commit: afea716210fdc48b59dfc58f107fc7086b53562b Author: Paul Zander <negril.nx+gentoo <AT> gmail <DOT> com> AuthorDate: Thu Dec 5 12:16:31 2024 +0000 Commit: Sam James <sam <AT> gentoo <DOT> org> CommitDate: Thu Dec 5 18:01:14 2024 +0000 URL: https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=afea7162
sci-libs/vtk: 9.3.1 update cuda handling Allow cuda-12.6 Closes: https://bugs.gentoo.org/938673 Closes: https://bugs.gentoo.org/930031 Signed-off-by: Paul Zander <negril.nx+gentoo <AT> gmail.com> Signed-off-by: Sam James <sam <AT> gentoo.org> sci-libs/vtk/vtk-9.3.1.ebuild | 211 +++++++++++++++++++++++++++++++++--------- 1 file changed, 165 insertions(+), 46 deletions(-) diff --git a/sci-libs/vtk/vtk-9.3.1.ebuild b/sci-libs/vtk/vtk-9.3.1.ebuild index 895bd1487f9e..0eea06d8c49f 100644 --- a/sci-libs/vtk/vtk-9.3.1.ebuild +++ b/sci-libs/vtk/vtk-9.3.1.ebuild @@ -51,7 +51,7 @@ REQUIRED_USE=" boost cgns ffmpeg freetype gdal imaging las mysql netcdf odbc opencascade openvdb pdal postgres rendering views ) - cuda? ( video_cards_nvidia vtkm !tbb ) + cuda? ( video_cards_nvidia vtkm ) java? ( rendering ) minimal? ( !rendering ) !minimal? ( cgns netcdf rendering ) @@ -156,6 +156,7 @@ PATCHES=( "${FILESDIR}/${PN}-9.3.0-opencascade.patch" "${FILESDIR}/${PN}-9.3.0-ThrustPatches.patch" "${FILESDIR}/${PN}-9.3.0-ThirdParty-gcc15.patch" + "${FILESDIR}/${PN}-9.3.0-update-for-cuda-12.6.patch" ) DOCS=( CONTRIBUTING.md README.md ) @@ -186,55 +187,154 @@ vtk_check_reqs() { "check-reqs_pkg_${EBUILD_PHASE}" } -vtk_check_compiler() { - [[ -z "$1" ]] && die "no compiler specified" - local compiler="$1" - local package="sys-devel/${compiler}" - local version="${package}" - local CUDAHOSTCXX_test - while - CUDAHOSTCXX="${CUDAHOSTCXX_test}" - version=$(best_version "${version}") - if [[ -z "${version}" ]]; then - if [[ -z "${CUDAHOSTCXX}" ]]; then - die "could not find supported version of ${package}" +cuda_get_host_compiler() { + if [[ -n "${NVCC_CCBIN}" ]]; then + echo "${NVCC_CCBIN}" + return + fi + + if [[ -n "${CUDAHOSTCXX}" ]]; then + echo "${CUDAHOSTCXX}" + return + fi + + einfo "Trying to find working CUDA host compiler" + + if ! tc-is-gcc && ! tc-is-clang; then + die "$(tc-get-compiler-type) compiler is not supported" + fi + + local compiler compiler_type compiler_version + local package package_version + local NVCC_CCBIN_default + + compiler_type="$(tc-get-compiler-type)" + compiler_version="$("${compiler_type}-major-version")" + + # try the default compiler first + NVCC_CCBIN="$(tc-getCXX)" + NVCC_CCBIN_default="${NVCC_CCBIN}-${compiler_version}" + + compiler="${NVCC_CCBIN/%-${compiler_version}}" + + # store the package so we can re-use it later + package="sys-devel/${compiler_type}" + package_version="${package}" + + ebegin "testing ${NVCC_CCBIN_default} (default)" + + while ! nvcc -v -ccbin "${NVCC_CCBIN}" - -x cu <<<"int main(){}" &>> "${T}/cuda_get_host_compiler.log" ; do + eend 1 + + while true; do + # prepare next version + if ! package_version="<$(best_version "${package_version}")"; then + die "could not find a supported version of ${compiler}" fi - break - fi - CUDAHOSTCXX_test="$( - dirname "$( - realpath "$( - which "${compiler}-$(echo "${version}" | grep -oP "(?<=${package}-)[0-9]*")" - )" - )" + + NVCC_CCBIN="${compiler}-$(ver_cut 1 "${package_version/#<${package}-/}")" + + [[ "${NVCC_CCBIN}" != "${NVCC_CCBIN_default}" ]] && break + done + ebegin "testing ${NVCC_CCBIN}" + done + eend $? + + # clean temp file + rm -f a.out + + echo "${NVCC_CCBIN}" + export NVCC_CCBIN +} + +cuda_get_host_native_arch() { + [[ -n ${CUDAARCHS} ]] && echo "${CUDAARCHS}" + + __nvcc_device_query || die "failed to query the native device" +} + +vtk_add_sandbox() { + local WRITE=() + + # mesa via virtx will make use of udmabuf if it exists + [[ -c "/dev/udmabuf" ]] && WRITE+=( "/dev/udmabuf" ) + + readarray -t dris <<<"$( + for dri in /sys/class/drm/*/dev; do + realpath "/dev/char/$(cat "${dri}")" + eqawarn "dri ${dri} $(cat "${dri}") $(realpath "/dev/char/$(cat "${dri}")")" + done + )" + + [[ -n "${dris[*]}" ]] && WRITE+=( "${dris[@]}" ) + + if [[ -d /sys/module/nvidia ]]; then + # /dev/nvidia{0-9} + readarray -t nvidia_devs <<<"$( + find /dev -regextype posix-extended -regex '/dev/nvidia(|-(nvswitch|vgpu))[0-9]*' )" - version="<${version}" - do ! echo "int main(){}" | nvcc "-ccbin=${CUDAHOSTCXX_test}" - -x cu &>/dev/null; done + [[ -n "${nvidia_devs[*]}" ]] && WRITE+=( "${nvidia_devs[@]}" ) + + WRITE+=( + "/dev/nvidiactl" + "/dev/nvidia-modeset" + + "/dev/nvidia-vgpuctl" + + "/dev/nvidia-nvlink" + "/dev/nvidia-nvswitchctl" + + "/dev/nvidia-uvm" + "/dev/nvidia-uvm-tools" + + # "/dev/nvidia-caps/nvidia-cap%d" + "/dev/nvidia-caps/" + # "/dev/nvidia-caps-imex-channels/channel%d" + "/dev/nvidia-caps-imex-channels/" + ) + fi + + # for portage + WRITE+=( "/proc/self/task/" ) + + local dev + for dev in "${WRITE[@]}"; do + [[ ! -e "${dev}" ]] && return + + [[ -w "${dev}" ]] && return + + eqawarn "addwrite ${dev}" + addwrite "${dev}" + if [[ ! -d "${dev}" ]] && [[ ! -w "${dev}" ]]; then + eerror "can not access ${dev} after addwrite" + fi + done } pkg_pretend() { [[ ${MERGE_TYPE} != binary ]] && has openmp && tc-check-openmp - if [[ $(tc-is-gcc) && $(gcc-majorversion) = 11 ]] && use cuda ; then - # FIXME: better use eerror? - ewarn "GCC 11 is know to fail building with CUDA support in some cases." - ewarn "See bug #820593" - fi - vtk_check_reqs + + # When building binpkgs you probably want to include all targets + if use cuda && [[ ${MERGE_TYPE} == "buildonly" ]] && [[ -n "${CUDA_GENERATION}" || -n "${CUDA_ARCH_BIN}" ]]; then + local info_message="When building a binary package it's recommended to unset CUDA_GENERATION and CUDA_ARCH_BIN" + einfo "$info_message so all available architectures are build." + fi } pkg_setup() { [[ ${MERGE_TYPE} != binary ]] && has openmp && tc-check-openmp - if [[ $(tc-is-gcc) && $(gcc-majorversion) = 11 ]] && use cuda ; then - # FIXME: better use eerror? - ewarn "GCC 11 is know to fail building with CUDA support in some cases." - ewarn "See bug #820593" - fi - vtk_check_reqs + if use cuda && [[ ! -e /dev/nvidia-uvm ]]; then + # NOTE We try to load nvidia-uvm and nvidia-modeset here, + # so __nvcc_device_query does not fail later. + + nvidia-modprobe -m -u -c 0 || true + fi + use java && java-pkg-opt-2_pkg_setup use python && python-single-r1_pkg_setup use web && webapp_pkg_setup @@ -436,15 +536,36 @@ src_configure() { if use cuda; then cuda_add_sandbox -w - tc-is-gcc && vtk_check_compiler "gcc" - tc-is-clang && vtk_check_compiler "clang" - [[ -z "${CUDAARCHS}" ]] && einfo "trying to determine host CUDAARCHS" - : "${CUDAARCHS:=$(__nvcc_device_query)}" - einfo "building for CUDAARCHS = ${CUDAARCHS}" + addwrite "/proc/self/task" + + if ! test -w /dev/nvidiactl; then + # eqawarn "Can't access the GPU at /dev/nvidiactl." + # eqawarn "User $(id -nu) is not in the group \"video\"." + if [[ -z "${CUDA_GENERATION}" ]] && [[ -z "${CUDA_ARCH_BIN}" ]]; then + # build all targets + mycmakeargs+=( + -DCUDA_GENERATION="" + ) + fi + else + local -x CUDAARCHS + : "${CUDAARCHS:="$(cuda_get_host_native_arch)"}" + fi - export CUDAARCHS - export CUDAHOSTCXX - unset NVCCFLAGS + # set NVCC_CCBIN + local -x CUDAHOSTCXX CUDAHOSTLD + CUDAHOSTCXX="$(cuda_get_host_compiler)" + CUDAHOSTLD="$(tc-getCXX)" + export NVCC_CCBIN="${CUDAHOSTCXX}" + + if tc-is-gcc; then + # Filter out IMPLICIT_LINK_DIRECTORIES picked up by CMAKE_DETERMINE_COMPILER_ABI(CUDA) + # See /usr/share/cmake/Help/variable/CMAKE_LANG_IMPLICIT_LINK_DIRECTORIES.rst + CMAKE_CUDA_IMPLICIT_LINK_DIRECTORIES_EXCLUDE=$( + "${CUDAHOSTLD}" -E -v - <<<"int main(){}" |& \ + grep LIBRARY_PATH | cut -d '=' -f 2 | cut -d ':' -f 1 + ) + fi fi if use debug; then @@ -695,9 +816,7 @@ src_compile() { } src_test() { - if use cuda; then - cuda_add_sandbox -w - fi + vtk_add_sandbox # don't work at all REALLY_BAD_TESTS=(
