diff options
Diffstat (limited to 'dev-libs')
| -rw-r--r-- | dev-libs/rccl/files/rccl-10.1.0-llvm-23-gfx942.patch | 19 | ||||
| -rw-r--r-- | dev-libs/rccl/rccl-10.1.0.ebuild | 1 | ||||
| -rw-r--r-- | dev-libs/rocm-opencl-runtime/rocm-opencl-runtime-10.1.0-r1.ebuild (renamed from dev-libs/rocm-opencl-runtime/rocm-opencl-runtime-10.1.0.ebuild) | 6 | ||||
| -rw-r--r-- | dev-libs/roct-thunk-interface/roct-thunk-interface-10.1.0.ebuild | 6 |
4 files changed, 31 insertions, 1 deletions
diff --git a/dev-libs/rccl/files/rccl-10.1.0-llvm-23-gfx942.patch b/dev-libs/rccl/files/rccl-10.1.0-llvm-23-gfx942.patch new file mode 100644 index 000000000000..8c391212936c --- /dev/null +++ b/dev-libs/rccl/files/rccl-10.1.0-llvm-23-gfx942.patch @@ -0,0 +1,19 @@ +Fix build with LLVM 23 for gfx942 and gfx950 + +LLVM 23 fails with "Cannot scavenge register in FI elimination" on some +kernels (fixed by llvm/llvm-project@c003bfba6158); together, disabling the +machine scheduler and the basic SGPR allocator avoid it. + +Bug: https://bugs.gentoo.org/984543 +--- a/cmake/DeviceLinker.cmake ++++ b/cmake/DeviceLinker.cmake +@@ -288,6 +288,9 @@ foreach(DL_GPU_TARGET ${DL_GPU_TARGETS}) + -fPIC + ${DL_INHERITED_FLAGS} + ) ++ if(DL_GPU_TARGET MATCHES "^gfx9(42|50)$") ++ target_compile_options(${_dev_target} PRIVATE -mllvm=-enable-misched=false -mllvm=-sgpr-regalloc=basic) ++ endif() + target_compile_definitions(${_dev_target} PRIVATE RCCL_DEVICE_LINKER) + target_link_libraries(${_dev_target} PRIVATE rccl_device_defs) + diff --git a/dev-libs/rccl/rccl-10.1.0.ebuild b/dev-libs/rccl/rccl-10.1.0.ebuild index 8cae2c180efc..7cb807abb404 100644 --- a/dev-libs/rccl/rccl-10.1.0.ebuild +++ b/dev-libs/rccl/rccl-10.1.0.ebuild @@ -48,6 +48,7 @@ RESTRICT="!test? ( test )" PATCHES=( "${FILESDIR}"/${P}-no-roctx.patch + "${FILESDIR}"/${P}-llvm-23-gfx942.patch ) src_prepare() { diff --git a/dev-libs/rocm-opencl-runtime/rocm-opencl-runtime-10.1.0.ebuild b/dev-libs/rocm-opencl-runtime/rocm-opencl-runtime-10.1.0-r1.ebuild index 797a3cde48e4..2d253cbb8659 100644 --- a/dev-libs/rocm-opencl-runtime/rocm-opencl-runtime-10.1.0.ebuild +++ b/dev-libs/rocm-opencl-runtime/rocm-opencl-runtime-10.1.0-r1.ebuild @@ -66,6 +66,12 @@ src_configure() { append-flags -fno-strict-aliasing filter-lto + use debug && CMAKE_BUILD_TYPE="Debug" + # upstream release builds have assertions disabled; some are reachable + # (e.g. Isa::isCompatible on gfx90c), aborting otherwise working programs + # Bug: https://bugs.gentoo.org/984577 + use debug || append-cppflags -DNDEBUG + # Fix ld.lld linker error: https://github.com/RadeonOpenCompute/ROCm-OpenCL-Runtime/issues/155 append-ldflags $(test-flags-CCLD -Wl,--undefined-version) diff --git a/dev-libs/roct-thunk-interface/roct-thunk-interface-10.1.0.ebuild b/dev-libs/roct-thunk-interface/roct-thunk-interface-10.1.0.ebuild index 8f05341deba2..21abeea777e9 100644 --- a/dev-libs/roct-thunk-interface/roct-thunk-interface-10.1.0.ebuild +++ b/dev-libs/roct-thunk-interface/roct-thunk-interface-10.1.0.ebuild @@ -88,6 +88,10 @@ src_configure() { filter-lto export LIBHSAKMT_PATH="${BUILD_DIR}" + mycmakeargs+=( + -DCMAKE_DISABLE_FIND_PACKAGE_hsakmt=ON + -DHSAKMT_LIBRARY="${BUILD_DIR}/libhsakmt.so" + ) test_wrapper "${S}/tests/kfdtest" cmake_src_configure fi } @@ -125,5 +129,5 @@ src_test() { KFDQMTest.QueueLatency ) local IFS=: - ./run_kfdtest.sh -e "${skipped_tests[*]}" || die + LD_LIBRARY_PATH="${BUILD_DIR}" ./run_kfdtest.sh -e "${skipped_tests[*]}" || die } |