[ONNXRuntime] Added builds with CUDA and TensorRT Execution Providers (…

…#4386) * [ONNXRuntime] Added builds with CUDA and TensorRT Execution Providers * Added support for aarch64-linux-gnu - for Jetson 4.6, i.e. CUDA 10.2 * Limited CUDA builds to a single CUDA version per major CUDA version * Enabled lazy artifacts * Moved ONNXRuntime to O/ONNXRuntime * Added comment wrt. CUDA platform expansion wrt. C++ string ABI * Reverted to looping over CUDA versions To support multiple versions in the future.
JuliaPackaging · Sep 12, 2022 · d6a5224 · d6a5224
1 parent 199ab2d
commit d6a5224
Show file tree

Hide file tree

Showing 2 changed files with 132 additions and 73 deletions.
diff --git a/O/ONNXRuntime/ONNXRuntime/build_tarballs.jl b/O/ONNXRuntime/ONNXRuntime/build_tarballs.jl
diff --git a/O/ONNXRuntime/build_tarballs.jl b/O/ONNXRuntime/build_tarballs.jl
@@ -0,0 +1,132 @@
+# Note that this script can accept some limited command-line arguments, run
+# `julia build_tarballs.jl --help` to see a usage message.
+using BinaryBuilder, Pkg
+
+name = "ONNXRuntime"
+version = v"1.10.0"
+
+# Cf. https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#requirements
+# Cf. https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html#requirements
+cuda_versions = [
+    # No CUDA 10.2, since pre-built x86_64 CUDA binaries are built for CUDA 11
+    v"11.3", # Using 11.3, and not 11.4, to be compatible with TensorRT (JLL) v8.0.1 (which includes aarch64 support)
+]
+cuda_aarch64_tag = "10.2"
+cudnn_version = v"8.2.4"
+tensorrt_version = v"8.0.1"
+
+cudnn_compat = string(cudnn_version.major)
+tensorrt_compat = string(tensorrt_version.major)
+
+# Collection of sources required to complete build
+sources = [
+    GitSource("https://github.com/microsoft/onnxruntime.git", "0d9030e79888d1d5828730b254fedc53c7b640c1"),
+    ArchiveSource("https://github.com/microsoft/onnxruntime/releases/download/v$version/onnxruntime-win-x64-$version.zip", "a0c6db3cff65bd282f6ba4a57789e619c27e55203321aa08c023019fe9da50d7"; unpack_target="onnxruntime-x86_64-w64-mingw32"),
+    ArchiveSource("https://github.com/microsoft/onnxruntime/releases/download/v$version/onnxruntime-win-x86-$version.zip", "fd1680fa7248ec334efc2564086e9c5e0d6db78337b55ec32e7b666164bdb88c"; unpack_target="onnxruntime-i686-w64-mingw32"),
+    ArchiveSource("https://github.com/microsoft/onnxruntime/releases/download/v$version/onnxruntime-linux-x64-gpu-$version.tgz", "bc880ba8a572acf79d50dcd35ba6dd8e5fb708d03883959ef60efbc15f5cdcb6"; unpack_target="onnxruntime-x86_64-linux-gnu-cuda"),
+    ArchiveSource("https://github.com/microsoft/onnxruntime/releases/download/v$version/onnxruntime-win-x64-gpu-$version.zip", "0da11b8d953fad4ec75f87bb894f72dea511a3940cff2f4dad37451586d1ebbc"; unpack_target="onnxruntime-x86_64-w64-mingw32-cuda"),
+    # aarch64-linux-gnu binaries for NVIDIA Jetson from NVIDIA-managed Jetson Zoo: https://elinux.org/Jetson_Zoo#ONNX_Runtime
+    FileSource("https://nvidia.box.com/shared/static/jy7nqva7l88mq9i8bw3g3sklzf4kccn2.whl", "a608b7a4a4fc6ad5c90d6005edbfe0851847b991b08aafff4549bbbbdb938bf6"; filename = "onnxruntime-aarch64-linux-gnu-cuda.whl"),
+]
+
+# Bash recipe for building across all platforms
+script = raw"""
+cd $WORKSPACE/srcdir
+
+if [[ $bb_full_target != *cuda*
+    || $bb_full_target == aarch64-linux-gnu*cuda* ]]; then # pre-built aarch64-linux-gnu binaries needs separate libonnxruntime (for CPU).
+    if [[ $target == *-w64-mingw32* ]]; then
+        chmod 755 onnxruntime-$target/onnxruntime-*/lib/*
+        mkdir -p $includedir $libdir
+        cp -av onnxruntime-$target/onnxruntime-*/include/* $includedir
+        cp -av onnxruntime-$target/onnxruntime-*/lib/* $libdir
+        install_license onnxruntime-$target/onnxruntime-*/LICENSE
+    else
