rocmPackages: 6.0.2 -> 6.3.3, and various ROCm build fixes and new packages (#367695)

nixos/doc/manual/release-notes/rl-2505.section.md

@@ -389,6 +389,10 @@
 
 - `python3Packages.jaeger-client` was removed because it was deprecated upstream. [OpenTelemetry](https://opentelemetry.io) is the recommended replacement.
 
+- `rocmPackages.rocm-thunk` has been removed and its functionality has been integrated with the ROCm CLR. Use `rocmPackages.clr` instead.
+
+- `rocmPackages.clang-ocl` has been removed. [It was deprecated by AMD in 2023.](https://github.com/ROCm/clang-ocl)
+
 - `nodePackages.meshcommander` has been removed, as the package was deprecated by Intel.
 
 - The default version of `z3` has been updated from 4.8 to 4.13. There are still a few packages that need specific older versions; those will continue to be maintained as long as other packages depend on them but may be removed in the future.

pkgs/by-name/ll/llama-cpp/package.nix

@@ -152,16 +152,11 @@ effectiveStdenv.mkDerivation (finalAttrs: {
     ++ optionals cudaSupport [
       (cmakeFeature "CMAKE_CUDA_ARCHITECTURES" cudaPackages.flags.cmakeCudaArchitecturesString)
     ]
-    ++ optionals rocmSupport [
-      (cmakeFeature "CMAKE_C_COMPILER" "hipcc")
-      (cmakeFeature "CMAKE_CXX_COMPILER" "hipcc")
-
-      # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM
-      # in https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
-      # and select the line that matches the current nixpkgs version of rocBLAS.
-      # Should likely use `rocmPackages.clr.gpuTargets`.
-      "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
-    ]
+    ++ optionals rocmSupport ([
+      (cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.clr.hipClangPath}/clang++")
+      # TODO: this should become `clr.gpuTargets` in the future.
+      (cmakeFeature "CMAKE_HIP_ARCHITECTURES" rocmPackages.rocblas.amdgpu_targets)
+    ])
     ++ optionals metalSupport [
       (cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
       (cmakeBool "LLAMA_METAL_EMBED_LIBRARY" true)

pkgs/by-name/ol/ollama/package.nix

@@ -70,6 +70,7 @@ let
 
   rocmLibs = [
     rocmPackages.clr
+    rocmPackages.hipblas-common
     rocmPackages.hipblas
     rocmPackages.rocblas
     rocmPackages.rocsolver
@@ -77,10 +78,9 @@ let
     rocmPackages.rocm-device-libs
     rocmPackages.rocm-smi
   ];
-  rocmClang = linkFarm "rocm-clang" { llvm = rocmPackages.llvm.clang; };
   rocmPath = buildEnv {
     name = "rocm-path";
-    paths = rocmLibs ++ [ rocmClang ];
+    paths = rocmLibs;
   };
 
   cudaLibs = [
@@ -131,6 +131,8 @@ let
   goBuild =
     if enableCuda then
       buildGoModule.override { stdenv = cudaPackages.backendStdenv; }
+    else if enableRocm then
+      buildGoModule.override { stdenv = rocmPackages.stdenv; }
     else
       buildGoModule;
   inherit (lib) licenses platforms maintainers;
@@ -148,6 +150,14 @@ goBuild {
       ROCM_PATH = rocmPath;
       CLBlast_DIR = "${clblast}/lib/cmake/CLBlast";
       HIP_PATH = rocmPath;
+      CFLAGS = "-Wno-c++17-extensions -I${rocmPath}/include";
+      CXXFLAGS = "-Wno-c++17-extensions -I${rocmPath}/include";
+    }
+    // lib.optionalAttrs (enableRocm && (rocmPackages.clr.localGpuTargets or false)) {
+
+      # If rocm CLR is set to build for an exact set of targets reuse that target list,
+      # otherwise let ollama use its builtin defaults
+      HIP_ARCHS = lib.concatStringsSep ";" rocmPackages.clr.localGpuTargets;
     }
     // lib.optionalAttrs enableCuda { CUDA_PATH = cudaPath; };
 

pkgs/by-name/uc/ucx/package.nix

@@ -33,8 +33,10 @@ let
     paths = rocmList;
   };
 
+  # rocm build fails with gcc stdenv due to unrecognised arg parallel-jobs
+  stdenv' = if enableRocm then rocmPackages.stdenv else stdenv;
 in
-stdenv.mkDerivation rec {
+stdenv'.mkDerivation rec {
   pname = "ucx";
   version = "1.18.0";
 

pkgs/development/libraries/science/math/magma/generic.nix

@@ -117,18 +117,11 @@ stdenv.mkDerivation {
     "test"
   ];
 
-  patches = lib.optionals (version == "2.9.0") [
-    # get ROCm version directly
-    # https://github.com/icl-utk-edu/magma/pull/27
-    (fetchpatch {
-      url = "https://github.com/icl-utk-edu/magma/commit/10fe816b763c41099fa1c978a79d6869246671cf.patch";
-      hash = "sha256-qSY5ACMHyHofJdQKyPqx8sI8GbPD6IZezmCd8qOS5OM=";
-    })
-  ];
-
-  # Fixup for the python test runners
   postPatch =
     ''
+      # For rocm version script invoked by cmake
+      patchShebangs tools/
+      # Fixup for the python test runners
       patchShebangs ./testing/run_{tests,summarize}.py
     ''
     + lib.optionalString (strings.versionOlder version "2.9.0") ''
@@ -201,6 +194,10 @@ stdenv.mkDerivation {
       (strings.cmakeFeature "MIN_ARCH" minArch) # Disarms magma's asserts
     ]
     ++ lists.optionals rocmSupport [
+      # Can be removed once https://github.com/icl-utk-edu/magma/pull/27 is merged
+      # Can't easily apply the PR as a patch because we rely on the tarball with pregenerated
+      # hipified files ∴ fetchpatch of the PR will apply cleanly but fail to build
+      (strings.cmakeFeature "ROCM_CORE" "${effectiveRocmPackages.clr}")
       (strings.cmakeFeature "CMAKE_C_COMPILER" "${effectiveRocmPackages.clr}/bin/hipcc")
       (strings.cmakeFeature "CMAKE_CXX_COMPILER" "${effectiveRocmPackages.clr}/bin/hipcc")
     ];

pkgs/development/python-modules/accelerate/default.nix

@@ -92,8 +92,9 @@ buildPythonPackage rec {
       # set the environment variable, CC, which conflicts with standard environment
       "test_patch_environment_key_exists"
     ]
-    ++ lib.optionals (pythonAtLeast "3.13") [
+    ++ lib.optionals ((pythonAtLeast "3.13") || (torch.rocmSupport or false)) [
       # RuntimeError: Dynamo is not supported on Python 3.13+
+      # OR torch.compile tests broken on torch 2.5 + rocm
       "test_can_unwrap_distributed_compiled_model_keep_torch_compile"
       "test_can_unwrap_distributed_compiled_model_remove_torch_compile"
       "test_convert_to_fp32"

pkgs/development/python-modules/torch/default.nix

@@ -36,6 +36,7 @@
   symlinkJoin,
   which,
   pybind11,
+  pkg-config,
   removeReferencesTo,
 
   # Build inputs
@@ -69,7 +70,13 @@
   #          (dependencies without cuda support).
   #          Instead we should rely on overlays and nixpkgsFun.
   # (@SomeoneSerge)
-  _tritonEffective ? if cudaSupport then triton-cuda else triton,
+  _tritonEffective ?
+    if cudaSupport then
+      triton-cuda
+    else if rocmSupport then
+      rocmPackages.triton
+    else
+      triton,
   triton-cuda,
 
   # Disable MKLDNN on aarch64-darwin, it negatively impacts performance,
@@ -91,7 +98,7 @@
 
   # ROCm dependencies
   rocmSupport ? config.rocmSupport,
-  rocmPackages_5,
+  rocmPackages,
   gpuTargets ? [ ],
 
   vulkanSupport ? false,
@@ -111,8 +118,6 @@ let
 
   triton = throw "python3Packages.torch: use _tritonEffective instead of triton to avoid divergence";
 
-  rocmPackages = rocmPackages_5;
-
   setBool = v: if v then "1" else "0";
 
   # https://github.com/pytorch/pytorch/blob/v2.4.0/torch/utils/cpp_extension.py#L1953
@@ -180,7 +185,7 @@ let
       clr
       rccl
       miopen
-      miopengemm
+      aotriton
       rocrand
       rocblas
       rocsparse
@@ -192,10 +197,12 @@ let
       rocfft
       rocsolver
       hipfft
+      hiprand
       hipsolver
+      hipblas-common
       hipblas
+      hipblaslt
       rocminfo
-      rocm-thunk
       rocm-comgr
       rocm-device-libs
       rocm-runtime
@@ -225,8 +232,6 @@ let
     # In particular, this triggered warnings from cuda's `aliases.nix`
     "Magma cudaPackages does not match cudaPackages" =
       cudaSupport && (effectiveMagma.cudaPackages.cudaVersion != cudaPackages.cudaVersion);
-    "Rocm support is currently broken because `rocmPackages.hipblaslt` is unpackaged. (2024-06-09)" =
-      rocmSupport;
   };
 
   unroll-src = writeShellScript "unroll-src" ''
@@ -293,6 +298,11 @@ buildPythonPackage rec {
 
       # annotations (3.7), print_function (3.0), with_statement (2.6) are all supported
       sed -i -e "/from __future__ import/d" **.py
+      substituteInPlace third_party/NNPACK/CMakeLists.txt \
+        --replace-fail "PYTHONPATH=" 'PYTHONPATH=$ENV{PYTHONPATH}:'
+      # flag from cmakeFlags doesn't work, not clear why
+      # setting it at the top of NNPACK's own CMakeLists does
+      sed -i '2s;^;set(PYTHON_SIX_SOURCE_DIR ${six.src})\n;' third_party/NNPACK/CMakeLists.txt
     ''
     + lib.optionalString rocmSupport ''
       # https://github.com/facebookincubator/gloo/pull/297
@@ -365,6 +375,10 @@ buildPythonPackage rec {
   # We only do an imports check, so do not build tests either.
   BUILD_TEST = setBool false;
 
+  # ninja hook doesn't automatically turn on ninja
+  # because pytorch setup.py is responsible for this
+  CMAKE_GENERATOR = "Ninja";
+
   # Unlike MKL, oneDNN (née MKLDNN) is FOSS, so we enable support for
   # it by default. PyTorch currently uses its own vendored version
   # of oneDNN through Intel iDeep.
@@ -375,14 +389,15 @@ buildPythonPackage rec {
   # Also avoids pytorch exporting the headers of pybind11
   USE_SYSTEM_PYBIND11 = true;
 
-  # NB technical debt: building without NNPACK as workaround for missing `six`
-  USE_NNPACK = 0;
+  # Multicore CPU convnet support
+  USE_NNPACK = 1;
 
   # Explicitly enable MPS for Darwin
   USE_MPS = setBool stdenv.hostPlatform.isDarwin;
 
   cmakeFlags =
     [
+      (lib.cmakeFeature "PYTHON_SIX_SOURCE_DIR" "${six.src}")
       # (lib.cmakeBool "CMAKE_FIND_DEBUG_MODE" true)
       (lib.cmakeFeature "CUDAToolkit_VERSION" cudaPackages.cudaVersion)
     ]
@@ -440,6 +455,9 @@ buildPythonPackage rec {
     }
     // lib.optionalAttrs vulkanSupport {
       VULKAN_SDK = shaderc.bin;
+    }
+    // lib.optionalAttrs rocmSupport {
+      AOTRITON_INSTALLED_PREFIX = "${rocmPackages.aotriton}";
     };
 
   nativeBuildInputs =
@@ -448,6 +466,7 @@ buildPythonPackage rec {
       which
       ninja
       pybind11
+      pkg-config
       removeReferencesTo
     ]
     ++ lib.optionals cudaSupport (
@@ -501,7 +520,10 @@ buildPythonPackage rec {
     ]
     ++ lib.optionals tritonSupport [ _tritonEffective ]
     ++ lib.optionals MPISupport [ mpi ]
-    ++ lib.optionals rocmSupport [ rocmtoolkit_joined ];
+    ++ lib.optionals rocmSupport [
+      rocmtoolkit_joined
+      rocmPackages.clr # Added separately so setup hook applies
+    ];
 
   pythonRelaxDeps = [
     "sympy"

pkgs/development/python-modules/torchaudio/default.nix

@@ -36,7 +36,6 @@ let
       clr
       rccl
       miopen
-      miopengemm
       rocrand
       rocblas
       rocsparse
@@ -49,9 +48,9 @@ let
       rocsolver
       hipfft
       hipsolver
+      hipblas-common
       hipblas
       rocminfo
-      rocm-thunk
       rocm-comgr
       rocm-device-libs
       rocm-runtime

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

@@ -0,0 +1,89 @@
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  fetchpatch,
+  rocmUpdateScript,
+  cmake,
+  pkg-config,
+  libdrm,
+  wrapPython,
+  autoPatchelfHook,
+}:
+
+let
+  esmi_ib_src = fetchFromGitHub {
+    owner = "amd";
+    repo = "esmi_ib_library";
+    rev = "esmi_pkg_ver-3.0.3";
+    hash = "sha256-q0w5c5c+CpXkklmSyfzc+sbkt4cHNxscGJA3AXwvHxQ=";
+  };
+in
+stdenv.mkDerivation (finalAttrs: {
+  pname = "amdsmi";
+  version = "6.3.3";
+  src = fetchFromGitHub {
+    owner = "rocm";
+    repo = "amdsmi";
+    rev = "rocm-${finalAttrs.version}";
+    hash = "sha256-hrPqd4ZWqzTG7JRfVwc1SZx6TNS0Q/LFg8yDxrM3mPo=";
+  };
+
+  postPatch = ''
+    substituteInPlace goamdsmi_shim/CMakeLists.txt \
+      --replace-fail "amd_smi)" ${"'"}''${AMD_SMI_TARGET})' \
+      --replace-fail 'target_link_libraries(''${GOAMDSMI_SHIM_TARGET} -L' '#'
+
+    # Manually unpack esmi_ib_src and add amd_hsmp.h so execute-process git clone doesn't run
+    cp -rf --no-preserve=mode ${esmi_ib_src} ./esmi_ib_library
+    mkdir -p ./esmi_ib_library/include/asm
+    cp ./include/amd_smi/impl/amd_hsmp.h ./esmi_ib_library/include/asm/amd_hsmp.h
+  '';
+
+  patches = [
+    # Fix ld.lld undefined reference: drmGetVersion
+    (fetchpatch {
+      url = "https://github.com/ROCm/amdsmi/commit/c3864bf6171970d86dc50fd23f06377736823997.patch";
+      hash = "sha256-zRG1tBD8sIQCWdKfCbXC/Z/6d6NTrRYvRpddPWdM4j8=";
+    })
+  ];
+
+  nativeBuildInputs = [
+    cmake
+    pkg-config
+    wrapPython
+    autoPatchelfHook
+  ];
+
+  buildInputs = [
+    libdrm
+  ];
+
+  cmakeFlags = [
+    # Manually define CMAKE_INSTALL_<DIR>
+    # See: https://github.com/NixOS/nixpkgs/pull/197838
+    "-DCMAKE_INSTALL_BINDIR=bin"
+    "-DCMAKE_INSTALL_LIBDIR=lib"
+    "-DCMAKE_INSTALL_INCLUDEDIR=include"
+  ];
+
+  postInstall = ''
+    wrapPythonProgramsIn $out
+    rm $out/bin/amd-smi
+    ln -sf $out/libexec/amdsmi_cli/amdsmi_cli.py $out/bin/amd-smi
+  '';
+
+  passthru.updateScript = rocmUpdateScript {
+    name = finalAttrs.pname;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
+  };
+
+  meta = with lib; {
+    description = "System management interface for AMD GPUs supported by ROCm";
+    homepage = "https://github.com/ROCm/rocm_smi_lib";
+    license = with licenses; [ mit ];
+    maintainers = with maintainers; [ lovesegfault ] ++ teams.rocm.members;
+    platforms = [ "x86_64-linux" ];
+  };
+})

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

@@ -0,0 +1,226 @@
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  cmake,
+  rocm-cmake,
+  clr,
+  rocblas,
+  rocsolver,
+  gtest,
+  msgpack,
+  libxml2,
+  python3,
+  python3Packages,
+  openmp,
+  hipblas-common,
+  hipblas,
+  nlohmann_json,
+  triton-llvm,
+  rocmlir,
+  lapack-reference,
+  ninja,
+  ncurses,
+  libffi,
+  zlib,
+  zstd,
+  xz,
+  pkg-config,
+  buildTests ? false,
+  buildBenchmarks ? false,
+  buildSamples ? false,
+  gpuTargets ? [
+    # aotriton GPU support list:
+    # https://github.com/ROCm/aotriton/blob/main/v2python/gpu_targets.py
+    "gfx90a"
+    "gfx942"
+    "gfx1100"
+    "gfx1101"
+  ],
+}:
+
+stdenv.mkDerivation (
+  finalAttrs:
+  let
+    py = python3.withPackages (ps: [
+      ps.pyyaml
+      ps.distutils
+      ps.setuptools
+      ps.packaging
+      ps.numpy
+      ps.wheel
+      ps.filelock
+      ps.iniconfig
+      ps.pluggy
+      ps.pybind11
+    ]);
+    gpuTargets' = lib.concatStringsSep ";" gpuTargets;
+    compiler = "amdclang++";
+    cFlags = "-O3 -DNDEBUG";
+    triton-llvm' = triton-llvm;
+  in
+  {
+    pname = "aotriton";
+    version = "0.8.2b";
+
+    src = fetchFromGitHub {
+      owner = "ROCm";
+      repo = "aotriton";
+      rev = "${finalAttrs.version}";
+      hash = "sha256-gSzGYWfyUNLyzqpu3BM8rjFFL7cRVZ+w9L5pnh9QGz4=";
+      fetchSubmodules = true;
+    };
+    env.CXX = compiler;
+    env.ROCM_PATH = "${clr}";
+    requiredSystemFeatures = [ "big-parallel" ];
+
+    outputs =
+      [
+        "out"
+      ]
+      ++ lib.optionals buildTests [
+        "test"
+      ]
+      ++ lib.optionals buildBenchmarks [
+        "benchmark"
+      ]
+      ++ lib.optionals buildSamples [
+        "sample"
+      ];
+
+    # Need an empty cuda.h for this to compile
+    # Better than pulling in unfree cuda headers
+    postPatch = ''
+      touch third_party/triton/third_party/nvidia/include/cuda.h
+    '';
+
+    doCheck = false;
+    doInstallCheck = false;
+
+    nativeBuildInputs = [
+      cmake
+      rocm-cmake
+      pkg-config
+      py
+      clr
+      ninja
+    ];
+
+    buildInputs =
+      [
+        rocblas
+        rocsolver
+        hipblas-common
+        hipblas
+        openmp
+        libffi
+        ncurses
+        xz
+        nlohmann_json
+        rocmlir
+
+        msgpack
+        libxml2
+        python3Packages.msgpack
+        zlib
+        zstd
+      ]
+      ++ lib.optionals buildTests [
+        gtest
+      ]
+      ++ lib.optionals (buildTests || buildBenchmarks) [
+        lapack-reference
+      ];
+
+    env.TRITON_OFFLINE_BUILD = 1;
+    env.LLVM_SYSPATH = "${triton-llvm'}";
+    env.JSON_SYSPATH = nlohmann_json;
+    env.MLIR_DIR = "${triton-llvm'}/lib/cmake/mlir";
+    env.CXXFLAGS = "-I/build/source/third_party/triton/third_party/nvidia/backend/include";
+
+    # Fix up header issues in triton: https://github.com/triton-lang/triton/pull/3985/files
+    preConfigure = ''
+      mkdir third_party/triton/third_party/nvidia/backend/include/
+      touch third_party/triton/third_party/nvidia/backend/include/cuda.h
+      find third_party/triton -type f -exec sed -i 's|[<]cupti.h[>]|"cupti.h"|g' {} +
+      find third_party/triton -type f -exec sed -i 's|[<]cuda.h[>]|"cuda.h"|g' {} +
+
+      sed -i '2s;^;set(CMAKE_SUPPRESS_DEVELOPER_WARNINGS ON CACHE BOOL "ON")\n;' CMakeLists.txt
+      sed -i '2s;^;set(CMAKE_VERBOSE_MAKEFILE ON CACHE BOOL "ON")\n;' CMakeLists.txt
+      sed -i '2s;^;set(CMAKE_SUPPRESS_DEVELOPER_WARNINGS ON CACHE BOOL "ON")\n;' third_party/triton/CMakeLists.txt
+      sed -i '2s;^;set(CMAKE_VERBOSE_MAKEFILE ON CACHE BOOL "ON")\n;' third_party/triton/CMakeLists.txt
+      substituteInPlace third_party/triton/python/setup.py \
+        --replace-fail "from distutils.command.clean import clean" "import setuptools;from distutils.command.clean import clean" \
+        --replace-fail 'system == "Linux"' 'False'
+
+      cmakeFlagsArray+=(
+        '-DCMAKE_C_FLAGS_RELEASE=${cFlags}'
+        '-DCMAKE_CXX_FLAGS_RELEASE=${cFlags}'
+      )
+      prependToVar cmakeFlags "-GNinja"
+      mkdir -p /build/tmp-home
+      export HOME=/build/tmp-home
+    '';
+
+    # Excerpt from README:
+    # Note: do not run ninja separately, due to the limit of the current build system,
+    # ninja install will run the whole build process unconditionally.
+    dontBuild = true;
+
+    installPhase = ''
+      runHook preInstall
+      ninja -v install
+      runHook postInstall
+    '';
+
+    cmakeFlags =
+      [
+        "-Wno-dev"
+        "-DAOTRITON_NOIMAGE_MODE=ON" # FIXME: Should be able to build with object code but generate_shim is failing
+        "-DCMAKE_BUILD_TYPE=Release"
+        "-DCMAKE_VERBOSE_MAKEFILE=ON"
+        "-DVIRTUALENV_PYTHON_EXENAME=${lib.getExe py}"
+        "-DCMAKE_CXX_COMPILER=${compiler}"
+        # Manually define CMAKE_INSTALL_<DIR>
+        # See: https://github.com/NixOS/nixpkgs/pull/197838
+        "-DCMAKE_INSTALL_BINDIR=bin"
+        "-DCMAKE_INSTALL_LIBDIR=lib"
+        "-DCMAKE_INSTALL_INCLUDEDIR=include"
+        "-DAMDGPU_TARGETS=${gpuTargets'}"
+        "-DGPU_TARGETS=${gpuTargets'}"
+      ]
+      ++ lib.optionals buildTests [
+        "-DBUILD_CLIENTS_TESTS=ON"
+      ]
+      ++ lib.optionals buildBenchmarks [
+        "-DBUILD_CLIENTS_BENCHMARKS=ON"
+      ]
+      ++ lib.optionals buildSamples [
+        "-DBUILD_CLIENTS_SAMPLES=ON"
+      ];
+
+    postInstall =
+      lib.optionalString buildTests ''
+        mkdir -p $test/bin
+        mv $out/bin/hipblas-test $test/bin
+      ''
+      + lib.optionalString buildBenchmarks ''
+        mkdir -p $benchmark/bin
+        mv $out/bin/hipblas-bench $benchmark/bin
+      ''
+      + lib.optionalString buildSamples ''
+        mkdir -p $sample/bin
+        mv $out/bin/example-* $sample/bin
+      ''
+      + lib.optionalString (buildTests || buildBenchmarks || buildSamples) ''
+        rmdir $out/bin
+      '';
+    meta = with lib; {
+      description = "ROCm Ahead of Time (AOT) Triton Math Library ";
+      homepage = "https://github.com/ROCm/aotriton";
+      license = with licenses; [ mit ];
+      maintainers = teams.rocm.members;
+      platforms = platforms.linux;
+    };
+  }
+)

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

@@ -1,45 +0,0 @@
-{
-  lib,
-  stdenv,
-  fetchFromGitHub,
-  rocmUpdateScript,
-  cmake,
-  rocm-cmake,
-  rocm-device-libs,
-}:
-
-stdenv.mkDerivation (finalAttrs: {
-  pname = "clang-ocl";
-  version = "6.0.2";
-
-  src = fetchFromGitHub {
-    owner = "ROCm";
-    repo = "clang-ocl";
-    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-uMSvcVJj+me2E+7FsXZ4l4hTcK6uKEegXpkHGcuist0=";
-  };
-
-  nativeBuildInputs = [
-    cmake
-    rocm-cmake
-  ];
-
-  buildInputs = [ rocm-device-libs ];
-
-  passthru.updateScript = rocmUpdateScript {
-    name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
-  };
-
-  meta = with lib; {
-    description = "OpenCL compilation with clang compiler";
-    homepage = "https://github.com/ROCm/clang-ocl";
-    license = with licenses; [ mit ];
-    maintainers = teams.rocm.members;
-    platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
-  };
-})

pkgs/development/rocm-modules/6/clr/cmake-find-x11-libgl.patch

@@ -0,0 +1,40 @@
+diff --git a/rocclr/cmake/ROCclr.cmake b/rocclr/cmake/ROCclr.cmake
+index 3f233b72f..67bdc62ee 100644
+--- a/rocclr/cmake/ROCclr.cmake
++++ b/rocclr/cmake/ROCclr.cmake
+@@ -44,6 +44,19 @@ find_package(Threads REQUIRED)
+ 
+ find_package(AMD_OPENCL)
+ 
++# Find X11 package
++find_package(X11 REQUIRED)
++if(NOT X11_FOUND)
++    message(FATAL_ERROR "X11 libraries not found")
++endif()
++
++# Find OpenGL package
++find_package(OpenGL REQUIRED)
++if(NOT OpenGL_FOUND)
++    message(FATAL_ERROR "OpenGL not found")
++endif()
++
++
+ add_library(rocclr STATIC)
+ 
+ include(ROCclrCompilerOptions)
+@@ -123,9 +136,14 @@ target_include_directories(rocclr PUBLIC
+   ${ROCCLR_SRC_DIR}/device
+   ${ROCCLR_SRC_DIR}/elf
+   ${ROCCLR_SRC_DIR}/include
++  ${X11_INCLUDE_DIR}
++  ${OPENGL_INCLUDE_DIR}
+   ${AMD_OPENCL_INCLUDE_DIRS})
+ 
+-target_link_libraries(rocclr PUBLIC Threads::Threads)
++target_link_libraries(rocclr PUBLIC
++  Threads::Threads
++  ${X11_LIBRARIES}
++  ${OPENGL_LIBRARIES})
+ # IPC on Windows is not supported
+ if(UNIX)
+   target_link_libraries(rocclr PUBLIC rt)

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

@@ -4,52 +4,58 @@
   callPackage,
   fetchFromGitHub,
   fetchpatch,
-  fetchurl,
   rocmUpdateScript,
   makeWrapper,
   cmake,
   perl,
-  clang,
   hip-common,
   hipcc,
   rocm-device-libs,
   rocm-comgr,
   rocm-runtime,
+  rocm-core,
   roctracer,
   rocminfo,
   rocm-smi,
   numactl,
+  libffi,
+  zstd,
+  zlib,
   libGL,
   libxml2,
   libX11,
   python3Packages,
+  rocm-merged-llvm,
+  khronos-ocl-icd-loader,
+  gcc-unwrapped,
+  writeShellScriptBin,
+  localGpuTargets ? null,
 }:
 
 let
+  inherit (rocm-core) ROCM_LIBPATCH_VERSION;
+  hipClang = rocm-merged-llvm;
+  hipClangPath = "${hipClang}/bin";
   wrapperArgs = [
     "--prefix PATH : $out/bin"
     "--prefix LD_LIBRARY_PATH : ${rocm-runtime}"
     "--set HIP_PLATFORM amd"
     "--set HIP_PATH $out"
-    "--set HIP_CLANG_PATH ${clang}/bin"
+    "--set HIP_CLANG_PATH ${hipClangPath}"
     "--set DEVICE_LIB_PATH ${rocm-device-libs}/amdgcn/bitcode"
     "--set HSA_PATH ${rocm-runtime}"
     "--set ROCM_PATH $out"
   ];
-
-  # https://github.com/NixOS/nixpkgs/issues/305641
-  # Not needed when 3.29.2 is in unstable
-  cmake' = cmake.overrideAttrs (old: rec {
-    version = "3.29.2";
-    src = fetchurl {
-      url = "https://cmake.org/files/v${lib.versions.majorMinor version}/cmake-${version}.tar.gz";
-      hash = "sha256-NttLaSaqt0G6bksuotmckZMiITIwi03IJNQSPLcwNS4=";
-    };
-  });
+  amdclang = writeShellScriptBin "amdclang" ''
+    exec clang "$@"
+  '';
+  amdclangxx = writeShellScriptBin "amdclang++" ''
+    exec clang++ "$@"
+  '';
 in
 stdenv.mkDerivation (finalAttrs: {
   pname = "clr";
-  version = "6.0.2";
+  version = "6.3.3";
 
   outputs = [
     "out"
@@ -60,15 +66,17 @@ stdenv.mkDerivation (finalAttrs: {
     owner = "ROCm";
     repo = "clr";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-ZMpA7vCW2CcpGdBLZfPimMHcgjhN1PHuewJiYwZMgGY=";
+    hash = "sha256-4qjfnn0kto2sNaSumXxHRHFrf3a3RZILOdhVSxkEs1I=";
   };
 
   nativeBuildInputs = [
     makeWrapper
-    cmake'
+    cmake
     perl
     python3Packages.python
     python3Packages.cppheaderparser
+    amdclang
+    amdclangxx
   ];
 
   buildInputs = [
@@ -76,9 +84,15 @@ stdenv.mkDerivation (finalAttrs: {
     libGL
     libxml2
     libX11
+    khronos-ocl-icd-loader
+    hipClang
+    libffi
+    zstd
+    zlib
   ];
 
   propagatedBuildInputs = [
+    rocm-core
     rocm-device-libs
     rocm-comgr
     rocm-runtime
@@ -86,6 +100,7 @@ stdenv.mkDerivation (finalAttrs: {
   ];
 
   cmakeFlags = [
+    "-DCMAKE_BUILD_TYPE=Release"
     "-DCMAKE_POLICY_DEFAULT_CMP0072=NEW" # Prefer newer OpenGL libraries
     "-DCLR_BUILD_HIP=ON"
     "-DCLR_BUILD_OCL=ON"
@@ -94,6 +109,9 @@ stdenv.mkDerivation (finalAttrs: {
     "-DHIP_PLATFORM=amd"
     "-DPROF_API_HEADER_PATH=${roctracer.src}/inc/ext"
     "-DROCM_PATH=${rocminfo}"
+    "-DBUILD_ICD=ON"
+    "-DHIP_ENABLE_ROCPROFILER_REGISTER=OFF" # circular dep - may need -minimal and -full builds?
+    "-DAMD_ICD_LIBRARY_DIR=${khronos-ocl-icd-loader}"
 
