu_targets = ( + 'gfx701', 'gfx801', 'gfx802', 'gfx803', + 'gfx900', 'gfx906', 'gfx908', 'gfx1010', + 'gfx1011', 'gfx1012', 'none' + ) + + variant('hip', default=False, description='Enable HIP support') + + # possible amd gpu targets for hip builds + variant('amdgpu_target', default='none', values=amdgpu_targets) + + depends_on('llvm-amdgpu', when='+hip') + depends_on('hsa-rocr-dev', when='+hip') + depends_on('hip', when='+hip') + + # need amd gpu type for hip builds + conflicts('amdgpu_target=none', when='+hip') + + # Make sure non-'none' amdgpu_targets cannot be used without +hip + for value in amdgpu_targets[:-1]: + conflicts('~hip', when='amdgpu_target=' + value) + + # https://github.com/ROCm-Developer-Tools/HIP/blob/master/bin/hipcc + # It seems that hip-clang does not (yet?) accept this flag, in which case + # we will still need to set the HCC_AMDGPU_TARGET environment flag in the + # hip package file. But I will leave this here for future development. + @staticmethod + def hip_flags(amdgpu_target): + return '--amdgpu-target={0}'.format(amdgpu_target) + + # https://llvm.org/docs/AMDGPUUsage.html + # Possible architectures (not including 'none' option) + @staticmethod + def amd_gputargets_list(): + return ( + 'gfx701', 'gfx801', 'gfx802', 'gfx803', + 'gfx900', 'gfx906', 'gfx908', 'gfx1010', + 'gfx1011', 'gfx1012' + ) + + # HIP version vs Architecture + + # TODO: add a bunch of lines like: + # depends_on('hip@:6.0', when='amdgpu_target=gfx701') + # to indicate minimum version for each architecture. + + # Compiler conflicts + + # TODO: add conflicts statements along the lines of + # arch_platform = ' target=x86_64: platform=linux' + # conflicts('%gcc@5:', when='+cuda ^cuda@:7.5' + arch_platform) + # conflicts('platform=darwin', when='+cuda ^cuda@11.0.2:') + # for hip-related limitations. diff --git a/lib/spack/spack/pkgkit.py b/lib/spack/spack/pkgkit.py index 4f25d41dfb..da519f2e16 100644 --- a/lib/spack/spack/pkgkit.py +++ b/lib/spack/spack/pkgkit.py @@ -20,6 +20,7 @@ from spack.build_systems.aspell_dict import AspellDictPackage from spack.build_systems.autotools import AutotoolsPackage from spack.build_systems.cmake import CMakePackage from spack.build_systems.cuda import CudaPackage +from spack.build_systems.hip import HipPackage from spack.build_systems.qmake import QMakePackage from spack.build_systems.maven import MavenPackage from spack.build_systems.scons import SConsPackage -- cgit v1.2.3-70-g09d2