Skip to content

Commit

Permalink
Update CUDA with Tim's improvements, and to new BB syntax
Browse files Browse the repository at this point in the history
  • Loading branch information
@staticfloat
staticfloat committed Oct 4, 2019
1 parent e0579fc commit 4b2fe25
Showing 1 changed file with 92 additions and 56 deletions.
148 changes: 92 additions & 56 deletions C/CUDA/build_tarballs.jl
View file
Original file line number Diff line number Diff line change
@@ -1,69 +1,105 @@
using BinaryBuilder

name = "CUDA"
version = v"10.1.168"
version = v"10.1.243"

sources = [
"https://developer.nvidia.com/compute/cuda/10.1/Prod/cluster_management/cuda_cluster_pkgs_10.1.168_418.67_rhel6.tar.gz" =>
"965570c92de387cec04d77a2bdce26b6457b027c0b2b12dc537a5ca1c1aa82b3",
"https://developer.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.168_mac.dmg" =>
"a53d17c92b81bb8b8f812d0886a8c2ddf2730be6f5f2659aee11c0da207c2331",
"https://developer.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.168_425.25_win10.exe" =>
"52450b81a699cb75086e9d3d62abb2a33f823fcf5395444e57ebb5864cc2fd51",
"http://developer.download.nvidia.com/compute/cuda/10.1/Prod/cluster_management/cuda_cluster_pkgs_10.1.243_418.87.00_rhel6.tar.gz" =>
"024b61d193105aef37241c89e511f0fec9dcecc2af416f2a1151f2a4dbbb3c29",
"http://developer.download.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.243_mac.dmg" =>
"432a2f07a793f21320edc5d10e7f68a8e4e89465c31e1696290bdb0ca7c8c997",
"http://developer.download.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.243_426.00_win10.exe" =>
"35d3c99c58dd601b2a2caa28f44d828cae1eaf8beb70702732585fa001cd8ad7",
]

# CUDA is weirdly organized, with several tools in bin/lib directories, some in dedicated
# subproject folders, and others in a catch-all extras/ directory. to simplify using
# the resulting binaries, we reorganize everything using a flat bin/lib structure.

script = raw"""
cd ${WORKSPACE}/srcdir
apk add p7zip rpm
# Make temporary space for extraction
mkdir -p $(pwd)/.tmp
if [[ ${target} == x86_64-linux-gnu ]]; then
cd cuda_cluster_pkgs*
# extract cluster packages
rpm2cpio cuda-cluster-devel*.rpm | cpio -idmv
rpm2cpio cuda-cluster-runtime*.rpm | cpio -idmv
rpm2cpio cuda-cluster-devel*.rpm | cpio -idmv
cd usr/local/cuda*
# Install things we like
for subdir in bin lib64 include extras nvvm nvml targets; do
mkdir -p ${prefix}/${subdir}
mv usr/local/cuda*/${subdir}/* ${prefix}/${subdir}/
done
# toplevel
mv bin ${prefix}
mv targets/x86_64-linux/lib ${prefix}
mkdir ${prefix}/share
# We need to maintain the "targets" directories
for dir in include lib; do
rm -rf ${prefix}/targets/x86_64-linux/${dir}
ln -s ../../${dir} ${prefix}/targets/x86_64-linux/${dir}
# nested
for project in nvvm extras/CUPTI; do
[[ -d ${project}/bin ]] && mv ${project}/bin/* ${prefix}/bin
[[ -d ${project}/lib64 ]] && mv ${project}/lib64/* ${prefix}/lib
done
mv nvvm/libdevice ${prefix}/share
# clean up
rm ${prefix}/bin/{nvcc,nvcc.profile,cicc,cudafe++} # CUDA C/C++ compiler
rm -r ${prefix}/bin/crt/
rm ${prefix}/bin/{gpu-library-advisor,bin2c} # C/C++ utilities
rm ${prefix}/bin/{nvvp,nsight,computeprof} # requires Java
rm ${prefix}/lib/*.a # we can't link statically
rm -r ${prefix}/lib/stubs/ # stubs are a C/C++ thing
rm ${prefix}/bin/cuda-install-samples-*.sh
rm ${prefix}/bin/nsight_ee_plugins_manage.sh
elif [[ ${target} == x86_64-w64-mingw32 ]]; then
cd .tmp
7z x ${WORKSPACE}/srcdir/cuda_*_win10.exe
# Install things
mkdir -p ${prefix}/bin ${prefix}/include ${prefix}/lib/x64
for project in curand cusparse npp cufft cublas cudart cusolver nvrtc; do
mv ${project}/bin/* ${prefix}/bin/
[[ -d ${project}_dev/include ]] && mv ${project}_dev/include/* ${prefix}/include/
[[ -d ${project}_dev/lib ]] && mv ${project}_dev/lib/x64/* ${prefix}/lib/x64/
7z x ${WORKSPACE}/srcdir/*-cuda_*_win10.exe -bb
# toplevel
mkdir -p ${prefix}/bin ${prefix}/share
# no lib folder; we don't ship static libs
# nested
for project in cuobjdump memcheck nvcc nvcc/nvvm nvdisasm curand cusparse npp cufft cublas cudart cusolver nvrtc nvgraph nvprof nvprune; do
[[ -d ${project}/bin ]] && mv ${project}/bin/* ${prefix}/bin
done
elif [[ ${target} == x86_64-apple-darwin* ]]; then
cd .tmp
7z x ${WORKSPACE}/srcdir/cuda_*_mac.dmg
7z x 5.hfs
mv nvcc/nvvm/libdevice ${prefix}/share
mv cupti/extras/CUPTI/lib64/* ${prefix}/bin/
# Extract embedded tarball into prefix
tar -C ${prefix} -zxvf CUDAMacOSXInstaller/CUDAMacOSXInstaller.app/Contents/Resources/payload/cuda_mac_installer_tk.tar.gz
# clean up
rm ${prefix}/bin/{nvcc,cicc,cudafe++}.exe # CUDA C/C++ compiler
rm ${prefix}/bin/nvcc.profile
rm -r ${prefix}/bin/crt/
rm ${prefix}/bin/bin2c.exe # C/C++ utilities
rm ${prefix}/bin/*.lib # we can't link statically
# Clean up things we don't care about
rm -rf ${prefix}/usr
# Make .exe's executable
chmod +x ${prefix}/bin/*.exe
for subdir in bin lib include extras nvvm; do
mv ${prefix}/Developer/NVIDIA/CUDA-*/${subdir} ${prefix}/
elif [[ ${target} == x86_64-apple-darwin* ]]; then
7z x ${WORKSPACE}/srcdir/*-cuda_*_mac.dmg
7z x 5.hfs
tar -zxvf CUDAMacOSXInstaller/CUDAMacOSXInstaller.app/Contents/Resources/payload/cuda_mac_installer_tk.tar.gz
cd Developer/NVIDIA/CUDA-*/
# toplevel
mv bin ${prefix}
mv lib ${prefix}
mkdir ${prefix}/share
# nested
for project in nvvm extras/CUPTI; do
[[ -d ${project}/bin ]] && mv ${project}/bin/* ${prefix}/bin
[[ -d ${project}/lib ]] && mv ${project}/lib/* ${prefix}/lib
done
rm -rf ${prefix}/Developer
mv nvvm/libdevice ${prefix}/share
# clean up
rm ${prefix}/bin/{nvcc,nvcc.profile,cicc,cudafe++} # CUDA C/C++ compiler
rm -r ${prefix}/bin/crt/
rm ${prefix}/bin/{gpu-library-advisor,bin2c} # C/C++ utilities
rm ${prefix}/bin/{nvvp,nsight,computeprof} # requires Java
rm ${prefix}/lib/*.a # we can't link statically
rm -r ${prefix}/lib/stubs/ # stubs are a C/C++ thing
rm ${prefix}/bin/uninstall_cuda_*.pl
rm ${prefix}/bin/nsight_ee_plugins_manage.sh
rm ${prefix}/bin/.cuda_toolkit_uninstall_manifest_do_not_delete.txt
fi
"""