     # Temporarily set variables to work around upstream CMakeLists issue
     # Can be removed once https://github.com/ROCm/rocm-cmake/issues/121 is fixed
@@ -102,26 +120,36 @@ stdenv.mkDerivation (finalAttrs: {
     "-DCMAKE_INSTALL_LIBDIR=lib"
   ];
 
+  env.LLVM_DIR = "";
+
   patches = [
+    ./cmake-find-x11-libgl.patch
+
     (fetchpatch {
-      name = "add-missing-operators.patch";
-      url = "https://github.com/ROCm/clr/commit/86bd518981b364c138f9901b28a529899d8654f3.patch";
-      hash = "sha256-lbswri+zKLxif0hPp4aeJDeVfadhWZz4z+m+G2XcCPI=";
+      # Fix handling of old fatbin version https://github.com/ROCm/clr/issues/99
+      sha256 = "sha256-CK/QwgWJQEruiG4DqetF9YM0VEWpSiUMxAf1gGdJkuA=";
+      url = "https://src.fedoraproject.org/rpms/rocclr/raw/rawhide/f/0001-handle-v1-of-compressed-fatbins.patch";
     })
     (fetchpatch {
-      name = "static-functions.patch";
-      url = "https://github.com/ROCm/clr/commit/77c581a3ebd47b5e2908973b70adea66891159ee.patch";
-      hash = "sha256-auBedbd7rghlKav7A9V6l64J7VmtE9GizIdi5gWj+fs=";
+      # improve rocclr isa compatibility check
+      sha256 = "sha256-wUrhpYN68AbEXeFU5f366C6peqHyq25kujJXY/bBJMs=";
+      url = "https://github.com/GZGavinZhao/clr/commit/22c17a0ac09c6b77866febf366591f669a1ed133.patch";
     })
     (fetchpatch {
-      name = "extend-hip-isa-compatibility-check.patch";
-      url = "https://salsa.debian.org/rocm-team/rocm-hipamd/-/raw/d6d20142c37e1dff820950b16ff8f0523241d935/debian/patches/0026-extend-hip-isa-compatibility-check.patch";
-      hash = "sha256-eG0ALZZQLRzD7zJueJFhi2emontmYy6xx8Rsm346nQI=";
+      # [PATCH] Improve hipamd compat check
+      sha256 = "sha256-uZQ8rMrWH61CCbxwLqQGggDmXFmYTi6x8OcgYPrZRC8=";
+      url = "https://github.com/GZGavinZhao/clr/commit/63c6ee630966744d4199fdfb854e98d2da9e1122.patch";
     })
     (fetchpatch {
-      name = "improve-rocclr-isa-compatibility-check.patch";
-      url = "https://salsa.debian.org/rocm-team/rocm-hipamd/-/raw/d6d20142c37e1dff820950b16ff8f0523241d935/debian/patches/0025-improve-rocclr-isa-compatibility-check.patch";
-      hash = "sha256-8eowuRiOAdd9ucKv4Eg9FPU7c6367H3eP3fRAGfXc6Y=";
+      # [PATCH] SWDEV-504340 - Move cast of cl_mem inside the condition
+      # Fixes crash due to UB in KernelBlitManager::setArgument
+      sha256 = "sha256-nL4CZ7EOXqsTVUtYhuu9DLOMpnMeMRUhkhylEQLTg9I=";
+      url = "https://github.com/ROCm/clr/commit/fa63919a6339ea2a61111981ba2362c97fbdf743.patch";
+    })
+    (fetchpatch {
+      # [PATCH] SWDEV-507104 - Removes alignment requirement for Semaphore class to resolve runtime misaligned memory issues
+      sha256 = "sha256-nStJ22B/CM0fzQTvYjbHDbQt0GlE8DXxVK+UDU9BAx4=";
+      url = "https://github.com/ROCm/clr/commit/21d764518363d74187deaef2e66c1a127bc5aa64.patch";
     })
   ];
 
@@ -131,83 +159,121 @@ stdenv.mkDerivation (finalAttrs: {
 
     # We're not on Windows so these are never installed to hipcc...
     substituteInPlace hipamd/CMakeLists.txt \
-      --replace "install(PROGRAMS \''${HIPCC_BIN_DIR}/hipcc.bat DESTINATION bin)" "" \
-      --replace "install(PROGRAMS \''${HIPCC_BIN_DIR}/hipconfig.bat DESTINATION bin)" ""
+      --replace-fail "install(PROGRAMS \''${HIPCC_BIN_DIR}/hipcc.bat DESTINATION bin)" "" \
+      --replace-fail "install(PROGRAMS \''${HIPCC_BIN_DIR}/hipconfig.bat DESTINATION bin)" ""
 
     substituteInPlace hipamd/src/hip_embed_pch.sh \
-      --replace "\''$LLVM_DIR/bin/clang" "${clang}/bin/clang"
-
-    # https://lists.debian.org/debian-ai/2024/02/msg00178.html
-    substituteInPlace rocclr/utils/flags.hpp \
-      --replace-fail "HIP_USE_RUNTIME_UNBUNDLER, false" "HIP_USE_RUNTIME_UNBUNDLER, true"
+      --replace-fail "\''$LLVM_DIR/bin/clang" "${hipClangPath}/clang"
 
     substituteInPlace opencl/khronos/icd/loader/icd_platform.h \
       --replace-fail '#define ICD_VENDOR_PATH "/etc/OpenCL/vendors/";' \
                      '#define ICD_VENDOR_PATH "/run/opengl-driver/etc/OpenCL/vendors/";'
+
+    # new unbundler has better error messages, defaulting it on
+    substituteInPlace rocclr/utils/flags.hpp \
+      --replace-fail "HIP_ALWAYS_USE_NEW_COMGR_UNBUNDLING_ACTION, false" "HIP_ALWAYS_USE_NEW_COMGR_UNBUNDLING_ACTION, true"
   '';
 
   postInstall = ''
+    chmod +x $out/bin/*
     patchShebangs $out/bin
 
-    # hipcc.bin and hipconfig.bin is mysteriously never installed
-    cp -a ${hipcc}/bin/{hipcc.bin,hipconfig.bin} $out/bin
+    cp ${amdclang}/bin/* $out/bin/
+    cp ${amdclangxx}/bin/* $out/bin/
 
-    wrapProgram $out/bin/hipcc.bin ${lib.concatStringsSep " " wrapperArgs}
-    wrapProgram $out/bin/hipconfig.bin ${lib.concatStringsSep " " wrapperArgs}
-    wrapProgram $out/bin/hipcc.pl ${lib.concatStringsSep " " wrapperArgs}
-    wrapProgram $out/bin/hipconfig.pl ${lib.concatStringsSep " " wrapperArgs}
+    for prog in hip{cc,config}{,.pl}; do
+      wrapProgram $out/bin/$prog ${lib.concatStringsSep " " wrapperArgs}
+    done
+
+    mkdir -p $out/nix-support/
+    echo '
+    export HIP_PATH="${placeholder "out"}"
+    export HIP_PLATFORM=amd
+    export HIP_DEVICE_LIB_PATH="${rocm-device-libs}/amdgcn/bitcode"
+    export NIX_CC_USE_RESPONSE_FILE=0
+    export HIP_CLANG_PATH="${hipClangPath}"
+    export ROCM_LIBPATCH_VERSION="${ROCM_LIBPATCH_VERSION}"
+    export HSA_PATH="${rocm-runtime}"' > $out/nix-support/setup-hook
 
     # Just link rocminfo, it's easier
     ln -s ${rocminfo}/bin/* $out/bin
+    ln -s ${rocm-core}/include/* $out/include/
 
     # Replace rocm-opencl-icd functionality
     mkdir -p $icd/etc/OpenCL/vendors
     echo "$out/lib/libamdocl64.so" > $icd/etc/OpenCL/vendors/amdocl64.icd
 
     # add version info to output (downstream rocmPackages look for this)
-    mkdir $out/.info
-    echo "${finalAttrs.version}" > $out/.info/version
+    ln -s ${rocm-core}/.info/ $out/.info
+
+    ln -s ${hipClang} $out/llvm
   '';
 
-  passthru = {
-    # All known and valid general GPU targets
-    # We cannot use this for each ROCm library, as each defines their own supported targets
-    # See: https://github.com/ROCm/ROCm/blob/77cbac4abab13046ee93d8b5bf410684caf91145/README.md#library-target-matrix
-    gpuTargets = lib.forEach [
-      "803"
-      "900"
-      "906"
-      "908"
-      "90a"
-      "940"
-      "941"
-      "942"
-      "1010"
-      "1012"
-      "1030"
-      "1100"
-      "1101"
-      "1102"
-    ] (target: "gfx${target}");
+  disallowedRequisites = [
+    gcc-unwrapped
+  ];
 
-    updateScript = rocmUpdateScript {
-      name = finalAttrs.pname;
-      owner = finalAttrs.src.owner;
-      repo = finalAttrs.src.repo;
-      page = "tags?per_page=1";
-      filter = ".[0].name | split(\"-\") | .[1]";
-    };
+  passthru =
+    {
+      # All known and valid general GPU targets
+      # We cannot use this for each ROCm library, as each defines their own supported targets
+      # See: https://github.com/ROCm/ROCm/blob/77cbac4abab13046ee93d8b5bf410684caf91145/README.md#library-target-matrix
+      # Generic targets are not yet available in rocm-6.3.1 llvm
+      gpuTargets = lib.forEach [
+        # "9-generic"
+        "900" # MI25, Vega 56/64
+        "906" # MI50/60, Radeon VII
+        "908" # MI100
+        "90a" # MI210 / MI250
+        # "9-4-generic"
+        # 940/1 - never released publicly, maybe HPE cray specific MI3xx?
+        "942" # MI300
+        # "10-1-generic"
+        "1010"
+        "1012"
+        # "10-3-generic"
+        "1030" # W6800, various Radeon cards
+        # "11-generic"
+        "1100"
+        "1101"
+        "1102"
+      ] (target: "gfx${target}");
 
-    impureTests = {
-      rocm-smi = callPackage ./test-rocm-smi.nix {
-        inherit rocm-smi;
-        clr = finalAttrs.finalPackage;
+      inherit hipClangPath;
+
+      updateScript = rocmUpdateScript {
+        name = finalAttrs.pname;
+        inherit (finalAttrs.src) owner;
+        inherit (finalAttrs.src) repo;
+        page = "tags?per_page=4";
       };
-      opencl-example = callPackage ./test-opencl-example.nix {
-        clr = finalAttrs.finalPackage;
+
+      impureTests = {
+        rocm-smi = callPackage ./test-rocm-smi.nix {
+          inherit rocm-smi;
+          clr = finalAttrs.finalPackage;
+        };
+        opencl-example = callPackage ./test-opencl-example.nix {
+          clr = finalAttrs.finalPackage;
+        };
       };
+
+      selectGpuTargets =
+        {
+          supported ? [ ],
+        }:
+        supported;
+      gpuArchSuffix = "";
+    }
+    // lib.optionalAttrs (localGpuTargets != null) {
+      inherit localGpuTargets;
+      gpuArchSuffix = "-" + (builtins.concatStringsSep "-" localGpuTargets);
+      selectGpuTargets =
+        {
+          supported ? [ ],
+        }:
+        if supported == [ ] then localGpuTargets else lib.lists.intersectLists localGpuTargets supported;
     };
-  };
 
   meta = with lib; {
     description = "AMD Common Language Runtime for hipamd, opencl, and rocclr";
@@ -215,8 +281,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = with maintainers; [ lovesegfault ] ++ teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

pkgs/development/rocm-modules/6/composable_kernel/base.nix

@@ -0,0 +1,169 @@
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  rocmUpdateScript,
+  cmake,
+  rocm-cmake,
+  rocm-merged-llvm,
+  clr,
+  rocminfo,
+  hipify,
+  gitMinimal,
+  gtest,
+  zstd,
+  buildTests ? false,
+  buildExamples ? false,
+  gpuTargets ? (
+    clr.localGpuTargets or [
+      "gfx900"
+      "gfx906"
+      "gfx908"
+      "gfx90a"
+      "gfx942"
+      "gfx1030"
+      "gfx1100"
+      "gfx1101"
+      "gfx1102"
+    ]
+  ),
+}:
+
+stdenv.mkDerivation (finalAttrs: {
+  preBuild = ''
+    echo "This derivation isn't intended to be built directly and only exists to be overridden and built in chunks";
+    exit 1
+  '';
+
+  pname = "composable_kernel_base";
+  # Picked this version over 6.3 because much easier to get to build
+  # and it matches the version torch 2.6 wants
+  version = "6.4.0-unstable-20241220";
+
+  outputs =
+    [
+      "out"
+    ]
+    ++ lib.optionals buildTests [
+      "test"
+    ]
+    ++ lib.optionals buildExamples [
+      "example"
+    ];
+
+  src = fetchFromGitHub {
+    owner = "ROCm";
+    repo = "composable_kernel";
+    rev = "07339c738396ebeae57374771ded4dcf11bddf1e";
+    hash = "sha256-EvEBxlOpQ71BF57VW79WBo/cdxAwTKFXFMiYKyGyyEs=";
+  };
+
+  nativeBuildInputs = [
+    # Deliberately not using ninja
+    # because we're jankily composing build outputs from multiple drvs
+    # ninja won't believe they're up to date
+    gitMinimal
+    cmake
+    rocminfo
+    clr
+    hipify
+    zstd
+  ];
+
+  buildInputs = [
+    rocm-cmake
+    clr
+    zstd
+  ];
+
+  strictDeps = true;
+  enableParallelBuilding = true;
+  env.ROCM_PATH = clr;
+  env.HIP_CLANG_PATH = "${rocm-merged-llvm}/bin";
+
+  cmakeFlags =
+    [
+      "-DCMAKE_MODULE_PATH=${clr}/hip/cmake"
+      "-DCMAKE_BUILD_TYPE=Release"
+      "-DCMAKE_POLICY_DEFAULT_CMP0069=NEW"
+      # "-DDL_KERNELS=ON" # Not needed, slow to build
+      # CK_USE_CODEGEN Required for migraphx which uses device_gemm_multiple_d.hpp
+      # but migraphx requires an incompatible fork of CK and fails anyway
+      # "-DCK_USE_CODEGEN=ON"
+      # It might be worth skipping fp64 in future with this:
+      # "-DDTYPES=fp32;fp16;fp8;bf16;int8"
+      # Manually define CMAKE_INSTALL_<DIR>
+      # See: https://github.com/NixOS/nixpkgs/pull/197838
+      "-DCMAKE_INSTALL_BINDIR=bin"
+      "-DCMAKE_INSTALL_LIBDIR=lib"
+      "-DCMAKE_INSTALL_INCLUDEDIR=include"
+      "-DBUILD_DEV=OFF"
+      "-DROCM_PATH=${clr}"
+      "-DCMAKE_HIP_COMPILER_ROCM_ROOT=${clr}"
+
+      # FP8 can build for 908/90a but very slow build
+      # and produces unusably slow kernels that are huge
+      "-DCK_USE_FP8_ON_UNSUPPORTED_ARCH=OFF"
+    ]
+    ++ lib.optionals (gpuTargets != [ ]) [
+      # We intentionally set GPU_ARCHS and not AMD/GPU_TARGETS
+      # per readme this is required if archs are dissimilar
+      # In rocm-6.3.x not setting any arch flag worked
+      # but setting dissimilar arches always failed
+      "-DGPU_ARCHS=${lib.concatStringsSep ";" gpuTargets}"
+    ]
+    ++ lib.optionals buildTests [
+      "-DGOOGLETEST_DIR=${gtest.src}" # Custom linker names
+    ];
+
+  # No flags to build selectively it seems...
+  postPatch =
+    # Reduce configure time by preventing thousands of clang-tidy targets being added
+    # We will never call them
+    # Never build profiler
+    ''
+      substituteInPlace library/src/utility/CMakeLists.txt library/src/tensor_operation_instance/gpu/CMakeLists.txt \
+        --replace-fail clang_tidy_check '#clang_tidy_check'
+      substituteInPlace CMakeLists.txt \
+        --replace-fail "add_subdirectory(profiler)" ""
+    ''
+    # Optionally remove tests
+    + lib.optionalString (!buildTests) ''
+      substituteInPlace CMakeLists.txt \
+        --replace-fail "add_subdirectory(test)" ""
+      substituteInPlace codegen/CMakeLists.txt \
+        --replace-fail "include(ROCMTest)" ""
+    ''
+    # Optionally remove examples
+    + lib.optionalString (!buildExamples) ''
+      substituteInPlace CMakeLists.txt \
+        --replace-fail "add_subdirectory(example)" ""
+    '';
+
+  postInstall =
+    lib.optionalString buildTests ''
+      mkdir -p $test/bin
+      mv $out/bin/test_* $test/bin
+    ''
+    + lib.optionalString buildExamples ''
+      mkdir -p $example/bin
+      mv $out/bin/example_* $example/bin
+    '';
+
+  passthru.updateScript = rocmUpdateScript {
+    name = finalAttrs.pname;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
+  };
+
+  passthru.anyGfx9Target = lib.lists.any (lib.strings.hasPrefix "gfx9") gpuTargets;
+
+  meta = with lib; {
+    description = "Performance portable programming model for machine learning tensor operators";
+    homepage = "https://github.com/ROCm/composable_kernel";
+    license = with licenses; [ mit ];
+    maintainers = teams.rocm.members;
+    platforms = platforms.linux;
+    broken = true;
+  };
+})

pkgs/development/rocm-modules/6/composable_kernel/ck4inductor.nix

@@ -0,0 +1,43 @@
+{
+  buildPythonPackage,
+  python,
+  composable_kernel,
+  lib,
+  setuptools,
+  setuptools-scm,
+  rocm-merged-llvm,
+}:
+buildPythonPackage {
+  pyproject = true;
+  pname = "ck4inductor";
+  build-system = [
+    setuptools
+    setuptools-scm
+  ];
+  version = "6.4.0";
+  inherit (composable_kernel) src;
+  pythonImportsCheck = [
+    "ck4inductor"
+    "ck4inductor.universal_gemm.gen_instances"
+    "ck4inductor.universal_gemm.gen_instances"
+    "ck4inductor.universal_gemm.op"
+  ];
+  propagatedBuildInputs = [
+    # At runtime will fail to compile anything with ck4inductor without this
+    # can't easily use in checks phase because most of the compiler machinery is in torch
+    rocm-merged-llvm
+  ];
+  checkPhase = ''
+    if [ ! -d "$out/${python.sitePackages}/ck4inductor" ]; then
+      echo "ck4inductor isn't at the expected location in $out/${python.sitePackages}/ck4inductor"
+      exit 1
+    fi
+  '';
+  meta = with lib; {
+    description = "pytorch inductor backend which uses composable_kernel universal GEMM implementations";
+    homepage = "https://github.com/ROCm/composable_kernel";
+    license = with licenses; [ mit ];
+    maintainers = teams.rocm.members;
+    platforms = platforms.linux;
+  };
+}

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

@@ -1,112 +1,241 @@
 {
   lib,
-  stdenv,
-  fetchFromGitHub,
-  rocmUpdateScript,
-  cmake,
-  rocm-cmake,
   clr,
-  openmp,
-  clang-tools-extra,
-  git,
-  gtest,
-  zstd,
-  buildTests ? false,
-  buildExamples ? false,
-  gpuTargets ? [ ], # gpuTargets = [ "gfx803" "gfx900" "gfx1030" ... ]
+  composable_kernel_base,
 }:
 
-stdenv.mkDerivation (finalAttrs: {
-  pname = "composable_kernel";
-  version = "6.0.2";
-
-  outputs =
-    [
-      "out"
-    ]
-    ++ lib.optionals buildTests [
-      "test"
-    ]
-    ++ lib.optionals buildExamples [
-      "example"
-    ];
-
-  src = fetchFromGitHub {
-    owner = "ROCm";
-    repo = "composable_kernel";
-    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-NCqMganmNyQfz3X+KQOrfrimnrgd3HbAGK5DeC4+J+o=";
+let
+  parts = {
+    _mha = {
+      # mha takes ~3hrs on 64 cores on an EPYC milan system at ~2.5GHz
+      # big-parallel builders are one gen newer and clocked ~30% higher but only 24 cores
+      # Should be <10h timeout but might be cutting it close
+      # TODO: work out how to split this into smaller chunks instead of all 3k mha instances together
+      # mha_0,1,2, search ninja target file for the individual instances, split by the index?
+      # TODO: can we prune the generated instances down to only what in practice are used with popular models
+      # when using flash-attention + MHA kernels?
+      targets = [
+        "device_mha_instance"
+      ];
+      requiredSystemFeatures = [ "big-parallel" ];
+      extraCmakeFlags = [ "-DHIP_CLANG_NUM_PARALLEL_JOBS=2" ];
+    };
+    gemm_multiply_multiply = {
+      targets = [
+        "device_gemm_multiply_multiply_instance"
+      ];
+      requiredSystemFeatures = [ "big-parallel" ];
+      extraCmakeFlags = [ "-DHIP_CLANG_NUM_PARALLEL_JOBS=2" ];
+    };
+    grouped_conv = {
+      targets = [
+        "device_grouped_conv1d_bwd_weight_instance"
+        "device_grouped_conv2d_bwd_data_instance"
+        "device_grouped_conv2d_bwd_weight_instance"
+        "device_grouped_conv1d_fwd_instance"
+        "device_grouped_conv2d_fwd_instance"
+        "device_grouped_conv2d_fwd_dynamic_op_instance"
+      ];
+    };
+    grouped_conv_bwd_3d = {
+      targets = [
+        "device_grouped_conv3d_bwd_data_instance"
+        "device_grouped_conv3d_bwd_data_bilinear_instance"
+        "device_grouped_conv3d_bwd_data_scale_instance"
+        "device_grouped_conv3d_bwd_weight_instance"
+        "device_grouped_conv3d_bwd_weight_bilinear_instance"
+        "device_grouped_conv3d_bwd_weight_scale_instance"
+      ];
+    };
+    grouped_conv_fwd_3d = {
+      targets = [
+        "device_grouped_conv3d_fwd_instance"
+        "device_grouped_conv3d_fwd_bilinear_instance"
+        "device_grouped_conv3d_fwd_convinvscale_instance"
+        "device_grouped_conv3d_fwd_convscale_instance"
+        "device_grouped_conv3d_fwd_convscale_add_instance"
+        "device_grouped_conv3d_fwd_convscale_relu_instance"
+        "device_grouped_conv3d_fwd_dynamic_op_instance"
+        "device_grouped_conv3d_fwd_scale_instance"
+        "device_grouped_conv3d_fwd_scaleadd_ab_instance"
+        "device_grouped_conv3d_fwd_scaleadd_scaleadd_relu_instance"
+      ];
+    };
+    batched_gemm = {
+      targets = [
+        "device_batched_gemm_instance"
+        "device_batched_gemm_add_relu_gemm_add_instance"
+        "device_batched_gemm_bias_permute_instance"
+        "device_batched_gemm_gemm_instance"
+        "device_batched_gemm_reduce_instance"
+        "device_batched_gemm_softmax_gemm_instance"
+        "device_batched_gemm_softmax_gemm_permute_instance"
+        "device_grouped_gemm_instance"
+        "device_grouped_gemm_bias_instance"
+        "device_grouped_gemm_fastgelu_instance"
+        "device_grouped_gemm_fixed_nk_instance"
+        "device_grouped_gemm_fixed_nk_multi_abd_instance"
+        "device_grouped_gemm_tile_loop_instance"
+      ];
+    };
+    gemm_universal = {
+      targets = [
+        "device_gemm_universal_instance"
+        "device_gemm_universal_batched_instance"
+        "device_gemm_universal_reduce_instance"
+        "device_gemm_universal_streamk_instance"
+      ];
+      requiredSystemFeatures = [ "big-parallel" ];
+      extraCmakeFlags = [ "-DHIP_CLANG_NUM_PARALLEL_JOBS=2" ];
+    };
+    gemm_other = {
+      targets = [
+        "device_gemm_instance"
+        "device_gemm_ab_scale_instance"
+        "device_gemm_add_instance"
+        "device_gemm_add_add_fastgelu_instance"
+        "device_gemm_add_fastgelu_instance"
+        "device_gemm_add_multiply_instance"
+        "device_gemm_add_relu_instance"
+        "device_gemm_add_relu_add_layernorm_instance"
+        "device_gemm_add_silu_instance"
+        "device_gemm_bias_add_reduce_instance"
+        "device_gemm_bilinear_instance"
+        "device_gemm_fastgelu_instance"
+        "device_gemm_multi_abd_instance"
+        "device_gemm_multiply_add_instance"
+        "device_gemm_reduce_instance"
+        "device_gemm_splitk_instance"
+        "device_gemm_streamk_instance"
+      ];
+    };
+    conv = {
+      targets = [
+        "device_conv1d_bwd_data_instance"
+        "device_conv2d_bwd_data_instance"
+        "device_conv2d_fwd_instance"
+        "device_conv2d_fwd_bias_relu_instance"
+        "device_conv2d_fwd_bias_relu_add_instance"
+        "device_conv3d_bwd_data_instance"
+      ];
+    };
+    pool = {
+      targets = [
+        "device_avg_pool2d_bwd_instance"
+        "device_avg_pool3d_bwd_instance"
+        "device_pool2d_fwd_instance"
+        "device_pool3d_fwd_instance"
+        "device_max_pool_bwd_instance"
+      ];
+    };
+    other1 = {
+      targets = [
+        "device_batchnorm_instance"
+        "device_contraction_bilinear_instance"
+        "device_contraction_scale_instance"
+        "device_elementwise_instance"
+        "device_elementwise_normalization_instance"
+        "device_normalization_bwd_data_instance"
+        "device_normalization_bwd_gamma_beta_instance"
+        "device_normalization_fwd_instance"
+      ];
+    };
+    other2 = {
+      targets = [
+        "device_column_to_image_instance"
+        "device_image_to_column_instance"
+        "device_permute_scale_instance"
+        "device_quantization_instance"
+        "device_reduce_instance"
+        "device_softmax_instance"
+        "device_transpose_instance"
+      ];
+    };
   };
+  tensorOpBuilder =
+    {
+      part,
+      targets,
+      extraCmakeFlags ? [ ],
+      requiredSystemFeatures ? [ ],
+    }:
+    composable_kernel_base.overrideAttrs (old: {
+      inherit requiredSystemFeatures;
+      pname = "composable_kernel${clr.gpuArchSuffix}-${part}";
+      makeTargets = targets;
+      preBuild = ''
+        echo "Building ${part}"
+        makeFlagsArray+=($makeTargets)
+        substituteInPlace Makefile \
+          --replace-fail '.NOTPARALLEL:' ""
+      '';
 
-  nativeBuildInputs = [
-    git
-    cmake
-    rocm-cmake
-    clr
-    clang-tools-extra
-    zstd
-  ];
+      # Compile parallelism adjusted based on available RAM
+      # Never uses less than NIX_BUILD_CORES/4, never uses more than NIX_BUILD_CORES
+      # CK uses an unusually high amount of memory per core in the build step
+      # Nix/nixpkgs doesn't really have any infra to tell it that this build is unusually memory hungry
+      # So, bodge. Otherwise you end up having to build all of ROCm with a low core limit when
+      # it's only this package that has trouble.
+      preConfigure =
+        old.preConfigure or ""
+        + ''
+          MEM_GB_TOTAL=$(awk '/MemTotal/ { printf "%d \n", $2/1024/1024 }' /proc/meminfo)
+          MEM_GB_AVAILABLE=$(awk '/MemAvailable/ { printf "%d \n", $2/1024/1024 }' /proc/meminfo)
+          APPX_GB=$((MEM_GB_AVAILABLE > MEM_GB_TOTAL ? MEM_GB_TOTAL : MEM_GB_AVAILABLE))
+          MAX_CORES=$((1 + APPX_GB/3))
+          MAX_CORES=$((MAX_CORES < NIX_BUILD_CORES/3 ? NIX_BUILD_CORES/3 : MAX_CORES))
+          export NIX_BUILD_CORES="$((NIX_BUILD_CORES > MAX_CORES ? MAX_CORES : NIX_BUILD_CORES))"
+          echo "Picked new core limit NIX_BUILD_CORES=$NIX_BUILD_CORES based on available mem: $APPX_GB GB"
+          cmakeFlagsArray+=(
+            "-DCK_PARALLEL_COMPILE_JOBS=$NIX_BUILD_CORES"
+          )
+        '';
+      cmakeFlags = old.cmakeFlags ++ extraCmakeFlags;
+      # Early exit after build phase with success, skips fixups etc
+      # Will get copied back into /build of the final CK
+      postBuild = ''
+        find . -name "*.o" -type f | while read -r file; do
+          mkdir -p "$out/$(dirname "$file")"
+          cp --reflink=auto "$file" "$out/$file"
+        done
+        exit 0
+      '';
+      meta = old.meta // {
+        broken = false;
+      };
+    });
+  composable_kernel_parts = builtins.mapAttrs (
+    part: targets: tensorOpBuilder (targets // { inherit part; })
+  ) parts;
+in
 
-  buildInputs = [ openmp ];
+composable_kernel_base.overrideAttrs (
+  finalAttrs: old: {
+    pname = "composable_kernel${clr.gpuArchSuffix}";
+    parts_dirs = builtins.attrValues composable_kernel_parts;
+    disallowedReferences = builtins.attrValues composable_kernel_parts;
+    preBuild = ''
+      for dir in $parts_dirs; do
+        find "$dir" -type f -name "*.o" | while read -r file; do
+          # Extract the relative path by removing the output directory prefix
+          rel_path="''${file#"$dir/"}"
 
