Add TACUDA package

overlay.nix

@@ -58,6 +58,7 @@ let
     tagaspi = callPackage ./pkgs/tagaspi/default.nix { };
     tampi = callPackage ./pkgs/tampi/default.nix { };
     upc-qaire-exporter = prev.callPackage ./pkgs/upc-qaire-exporter/default.nix { };
+    tacuda = callPackage ./pkgs/tacuda/default.nix { };
     wxparaver = callPackage ./pkgs/paraver/default.nix { };
   };
 

pkgs/extrae/default.nix

@@ -20,7 +20,6 @@
 #, python3Packages
 , installShellFiles
 , symlinkJoin
-, enablePapi ? stdenv.hostPlatform == stdenv.buildPlatform # Disabled when cross-compiling
 }:
 
 let
@@ -88,7 +87,7 @@ stdenv.mkDerivation rec {
       --enable-sampling
       --with-unwind=${libunwind.dev}
       --with-xml-prefix=${libxml2.dev}
-      ${lib.optionalString enablePapi "--with-papi=${papi}"}
+      --with-papi=${papi}
       ${if (mpi != null) then ''--with-mpi=${mpi}''
         else ''--without-mpi''}
       --without-dyninst)

pkgs/lmbench/default.nix

@@ -22,11 +22,7 @@ stdenv.mkDerivation rec {
   '';
 
   buildInputs = [ libtirpc ];
-  patches = [
-    ./fix-install.patch
-    ./gcc-14.patch
-    ./fix-cross.patch
-  ];
+  patches = [ ./fix-install.patch ./gcc-14.patch ];
 
   hardeningDisable = [ "all" ];
 
@@ -39,8 +35,6 @@ stdenv.mkDerivation rec {
       CFLAGS=-Wno-implicit-int
       CPPFLAGS=-I${libtirpc.dev}/include/tirpc
       LDFLAGS=-ltirpc
-      CC=${stdenv.cc.targetPrefix}cc
-      AR=${stdenv.cc.targetPrefix}ar
     )
   '';
 

pkgs/lmbench/fix-cross.patch

@@ -1,13 +0,0 @@
-diff --git a/src/Makefile b/src/Makefile
-index 2555014..356eeb3 100644
---- a/src/Makefile
-+++ b/src/Makefile
-@@ -36,7 +36,7 @@ SHELL=/bin/sh
- 
- CC=`../scripts/compiler`
- MAKE=`../scripts/make`
--AR=ar
-+AR?=ar
- ARCREATE=cr
- 
- # base of installation location

pkgs/nanos6/default.nix

@@ -16,7 +16,6 @@
 , jemallocNanos6 ? null
 , cachelineBytes ? 64
 , enableGlibcxxDebug ? false
-, enablePapi ? stdenv.hostPlatform == stdenv.buildPlatform # Disabled when cross-compiling
 , useGit ? false
 , gitUrl ? "ssh://git@bscpm04.bsc.es/nanos6/nanos6"
 , gitBranch ? "master"
@@ -48,8 +47,6 @@ let
   };
 
   source = if (useGit) then git else release;
-
-  isCross = stdenv.hostPlatform != stdenv.buildPlatform;
 in
   stdenv.mkDerivation (source // {
     pname = "nanos6";
@@ -74,11 +71,9 @@ in
       "--disable-all-instrumentations"
       "--enable-ovni-instrumentation"
       "--with-ovni=${ovni}"
-      "--with-boost=${boost.dev}"
     ] ++
       (optional enableJemalloc "--with-jemalloc=${jemallocNanos6}") ++
-      (optional enableGlibcxxDebug "CXXFLAGS=-D_GLIBCXX_DEBUG") ++
-      (optional isCross "--with-symbol-resolution=ifunc");
+      (optional enableGlibcxxDebug "CXXFLAGS=-D_GLIBCXX_DEBUG");
 
     postConfigure = lib.optionalString (!enableDebug) ''
       # Disable debug
@@ -102,14 +97,16 @@ in
       # TODO: papi_version is needed for configure:
       # ./configure: line 27378: papi_version: command not found
       # This probably breaks cross-compilation
-    ] ++ lib.optionals enablePapi [ papi ];
+      papi
+    ];
 
     buildInputs = [
       boost
       numactl
       hwloc
+      papi
       ovni
-    ] ++ lib.optionals enablePapi [ papi ];
+    ];
 
     # Create a script that sets NANOS6_HOME
     postInstall = ''

pkgs/nosv/default.nix

@@ -7,7 +7,7 @@
 , numactl
 , hwloc
 , papi
-, enablePapi ? stdenv.hostPlatform == stdenv.buildPlatform # Disabled when cross-compiling
+, enablePapi ? true
 , cacheline ? 64 # bits
 , ovni ? null
 , useGit ? false

pkgs/tacuda/default.nix

@@ -0,0 +1,68 @@
+{ stdenv
+, fetchFromGitHub
+, automake
+, autoconf
+, libtool
+, gnumake
+, autoreconfHook
+, boost
+, cudaPackages
+
+, useGit ? false
+, gitUrl ? "git@gitlab-internal.bsc.es:task-awareness/tacuda/tacuda.git"
+, gitBranch ? "main"
+, gitCommit ? "35234f9445e6149a2bd38d119841e2485d6ee05e"
+}:
+
+let
+
+  release_ver = "2.1.0";
+  release = {
+    version = release_ver;
+    src = fetchFromGitHub {
+      owner = "bsc-pm";
+      repo = "tacuda";
+      rev = release_ver;
+      hash = "sha256-Cj3EiLVJSLvRv0ydeg7Vp4SpkniEqHkcWF+YOJQ8EcM=";
+    };
+  };
+
+  git = rec {
+    version = src.shortRev;
+    src = builtins.fetchGit {
+      url = gitUrl;
+      ref = gitBranch;
+      rev = gitCommit;
+    };
+  };
+
+  source = if (useGit) then git else release;
+
+in
+
+stdenv.mkDerivation {
+  pname = "tacuda";
+  inherit (source) src version;
+
+  enableParallelBuilding = true;
+  separateDebugInfo = true;
+
+  nativeBuildInputs = [
+    autoreconfHook
+    automake
+    autoconf
+    libtool
+    gnumake
+  ];
+
+  patches = [ ./fix_config.patch ];
+
+  configureFlags = [ "--with-cuda-include=${cudaPackages.cudatoolkit}/include" ];
+
+  buildInputs = [
+    boost
+    cudaPackages.libcublas
+    cudaPackages.cuda_cudart
+  ];
+
+}

pkgs/tacuda/fix_config.patch

@@ -0,0 +1,13 @@
+diff --git a/m4/cuda.m4 b/m4/cuda.m4
+index 23f5c94..8f9b534 100644
+--- a/m4/cuda.m4
++++ b/m4/cuda.m4
+@@ -40,7 +40,7 @@ search_libs="cuda cublas cudart"
+ required_libs=""
+ 
+ m4_foreach([function],
+-           [cuInit,
++           [
+ 	    cublasSgemm,
+             cudaStreamCreate,
+             cudaLaunchKernel,