{ stdenv, lib, fetchFromGitHub, cmake }: # This was originally called mkl-dnn, then it was renamed to dnnl, and it has # just recently been renamed again to oneDNN. See here for details: # https://github.com/oneapi-src/oneDNN#oneapi-deep-neural-network-library-onednn stdenv.mkDerivation rec { pname = "oneDNN"; version = "1.5.1"; src = fetchFromGitHub { owner = "oneapi-src"; repo = "oneDNN"; rev = "v${version}"; sha256 = "1l66gkidldjpznp8pb01wdgrmm0rmrbndv8lzidz8fp9hf473zgl"; }; outputs = [ "out" "dev" "doc" ]; nativeBuildInputs = [ cmake ]; doCheck = true; cmakeFlags = [ # oneDNN compiles with -msse4.1 by default, but not all x86_64 # CPUs support SSE 4.1. "-DDNNL_ARCH_OPT_FLAGS=" ]; # The test driver doesn't add an RPath to the build libdir preCheck = '' export LD_LIBRARY_PATH=$LD_LIBRARY_PATH''${LD_LIBRARY_PATH:+:}$PWD/src export DYLD_LIBRARY_PATH=$DYLD_LIBRARY_PATH''${DYLD_LIBRARY_PATH:+:}$PWD/src '' + lib.optionalString stdenv.isx86_64 '' # Use baseline SIMD in case CPU features get misdetected. export DNNL_MAX_CPU_ISA=SSE41 ''; # The cmake install gets tripped up and installs a nix tree into $out, in # addition to the correct install; clean it up. postInstall = '' rm -r $out/nix ''; meta = with lib; { description = "oneAPI Deep Neural Network Library (oneDNN)"; homepage = "https://01.org/oneDNN"; changelog = "https://github.com/oneapi-src/oneDNN/releases/tag/v${version}"; license = licenses.asl20; platforms = platforms.all; maintainers = with maintainers; [ alexarice bhipple ]; }; }