-  cmakeFlags =
-    [
-      "-DCMAKE_C_COMPILER=hipcc"
-      "-DCMAKE_CXX_COMPILER=hipcc"
-    ]
-    ++ lib.optionals (gpuTargets != [ ]) [
-      "-DGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
-      "-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
-    ]
-    ++ lib.optionals buildTests [
-      "-DGOOGLETEST_DIR=${gtest.src}" # Custom linker names
-    ];
+          # Create parent directory if it doesn't exist
+          mkdir -p "$(dirname "$rel_path")"
 
-  # No flags to build selectively it seems...
-  postPatch =
-    lib.optionalString (!buildTests) ''
-      substituteInPlace CMakeLists.txt \
-        --replace "add_subdirectory(test)" ""
-    ''
-    + lib.optionalString (!buildExamples) ''
-      substituteInPlace CMakeLists.txt \
-        --replace "add_subdirectory(example)" ""
-    ''
-    + ''
-      substituteInPlace CMakeLists.txt \
-        --replace "add_subdirectory(profiler)" ""
+          # Copy the file back to its original location, give it a future timestamp
+          # so make treats it as up to date
+          cp --reflink=auto --no-preserve=all "$file" "$rel_path"
+          touch -d "now +10 hours" "$rel_path"
+        done
+      done
     '';
-
-  postInstall =
-    ''
-      zstd --rm $out/lib/libdevice_operations.a
-    ''
-    + lib.optionalString buildTests ''
-      mkdir -p $test/bin
-      mv $out/bin/test_* $test/bin
-    ''
-    + lib.optionalString buildExamples ''
-      mkdir -p $example/bin
-      mv $out/bin/example_* $example/bin
-    '';
-
-  passthru.updateScript = rocmUpdateScript {
-    name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
-  };
-
-  # Times out otherwise
-  requiredSystemFeatures = [ "big-parallel" ];
-
-  meta = with lib; {
-    description = "Performance portable programming model for machine learning tensor operators";
-    homepage = "https://github.com/ROCm/composable_kernel";
-    license = with licenses; [ mit ];
-    maintainers = teams.rocm.members;
-    platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
-  };
-})
+    passthru = old.passthru // {
+      parts = composable_kernel_parts;
+    };
+    meta = old.meta // {
+      # Builds which don't don't target any gfx9 cause cmake errors in dependent projects
+      broken = !finalAttrs.passthru.anyGfx9Target;
+    };
+  }
+)

pkgs/development/rocm-modules/6/composable_kernel/unpack.nix

@@ -1,20 +0,0 @@
-{
-  runCommandLocal,
-  composable_kernel_build,
-  zstd,
-}:
-let
-  ck = composable_kernel_build;
-in
-runCommandLocal "unpack-${ck.name}"
-  {
-    nativeBuildInputs = [ zstd ];
-    meta = ck.meta;
-  }
-  ''
-    mkdir -p $out
-    cp -r --no-preserve=mode ${ck}/* $out
-    zstd -dv --rm $out/lib/libdevice_operations.a.zst -o $out/lib/libdevice_operations.a
-    substituteInPlace $out/lib/cmake/composable_kernel/*.cmake \
-      --replace "${ck}" "$out"
-  ''

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

@@ -1,530 +1,507 @@
-{ stdenv
-, lib
-, config
-, callPackage
-, recurseIntoAttrs
-, symlinkJoin
-, fetchFromGitHub
-, cudaPackages
-, python3Packages
-, elfutils
-, boost179
-, opencv
-, ffmpeg_4
-, libjpeg_turbo
+{
+  lib,
+  config,
+  callPackage,
+  newScope,
+  recurseIntoAttrs,
+  symlinkJoin,
+  fetchFromGitHub,
+  ffmpeg_4,
+  boost179,
+  opencv,
+  libjpeg_turbo,
+  python3Packages,
+  triton-llvm,
+  openmpi,
+  rocmGpuArches ? [ ],
 }:
 
 let
-  rocmUpdateScript = callPackage ./update.nix { };
-in rec {
-  ## ROCm ##
-  llvm = recurseIntoAttrs (callPackage ./llvm/default.nix { inherit rocmUpdateScript rocm-device-libs rocm-runtime rocm-thunk clr; });
-
-  rocm-core = callPackage ./rocm-core {
-    inherit rocmUpdateScript;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocm-cmake = callPackage ./rocm-cmake {
-    inherit rocmUpdateScript;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocm-thunk = callPackage ./rocm-thunk {
-    inherit rocmUpdateScript;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocm-smi = python3Packages.callPackage ./rocm-smi {
-    inherit rocmUpdateScript;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  # Eventually will be in the LLVM repo
-  rocm-device-libs = callPackage ./rocm-device-libs {
-    inherit rocmUpdateScript rocm-cmake;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocm-runtime = callPackage ./rocm-runtime {
-    inherit rocmUpdateScript rocm-device-libs rocm-thunk;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  # Eventually will be in the LLVM repo
-  rocm-comgr = callPackage ./rocm-comgr {
-    inherit rocmUpdateScript rocm-cmake rocm-device-libs;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocminfo = callPackage ./rocminfo {
-    inherit rocmUpdateScript rocm-cmake rocm-runtime;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  clang-ocl = callPackage ./clang-ocl {
-    inherit rocmUpdateScript rocm-cmake rocm-device-libs;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  # Unfree
-  hsa-amd-aqlprofile-bin = callPackage ./hsa-amd-aqlprofile-bin {
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  # Broken, too many errors
-  rdc = callPackage ./rdc {
-    inherit rocmUpdateScript rocm-smi rocm-runtime stdenv;
-    # stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocm-docs-core = python3Packages.callPackage ./rocm-docs-core { inherit stdenv; };
-
-  hip-common = callPackage ./hip-common {
-    inherit rocmUpdateScript;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  # Eventually will be in the LLVM repo
-  hipcc = callPackage ./hipcc {
-    inherit rocmUpdateScript;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  # Replaces hip, opencl-runtime, and rocclr
-  clr = callPackage ./clr {
-    inherit rocmUpdateScript hip-common hipcc rocm-device-libs rocm-comgr rocm-runtime roctracer rocminfo rocm-smi;
-    inherit (llvm) clang;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  hipify = callPackage ./hipify {
-    inherit rocmUpdateScript;
-    inherit (llvm) clang;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  # Needs GCC
-  rocprofiler = callPackage ./rocprofiler {
-    inherit rocmUpdateScript clr rocm-core rocm-thunk rocm-device-libs roctracer rocdbgapi rocm-smi hsa-amd-aqlprofile-bin stdenv;
-    inherit (llvm) clang;
-  };
-
-  # Needs GCC
-  roctracer = callPackage ./roctracer {
-    inherit rocmUpdateScript rocm-device-libs rocm-runtime clr stdenv;
-  };
-
-  rocgdb = callPackage ./rocgdb {
-    inherit rocmUpdateScript rocdbgapi;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocdbgapi = callPackage ./rocdbgapi {
-    inherit rocmUpdateScript rocm-cmake rocm-comgr rocm-runtime;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocr-debug-agent = callPackage ./rocr-debug-agent {
-    inherit rocmUpdateScript clr rocdbgapi;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocprim = callPackage ./rocprim {
-    inherit rocmUpdateScript rocm-cmake clr;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocsparse = callPackage ./rocsparse {
-    inherit rocmUpdateScript rocm-cmake rocprim clr;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocthrust = callPackage ./rocthrust {
-    inherit rocmUpdateScript rocm-cmake rocprim clr;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocrand = callPackage ./rocrand {
-    inherit rocmUpdateScript rocm-cmake clr;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  hiprand = callPackage ./hiprand {
-    inherit rocmUpdateScript rocm-cmake clr rocrand;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocfft = callPackage ./rocfft {
-    inherit rocmUpdateScript rocm-cmake rocrand rocfft clr;
-    inherit (llvm) openmp;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rccl = callPackage ./rccl {
-    inherit rocmUpdateScript rocm-cmake rocm-smi clr hipify;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  hipcub = callPackage ./hipcub {
-    inherit rocmUpdateScript rocm-cmake rocprim clr;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  hipsparse = callPackage ./hipsparse {
-    inherit rocmUpdateScript rocm-cmake rocsparse clr;
-    inherit (llvm) openmp;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  hipfort = callPackage ./hipfort {
-    inherit rocmUpdateScript rocm-cmake;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  hipfft = callPackage ./hipfft {
-    inherit rocmUpdateScript rocm-cmake rocfft clr;
-    inherit (llvm) openmp;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  tensile = python3Packages.callPackage ./tensile {
-    inherit rocmUpdateScript rocminfo;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocblas = callPackage ./rocblas {
-    inherit rocmUpdateScript rocm-cmake clr tensile;
-    inherit (llvm) openmp;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocsolver = callPackage ./rocsolver {
-    inherit rocmUpdateScript rocm-cmake rocblas rocsparse clr;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocwmma = callPackage ./rocwmma {
-    inherit rocmUpdateScript rocm-cmake rocm-smi rocblas clr;
-    inherit (llvm) openmp;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocalution = callPackage ./rocalution {
-    inherit rocmUpdateScript rocm-cmake rocprim rocsparse rocrand rocblas clr;
-    inherit (llvm) openmp;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocmlir = callPackage ./rocmlir {
-    inherit rocmUpdateScript rocm-cmake rocminfo clr;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rocmlir-rock = rocmlir.override {
-    buildRockCompiler = true;
-  };
-
-  hipsolver = callPackage ./hipsolver {
-    inherit rocmUpdateScript rocm-cmake rocblas rocsolver clr;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  hipblas = callPackage ./hipblas {
-    inherit rocmUpdateScript rocm-cmake rocblas rocsolver clr;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  # hipBlasLt - Very broken with Tensile at the moment, only supports GFX9
-  # hipTensor - Only supports GFX9
-
-  composable_kernel = callPackage ./composable_kernel/unpack.nix {
-    composable_kernel_build = callPackage ./composable_kernel {
-      inherit rocmUpdateScript rocm-cmake clr;
-      inherit (llvm) openmp clang-tools-extra;
+  outer = lib.makeScope newScope (
+    self:
+    let
+      inherit (self) llvm;
+      pyPackages = python3Packages;
+      openmpi-orig = openmpi;
+    in
+    {
+      inherit rocmGpuArches;
+      buildTests = false;
+      buildBenchmarks = false;
       stdenv = llvm.rocmClangStdenv;
-    };
-  };
 
-  half = callPackage ./half {
-    inherit rocmUpdateScript rocm-cmake;
-    stdenv = llvm.rocmClangStdenv;
-  };
+      rocmPath = self.callPackage ./rocm-path { };
+      rocmUpdateScript = self.callPackage ./update.nix { };
 
-  miopen = callPackage ./miopen {
-    inherit rocmUpdateScript rocm-cmake rocblas clang-ocl composable_kernel rocm-comgr clr rocm-docs-core half roctracer;
-    inherit (llvm) clang-tools-extra;
-    stdenv = llvm.rocmClangStdenv;
-    rocmlir = rocmlir-rock;
-    boost = boost179.override { enableStatic = true; };
-  };
+      ## ROCm ##
+      llvm = recurseIntoAttrs (
+        callPackage ./llvm/default.nix {
+          inherit (self) rocm-device-libs rocm-runtime;
+        }
+      );
+      inherit (self.llvm) rocm-merged-llvm clang openmp;
 
-  miopen-hip = miopen;
-
-  migraphx = callPackage ./migraphx {
-    inherit rocmUpdateScript rocm-cmake rocblas composable_kernel miopen clr half rocm-device-libs;
-    inherit (llvm) openmp clang-tools-extra;
-    stdenv = llvm.rocmClangStdenv;
-    rocmlir = rocmlir-rock;
-  };
-
-  rpp = callPackage ./rpp {
-    inherit rocmUpdateScript rocm-cmake rocm-docs-core clr half;
-    inherit (llvm) openmp;
-    stdenv = llvm.rocmClangStdenv;
-  };
-
-  rpp-hip = rpp.override {
-    useOpenCL = false;
-    useCPU = false;
-  };
-
-  rpp-opencl = rpp.override {
-    useOpenCL = true;
-    useCPU = false;
-  };
-
-  rpp-cpu = rpp.override {
-    useOpenCL = false;
-    useCPU = true;
-  };
-
-  mivisionx = callPackage ./mivisionx {
-    inherit rocmUpdateScript rocm-cmake rocm-device-libs clr rpp rocblas miopen migraphx half rocm-docs-core;
-    inherit (llvm) clang openmp;
-    opencv = opencv.override { enablePython = true; };
-    ffmpeg = ffmpeg_4;
-    stdenv = llvm.rocmClangStdenv;
-
-    # Unfortunately, rocAL needs a custom libjpeg-turbo until further notice
-    # See: https://github.com/ROCm/MIVisionX/issues/1051
-    libjpeg_turbo = libjpeg_turbo.overrideAttrs {
-      version = "2.0.6.1";
-
-      src = fetchFromGitHub {
-        owner = "rrawther";
-        repo = "libjpeg-turbo";
-        rev = "640d7ee1917fcd3b6a5271aa6cf4576bccc7c5fb";
-        sha256 = "sha256-T52whJ7nZi8jerJaZtYInC2YDN0QM+9tUDqiNr6IsNY=";
+      rocm-core = self.callPackage ./rocm-core { };
+      amdsmi = pyPackages.callPackage ./amdsmi {
+        inherit (self) rocmUpdateScript;
       };
 
-      # overwrite all patches, since patches for newer version do not apply
-      patches = [ ./0001-Compile-transupp.c-as-part-of-the-library.patch ];
-    };
-  };
+      rocm-cmake = self.callPackage ./rocm-cmake { };
 
-  mivisionx-hip = mivisionx.override {
-    rpp = rpp-hip;
-    useOpenCL = false;
-    useCPU = false;
-  };
+      rocm-smi = pyPackages.callPackage ./rocm-smi {
+        inherit (self) rocmUpdateScript;
+      };
 
-  mivisionx-cpu = mivisionx.override {
-    rpp = rpp-cpu;
-    useOpenCL = false;
-    useCPU = true;
-  };
+      rocm-device-libs = self.callPackage ./rocm-device-libs {
+        inherit (llvm) rocm-merged-llvm;
+      };
 
-  ## Meta ##
-  # Emulate common ROCm meta layout
-  # These are mainly for users. I strongly suggest NOT using these in nixpkgs derivations
-  # Don't put these into `propagatedBuildInputs` unless you want PATH/PYTHONPATH issues!
-  # See: https://rocm.docs.amd.com/en/docs-5.7.1/_images/image.004.png
-  # See: https://rocm.docs.amd.com/en/docs-5.7.1/deploy/linux/os-native/package_manager_integration.html
-  meta = rec {
-    rocm-developer-tools = symlinkJoin {
-      name = "rocm-developer-tools-meta";
+      rocm-runtime = self.callPackage ./rocm-runtime {
+        inherit (llvm) rocm-merged-llvm;
+      };
 
-      paths = [
-        hsa-amd-aqlprofile-bin
-        rocm-core
-        rocr-debug-agent
-        roctracer
-        rocdbgapi
-        rocprofiler
-        rocgdb
-        rocm-language-runtime
-      ];
-    };
+      rocm-comgr = self.callPackage ./rocm-comgr {
+        inherit (llvm) rocm-merged-llvm;
+      };
 
-    rocm-ml-sdk = symlinkJoin {
-      name = "rocm-ml-sdk-meta";
+      rocminfo = self.callPackage ./rocminfo { };
 
-      paths = [
-        rocm-core
-        miopen-hip
-        rocm-hip-sdk
-        rocm-ml-libraries
-      ];
-    };
+      # Unfree
+      hsa-amd-aqlprofile-bin = self.callPackage ./hsa-amd-aqlprofile-bin { };
 
-    rocm-ml-libraries = symlinkJoin {
-      name = "rocm-ml-libraries-meta";
+      rdc = self.callPackage ./rdc { };
 
-      paths = [
-        llvm.clang
-        llvm.mlir
-        llvm.openmp
-        rocm-core
-        miopen-hip
-        rocm-hip-libraries
-      ];
-    };
+      rocm-docs-core = python3Packages.callPackage ./rocm-docs-core { };
 
-    rocm-hip-sdk = symlinkJoin {
-      name = "rocm-hip-sdk-meta";
+      hip-common = self.callPackage ./hip-common { };
 
-      paths = [
-        rocprim
-        rocalution
-        hipfft
-        rocm-core
-        hipcub
-        hipblas
-        rocrand
-        rocfft
-        rocsparse
-        rccl
-        rocthrust
-        rocblas
-        hipsparse
-        hipfort
-        rocwmma
-        hipsolver
-        rocsolver
-        rocm-hip-libraries
-        rocm-hip-runtime-devel
-      ];
-    };
+      # Eventually will be in the LLVM repo
+      hipcc = self.callPackage ./hipcc {
+        inherit (llvm) rocm-merged-llvm;
+      };
 
-    rocm-hip-libraries = symlinkJoin {
-      name = "rocm-hip-libraries-meta";
+      # Replaces hip, opencl-runtime, and rocclr
+      clr = self.callPackage ./clr { };
 
-      paths = [
-        rocblas
-        hipfort
-        rocm-core
-        rocsolver
-        rocalution
-        rocrand
-        hipblas
-        rocfft
-        hipfft
-        rccl
-        rocsparse
-        hipsparse
-        hipsolver
-        rocm-hip-runtime
-      ];
-    };
+      aotriton = self.callPackage ./aotriton { };
 
-    rocm-openmp-sdk = symlinkJoin {
-      name = "rocm-openmp-sdk-meta";
+      hipify = self.callPackage ./hipify {
+        inherit (llvm)
+          clang
+          rocm-merged-llvm
+          ;
+      };
 
-      paths = [
-        rocm-core
-        llvm.clang
-        llvm.mlir
-        llvm.openmp # openmp-extras-devel (https://github.com/ROCm/aomp)
-        rocm-language-runtime
-      ];
-    };
+      # hsakmt was merged into rocm-runtime
+      hsakmt = self.rocm-runtime;
 
-    rocm-opencl-sdk = symlinkJoin {
-      name = "rocm-opencl-sdk-meta";
+      rocprofiler = self.callPackage ./rocprofiler {
+        inherit (llvm) clang;
+      };
+      rocprofiler-register = self.callPackage ./rocprofiler-register {
+        inherit (llvm) clang;
+      };
 
-      paths = [
-        rocm-core
-        rocm-runtime
-        clr
-        clr.icd
-        rocm-thunk
-        rocm-opencl-runtime
-      ];
-    };
+      # Needs GCC
+      roctracer = self.callPackage ./roctracer { };
 
-    rocm-opencl-runtime = symlinkJoin {
-      name = "rocm-opencl-runtime-meta";
+      rocgdb = self.callPackage ./rocgdb { };
 
-      paths = [
-        rocm-core
-        clr
-        clr.icd
-        rocm-language-runtime
-      ];
-    };
+      rocdbgapi = self.callPackage ./rocdbgapi { };
 
-    rocm-hip-runtime-devel = symlinkJoin {
-      name = "rocm-hip-runtime-devel-meta";
+      rocr-debug-agent = self.callPackage ./rocr-debug-agent { };
 
-      paths = [
-        clr
-        rocm-core
-        hipify
-        rocm-cmake
-        llvm.clang
-        llvm.mlir
-        llvm.openmp
-        rocm-thunk
-        rocm-runtime
-        rocm-hip-runtime
-      ];
-    };
+      rocprim = self.callPackage ./rocprim { };
 
-    rocm-hip-runtime = symlinkJoin {
-      name = "rocm-hip-runtime-meta";
+      rocsparse = self.callPackage ./rocsparse { };
 
-      paths = [
-        rocm-core
-        rocminfo
-        clr
-        rocm-language-runtime
-      ];
-    };
+      rocthrust = self.callPackage ./rocthrust { };
 
-    rocm-language-runtime = symlinkJoin {
-      name = "rocm-language-runtime-meta";
+      rocrand = self.callPackage ./rocrand { };
 
-      paths = [
-        rocm-runtime
-        rocm-core
-        rocm-comgr
-        llvm.openmp # openmp-extras-runtime (https://github.com/ROCm/aomp)
-      ];
-    };
+      hiprand = self.callPackage ./hiprand { };
 
-    rocm-all = symlinkJoin {
-      name = "rocm-all-meta";
+      rocfft = self.callPackage ./rocfft { };
 
-      paths = [
-        rocm-developer-tools
-        rocm-ml-sdk
-        rocm-ml-libraries
-        rocm-hip-sdk
-        rocm-hip-libraries
-        rocm-openmp-sdk
-        rocm-opencl-sdk
-        rocm-opencl-runtime
-        rocm-hip-runtime-devel
-        rocm-hip-runtime
-        rocm-language-runtime
-      ];
-    };
-  };
-} // lib.optionalAttrs config.allowAliases {
-  miopengemm= throw ''
-    'miopengemm' has been deprecated.
-    It is still available for some time as part of rocmPackages_5.
-  ''; # Added 2024-3-3
+      mscclpp = self.callPackage ./mscclpp { };
 
-  miopen-opencl= throw ''
-    'miopen-opencl' has been deprecated.
-    It is still available for some time as part of rocmPackages_5.
-  ''; # Added 2024-3-3
+      rccl = self.callPackage ./rccl { };
 
-  mivisionx-opencl = throw ''
-    'mivisionx-opencl' has been deprecated.
-    Other versions of mivisionx are still available.
-    It is also still available for some time as part of rocmPackages_5.
-  ''; # Added 2024-3-24
+      # RCCL with sanitizers and tests
+      # Can't have with sanitizer build as dep of other packages without
+      # runtime crashes due to ASAN not loading first
+      rccl-tests = self.callPackage ./rccl {
+        buildTests = true;
+      };
+
+      hipcub = self.callPackage ./hipcub { };
+
+      hipsparse = self.callPackage ./hipsparse { };
+
+      hipfort = self.callPackage ./hipfort { };
+
+      hipfft = self.callPackage ./hipfft { };
+
+      tensile = pyPackages.callPackage ./tensile {
+        inherit (self)
+          rocmUpdateScript
+          clr
+          ;
+      };
+
+      rocblas = self.callPackage ./rocblas {
+        buildTests = true;
+        buildBenchmarks = true;
+      };
+
+      rocsolver = self.callPackage ./rocsolver { };
+
+      rocwmma = self.callPackage ./rocwmma { };
+
+      rocalution = self.callPackage ./rocalution { };
+
+      rocmlir-rock = self.callPackage ./rocmlir {
+        buildRockCompiler = true;
+      };
+      rocmlir = self.rocmlir-rock;
+
+      hipsolver = self.callPackage ./hipsolver { };
+
+      hipblas-common = self.callPackage ./hipblas-common { };
+
+      hipblas = self.callPackage ./hipblas { };
+
+      hipblaslt = self.callPackage ./hipblaslt { };
+
+      # hipTensor - Only supports GFX9
+
+      composable_kernel_base = self.callPackage ./composable_kernel/base.nix { };
+      composable_kernel = self.callPackage ./composable_kernel { };
+
+      ck4inductor = pyPackages.callPackage ./composable_kernel/ck4inductor.nix {
+        inherit (self) composable_kernel;
+        inherit (llvm) rocm-merged-llvm;
+      };
+
+      half = self.callPackage ./half { };
+
+      miopen = self.callPackage ./miopen {
+        boost = boost179.override { enableStatic = true; };
+      };
+
+      miopen-hip = self.miopen;
+
+      migraphx = self.callPackage ./migraphx { };
+
+      rpp = self.callPackage ./rpp { };
+
+      rpp-hip = self.rpp.override {
+        useOpenCL = false;
+        useCPU = false;
+      };
+
+      rpp-opencl = self.rpp.override {
+        useOpenCL = true;
+        useCPU = false;
+      };
+
+      rpp-cpu = self.rpp.override {
+        useOpenCL = false;
+        useCPU = true;
+      };
+
+      mivisionx = self.callPackage ./mivisionx {
+        opencv = opencv.override { enablePython = true; };
+        # TODO: Remove this pin in ROCm 6.4+
+        # FFMPEG support was improved in https://github.com/ROCm/MIVisionX/pull/1460
+        ffmpeg = ffmpeg_4;
+        # Unfortunately, rocAL needs a custom libjpeg-turbo until further notice
+        # See: https://github.com/ROCm/MIVisionX/issues/1051
+        libjpeg_turbo = libjpeg_turbo.overrideAttrs {
+          version = "2.0.6.1";
+          src = fetchFromGitHub {
+            owner = "rrawther";
+            repo = "libjpeg-turbo";
+            rev = "640d7ee1917fcd3b6a5271aa6cf4576bccc7c5fb";
+            sha256 = "sha256-T52whJ7nZi8jerJaZtYInC2YDN0QM+9tUDqiNr6IsNY=";
+          };
+          # overwrite all patches, since patches for newer version do not apply
+          patches = [ ./0001-Compile-transupp.c-as-part-of-the-library.patch ];
+        };
+      };
+
+      mivisionx-hip = self.mivisionx.override {
+        rpp = self.rpp-hip;
+        useOpenCL = false;
+        useCPU = false;
+      };
+
+      mivisionx-cpu = self.mivisionx.override {
+        rpp = self.rpp-cpu;
+        useOpenCL = false;
+        useCPU = true;
+      };
+
+      # Even if config.rocmSupport is false we need rocmSupport true
+      # version of ucc/ucx in openmpi in this package set
+      openmpi = openmpi-orig.override (
+        prev:
+        let
+          ucx = prev.ucx.override {
+            enableCuda = false;
+            enableRocm = true;
+          };
+        in
+        {
+          inherit ucx;
+          ucc = prev.ucc.override {
+            enableCuda = false;
+            inherit ucx;
+          };
+        }
+      );
+      mpi = self.openmpi;
+
+      triton-llvm = triton-llvm.overrideAttrs {
+        src = fetchFromGitHub {
+          owner = "llvm";
+          repo = "llvm-project";
+          # make sure this matches triton llvm rel branch hash for now
+          # https://github.com/triton-lang/triton/blob/release/3.2.x/cmake/llvm-hash.txt
+          rev = "86b69c31642e98f8357df62c09d118ad1da4e16a";
+          hash = "sha256-W/mQwaLGx6/rIBjdzUTIbWrvGjdh7m4s15f70fQ1/hE=";
+        };
+        pname = "triton-llvm-rocm";
+        patches = [ ]; # FIXME: https://github.com/llvm/llvm-project//commit/84837e3cc1cf17ed71580e3ea38299ed2bfaa5f6.patch doesn't apply, may need to rebase
+      };
+
+      triton = pyPackages.callPackage ./triton { rocmPackages = self; };
+
+      ## Meta ##
+      # Emulate common ROCm meta layout
+      # These are mainly for users. I strongly suggest NOT using these in nixpkgs derivations
+      # Don't put these into `propagatedBuildInputs` unless you want PATH/PYTHONPATH issues!
+      # See: https://rocm.docs.amd.com/en/docs-5.7.1/_images/image.004.png
+      # See: https://rocm.docs.amd.com/en/docs-5.7.1/deploy/linux/os-native/package_manager_integration.html
+      meta = with self; rec {
+        rocm-developer-tools = symlinkJoin {
+          name = "rocm-developer-tools-meta";
+          paths = [
+            hsa-amd-aqlprofile-bin
+            rocm-core
+            rocr-debug-agent
+            roctracer
+            rocdbgapi
+            rocprofiler
+            rocgdb
+            rocm-language-runtime
+          ];
+        };
+        rocm-ml-sdk = symlinkJoin {
+          name = "rocm-ml-sdk-meta";
+          paths = [
+            rocm-core
+            miopen-hip
+            rocm-hip-sdk
+            rocm-ml-libraries
+          ];
+        };
+        rocm-ml-libraries = symlinkJoin {
+          name = "rocm-ml-libraries-meta";
+          paths = [
+            llvm.clang
+            llvm.mlir
+            llvm.openmp
+            rocm-core
+            miopen-hip
+            rocm-hip-libraries
+          ];
+        };
+        rocm-hip-sdk = symlinkJoin {
+          name = "rocm-hip-sdk-meta";
+          paths = [
+            rocprim
+            rocalution
+            hipfft
+            rocm-core
+            hipcub
+            hipblas
+            hipblaslt
+            rocrand
+            rocfft
+            rocsparse
+            rccl
+            rocthrust
+            rocblas
+            hipsparse
+            hipfort
+            rocwmma
+            hipsolver
+            rocsolver
+            rocm-hip-libraries
+            rocm-hip-runtime-devel
+          ];
+        };
+        rocm-hip-libraries = symlinkJoin {
+          name = "rocm-hip-libraries-meta";
+          paths = [
+            rocblas
+            hipfort
+            rocm-core
+            rocsolver
+            rocalution
+            rocrand
+            hipblas
+            hipblaslt
+            rocfft
+            hipfft
+            rccl
+            rocsparse
+            hipsparse
+            hipsolver
+            rocm-hip-runtime
+          ];
+        };
+        rocm-openmp-sdk = symlinkJoin {
+          name = "rocm-openmp-sdk-meta";
+          paths = [
+            rocm-core
+            llvm.clang
+            llvm.mlir
+            llvm.openmp # openmp-extras-devel (https://github.com/ROCm/aomp)
+            rocm-language-runtime
+          ];
+        };
+        rocm-opencl-sdk = symlinkJoin {
+          name = "rocm-opencl-sdk-meta";
+          paths = [
+            rocm-core
+            rocm-runtime
+            clr
+            clr.icd
+            rocm-opencl-runtime
+          ];
+        };
+        rocm-opencl-runtime = symlinkJoin {
+          name = "rocm-opencl-runtime-meta";
+          paths = [
+            rocm-core
+            clr
+            clr.icd
+            rocm-language-runtime
+          ];
+        };
+        rocm-hip-runtime-devel = symlinkJoin {
+          name = "rocm-hip-runtime-devel-meta";
+          paths = [
+            clr
+            rocm-core
+            hipify
+            rocm-cmake
+            llvm.clang
+            llvm.mlir
+            llvm.openmp
+            rocm-runtime
+            rocm-hip-runtime
+          ];
+        };
+        rocm-hip-runtime = symlinkJoin {
+          name = "rocm-hip-runtime-meta";
+          paths = [
+            rocm-core
+            rocminfo
+            clr
+            rocm-language-runtime
+          ];
+        };
+        rocm-language-runtime = symlinkJoin {
+          name = "rocm-language-runtime-meta";
+          paths = [
+            rocm-runtime
+            rocm-core
+            rocm-comgr
+            llvm.openmp # openmp-extras-runtime (https://github.com/ROCm/aomp)
+          ];
+        };
+        rocm-all = symlinkJoin {
+          name = "rocm-all-meta";
+          paths = [
+            rocm-developer-tools
+            rocm-ml-sdk
+            rocm-ml-libraries
+            rocm-hip-sdk
+            rocm-hip-libraries
+            rocm-openmp-sdk
+            rocm-opencl-sdk
+            rocm-opencl-runtime
+            rocm-hip-runtime-devel
+            rocm-hip-runtime
+            rocm-language-runtime
+          ];
+        };
+      };
+
+      rocm-tests = self.callPackage ./rocm-tests {
+        rocmPackages = self;
+      };
+    }
+    // lib.optionalAttrs config.allowAliases {
+      rocm-thunk = throw ''
+        'rocm-thunk' has been removed. It's now part of the ROCm runtime.
+      ''; # Added 2025-3-16
+
+      clang-ocl = throw ''
+        'clang-ocl' has been deprecated upstream. Use ROCm's clang directly.
+      ''; # Added 2025-3-16
+
+      miopengemm = throw ''
+        'miopengemm' has been deprecated.
+        It is still available for some time as part of rocmPackages_5.
+      ''; # Added 2024-3-3
+
+      miopen-opencl = throw ''
+        'miopen-opencl' has been deprecated.
+        It is still available for some time as part of rocmPackages_5.
+      ''; # Added 2024-3-3
+
+      mivisionx-opencl = throw ''
+        'mivisionx-opencl' has been deprecated.
+        Other versions of mivisionx are still available.
+        It is also still available for some time as part of rocmPackages_5.
+      ''; # Added 2024-3-24
+    }
+  );
+  scopeForArches =
+    arches:
+    outer.overrideScope (
+      _final: prev: {
+        clr = prev.clr.override {
+          localGpuTargets = arches;
+        };
+      }
+    );
+in
+outer
+// builtins.listToAttrs (
+  builtins.map (arch: {
+    name = arch;
+    value = scopeForArches [ arch ];
+  }) outer.clr.gpuTargets
+)
+// {
+  gfx9 = scopeForArches [
+    "gfx906"
+    "gfx908"
+    "gfx90a"
+    "gfx942"
+  ];
+  gfx10 = scopeForArches [
+    "gfx1010"
+    "gfx1030"
+  ];
+  gfx11 = scopeForArches [
+    "gfx1100"
+    "gfx1101"
+    "gfx1102"
+  ];
 }

