{ lib, stdenv, fetchFromGitHub, rocmUpdateScript, cmake, rocm-cmake, rocblas, rocprim, rocsparse, clr, fmt, gtest, gfortran, lapack-reference, buildTests ? false, buildBenchmarks ? false, gpuTargets ? ( clr.localGpuTargets or [ "gfx900" "gfx906" "gfx908" "gfx90a" "gfx942" "gfx1010" "gfx1030" "gfx1100" "gfx1101" "gfx1102" ] ), }: stdenv.mkDerivation (finalAttrs: { pname = "rocsolver${clr.gpuArchSuffix}"; version = "6.3.3"; outputs = [ "out" ] ++ lib.optionals buildTests [ "test" ] ++ lib.optionals buildBenchmarks [ "benchmark" ]; src = fetchFromGitHub { owner = "ROCm"; repo = "rocSOLVER"; rev = "rocm-${finalAttrs.version}"; hash = "sha256-+sGU+0CB48iolJSyYo+xH36q5LCUp+nKtOYbguzMuhg="; }; nativeBuildInputs = [ cmake # no ninja, it buffers console output and nix times out long periods of no output rocm-cmake clr ] ++ lib.optionals (buildTests || buildBenchmarks) [ gfortran ]; buildInputs = [ # FIXME: rocblas and rocsolver can't build in parallel # but rocsolver doesn't need rocblas' offload builds at build time # could we build against a rocblas-minimal? rocblas rocprim rocsparse fmt ] ++ lib.optionals buildTests [ gtest ] ++ lib.optionals (buildTests || buildBenchmarks) [ lapack-reference ]; cmakeFlags = [ "-DHIP_CLANG_NUM_PARALLEL_JOBS=4" "-DCMAKE_BUILD_TYPE=Release" "-DCMAKE_VERBOSE_MAKEFILE=ON" # Manually define CMAKE_INSTALL_