diff options
Diffstat (limited to 'PKGBUILD')
| -rw-r--r-- | PKGBUILD | 381 |
1 files changed, 290 insertions, 91 deletions
@@ -1,137 +1,336 @@ -# Maintainer: Chih-Hsuan Yen <yan12125@gmail.com> +# Maintainer: Gustavo Alvarez <sl1pkn07@gmail.com> +# Contributor: Chih-Hsuan Yen <yan12125@gmail.com> + +# MAKEFLAGS=-j2 # NOTE Can be usseful when got OOM + +_ENABLE_CUDA=1 +_ENABLE_TENSORRT=0 # NOTE: not working due https://github.com/microsoft/onnxruntime/issues/15131 pkgbase=python-onnxruntime # Not split DNNL EP to another package as it's needed unconditionally at runtime if built at compile time # https://github.com/microsoft/onnxruntime/blob/v1.9.1/onnxruntime/python/onnxruntime_pybind_state.cc#L533 -pkgname=(python-onnxruntime python-onnxruntime-cuda) -pkgver=1.9.1 +pkgname=( + 'onnxruntime' + 'python-onnxruntime' +) +pkgver=1.16.3 pkgdesc='Cross-platform, high performance scoring engine for ML models' -pkgrel=4 -arch=(x86_64) +pkgrel=1 +arch=('x86_64') url='https://github.com/microsoft/onnxruntime' -license=(MIT) -depends=(nsync re2 python-flatbuffers python-numpy python-onnx python-protobuf openmpi onednn) -makedepends=(git cmake gtest gmock pybind11 python-setuptools nlohmann-json chrono-date boost eigen flatbuffers cuda cudnn nccl clang) +license=('MIT') +makedepends=( + 'git' + 'cmake' + 'ninja' + 'gcc-libs' + 'glibc' + 'cxxopts' + 'pybind11' + 'abseil-cpp' + 'nlohmann-json' + 'chrono-date' + 'boost' + 'eigen' +# 'flatbuffers' + 'onednn' +# 're2' +# 'protobuf' + 'nsync' + 'openmpi' + 'python-coloredlogs' + 'python-flatbuffers' + 'python-numpy' +# 'python-protobuf' + 'python-sympy' + 'python-setuptools' + 'python-installer' + 'python-wheel' + 'python-build' + + 'chrpath' +) # not de-vendored libraries # onnx: needs shared libonnx (https://github.com/onnx/onnx/issues/3030) -source=("git+https://github.com/microsoft/onnxruntime#tag=v$pkgver" - "git+https://github.com/onnx/onnx.git" - "git+https://github.com/dcleblanc/SafeInt.git" - "git+https://github.com/martinmoene/optional-lite.git" - "git+https://github.com/tensorflow/tensorboard.git" - "git+https://github.com/dmlc/dlpack.git" - "git+https://github.com/jarro2783/cxxopts.git" - "pytorch_cpuinfo::git+https://github.com/pytorch/cpuinfo.git" - build-fixes.patch - clang.patch - system-dnnl.diff) -sha512sums=('SKIP' - 'SKIP' - 'SKIP' - 'SKIP' - 'SKIP' - 'SKIP' - 'SKIP' - 'SKIP' - '685f0235abed6e1277dd0eb9bda56c464d1987fe7fc90a3550e17ec70cc49fd15f34996a0e159f9622c4ca3e6bf29917fe51b7849342531fa2a6808d782f1e06' - 'ad94af8bb25744b244c4f82e9a06189741f82b295a88523ca0e8005568fac710c2299d783989457e9cf96ef8da0593fb4f70c8792d416f44ab29d6493e204f13' - '6735c7aca2ba2f1f2a5286eb064125bf7f2c68a575d572dd157769d15778ff3e717b3a53d696c767748229f23ee6c3a7c82679df1d86283d7c4dd0ec9103ae08') -# CUDA seems not working with LTO -options+=('!lto') +source=( + "git+https://github.com/microsoft/onnxruntime#tag=v${pkgver}" + 