sci-libs/caffe2: add 2.3.0

Signed-off-by: Alfredo Tupone <tupone@gentoo.org>

6 files changed, 374 insertions, 0 deletions

diff --git a/sci-libs/caffe2/Manifest b/sci-libs/caffe2/Manifest
index b72973ca6f63..f1df96691729 100644
--- a/sci-libs/caffe2/Manifest
+++ b/sci-libs/caffe2/Manifest
@@ -1,2 +1,3 @@
 DIST pytorch-2.1.2.tar.gz 116316469 BLAKE2B c5a55ee264bc3477d3556ba6376b5591117e992e56e0dd0c9ba93d12526e2727f7840f6f1e0730a38223b6492c9556840c4ebf22ffd220e97225c2abff303747 SHA512 a8961d78ad785b13c959a0612563a60e0de17a7c8bb9822ddea9a24072796354d07e81c47b6cc8761b21a6448845b088cf80e1661d9e889b0ed5474d3dc76756
 DIST pytorch-2.2.2.tar.gz 116367503 BLAKE2B 0be22f2ec4b9aac6f5e976664cae01facf07929a32565cd57d7cc5b2d9888e9ae71ca301853752fe8f31d174d04c9974eb9ed2f3d452360a50ccf024f200726a SHA512 7990e0f9484038c3458c0bda2c863bf2b19e56edab81fc5938c6e0f08b17558287f853bb67350e8cca8f42bec0f1d4ba0e94e50a145db8da44bdd4bd703d91d0
+DIST pytorch-2.3.0.tar.gz 117029829 BLAKE2B 8f9c0d71ee0a9219b495eddccdcc65107f7ad537c43c68100b229f3d27b0e6c01ccb1659c7fffc356a48d80f2adc0a10361305dc8f1df20446de837d380f89f6 SHA512 67f7e9a096c3ffb952206ebf9105bedebb68c24ad82456083adf1d1d210437fcaa9dd52b68484cfc97d408c9eebc9541c76868c34a7c9982494dc3f424cfb07c
diff --git a/sci-libs/caffe2/caffe2-2.3.0.ebuild b/sci-libs/caffe2/caffe2-2.3.0.ebuild
new file mode 100644
index 000000000000..c69c5e8fdb58
--- /dev/null
+++ b/sci-libs/caffe2/caffe2-2.3.0.ebuild
@@ -0,0 +1,270 @@
+# Copyright 2022-2024 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+PYTHON_COMPAT=( python3_{10..12} )
+ROCM_VERSION=5.7
+inherit python-single-r1 cmake cuda flag-o-matic prefix rocm
+
+MYPN=pytorch
+MYP=${MYPN}-${PV}
+
+DESCRIPTION="A deep learning framework"
+HOMEPAGE="https://pytorch.org/"
+SRC_URI="https://github.com/pytorch/${MYPN}/archive/refs/tags/v${PV}.tar.gz
+	-> ${MYP}.tar.gz"
+
+S="${WORKDIR}"/${MYP}
+
+LICENSE="BSD"
+SLOT="0"
+KEYWORDS="~amd64"
+IUSE="cuda distributed fbgemm ffmpeg gloo mkl mpi nnpack +numpy onednn openblas opencl opencv openmp qnnpack rocm xnnpack"
+RESTRICT="test"
+REQUIRED_USE="
+	${PYTHON_REQUIRED_USE}
+	ffmpeg? ( opencv )
+	mpi? ( distributed )
+	gloo? ( distributed )
+	?? ( cuda rocm )
+	rocm? ( || ( ${ROCM_REQUIRED_USE} ) )
+"
+
+# CUDA 12 not supported yet: https://github.com/pytorch/pytorch/issues/91122
+RDEPEND="
+	${PYTHON_DEPS}
+	dev-cpp/gflags:=
+	>=dev-cpp/glog-0.5.0
+	dev-libs/cpuinfo
+	dev-libs/libfmt
+	dev-libs/protobuf:=
+	dev-libs/pthreadpool
+	dev-libs/sleef
+	virtual/lapack
+	>=sci-libs/onnx-1.12.0