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

@@ -1,20 +1,21 @@
-{ lib
-, stdenv
-, fetchFromGitHub
-, rocmUpdateScript
-, cmake
-, rocm-cmake
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  rocmUpdateScript,
+  cmake,
+  rocm-cmake,
 }:
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "half";
-  version = "6.0.2";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "half";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-wvl8ny7pbY9hUGGtJ70R7/4YIsahgI7qcVzUnxmUfZM=";
+    hash = "sha256-H8Ogm4nxaxDB0WHx+KhRjUO3vzp3AwCqrIQ6k8R+xkc=";
   };
 
   nativeBuildInputs = [
@@ -24,8 +25,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -34,6 +35,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.unix;
-    broken = versions.minor finalAttrs.version != versions.minor stdenv.cc.version || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -1,18 +1,19 @@
-{ lib
-, stdenv
-, fetchFromGitHub
-, rocmUpdateScript
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  rocmUpdateScript,
 }:
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "hip-common";
-  version = "6.0.2";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "HIP";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-51u3By0R4LKoWiklNacFP6HILL845jxpN6FD7rQB+zQ=";
+    hash = "sha256-eKDbuG57KCg/nZpy+RcDsjyw0pP68qO9K/3ZpSPuMNw=";
   };
 
   dontConfigure = true;
@@ -29,8 +30,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -39,6 +40,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = with maintainers; [ lovesegfault ] ++ teams.rocm.members;
     platforms = platforms.linux;
-    broken = versions.minor finalAttrs.version != versions.minor stdenv.cc.version || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -0,0 +1,37 @@
+{
+  lib,
+  stdenv,
+  cmake,
+  fetchFromGitHub,
+  rocm-cmake,
+  rocmUpdateScript,
+}:
+stdenv.mkDerivation (finalAttrs: {
+  pname = "hipblas-common";
+  version = "6.3.3";
+
+  src = fetchFromGitHub {
+    owner = "ROCm";
+    repo = "hipBLAS-common";
+    rev = "rocm-${finalAttrs.version}";
+    hash = "sha256-tvNz4ymQ1y3YSUQxAtNu2who79QzSKR+3JEevr+GDWo=";
+  };
+
+  nativeBuildInputs = [
+    cmake
+    rocm-cmake
+  ];
+
+  passthru.updateScript = rocmUpdateScript {
+    name = finalAttrs.pname;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
+  };
+  meta = with lib; {
+    description = "Common files shared by hipBLAS and hipBLASLt";
+    homepage = "https://github.com/ROCm/hipBLASlt";
+    license = with licenses; [ mit ];
+    maintainers = teams.rocm.members;
+    platforms = platforms.linux;
+  };
+})

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

@@ -2,13 +2,17 @@
   lib,
   stdenv,
   fetchFromGitHub,
+  fetchpatch,
   rocmUpdateScript,
   cmake,
   rocm-cmake,
   clr,
   gfortran,
+  hipblas-common,
   rocblas,
   rocsolver,
+  rocsparse,
+  rocprim,
   gtest,
   lapack-reference,
   buildTests ? false,
@@ -19,7 +23,7 @@
 # Can also use cuBLAS
 stdenv.mkDerivation (finalAttrs: {
   pname = "hipblas";
-  version = "6.0.2";
+  version = "6.3.3";
 
   outputs =
     [
@@ -39,9 +43,23 @@ stdenv.mkDerivation (finalAttrs: {
     owner = "ROCm";
     repo = "hipBLAS";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-Fq7o2sMmHlHIv9UKJw+u/h9K/ZhKVJWwosYTdYIsscA=";
+    hash = "sha256-Rz1KAhBUbvErHTF2PM1AkVhqo4OHldfSNMSpp5Tx9yk=";
   };
 
+  patches = [
+    # https://github.com/ROCm/hipBLAS/pull/952
+    (fetchpatch {
+      name = "transitively-depend-hipblas-common.patch";
+      url = "https://github.com/ROCm/hipBLAS/commit/54220fdaebf0fb4fd0921ee9e418ace5b143ec8f.patch";
+      hash = "sha256-MFEhv8Bkrd2zD0FFIDg9oJzO7ztdyMAF+R9oYA0rmwQ=";
+    })
+  ];
+
+  postPatch = ''
+    substituteInPlace library/CMakeLists.txt \
+      --replace-fail "find_package(Git REQUIRED)" ""
+  '';
+
   nativeBuildInputs = [
     cmake
     rocm-cmake
@@ -49,9 +67,13 @@ stdenv.mkDerivation (finalAttrs: {
     gfortran
   ];
 
+  propagatedBuildInputs = [ hipblas-common ];
+
   buildInputs =
     [
       rocblas
+      rocprim
+      rocsparse
       rocsolver
     ]
     ++ lib.optionals buildTests [
@@ -63,13 +85,16 @@ stdenv.mkDerivation (finalAttrs: {
 
   cmakeFlags =
     [
-      "-DCMAKE_C_COMPILER=hipcc"
-      "-DCMAKE_CXX_COMPILER=hipcc"
+      "-DCMAKE_BUILD_TYPE=Release"
+      "-DCMAKE_CXX_COMPILER=${lib.getExe' clr "hipcc"}"
+      # Upstream is migrating to amdclang++, it is likely this will be correct in next version bump
+      #"-DCMAKE_CXX_COMPILER=${lib.getBin clr}/bin/amdclang++"
       # Manually define CMAKE_INSTALL_<DIR>
       # See: https://github.com/NixOS/nixpkgs/pull/197838
       "-DCMAKE_INSTALL_BINDIR=bin"
       "-DCMAKE_INSTALL_LIBDIR=lib"
       "-DCMAKE_INSTALL_INCLUDEDIR=include"
+      "-DAMDGPU_TARGETS=${rocblas.amdgpu_targets}"
     ]
     ++ lib.optionals buildTests [
       "-DBUILD_CLIENTS_TESTS=ON"
@@ -100,8 +125,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -110,8 +135,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -0,0 +1,235 @@
+{
+  lib,
+  stdenv,
+  fetchpatch,
+  fetchFromGitHub,
+  cmake,
+  rocm-cmake,
+  clr,
+  gfortran,
+  gtest,
+  msgpack,
+  libxml2,
+  python3,
+  python3Packages,
+  openmp,
+  hipblas-common,
+  tensile,
+  lapack-reference,
+  ncurses,
+  libffi,
+  zlib,
+  zstd,
+  rocmUpdateScript,
+  buildTests ? false,
+  buildBenchmarks ? false,
+  buildSamples ? false,
+  # hipblaslt supports only devices with MFMA or WMMA
+  # WMMA on gfx1100 may be broken
+  # MFMA on MI100 may be broken
+  # MI200/MI300 known to work
+  gpuTargets ? (
+    clr.localGpuTargets or [
+      # "gfx908" FIXME: confirm MFMA on MI100 works
+      "gfx90a"
+      "gfx942"
+      # "gfx1100" FIXME: confirm WMMA targets work
+    ]
+  ),
+}:
+
+stdenv.mkDerivation (
+  finalAttrs:
+  let
+    supportsTargetArches =
+      (builtins.any (lib.strings.hasPrefix "gfx9") gpuTargets)
+      || (builtins.any (lib.strings.hasPrefix "gfx11") gpuTargets);
+    tensile' = (tensile.override { isTensileLite = true; }).overrideAttrs {
+      inherit (finalAttrs) src;
+      sourceRoot = "${finalAttrs.src.name}/tensilelite";
+    };
+    py = python3.withPackages (ps: [
+      ps.pyyaml
+      ps.setuptools
+      ps.packaging
+    ]);
+    gpuTargets' = lib.optionalString supportsTargetArches (lib.concatStringsSep ";" gpuTargets);
+    compiler = "amdclang++";
+    cFlags = "-O3 -I${msgpack}/include"; # FIXME: cmake files need patched to include this properly
+  in
+  {
+    pname = "hipblaslt${clr.gpuArchSuffix}";
+    version = "6.3.3";
+
+    src = fetchFromGitHub {
+      owner = "ROCm";
+      repo = "hipBLASLt";
+      rev = "rocm-${finalAttrs.version}";
+      hash = "sha256-ozfHwsxcczzYXN9SIkyfRvdtaCqlDN4bh3UHZNS2oVQ=";
+    };
+    env.CXX = compiler;
+    env.CFLAGS = cFlags;
+    env.CXXFLAGS = cFlags;
+    env.ROCM_PATH = "${clr}";
+    env.TENSILE_ROCM_ASSEMBLER_PATH = lib.getExe' clr "amdclang++";
+    env.TENSILE_GEN_ASSEMBLY_TOOLCHAIN = lib.getExe' clr "amdclang++";
+    # Some tensile scripts look for this as an env var rather than a cmake flag
+    env.CMAKE_CXX_COMPILER = lib.getExe' clr "amdclang++";
+    requiredSystemFeatures = [ "big-parallel" ];
+
+    outputs =
+      [
+        "out"
+      ]
+      ++ lib.optionals buildTests [
+        "test"
+      ]
+      ++ lib.optionals buildBenchmarks [
+        "benchmark"
+      ]
+      ++ lib.optionals buildSamples [
+        "sample"
+      ];
+
+    postPatch = ''
+      mkdir -p build/Tensile/library
+      # git isn't needed and we have no .git
+      substituteInPlace cmake/Dependencies.cmake \
+        --replace-fail "find_package(Git REQUIRED)" ""
+      substituteInPlace CMakeLists.txt \
+        --replace-fail "include(virtualenv)" "" \
+        --replace-fail "virtualenv_install(\''${Tensile_TEST_LOCAL_PATH})" "" \
+        --replace-fail "virtualenv_install(\''${CMAKE_SOURCE_DIR}/tensilelite)" "" \
+        --replace-fail 'find_package(Tensile 4.33.0 EXACT REQUIRED HIP LLVM OpenMP PATHS "''${INSTALLED_TENSILE_PATH}")' "find_package(Tensile)" \
+        --replace-fail 'Tensile_CPU_THREADS ""' 'Tensile_CPU_THREADS "$ENV{NIX_BUILD_CORES}"'
+      # FIXME: TensileCreateExtOpLibraries build failure due to unsupported null operand
+      # Working around for now by disabling the ExtOp libs
+      substituteInPlace library/src/amd_detail/rocblaslt/src/CMakeLists.txt \
+        --replace-fail 'TensileCreateExtOpLibraries("' '# skipping TensileCreateExtOpLibraries'
+      substituteInPlace library/src/amd_detail/rocblaslt/src/kernels/compile_code_object.sh \
+        --replace-fail '${"\${rocm_path}"}/bin/' ""
+    '';
+
+    # Apply patches to allow building without a target arch if we need to do that
+    patches = lib.optionals (!supportsTargetArches) [
+      # Add ability to build without specitying any arch.
+      (fetchpatch {
+        sha256 = "sha256-VW3bPzmQvfo8+iKsVfpn4sbqAe41fLzCEUfBh9JxVyk=";
+        url = "https://raw.githubusercontent.com/gentoo/gentoo/refs/heads/master/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-no-arch.patch";
+      })
+      # Followup to above patch for 6.3.x
+      (fetchpatch {
+        sha256 = "sha256-GCsrne6BiWzwj8TMAfFuaYz1Pij97hoCc6E3qJhWb10=";
+        url = "https://raw.githubusercontent.com/gentoo/gentoo/refs/heads/master/sci-libs/hipBLASLt/files/hipBLASLt-6.3.0-no-arch-extra.patch";
+      })
+    ];
+
+    doCheck = false;
+    doInstallCheck = false;
+
+    nativeBuildInputs = [
+      cmake
+      rocm-cmake
+      py
+      clr
+      gfortran
+      # need make to get streaming console output so nix knows build is still running
+      # so deliberately not using ninja
+    ];
+
+    buildInputs =
+      [
+        hipblas-common
+        tensile'
+        openmp
+        libffi
+        ncurses
+
+        # Tensile deps - not optional, building without tensile isn't actually supported
+        msgpack # FIXME: not included in cmake!
+        libxml2
+        python3Packages.msgpack
+        python3Packages.joblib
+        zlib
+        zstd
+      ]
+      ++ lib.optionals buildTests [
+        gtest
+      ]
+      ++ lib.optionals (buildTests || buildBenchmarks) [
+        lapack-reference
+      ];
+
+    cmakeFlags =
+      [
+        "-Wno-dev"
+        "-DCMAKE_BUILD_TYPE=Release"
+        "-DCMAKE_VERBOSE_MAKEFILE=ON"
+        "-DVIRTUALENV_PYTHON_EXENAME=${lib.getExe py}"
+        "-DTENSILE_USE_HIP=ON"
+        "-DTENSILE_BUILD_CLIENT=OFF"
+        "-DTENSILE_USE_FLOAT16_BUILTIN=ON"
+        "-DCMAKE_CXX_COMPILER=${compiler}"
+        # Manually define CMAKE_INSTALL_<DIR>
+        # See: https://github.com/NixOS/nixpkgs/pull/197838
+        "-DCMAKE_INSTALL_BINDIR=bin"
+        "-DCMAKE_INSTALL_LIBDIR=lib"
+        "-DCMAKE_INSTALL_INCLUDEDIR=include"
+        "-DHIPBLASLT_ENABLE_MARKER=Off"
+        # FIXME what are the implications of hardcoding this?
+        "-DTensile_CODE_OBJECT_VERSION=V5"
+        "-DTensile_COMPILER=${compiler}"
+        "-DAMDGPU_TARGETS=${gpuTargets'}"
+        "-DGPU_TARGETS=${gpuTargets'}"
+        "-DTensile_LIBRARY_FORMAT=msgpack"
+      ]
+      ++ lib.optionals (!supportsTargetArches) [
+        "-DBUILD_WITH_TENSILE=OFF"
+      ]
+      ++ lib.optionals buildTests [
+        "-DBUILD_CLIENTS_TESTS=ON"
+      ]
+      ++ lib.optionals buildBenchmarks [
+        "-DBUILD_CLIENTS_BENCHMARKS=ON"
+      ]
+      ++ lib.optionals buildSamples [
+        "-DBUILD_CLIENTS_SAMPLES=ON"
+      ];
+
+    postInstall =
+      lib.optionalString buildTests ''
+        mkdir -p $test/bin
+        mv $out/bin/hipblas-test $test/bin
+      ''
+      + lib.optionalString buildBenchmarks ''
+        mkdir -p $benchmark/bin
+        mv $out/bin/hipblas-bench $benchmark/bin
+      ''
+      + lib.optionalString buildSamples ''
+        mkdir -p $sample/bin
+        mv $out/bin/example-* $sample/bin
+      ''
+      + lib.optionalString (buildTests || buildBenchmarks || buildSamples) ''
+        rmdir $out/bin
+      '';
+    # If this is false there are no kernels in the output lib
+    # and it's useless at runtime
+    # so if it's an optional dep it's best to not depend on it
+    # Some packages like torch need hipblaslt to compile
+    # and are fine ignoring it at runtime if it's not supported
+    # so we have to support building an empty hipblaslt
+    passthru.supportsTargetArches = supportsTargetArches;
+    passthru.updateScript = rocmUpdateScript {
+      name = finalAttrs.pname;
+      inherit (finalAttrs.src) owner repo;
+    };
+    passthru.tensilelite = tensile';
+    meta = with lib; {
+      description = "hipBLASLt is a library that provides general matrix-matrix operations with a flexible API";
+      homepage = "https://github.com/ROCm/hipBLASlt";
+      license = with licenses; [ mit ];
+      maintainers = teams.rocm.members;
+      platforms = platforms.linux;
+    };
+  }
+)

pkgs/development/rocm-modules/6/hipcc/0001-hipcc-Remove-extra-definition-of-hipBinUtilPtr_-in-d.patch

@@ -0,0 +1,39 @@
+From f259eca77c592813e11752a46c4e1f9a74c64091 Mon Sep 17 00:00:00 2001
+From: Luna Nova <git@lunnova.dev>
+Date: Fri, 11 Oct 2024 02:56:22 -0700
+Subject: [PATCH] [hipcc] Remove extra definition of hipBinUtilPtr_ in derived
+ platforms
+
+Fixes UB when hipBinUtilPtr_ is used.
+---
+ amd/hipcc/src/hipBin_amd.h    | 1 -
+ amd/hipcc/src/hipBin_nvidia.h | 1 -
+ 2 files changed, 2 deletions(-)
+
+diff --git a/amd/hipcc/src/hipBin_amd.h b/amd/hipcc/src/hipBin_amd.h
+index 0a782d1beab9..36cd625ae8bc 100644
+--- a/src/hipBin_amd.h
++++ b/src/hipBin_amd.h
+@@ -42,7 +42,6 @@ THE SOFTWARE.
+ 
+ class HipBinAmd : public HipBinBase {
+  private:
+-  HipBinUtil* hipBinUtilPtr_;
+   string hipClangPath_ = "";
+   string roccmPathEnv_, hipRocclrPathEnv_, hsaPathEnv_;
+   PlatformInfo platformInfoAMD_;
+diff --git a/amd/hipcc/src/hipBin_nvidia.h b/amd/hipcc/src/hipBin_nvidia.h
+index ff142cc1cea2..09b7b80979c7 100644
+--- a/src/hipBin_nvidia.h
++++ b/src/hipBin_nvidia.h
+@@ -31,7 +31,6 @@ THE SOFTWARE.
+ 
+ class HipBinNvidia : public HipBinBase {
+  private:
+-  HipBinUtil* hipBinUtilPtr_;
+   string cudaPath_ = "";
+   PlatformInfo platformInfoNV_;
+   string hipCFlags_, hipCXXFlags_, hipLdFlags_;
+-- 
+2.46.0
+

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

@@ -1,49 +1,46 @@
 {
   lib,
   stdenv,
-  fetchFromGitHub,
-  rocmUpdateScript,
+  rocm-merged-llvm,
   cmake,
   lsb-release,
 }:
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "hipcc";
-  version = "6.0.2";
-
-  src = fetchFromGitHub {
-    owner = "ROCm";
-    repo = "HIPCC";
-    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-/LRQN+RSMBPk2jS/tdp3psUL/B0RJZQhRri7e67KsG4=";
-  };
+  # In-tree with ROCm LLVM
+  inherit (rocm-merged-llvm) version;
+  src = rocm-merged-llvm.llvm-src;
+  sourceRoot = "${finalAttrs.src.name}/amd/hipcc";
 
   nativeBuildInputs = [ cmake ];
 
+  buildInputs = [ rocm-merged-llvm ];
+
+  patches = [
+    # https://github.com/ROCm/llvm-project/pull/183
+    # Fixes always-invoked UB in hipcc
+    ./0001-hipcc-Remove-extra-definition-of-hipBinUtilPtr_-in-d.patch
+  ];
+
   postPatch = ''
     substituteInPlace src/hipBin_amd.h \
-      --replace "/usr/bin/lsb_release" "${lsb-release}/bin/lsb_release"
+      --replace-fail "/usr/bin/lsb_release" "${lsb-release}/bin/lsb_release"
   '';
 
+  cmakeFlags = [
+    "-DCMAKE_BUILD_TYPE=Release"
+  ];
   postInstall = ''
     rm -r $out/hip/bin
     ln -s $out/bin $out/hip/bin
   '';
 
-  passthru.updateScript = rocmUpdateScript {
-    name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
-  };
-
   meta = with lib; {
     description = "Compiler driver utility that calls clang or nvcc";
     homepage = "https://github.com/ROCm/HIPCC";
     license = with licenses; [ mit ];
     maintainers = with maintainers; [ lovesegfault ] ++ teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -1,36 +1,40 @@
-{ lib
-, stdenv
-, fetchFromGitHub
-, rocmUpdateScript
-, cmake
-, rocm-cmake
-, rocprim
-, clr
-, gtest
-, gbenchmark
-, buildTests ? false
-, buildBenchmarks ? false
-, gpuTargets ? [ ]
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  rocmUpdateScript,
+  cmake,
+  rocm-cmake,
+  rocprim,
+  clr,
+  gtest,
+  gbenchmark,
+  buildTests ? false,
+  buildBenchmarks ? false,
+  gpuTargets ? [ ],
 }:
 
 # CUB can also be used as a backend instead of rocPRIM.
 stdenv.mkDerivation (finalAttrs: {
   pname = "hipcub";
-  version = "6.0.2";
+  version = "6.3.3";
 
-  outputs = [
-    "out"
-  ] ++ lib.optionals buildTests [
-    "test"
-  ] ++ lib.optionals buildBenchmarks [
-    "benchmark"
-  ];
+  outputs =
+    [
+      "out"
+    ]
+    ++ lib.optionals buildTests [
+      "test"
+    ]
+    ++ lib.optionals buildBenchmarks [
+      "benchmark"
+    ];
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "hipCUB";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-8QzVgj0JSb86zEG3sj5AAt9pG3frw+xrjEOTo7xCIrc=";
+    hash = "sha256-uECOQWG9C64tg5YZdm9/3+fZXaZVGslu8vElK3m23GY=";
   };
 
   nativeBuildInputs = [
@@ -39,44 +43,53 @@ stdenv.mkDerivation (finalAttrs: {
     clr
   ];
 
-  buildInputs = [
-    rocprim
-  ] ++ lib.optionals buildTests [
-    gtest
-  ] ++ lib.optionals buildBenchmarks [
-    gbenchmark
-  ];
+  buildInputs =
+    [
+      rocprim
+    ]
+    ++ lib.optionals buildTests [
+      gtest
+    ]
+    ++ lib.optionals buildBenchmarks [
+      gbenchmark
+    ];
 
-  cmakeFlags = [
-    "-DCMAKE_CXX_COMPILER=hipcc"
-    "-DHIP_ROOT_DIR=${clr}"
-    # Manually define CMAKE_INSTALL_<DIR>
-    # See: https://github.com/NixOS/nixpkgs/pull/197838
-    "-DCMAKE_INSTALL_BINDIR=bin"
-    "-DCMAKE_INSTALL_LIBDIR=lib"
-    "-DCMAKE_INSTALL_INCLUDEDIR=include"
-  ] ++ lib.optionals (gpuTargets != [ ]) [
-    "-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
-  ] ++ lib.optionals buildTests [
-    "-DBUILD_TEST=ON"
-  ] ++ lib.optionals buildBenchmarks [
-    "-DBUILD_BENCHMARK=ON"
-  ];
+  cmakeFlags =
+    [
+      "-DHIP_ROOT_DIR=${clr}"
+      # Manually define CMAKE_INSTALL_<DIR>
+      # See: https://github.com/NixOS/nixpkgs/pull/197838
+      "-DCMAKE_INSTALL_BINDIR=bin"
+      "-DCMAKE_INSTALL_LIBDIR=lib"
+      "-DCMAKE_INSTALL_INCLUDEDIR=include"
+    ]
+    ++ lib.optionals (gpuTargets != [ ]) [
+      "-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
+    ]
+    ++ lib.optionals buildTests [
+      "-DBUILD_TEST=ON"
+    ]
+    ++ lib.optionals buildBenchmarks [
+      "-DBUILD_BENCHMARK=ON"
+    ];
 
-  postInstall = lib.optionalString buildTests ''
-    mkdir -p $test/bin
-    mv $out/bin/test_* $test/bin
-  '' + lib.optionalString buildBenchmarks ''
-    mkdir -p $benchmark/bin
-    mv $out/bin/benchmark_* $benchmark/bin
-  '' + lib.optionalString (buildTests || buildBenchmarks) ''
-    rmdir $out/bin
-  '';
+  postInstall =
+    lib.optionalString buildTests ''
+      mkdir -p $test/bin
+      mv $out/bin/test_* $test/bin
+    ''
+    + lib.optionalString buildBenchmarks ''
+      mkdir -p $benchmark/bin
+      mv $out/bin/benchmark_* $benchmark/bin
+    ''
+    + lib.optionalString (buildTests || buildBenchmarks) ''
+      rmdir $out/bin
+    '';
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -85,6 +98,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ bsd3 ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken = versions.minor finalAttrs.version != versions.minor stdenv.cc.version || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -22,7 +22,7 @@
 # Can also use cuFFT
 stdenv.mkDerivation (finalAttrs: {
   pname = "hipfft";
-  version = "6.0.2";
+  version = "6.3.3";
 
   outputs =
     [
@@ -42,7 +42,7 @@ stdenv.mkDerivation (finalAttrs: {
     owner = "ROCm";
     repo = "hipFFT";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-DjjNQryJdl7RmaMQRQPWkleweEWMIwH/xXU84GGjoC0=";
+    hash = "sha256-Jq/YHEtOo7a0/Ki7gxZATKmSqPU6cyLf5gx3A4MAZNw=";
     fetchSubmodules = true;
   };
 
@@ -111,8 +111,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -121,8 +121,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -1,21 +1,22 @@
-{ lib
-, stdenv
-, fetchFromGitHub
-, rocmUpdateScript
-, cmake
-, rocm-cmake
-, gfortran
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  rocmUpdateScript,
+  cmake,
+  rocm-cmake,
+  gfortran,
 }:
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "hipfort";
-  version = "6.0.2";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "hipfort";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-3PIqSDyDlY0oVSEx20EPlKGYNkc9xPZtIG3Sbw69esE=";
+    hash = "sha256-V5XDNM0bYHKnpkcnaDyxIS1zwsgaByJj+znFxJ6VxR0=";
   };
 
   nativeBuildInputs = [
@@ -51,8 +52,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -61,6 +62,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ]; # mitx11
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken = versions.minor finalAttrs.version != versions.minor stdenv.cc.version || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -1,39 +1,57 @@
-{ lib
-, stdenv
-, fetchFromGitHub
-, rocmUpdateScript
-, cmake
-, clang
-, libxml2
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  rocmUpdateScript,
+  cmake,
+  clang,
+  libxml2,
+  rocm-merged-llvm,
+  zlib,
+  zstd,
+  perl,
 }:
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "hipify";
-  version = "6.0.2";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "HIPIFY";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-nNyWrPPhUwT7FyASzc3kf5NCTzeqvHybVOc+6hBzkA4=";
+    hash = "sha256-o/1LNsNtAyQcSug1gf7ujGNRRbvC33kwldrJKZi2LA0=";
   };
 
