hlrs-software-stack/spack
7
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

dbcsr: add opencl as a third backend (#21468)

Browse source
This commit is contained in:
Harmen Stoppels 2021-02-04 14:44:59 +01:00 committed by GitHub
parent 694d633a2c
commit 39a429b2a3
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 23 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

29
var/spack/repos/builtin/packages/dbcsr/package.py
View file

@ -26,6 +26,8 @@ class Dbcsr(CMakePackage, CudaPackage, ROCmPackage):
' different GPU models. Enable this when building' ' different GPU models. Enable this when building'
' with cuda_arch=35 for a K20x instead of a K40')) ' with cuda_arch=35 for a K20x instead of a K40'))
variant('opencl', default=False, description='Enable OpenCL backend')
depends_on('blas') depends_on('blas')
depends_on('lapack') depends_on('lapack')
depends_on('mpi', when='+mpi') depends_on('mpi', when='+mpi')
@ -38,6 +40,8 @@ class Dbcsr(CMakePackage, CudaPackage, ROCmPackage):
depends_on('hipblas', when='+rocm') depends_on('hipblas', when='+rocm')
depends_on('opencl', when='+opencl')
# We only support specific gpu archs for which we have parameter files # We only support specific gpu archs for which we have parameter files
# for optimal kernels. Note that we don't override the parent class arch # for optimal kernels. Note that we don't override the parent class arch
# properties, since the parent class defines constraints for different archs # properties, since the parent class defines constraints for different archs
@ -58,13 +62,17 @@ class Dbcsr(CMakePackage, CudaPackage, ROCmPackage):
if arch not in dbcsr_amdgpu_targets: if arch not in dbcsr_amdgpu_targets:
conflicts('+rocm', when='amdgpu_target={0}'.format(arch), msg=amd_msg) conflicts('+rocm', when='amdgpu_target={0}'.format(arch), msg=amd_msg)
conflicts('+cuda', when='+rocm', msg="CUDA and ROCm are mutually exclusive") accel_msg = "CUDA, ROCm and OpenCL support are mutually exlusive"
conflicts('+rocm', when='@:2.0', msg="ROCm support is available in DBCSR v2.1 and later") conflicts('+cuda', when='+rocm', msg=accel_msg)
conflicts('+cuda', when='+opencl', msg=accel_msg)
conflicts('+rocm', when='+opencl', msg=accel_msg)
# Require openmp threading for OpenBLAS by making other options conflict # Require openmp threading for OpenBLAS by making other options conflict
conflicts('^openblas threads=pthreads', when='+openmp') conflicts('^openblas threads=pthreads', when='+openmp')
conflicts('^openblas threads=none', when='+openmp') conflicts('^openblas threads=none', when='+openmp')
conflicts('smm=blas', when='+opencl')
generator = 'Ninja' generator = 'Ninja'
depends_on('ninja@1.10:', type='build') depends_on('ninja@1.10:', type='build')
@ -89,8 +97,6 @@ def cmake_args(self):
'-DLAPACK_LIBRARIES=%s' % (spec['lapack'].libs.joined(';')), '-DLAPACK_LIBRARIES=%s' % (spec['lapack'].libs.joined(';')),
'-DWITH_EXAMPLES=ON', '-DWITH_EXAMPLES=ON',
self.define_from_variant('BUILD_SHARED_LIBS', 'shared'), self.define_from_variant('BUILD_SHARED_LIBS', 'shared'),
self.define_from_variant('USE_HIP', 'rocm'),
self.define_from_variant('USE_CUDA', 'cuda'),
] ]
if self.spec.satisfies('+cuda'): if self.spec.satisfies('+cuda'):
@ -107,7 +113,10 @@ def cmake_args(self):
and self.spec.satisfies('+cuda_arch_35_k20x')): and self.spec.satisfies('+cuda_arch_35_k20x')):
gpuver = 'K20X' gpuver = 'K20X'
args += ['-DWITH_GPU=%s' % gpuver] args += [
'-DWITH_GPU=%s' % gpuver,
'-DUSE_ACCEL=cuda'
]
if self.spec.satisfies('+rocm'): if self.spec.satisfies('+rocm'):
amd_arch = self.spec.variants['amdgpu_target'].value[0] amd_arch = self.spec.variants['amdgpu_target'].value[0]
@ -116,7 +125,15 @@ def cmake_args(self):
'gfx906': 'Mi50' 'gfx906': 'Mi50'
}[amd_arch] }[amd_arch]
args += ['-DWITH_GPU={0}'.format(gpuver)] args += [
'-DWITH_GPU={0}'.format(gpuver),
'-DUSE_ACCEL=hip'
]
if self.spec.satisfies('+opencl'):
args += [
'-DUSE_ACCEL=opencl'
]
return args return args