Merge pull request #9046 from fwyzard/IB/CMSSW_14_1_X/master_cuda-12.4.0

Update to CUDA 12.4.0
cms-sw · Mar 13, 2024 · bd55ddd · bd55ddd
2 parents 54a5f9c + a00c4c1
commit bd55ddd
Show file tree

Hide file tree

Showing 12 changed files with 41 additions and 29 deletions.
diff --git a/celeritas.spec b/celeritas.spec
@@ -9,6 +9,7 @@ Source: git+https://github.com/celeritas-project/celeritas?obj=develop/%{tag}&ex
 BuildRequires: cmake
 
 %define build_flags -Wall -Wextra -pedantic %{?arch_build_flags} %{?lto_build_flags} %{?pgo_build_flags}
+Requires: python3
 Requires: json
 Requires: geant4
 Requires: vecgeom

diff --git a/cmake.spec b/cmake.spec
@@ -1,4 +1,4 @@
-### RPM external cmake 3.25.2
+### RPM external cmake 3.28.3
 %define downloaddir %(echo %realversion | cut -d. -f1,2)
 Source: http://www.cmake.org/files/v%{downloaddir}/%n-%realversion.tar.gz
 Requires: bz2lib curl expat zlib
@@ -20,6 +20,7 @@ cat > build-flags.cmake <<- EOF
 	SET(BUILD_CursesDialog FALSE CACHE BOOL "" FORCE)
 
 	# Use system libraries, not cmake bundled ones.
+	SET(CMAKE_USE_OPENSSL TRUE CACHE BOOL "" FORCE)
 	SET(CMAKE_USE_SYSTEM_LIBRARY_CURL TRUE CACHE BOOL "" FORCE)
 	SET(CMAKE_USE_SYSTEM_LIBRARY_ZLIB TRUE CACHE BOOL "" FORCE)
 	SET(CMAKE_USE_SYSTEM_LIBRARY_BZIP2 TRUE CACHE BOOL "" FORCE)

diff --git a/cuda-compatible-runtime.spec b/cuda-compatible-runtime.spec
@@ -11,11 +11,11 @@ Requires: cuda
 
 %build
 ## INCLUDE cuda-flags
-# defines nvcc_stdcxx and cuda_flags_4
+# defines nvcc_flags_stdcxx and nvcc_flags_cuda_archs
 
 rm -rf %{_builddir}/build && mkdir %{_builddir}/build
 if
-  $CUDA_ROOT/bin/nvcc %{nvcc_stdcxx} -O2 -g %{cuda_flags_4} test.cu -I $CUDA_ROOT/include -L $CUDA_ROOT/lib64 -L $CUDA_ROOT/lib64/stubs --cudart static -ldl -lrt --compiler-options '-Wall -pthread' -o %{_builddir}/build/cuda-compatible-runtime
+  $CUDA_ROOT/bin/nvcc %{nvcc_flags_stdcxx} -O2 -g %{nvcc_flags_cuda_archs} test.cu -I $CUDA_ROOT/include -L $CUDA_ROOT/lib64 -L $CUDA_ROOT/lib64/stubs --cudart static -ldl -lrt --compiler-options '-Wall -pthread' -o %{_builddir}/build/cuda-compatible-runtime
 then
   true
 else

diff --git a/cuda-flags.file b/cuda-flags.file
@@ -5,32 +5,39 @@
 # build support for Pascal (6.x), Volta (7.0), Turing (7.5), Ampere (8.x) and Lovelace (8.9)
 %define cuda_arch 60 70 75 80 89
 
+# LIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES style for listing the supported CUDA compute architectures
+%define omptarget_cuda_archs %(echo $(for ARCH in %cuda_arch; do echo "$ARCH"; done) | sed -e"s/ /,/g")
+
 # LLVM style for listing the supported CUDA compute architectures
-%define llvm_cuda_arch %(echo $(for ARCH in %cuda_arch; do echo "$ARCH"; done) | sed -e"s/ /,/g")
+%define llvm_flags_cuda_archs %(echo $(for ARCH in %cuda_arch; do echo "--offload-arch=sm_$ARCH"; done)) -Wunknown-cuda-version
 
 # C++ standard to use for building host and device code with nvcc
-%define nvcc_stdcxx -std=c++%{cms_cxx_standard}
+%define nvcc_flags_stdcxx -std=c++%{cms_cxx_standard}
 
 # generate optimised code
-%define cuda_flags_0 -O3
+%define nvcc_flags_opt -O3
 
 # generate debugging information for device code
-%define cuda_flags_1 --generate-line-info --source-in-ptx --display-error-number
+%define nvcc_flags_debug --generate-line-info --source-in-ptx --display-error-number
 
 # imply __host__, __device__ attributes in constexpr functions
-%define cuda_flags_2 --expt-relaxed-constexpr
+%define nvcc_flags_constexpr --expt-relaxed-constexpr
 
