zhangyiss/spack
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
5fa8890bd3
spack/var/spack/repos/builtin/packages/babelstream/package.py
Adam J. Stewart 5fa8890bd3
CUDA: support Grace Hopper 9.0a compute capability (#45540) 
* CUDA: support Grace Hopper 9.0a compute capability

* Fix other packages

* Add type annotations

* Support ancient Python versions

* isort

* spec -> self.spec

Co-authored-by: Andrew W Elble <aweits@rit.edu>

* [@spackbot] updating style on behalf of adamjstewart

---------

Co-authored-by: Andrew W Elble <aweits@rit.edu>
Co-authored-by: adamjstewart <adamjstewart@users.noreply.github.com>
2024-09-11 17:43:20 +02:00

418 lines
19 KiB
Python
Raw Blame History

# Copyright 2013-2024 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)
import re # To get the variant name after (+)
from spack.package import *
def find_model_flag(str):
res = re.findall(r"\+(\w+)", str)
if not res:
return ""
return res
class Babelstream(CMakePackage, CudaPackage, ROCmPackage):
"""Measure memory transfer rates to/from global device memory on GPUs.
This benchmark is similar in spirit, and based on, the STREAM benchmark for CPUs."""
homepage = "https://github.com/UoB-HPC/BabelStream"
url = "https://github.com/UoB-HPC/BabelStream/archive/refs/tags/v4.0.tar.gz"
git = "https://github.com/UoB-HPC/BabelStream.git"
version("4.0", sha256="a9cd39277fb15d977d468435eb9b894f79f468233f0131509aa540ffda4f5953")
version("main", branch="main")
version("develop", branch="develop")
depends_on("cxx", type="build") # generated
maintainers("tomdeakin", "kaanolgu", "tom91136", "robj0nes")
# Languages
# Also supported variants are cuda and rocm (for HIP)
variant("sycl", default=False, description="Enable SYCL support")
variant("sycl2020", default=False, description="Enable SYCL support")
variant("omp", default=False, description="Enable OpenMP support")
variant("ocl", default=False, description="Enable OpenCL support")
variant("tbb", default=False, description="Enable TBB support")
variant("acc", default=False, description="Enable OpenACC support")
variant("thrust", default=False, description="Enable THRUST support")
variant("raja", default=False, description="Enable RAJA support")
variant("stddata", default=False, description="Enable STD-data support")
variant("stdindices", default=False, description="Enable STD-indices support")
variant("stdranges", default=False, description="Enable STD-ranges support")
# Some models need to have the programming model abstraction downloaded -
# this variant enables a path to be provided.
variant("dir", values=str, default="none", description="Enable Directory support")
# Kokkos conflict and variant
conflicts(
"dir=none", when="+kokkos", msg="KOKKKOS requires architecture to be specfied by dir="
)
variant("kokkos", default=False, description="Enable KOKKOS support")
# ACC conflict
variant("cpu_arch", values=str, default="none", description="Enable CPU Target for ACC")
variant("acc_target", values=str, default="none", description="Enable CPU Target for ACC")
# STD conflicts
conflicts("+stddata", when="%gcc@:10.1.0", msg="STD-data requires newer version of GCC")
conflicts("+stdindices", when="%gcc@:10.1.0", msg="STD-indices requires newer version of GCC")
conflicts("+stdranges", when="%gcc@:10.1.0", msg="STD-ranges requires newer version of GCC")
# CUDA conflict
conflicts(
"cuda_arch=none",
when="+cuda",
msg="CUDA requires architecture to be specfied by cuda_arch=",
)
variant("mem", values=str, default="DEFAULT", description="Enable MEM Target for CUDA")
# Raja Conflict
variant(
"offload", values=str, default="none", description="Enable RAJA Target [CPU or NVIDIA]"
)
conflicts(
"offload=none",
when="+raja",
msg="RAJA requires architecture to be specfied by acc_target=[CPU,NVIDIA]",
)
# download raja from https://github.com/LLNL/RAJA
conflicts(
"dir=none",
when="+raja",
msg="RAJA implementation requires architecture to be specfied by dir=",
)
# Thrust Conflict
# conflicts("~cuda", when="+thrust", msg="Thrust requires +cuda variant")
depends_on("thrust", when="+thrust")