+	<sci-libs/onnx-1.15.0
+	sci-libs/foxi
+	cuda? (
+		=dev-libs/cudnn-8*
+		>=dev-libs/cudnn-frontend-1.0.3:0/8
+		<dev-util/nvidia-cuda-toolkit-12.4.0:=[profiler]
+	)
+	fbgemm? ( >=dev-libs/FBGEMM-2023.12.01 )
+	ffmpeg? ( media-video/ffmpeg:= )
+	gloo? ( sci-libs/gloo[cuda?] )
+	mpi? ( virtual/mpi )
+	nnpack? ( sci-libs/NNPACK )
+	numpy? ( $(python_gen_cond_dep '
+		dev-python/numpy[${PYTHON_USEDEP}]
+		') )
+	onednn? ( dev-libs/oneDNN )
+	opencl? ( virtual/opencl )
+	opencv? ( media-libs/opencv:= )
+	qnnpack? ( sci-libs/QNNPACK )
+	rocm? (
+		>=dev-util/hip-5.7
+		>=dev-libs/rccl-5.7[${ROCM_USEDEP}]
+		>=sci-libs/rocThrust-5.7[${ROCM_USEDEP}]
+		>=sci-libs/rocPRIM-5.7[${ROCM_USEDEP}]
+		>=sci-libs/hipBLAS-5.7[${ROCM_USEDEP}]
+		>=sci-libs/hipFFT-5.7[${ROCM_USEDEP}]
+		>=sci-libs/hipSPARSE-5.7[${ROCM_USEDEP}]
+		>=sci-libs/hipRAND-5.7[${ROCM_USEDEP}]
+		>=sci-libs/hipCUB-5.7[${ROCM_USEDEP}]
+		>=sci-libs/hipSOLVER-5.7[${ROCM_USEDEP}]
+		>=sci-libs/miopen-5.7[${ROCM_USEDEP}]
+		>=dev-util/roctracer-5.7[${ROCM_USEDEP}]
+	)
+	distributed? ( sci-libs/tensorpipe[cuda?] )
+	xnnpack? ( >=sci-libs/XNNPACK-2022.12.22 )
+	mkl? ( sci-libs/mkl )
+	openblas? ( sci-libs/openblas )
+"
+DEPEND="
+	${RDEPEND}
+	cuda? ( >=dev-libs/cutlass-3.4.1 )
+	onednn? ( sci-libs/ideep )
+	dev-libs/psimd
+	dev-libs/FP16
+	dev-libs/FXdiv
+	dev-libs/pocketfft
+	dev-libs/flatbuffers
+	>=sci-libs/kineto-0.4.0_p20231031
+	$(python_gen_cond_dep '
+		dev-python/pyyaml[${PYTHON_USEDEP}]
+		dev-python/pybind11[${PYTHON_USEDEP}]
+	')
+"
+
+PATCHES=(
+	"${FILESDIR}"/${PN}-2.2.1-gentoo.patch
+	"${FILESDIR}"/${PN}-1.13.0-install-dirs.patch
+	"${FILESDIR}"/${PN}-1.12.0-glog-0.6.0.patch
+	"${FILESDIR}"/${PN}-1.13.1-tensorpipe.patch
+	"${FILESDIR}"/${P}-cudnn_include_fix.patch
+	"${FILESDIR}"/${PN}-2.1.2-fix-rpath.patch
+	"${FILESDIR}"/${PN}-2.1.2-fix-openmp-link.patch
+	"${FILESDIR}"/${P}-rocm-fix-std-cpp17.patch
+	"${FILESDIR}"/${PN}-2.2.2-musl.patch
+	"${FILESDIR}"/${P}-CMakeFix.patch
+)
+
+src_prepare() {
+	filter-lto #bug 862672
+	sed -i \
+		-e "/third_party\/gloo/d" \
+		cmake/Dependencies.cmake \
+		|| die
+	cmake_src_prepare
+	pushd torch/csrc/jit/serialization || die
+	flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
+	popd
+	# prefixify the hardcoded paths, after all patches are applied
+	hprefixify \
+		aten/CMakeLists.txt \
+		caffe2/CMakeLists.txt \
+		cmake/Metal.cmake \
+		cmake/Modules/*.cmake \
+		cmake/Modules_CUDA_fix/FindCUDNN.cmake \
+		cmake/Modules_CUDA_fix/upstream/FindCUDA/make2cmake.cmake \