+        # Cross-compiling for aarch64-apple-darwin on x86_64 requires setting arch.: https://github.com/microsoft/onnxruntime/blob/v1.10.0/cmake/CMakeLists.txt#L186
+        if [[ $target == aarch64-apple-darwin* ]]; then
+            cmake_extra_args="-DCMAKE_OSX_ARCHITECTURES='arm64'"
+        fi
+
+        cd onnxruntime
+        git submodule update --init --recursive
+        mkdir -p build
+        cd build
+        cmake $WORKSPACE/srcdir/onnxruntime/cmake \
+            -DCMAKE_INSTALL_PREFIX=$prefix \
+            -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TARGET_TOOLCHAIN} \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DONNX_CUSTOM_PROTOC_EXECUTABLE=$host_bindir/protoc \
+            -Donnxruntime_BUILD_SHARED_LIB=ON \
+            -Donnxruntime_BUILD_UNIT_TESTS=OFF \
+            -Donnxruntime_DISABLE_RTTI=OFF \
+            $cmake_extra_args
+        make -j $nproc
+        make install
+        install_license $WORKSPACE/srcdir/onnxruntime/LICENSE
+    fi
+else
+    if [[ $target == *-w64-mingw32* ]]; then
+        chmod 755 onnxruntime-$target-cuda/onnxruntime-*/lib/*
+    fi
+    mkdir -p $includedir $libdir
+    cp -av onnxruntime-$target-cuda/onnxruntime-*/include/* $includedir
+    find onnxruntime-$target-cuda/onnxruntime-*/lib -not -type d | xargs -Isrc cp -av src $libdir
+    install_license onnxruntime-$target-cuda/onnxruntime-*/LICENSE
+fi
+
+if [[ $bb_full_target == aarch64-linux-gnu*cuda* ]]; then
+    cd $WORKSPACE/srcdir
+    unzip -d onnxruntime-$target-cuda onnxruntime-$target-cuda.whl
+    mkdir -p $libdir
+    find onnxruntime-$target-cuda/onnxruntime_gpu*.data/purelib/onnxruntime/capi -name *.so* -not -name *py* | xargs -Isrc cp -av src $libdir
+fi
+"""
+
+# These are the platforms we will build for by default, unless further
+# platforms are passed in on the command line
+function platform_exclude_filter(p::Platform)
+    libc(p) == "musl" ||
+    p == Platform("i686", "Linux") || # No binary - and source build fails linking CXX shared library libonnxruntime.so
+    Sys.isfreebsd(p)
+end
+platforms = supported_platforms(; exclude=platform_exclude_filter)
+platforms = expand_cxxstring_abis(platforms; skip=!Sys.islinux)
+
+cuda_platforms = Platform[]
+for cuda_version in cuda_versions
+    cuda_tag = "$(cuda_version.major).$(cuda_version.minor)"
+    for p in [
+        Platform("x86_64", "Linux"; cuda = cuda_tag),
+        Platform("x86_64", "Windows"; cuda = cuda_tag)
+    ]
+        push!(platforms, p)
+        push!(cuda_platforms, p)
+    end
+end
+push!(platforms, Platform("aarch64", "Linux"; cuda = cuda_aarch64_tag))
+push!(cuda_platforms, Platform("aarch64", "Linux"; cuda = cuda_aarch64_tag))
+
+# Not expanding cuda_platforms wrt. cxxstring ABIs, since pre-built x86_64 CUDA binaries are built for cxx03 (and the tensorrt provider for both x86_64 and aarch64 is built for cxx11)
+
+# The products that we will ensure are always built
+products = [
+    LibraryProduct(["libonnxruntime", "onnxruntime"], :libonnxruntime)
+]
+
+# Dependencies that must be installed before this package can be built
+dependencies = [
+    Dependency("CUDNN_jll", cudnn_version;
+        compat = cudnn_compat,
+        platforms = cuda_platforms),
+    Dependency("TensorRT_jll", tensorrt_version;
+        compat = tensorrt_compat,
+        platforms = cuda_platforms),
+    Dependency("Zlib_jll"; platforms = cuda_platforms),
+    HostBuildDependency(PackageSpec("protoc_jll", Base.UUID("c7845625-083e-5bbe-8504-b32d602b7110"), v"3.16.1"))
+]
+
+# Build the tarballs, and possibly a `build.jl` as well.
+build_tarballs(ARGS, name, version, sources, script, platforms, products, dependencies;
+    lazy_artifacts = true,
+    julia_compat = "1.6",
+    preferred_gcc_version = v"8")