depends_on("rocthrust", when="+thrust implementation=rocm")
# TBB Dependency
depends_on("intel-oneapi-tbb", when="+tbb")
partitioner_vals = ["auto", "affinity", "static", "simple"]
variant(
"partitioner",
values=partitioner_vals,
default="auto",
description="Partitioner specifies how a loop template should partition its work among threads.\
Possible values are:\
AUTO - Optimize range subdivision based on work-stealing events.\
AFFINITY - Proportional splitting that optimizes for cache affinity.\
STATIC - Distribute work uniformly with no additional load balancing.\
SIMPLE - Recursively split its range until it cannot be further subdivided.\
See https://spec.oneapi.com/versions/latest/elements/oneTBB/source/algorithms.html#partitioners for more details.",
)
# Kokkos Dependency
depends_on("kokkos@3.7.1", when="+kokkos")
# OpenCL Dependency
backends = {
"ocl": [
("amd", "rocm-opencl", "enable ROCM backend"),
("cuda", "cuda", "enable Cuda backend"),
("intel", "intel-oneapi-compilers", "enable Intel backend"),
("pocl", "pocl@1.5", "enable POCL backend"),
],
"kokkos": [
("cuda", "cuda", "enable Cuda backend"),
("omp", "none", "enable Cuda backend"),
],
}
backend_vals = ["none"]
for lang in backends:
for item in backends[lang]:
backend, dpdncy, descr = item
backend_vals.append(backend.lower())
variant("backend", values=backend_vals, default="none", description="Enable backend support")
for lang in backends:
for item in backends[lang]:
backend, dpdncy, descr = item
if dpdncy.lower() != "none":
depends_on("%s" % dpdncy.lower(), when="backend=%s" % backend.lower())
# this flag could be used in all required languages
variant("flags", values=str, default="none", description="Additional CXX flags to be provided")
# comp_impl_vals=["ONEAPI-DPCPP","DPCPP","HIPSYCL","COMPUTECPP"]
variant(
"implementation",
values=str,
default="none",
description="Compile using the specified SYCL compiler option",
)
conflicts(
"implementation=none",
when="+sycl",
msg="SYCL requires compiler implementation to be specified by option=",
)
conflicts(
"implementation=none",
when="+thrust",
msg="Which Thrust implementation to use, supported options include:\
- CUDA (via https://github.com/NVIDIA/thrust)\
- ROCM (via https://github.com/ROCm/rocThrust)",
)
# This applies to all
depends_on("cmake@3.14.0:", type="build")
depends_on("opencl-c-headers", when="+ocl")
def cmake_args(self):
# convert spec to string to work on it
spec_string = str(self.spec)
# take only the first portion of the spec until space
spec_string_truncate = spec_string.split(" ", 1)[0]
model_list = find_model_flag(spec_string_truncate) # Prints out ['cuda', 'thrust']
if len(model_list) > 1:
ignore_list = ["cuda"] # if +acc is provided ignore the cuda model
model = list(set(model_list) - set(ignore_list))
# We choose 'thrust' from the list of ['cuda', 'thrust']
args = ["-DMODEL=" + model[0]]
else:
# if it is +stddata,indices etc. we need to pass it
# as std-data to the CMake compiler
# do some alterations here
if "std" in model_list[0]:
args = ["-DMODEL=" + "std-" + model_list[0].split("d", 1)[1]]
else:
args = ["-DMODEL=" + model_list[0]]
# ===================================
# ACC
# ===================================
if ("+acc" in self.spec) and ("~cuda" in self.spec):
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
if "cuda_arch" in self.spec.variants:
cuda_arch_list = self.spec.variants["cuda_arch"].value
# the architecture value is only number so append sm_ to the name
cuda_arch = "cc" + cuda_arch_list[0]
args.append("-DTARGET_DEVICE=gpu")
args.append("-DCUDA_ARCH=" + cuda_arch)
elif "cpu_arch" in self.spec.variants:
cpu_arch_list = self.spec.variants["cpu_arch"].value
# the architecture value is only number so append sm_ to the name
cpu_arch = cpu_arch_list[0]
args.append("-DTARGET_DEVICE=multicore")
args.append("-DTARGET_PROCESSOR=" + cpu_arch)
# ===================================
# STDdata,STDindices,STDranges
# ===================================
std_list = ["+stddata", "+stdindices", "+stdranges"]
if spec_string.startswith(tuple(std_list)):