-  nativeBuildInputs = [ cmake ];
-  buildInputs = [ libxml2 ];
+  nativeBuildInputs = [
+    cmake
+  ];
+
+  buildInputs = [
+    libxml2
+    rocm-merged-llvm
+    zlib
+    zstd
+    perl
+  ];
 
   postPatch = ''
     substituteInPlace CMakeLists.txt \
       --replace "\''${LLVM_TOOLS_BINARY_DIR}/clang" "${clang}/bin/clang"
+    chmod +x bin/*
   '';
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   postInstall = ''
-    patchShebangs $out/bin
+    chmod +x $out/bin/*
+    chmod +x $out/libexec/*
+    patchShebangs $out/bin/
+    patchShebangs $out/libexec/
   '';
 
   meta = with lib; {
@@ -42,6 +60,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken = versions.minor finalAttrs.version != versions.minor stdenv.cc.version || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -14,7 +14,7 @@
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "hiprand";
-  version = "6.0.2";
+  version = "6.3.3";
 
   outputs =
     [
@@ -28,7 +28,7 @@ stdenv.mkDerivation (finalAttrs: {
     owner = "ROCm";
     repo = "hipRAND";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-uGHzOhUX5JEknVFwhHhWFdPmwLS/TuaXYMeItS7tXIg=";
+    hash = "sha256-TVc+qFwRiS5tAo1OKI1Wu5hadlwPZmSVZ9SvVvH1w7Y=";
   };
 
   nativeBuildInputs = [
@@ -41,8 +41,6 @@ stdenv.mkDerivation (finalAttrs: {
 
   cmakeFlags =
     [
-      "-DCMAKE_C_COMPILER=hipcc"
-      "-DCMAKE_CXX_COMPILER=hipcc"
       "-DHIP_ROOT_DIR=${clr}"
       # Manually define CMAKE_INSTALL_<DIR>
       # See: https://github.com/NixOS/nixpkgs/pull/197838
@@ -67,8 +65,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -77,8 +75,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -9,6 +9,8 @@
   gfortran,
   rocblas,
   rocsolver,
+  rocsparse,
+  suitesparse,
   gtest,
   lapack-reference,
   buildTests ? false,
@@ -19,7 +21,7 @@
 # Can also use cuSOLVER
 stdenv.mkDerivation (finalAttrs: {
   pname = "hipsolver";
-  version = "6.0.2";
+  version = "6.3.3";
 
   outputs =
     [
@@ -39,7 +41,7 @@ stdenv.mkDerivation (finalAttrs: {
     owner = "ROCm";
     repo = "hipSOLVER";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-iMfaOv4TdTkmaRHCZOuqUfjO081J6on71+s8nIwwV00=";
+    hash = "sha256-ZQUKU3L4DgZ5zM7pCYEix0ulRkl78x/5wJnyCndTAwk=";
   };
 
   nativeBuildInputs = [
@@ -53,6 +55,8 @@ stdenv.mkDerivation (finalAttrs: {
     [
       rocblas
       rocsolver
+      rocsparse
+      suitesparse
     ]
     ++ lib.optionals buildTests [
       gtest
@@ -63,13 +67,13 @@ stdenv.mkDerivation (finalAttrs: {
 
   cmakeFlags =
     [
-      "-DCMAKE_C_COMPILER=hipcc"
       "-DCMAKE_CXX_COMPILER=hipcc"
       # Manually define CMAKE_INSTALL_<DIR>
       # See: https://github.com/NixOS/nixpkgs/pull/197838
       "-DCMAKE_INSTALL_BINDIR=bin"
       "-DCMAKE_INSTALL_LIBDIR=lib"
       "-DCMAKE_INSTALL_INCLUDEDIR=include"
+      "-DBUILD_WITH_SPARSE=OFF" # FIXME: broken - can't find suitesparse/cholmod, looks fixed in master
     ]
     ++ lib.optionals buildTests [
       "-DBUILD_CLIENTS_TESTS=ON"
@@ -101,8 +105,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -111,8 +115,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -12,6 +12,7 @@
   gtest,
   openmp,
   buildTests ? false,
+  buildBenchmarks ? false,
   buildSamples ? false,
   gpuTargets ? [ ],
 }:
@@ -19,7 +20,7 @@
 # This can also use cuSPARSE as a backend instead of rocSPARSE
 stdenv.mkDerivation (finalAttrs: {
   pname = "hipsparse";
-  version = "6.0.2";
+  version = "6.3.3";
 
   outputs =
     [
@@ -36,7 +37,7 @@ stdenv.mkDerivation (finalAttrs: {
     owner = "ROCm";
     repo = "hipSPARSE";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-fi5b0IF++OiezpM3JuUkhwpmW2apeFH4r5g6CcFseNY=";
+    hash = "sha256-3a7fKpYyiqG3aGOg7YrTHmKoH4rgTVLD16DvrZ3YY1g=";
   };
 
   nativeBuildInputs = [
@@ -51,7 +52,7 @@ stdenv.mkDerivation (finalAttrs: {
       rocsparse
       git
     ]
-    ++ lib.optionals buildTests [
+    ++ lib.optionals (buildTests || buildBenchmarks) [
       gtest
     ]
     ++ lib.optionals (buildTests || buildSamples) [
@@ -60,20 +61,17 @@ stdenv.mkDerivation (finalAttrs: {
 
   cmakeFlags =
     [
-      "-DCMAKE_C_COMPILER=hipcc"
-      "-DCMAKE_CXX_COMPILER=hipcc"
-      "-DBUILD_CLIENTS_SAMPLES=${if buildSamples then "ON" else "OFF"}"
       # Manually define CMAKE_INSTALL_<DIR>
       # See: https://github.com/NixOS/nixpkgs/pull/197838
       "-DCMAKE_INSTALL_BINDIR=bin"
       "-DCMAKE_INSTALL_LIBDIR=lib"
       "-DCMAKE_INSTALL_INCLUDEDIR=include"
+      (lib.cmakeBool "BUILD_CLIENTS_TESTS" buildTests)
+      (lib.cmakeBool "BUILD_CLIENTS_BENCHMARKS" buildBenchmarks)
+      (lib.cmakeBool "BUILD_CLIENTS_SAMPLES" buildSamples)
     ]
     ++ lib.optionals (gpuTargets != [ ]) [
       "-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
-    ]
-    ++ lib.optionals buildTests [
-      "-DBUILD_CLIENTS_TESTS=ON"
     ];
 
   # We have to manually generate the matrices
@@ -140,8 +138,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -150,8 +148,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

pkgs/development/rocm-modules/6/hsa-amd-aqlprofile-bin/default.nix

@@ -4,22 +4,25 @@
   fetchurl,
   callPackage,
   dpkg,
+  rocm-core,
 }:
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "hsa-amd-aqlprofile-bin";
-  version = "6.0.2";
+  version = "6.3.3";
 
   src =
     let
-      version = finalAttrs.version;
-      dotless = builtins.replaceStrings [ "." ] [ "0" ] version;
-      incremental = "115";
+      inherit (finalAttrs) version;
+      patch = rocm-core.ROCM_LIBPATCH_VERSION;
+      majorMinor = lib.versions.majorMinor version;
+      poolVersion = if majorMinor + ".0" == version then majorMinor else version;
+      incremental = "74";
       osRelease = "22.04";
     in
     fetchurl {
-      url = "https://repo.radeon.com/rocm/apt/${version}/pool/main/h/hsa-amd-aqlprofile/hsa-amd-aqlprofile_1.0.0.${dotless}.${dotless}-${incremental}~${osRelease}_amd64.deb";
-      hash = "sha256-0XeKUKaof5pSMS/UgLwumBDBYgyH/pCex9jViUKENXY=";
+      url = "https://repo.radeon.com/rocm/apt/${poolVersion}/pool/main/h/hsa-amd-aqlprofile/hsa-amd-aqlprofile_1.0.0.${patch}-${incremental}~${osRelease}_amd64.deb";
+      hash = "sha256-Lo6gU9ywkujtsKvnOAwL3L8qQNPwjjm0Pm4OyzoUYao=";
     };
 
   nativeBuildInputs = [ dpkg ];
@@ -31,7 +34,7 @@ stdenv.mkDerivation (finalAttrs: {
     runHook preInstall
 
     mkdir -p $out
-    cp -a opt/rocm-${finalAttrs.version}/* $out
+    cp -a opt/rocm-${finalAttrs.version}*/* $out
     chmod +x $out/lib/libhsa-amd-aqlprofile64.so.1.*
     chmod +x $out/lib/hsa-amd-aqlprofile/librocprofv2_att.so
 
@@ -46,8 +49,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ unfree ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

pkgs/development/rocm-modules/6/llvm/base.nix

@@ -1,216 +0,0 @@
-{
-  lib,
-  stdenv,
-  gcc12Stdenv,
-  fetchFromGitHub,
-  rocmUpdateScript,
-  pkg-config,
-  cmake,
-  ninja,
-  git,
-  doxygen,
-  sphinx,
-  lit,
-  libxml2,
-  libxcrypt,
-  libedit,
-  libffi,
-  mpfr,
-  zlib,
-  ncurses,
-  python3Packages,
-  buildDocs ? true,
-  buildMan ? true,
-  buildTests ? true,
-  targetName ? "llvm",
-  targetDir ? "llvm",
-  targetProjects ? [ ],
-  targetRuntimes ? [ ],
-  llvmTargetsToBuild ? [ "NATIVE" ], # "NATIVE" resolves into x86 or aarch64 depending on stdenv
-  extraPatches ? [ ],
-  extraNativeBuildInputs ? [ ],
-  extraBuildInputs ? [ ],
-  extraCMakeFlags ? [ ],
-  extraPostPatch ? "",
-  checkTargets ? [
-    (lib.optionalString buildTests (if targetDir == "runtimes" then "check-runtimes" else "check-all"))
-  ],
-  extraPostInstall ? "",
-  hardeningDisable ? [ ],
-  requiredSystemFeatures ? [ ],
-  extraLicenses ? [ ],
-  isBroken ? false,
-}:
-
-let
-  stdenv' = stdenv;
-in
-let
-  stdenv =
-    if stdenv'.cc.cc.isGNU or false && lib.versionAtLeast stdenv'.cc.cc.version "13.0" then
-      gcc12Stdenv
-    else
-      stdenv';
-in
-
-let
-  llvmNativeTarget =
-    if stdenv.hostPlatform.isx86_64 then
-      "X86"
-    else if stdenv.hostPlatform.isAarch64 then
-      "AArch64"
-    else
-      throw "Unsupported ROCm LLVM platform";
-  inferNativeTarget = t: if t == "NATIVE" then llvmNativeTarget else t;
-  llvmTargetsToBuild' = [ "AMDGPU" ] ++ builtins.map inferNativeTarget llvmTargetsToBuild;
-in
-stdenv.mkDerivation (finalAttrs: {
-  pname = "rocm-llvm-${targetName}";
-  version = "6.0.2";
-
-  outputs =
-    [
-      "out"
-    ]
-    ++ lib.optionals buildDocs [
-      "doc"
-    ]
-    ++ lib.optionals buildMan [
-      "man"
-      "info" # Avoid `attribute 'info' missing` when using with wrapCC
-    ];
-
-  patches = extraPatches;
-
-  src = fetchFromGitHub {
-    owner = "ROCm";
-    repo = "llvm-project";
-    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-uGxalrwMNCOSqSFVrYUBi3ijkMEFFTrzFImmvZKQf6I=";
-  };
-
-  nativeBuildInputs =
-    [
-      pkg-config
-      cmake
-      ninja
-      git
-      (python3Packages.python.withPackages (p: [ p.setuptools ]))
-    ]
-    ++ lib.optionals (buildDocs || buildMan) [
-      doxygen
-      sphinx
-      python3Packages.recommonmark
-    ]
-    ++ lib.optionals (buildTests && !finalAttrs.passthru.isLLVM) [
-      lit
-    ]
-    ++ extraNativeBuildInputs;
-
-  buildInputs = [
-    libxml2
-    libxcrypt
-    libedit
-    libffi
-    mpfr
-  ] ++ extraBuildInputs;
-
-  propagatedBuildInputs = lib.optionals finalAttrs.passthru.isLLVM [
-    zlib
-    ncurses
-  ];
-
-  sourceRoot = "${finalAttrs.src.name}/${targetDir}";
-
-  cmakeFlags =
-    [
-      "-DLLVM_TARGETS_TO_BUILD=${builtins.concatStringsSep ";" llvmTargetsToBuild'}"
-    ]
-    ++ lib.optionals (finalAttrs.passthru.isLLVM && targetProjects != [ ]) [
-      "-DLLVM_ENABLE_PROJECTS=${lib.concatStringsSep ";" targetProjects}"
-    ]
-    ++
-      lib.optionals ((finalAttrs.passthru.isLLVM || targetDir == "runtimes") && targetRuntimes != [ ])
-        [
-          "-DLLVM_ENABLE_RUNTIMES=${lib.concatStringsSep ";" targetRuntimes}"
-        ]
-    ++ lib.optionals finalAttrs.passthru.isLLVM [
-      "-DLLVM_INSTALL_UTILS=ON"
-      "-DLLVM_INSTALL_GTEST=ON"
-    ]
-    ++ lib.optionals (buildDocs || buildMan) [
-      "-DLLVM_INCLUDE_DOCS=ON"
-      "-DLLVM_BUILD_DOCS=ON"
-      # "-DLLVM_ENABLE_DOXYGEN=ON" Way too slow, only uses one core
-      "-DLLVM_ENABLE_SPHINX=ON"
-      "-DSPHINX_OUTPUT_HTML=ON"
-      "-DSPHINX_OUTPUT_MAN=ON"
-      "-DSPHINX_WARNINGS_AS_ERRORS=OFF"
-    ]
-    ++ lib.optionals buildTests [
-      "-DLLVM_INCLUDE_TESTS=ON"
-      "-DLLVM_BUILD_TESTS=ON"
-      "-DLLVM_EXTERNAL_LIT=${lit}/bin/.lit-wrapped"
-    ]
-    ++ extraCMakeFlags;
-
-  prePatch = ''
-    cd ../
-    chmod -R u+w .
-  '';
-
-  postPatch =
-    ''
-      cd ${targetDir}
-    ''
-    + lib.optionalString finalAttrs.passthru.isLLVM ''
-      patchShebangs lib/OffloadArch/make_generated_offload_arch_h.sh
-    ''
-    + lib.optionalString (buildTests && finalAttrs.passthru.isLLVM) ''
-      # FileSystem permissions tests fail with various special bits
-      rm test/tools/llvm-objcopy/ELF/mirror-permissions-unix.test
-      rm unittests/Support/Path.cpp
-
-      substituteInPlace unittests/Support/CMakeLists.txt \
-        --replace-fail "Path.cpp" ""
-    ''
-    + extraPostPatch;
-
-  doCheck = buildTests;
-  checkTarget = lib.concatStringsSep " " checkTargets;
-
-  postInstall =
-    lib.optionalString buildMan ''
-      mkdir -p $info
-    ''
-    + extraPostInstall;
-
-  passthru = {
-    isLLVM = targetDir == "llvm";
-    isClang = targetDir == "clang" || builtins.elem "clang" targetProjects;
-    isROCm = true;
-
-    updateScript = rocmUpdateScript {
-      name = finalAttrs.pname;
-      owner = finalAttrs.src.owner;
-      repo = finalAttrs.src.repo;
-    };
-  };
-
-  inherit hardeningDisable requiredSystemFeatures;
-
-  meta = with lib; {
-    description = "ROCm fork of the LLVM compiler infrastructure";
-    homepage = "https://github.com/ROCm/llvm-project";
-    license = with licenses; [ ncsa ] ++ extraLicenses;
-    maintainers =
-      with maintainers;
-      [
-        acowley
-        lovesegfault
-      ]
-      ++ teams.rocm.members;
-    platforms = platforms.linux;
-    broken = isBroken || versionAtLeast finalAttrs.version "7.0.0";
-  };
-})

pkgs/development/rocm-modules/6/llvm/clang-bodge-ignore-systemwide-incls.diff

@@ -0,0 +1,23 @@
+diff --git a/lib/Driver/ToolChains/Linux.cpp b/lib/Driver/ToolChains/Linux.cpp
+index 57368104c914..71c57f72078e 100644
+--- a/lib/Driver/ToolChains/Linux.cpp
++++ b/lib/Driver/ToolChains/Linux.cpp
+@@ -640,6 +640,7 @@ void Linux::AddClangSystemIncludeArgs(const ArgList &DriverArgs,
+     return;
+ 
+   // LOCAL_INCLUDE_DIR
++  if (!SysRoot.empty())
+   addSystemInclude(DriverArgs, CC1Args, concat(SysRoot, "/usr/local/include"));
+   // TOOL_INCLUDE_DIR
+   AddMultilibIncludeArgs(DriverArgs, CC1Args);
+@@ -672,8 +673,10 @@ void Linux::AddClangSystemIncludeArgs(const ArgList &DriverArgs,
+   // Add an include of '/include' directly. This isn't provided by default by
+   // system GCCs, but is often used with cross-compiling GCCs, and harmless to
+   // add even when Clang is acting as-if it were a system compiler.
++  if (!SysRoot.empty())
+   addExternCSystemInclude(DriverArgs, CC1Args, concat(SysRoot, "/include"));
+ 
++  if (!SysRoot.empty())
+   addExternCSystemInclude(DriverArgs, CC1Args, concat(SysRoot, "/usr/include"));
+ 
+   if (!DriverArgs.hasArg(options::OPT_nobuiltininc) && getTriple().isMusl())

pkgs/development/rocm-modules/6/llvm/clang-log-jobs.diff

@@ -0,0 +1,40 @@
+diff --git a/clang/lib/Driver/Compilation.cpp b/clang/lib/Driver/Compilation.cpp
+index 06f5e7e7e335..8407d664886a 100644
+--- a/lib/Driver/Compilation.cpp
++++ b/lib/Driver/Compilation.cpp
+@@ -340,6 +340,9 @@ private:
+ void Compilation::ExecuteJobs(const JobList &Jobs,
+                               FailingCommandList &FailingCommands,
+                               bool LogOnly) const {
++  // If >1 job, log as each job finishes so can see progress while building many offloads
++  const bool logJobs = Jobs.size() > 1;
++  auto start_time = std::chrono::steady_clock::now();
+   // According to UNIX standard, driver need to continue compiling all the
+   // inputs on the command line even one of them failed.
+   // In all but CLMode, execute all the jobs unless the necessary inputs for the
+@@ -364,11 +367,25 @@ void Compilation::ExecuteJobs(const JobList &Jobs,
+ 
+     JS.setJobState(Next, JobScheduler::JS_RUN);
+     auto Work = [&, Next]() {
++      auto job_start_time = std::chrono::steady_clock::now();
+       const Command *FailingCommand = nullptr;
+       if (int Res = ExecuteCommand(*Next, FailingCommand, LogOnly)) {
+         FailingCommands.push_back(std::make_pair(Res, FailingCommand));
+         JS.setJobState(Next, JobScheduler::JS_FAIL);
+       } else {
++        if (logJobs && Next) {
++          auto now = std::chrono::steady_clock::now();
++          auto job_duration = std::chrono::duration_cast<std::chrono::seconds>(now - job_start_time).count();
++          auto duration = std::chrono::duration_cast<std::chrono::seconds>(now - start_time).count();
++          if (duration > 10 && job_duration > 0) {
++            if (Next->getOutputFilenames().empty())
++              if (Next->getExecutable()) llvm::errs() << "Job completed: " << Next->getExecutable() << "\n";
++              else (llvm::errs() << "Job completed: "), Next->Print(llvm::errs(), "\n", true);
++            else
++              llvm::errs() << "Job completed: " << Next->getOutputFilenames().front().c_str() << "\n";
++          }
++        }
++
+         JS.setJobState(Next, JobScheduler::JS_DONE);
+       }
+     };

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

@@ -1,142 +1,520 @@
 {
-  # stdenv FIXME: Try changing back to this with a new ROCm release https://github.com/NixOS/nixpkgs/issues/271943
-  gcc12Stdenv,
-  callPackage,
-  rocmUpdateScript,
-  wrapBintoolsWith,
+  lib,
+  stdenv,
+  llvmPackages_18,
   overrideCC,
   rocm-device-libs,
   rocm-runtime,
-  rocm-thunk,
-  clr,
+  fetchFromGitHub,
+  runCommand,
+  symlinkJoin,
+  rdfind,
+  wrapBintoolsWith,
+  emptyDirectory,
+  zstd,
+  zlib,
+  gcc-unwrapped,
+  glibc,
+  substituteAll,
+  libffi,
+  libxml2,
+  removeReferencesTo,
+  fetchpatch,
+  # Build compilers and stdenv suitable for profiling
+  # compressed line tables (-g1 -gz) and
+  # frame pointers for sampling profilers (-fno-omit-frame-pointer -momit-leaf-frame-pointer)
+  # TODO: Should also apply to downstream packages which use rocmClangStdenv
+  profilableStdenv ? false,
 }:
 
 let
-  ## Stage 1 ##
-  # Projects
-  llvm = callPackage ./stage-1/llvm.nix {
-    inherit rocmUpdateScript;
-    stdenv = gcc12Stdenv;
-  };
-  clang-unwrapped = callPackage ./stage-1/clang-unwrapped.nix {
-    inherit rocmUpdateScript llvm;
-    stdenv = gcc12Stdenv;
-  };
-  lld = callPackage ./stage-1/lld.nix {
-    inherit rocmUpdateScript llvm;
-    stdenv = gcc12Stdenv;
+  llvmPackagesNoBintools = llvmPackages_18.override {
+    bootBintools = null;
+    bootBintoolsNoLibc = null;
   };
+  useLibcxx = false; # whether rocm stdenv uses libcxx (clang c++ stdlib) instead of gcc stdlibc++
 
-  # Runtimes
-  runtimes = callPackage ./stage-1/runtimes.nix {
-    inherit rocmUpdateScript llvm;
-    stdenv = gcc12Stdenv;
-  };
+  llvmStdenv = overrideCC llvmPackagesNoBintools.libcxxStdenv llvmPackagesNoBintools.clangUseLLVM;
+  llvmLibstdcxxStdenv = overrideCC llvmPackagesNoBintools.stdenv (
+    llvmPackagesNoBintools.libstdcxxClang.override {
+      inherit (llvmPackages_18) bintools;
+    }
+  );
+  stdenvToBuildRocmLlvm = if useLibcxx then llvmStdenv else llvmLibstdcxxStdenv;
+  gcc-include = runCommand "gcc-include" { } ''
+    mkdir -p $out
+    ln -s ${gcc-unwrapped}/include/ $out/
+    ln -s ${gcc-unwrapped}/lib/ $out/
+  '';
 