'install-orttraining-files.diff' + 'system-dnnl.diff' +# 'system-flatbuffers.patch' +) +sha512sums=( + 'SKIP' + 'SKIP' + 'SKIP' +# 'SKIP' +) +options=('debug') + +if [[ $_ENABLE_CUDA = 1 ]]; then + pkgname+=('onnxruntime-cuda') + makedepends+=( + 'cuda' + 'cudnn' + 'nccl' + ) +fi + +if [[ $_ENABLE_TENSORRT = 1 ]]; then + pkgname+=('onnxruntime-tensorrt') + makedepends+=('tensorrt') +# depends+=('protobuf' 'libprotobuf.so') +fi # Check PKGBUILDs of python-pytorch and tensorflow for CUDA architectures built by official packages -_CUDA_ARCHITECTURES="52-real;53-real;60-real;61-real;62-real;70-real;72-real;75-real;80-real;86-real;86-virtual" +_CUDA_ARCHITECTURES="52-real;53-real;60-real;61-real;62-real;70-real;72-real;75-real;80-real;86-real;87-real;89-real;90-real;90-virtual" prepare() { cd onnxruntime - patch -Np1 -i ../build-fixes.patch - patch -Np1 -i ../clang.patch - patch -Np1 -i ../system-dnnl.diff + # Use System Dlnn + patch -Np1 -i "${srcdir}/system-dnnl.diff" + + # Find system nlohmann-json + sed 's|3.10 ||g' \ + -i cmake/external/onnxruntime_external_deps.cmake + + # Find system chrono-date + sed -e 's|${DEP_SHA1_date}|&\n \ \ \ \ \ \FIND_PACKAGE_ARGS NAMES date|g' \ + -e 's|date_interface|date::date-tz|g' \ + -i cmake/external/onnxruntime_external_deps.cmake \ + -i cmake/onnxruntime_common.cmake \ + -i cmake/onnxruntime_unittests.cmake + + # Find system abseil-cpp + sed 's|ABSL_PATCH_COMMAND}|&\n\ \ \ \ \FIND_PACKAGE_ARGS NAMES absl|g' \ + -i cmake/external/abseil-cpp.cmake + + # Find system cxxopts + sed 's|${DEP_SHA1_cxxopts}|&\n\ \ \ \ \FIND_PACKAGE_ARGS NAMES cxxopts|g' \ + -i cmake/external/onnxruntime_external_deps.cmake + +# # Find system mimalloc +# sed 's|${DEP_SHA1_mimalloc}|&\n\ \ \ \ \ \ \FIND_PACKAGE_ARGS NAMES mimalloc|g' \ +# -i cmake/external/onnxruntime_external_deps.cmake + + # Find system nsync + sed -e 's|NAMES nsync|&_cpp|g' \ + -e '295aadd_library(nsync::nsync_cpp ALIAS nsync_cpp)' \ + -i cmake/external/onnxruntime_external_deps.cmake - git submodule init - for mod in onnx SafeInt optional-lite tensorboard dlpack cxxopts pytorch_cpuinfo; do - git config submodule.cmake/external/$mod.url "$srcdir"/$mod - git submodule update cmake/external/$mod - done + if [[ $_ENABLE_TENSORRT = 1 ]]; then + # Update Tensorboard 8343cad89d984c199637ead11c8d4c053191673a (2.15.1) +# # Update Tensorboard a01ceb5957d9ecd56314df115c09e3ddb60d12f7 +# sed -e 's|373eb09e4c5d2b3cc2493f0949dc4be6b6a45e81|a01ceb5957d9ecd56314df115c09e3ddb60d12f7|g' \ +# -e 's|ff427b6a135344d86b65fa2928fbd29886eefaec|113750f323d131859ac4e17070d2c9417e80d701|g' \ +# -i cmake/deps.txt + + + # Update onnx_tensorrt 6ba67d3428e05f690145373ca87fb8d32f98df45 (8.6 GA) + sed -e 's|0462dc31ae78f48744b6141ae376df1f96d3f459|6ba67d3428e05f690145373ca87fb8d32f98df45|g' \ + -e 's|67b833913605a4f3f499894ab11528a702c2b381|67b833913605a4f3f499894ab11528a702c2b381|g' \ + -i cmake/deps.txt + fi + + patch -Np1 -i "${srcdir}/install-orttraining-files.diff" +# patch -Np1 -i "${srcdir}/system-flatbuffers.patch" + + # fix build with gcc12(?), take idea from https://github.com/microsoft/onnxruntime/pull/11667 and https://github.com/microsoft/onnxruntime/pull/10014 + sed 's|dims)|TensorShape(dims))|g' \ + -i onnxruntime/contrib_ops/cuda/quantization/qordered_ops/qordered_qdq.cc + + # fix missing #include <iostream> + sed '11a#include <iostream>' \ + -i orttraining/orttraining/test/training_api/trainer/trainer.cc + +# cd onnxruntime/core/flatbuffers/schema +# python compile_schema.py --flatc /usr/bin/flatc } build() { - cd "$srcdir"/onnxruntime - local cmake_args=( + if [[ ${_ENABLE_CUDA} = 1 ]]; then + export CC="/opt/cuda/bin/gcc" + export CXX="/opt/cuda/bin/g++" + export CUDAHOSTCXX="${CXX}" + fi + + # Gcc 12+ + CXXFLAGS+=" -Wno-maybe-uninitialized -Wno-error=restrict" + CFLAGS="${CFLAGS/_FORTIFY_SOURCE=2/_FORTIFY_SOURCE=0}" + CXXFLAGS="${CXXFLAGS/_FORTIFY_SOURCE=2/_FORTIFY_SOURCE=0}" + + + # Use -Donnxruntime_ENABLE_LAZY_TENSOR=OFF as it requires patched python-pytorch + # See: https://github.com/microsoft/onnxruntime/pull/10460 https://github.com/pytorch/pytorch/pulls/wschin + local _cmake_args=( + -DCMAKE_BUILD_TYPE=Debug -DCMAKE_INSTALL_PREFIX=/usr + -DCMAKE_SKIP_INSTALL_RPATH=OFF + -DCMAKE_SKIP_RPATH=OFF -Donnxruntime_ENABLE_PYTHON=ON - -Donnxruntime_PREFER_SYSTEM_LIB=ON -Donnxruntime_BUILD_SHARED_LIB=ON + -Donnxruntime_BUILD_UNIT_TESTS=OFF -Donnxruntime_ENABLE_TRAINING=ON + -Donnxruntime_ENABLE_LAZY_TENSOR=OFF -Donnxruntime_USE_MPI=ON - -Donnxruntime_USE_PREINSTALLED_EIGEN=ON -Donnxruntime_USE_DNNL=ON - -Deigen_SOURCE_PATH=/usr/include/eigen3 + -Donnxruntime_USE_PREINSTALLED_EIGEN=ON + -Deigen_SOURCE_PATH="$(pkg-config --cflags eigen3 | sed 's|-I||g')" + -DCMAKE_CXX_STANDARD=17 + -DCMAKE_IGNORE_PATH=/usr/lib/cmake/flatbuffers/\;/lib/cmake/flatbuffers/\;/usr/lib/cmake/protobuf/\;/lib/cmake/protobuf/ + -DBUILD_TESTING=OFF ) # Use protobuf-lite instead of full protobuf to workaround symbol conflicts # with onnx; see https://github.com/onnx/onnx/issues/1277 for details. - cmake_args+=( - -DONNX_CUSTOM_PROTOC_EXECUTABLE=/usr/bin/protoc + _cmake_args+=( -Donnxruntime_USE_FULL_PROTOBUF=OFF ) - # 1. Redefine ___is_signed to ___is_signed to workaround a regression - # from CUDA 11.3 -> 11.3.1 [1]. - # 2. Enable parallel builds for NVCC via -t0, which spawns multiple - # cicc and ptxas processes for each nvcc invocation. The number of - # total processes may be much larger than the number of cores - let - # the scheduler handle it. - # [1] https://forums.developer.nvidia.com/t/182176 - cmake_args+=( - -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/clang - -DCMAKE_CUDA_FLAGS="-D__is_signed=___is_signed -t0" - -DCMAKE_CUDA_ARCHITECTURES="$_CUDA_ARCHITECTURES" - -Donnxruntime_USE_CUDA=ON - -Donnxruntime_CUDA_HOME=/opt/cuda - -Donnxruntime_CUDNN_HOME=/usr - -Donnxruntime_USE_NCCL=ON - ) + if [[ ${_ENABLE_CUDA} = 1 ]]; then + _cmake_args+=( + -DCMAKE_CUDA_ARCHITECTURES="${_CUDA_ARCHITECTURES}" + -DCMAKE_CUDA_STANDARD_REQUIRED=ON + -DCMAKE_CXX_STANDARD_REQUIRED=ON + -Donnxruntime_USE_CUDA=ON + -Donnxruntime_CUDA_HOME=/opt/cuda + -Donnxruntime_CUDNN_HOME=/usr + -Donnxruntime_USE_NCCL=ON + -Donnxruntime_NVCC_THREADS=1 + ) + fi - # Use clang as GCC does not work. GCC 11 crashes with internal - # compiler errors. GCC 10 does not work as some dependent packages - # (ex: re2) are built with libstdc++ from GCC 11, and thus linking - # onnxruntime with libstdc++ 10 fails. - CC=/usr/bin/clang CXX=/usr/bin/clang++ \ - cmake -B build -S cmake "${cmake_args[@]}" "$@" + if [[ ${_ENABLE_TENSORRT} = 1 ]]; then + _cmake_args+=( + -Donnxruntime_USE_TENSORRT=ON + -Donnxruntime_USE_TENSORRT_BUILTIN_PARSER=ON + ) + fi + cmake -S onnxruntime/cmake -B build \ + "${_cmake_args[@]}" \ + "$@" \ + -G Ninja + + LC_ALL=C cmake --build build #-v + + ( cd build - make - python ../setup.py build + install -Dm644 ../onnxruntime/docs/python/README.rst docs/python/README.rst + ln -s ../onnxruntime/setup.py . + python -m build --wheel --no-isolation + ) + } -package_python-onnxruntime() { - cd onnxruntime/build +package_onnxruntime() { + depends=( + 'gcc-libs' # libgcc_s.so libstdc++.so + 'glibc' # ld-linux-x86-64.so libc.so ibm.so + 'onednn' # libdnnl.so + 'openmpi' 'libmpi.so' + 'abseil-cpp' # libabsl_hash.so libabsl_raw_hash_set.so libabsl_raw_logging_internal.so libabsl_throw_delegate.so + 'nsync' # libnsync_cpp.so +# 'protobuf' 'libprotobuf-lite.so' + ) + provides=( + 'libonnxruntime.so' + 'libonnxruntime_providers_shared.so' + ) - make install DESTDIR="$pkgdir" + DESTDIR="${pkgdir}" cmake --install build - python ../setup.py install --root="$pkgdir" --skip-build --optimize=1 + # installed as split packages + rm -vf "${pkgdir}/usr/lib/"libonnxruntime_providers_{tensorrt,cuda}.so - PY_ORT_DIR="$(python -c 'import site; print(site.getsitepackages()[0])')/onnxruntime" - install -Ddm755 "$pkgdir"/usr/share/licenses/$pkgname - for f in LICENSE ThirdPartyNotices.txt ; do - ln -s "$PY_ORT_DIR/$f" "$pkgdir"/usr/share/licenses/$pkgname/$f - done - # already installed by `make install`, and not useful as this path is not looked up by the linker - rm -vf "$pkgdir/$PY_ORT_DIR"/capi/libonnxruntime_providers_* + chrpath -d "${pkgdir}/usr/lib/"libonnxruntime.so.* - # installed as split packages - rm -vf "$pkgdir"/usr/lib/libonnxruntime_providers_cuda.so + install -Dm644 onnxruntime/LICENSE "${pkgdir}/usr/share/licenses/${pkgname}/LICENSE" + install -Dm644 onnxruntime/ThirdPartyNotices.txt "${pkgdir}/usr/share/licenses/${pkgname}/ThirdPartyNotices.txt" +} + +package_python-onnxruntime() { + pkgdesc+=' (Python Bindings)' + depends=( + 'onnxruntime' + 'gcc-libs' # libgcc_s.so libstdc++.so + 'glibc' # ld-linux-x86-64.so libc.so libm.so + 'abseil-cpp' # libabsl_hash.so libabsl_raw_hash_set.so