+		cmake/Modules_CUDA_fix/upstream/FindPackageHandleStandardArgs.cmake \
+		cmake/public/LoadHIP.cmake \
+		cmake/public/cuda.cmake \
+		cmake/Dependencies.cmake \
+		torch/CMakeLists.txt \
+		CMakeLists.txt
+
+	if use rocm; then
+		sed -e "s:/opt/rocm:/usr:" \
+			-e "s:lib/cmake:$(get_libdir)/cmake:g" \
+			-e "s/HIP 1.0/HIP 1.0 REQUIRED/" \
+			-i cmake/public/LoadHIP.cmake || die
+
+		ebegin "HIPifying cuda sources"
+		${EPYTHON} tools/amd_build/build_amd.py || die
+		eend $?
+	fi
+}
+
+src_configure() {
+	if use cuda && [[ -z ${TORCH_CUDA_ARCH_LIST} ]]; then
+		ewarn "WARNING: caffe2 is being built with its default CUDA compute capabilities: 3.5 and 7.0."
+		ewarn "These may not be optimal for your GPU."
+		ewarn ""
+		ewarn "To configure caffe2 with the CUDA compute capability that is optimal for your GPU,"
+		ewarn "set TORCH_CUDA_ARCH_LIST in your make.conf, and re-emerge caffe2."
+		ewarn "For example, to use CUDA capability 7.5 & 3.5, add: TORCH_CUDA_ARCH_LIST=7.5 3.5"
+		ewarn "For a Maxwell model GPU, an example value would be: TORCH_CUDA_ARCH_LIST=Maxwell"
+		ewarn ""
+		ewarn "You can look up your GPU's CUDA compute capability at https://developer.nvidia.com/cuda-gpus"
+		ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery | grep 'CUDA Capability'"
+	fi
+
+	local mycmakeargs=(
+		-DBUILD_CUSTOM_PROTOBUF=OFF
+		-DBUILD_SHARED_LIBS=ON
+
+		-DUSE_CCACHE=OFF
+		-DUSE_CUDA=$(usex cuda)
+		-DUSE_DISTRIBUTED=$(usex distributed)
+		-DUSE_MPI=$(usex mpi)
+		-DUSE_FAKELOWP=OFF
+		-DUSE_FBGEMM=$(usex fbgemm)
+		-DUSE_FFMPEG=$(usex ffmpeg)
+		-DUSE_GFLAGS=ON
+		-DUSE_GLOG=ON
+		-DUSE_GLOO=$(usex gloo)
+		-DUSE_KINETO=OFF # TODO
+		-DUSE_LEVELDB=OFF
+		-DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
+		-DUSE_MKLDNN=$(usex onednn)
+		-DUSE_NNPACK=$(usex nnpack)
+		-DUSE_QNNPACK=$(usex qnnpack)
+		-DUSE_XNNPACK=$(usex xnnpack)
+		-DUSE_SYSTEM_XNNPACK=$(usex xnnpack)
+		-DUSE_TENSORPIPE=$(usex distributed)
+		-DUSE_PYTORCH_QNNPACK=OFF
+		-DUSE_NUMPY=$(usex numpy)
+		-DUSE_OPENCL=$(usex opencl)
+		-DUSE_OPENCV=$(usex opencv)
+		-DUSE_OPENMP=$(usex openmp)
+		-DUSE_ROCM=$(usex rocm)
+		-DUSE_SYSTEM_CPUINFO=ON
+		-DUSE_SYSTEM_PYBIND11=ON
+		-DUSE_UCC=OFF
+		-DUSE_VALGRIND=OFF
+		-DPYBIND11_PYTHON_VERSION="${EPYTHON#python}"
+		-DPYTHON_EXECUTABLE="${PYTHON}"
+		-DUSE_ITT=OFF
+		-DUSE_SYSTEM_PTHREADPOOL=ON
+		-DUSE_SYSTEM_FXDIV=ON
+		-DUSE_SYSTEM_FP16=ON
+		-DUSE_SYSTEM_GLOO=ON
+		-DUSE_SYSTEM_ONNX=ON
+		-DUSE_SYSTEM_SLEEF=ON
+		-DUSE_METAL=OFF
+
