zhangyiss/spack
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

cp2k: fine graining control of the GPU modules (#48925)

Browse Source 
Co-authored-by: Mathieu Taillefumier <mathieu.taillefumier@free.fr>
Co-authored-by: Mikael Simberg <mikael.simberg@iki.fi>
Co-authored-by: Rocco Meli <r.meli@bluemail.ch>
This commit is contained in:
Taillefumier Mathieu 2025-02-19 03:29:31 -05:00 committed by GitHub
parent 2427b9649d
commit 49bf5a349e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 35 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

40
var/spack/repos/builtin/packages/cp2k/package.py
View File

@ -119,7 +119,23 @@ class Cp2k(MakefilePackage, CMakePackage, CudaPackage, ROCmPackage):
variant("dftd4", when="@2024.2:", default=False, description="Enable DFT-D4 support")
variant("mpi_f08", default=False, description="Use MPI F08 module")
variant("smeagol", default=False, description="Enable libsmeagol support", when="@2025.2:")
variant(
"pw_gpu", default=True, description="Enable FFT calculations on GPU", when="@2025.2: +cuda"
)
variant("grid_gpu", default=True, description="Enable grid GPU backend", when="@2025.2:")
variant("dbm_gpu", default=True, description="Enable DBM GPU backend", when="@2025.2:")
variant(
"pw_gpu",
default=False,
description="Enable FFT calculations on GPU",
when="@2025.2: +rocm",
)
variant(
"hip_backend_cuda",
default=False,
description="Enable HIP backend on Nvidia GPU",
when="@2025.2: +cuda",
)
variant(
"enable_regtests",
default=False,
@ -148,6 +164,12 @@ class Cp2k(MakefilePackage, CMakePackage, CudaPackage, ROCmPackage):
description="Use CUBLAS for general matrix operations in DBCSR",
)
with when("+hip_backend_cuda"):
depends_on("hipcc")
depends_on("hip+cuda")
depends_on("hipfft+cuda")
depends_on("hipblas+cuda")
HFX_LMAX_RANGE = range(4, 8)
variant(
@ -983,10 +1005,15 @@ def cmake_args(self):
raise InstallError("CP2K supports only one cuda_arch at a time.")
else:
gpu_ver = GPU_MAP[spec.variants["cuda_arch"].value[0]]
args += [
self.define("CP2K_USE_ACCEL", "CUDA"),
self.define("CP2K_WITH_GPU", gpu_ver),
]
if spec.satisfies("+hip_backend_cuda"):
args += [
self.define("CP2K_USE_ACCEL", "HIP"),
self.define("CMAKE_HIP_PLATFORM", "nvidia"),
]
else:
args += [self.define("CP2K_USE_ACCEL", "CUDA")]
args += [self.define("CP2K_WITH_GPU", gpu_ver)]
if spec.satisfies("+rocm"):
if len(spec.variants["amdgpu_target"].value) > 1:
@ -1018,6 +1045,9 @@ def cmake_args(self):
self.define_from_variant("CP2K_USE_DFTD4", "dftd4"),
self.define_from_variant("CP2K_USE_MPI_F08", "mpi_f08"),
self.define_from_variant("CP2K_USE_LIBSMEAGOL", "smeagol"),
self.define_from_variant("CP2K_ENABLE_GRID_GPU", "grid_gpu"),
self.define_from_variant("CP2K_ENABLE_DBM_GPU", "dbm_gpu"),
self.define_from_variant("CP2K_ENABLE_PW_GPU", "pw_gpu"),
]
# we force the use elpa openmp threading support. might need to be revisited though