spack/lib/spack/spack/build_systems/cuda.py
Adam J. Stewart 6153eea2c2
CudaPackage: add support for Tesla K80 and older CUDA (#15031)
* CudaPackage: add support for Tesla K80 and older CUDA

* Flake8 fixes

* Fix cuda_arch when no arch is set

* Fine-tune cuda_arch=37,50 supported CUDA versions

* CUDA 6.5+ supports SM_37

* Add @svenevs as a maintainer
2020-02-27 12:16:20 -06:00

164 lines
7.6 KiB
Python

# 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)
from spack.package import PackageBase
from spack.directives import depends_on, variant, conflicts
import spack.variant
class CudaPackage(PackageBase):
"""Auxiliary class which contains CUDA variant, dependencies and conflicts
and is meant to unify and facilitate its usage.
"""
maintainers = ['ax3l', 'svenevs']
# https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#gpu-feature-list
# https://developer.nvidia.com/cuda-gpus
# https://en.wikipedia.org/wiki/CUDA#GPUs_supported
cuda_arch_values = [
'10', '11', '12', '13',
'20', '21',
'30', '32', '35', '37',
'50', '52', '53',
'60', '61', '62',
'70', '72', '75',
]
# FIXME: keep cuda and cuda_arch separate to make usage easier until
# Spack has depends_on(cuda, when='cuda_arch!=None') or alike
variant('cuda', default=False,
description='Build with CUDA')
variant('cuda_arch',
description='CUDA architecture',
values=spack.variant.any_combination_of(*cuda_arch_values))
# https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#nvcc-examples
# https://llvm.org/docs/CompileCudaWithLLVM.html#compiling-cuda-code
@staticmethod
def cuda_flags(arch_list):
return [('--generate-code arch=compute_{0},code=sm_{0} '
'--generate-code arch=compute_{0},code=compute_{0}').format(s)
for s in arch_list]
depends_on('cuda', when='+cuda')
# CUDA version vs Architecture
# https://en.wikipedia.org/wiki/CUDA#GPUs_supported
depends_on('cuda@:6.0', when='cuda_arch=10')
depends_on('cuda@:6.5', when='cuda_arch=11')
depends_on('cuda@2.1:6.5', when='cuda_arch=12')
depends_on('cuda@2.1:6.5', when='cuda_arch=13')
depends_on('cuda@3.0:8.0', when='cuda_arch=20')
depends_on('cuda@3.2:8.0', when='cuda_arch=21')
depends_on('cuda@5.0:10.2', when='cuda_arch=30')
depends_on('cuda@5.0:10.2', when='cuda_arch=32')
depends_on('cuda@5.0:10.2', when='cuda_arch=35')
depends_on('cuda@6.5:10.2', when='cuda_arch=37')
depends_on('cuda@6.0:', when='cuda_arch=50')
depends_on('cuda@6.5:', when='cuda_arch=52')
depends_on('cuda@6.5:', when='cuda_arch=53')
depends_on('cuda@8.0:', when='cuda_arch=60')
depends_on('cuda@8.0:', when='cuda_arch=61')
depends_on('cuda@8.0:', when='cuda_arch=62')
depends_on('cuda@9.0:', when='cuda_arch=70')
depends_on('cuda@9.0:', when='cuda_arch=72')
depends_on('cuda@10.0:', when='cuda_arch=75')
# There are at least three cases to be aware of for compiler conflicts
# 1. Linux x86_64
# 2. Linux ppc64le
# 3. Mac OS X
# CUDA-compiler conflicts are version-to-version specific and are
# difficult to express with the current Spack conflict syntax
# Linux x86_64 compiler conflicts from here:
# https://gist.github.com/ax3l/9489132
arch_platform = ' target=x86_64: platform=linux'
conflicts('%gcc@5:', when='+cuda ^cuda@:7.5' + arch_platform)
conflicts('%gcc@6:', when='+cuda ^cuda@:8' + arch_platform)
conflicts('%gcc@7:', when='+cuda ^cuda@:9.1' + arch_platform)
conflicts('%gcc@8:', when='+cuda ^cuda@:10.0.130' + arch_platform)
conflicts('%gcc@9:', when='+cuda ^cuda@:10.2.89' + arch_platform)
conflicts('%pgi@:14.8', when='+cuda ^cuda@:7.0.27' + arch_platform)
conflicts('%pgi@:15.3,15.5:', when='+cuda ^cuda@7.5' + arch_platform)
conflicts('%pgi@:16.2,16.0:16.3', when='+cuda ^cuda@8' + arch_platform)