+		-Wno-dev
+		-DTORCH_INSTALL_LIB_DIR="${EPREFIX}"/usr/$(get_libdir)
+		-DLIBSHM_INSTALL_LIB_SUBDIR="${EPREFIX}"/usr/$(get_libdir)
+	)
+
+	if use mkl; then
+		mycmakeargs+=(-DBLAS=MKL)
+	elif use openblas; then
+		mycmakeargs+=(-DBLAS=OpenBLAS)
+	else
+		mycmakeargs+=(-DBLAS=Generic -DBLAS_LIBRARIES=)
+	fi
+
+	if use cuda; then
+		addpredict "/dev/nvidiactl" # bug 867706
+		addpredict "/dev/char"
+		addpredict "/proc/self/task" # bug 926116
+
+		mycmakeargs+=(
+			-DUSE_CUDNN=ON
+			-DTORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-3.5 7.0}"
+			-DUSE_NCCL=OFF # TODO: NVIDIA Collective Communication Library
+			-DCMAKE_CUDA_FLAGS="$(cuda_gccdir -f | tr -d \")"
+		)
+	elif use rocm; then
+		export PYTORCH_ROCM_ARCH="$(get_amdgpu_flags)"
+
+		mycmakeargs+=(
+			-DUSE_NCCL=ON
+			-DUSE_SYSTEM_NCCL=ON
+		)
+	fi
+
+	if use onednn; then
+		mycmakeargs+=(
+			-DUSE_MKLDNN=ON
+			-DMKLDNN_FOUND=ON
+			-DMKLDNN_LIBRARIES=dnnl
+			-DMKLDNN_INCLUDE_DIR="${ESYSROOT}/usr/include/oneapi/dnnl"
+		)
+	fi
+
+	cmake_src_configure
+
+	# do not rerun cmake and the build process in src_install
+	sed '/RERUN/,+1d' -i "${BUILD_DIR}"/build.ninja || die
+}
+
+src_install() {
+	cmake_src_install
+
+	insinto "/var/lib/${PN}"
+	doins "${BUILD_DIR}"/CMakeCache.txt
+
+	rm -rf python
+	mkdir -p python/torch/include || die
+	mv "${ED}"/usr/lib/python*/site-packages/caffe2 python/ || die
+	cp torch/version.py python/torch/ || die
+	python_domodule python/caffe2
+	python_domodule python/torch
+	ln -s ../../../../../include/torch \
+		"${D}$(python_get_sitedir)"/torch/include/torch || die # bug 923269
+}
diff --git a/sci-libs/caffe2/files/caffe2-2.2.2-musl.patch b/sci-libs/caffe2/files/caffe2-2.2.2-musl.patch
new file mode 100644
index 000000000000..f63e9f1df332
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.2.2-musl.patch
@@ -0,0 +1,13 @@
+--- a/torch/csrc/profiler/unwind/unwind.cpp	2024-04-29 12:05:40.895667482 +0200
++++ b/torch/csrc/profiler/unwind/unwind.cpp	2024-04-29 12:05:53.099524760 +0200
+@@ -112,8 +112,8 @@
+ }
+ 
+ struct Version {
+-  uint64_t adds_ = LONG_LONG_MAX;
+-  uint64_t subs_ = LONG_LONG_MAX;
++  uint64_t adds_ = LLONG_MAX;
++  uint64_t subs_ = LLONG_MAX;
+ };
+ 
+ struct UnwindCache {
diff --git a/sci-libs/caffe2/files/caffe2-2.3.0-CMakeFix.patch b/sci-libs/caffe2/files/caffe2-2.3.0-CMakeFix.patch
new file mode 100644
index 000000000000..eba37d933cac
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.3.0-CMakeFix.patch
@@ -0,0 +1,11 @@
+--- a/CMakeLists.txt	2024-04-29 20:32:26.259716769 +0200
