rocmPackages.rocblas: remove reference to tensile in librocblas, split outputs (#458711)

pkgs/development/rocm-modules/6/default.nix

@@ -26,8 +26,6 @@ let
     {
       inherit rocmClangStdenv;
       stdenv = rocmClangStdenv;
-      buildTests = false;
-      buildBenchmarks = false;
 
       rocmUpdateScript = self.callPackage ./update.nix { };
 

pkgs/development/rocm-modules/6/rocblas/default.nix

@@ -7,6 +7,7 @@
   cmake,
   rocm-cmake,
   clr,
+  diffutils,
   python3,
   tensile,
   boost,
@@ -24,9 +25,10 @@
   python3Packages,
   rocm-smi,
   pkg-config,
+  removeReferencesTo,
   buildTensile ? true,
-  buildTests ? false,
-  buildBenchmarks ? false,
+  buildTests ? true,
+  buildBenchmarks ? true,
   tensileSepArch ? true,
   tensileLazyLib ? true,
   withHipBlasLt ? true,
@@ -47,6 +49,8 @@ stdenv.mkDerivation (finalAttrs: {
     hash = "sha256-FCzo/BOk4xLEFkdOdqcCXh4a9t3/OIIBEy8oz6oOMWg=";
   };
 
+  outputs = [ "out" ] ++ lib.optional buildBenchmarks "benchmark" ++ lib.optional buildTests "test";
+
   nativeBuildInputs = [
     cmake
     # no ninja, it buffers console output and nix times out long periods of no output
@@ -54,6 +58,7 @@ stdenv.mkDerivation (finalAttrs: {
     clr
     gitMinimal
     pkg-config
+    removeReferencesTo
   ]
   ++ lib.optionals buildTensile [
     tensile
@@ -76,10 +81,8 @@ stdenv.mkDerivation (finalAttrs: {
     python3Packages.msgpack
     python3Packages.zstandard
   ]
-  ++ lib.optionals buildTests [
-    gtest
-  ]
   ++ lib.optionals (buildTests || buildBenchmarks) [
+    gtest
     gfortran
     rocm-smi
   ]
@@ -146,14 +149,39 @@ stdenv.mkDerivation (finalAttrs: {
     substituteInPlace CMakeLists.txt \
       --replace-fail "4.43.0" "4.44.0" \
       --replace-fail '0.10' '1.0'
+  ''
+  # Fixes sh: line 1: /usr/bin/diff: No such file or directory
+  # /build/source/clients/gtest/../include/testing_logging.hpp:1117: Failure
+  + lib.optionals buildTests ''
+    substituteInPlace clients/include/testing_logging.hpp \
+      --replace-fail "/usr/bin/diff" "${lib.getExe' diffutils "diff"}"
   '';
 
+  postInstall =
+    # tensile isn't needed at runtime and pulls in ~400MB of python deps
+    ''
+      remove-references-to -t ${tensile} \
+        "$out/lib/librocblas.so."*
+    ''
+    + lib.optionalString buildBenchmarks ''
+      moveToOutput "bin/*-tune" "$benchmark"
+      moveToOutput "bin/*-bench" "$benchmark"
+      moveToOutput "bin/*example*" "$benchmark"
+      cp "$out/bin/"*.{yaml,txt} "$benchmark/bin"
+    ''
+    + lib.optionalString buildTests ''
+      moveToOutput "bin/*test*" "$test"
+      cp "$out/bin/"*.{yaml,txt} "$test/bin"
+    ''
+    + ''
+      if [ -d $out/bin ]; then
+        rm $out/bin/*.{yaml,txt} || true
+        rmdir $out/bin
+      fi
+    '';
+
   passthru = {
     amdgpu_targets = gpuTargets';
-    tests.rocblas-tests = finalAttrs.finalPackage.override {
-      buildBenchmarks = true;
-      buildTests = true;
-    };
     updateScript = rocmUpdateScript {
       name = finalAttrs.pname;
       inherit (finalAttrs.src) owner;