Expand All @@ -73,21 +109,21 @@ platforms = [
MacOS(:x86_64),
]

products(prefix) = [
ExecutableProduct(prefix, "nvcc", :nvcc),
ExecutableProduct(prefix, "cudafe++", :cudafepp),
ExecutableProduct(prefix, "nvprof", :nvprof),
ExecutableProduct(prefix, "ptxas", :ptxas),
LibraryProduct(prefix, "libcudart", :libcudart),
LibraryProduct(prefix, "libcufft", :libcufft),
LibraryProduct(prefix, "libcufftw", :libcufftw),
LibraryProduct(prefix, "libcurand", :libcurand),
LibraryProduct(prefix, "libcublas", :libcublas),
LibraryProduct(prefix, "libcusolver", :libcusolver),
LibraryProduct(prefix, "libcusparse", :libcusparse),
LibraryProduct(prefix, "libnvrtc", :libnvrtc),
# cuda-gdb, libnvjpeg, libOpenCL, libaccinj(64), libnvperf_host, libnvperf_target only on linux

products = [
ExecutableProduct("nvprof", :nvprof),
ExecutableProduct("ptxas", :ptxas),
LibraryProduct(["libcudart", "cudart", "cudart64_101"], :libcudart),
LibraryProduct(["libcufft", "cufft", "cufft64_10"], :libcufft),
LibraryProduct(["libcufftw", "cufftw", "cufftw64_10"], :libcufftw),
LibraryProduct(["libcurand", "curand", "curand64_10"], :libcurand),
LibraryProduct(["libcublas", "cublas", "cublas64_10"], :libcublas),
LibraryProduct(["libcusolver", "cusolver", "cusolver64_10"], :libcusolver),
LibraryProduct(["libcusparse", "cusparse", "cusparse64_10"], :libcusparse),
LibraryProduct(["libnvrtc", "nvrtc", "nvrtc64_101_0"], :libnvrtc),
]

dependencies = []

build_tarballs(ARGS, name, version, sources, script, platforms, products, dependencies; skip_audit=true)
build_tarballs(ARGS, name, version, sources, script, platforms, products, dependencies)

2 comments on commit 4b2fe25

@staticfloat
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@maleadt just FYI, I have upgraded your CUDA recipe here to run with the latest BB, which gives us a JLL package that I just registered. Great job on squeezing the overall size down, only 600MB!

@maleadt
Copy link
Contributor

@maleadt maleadt commented on 4b2fe25 Oct 5, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Great! I'll be doing some more development soon, probably over at the CUDABuilder.jl repo to avoid noise here (I still need to figure out how to efficiently provide multiple versions of CUDA), but I'll be sure to push significant updates here too.

Please sign in to comment.