++++ b/CMakeLists.txt	2024-04-29 20:32:35.886384618 +0200
+@@ -50,7 +50,7 @@
+ 
+ # This define is needed to preserve behavior given anticpated changes to cccl/thrust
+ # https://nvidia.github.io/libcudacxx/standard_api/numerics_library/complex.html
+-string(APPEND CMAKE_CUDA_FLAGS "-DLIBCUDACXX_ENABLE_SIMPLIFIED_COMPLEX_OPERATIONS")
++string(APPEND CMAKE_CUDA_FLAGS " -DLIBCUDACXX_ENABLE_SIMPLIFIED_COMPLEX_OPERATIONS")
+ 
+ if(LINUX)
+   include(cmake/CheckAbi.cmake)
diff --git a/sci-libs/caffe2/files/caffe2-2.3.0-cudnn_include_fix.patch b/sci-libs/caffe2/files/caffe2-2.3.0-cudnn_include_fix.patch
new file mode 100644
index 000000000000..77905dbd1ac8
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.3.0-cudnn_include_fix.patch
@@ -0,0 +1,11 @@
+--- a/cmake/Dependencies.cmake	2024-04-29 18:37:34.005639858 +0200
++++ b/cmake/Dependencies.cmake	2024-04-29 18:39:29.126587738 +0200
+@@ -1235,7 +1235,7 @@
+   if(CUDNN_VERSION VERSION_LESS 8.5)
+     message(FATAL_ERROR "PyTorch needs CuDNN-8.5 or above, but found ${CUDNN_VERSION}. Builds are still possible with `USE_CUDNN=0`")
+   endif()
+-  set(CUDNN_FRONTEND_INCLUDE_DIR ${CMAKE_CURRENT_LIST_DIR}/../third_party/cudnn_frontend/include)
++  set(CUDNN_FRONTEND_INCLUDE_DIR /opt/cuda/include)
+   target_include_directories(torch::cudnn INTERFACE ${CUDNN_FRONTEND_INCLUDE_DIR})
+ endif()
+ 
diff --git a/sci-libs/caffe2/files/caffe2-2.3.0-rocm-fix-std-cpp17.patch b/sci-libs/caffe2/files/caffe2-2.3.0-rocm-fix-std-cpp17.patch
new file mode 100644
index 000000000000..127a31e4b225
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.3.0-rocm-fix-std-cpp17.patch
@@ -0,0 +1,68 @@
+Fix for error: invalid argument '-std=c++17' not allowed with 'C'
+https://github.com/pytorch/pytorch/issues/103222
+--- a/c10/hip/CMakeLists.txt
++++ b/c10/hip/CMakeLists.txt
+@@ -30,6 +30,7 @@ hip_add_library(c10_hip ${C10_HIP_SRCS} ${C10_HIP_HEADERS})
+ 
+ # Propagate HIP_CXX_FLAGS that were set from Dependencies.cmake
+ target_compile_options(c10_hip PRIVATE ${HIP_CXX_FLAGS})
++set_target_properties(c10_hip PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+ 
+ # caffe2_hip adds a bunch of dependencies like rocsparse, but c10/hip is supposed to be
+ # minimal.  I'm not sure if we need hip_hcc or not; for now leave it out
+--- a/caffe2/CMakeLists.txt
++++ b/caffe2/CMakeLists.txt
+@@ -1712,6 +1712,7 @@ if(USE_ROCM)
+ 
+   # Since PyTorch files contain HIP headers, these flags are required for the necessary definitions to be added.