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
# ===================================
# CUDA
# ===================================
if ("+cuda" in self.spec) and ("~kokkos" in self.spec) and ("~acc" in self.spec):
# Set up the cuda macros needed by the build
cuda_arch_list = self.spec.variants["cuda_arch"].value
# the architecture value is only number so append sm_ to the name
cuda_arch = "sm_" + cuda_arch_list[0]
args.append("-DCUDA_ARCH=" + cuda_arch)
cuda_dir = self.spec["cuda"].prefix
cuda_comp = cuda_dir + "/bin/nvcc"
args.append("-DCMAKE_CUDA_COMPILER=" + cuda_comp)
args.append("-DMEM=" + self.spec.variants["mem"].value)
if self.spec.variants["flags"].value != "none":
args.append("-DCUDA_EXTRA_FLAGS=" + self.spec.variants["flags"].value)
# ===================================
# OMP
# ===================================
# `~kokkos` option is there to prevent +kokkos +omp setting to use omp directly from here
# Same applies for raja
if ("+omp" in self.spec) and ("~kokkos" in self.spec) and ("~raja" in self.spec):
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
if "cuda_arch" in self.spec.variants:
cuda_arch_list = self.spec.variants["cuda_arch"].value
# the architecture value is only number so append sm_ to the name
cuda_arch = "sm_" + cuda_arch_list[0]
args.append("-DOFFLOAD= " + "NVIDIA:" + cuda_arch)
elif "amdgpu_target" in self.spec.variants:
rocm_arch = self.spec.variants["amdgpu_target"].value
# the architecture value is only number so append sm_ to the name
args.append("-DOFFLOAD=" + " AMD:" + rocm_arch)
else:
args.append("-DOFFLOAD=" + "INTEL")
# ===================================
# SYCL
# ===================================
if self.spec.satisfies("+sycl"):
args.append("-DSYCL_COMPILER=" + self.spec.variants["implementation"].value.upper())
if self.spec.variants["implementation"].value.upper() != "ONEAPI-DPCPP":
args.append(
"-DSYCL_COMPILER_DIR=" + self.spec.variants["implementation"].value.upper()
)
if self.spec.variants["implementation"].value.upper() == "COMPUTE-CPP":
args.append("-DOpenCL_LIBRARY=")
# ===================================
# SYCL 2020
# ===================================
if self.spec.satisfies("+sycl2020"):
if self.spec.satisfies("%oneapi"):
# -fsycl flag is required for setting up sycl/sycl.hpp seems like
# it doesn't get it from the CMake file
args.append("-DCXX_EXTRA_FLAGS= -fsycl -O3")
# this is required to enable -DCMAKE_CXX_COMPILER=icpx flag from CMake
args.append("-DSYCL_COMPILER=ONEAPI-ICPX")
else:
args.append(
"-DSYCL_COMPILER=" + self.spec.variants["implementation"].value.upper()
)
if self.spec.variants["implementation"].value.upper() != "ONEAPI-DPCPP":
args.append(
"-DSYCL_COMPILER_DIR=" + self.spec.variants["implementation"].value.upper()
)
if self.spec.variants["implementation"].value.upper() == "COMPUTE-CPP":
args.append("-DOpenCL_LIBRARY=")
# ===================================
# HIP(ROCM)
# ===================================
if self.spec.satisfies("+rocm"):
hip_comp = self.spec["rocm"].prefix + "/bin/hipcc"
args.append("-DCMAKE_CXX_COMPILER=" + hip_comp)
args.append(
"-DCXX_EXTRA_FLAGS= --offload-arch="
+ self.spec.variants["amdgpu_target"].value
+ " "
+ self.spec.variants["flags"].value
+ " -O3"
)
# ===================================
# TBB
# ===================================
if self.spec.satisfies("+tbb"):
args.append("-DONE_TBB_DIR=" + self.spec["tbb"].prefix + "/tbb/latest/")
args.append("-DPARTITIONER=" + self.spec.variants["partitioner"].value.upper())
# ===================================
# OpenCL (ocl)
# ===================================
if self.spec.satisfies("+ocl"):
if "backend" in self.spec.variants:
if "cuda" in self.spec.variants["backend"].value:
cuda_dir = self.spec["cuda"].prefix
args.append("-DOpenCL_LIBRARY=" + cuda_dir + "/lib64/libOpenCL.so")
elif "amd" in self.spec.variants["backend"].value:
rocm_dir = self.spec["rocm-opencl"].prefix
args.append("-DOpenCL_LIBRARY=" + rocm_dir + "/lib64/libOpenCL.so")
elif "intel" in self.spec.variants["backend"].value:
intel_lib = (
self.spec["intel-oneapi-compilers"].prefix