-  ## Stage 2 ##
-  # Helpers
-  bintools-unwrapped = callPackage ./stage-2/bintools-unwrapped.nix { inherit llvm lld; };
-  bintools = wrapBintoolsWith { bintools = bintools-unwrapped; };
-  rStdenv = callPackage ./stage-2/rstdenv.nix {
-    inherit
-      llvm
-      clang-unwrapped
-      lld
-      runtimes
-      bintools
-      ;
-    stdenv = gcc12Stdenv;
+  # A prefix for use as the GCC prefix when building rocmcxx
+  disallowedRefsForToolchain = [
+    stdenv.cc
+    stdenv.cc.cc
+    stdenv.cc.bintools
+    gcc-unwrapped
+    stdenvToBuildRocmLlvm
+  ];
+  gcc-prefix =
+    let
+      gccPrefixPaths = [
+        gcc-unwrapped
+        gcc-unwrapped.lib
+        glibc.dev
+      ];
+    in
+    symlinkJoin {
+      name = "gcc-prefix";
+      paths = gccPrefixPaths ++ [
+        glibc
+      ];
+      disallowedRequisites = gccPrefixPaths;
+      postBuild = ''
+        rm -rf $out/{bin,libexec,nix-support,lib64,share,etc}
+        rm $out/lib/gcc/x86_64-unknown-linux-gnu/*/plugin/include/auto-host.h
+
+        mkdir /build/tmpout
+        mv $out/* /build/tmpout
+        cp -Lr --no-preserve=mode /build/tmpout/* $out/
+        set -x
+        versionedIncludePath="$(echo $out/include/c++/*/)"
+        mv $versionedIncludePath/* $out/include/c++/
+        rm -rf $versionedIncludePath/
+
+        find $out/lib -type f -exec ${removeReferencesTo}/bin/remove-references-to -t ${gcc-unwrapped.lib} {} +
+
+        ln -s $out $out/x86_64-unknown-linux-gnu
+      '';
+    };
+  version = "6.3.1";
+  # major version of this should be the clang version ROCm forked from
+  rocmLlvmVersion = "18.0.0-${llvmSrc.rev}";
+  usefulOutputs =
+    drv:
+    builtins.filter (x: x != null) [
+      drv
+      (drv.lib or null)
+      (drv.dev or null)
+    ];
+  listUsefulOutputs = builtins.concatMap usefulOutputs;
+  llvmSrc = fetchFromGitHub {
+    # Performance improvements cherry-picked on top of rocm-6.3.x
+    # most importantly, amdgpu-early-alwaysinline memory usage fix
+    owner = "LunNova";
+    repo = "llvm-project-rocm";
+    rev = "4182046534deb851753f0d962146e5176f648893";
+    hash = "sha256-sPmYi1WiiAqnRnHVNba2nPUxGflBC01FWCTNLPlYF9c=";
   };
+  llvmSrcFixed = llvmSrc;
+  llvmMajorVersion = lib.versions.major rocmLlvmVersion;
+  # An llvmPackages (pkgs/development/compilers/llvm/) built from ROCm LLVM's source tree
+  # optionally using LLVM libcxx
+  llvmPackagesRocm = llvmPackages_18.override (_old: {
+    stdenv = stdenvToBuildRocmLlvm; # old.stdenv #llvmPackagesNoBintools.libcxxStdenv;
+
+    # not setting gitRelease = because that causes patch selection logic to use git patches
+    # ROCm LLVM is closer to 18 official
+    # gitRelease = {}; officialRelease = null;
+    officialRelease = { }; # Set but empty because we're overriding everything from it.
+    version = rocmLlvmVersion;
+    src = llvmSrcFixed;
+    monorepoSrc = llvmSrcFixed;
+    doCheck = false;
+  });
+  sysrootCompiler =
+    cc: name: paths:
+    let
+      linked = symlinkJoin { inherit name paths; };
+    in
+    runCommand name { } ''
+      set -x
+      mkdir -p $out/
+      cp --reflink=auto -rL ${linked}/* $out/
+      chmod -R +rw $out
+      mkdir -p $out/usr
+      ln -s $out/ $out/usr/local
+      mkdir -p $out/nix-support/
+      rm -rf $out/lib64 # we don't need mixed 32 bit
+      echo 'export CC=clang' >> $out/nix-support/setup-hook
+      echo 'export CXX=clang++' >> $out/nix-support/setup-hook
+      mkdir -p $out/lib/clang/${llvmMajorVersion}/lib/linux/
+      ln -s $out/lib/linux/libclang_rt.* $out/lib/clang/${llvmMajorVersion}/lib/linux/
+
+      find $out -type f -exec sed -i "s|${cc.out}|$out|g" {} +
+      find $out -type f -exec sed -i "s|${cc.dev}|$out|g" {} +
+
+      # our /include now has more than clang expects, so this specific dir still needs to point to cc.dev
+      # FIXME: could copy into a different subdir?
+      sed -i 's|set(CLANG_INCLUDE_DIRS.*$|set(CLANG_INCLUDE_DIRS "${cc.dev}/include")|g' $out/lib/cmake/clang/ClangConfig.cmake
+      ${lib.getExe rdfind} -makesymlinks true $out/ # create links *within* the sysroot to save space
+    '';
+  findClangNostdlibincPatch =
+    x:
+    (
+      (lib.strings.hasSuffix "add-nostdlibinc-flag.patch" (builtins.baseNameOf x))
+      || (lib.strings.hasSuffix "clang-at-least-16-LLVMgold-path.patch" (builtins.baseNameOf x))
+    );
+  llvmTargetsFlag = "-DLLVM_TARGETS_TO_BUILD=AMDGPU;${
+    {
+      "x86_64" = "X86";
+      "aarch64" = "AArch64";
+    }
+    .${llvmStdenv.targetPlatform.parsed.cpu.name}
+  }";
+  # -ffat-lto-objects = emit LTO object files that are compatible with non-LTO-supporting builds too
+  # FatLTO objects are a special type of fat object file that contain LTO compatible IR in addition to generated object code,
+  # instead of containing object code for multiple target architectures. This allows users to defer the choice of whether to
+  # use LTO or not to link-time, and has been a feature available in other compilers, like GCC, for some time.
+
+  tablegenUsage = x: !(lib.strings.hasInfix "llvm-tblgen" x);
+  addGccLtoCmakeFlags = !llvmPackagesRocm.stdenv.cc.isClang;
+  llvmExtraCflags =
+    "-O3 -DNDEBUG -march=skylake -mtune=znver3"
+    + (lib.optionalString addGccLtoCmakeFlags " -D_GLIBCXX_USE_CXX11_ABI=0 -flto -ffat-lto-objects -flto-compression-level=19 -Wl,-flto")
+    + (lib.optionalString llvmPackagesRocm.stdenv.cc.isClang " -flto=thin -ffat-lto-objects")
+    + (lib.optionalString profilableStdenv " -fno-omit-frame-pointer -momit-leaf-frame-pointer -gz -g1");
 in
 rec {
-  inherit
-    llvm
-    clang-unwrapped
-    lld
-    bintools
-    ;
-
-  # Runtimes
-  libc = callPackage ./stage-2/libc.nix {
-    inherit rocmUpdateScript;
-    stdenv = rStdenv;
+  inherit (llvmPackagesRocm) libunwind;
+  inherit (llvmPackagesRocm) libcxx;
+  # Pass through original attrs for debugging where non-overridden llvm/clang is getting used
+  # llvm-orig = llvmPackagesRocm.llvm; # nix why-depends --derivation .#rocmPackages.clr .#rocmPackages.llvm.llvm-orig
+  # clang-orig = llvmPackagesRocm.clang; # nix why-depends --derivation .#rocmPackages.clr .#rocmPackages.llvm.clang-orig
+  llvm = (llvmPackagesRocm.llvm.override { ninja = emptyDirectory; }).overrideAttrs (old: {
+    dontStrip = profilableStdenv;
+    nativeBuildInputs = old.nativeBuildInputs ++ [ removeReferencesTo ];
+    buildInputs = old.buildInputs ++ [
+      zstd
+      zlib
+    ];
+    env.NIX_BUILD_ID_STYLE = "fast";
+    postPatch = ''
+      ${old.postPatch or ""}
+      patchShebangs lib/OffloadArch/make_generated_offload_arch_h.sh
+    '';
+    LDFLAGS = "-Wl,--build-id=sha1,--icf=all,--compress-debug-sections=zlib";
+    cmakeFlags =
+      (builtins.filter tablegenUsage old.cmakeFlags)
+      ++ [
+        llvmTargetsFlag
+        "-DCMAKE_BUILD_TYPE=Release"
+        "-DLLVM_ENABLE_ZSTD=FORCE_ON"
+        "-DLLVM_ENABLE_ZLIB=FORCE_ON"
+        "-DLLVM_ENABLE_THREADS=ON"
+        "-DLLVM_ENABLE_LTO=Thin"
+        "-DLLVM_USE_LINKER=lld"
+        (lib.cmakeBool "LLVM_ENABLE_LIBCXX" useLibcxx)
+        "-DCLANG_DEFAULT_CXX_STDLIB=${if useLibcxx then "libc++" else "libstdc++"}"
+      ]
+      ++ lib.optionals addGccLtoCmakeFlags [
+        "-DCMAKE_AR=${gcc-unwrapped}/bin/gcc-ar"
+        "-DCMAKE_RANLIB=${gcc-unwrapped}/bin/gcc-ranlib"
+        "-DCMAKE_NM=${gcc-unwrapped}/bin/gcc-nm"
+      ]
+      ++ lib.optionals useLibcxx [
+        "-DLLVM_ENABLE_LTO=Thin"
+        "-DLLVM_USE_LINKER=lld"
+        "-DLLVM_ENABLE_LIBCXX=ON"
+      ];
+    preConfigure = ''
+      ${old.preConfigure or ""}
+      cmakeFlagsArray+=(
+        '-DCMAKE_C_FLAGS_RELEASE=${llvmExtraCflags}'
+        '-DCMAKE_CXX_FLAGS_RELEASE=${llvmExtraCflags}'
+      )
+    '';
+    # Ensure we don't leak refs to compiler that was used to bootstrap this LLVM
+    disallowedReferences = (old.disallowedReferences or [ ]) ++ disallowedRefsForToolchain;
+    postFixup = ''
+      ${old.postFixup or ""}
+      remove-references-to -t "${stdenv.cc}" "$lib/lib/libLLVMSupport.a"
+      find $lib -type f -exec remove-references-to -t ${stdenv.cc.cc} {} +
+      find $lib -type f -exec remove-references-to -t ${stdenvToBuildRocmLlvm.cc} {} +
+      find $lib -type f -exec remove-references-to -t ${stdenv.cc.bintools} {} +
+    '';
+  });
+  lld =
+    (llvmPackagesRocm.lld.override {
+      libllvm = llvm;
+      ninja = emptyDirectory;
+    }).overrideAttrs
+      (old: {
+        patches = builtins.filter (
+          x: !(lib.strings.hasSuffix "more-openbsd-program-headers.patch" (builtins.baseNameOf x))
+        ) old.patches;
+        dontStrip = profilableStdenv;
+        nativeBuildInputs = old.nativeBuildInputs ++ [
+          llvmPackagesNoBintools.lld
+          removeReferencesTo
+        ];
+        buildInputs = old.buildInputs ++ [
+          zstd
+          zlib
+        ];
+        env.NIX_BUILD_ID_STYLE = "fast";
+        LDFLAGS = "-Wl,--build-id=sha1,--icf=all,--compress-debug-sections=zlib";
+        cmakeFlags =
+          (builtins.filter tablegenUsage old.cmakeFlags)
+          ++ [
+            llvmTargetsFlag
+            "-DCMAKE_BUILD_TYPE=Release"
+            "-DLLVM_ENABLE_ZSTD=FORCE_ON"
+            "-DLLVM_ENABLE_ZLIB=FORCE_ON"
+            "-DLLVM_ENABLE_THREADS=ON"
+            "-DLLVM_ENABLE_LTO=Thin"
+            "-DLLVM_USE_LINKER=lld"
+            (lib.cmakeBool "LLVM_ENABLE_LIBCXX" useLibcxx)
+            "-DCLANG_DEFAULT_CXX_STDLIB=${if useLibcxx then "libc++" else "libstdc++"}"
+          ]
+          ++ lib.optionals addGccLtoCmakeFlags [
+            "-DCMAKE_AR=${gcc-unwrapped}/bin/gcc-ar"
+            "-DCMAKE_RANLIB=${gcc-unwrapped}/bin/gcc-ranlib"
+            "-DCMAKE_NM=${gcc-unwrapped}/bin/gcc-nm"
+          ]
+          ++ lib.optionals useLibcxx [
+            "-DLLVM_ENABLE_LIBCXX=ON"
+          ];
+        # Ensure we don't leak refs to compiler that was used to bootstrap this LLVM
+        disallowedReferences = (old.disallowedReferences or [ ]) ++ disallowedRefsForToolchain;
+        postFixup = ''
+          ${old.postFixup or ""}
+          find $lib -type f -exec remove-references-to -t ${stdenv.cc.cc} {} +
+          find $lib -type f -exec remove-references-to -t ${stdenv.cc.bintools} {} +
+        '';
+        preConfigure = ''
+          ${old.preConfigure or ""}
+          cmakeFlagsArray+=(
+            '-DCMAKE_C_FLAGS_RELEASE=${llvmExtraCflags}'
+            '-DCMAKE_CXX_FLAGS_RELEASE=${llvmExtraCflags}'
+          )
+        '';
+      });
+  clang-unwrapped =
+    (
+      (llvmPackagesRocm.clang-unwrapped.override {
+        libllvm = llvm;
+        ninja = emptyDirectory;
+      }).overrideAttrs
+      (
+        old:
+        let
+          filteredPatches = builtins.filter (x: !(findClangNostdlibincPatch x)) old.patches;
+        in
+        {
+          meta.platforms = [
+            "x86_64-linux"
+          ];
+          pname = "${old.pname}-rocm";
+          patches = filteredPatches ++ [
+            # Never add FHS include paths
+            ./clang-bodge-ignore-systemwide-incls.diff
+            # Prevents builds timing out if a single compiler invocation is very slow but
+            # per-arch jobs are completing by ensuring there's terminal output
+            ./clang-log-jobs.diff
+            (fetchpatch {
+              # [ClangOffloadBundler]: Add GetBundleIDsInFile to OffloadBundler
+              sha256 = "sha256-G/mzUdFfrJ2bLJgo4+mBcR6Ox7xGhWu5X+XxT4kH2c8=";
+              url = "https://github.com/GZGavinZhao/rocm-llvm-project/commit/6d296f879b0fed830c54b2a9d26240da86c8bb3a.patch";
+              relative = "clang";
+            })
+            # FIXME: Needed due to https://github.com/NixOS/nixpkgs/issues/375431
+            # Once we can switch to overrideScope this can be removed
+            (substituteAll {
+              src = ./../../../compilers/llvm/common/clang/clang-at-least-16-LLVMgold-path.patch;
+              libllvmLibdir = "${llvm.lib}/lib";
+            })
+          ];
+          nativeBuildInputs = old.nativeBuildInputs ++ [
+            llvmPackagesNoBintools.lld
+            removeReferencesTo
+          ];
+          buildInputs = old.buildInputs ++ [
+            zstd
+            zlib
+          ];
+          dontStrip = profilableStdenv;
+          LDFLAGS = "-Wl,--build-id=sha1,--icf=all,--compress-debug-sections=zlib";
+          env = (old.env or { }) // {
+            NIX_BUILD_ID_STYLE = "fast";
+          };
+          # Ensure we don't leak refs to compiler that was used to bootstrap this LLVM
+          disallowedReferences = (old.disallowedReferences or [ ]) ++ disallowedRefsForToolchain;
+          requiredSystemFeatures = (old.requiredSystemFeatures or [ ]) ++ [ "big-parallel" ];
+          # https://github.com/llvm/llvm-project/blob/6976deebafa8e7de993ce159aa6b82c0e7089313/clang/cmake/caches/DistributionExample-stage2.cmake#L9-L11
+          cmakeFlags =
+            (builtins.filter tablegenUsage old.cmakeFlags)
+            ++ [
+              llvmTargetsFlag
+              "-DCMAKE_BUILD_TYPE=Release"
+              "-DLLVM_ENABLE_ZSTD=FORCE_ON"
+              "-DLLVM_ENABLE_ZLIB=FORCE_ON"
+              "-DLLVM_ENABLE_THREADS=ON"
+              "-DLLVM_ENABLE_LTO=Thin"
+              "-DLLVM_USE_LINKER=lld"
+              (lib.cmakeBool "LLVM_ENABLE_LIBCXX" useLibcxx)
+              "-DCLANG_DEFAULT_CXX_STDLIB=${if useLibcxx then "libc++" else "libstdc++"}"
+            ]
+            ++ lib.optionals addGccLtoCmakeFlags [
+              "-DCMAKE_AR=${gcc-unwrapped}/bin/gcc-ar"
+              "-DCMAKE_RANLIB=${gcc-unwrapped}/bin/gcc-ranlib"
+              "-DCMAKE_NM=${gcc-unwrapped}/bin/gcc-nm"
+            ]
+            ++ lib.optionals useLibcxx [
+              "-DLLVM_ENABLE_LTO=Thin"
+              "-DLLVM_ENABLE_LIBCXX=ON"
+              "-DLLVM_USE_LINKER=lld"
+              "-DCLANG_DEFAULT_RTLIB=compiler-rt"
+            ]
+            ++ lib.optionals (!useLibcxx) [
+              # FIXME: Config file in rocmcxx instead of GCC_INSTALL_PREFIX?
+              "-DGCC_INSTALL_PREFIX=${gcc-prefix}"
+            ];
+          postFixup =
+            (old.postFixup or "")
+            + ''
+              find $lib -type f -exec remove-references-to -t ${stdenv.cc.cc} {} +
+              find $lib -type f -exec remove-references-to -t ${stdenv.cc.bintools} {} +
+            '';
+          preConfigure =
+            (old.preConfigure or "")
+            + ''
+              cmakeFlagsArray+=(
+                '-DCMAKE_C_FLAGS_RELEASE=${llvmExtraCflags}'
+                '-DCMAKE_CXX_FLAGS_RELEASE=${llvmExtraCflags}'
+              )
+            '';
+        }
+      )
+    )
+    // {
+      libllvm = llvm;
+    };
+  # A clang that understands standard include searching in a GNU sysroot and will put GPU libs in include path
+  # in the right order
+  # and expects its libc to be in the sysroot
+  rocmcxx =
+    (sysrootCompiler clang-unwrapped "rocmcxx" (
+      listUsefulOutputs (
+        [
+          clang-unwrapped
+          bintools
+          compiler-rt
+        ]
+        ++ (lib.optionals useLibcxx [
+          libcxx
+        ])
+        ++ (lib.optionals (!useLibcxx) [
+          gcc-include
+          glibc
+          glibc.dev
+        ])
+      )
+    ))
+    // {
+      version = llvmMajorVersion;
+      cc = rocmcxx;
+      libllvm = llvm;
+      isClang = true;
+      isGNU = false;
+    };
+  clang-tools = llvmPackagesRocm.clang-tools.override {
+    inherit clang-unwrapped clang;
   };
-  libunwind = callPackage ./stage-2/libunwind.nix {
-    inherit rocmUpdateScript;
-    stdenv = rStdenv;
-  };
-  libcxxabi = callPackage ./stage-2/libcxxabi.nix {
-    inherit rocmUpdateScript;
-    stdenv = rStdenv;
-  };
-  libcxx = callPackage ./stage-2/libcxx.nix {
-    inherit rocmUpdateScript;
-    stdenv = rStdenv;
-  };
-  compiler-rt = callPackage ./stage-2/compiler-rt.nix {
-    inherit rocmUpdateScript llvm;
-    stdenv = rStdenv;
+  compiler-rt-libc = llvmPackagesRocm.compiler-rt-libc.overrideAttrs (old: {
+    patches = old.patches ++ [
+      (fetchpatch {
+        name = "Fix-missing-main-function-in-float16-bfloat16-support-checks.patch";
+        url = "https://github.com/ROCm/llvm-project/commit/68d8b3846ab1e6550910f2a9a685690eee558af2.patch";
+        hash = "sha256-Db+L1HFMWVj4CrofsGbn5lnMoCzEcU+7q12KKFb17/g=";
+        relative = "compiler-rt";
+      })
+    ];
+  });
+  compiler-rt = compiler-rt-libc;
+  bintools = wrapBintoolsWith {
+    bintools = llvmPackagesRocm.bintools-unwrapped.override {
+      inherit lld llvm;
+    };
   };
 
-  ## Stage 3 ##
-  # Helpers
-  clang = callPackage ./stage-3/clang.nix {
-    inherit
-      llvm
-      lld
-      clang-unwrapped
-      bintools
-      libc
-      libunwind
-      libcxxabi
-      libcxx
-      compiler-rt
-      ;
-    stdenv = gcc12Stdenv;
+  clang = rocmcxx;
+
+  # Emulate a monolithic ROCm LLVM build to support building ROCm's in-tree LLVM projects
+  rocm-merged-llvm = symlinkJoin {
+    name = "rocm-llvm-merge";
+    paths =
+      [
+        llvm
+        llvm.dev
+        lld
+        lld.lib
+        lld.dev
+        libunwind
+        libunwind.dev
+        compiler-rt
+        compiler-rt.dev
+        rocmcxx
+      ]
+      ++ lib.optionals useLibcxx [
+        libcxx
+        libcxx.out
+        libcxx.dev
+      ];
+    postBuild = builtins.unsafeDiscardStringContext ''
+      found_files=$(find $out -name '*.cmake')
+      if [ -z "$found_files" ]; then
+          >&2 echo "Error: No CMake files found in $out"
+          exit 1
+      fi
+
+      for target in ${clang-unwrapped.out} ${clang-unwrapped.lib} ${clang-unwrapped.dev}; do
+        if grep "$target" $found_files; then
+            >&2 echo "Unexpected ref to $target (clang-unwrapped) found"
+            # exit 1
+            # # FIXME: enable this to reduce closure size
+        fi
+      done
+    '';
+    inherit version;
+    llvm-src = llvmSrc;
   };
-  rocmClangStdenv = overrideCC gcc12Stdenv clang;
+
+  rocmClangStdenv = overrideCC (
+    if useLibcxx then llvmPackagesRocm.libcxxStdenv else llvmPackagesRocm.stdenv
+  ) clang;
 
   # Projects
-  clang-tools-extra = callPackage ./stage-3/clang-tools-extra.nix {
-    inherit rocmUpdateScript llvm clang-unwrapped;
-    stdenv = rocmClangStdenv;
-  };
-  libclc = callPackage ./stage-3/libclc.nix {
-    inherit rocmUpdateScript llvm clang;
-    stdenv = rocmClangStdenv;
-  };
-  lldb = callPackage ./stage-3/lldb.nix {
-    inherit rocmUpdateScript clang;
-    stdenv = rocmClangStdenv;
-  };
-  mlir = callPackage ./stage-3/mlir.nix {
-    inherit rocmUpdateScript clr;
-    stdenv = rocmClangStdenv;
-  };
-  polly = callPackage ./stage-3/polly.nix {
-    inherit rocmUpdateScript;
-    stdenv = rocmClangStdenv;
-  };
-  flang = callPackage ./stage-3/flang.nix {
-    inherit rocmUpdateScript clang-unwrapped mlir;
-    stdenv = rocmClangStdenv;
-  };
-  openmp = callPackage ./stage-3/openmp.nix {
-    inherit
-      rocmUpdateScript
-      llvm
-      clang-unwrapped
-      clang
-      rocm-device-libs
-      rocm-runtime
-      rocm-thunk
-      ;
-    stdenv = rocmClangStdenv;
-  };
-
-  # Runtimes
-  pstl = callPackage ./stage-3/pstl.nix {
-    inherit rocmUpdateScript;
-    stdenv = rocmClangStdenv;
-  };
+  openmp =
+    (llvmPackagesRocm.openmp.override {
+      stdenv = rocmClangStdenv;
+      llvm = rocm-merged-llvm;
+      targetLlvm = rocm-merged-llvm;
+      clang-unwrapped = clang;
+    }).overrideAttrs
+      (old: {
+        disallowedReferences = (old.disallowedReferences or [ ]) ++ disallowedRefsForToolchain;
+        nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [ removeReferencesTo ];
+        cmakeFlags =
+          old.cmakeFlags
+          ++ [
+            "-DDEVICELIBS_ROOT=${rocm-device-libs.src}"
+            # OMPD support is broken in ROCm 6.3. Haven't investigated why.
+            "-DLIBOMP_OMPD_SUPPORT:BOOL=FALSE"
+            "-DLIBOMP_OMPD_GDB_SUPPORT:BOOL=FALSE"
+          ]
+          ++ lib.optionals addGccLtoCmakeFlags [
+            "-DCMAKE_AR=${gcc-unwrapped}/bin/gcc-ar"
+            "-DCMAKE_RANLIB=${gcc-unwrapped}/bin/gcc-ranlib"
+          ];
+        env.LLVM = "${rocm-merged-llvm}";
+        env.LLVM_DIR = "${rocm-merged-llvm}";
+        buildInputs = old.buildInputs ++ [
+          rocm-device-libs
+          rocm-runtime
+          zlib
+          zstd
+          libxml2
+          libffi
+        ];
+      });
 }

pkgs/development/rocm-modules/6/llvm/stage-1/clang-unwrapped.nix

@@ -1,48 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-  llvm,
-}:
-
-callPackage ../base.nix {
-  inherit stdenv rocmUpdateScript;
-  targetName = "clang-unwrapped";
-  targetDir = "clang";
-  extraBuildInputs = [ llvm ];
-
-  extraCMakeFlags = [
-    "-DCLANG_INCLUDE_DOCS=ON"
-    "-DCLANG_INCLUDE_TESTS=ON"
-  ];
-
-  extraPostPatch = ''
-    # Looks like they forgot to add finding libedit to the standalone build
-    ln -s ../cmake/Modules/FindLibEdit.cmake cmake/modules
-
-    substituteInPlace CMakeLists.txt \
-      --replace-fail "include(CheckIncludeFile)" "include(CheckIncludeFile)''\nfind_package(LibEdit)"
-
-    # `No such file or directory: '/build/source/clang/tools/scan-build/bin/scan-build'`
-    rm test/Analysis/scan-build/*.test
-    rm test/Analysis/scan-build/rebuild_index/rebuild_index.test
-
-    # `does not depend on a module exporting 'baz.h'`
-    rm test/Modules/header-attribs.cpp
-
-    # We do not have HIP or the ROCm stack available yet
-    rm test/Driver/hip-options.hip
-
-    # ???? `ld: cannot find crti.o: No such file or directory` linker issue?
-    rm test/Interpreter/dynamic-library.cpp
-
-    # `fatal error: 'stdio.h' file not found`
-    rm test/OpenMP/amdgcn_emit_llvm.c
-  '';
-
-  extraPostInstall = ''
-    mv bin/clang-tblgen $out/bin
-  '';
-
-  requiredSystemFeatures = [ "big-parallel" ];
-}

pkgs/development/rocm-modules/6/llvm/stage-1/lld.nix

@@ -1,15 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-  llvm,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  buildMan = false; # No man pages to build
-  targetName = "lld";
-  targetDir = targetName;
-  extraBuildInputs = [ llvm ];
-  checkTargets = [ "check-${targetName}" ];
-}

pkgs/development/rocm-modules/6/llvm/stage-1/llvm.nix

@@ -1,11 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-}:
-
-callPackage ../base.nix {
-  inherit stdenv rocmUpdateScript;
-  requiredSystemFeatures = [ "big-parallel" ];
-  isBroken = stdenv.hostPlatform.isAarch64; # https://github.com/ROCm/ROCm/issues/1831#issuecomment-1278205344
-}

pkgs/development/rocm-modules/6/llvm/stage-1/runtimes.nix

@@ -1,32 +0,0 @@
-{
-  lib,
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-  llvm,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  buildDocs = false;
-  buildMan = false;
-  buildTests = false;
-  targetName = "runtimes";
-  targetDir = targetName;
-
-  targetRuntimes = [
-    "libunwind"
-    "libcxxabi"
-    "libcxx"
-    "compiler-rt"
-  ];
-
-  extraBuildInputs = [ llvm ];
-
-  extraCMakeFlags = [
-    "-DLIBCXX_INCLUDE_BENCHMARKS=OFF"
-    "-DLIBCXX_CXX_ABI=libcxxabi"
-  ];
-
-  extraLicenses = [ lib.licenses.mit ];
-}

pkgs/development/rocm-modules/6/llvm/stage-2/1000-libcxx-failing-tests.list

@@ -1,176 +0,0 @@
-../libcxx/test/libcxx/containers/gnu_cxx/hash_map.pass.cpp
-../libcxx/test/libcxx/containers/gnu_cxx/hash_set.pass.cpp
-../libcxx/test/libcxx/depr/depr.default.allocator/allocator.members/allocate.cxx2a.pass.cpp
-../libcxx/test/libcxx/depr/depr.default.allocator/allocator.members/construct.cxx2a.pass.cpp
-../libcxx/test/libcxx/input.output/filesystems/class.directory_entry/directory_entry.mods/last_write_time.pass.cpp
-../libcxx/test/libcxx/input.output/filesystems/class.path/path.member/path.native.obs/string_alloc.pass.cpp
-../libcxx/test/libcxx/language.support/support.dynamic/libcpp_deallocate.sh.cpp
-../libcxx/test/libcxx/localization/locales/locale/locale.types/locale.facet/no_allocation.pass.cpp
-../libcxx/test/libcxx/utilities/utility/mem.res/mem.res.monotonic.buffer/mem.res.monotonic.buffer.mem/allocate_from_underaligned_buffer.pass.cpp
-../libcxx/test/libcxx/utilities/utility/mem.res/mem.res.monotonic.buffer/mem.res.monotonic.buffer.mem/allocate_in_geometric_progression.pass.cpp
-../libcxx/test/std/containers/associative/map/map.access/index_key.pass.cpp
-../libcxx/test/std/containers/associative/map/map.access/index_rv_key.pass.cpp
-../libcxx/test/std/containers/associative/map/map.modifiers/insert_and_emplace_allocator_requirements.pass.cpp
-../libcxx/test/std/containers/associative/multimap/multimap.modifiers/insert_allocator_requirements.pass.cpp
-../libcxx/test/std/containers/associative/multiset/insert_emplace_allocator_requirements.pass.cpp
-../libcxx/test/std/containers/associative/set/insert_and_emplace_allocator_requirements.pass.cpp
-../libcxx/test/std/containers/sequences/list/list.modifiers/insert_iter_iter_iter.pass.cpp
-../libcxx/test/std/containers/sequences/list/list.modifiers/insert_iter_size_value.pass.cpp
-../libcxx/test/std/containers/sequences/list/list.modifiers/insert_iter_value.pass.cpp
-../libcxx/test/std/containers/sequences/vector.bool/ctor_exceptions.pass.cpp
-../libcxx/test/std/containers/sequences/vector/vector.cons/exceptions.pass.cpp
-../libcxx/test/std/containers/unord/unord.map/unord.map.elem/index.pass.cpp
-../libcxx/test/std/containers/unord/unord.map/unord.map.modifiers/insert_and_emplace_allocator_requirements.pass.cpp
-../libcxx/test/std/containers/unord/unord.multimap/unord.multimap.modifiers/insert_allocator_requirements.pass.cpp
-../libcxx/test/std/containers/unord/unord.multiset/insert_emplace_allocator_requirements.pass.cpp
-../libcxx/test/std/containers/unord/unord.set/insert_and_emplace_allocator_requirements.pass.cpp
-../libcxx/test/std/experimental/memory/memory.resource.global/new_delete_resource.pass.cpp
-../libcxx/test/std/experimental/memory/memory.resource.global/null_memory_resource.pass.cpp
-../libcxx/test/std/input.output/file.streams/fstreams/filebuf.virtuals/pbackfail.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.cons/copy_assign.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.cons/copy.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.cons/move_assign.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.cons/move.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.cons/path.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.mods/assign.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.mods/refresh.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.mods/replace_filename.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.obs/file_size.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.obs/file_type_obs.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.obs/hard_link_count.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.obs/last_write_time.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.obs/status.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_entry/directory_entry.obs/symlink_status.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_iterator/directory_iterator.members/copy_assign.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_iterator/directory_iterator.members/copy.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_iterator/directory_iterator.members/ctor.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_iterator/directory_iterator.members/increment.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_iterator/directory_iterator.members/move_assign.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_iterator/directory_iterator.members/move.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.directory_iterator/directory_iterator.nonmembers/begin_end.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.append.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.assign/move.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.assign/source.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.compare.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.concat.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.construct/move.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.decompose/path.decompose.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.gen/lexically_normal.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.gen/lexically_relative_and_proximate.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.generic.obs/generic_string_alloc.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.generic.obs/named_overloads.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.modifiers/clear.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.modifiers/make_preferred.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.modifiers/remove_filename.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.modifiers/replace_extension.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.modifiers/replace_filename.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.modifiers/swap.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.member/path.native.obs/named_overloads.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.nonmember/path.factory.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.nonmember/path.io.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.path/path.nonmember/swap.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.rec.dir.itr/rec.dir.itr.members/copy_assign.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.rec.dir.itr/rec.dir.itr.members/copy.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.rec.dir.itr/rec.dir.itr.members/ctor.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.rec.dir.itr/rec.dir.itr.members/depth.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.rec.dir.itr/rec.dir.itr.members/disable_recursion_pending.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.rec.dir.itr/rec.dir.itr.members/increment.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.rec.dir.itr/rec.dir.itr.members/move_assign.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.rec.dir.itr/rec.dir.itr.members/move.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.rec.dir.itr/rec.dir.itr.members/pop.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.rec.dir.itr/rec.dir.itr.members/recursion_pending.pass.cpp
-../libcxx/test/std/input.output/filesystems/class.rec.dir.itr/rec.dir.itr.nonmembers/begin_end.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.canonical/canonical.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.copy_file/copy_file_large.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.copy_file/copy_file.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.copy_symlink/copy_symlink.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.copy/copy.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.create_directories/create_directories.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.create_directory_symlink/create_directory_symlink.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.create_directory/create_directory_with_attributes.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.create_directory/create_directory.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.create_hard_link/create_hard_link.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.create_symlink/create_symlink.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.current_path/current_path.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.equivalent/equivalent.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.exists/exists.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.file_size/file_size.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.hard_lk_ct/hard_link_count.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.is_block_file/is_block_file.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.is_char_file/is_character_file.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.is_directory/is_directory.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.is_empty/is_empty.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.is_fifo/is_fifo.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.is_other/is_other.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.is_regular_file/is_regular_file.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.is_socket/is_socket.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.is_symlink/is_symlink.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.last_write_time/last_write_time.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.permissions/permissions.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.proximate/proximate.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.read_symlink/read_symlink.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.relative/relative.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.remove_all/remove_all.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.remove_all/toctou.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.remove/remove.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.rename/rename.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.resize_file/resize_file.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.space/space.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.status/status.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.symlink_status/symlink_status.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.temp_dir_path/temp_directory_path.pass.cpp
-../libcxx/test/std/input.output/filesystems/fs.op.funcs/fs.op.weakly_canonical/weakly_canonical.pass.cpp
-../libcxx/test/std/localization/locale.categories/category.ctype/facet.ctype.special/facet.ctype.char.dtor/dtor.pass.cpp
-../libcxx/test/std/localization/locale.stdcvt/codecvt_utf16.pass.cpp
-../libcxx/test/std/localization/locale.stdcvt/codecvt_utf8.pass.cpp
-../libcxx/test/std/localization/locales/locale.convenience/conversions/conversions.buffer/ctor.pass.cpp
-../libcxx/test/std/localization/locales/locale/locale.members/combine.pass.cpp
-../libcxx/test/std/strings/basic.string/string.cons/substr_rvalue.pass.cpp
-../libcxx/test/std/utilities/any/any.class/any.assign/copy.pass.cpp
-../libcxx/test/std/utilities/any/any.class/any.assign/value.pass.cpp
-../libcxx/test/std/utilities/any/any.class/any.cons/copy.pass.cpp
-../libcxx/test/std/utilities/any/any.class/any.cons/default.pass.cpp
-../libcxx/test/std/utilities/any/any.class/any.cons/in_place_type.pass.cpp
-../libcxx/test/std/utilities/any/any.class/any.cons/move.pass.cpp
-../libcxx/test/std/utilities/any/any.class/any.cons/value.pass.cpp
-../libcxx/test/std/utilities/any/any.class/any.modifiers/emplace.pass.cpp
-../libcxx/test/std/utilities/any/any.nonmembers/any.cast/any_cast_reference.pass.cpp
-../libcxx/test/std/utilities/any/any.nonmembers/make_any.pass.cpp
-../libcxx/test/std/utilities/function.objects/func.wrap/func.wrap.func/func.wrap.func.alg/swap.pass.cpp
-../libcxx/test/std/utilities/function.objects/func.wrap/func.wrap.func/func.wrap.func.con/copy_assign.pass.cpp
-../libcxx/test/std/utilities/function.objects/func.wrap/func.wrap.func/func.wrap.func.con/copy_move.pass.cpp
-../libcxx/test/std/utilities/function.objects/func.wrap/func.wrap.func/func.wrap.func.con/F_assign.pass.cpp
-../libcxx/test/std/utilities/function.objects/func.wrap/func.wrap.func/func.wrap.func.con/F.pass.cpp
-../libcxx/test/std/utilities/function.objects/func.wrap/func.wrap.func/func.wrap.func.con/nullptr_t_assign.pass.cpp
-../libcxx/test/std/utilities/function.objects/func.wrap/func.wrap.func/func.wrap.func.mod/swap.pass.cpp
-../libcxx/test/std/utilities/memory/default.allocator/allocator.members/allocate_at_least.pass.cpp
-../libcxx/test/std/utilities/memory/default.allocator/allocator.members/allocate.pass.cpp
-../libcxx/test/std/utilities/memory/util.smartptr/util.smartptr.enab/enable_shared_from_this.pass.cpp
-../libcxx/test/std/utilities/memory/util.smartptr/util.smartptr.shared/util.smartptr.shared.const/nullptr_t_deleter_throw.pass.cpp
-../libcxx/test/std/utilities/memory/util.smartptr/util.smartptr.shared/util.smartptr.shared.const/pointer_deleter_throw.pass.cpp
-../libcxx/test/std/utilities/memory/util.smartptr/util.smartptr.shared/util.smartptr.shared.const/pointer_throw.pass.cpp
-../libcxx/test/std/utilities/memory/util.smartptr/util.smartptr.shared/util.smartptr.shared.const/unique_ptr.pass.cpp
-../libcxx/test/std/utilities/memory/util.smartptr/util.smartptr.shared/util.smartptr.shared.create/make_shared.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.global/new_delete_resource.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.global/null_memory_resource.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.monotonic.buffer/mem.res.monotonic.buffer.ctor/without_buffer.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.monotonic.buffer/mem.res.monotonic.buffer.mem/allocate_deallocate.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.monotonic.buffer/mem.res.monotonic.buffer.mem/allocate_exception_safety.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.monotonic.buffer/mem.res.monotonic.buffer.mem/allocate_from_initial_buffer.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.monotonic.buffer/mem.res.monotonic.buffer.mem/allocate_from_zero_sized_buffer.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.monotonic.buffer/mem.res.monotonic.buffer.mem/allocate_in_geometric_progression.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.monotonic.buffer/mem.res.monotonic.buffer.mem/allocate_overaligned_request.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.monotonic.buffer/mem.res.monotonic.buffer.mem/allocate_with_initial_size.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.pool/mem.res.pool.ctor/ctor_does_not_allocate.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.pool/mem.res.pool.mem/equality.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.pool/mem.res.pool.mem/sync_allocate_overaligned_request.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.pool/mem.res.pool.mem/sync_allocate_reuse_blocks.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.pool/mem.res.pool.mem/sync_allocate.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.pool/mem.res.pool.mem/unsync_allocate_overaligned_request.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.pool/mem.res.pool.mem/unsync_allocate_reuse_blocks.pass.cpp
-../libcxx/test/std/utilities/utility/mem.res/mem.res.pool/mem.res.pool.mem/unsync_allocate.pass.cpp
-../libcxx/test/std/language.support/support.dynamic/hardware_inference_size.compile.pass.cpp
-../libcxx/test/std/language.support/support.dynamic/new.delete/new.delete.array/sized_delete_array14.pass.cpp
-../libcxx/test/std/language.support/support.dynamic/new.delete/new.delete.single/sized_delete14.pass.cpp
-../libcxx/test/libcxx/selftest/sh.cpp/empty.sh.cpp
-../libcxx/test/libcxx/transitive_includes.sh.cpp