conflicts('%pgi@:15,18:', when='+cuda ^cuda@9.0:9.1' + arch_platform)
conflicts('%pgi@:16', when='+cuda ^cuda@9.2.88:10' + arch_platform)
conflicts('%pgi@:17', when='+cuda ^cuda@10.2.89' + arch_platform)
conflicts('%clang@:3.4', when='+cuda ^cuda@:7.5' + arch_platform)
conflicts('%clang@:3.7,4:',
when='+cuda ^cuda@8.0:9.0' + arch_platform)
conflicts('%clang@:3.7,4.1:',
when='+cuda ^cuda@9.1' + arch_platform)
conflicts('%clang@:3.7,5.1:', when='+cuda ^cuda@9.2' + arch_platform)
conflicts('%clang@:3.7,6.1:', when='+cuda ^cuda@10.0.130' + arch_platform)
conflicts('%clang@:3.7,7.1:', when='+cuda ^cuda@10.1.105' + arch_platform)
conflicts('%clang@:3.7,8.1:',
when='+cuda ^cuda@10.1.105:10.1.243' + arch_platform)
conflicts('%clang@:3.2,9.0:', when='+cuda ^cuda@10.2.89' + arch_platform)
# x86_64 vs. ppc64le differ according to NVidia docs
# Linux ppc64le compiler conflicts from Table from the docs below:
# https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html
# https://docs.nvidia.com/cuda/archive/9.2/cuda-installation-guide-linux/index.html
# https://docs.nvidia.com/cuda/archive/9.1/cuda-installation-guide-linux/index.html
# https://docs.nvidia.com/cuda/archive/9.0/cuda-installation-guide-linux/index.html
# https://docs.nvidia.com/cuda/archive/8.0/cuda-installation-guide-linux/index.html
arch_platform = ' target=ppc64le: platform=linux'
# information prior to CUDA 9 difficult to find
conflicts('%gcc@6:', when='+cuda ^cuda@:9' + arch_platform)
conflicts('%gcc@8:', when='+cuda ^cuda@:10.0.130' + arch_platform)
conflicts('%gcc@9:', when='+cuda ^cuda@:10.1.243' + arch_platform)
conflicts('%pgi', when='+cuda ^cuda@:8' + arch_platform)
conflicts('%pgi@:16', when='+cuda ^cuda@:9.1.185' + arch_platform)
conflicts('%pgi@:17', when='+cuda ^cuda@:10' + arch_platform)
conflicts('%clang@4:', when='+cuda ^cuda@:9.0.176' + arch_platform)
conflicts('%clang@5:', when='+cuda ^cuda@:9.1' + arch_platform)
conflicts('%clang@6:', when='+cuda ^cuda@:9.2' + arch_platform)
conflicts('%clang@7:', when='+cuda ^cuda@10.0.130' + arch_platform)
conflicts('%clang@7.1:', when='+cuda ^cuda@:10.1.105' + arch_platform)
conflicts('%clang@8.1:', when='+cuda ^cuda@:10.2.89' + arch_platform)
# Intel is mostly relevant for x86_64 Linux, even though it also
# exists for Mac OS X. No information prior to CUDA 3.2 or Intel 11.1
conflicts('%intel@:11.0', when='+cuda ^cuda@:3.1')
conflicts('%intel@:12.0', when='+cuda ^cuda@5.5:')
conflicts('%intel@:13.0', when='+cuda ^cuda@6.0:')
conflicts('%intel@:13.2', when='+cuda ^cuda@6.5:')
conflicts('%intel@:14.9', when='+cuda ^cuda@7:')
# Intel 15.x is compatible with CUDA 7 thru current CUDA
conflicts('%intel@16.0:', when='+cuda ^cuda@:8.0.43')
conflicts('%intel@17.0:', when='+cuda ^cuda@:8.0.60')
conflicts('%intel@18.0:', when='+cuda ^cuda@:9.9')
conflicts('%intel@19.0:', when='+cuda ^cuda@:10.2.89')
# XL is mostly relevant for ppc64le Linux
conflicts('%xl@:12,14:', when='+cuda ^cuda@:9.1')
conflicts('%xl@:12,14:15,17:', when='+cuda ^cuda@9.2')
conflicts('%xl@17:', when='+cuda ^cuda@10.0.130:10.2.89')
# Mac OS X
# platform = ' platform=darwin'
# Apple XCode clang vs. LLVM clang are difficult to specify
# with spack syntax. Xcode clang name is `clang@x.y.z-apple`
# which precludes ranges being specified. We have proposed
# rename XCode clang to `clang@apple-x.y.z` or even
# `clang-apple@x.y.z as a possible fix.
# Compiler conflicts will be eventual taken from here:
# https://docs.nvidia.com/cuda/cuda-installation-guide-mac-os-x/index.html#abstract
conflicts('platform=darwin', when='+cuda ^cuda@11.0:')
# Make sure cuda_arch can not be used without +cuda
for value in cuda_arch_values:
conflicts('~cuda', when='cuda_arch=' + value)