libabsl_raw_logging_internal.so libabsl_throw_delegate.so + 'openmpi' 'libmpi.so' + 'nsync' # libnsync_cpp.so +# 'protobuf' 'libprotobuf-lite.so' + 'python-coloredlogs' + 'python-flatbuffers' + 'python-numpy' +# 'python-protobuf' + 'python-sympy' + 'python-packaging' + 'python-setuptools' + 'python-requests' + ) + optdepends=( + # https://github.com/microsoft/onnxruntime/pull/9969 + 'python-onnx: for the backend API, quantization, orttraining, transformers and various tools' + 'python-psutil: for transformers' + 'python-py-cpuinfo: for transformers' + 'python-py3nvml: for transformers' + 'python-transformers: for transformers' + 'python-scipy: for transformers and various tools' + 'python-pytorch: for transformers, orttraining and various tools' + 'python-pytorch-cuda' + 'python-cerberus: for orttraining' + 'python-h5py: for orttraining' + 'python-matplotlib' + 'python-tensorflow-opt-cuda' + 'python-importlib-metadata' + ) + + python -m installer --destdir="${pkgdir}" build/dist/*.whl + + _PY_ORT_DIR="$(python -c 'import site; print(site.getsitepackages()[0])')/onnxruntime" + # already installed by `cmake --install`, and not useful as this path is not looked up by the linker + rm -vf "${pkgdir}/${_PY_ORT_DIR}"/capi/libonnxruntime_providers_* + chrpath -d "${pkgdir}/${_PY_ORT_DIR}/capi/onnxruntime_pybind11_state.so" + + install -Ddm755 "${pkgdir}/usr/share/licenses" + ln -s onnxruntime "${pkgdir}/usr/share/licenses/${pkgname}" } -package_python-onnxruntime-cuda() { - depends+=(cuda cudnn nccl python-onnxruntime) +package_onnxruntime-cuda() { pkgdesc+=' (CUDA execution provider)' + depends=( + 'gcc-libs' # libgcc_s.so libstdc++.so + 'glibc' # ld-linux-x86-64.so libc.so libm.so + 'cudnn' # libcudnn.so + 'nccl' # libnccl.so + 'openmpi' 'libmpi.so' + 'nsync' # libnsync_cpp.so + 'abseil-cpp' # libabsl_hash.so libabsl_raw_hash_set.so libabsl_raw_logging_internal.so libabsl_throw_delegate.so + 'cuda' 'libcublas.so' 'libcudart.so' # libcublasLt.so libcufft.so +# 'protobuf' 'libprotobuf-lite.so' + ) + conflicts=('python-onnxruntime-cuda') + replaces=('python-onnxruntime-cuda') + + install -Dm755 build/libonnxruntime_providers_cuda.so -t "${pkgdir}/usr/lib" + + install -Ddm755 "${pkgdir}/usr/share/licenses" + ln -s onnxruntime "${pkgdir}/usr/share/licenses/${pkgname}" +} + +package_onnxruntime-tensorrt() { + pkgdesc+=' (TensorRT execution provider)' + depends=( + 'tensorrt' +# 'protobuf' 'libprotobuf-lite.so' + 'nsync' +# 'flatbuffers' + ) + pkgdesc+=' (TENSORRT execution provider)' + + install -Dm755 build/libonnxruntime_providers_tensorrt.so -t "${pkgdir}/usr/lib" - cd onnxruntime/build - install -Dm755 libonnxruntime_providers_cuda.so -t "$pkgdir"/usr/lib - install -Ddm755 "$pkgdir"/usr/share/licenses - ln -s python-onnxruntime "$pkgdir"/usr/share/licenses/$pkgname + install -Ddm755 "${pkgdir}/usr/share/licenses" + ln -s onnxruntime "${pkgdir}/usr/share/licenses/${pkgname}" } |