pkgs/development/rocm-modules/6/llvm/stage-2/bintools-unwrapped.nix

@@ -1,29 +0,0 @@
-{
-  runCommand,
-  llvm,
-  lld,
-}:
-
-runCommand "rocm-llvm-binutils-${llvm.version}" { preferLocalBuild = true; } ''
-  mkdir -p $out/bin
-
-  for prog in ${lld}/bin/*; do
-    ln -s $prog $out/bin/$(basename $prog)
-  done
-
-  for prog in ${llvm}/bin/*; do
-    ln -sf $prog $out/bin/$(basename $prog)
-  done
-
-  ln -s ${llvm}/bin/llvm-ar $out/bin/ar
-  ln -s ${llvm}/bin/llvm-as $out/bin/as
-  ln -s ${llvm}/bin/llvm-dwp $out/bin/dwp
-  ln -s ${llvm}/bin/llvm-nm $out/bin/nm
-  ln -s ${llvm}/bin/llvm-objcopy $out/bin/objcopy
-  ln -s ${llvm}/bin/llvm-objdump $out/bin/objdump
-  ln -s ${llvm}/bin/llvm-ranlib $out/bin/ranlib
-  ln -s ${llvm}/bin/llvm-readelf $out/bin/readelf
-  ln -s ${llvm}/bin/llvm-size $out/bin/size
-  ln -s ${llvm}/bin/llvm-strip $out/bin/strip
-  ln -s ${lld}/bin/lld $out/bin/ld
-''

pkgs/development/rocm-modules/6/llvm/stage-2/compiler-rt.nix

@@ -1,64 +0,0 @@
-{
-  lib,
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-  llvm,
-  glibc,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  buildDocs = false; # No documentation to build
-  buildMan = false; # No man pages to build
-  targetName = "compiler-rt";
-  targetDir = "runtimes";
-
-  targetRuntimes = [
-    "libunwind"
-    "libcxxabi"
-    "libcxx"
-    targetName
-  ];
-
-  extraCMakeFlags = [
-    "-DCOMPILER_RT_INCLUDE_TESTS=ON"
-    "-DCOMPILER_RT_USE_LLVM_UNWINDER=ON"
-    "-DCOMPILER_RT_CXX_LIBRARY=libcxx"
-    "-DCOMPILER_RT_CAN_EXECUTE_TESTS=OFF" # We can't run most of these
-
-    # Workaround having to build combined
-    "-DLIBUNWIND_INCLUDE_DOCS=OFF"
-    "-DLIBUNWIND_INCLUDE_TESTS=OFF"
-    "-DLIBUNWIND_USE_COMPILER_RT=ON"
-    "-DLIBUNWIND_INSTALL_LIBRARY=OFF"
-    "-DLIBUNWIND_INSTALL_HEADERS=OFF"
-    "-DLIBCXXABI_INCLUDE_TESTS=OFF"
-    "-DLIBCXXABI_USE_LLVM_UNWINDER=ON"
-    "-DLIBCXXABI_USE_COMPILER_RT=ON"
-    "-DLIBCXXABI_INSTALL_LIBRARY=OFF"
-    "-DLIBCXXABI_INSTALL_HEADERS=OFF"
-    "-DLIBCXX_INCLUDE_DOCS=OFF"
-    "-DLIBCXX_INCLUDE_TESTS=OFF"
-    "-DLIBCXX_USE_COMPILER_RT=ON"
-    "-DLIBCXX_CXX_ABI=libcxxabi"
-    "-DLIBCXX_INSTALL_LIBRARY=OFF"
-    "-DLIBCXX_INSTALL_HEADERS=OFF"
-  ];
-
-  extraPostPatch = ''
-    # `No such file or directory: 'ldd'`
-    substituteInPlace ../compiler-rt/test/lit.common.cfg.py \
-      --replace "'ldd'," "'${glibc.bin}/bin/ldd',"
-
-    # We can run these
-    substituteInPlace ../compiler-rt/test/CMakeLists.txt \
-      --replace "endfunction()" "endfunction()''\nadd_subdirectory(builtins)''\nadd_subdirectory(shadowcallstack)"
-
-    # Could not launch llvm-config in /build/source/runtimes/build/bin
-    mkdir -p build/bin
-    ln -s ${llvm}/bin/llvm-config build/bin
-  '';
-
-  extraLicenses = [ lib.licenses.mit ];
-}

pkgs/development/rocm-modules/6/llvm/stage-2/libc.nix

@@ -1,27 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  buildMan = false; # No man pages to build
-  targetName = "libc";
-  targetDir = "runtimes";
-  targetRuntimes = [ targetName ];
-
-  extraPostPatch = ''
-    # `Failed to match ... against ...` `Match value not within tolerance value of MPFR result:`
-    # We need a better way, but I don't know enough sed magic and patching `CMakeLists.txt` isn't working...
-    substituteInPlace ../libc/test/src/math/log10_test.cpp \
-      --replace-fail "i < N" "i < 0" \
-      --replace-fail "test(mpfr::RoundingMode::Nearest);" "" \
-      --replace-fail "test(mpfr::RoundingMode::Downward);" "" \
-      --replace-fail "test(mpfr::RoundingMode::Upward);" "" \
-      --replace-fail "test(mpfr::RoundingMode::TowardZero);" ""
-  '';
-
-  checkTargets = [ "check-${targetName}" ];
-  hardeningDisable = [ "fortify" ]; # Prevent `error: "Assumed value of MB_LEN_MAX wrong"`
-}

pkgs/development/rocm-modules/6/llvm/stage-2/libcxx.nix

@@ -1,43 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  buildMan = false; # No man pages to build
-  targetName = "libcxx";
-  targetDir = "runtimes";
-
-  targetRuntimes = [
-    "libunwind"
-    "libcxxabi"
-    targetName
-  ];
-
-  extraCMakeFlags = [
-    "-DLIBCXX_INCLUDE_DOCS=ON"
-    "-DLIBCXX_INCLUDE_TESTS=ON"
-    "-DLIBCXX_USE_COMPILER_RT=ON"
-    "-DLIBCXX_CXX_ABI=libcxxabi"
-
-    # Workaround having to build combined
-    "-DLIBUNWIND_INCLUDE_DOCS=OFF"
-    "-DLIBUNWIND_INCLUDE_TESTS=OFF"
-    "-DLIBUNWIND_USE_COMPILER_RT=ON"
-    "-DLIBUNWIND_INSTALL_LIBRARY=OFF"
-    "-DLIBUNWIND_INSTALL_HEADERS=OFF"
-    "-DLIBCXXABI_INCLUDE_TESTS=OFF"
-    "-DLIBCXXABI_USE_LLVM_UNWINDER=ON"
-    "-DLIBCXXABI_USE_COMPILER_RT=ON"
-    "-DLIBCXXABI_INSTALL_LIBRARY=OFF"
-    "-DLIBCXXABI_INSTALL_HEADERS=OFF"
-  ];
-
-  # Most of these can't find `bash` or `mkdir`, might just be hard-coded paths, or PATH is altered
-  extraPostPatch = ''
-    chmod +w -R ../libcxx/test/{libcxx,std}
-    cat ${./1000-libcxx-failing-tests.list} | xargs -d \\n rm
-  '';
-}

pkgs/development/rocm-modules/6/llvm/stage-2/libcxxabi.nix

@@ -1,38 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  buildDocs = false; # No documentation to build
-  buildMan = false; # No man pages to build
-  targetName = "libcxxabi";
-  targetDir = "runtimes";
-
-  targetRuntimes = [
-    "libunwind"
-    targetName
-    "libcxx"
-  ];
-
-  extraCMakeFlags = [
-    "-DLIBCXXABI_INCLUDE_TESTS=ON"
-    "-DLIBCXXABI_USE_LLVM_UNWINDER=ON"
-    "-DLIBCXXABI_USE_COMPILER_RT=ON"
-
-    # Workaround having to build combined
-    "-DLIBUNWIND_INCLUDE_DOCS=OFF"
-    "-DLIBUNWIND_INCLUDE_TESTS=OFF"
-    "-DLIBUNWIND_USE_COMPILER_RT=ON"
-    "-DLIBUNWIND_INSTALL_LIBRARY=OFF"
-    "-DLIBUNWIND_INSTALL_HEADERS=OFF"
-    "-DLIBCXX_INCLUDE_DOCS=OFF"
-    "-DLIBCXX_INCLUDE_TESTS=OFF"
-    "-DLIBCXX_USE_COMPILER_RT=ON"
-    "-DLIBCXX_CXX_ABI=libcxxabi"
-    "-DLIBCXX_INSTALL_LIBRARY=OFF"
-    "-DLIBCXX_INSTALL_HEADERS=OFF"
-  ];
-}

pkgs/development/rocm-modules/6/llvm/stage-2/libunwind.nix

@@ -1,27 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  buildMan = false; # No man pages to build
-  targetName = "libunwind";
-  targetDir = "runtimes";
-  targetRuntimes = [ targetName ];
-
-  extraCMakeFlags = [
-    "-DLIBUNWIND_INCLUDE_DOCS=ON"
-    "-DLIBUNWIND_INCLUDE_TESTS=ON"
-    "-DLIBUNWIND_USE_COMPILER_RT=ON"
-  ];
-
-  extraPostPatch = ''
-    # `command had no output on stdout or stderr` (Says these unsupported tests)
-    chmod +w -R ../libunwind/test
-    rm ../libunwind/test/floatregister.pass.cpp
-    rm ../libunwind/test/unwind_leaffunction.pass.cpp
-    rm ../libunwind/test/libunwind_02.pass.cpp
-  '';
-}

pkgs/development/rocm-modules/6/llvm/stage-2/rstdenv.nix

@@ -1,37 +0,0 @@
-{
-  stdenv,
-  overrideCC,
-  wrapCCWith,
-  llvm,
-  clang-unwrapped,
-  lld,
-  runtimes,
-  bintools,
-}:
-
-overrideCC stdenv (wrapCCWith rec {
-  inherit bintools;
-  libcxx = runtimes;
-  cc = clang-unwrapped;
-  gccForLibs = stdenv.cc.cc;
-
-  extraPackages = [
-    llvm
-    lld
-  ];
-
-  nixSupport.cc-cflags = [
-    "-resource-dir=$out/resource-root"
-    "-fuse-ld=lld"
-    "-rtlib=compiler-rt"
-    "-unwindlib=libunwind"
-    "-Wno-unused-command-line-argument"
-  ];
-
-  extraBuildCommands = ''
-    clang_version=`${cc}/bin/clang -v 2>&1 | grep "clang version " | grep -E -o "[0-9.-]+"`
-    mkdir -p $out/resource-root
-    ln -s ${cc}/lib/clang/$clang_version/include $out/resource-root
-    ln -s ${runtimes}/lib $out/resource-root
-  '';
-})

pkgs/development/rocm-modules/6/llvm/stage-3/1000-openmp-failing-tests.list

@@ -1,122 +0,0 @@
-runtime/test/tasking/hidden_helper_task/gtid.cpp
-runtime/test/ompt/parallel/parallel_if0.c
-runtime/test/ompt/parallel/serialized.c
-runtime/test/ompt/teams/parallel_team.c
-runtime/test/ompt/teams/serial_teams.c
-runtime/test/ompt/teams/serialized.c
-runtime/test/ompt/teams/team.c
-libomptarget/test/api/assert.c
-libomptarget/test/api/omp_device_managed_memory.c
-libomptarget/test/api/omp_device_memory.c
-libomptarget/test/api/omp_get_device_num.c
-libomptarget/test/api/omp_host_pinned_memory.c
-libomptarget/test/api/omp_host_pinned_memory_alloc.c
-libomptarget/test/api/omp_target_memcpy_async1.c
-libomptarget/test/api/omp_target_memcpy_async2.c
-libomptarget/test/api/omp_target_memcpy_rect_async1.c
-libomptarget/test/api/omp_target_memcpy_rect_async2.c
-libomptarget/test/mapping/array_section_implicit_capture.c
-libomptarget/test/mapping/data_absent_at_exit.c
-libomptarget/test/mapping/data_member_ref.cpp
-libomptarget/test/mapping/declare_mapper_api.cpp
-libomptarget/test/mapping/declare_mapper_target.cpp
-libomptarget/test/mapping/declare_mapper_target_data.cpp
-libomptarget/test/mapping/declare_mapper_target_data_enter_exit.cpp
-libomptarget/test/mapping/firstprivate_aligned.cpp
-libomptarget/test/mapping/has_device_addr.cpp
-libomptarget/test/mapping/implicit_device_ptr.c
-libomptarget/test/mapping/is_device_ptr.cpp
-libomptarget/test/mapping/lambda_mapping.cpp
-libomptarget/test/mapping/low_alignment.c
-libomptarget/test/mapping/map_back_race.cpp
-libomptarget/test/mapping/power_of_two_alignment.c
-libomptarget/test/mapping/pr38704.c
-libomptarget/test/mapping/prelock.cpp
-libomptarget/test/mapping/present/target_data_at_exit.c
-libomptarget/test/mapping/private_mapping.c
-libomptarget/test/mapping/ptr_and_obj_motion.c
-libomptarget/test/mapping/reduction_implicit_map.cpp
-libomptarget/test/mapping/target_derefence_array_pointrs.cpp
-libomptarget/test/mapping/target_map_for_member_data.cpp
-libomptarget/test/mapping/target_update_array_extension.c
-libomptarget/test/mapping/target_use_device_addr.c
-libomptarget/test/offloading/atomic-compare-signedness.c
-libomptarget/test/offloading/bug47654.cpp
-libomptarget/test/offloading/bug49021.cpp
-libomptarget/test/offloading/bug49779.cpp
-libomptarget/test/offloading/bug50022.cpp
-libomptarget/test/offloading/bug51781.c
-libomptarget/test/offloading/bug51982.c
-libomptarget/test/offloading/bug53727.cpp
-libomptarget/test/offloading/complex_reduction.cpp
-libomptarget/test/offloading/cuda_no_devices.c
-libomptarget/test/offloading/d2d_memcpy.c
-libomptarget/test/offloading/dynamic_module.c
-libomptarget/test/offloading/dynamic_module_load.c
-libomptarget/test/offloading/global_constructor.cpp
-libomptarget/test/offloading/lone_target_exit_data.c
-libomptarget/test/offloading/memory_manager.cpp
-libomptarget/test/offloading/parallel_offloading_map.cpp
-libomptarget/test/offloading/static_linking.c
-libomptarget/test/offloading/std_complex_arithmetic.cpp
-libomptarget/test/offloading/target-teams-atomic.c
-libomptarget/test/offloading/target_constexpr_mapping.cpp
-libomptarget/test/offloading/target_critical_region.cpp
-libomptarget/test/offloading/target_depend_nowait.cpp
-libomptarget/test/offloading/target_nowait_target.cpp
-libomptarget/test/offloading/taskloop_offload_nowait.cpp
-libomptarget/test/offloading/test_libc.cpp
-libomptarget/test/ompt/veccopy.c
-libomptarget/test/ompt/veccopy_disallow_both.c
-libomptarget/test/ompt/veccopy_emi.c
-libomptarget/test/ompt/veccopy_emi_map.c
-libomptarget/test/ompt/veccopy_map.c
-libomptarget/test/ompt/veccopy_no_device_init.c
-libomptarget/test/ompt/veccopy_wrong_return.c
-libomptarget/test/api/is_initial_device.c
-libomptarget/test/mapping/declare_mapper_nested_default_mappers_array_subscript.cpp
-libomptarget/test/mapping/declare_mapper_nested_default_mappers_ptr_subscript.cpp
-libomptarget/test/mapping/declare_mapper_nested_default_mappers_var.cpp
-libomptarget/test/mapping/target_pointers_members_map.cpp
-libomptarget/test/api/omp_dynamic_shared_memory_mixed.c
-libomptarget/test/api/omp_env_vars.c
-libomptarget/test/api/omp_get_mapped_ptr.c
-libomptarget/test/api/omp_get_num_devices.c
-libomptarget/test/api/omp_get_num_devices_with_empty_target.c
-libomptarget/test/mapping/alloc_fail.c
-libomptarget/test/mapping/array_section_use_device_ptr.c
-libomptarget/test/mapping/declare_mapper_nested_default_mappers.cpp
-libomptarget/test/mapping/declare_mapper_nested_mappers.cpp
-libomptarget/test/mapping/declare_mapper_target_update.cpp
-libomptarget/test/mapping/delete_inf_refcount.c
-libomptarget/test/mapping/lambda_by_value.cpp
-libomptarget/test/mapping/ompx_hold/omp_target_disassociate_ptr.c
-libomptarget/test/mapping/ompx_hold/struct.c
-libomptarget/test/mapping/ompx_hold/target-data.c
-libomptarget/test/mapping/ompx_hold/target.c
-libomptarget/test/mapping/present/target.c
-libomptarget/test/mapping/present/target_array_extension.c
-libomptarget/test/mapping/present/target_data.c
-libomptarget/test/mapping/present/target_data_array_extension.c
-libomptarget/test/mapping/present/target_enter_data.c
-libomptarget/test/mapping/present/target_exit_data_delete.c
-libomptarget/test/mapping/present/target_exit_data_release.c
-libomptarget/test/mapping/present/target_update.c
-libomptarget/test/mapping/present/target_update_array_extension.c
-libomptarget/test/mapping/present/zero_length_array_section.c
-libomptarget/test/mapping/present/zero_length_array_section_exit.c
-libomptarget/test/mapping/target_data_array_extension_at_exit.c
-libomptarget/test/mapping/target_has_device_addr.c
-libomptarget/test/mapping/target_implicit_partial_map.c
-libomptarget/test/mapping/target_wrong_use_device_addr.c
-libomptarget/test/offloading/host_as_target.c
-libomptarget/test/offloading/info.c
-libomptarget/test/offloading/offloading_success.c
-libomptarget/test/offloading/offloading_success.cpp
-libomptarget/test/offloading/wtime.c
-libomptarget/test/unified_shared_memory/api.c
-libomptarget/test/unified_shared_memory/associate_ptr.c
-libomptarget/test/unified_shared_memory/close_enter_exit.c
-libomptarget/test/unified_shared_memory/close_manual.c
-libomptarget/test/unified_shared_memory/close_member.c
-libomptarget/test/unified_shared_memory/close_modifier.c

pkgs/development/rocm-modules/6/llvm/stage-3/1001-mlir-failing-tests.list

@@ -1,11 +0,0 @@
-./test/Target/LLVMIR/openmp-llvm.mlir
-./test/mlir-spirv-cpu-runner/double.mlir
-./test/mlir-spirv-cpu-runner/simple_add.mlir
-./test/mlir-vulkan-runner/addf.mlir
-./test/mlir-vulkan-runner/addi.mlir
-./test/mlir-vulkan-runner/addi8.mlir
-./test/mlir-vulkan-runner/mulf.mlir
-./test/mlir-vulkan-runner/smul_extended.mlir
-./test/mlir-vulkan-runner/subf.mlir
-./test/mlir-vulkan-runner/time.mlir
-./test/mlir-vulkan-runner/umul_extended.mlir

pkgs/development/rocm-modules/6/llvm/stage-3/clang-tools-extra.nix

@@ -1,43 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-  llvm,
-  clang-unwrapped,
-  gtest,
-}:
-
-callPackage ../base.nix {
-  inherit stdenv rocmUpdateScript;
-  buildTests = false; # `invalid operands to binary expression ('std::basic_stringstream<char>' and 'const llvm::StringRef')`
-  targetName = "clang-tools-extra";
-
-  targetProjects = [
-    "clang"
-    "clang-tools-extra"
-  ];
-
-  extraBuildInputs = [ gtest ];
-
-  extraCMakeFlags = [
-    "-DLLVM_INCLUDE_DOCS=OFF"
-    "-DLLVM_INCLUDE_TESTS=OFF"
-    "-DCLANG_INCLUDE_DOCS=OFF"
-    "-DCLANG_INCLUDE_TESTS=ON"
-    "-DCLANG_TOOLS_EXTRA_INCLUDE_DOCS=ON"
-  ];
-
-  extraPostInstall = ''
-    # Remove LLVM and Clang
-    for path in `find ${llvm} ${clang-unwrapped}`; do
-      if [ $path != ${llvm} ] && [ $path != ${clang-unwrapped} ]; then
-        rm -f $out''${path#${llvm}} $out''${path#${clang-unwrapped}} || true
-      fi
-    done
-
-    # Cleanup empty directories
-    find $out -type d -empty -delete
-  '';
-
-  requiredSystemFeatures = [ "big-parallel" ];
-}

pkgs/development/rocm-modules/6/llvm/stage-3/clang.nix

@@ -1,77 +0,0 @@
-{
-  stdenv,
-  wrapCCWith,
-  llvm,
-  lld,
-  clang-unwrapped,
-  bintools,
-  libc,
-  libunwind,
-  libcxxabi,
-  libcxx,
-  compiler-rt,
-}:
-
-wrapCCWith rec {
-  inherit libcxx bintools;
-
-  # We do this to avoid HIP pathing problems, and mimic a monolithic install
-  cc = stdenv.mkDerivation (finalAttrs: {
-    inherit (clang-unwrapped) version;
-    pname = "rocm-llvm-clang";
-    dontUnpack = true;
-
-    installPhase = ''
-      runHook preInstall
-
-      clang_version=`${clang-unwrapped}/bin/clang -v 2>&1 | grep "clang version " | grep -E -o "[0-9.-]+"`
-      mkdir -p $out/{bin,include/c++/v1,lib/{cmake,clang/$clang_version/{include,lib}},libexec,share}
-
-      for path in ${llvm} ${clang-unwrapped} ${lld} ${libc} ${libunwind} ${libcxxabi} ${libcxx} ${compiler-rt}; do
-        cp -as $path/* $out
-        chmod +w $out/{*,include/c++/v1,lib/{clang/$clang_version/include,cmake}}
-        rm -f $out/lib/libc++.so
-      done
-
-      ln -s $out/lib/* $out/lib/clang/$clang_version/lib
-      ln -sf $out/include/* $out/lib/clang/$clang_version/include
-
-      runHook postInstall
-    '';
-
-    passthru.isClang = true;
-    passthru.isROCm = true;
-  });
-
-  gccForLibs = stdenv.cc.cc;
-
-  extraPackages = [
-    llvm
-    lld
-    libc
-    libunwind
-    libcxxabi
-    compiler-rt
-  ];
-
-  nixSupport.cc-cflags = [
-    "-resource-dir=$out/resource-root"
-    "-fuse-ld=lld"
-    "-rtlib=compiler-rt"
-    "-unwindlib=libunwind"
-    "-Wno-unused-command-line-argument"
-  ];
-
-  extraBuildCommands = ''
-    clang_version=`${cc}/bin/clang -v 2>&1 | grep "clang version " | grep -E -o "[0-9.-]+"`
-    mkdir -p $out/resource-root
-    ln -s ${cc}/lib/clang/$clang_version/{include,lib} $out/resource-root
-
-    # Not sure why, but hardening seems to make things break
-    echo "" > $out/nix-support/add-hardening.sh
-
-    # GPU compilation uses builtin `lld`
-    substituteInPlace $out/bin/{clang,clang++} \
-      --replace-fail "-MM) dontLink=1 ;;" "-MM | --cuda-device-only) dontLink=1 ;;''\n--cuda-host-only | --cuda-compile-host-device) dontLink=0 ;;"
-  '';
-}

pkgs/development/rocm-modules/6/llvm/stage-3/flang.nix

@@ -1,32 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-  clang-unwrapped,
-  mlir,
-  graphviz,
-  python3Packages,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  targetName = "flang";
-  targetDir = targetName;
-
-  extraNativeBuildInputs = [
-    graphviz
-    python3Packages.sphinx-markdown-tables
-  ];
-
-  extraBuildInputs = [ mlir ];
-
-  extraCMakeFlags = [
-    "-DCLANG_DIR=${clang-unwrapped}/lib/cmake/clang"
-    "-DMLIR_TABLEGEN_EXE=${mlir}/bin/mlir-tblgen"
-    "-DCLANG_TABLEGEN_EXE=${clang-unwrapped}/bin/clang-tblgen"
-    "-DFLANG_INCLUDE_TESTS=OFF" # `The dependency target "Bye" of target ...`
-  ];
-
-  # `flang/lib/Semantics/check-omp-structure.cpp:1905:1: error: no member named 'v' in 'Fortran::parser::OmpClause::OmpxDynCgroupMem'`
-  isBroken = true;
-}

pkgs/development/rocm-modules/6/llvm/stage-3/libclc.nix

@@ -1,38 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-  llvm,
-  clang,
-  spirv-llvm-translator,
-}:
-
-let
-  spirv = (spirv-llvm-translator.override { inherit llvm; });
-in
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  buildDocs = false; # No documentation to build
-  buildMan = false; # No man pages to build
-  targetName = "libclc";
-  targetDir = targetName;
-  extraBuildInputs = [ spirv ];
-
-  # `spirv-mesa3d` isn't compiling with LLVM 15.0.0, it does with LLVM 14.0.0
-  # Try removing the `spirv-mesa3d` and `clspv` patches next update
-  # `clspv` tests fail, unresolved calls
-  extraPostPatch = ''
-    substituteInPlace CMakeLists.txt \
-      --replace-fail "find_program( LLVM_CLANG clang PATHS \''${LLVM_BINDIR} NO_DEFAULT_PATH )" \
-        "find_program( LLVM_CLANG clang PATHS \"${clang}/bin\" NO_DEFAULT_PATH )" \
-      --replace-fail "find_program( LLVM_SPIRV llvm-spirv PATHS \''${LLVM_BINDIR} NO_DEFAULT_PATH )" \
-        "find_program( LLVM_SPIRV llvm-spirv PATHS \"${spirv}/bin\" NO_DEFAULT_PATH )" \
-      --replace-fail "  spirv-mesa3d-" "" \
-      --replace-fail "  spirv64-mesa3d-" "" \
-      --replace-fail "NOT \''${t} MATCHES" \
-        "NOT \''${ARCH} STREQUAL \"clspv\" AND NOT \''${ARCH} STREQUAL \"clspv64\" AND NOT \''${t} MATCHES"
-  '';
-
-  checkTargets = [ ];
-  isBroken = true; # ROCm 5.7.0 doesn't have IR/AttributeMask.h yet...?
-}

pkgs/development/rocm-modules/6/llvm/stage-3/lldb.nix

@@ -1,40 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-  clang,
-  xz,
-  swig,
-  lua5_3,
-  graphviz,
-  gtest,
-  python3Packages,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  buildTests = false; # FIXME: Bad pathing for clang executable in tests, using relative path most likely
-  targetName = "lldb";
-  targetDir = targetName;
-  extraNativeBuildInputs = [ python3Packages.sphinx-automodapi ];
-
-  extraBuildInputs = [
-    xz
-    swig
-    lua5_3
-    graphviz
-    gtest
-  ];
-
-  extraCMakeFlags = [
-    "-DLLDB_EXTERNAL_CLANG_RESOURCE_DIR=${clang}/resource-root/lib/clang/$clang_version"
-    "-DLLDB_INCLUDE_TESTS=ON"
-    "-DLLDB_INCLUDE_UNITTESTS=ON"
-  ];
-
-  extraPostPatch = ''
-    export clang_version=`clang -v 2>&1 | grep "clang version " | grep -E -o "[0-9.-]+"`
-  '';
-
-  checkTargets = [ "check-${targetName}" ];
-}

pkgs/development/rocm-modules/6/llvm/stage-3/mlir.nix

@@ -1,61 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-  clr,
-  vulkan-headers,
-  vulkan-loader,
-  glslang,
-  shaderc,
-  fetchpatch,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  buildDocs = false; # No decent way to hack this to work
-  buildMan = false; # No man pages to build
-  targetName = "mlir";
-  targetDir = targetName;
-
-  # Fix `DebugTranslation.cpp:139:10: error: no matching function for call to 'get'`
-  extraPatches = [
-    (fetchpatch {
-      url = "https://github.com/ROCm/llvm-project/commit/f1d1e10ec7e1061bf0b90abbc1e298d9438a5e74.patch";
-      hash = "sha256-3c91A9InMKxm+JcnWxoUeOU68y5I6w1AAXx6T9UByqI=";
-    })
-  ];
-  extraNativeBuildInputs = [ clr ];
-
-  extraBuildInputs = [
-    vulkan-headers
-    vulkan-loader
-    glslang
-    shaderc
-  ];
-
-  extraCMakeFlags = [
-    "-DMLIR_INCLUDE_DOCS=ON"
-    "-DMLIR_INCLUDE_TESTS=ON"
-    "-DMLIR_ENABLE_ROCM_RUNNER=ON"
-    "-DMLIR_ENABLE_SPIRV_CPU_RUNNER=ON"
-    "-DMLIR_ENABLE_VULKAN_RUNNER=ON"
-    "-DROCM_TEST_CHIPSET=gfx000" # CPU runner
-  ];
-
-  extraPostPatch = ''
-    # `add_library cannot create target "llvm_gtest" because an imported target with the same name already exists`
-    substituteInPlace CMakeLists.txt \
-      --replace-fail "EXISTS \''${UNITTEST_DIR}/googletest/include/gtest/gtest.h" "FALSE"
-
-    # Mainly `No such file or directory`
-    cat ${./1001-mlir-failing-tests.list} | xargs -d \\n rm
-  '';
-
-  extraPostInstall = ''
-    mkdir -p $out/bin
-    mv bin/mlir-tblgen $out/bin
-  '';
-
-  checkTargets = [ "check-${targetName}" ];
-  requiredSystemFeatures = [ "big-parallel" ];
-}

pkgs/development/rocm-modules/6/llvm/stage-3/openmp.nix

@@ -1,55 +0,0 @@
-{
-  lib,
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-  llvm,
-  clang,
-  clang-unwrapped,
-  rocm-device-libs,
-  rocm-runtime,
-  rocm-thunk,
-  perl,
-  elfutils,
-  libdrm,
-  numactl,
-  lit,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  targetName = "openmp";
-  targetDir = targetName;
-  extraNativeBuildInputs = [ perl ];
-
-  extraBuildInputs = [
-    rocm-device-libs
-    rocm-runtime
-    rocm-thunk
-    elfutils
-    libdrm
-    numactl
-  ];
-
-  extraCMakeFlags = [
-    "-DCMAKE_MODULE_PATH=/build/source/llvm/cmake/modules" # For docs
-    "-DCLANG_TOOL=${clang}/bin/clang"
-    "-DCLANG_OFFLOAD_BUNDLER_TOOL=${clang-unwrapped}/bin/clang-offload-bundler"
-    "-DPACKAGER_TOOL=${clang-unwrapped}/bin/clang-offload-packager"
-    "-DOPENMP_LLVM_TOOLS_DIR=${llvm}/bin"
-    "-DOPENMP_LLVM_LIT_EXECUTABLE=${lit}/bin/.lit-wrapped"
-    "-DDEVICELIBS_ROOT=${rocm-device-libs.src}"
-  ];
-
-  extraPostPatch = ''
-    # We can't build this target at the moment
-    substituteInPlace libomptarget/DeviceRTL/CMakeLists.txt \
-      --replace "gfx1010" ""
-
-    # No idea what's going on here...
-    cat ${./1000-openmp-failing-tests.list} | xargs -d \\n rm
-  '';
-
-  checkTargets = [ "check-${targetName}" ];
-  extraLicenses = [ lib.licenses.mit ];
-}

pkgs/development/rocm-modules/6/llvm/stage-3/polly.nix

@@ -1,19 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  targetName = "polly";
-  targetDir = targetName;
-
-  extraPostPatch = ''
-    # `add_library cannot create target "llvm_gtest" because an imported target with the same name already exists`
-    substituteInPlace CMakeLists.txt \
-      --replace-fail "NOT TARGET gtest" "FALSE"
-  '';
-
-  checkTargets = [ "check-${targetName}" ];
-}

pkgs/development/rocm-modules/6/llvm/stage-3/pstl.nix

@@ -1,16 +0,0 @@
-{
-  stdenv,
-  callPackage,
-  rocmUpdateScript,
-}:
-
-callPackage ../base.nix rec {
-  inherit stdenv rocmUpdateScript;
-  buildDocs = false; # No documentation to build
-  buildMan = false; # No man pages to build
-  buildTests = false; # Too many errors
-  targetName = "pstl";
-  targetDir = "runtimes";
-  targetRuntimes = [ targetName ];
-  checkTargets = [ "check-${targetName}" ];
-}