+   target_compile_options(torch_hip PUBLIC ${HIP_CXX_FLAGS})  # experiment
++  set_target_properties(torch_hip PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+   target_link_libraries(torch_hip PUBLIC c10_hip)
+ 
+   if(NOT INTERN_BUILD_MOBILE)
+@@ -1908,6 +1909,7 @@ if(BUILD_TEST)
+       target_include_directories(${test_name} PRIVATE $<INSTALL_INTERFACE:include>)
+       target_include_directories(${test_name} PRIVATE ${Caffe2_CPU_INCLUDE} ${Caffe2_HIP_INCLUDE})
+       target_compile_options(${test_name} PRIVATE ${HIP_CXX_FLAGS})
++      set_target_properties(${test_name} PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+       add_test(NAME ${test_name} COMMAND $<TARGET_FILE:${test_name}>)
+       if(INSTALL_TEST)
+         install(TARGETS ${test_name} DESTINATION test)
+@@ -2092,6 +2094,7 @@ if(BUILD_PYTHON)
+     endif()
+     if(NOT MSVC)
+       target_compile_options(caffe2_pybind11_state_hip PRIVATE ${HIP_CXX_FLAGS} -fvisibility=hidden)
++      set_target_properties(caffe2_pybind11_state_hip PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+     endif()
+     set_target_properties(caffe2_pybind11_state_hip PROPERTIES PREFIX "")
+     set_target_properties(caffe2_pybind11_state_hip PROPERTIES SUFFIX ${PY_EXT_SUFFIX})
+--- a/cmake/Dependencies.cmake
++++ b/cmake/Dependencies.cmake
+@@ -1278,7 +1278,6 @@
+     list(APPEND HIP_CXX_FLAGS -Wno-duplicate-decl-specifier)
+     list(APPEND HIP_CXX_FLAGS -DCAFFE2_USE_MIOPEN)
+     list(APPEND HIP_CXX_FLAGS -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_HIP)
+-    list(APPEND HIP_CXX_FLAGS -std=c++17)
+     if(ROCM_VERSION_DEV VERSION_GREATER_EQUAL "6.0.0")
+       list(APPEND HIP_CXX_FLAGS -DHIPBLAS_V2)
+     endif()
+--- a/cmake/public/utils.cmake
++++ b/cmake/public/utils.cmake
+@@ -335,6 +335,7 @@ function(caffe2_hip_binary_target target_name_or_src)
+   caffe2_binary_target(${target_name_or_src})
+ 
+   target_compile_options(${__target} PRIVATE ${HIP_CXX_FLAGS})
++  set_target_properties(${__target} PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+   target_include_directories(${__target} PRIVATE ${Caffe2_HIP_INCLUDE})
+ endfunction()
+ 
+--- a/modules/detectron/CMakeLists.txt
++++ b/modules/detectron/CMakeLists.txt
+@@ -31,6 +31,7 @@ if(BUILD_CAFFE2_OPS)
+         ${Detectron_CPU_SRCS}
+         ${Detectron_HIP_SRCS})
+     target_compile_options(caffe2_detectron_ops_hip PRIVATE ${HIP_CXX_FLAGS})
++    set_target_properties(caffe2_detectron_ops_hip PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+     if(USE_MKLDNN)
+       target_link_libraries(caffe2_detectron_ops_hip PRIVATE caffe2::mkldnn)
+     endif()