 # allow __host__, __device__ attributes in lambda declaration
-%define cuda_flags_3 --extended-lambda
+%define nvcc_flags_lambda --extended-lambda
 
 # build support for the various compute architectures
-%define cuda_flags_4 %(echo $(for ARCH in %cuda_arch; do echo "-gencode arch=compute_$ARCH,code=[sm_$ARCH,compute_$ARCH]"; done)) -Wno-deprecated-gpu-targets
+%define nvcc_flags_cuda_archs %(echo $(for ARCH in %cuda_arch; do echo "-gencode arch=compute_$ARCH,code=[sm_$ARCH,compute_$ARCH]"; done)) -Wno-deprecated-gpu-targets
 
 # disable warnings about attributes on defaulted methods
-%define cuda_flags_5 -Xcudafe --diag_suppress=esa_on_defaulted_function_ignored
+%define nvcc_flags_cudafe_diag -Xcudafe --diag_suppress=esa_on_defaulted_function_ignored
+
+# override the version of GCC passed to cudafe++
+%define override_gnu_version $(gcc -dumpfullversion | { IFS=.; read MAJOR MINOR PATCH; echo $(((MAJOR * 100 + MINOR) * 100)); })
+%define nvcc_flags_gnu_version -Xcudafe --gnu_version=%{override_gnu_version}
 
 # link the CUDA runtime shared library
-%define cuda_flags_6 --cudart shared
+%define nvcc_flags_cudart --cudart shared
 
 # collect all CUDA flags
-%define nvcc_cuda_flags %{nvcc_stdcxx} %{cuda_flags_0} %{cuda_flags_1} %{cuda_flags_2} %{cuda_flags_3} %{cuda_flags_4} %{cuda_flags_5} %{cuda_flags_6}
+%define nvcc_cuda_flags %{nvcc_flags_stdcxx} %{nvcc_flags_opt} %{nvcc_flags_debug} %{nvcc_flags_constexpr} %{nvcc_flags_lambda} %{nvcc_flags_cuda_archs} %{nvcc_flags_cudafe_diag} %{nvcc_flags_gnu_version} %{nvcc_flags_cudart}
diff --git a/cuda.spec b/cuda.spec
@@ -1,8 +1,8 @@
-### RPM external cuda 12.2.1
+### RPM external cuda 12.4.0
 ## INITENV +PATH LD_LIBRARY_PATH %i/lib64
 
 %define runpath_opts -m compute-sanitizer -m drivers -m nvvm
-%define driversversion 535.86.10
+%define driversversion 550.54.14
 
 %ifarch x86_64
 Source0: https://developer.download.nvidia.com/compute/cuda/%{realversion}/local_installers/%{n}_%{realversion}_%{driversversion}_linux.run

diff --git a/cudnn.spec b/cudnn.spec
@@ -1,4 +1,4 @@
-### RPM external cudnn 8.8.0.121
+### RPM external cudnn 8.9.7.29
 ## INITENV +PATH LD_LIBRARY_PATH %i/lib64
 
 %define cudaver 12

diff --git a/gdrcopy.spec b/gdrcopy.spec
@@ -1,4 +1,4 @@
-### RPM external gdrcopy 2.3
+### RPM external gdrcopy 2.4.1
 ## INITENV +PATH LD_LIBRARY_PATH %i/lib64
 Source: https://github.com/NVIDIA/%{n}/archive/v%{realversion}.tar.gz
 Requires: cuda

diff --git a/geant4.spec b/geant4.spec
@@ -2,7 +2,7 @@
 ## INCLUDE compilation_flags
 ## INCLUDE compilation_flags_lto
 ## INCLUDE cpp-standard
-%define tag 2d5992376773d172321f01710a82943ead5bb435
+%define tag 0f1b33cb3d004830d88b4f9fab2c12b31ed69164
 %define branch cms/v%{realversion}
 %define github_user cms-externals
 Source: git+https://github.com/%github_user/%{n}.git?obj=%{branch}/%{tag}&export=%{n}.%{realversion}&output=/%{n}.%{realversion}-%{tag}.tgz

diff --git a/llvm.spec b/llvm.spec
@@ -29,7 +29,7 @@ sed -ibak '/add_clang_subdirectory(libclang)/a add_subdirectory(include-what-you
 