+ "/compiler/2023.0.0/linux/lib/libOpenCL.so"
)
args.append("-DOpenCL_LIBRARY=" + intel_lib)
elif "pocl" in self.spec.variants["backend"].value:
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
pocl_lib = self.spec["pocl"].prefix + "/lib64/libOpenCL.so"
args.append("-DOpenCL_LIBRARY=" + pocl_lib)
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
# ===================================
# RAJA
# ===================================
if self.spec.satisfies("+raja"):
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
args.append("-DRAJA_IN_TREE=" + self.spec.variants["dir"].value)
if "offload" in self.spec.variants:
if "nvidia" in self.spec.variants["offload"].value:
cuda_dir = self.spec["cuda"].prefix
cuda_comp = cuda_dir + "/bin/nvcc"
args.append("-DCMAKE_CUDA_COMPILER=" + cuda_comp)
args.append("-DTARGET=NVIDIA")
cuda_arch_list = self.spec.variants["cuda_arch"].value
cuda_arch = "sm_" + cuda_arch_list[0]
args.append("-DCUDA_ARCH=" + cuda_arch)
args.append("DCUDA_TOOLKIT_ROOT_DIR=" + self.spec["cuda"].prefix)
if self.spec.variants["flags"].value != "none":
args.append("-DCUDA_EXTRA_FLAGS=" + self.spec.variants["flags"].value)
# if("cpu" in self.spec.variants['offload'].value):
if "omp" in self.spec.variants["backend"].value:
args.append("-DENABLE_OPENMP=ON")
if "cuda" in self.spec.variants["backend"].value:
args.append("-DENABLE_CUDA=ON")
# ===================================
# THRUST
# ===================================
if self.spec.satisfies("+thrust"):
if "cuda" in self.spec.variants["implementation"].value:
args.append("-DTHRUST_IMPL=" + self.spec.variants["implementation"].value.upper())
args.append("-SDK_DIR=" + self.spec["thrust"].prefix + "/include")
cuda_arch_list = self.spec.variants["cuda_arch"].value
# the architecture value is only number so append sm_ to the name
cuda_arch = "sm_" + cuda_arch_list[0]
args.append("-DCUDA_ARCH=" + cuda_arch)
cuda_dir = self.spec["cuda"].prefix
cuda_comp = cuda_dir + "/bin/nvcc"
args.append("-DCMAKE_CUDA_COMPILER=" + cuda_comp)
args.append("-DBACKEND=" + self.spec.variants["backend"].value.upper())
if self.spec.variants["flags"].value != "none":
args.append("-DCUDA_EXTRA_FLAGS=" + self.spec.variants["flags"].value)
if "rocm" in self.spec.variants["implementation"].value:
args.append("-DTHRUST_IMPL=" + self.spec.variants["implementation"].value.upper())
args.append("-SDK_DIR=" + self.spec["rocthrust"].prefix)
args.append("-DBACKEND=" + self.spec.variants["backend"].value.upper())
# ===================================
# kokkos
# ===================================
# kokkos implementation is versatile and it could use cuda or omp architectures as backend
# The usage should be spack install babelstream +kokkos +cuda [or +omp]
if self.spec.satisfies("+kokkos"):
args.append("-DCMAKE_CXX_COMPILER=" + self.compiler.cxx)
args.append("-DKOKKOS_IN_TREE=" + self.spec.variants["dir"].value)
# args.append("-DKOKKOS_IN_PACKAGE=" + self.spec["kokkos"].prefix)
if "backend" in self.spec.variants:
if "cuda" in self.spec.variants["backend"].value:
args.append("-DKokkos_ENABLE_CUDA=ON")
cuda_arch_list = self.spec.variants["cuda_arch"].value
cuda_arch = cuda_arch_list[0]
# arhitecture kepler optimisations
if cuda_arch in ("30", "32", "35", "37"):
args.append("-D" + "Kokkos_ARCH_KEPLER" + cuda_arch + "=ON")
# arhitecture maxwell optimisations
if cuda_arch in ("50", "52", "53"):
args.append("-D" + "Kokkos_ARCH_MAXWELL" + cuda_arch + "=ON")
# arhitecture pascal optimisations
if cuda_arch in ("60", "61"):
args.append("-D" + "Kokkos_ARCH_PASCAL" + cuda_arch + "=ON")
# architecture volta optimisations
if cuda_arch in ("70", "72"):
args.append("-D" + "Kokkos_ARCH_VOLTA" + cuda_arch + "=ON")
if cuda_arch == "75":
args.append("-DKokkos_ARCH_TURING75=ON")
if "omp" in self.spec.variants["backend"].value:
args.append("-DKokkos_ENABLE_OPENMP=ON")
# not in ["kokkos", "raja", "acc", "hip"] then compiler forced true
if set(model_list).intersection(["kokkos", "raja", "acc", "hip"]) is True:
args.append("-DCMAKE_CXX_COMPILER_FORCED=True")
return args
Reference in New Issue View Git Blame Copy Permalink