 %build
 ## INCLUDE cuda-flags
-# defines llvm_cuda_arch
+# defines omptarget_cuda_archs
 
 rm -rf %{_builddir}/build
 mkdir -p %{_builddir}/build
@@ -51,7 +51,7 @@ cmake %{_builddir}/llvm-%{realversion}-%{llvmCommit}/llvm \
   -DLLVM_HOST_TRIPLE=$(gcc -dumpmachine) \
   -DLLVM_TARGETS_TO_BUILD:STRING="X86;PowerPC;AArch64;NVPTX" \
   -DLIBOMPTARGET_NVPTX_ALTERNATE_HOST_COMPILER=/usr/bin/gcc \
-  -DLIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES="%llvm_cuda_arch" \
+  -DLIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES="%omptarget_cuda_archs" \
   -DCMAKE_REQUIRED_INCLUDES="${ZLIB_ROOT}/include" \
   -DCMAKE_PREFIX_PATH="${ZLIB_ROOT}"
 

diff --git a/onnxruntime.spec b/onnxruntime.spec
@@ -1,9 +1,9 @@
-### RPM external onnxruntime 1.14.1
+### RPM external onnxruntime 1.17.1
 ## INITENV +PATH PYTHON3PATH %{i}/${PYTHON3_LIB_SITE_PACKAGES}
 ## INCLUDE cuda-flags
 %define github_user cms-externals
 %define branch cms/v%{realversion}
-%define tag e4c6aa2984c7c71409f4c6d0db865117afa66932
+%define tag 5353412812a6f0409a15faff9a18640456ab1ff8
 Source: git+https://github.com/%{github_user}/%{n}.git?obj=%{branch}/%{tag}&export=%{n}-%{realversion}&submodules=1&output=/%{n}-%{realversion}.tgz
 
 BuildRequires: cmake ninja
@@ -35,6 +35,7 @@ cmake ../%{n}-%{realversion}/cmake -GNinja \
    -Donnxruntime_USE_CUDA=${USE_CUDA} \
    -Donnxruntime_CUDA_HOME="${CUDA_ROOT}" \
    -Donnxruntime_CUDNN_HOME="${CUDNN_ROOT}" \
+   -Donnxruntime_NVCC_THREADS=0 \
    -Donnxruntime_BUILD_CSHARP=OFF \
    -Donnxruntime_USE_OPENMP=OFF \
    -Donnxruntime_USE_TVM=OFF \
@@ -53,10 +54,12 @@ cmake ../%{n}-%{realversion}/cmake -GNinja \
    -DCMAKE_TRY_COMPILE_PLATFORM_VARIABLES="CMAKE_CUDA_RUNTIME_LIBRARY" \
    -DCMAKE_PREFIX_PATH="${ZLIB_ROOT};${LIBPNG_ROOT};${PROTOBUF_ROOT};${PY3_PYBIND11_ROOT};${RE2_ROOT}" \
    -DRE2_INCLUDE_DIR="${RE2_ROOT}/include" \
-   -DCMAKE_CXX_FLAGS="-Wno-error=stringop-overflow"
+   -DCMAKE_CXX_FLAGS="-Wno-error=stringop-overflow -Wno-error=maybe-uninitialized"
 
-# False positive string overflow
-# https://github.com/google/flatbuffers/issues/7366
+# -Wno-error=stringop-overflow is needed to work around a false positive string overflow,
+# see https://github.com/google/flatbuffers/issues/7366
+
+# -Wno-error=maybe-uninitialized is needed for ONNX runtime 1.17.1 with cuDNN 8.9 or 9.0
 
 ninja -v %{makeprocesses}
 python3 ../%{n}-%{realversion}/setup.py build

diff --git a/pytorch.spec b/pytorch.spec
@@ -2,10 +2,10 @@
 ## INCLUDE cuda-flags
 
 %define cuda_arch_float $(echo %{cuda_arch} | tr ' ' '\\n' | sed -E 's|([0-9])$|.\\1|' | tr '\\n' ' ')
-%define tag v%{realversion}
-%define branch release/2.1
+%define tag bb938bbe9f53414dda1e1159795b7536dbffd041
+%define branch cms/v%{realversion}
 
-Source: git+https://github.com/pytorch/pytorch.git?obj=%{branch}/%{tag}&export=%{n}-%{realversion}&submodules=1&output=/%{n}-%{realversion}.tgz
+Source: git+https://github.com/cms-externals/pytorch.git?obj=%{branch}/%{tag}&export=%{n}-%{realversion}&submodules=1&output=/%{n}-%{realversion}.tgz
 Source1: FindEigen3.cmake
 Source2: FindFMT.cmake
 Source99: scram-tools.file/tools/eigen/env

diff --git a/scram-tools.file/tool-env.file b/scram-tools.file/tool-env.file
@@ -13,7 +13,7 @@ export COMPILER_CXXFLAGS="%{arch_build_flags}"
 export COMPILER_WARNINGS="%{?warning_flags}"
 export ORACLE_ENV_ROOT=""
 export CUDA_FLAGS="%{nvcc_cuda_flags}"
-export CUDA_HOST_CXXFLAGS="%{nvcc_stdcxx}"
+export CUDA_HOST_CXXFLAGS="%{nvcc_flags_stdcxx}"
 export LTO_FLAGS="%{?lto_build_flags}"
 export PGO_FLAGS="%{?pgo_build_flags}"
 if [ $(echo ' %requiredtools ' | grep ' python3 ' |wc -l) -gt 0 ] ; then export PYTHON3_LIB_SITE_PACKAGES ; fi