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

@@ -7,21 +7,21 @@
   cmake,
   rocm-cmake,
   clr,
-  clang-tools-extra,
   openmp,
   rocblas,
+  hipblas-common,
+  hipblas,
+  hipblaslt,
   rocmlir,
-  composable_kernel,
   miopen,
   protobuf,
+  abseil-cpp,
   half,
   nlohmann_json,
   msgpack,
   sqlite,
   oneDNN_2,
   blaze,
-  cppcheck,
-  rocm-device-libs,
   texliveSmall,
   doxygen,
   sphinx,
@@ -54,7 +54,7 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
   pname = "migraphx";
-  version = "6.0.2";
+  version = "6.3.3";
 
   outputs =
     [
@@ -71,7 +71,7 @@ stdenv.mkDerivation (finalAttrs: {
     owner = "ROCm";
     repo = "AMDMIGraphX";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-VDYUSpWYAdJ63SKVCO26DVAC3RtZM7otqN0sYUA6DBQ=";
+    hash = "sha256-h9cTbrMwHeRGVJS/uHQnCXplNcrBqxbhwz2AcAEso0M=";
   };
 
   nativeBuildInputs =
@@ -80,7 +80,6 @@ stdenv.mkDerivation (finalAttrs: {
       cmake
       rocm-cmake
       clr
-      clang-tools-extra
       python3Packages.python
     ]
     ++ lib.optionals buildDocs [
@@ -96,8 +95,10 @@ stdenv.mkDerivation (finalAttrs: {
   buildInputs = [
     openmp
     rocblas
+    hipblas-common
+    hipblas
+    hipblaslt
     rocmlir
-    composable_kernel
     miopen
     protobuf
     half
@@ -106,16 +107,31 @@ stdenv.mkDerivation (finalAttrs: {
     sqlite
     oneDNN_2
     blaze
-    cppcheck
     python3Packages.pybind11
     python3Packages.onnx
   ];
 
+  LDFLAGS = "-Wl,--allow-shlib-undefined";
+
   cmakeFlags = [
     "-DMIGRAPHX_ENABLE_GPU=ON"
     "-DMIGRAPHX_ENABLE_CPU=ON"
     "-DMIGRAPHX_ENABLE_FPGA=ON"
     "-DMIGRAPHX_ENABLE_MLIR=OFF" # LLVM or rocMLIR mismatch?
+    "-DCMAKE_C_COMPILER=amdclang"
+    "-DCMAKE_CXX_COMPILER=amdclang++"
+    "-DCMAKE_VERBOSE_MAKEFILE=ON"
+    "-DEMBED_USE=CArrays" # Fixes error with lld
+    "-DDMIGRAPHX_ENABLE_PYTHON=ON"
+    "-DROCM_PATH=${clr}"
+    "-DHIP_ROOT_DIR=${clr}"
+    # migraphx relies on an incompatible fork of composable_kernel
+    # migraphxs relies on miopen which relies on current composable_kernel
+    # impossible to build with this ON; we can't link both of them even if we package both
+    "-DMIGRAPHX_USE_COMPOSABLEKERNEL=OFF"
+    "-DOpenMP_C_INCLUDE_DIR=${openmp.dev}/include"
+    "-DOpenMP_CXX_INCLUDE_DIR=${openmp.dev}/include"
+    "-DOpenMP_omp_LIBRARY=${openmp}/lib"
     # Manually define CMAKE_INSTALL_<DIR>
     # See: https://github.com/NixOS/nixpkgs/pull/197838
     "-DCMAKE_INSTALL_BINDIR=bin"
@@ -126,20 +142,12 @@ stdenv.mkDerivation (finalAttrs: {
 
   postPatch =
     ''
-      # We need to not use hipcc and define the CXXFLAGS manually due to `undefined hidden symbol: tensorflow:: ...`
-      export CXXFLAGS+="--rocm-path=${clr} --rocm-device-lib-path=${rocm-device-libs}/amdgcn/bitcode"
+      export CXXFLAGS+=" -w -isystem${rocmlir}/include/rocmlir -I${half}/include -I${abseil-cpp}/include -I${hipblas-common}/include"
       patchShebangs tools
 
       # `error: '__clang_hip_runtime_wrapper.h' file not found [clang-diagnostic-error]`
       substituteInPlace CMakeLists.txt \
         --replace "set(MIGRAPHX_TIDY_ERRORS ALL)" ""
-
-      # JIT library was removed from composable_kernel...
-      # https://github.com/ROCm/composable_kernel/issues/782
-      substituteInPlace src/targets/gpu/CMakeLists.txt \
-        --replace " COMPONENTS jit_library" "" \
-        --replace " composable_kernel::jit_library" "" \
-        --replace "if(WIN32)" "if(TRUE)"
     ''
     + lib.optionalString (!buildDocs) ''
       substituteInPlace CMakeLists.txt \
@@ -172,8 +180,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -182,6 +190,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken = true;
   };
 })

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

@@ -10,9 +10,13 @@
   rocm-cmake,
   rocblas,
   rocmlir,
+  rocrand,
+  rocm-runtime,
+  rocm-merged-llvm,
+  hipblas-common,
+  hipblas,
+  hipblaslt,
   clr,
-  clang-tools-extra,
-  clang-ocl,
   composable_kernel,
   frugally-deep,
   rocm-docs-core,
@@ -30,43 +34,53 @@
   rocm-comgr,
   roctracer,
   python3Packages,
+  # FIXME: should be able to use all clr targets
+  gpuTargets ? [
+    "gfx900"
+    "gfx906"
+    "gfx908"
+    "gfx90a"
+    "gfx942"
+    "gfx1030"
+    "gfx1100"
+    "gfx1101"
+    "gfx1102"
+  ], # clr.gpuTargets
   buildDocs ? false, # Needs internet because of rocm-docs-core
   buildTests ? false,
+  withComposableKernel ? composable_kernel.anyGfx9Target,
 }:
 
 let
-  version = "6.0.2";
+  # FIXME: cmake files need patched to include this properly
+  cFlags = "-O3 -DNDEBUG -Wno-documentation-pedantic --offload-compress -I${hipblas-common}/include  -I${hipblas}/include -I${roctracer}/include -I${nlohmann_json}/include -I${sqlite.dev}/include -I${rocrand}/include";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "MIOpen";
     rev = "rocm-${version}";
-    hash = "sha256-mbOdlSb0ESKi9hMkq3amv70Xkp/YKnZYre24d/y5TD0=";
+    hash = "sha256-rX+BE6wBDMnLyc6eai3bDVvmfahomDO0s10n6HhWu7c=";
     fetchLFS = true;
+    fetchSubmodules = true;
+    # WORKAROUND: .lfsconfig is incorrectly set to exclude everything upstream
     leaveDotGit = true;
-
-    # If you're reading this, it's gonna take a bit of time.
-    # fetchSubModules doesn't work with postFetch???
-    # fetchLFS isn't actually fetching the LFS files...
     postFetch = ''
       export HOME=$(mktemp -d)
       cd $out
-
-      # We need more history to fetch LFS files
+      set -x
       git remote add origin $url
-      git fetch origin
+      git fetch origin +refs/tags/rocm-${version}:refs/tags/rocm-${version}
       git clean -fdx
-      git checkout rocm-${version}
-
-      # We need to do this manually since using leaveDotGit and fetchSubmodules errors
-      git submodule update --init
-
-      # Fetch the LFS files
+      git switch -c rocm-${version} refs/tags/rocm-${version}
+      git config lfs.fetchexclude "none"
+      rm .lfsconfig
       git lfs install
-      git lfs fetch --all
+      git lfs track "*.kdb.bz2"
+      GIT_TRACE=1 git lfs fetch --include="src/kernels/**"
+      GIT_TRACE=1 git lfs pull --include="src/kernels/**"
       git lfs checkout
 
-      # Remove the defunct .git folder
       rm -rf .git
     '';
   };
@@ -112,8 +126,13 @@ stdenv.mkDerivation (finalAttrs: {
   inherit version src;
   pname = "miopen";
 
+  env.CFLAGS = cFlags;
+  env.CXXFLAGS = cFlags;
+
   # Find zstd and add to target. Mainly for torch.
   patches = [
+    ./skip-preexisting-dbs.patch
+    ./fix-isnan.patch # https://github.com/ROCm/MIOpen/pull/3448
     (fetchpatch {
       url = "https://github.com/ROCm/MIOpen/commit/e608b4325646afeabb5e52846997b926d2019d19.patch";
       hash = "sha256-oxa3qlIC2bzbwGxrQOZXoY/S7CpLsMrnWRB7Og0tk0M=";
@@ -122,11 +141,14 @@ stdenv.mkDerivation (finalAttrs: {
       url = "https://github.com/ROCm/MIOpen/commit/3413d2daaeb44b7d6eadcc03033a5954a118491e.patch";
       hash = "sha256-ST4snUcTmmSI1Ogx815KEX9GdMnmubsavDzXCGJkiKs=";
     })
-    (fetchpatch {
-      name = "Extend-MIOpen-ISA-compatibility.patch";
-      url = "https://github.com/GZGavinZhao/MIOpen/commit/416088b534618bd669a765afce59cfc7197064c1.patch";
-      hash = "sha256-OwONCA68y8s2GqtQj+OtotXwUXQ5jM8tpeM92iaD4MU=";
-    })
+    # FIXME: We need to rebase or drop this arch compat patch
+    # https://github.com/ROCm/MIOpen/issues/3540 suggests that
+    # arch compat patching doesn't work correctly for gfx1031
+    # (fetchpatch {
+    #   name = "Extend-MIOpen-ISA-compatibility.patch";
+    #   url = "https://github.com/GZGavinZhao/MIOpen/commit/416088b534618bd669a765afce59cfc7197064c1.patch";
+    #   hash = "sha256-OwONCA68y8s2GqtQj+OtotXwUXQ5jM8tpeM92iaD4MU=";
+    # })
   ];
 
   outputs =
@@ -139,21 +161,24 @@ stdenv.mkDerivation (finalAttrs: {
     ++ lib.optionals buildTests [
       "test"
     ];
+  enableParallelBuilding = true;
+  env.ROCM_PATH = clr;
+  env.LD_LIBRARY_PATH = lib.makeLibraryPath [ rocm-runtime ];
+  env.HIP_CLANG_PATH = "${rocm-merged-llvm}/bin";
 
   nativeBuildInputs = [
     pkg-config
     cmake
     rocm-cmake
     clr
-    clang-tools-extra
   ];
 
   buildInputs =
     [
+      hipblas
+      hipblas-common
       rocblas
       rocmlir
-      clang-ocl
-      composable_kernel
       half
       boost
       sqlite
@@ -161,6 +186,11 @@ stdenv.mkDerivation (finalAttrs: {
       nlohmann_json
       frugally-deep
       roctracer
+      rocrand
+      hipblaslt
+    ]
+    ++ lib.optionals withComposableKernel [
+      composable_kernel
     ]
     ++ lib.optionals buildDocs [
       latex
@@ -178,15 +208,32 @@ stdenv.mkDerivation (finalAttrs: {
 
   cmakeFlags =
     [
-      "-DCMAKE_CXX_FLAGS=-Wno-#warnings" # <half> -> <half/half.hpp>
-      "-DUNZIPPER=${bzip2}/bin/bunzip2"
+      "-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
+      "-DGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
+      "-DGPU_ARCHS=${lib.concatStringsSep ";" gpuTargets}"
+      "-DMIOPEN_USE_SQLITE_PERFDB=ON"
+      "-DCMAKE_VERBOSE_MAKEFILE=ON"
+      "-DCMAKE_MODULE_PATH=${clr}/hip/cmake"
+      "-DCMAKE_BUILD_TYPE=Release"
+
+      # needs to stream to stdout so bzcat rather than bunzip2
+      "-DUNZIPPER=${bzip2}/bin/bzcat"
+
+      "-DCMAKE_C_COMPILER=amdclang"
+      "-DCMAKE_CXX_COMPILER=amdclang++"
+      "-DROCM_PATH=${clr}"
+      "-DHIP_ROOT_DIR=${clr}"
+      (lib.cmakeBool "MIOPEN_USE_ROCBLAS" true)
+      (lib.cmakeBool "MIOPEN_USE_HIPBLASLT" true)
+      (lib.cmakeBool "MIOPEN_USE_COMPOSABLEKERNEL" withComposableKernel)
+      (lib.cmakeBool "MIOPEN_USE_HIPRTC" true)
+      (lib.cmakeBool "MIOPEN_USE_COMGR" true)
+      "-DCMAKE_HIP_COMPILER_ROCM_ROOT=${clr}"
       # Manually define CMAKE_INSTALL_<DIR>
       # See: https://github.com/NixOS/nixpkgs/pull/197838
       "-DCMAKE_INSTALL_BINDIR=bin"
       "-DCMAKE_INSTALL_LIBDIR=lib"
       "-DCMAKE_INSTALL_INCLUDEDIR=include"
-      "-DCMAKE_C_COMPILER=hipcc"
-      "-DCMAKE_CXX_COMPILER=hipcc"
       "-DMIOPEN_BACKEND=HIP"
     ]
     ++ lib.optionals buildTests [
@@ -195,24 +242,26 @@ stdenv.mkDerivation (finalAttrs: {
     ];
 
   postPatch = ''
+    substituteInPlace cmake/ClangTidy.cmake \
+      --replace-fail 'macro(enable_clang_tidy)' 'macro(enable_clang_tidy)
+      endmacro()
+      macro(enable_clang_tidy_unused)' \
+      --replace-fail 'function(clang_tidy_check TARGET)' 'function(clang_tidy_check TARGET)
+      return()'
+
     patchShebangs test src/composable_kernel fin utils install_deps.cmake
 
-    substituteInPlace CMakeLists.txt \
-      --replace "unpack_db(\"\''${CMAKE_SOURCE_DIR}/src/kernels/\''${FILE_NAME}.kdb.bz2\")" "" \
-      --replace "MIOPEN_HIP_COMPILER MATCHES \".*clang\\\\+\\\\+$\"" "true" \
-      --replace "set(MIOPEN_TIDY_ERRORS ALL)" "" # error: missing required key 'key'
-
-    substituteInPlace test/gtest/CMakeLists.txt \
-      --replace "include(googletest)" ""
-
-    substituteInPlace test/gtest/CMakeLists.txt \
-      --replace-fail " gtest_main " " ${gtest}/lib/libgtest.so ${gtest}/lib/libgtest_main.so "
-
     ln -sf ${gfx900} src/kernels/gfx900.kdb
     ln -sf ${gfx906} src/kernels/gfx906.kdb
     ln -sf ${gfx908} src/kernels/gfx908.kdb
     ln -sf ${gfx90a} src/kernels/gfx90a.kdb
     ln -sf ${gfx1030} src/kernels/gfx1030.kdb
+    mkdir -p build/share/miopen/db/
+    ln -sf ${gfx900} build/share/miopen/db/gfx900.kdb
+    ln -sf ${gfx906} build/share/miopen/db/gfx906.kdb
+    ln -sf ${gfx908} build/share/miopen/db/gfx908.kdb
+    ln -sf ${gfx90a} build/share/miopen/db/gfx90a.kdb
+    ln -sf ${gfx1030} build/share/miopen/db/gfx1030.kdb
   '';
 
   # Unfortunately, it seems like we have to call make on these manually
@@ -254,8 +303,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -264,8 +313,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

pkgs/development/rocm-modules/6/miopen/fix-isnan.patch

@@ -0,0 +1,31 @@
+From 17f67e0aa31cd2f1c1cb012d3858abf6956acc72 Mon Sep 17 00:00:00 2001
+From: "Sv. Lockal" <lockalsash@gmail.com>
+Date: Tue, 24 Dec 2024 14:43:10 +0000
+Subject: [PATCH] Fix missing isnan definition on libstdc++ >=14 systems
+
+Closes #3441
+---
+ driver/reducecalculation_driver.hpp | 3 ++-
+ 1 file changed, 2 insertions(+), 1 deletion(-)
+
+diff --git a/driver/reducecalculation_driver.hpp b/driver/reducecalculation_driver.hpp
+index 8226b3c953..2001969509 100644
+--- a/driver/reducecalculation_driver.hpp
++++ b/driver/reducecalculation_driver.hpp
+@@ -33,6 +33,7 @@
+ #include "random.hpp"
+ #include <algorithm>
+ #include <cfloat>
++#include <cmath>
+ #include <cstdlib>
+ #include <memory>
+ #include <miopen/miopen.h>
+@@ -77,7 +78,7 @@ int32_t mloReduceCalculationForwardRunHost(miopenTensorDescriptor_t inputDesc,
+         for(size_t i = 0; i < reduce_size; ++i)
+         {
+             Tcheck val = static_cast<Tcheck>(input[input_idx]);
+-            if(nanPropagation && isnan(val))
++            if(nanPropagation && std::isnan(val))
+             {
+                 val = 0.0f;
+             }

pkgs/development/rocm-modules/6/miopen/skip-preexisting-dbs.patch

@@ -0,0 +1,22 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index d0ffaf983..0b9ed0952 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -554,7 +554,7 @@ endif()
+ function(unpack_db db_bzip2_file)
+     get_filename_component(__fname ${db_bzip2_file} NAME_WLE)
+     add_custom_command(OUTPUT ${KERNELS_BINARY_DIR}/${__fname}
+-                       COMMAND ${UNZIPPER} -dc -k ${db_bzip2_file} > ${KERNELS_BINARY_DIR}/${__fname})
++                       COMMAND test -e ${KERNELS_BINARY_DIR}/${__fname} || ${UNZIPPER} -dc -k ${db_bzip2_file} > ${KERNELS_BINARY_DIR}/${__fname})
+     string(REPLACE "." "_" __tname ${__fname})
+     add_custom_target(generate_${__tname} ALL DEPENDS ${KERNELS_BINARY_DIR}/${__fname})
+ 
+@@ -563,7 +563,7 @@ function(unpack_db db_bzip2_file)
+     if(NOT MIOPEN_USE_SQLITE_PERFDB AND __extension STREQUAL ".db")
+         add_custom_command(OUTPUT ${KERNELS_BINARY_DIR}/${__fname}.txt
+                            DEPENDS sqlite2txt generate_${__tname}
+-                           COMMAND $<TARGET_FILE:sqlite2txt> ${KERNELS_BINARY_DIR}/${__fname} ${KERNELS_BINARY_DIR}/${__fname}.txt
++                           COMMAND test -e ${KERNELS_BINARY_DIR}/${__fname}.txt || $<TARGET_FILE:sqlite2txt> ${KERNELS_BINARY_DIR}/${__fname} ${KERNELS_BINARY_DIR}/${__fname}.txt
+         )
+         add_custom_target(generate_${__tname}_txt ALL DEPENDS ${KERNELS_BINARY_DIR}/${__fname}.txt)
+         add_dependencies(generate_kernels generate_${__tname}_txt)

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

@@ -12,7 +12,6 @@
   rocblas,
   miopen,
   migraphx,
-  clang,
   openmp,
   protobuf,
   qtcreator,
@@ -43,13 +42,13 @@ stdenv.mkDerivation (finalAttrs: {
         "cpu"
     );
 
-  version = "6.0.2";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "MIVisionX";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-d32lcJq24MXeIWbNbo6putWaol5kF2io6cz4ZuL+DbE=";
+    hash = "sha256-SisCbUDCAiWQ1Ue7qrtoT6vO/1ztzqji+3cJD6MXUNw=";
   };
 
   patches = [
@@ -98,6 +97,9 @@ stdenv.mkDerivation (finalAttrs: {
       "-DCMAKE_INSTALL_LIBDIR=lib"
       "-DCMAKE_INSTALL_INCLUDEDIR=include"
       "-DCMAKE_INSTALL_PREFIX_PYTHON=lib"
+      "-DOpenMP_C_INCLUDE_DIR=${openmp.dev}/include"
+      "-DOpenMP_CXX_INCLUDE_DIR=${openmp.dev}/include"
+      "-DOpenMP_omp_LIBRARY=${openmp}/lib"
       # "-DAMD_FP16_SUPPORT=ON" `error: typedef redefinition with different types ('__half' vs 'half_float::half')`
     ]
     ++ lib.optionals (gpuTargets != [ ]) [
@@ -115,37 +117,26 @@ stdenv.mkDerivation (finalAttrs: {
 
   postPatch = ''
     # We need to not use hipcc and define the CXXFLAGS manually due to `undefined hidden symbol: tensorflow:: ...`
-    export CXXFLAGS+="--rocm-path=${clr} --rocm-device-lib-path=${rocm-device-libs}/amdgcn/bitcode"
-    patchShebangs rocAL/rocAL_pybind/examples
-
-    # Properly find miopen
+    export CXXFLAGS+=" --rocm-path=${clr} --rocm-device-lib-path=${rocm-device-libs}/amdgcn/bitcode"
+    # Properly find miopen, fix ffmpeg version detection
     substituteInPlace amd_openvx_extensions/CMakeLists.txt \
-      --replace "miopen     PATHS \''${ROCM_PATH} QUIET" "miopen PATHS ${miopen} QUIET" \
-      --replace "\''${ROCM_PATH}/include/miopen/config.h" "${miopen}/include/miopen/config.h"
+      --replace-fail "miopen     PATHS \''${ROCM_PATH} QUIET" "miopen PATHS ${miopen} QUIET" \
+      --replace-fail "\''${ROCM_PATH}/include/miopen/config.h" "${miopen}/include/miopen/config.h"
 
     # Properly find turbojpeg
-    substituteInPlace amd_openvx/cmake/FindTurboJpeg.cmake \
-      --replace "\''${TURBO_JPEG_PATH}/include" "${libjpeg_turbo.dev}/include" \
-      --replace "\''${TURBO_JPEG_PATH}/lib" "${libjpeg_turbo.out}/lib"
-
-    # Fix bad paths
-    substituteInPlace rocAL/rocAL/rocAL_hip/CMakeLists.txt amd_openvx_extensions/amd_nn/nn_hip/CMakeLists.txt amd_openvx/openvx/hipvx/CMakeLists.txt \
-      --replace "COMPILER_FOR_HIP \''${ROCM_PATH}/llvm/bin/clang++" "COMPILER_FOR_HIP ${clang}/bin/clang++"
+    substituteInPlace cmake/FindTurboJpeg.cmake \
+      --replace-fail "\''${TURBO_JPEG_PATH}/include" "${libjpeg_turbo.dev}/include" \
+      --replace-fail "\''${TURBO_JPEG_PATH}/lib" "${libjpeg_turbo.out}/lib"
   '';
 
   postBuild = lib.optionalString buildDocs ''
     python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en ../docs _build/html
   '';
 
-  postInstall = lib.optionalString (!useOpenCL && !useCPU) ''
-    patchelf $out/lib/rocal_pybind*.so --shrink-rpath --allowed-rpath-prefixes "$NIX_STORE"
-    chmod +x $out/lib/rocal_pybind*.so
-  '';
-
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -154,8 +145,6 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
+    broken = useOpenCL;
   };
 })

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

@@ -0,0 +1,42 @@
+{
+  fetchFromGitHub,
+  stdenv,
+  cmake,
+  clr,
+  numactl,
+  nlohmann_json,
+}:
+stdenv.mkDerivation {
+  pname = "mscclpp";
+  version = "unstable-2024-12-13";
+  src = fetchFromGitHub {
+    owner = "microsoft";
+    repo = "mscclpp";
+    rev = "ee75caf365a27b9ab7521cfdda220b55429e5c37";
+    hash = "sha256-/mi9T9T6OIVtJWN3YoEe9az/86rz7BrX537lqaEh3ig=";
+  };
+  nativeBuildInputs = [
+    cmake
+  ];
+  buildInputs = [
+    clr
+    numactl
+  ];
+  postPatch = ''
+    substituteInPlace CMakeLists.txt \
+      --replace-fail "gfx90a gfx941 gfx942" "gfx908 gfx90a gfx942 gfx1030 gfx1100"
+  '';
+  cmakeFlags = [
+    "-DMSCCLPP_BYPASS_GPU_CHECK=ON"
+    "-DMSCCLPP_USE_ROCM=ON"
+    "-DMSCCLPP_BUILD_TESTS=OFF"
+    "-DGPU_TARGETS=gfx908;gfx90a;gfx942;gfx1030;gfx1100"
+    "-DAMDGPU_TARGETS=gfx908;gfx90a;gfx942;gfx1030;gfx1100"
+    "-DMSCCLPP_BUILD_APPS_NCCL=ON"
+    "-DMSCCLPP_BUILD_PYTHON_BINDINGS=OFF"
+    "-DFETCHCONTENT_QUIET=OFF"
+    "-DFETCHCONTENT_TRY_FIND_PACKAGE_MODE=ALWAYS"
+    "-DFETCHCONTENT_SOURCE_DIR_JSON=${nlohmann_json.src}"
+  ];
+  env.ROCM_PATH = clr;
+}

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

@@ -6,18 +6,35 @@
   cmake,
   rocm-cmake,
   rocm-smi,
+  rocm-core,
   clr,
+  mscclpp,
   perl,
   hipify,
   gtest,
   chrpath,
+  rocprofiler,
+  rocprofiler-register,
+  autoPatchelfHook,
   buildTests ? false,
-  gpuTargets ? [ ],
+  gpuTargets ? (clr.localGpuTargets or [ ]),
 }:
 
+let
+  useAsan = buildTests;
+  useUbsan = buildTests;
+  san = lib.optionalString (useAsan || useUbsan) (
+    "-fno-gpu-sanitize -fsanitize=undefined "
+    + (lib.optionalString useAsan "-fsanitize=address -shared-libsan ")
+  );
+in
+# Note: we can't properly test or make use of multi-node collective ops
+# https://github.com/NixOS/nixpkgs/issues/366242 tracks kernel support
+# kfd_peerdirect support which is on out-of-tree amdkfd in ROCm/ROCK-Kernel-Driver
+# infiniband ib_peer_mem support isn't in the mainline kernel but is carried by some distros
 stdenv.mkDerivation (finalAttrs: {
-  pname = "rccl";
-  version = "6.0.2";
+  pname = "rccl${clr.gpuArchSuffix}";
+  version = "6.3.3";
 
   outputs =
     [
@@ -27,11 +44,17 @@ stdenv.mkDerivation (finalAttrs: {
       "test"
     ];
 
+  patches = [
+    ./fix-mainline-support-and-ub.diff
+    ./enable-mscclpp-on-all-gfx9.diff
+    ./rccl-test-missing-iomanip.diff
+  ];
+
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "rccl";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-Oyml47yGEB7fALxBcDjqFngS38cnI39sDj94/JV7wE0=";
+    hash = "sha256-998tDiC0Qp9hhcXtFpiCWqwdKPVT2vNp0GU/rng03Bw=";
   };
 
   nativeBuildInputs = [
@@ -40,12 +63,16 @@ stdenv.mkDerivation (finalAttrs: {
     clr
     perl
     hipify
+    autoPatchelfHook # ASAN doesn't add rpath without this
   ];
 
   buildInputs =
     [
       rocm-smi
       gtest
+      rocprofiler
+      rocprofiler-register
+      mscclpp
     ]
     ++ lib.optionals buildTests [
       chrpath
@@ -53,8 +80,17 @@ stdenv.mkDerivation (finalAttrs: {
 
   cmakeFlags =
     [
-      "-DCMAKE_CXX_COMPILER=hipcc"
+      "-DHIP_CLANG_NUM_PARALLEL_JOBS=4"
+      "-DCMAKE_BUILD_TYPE=Release"
+      "-DROCM_PATH=${clr}"
+      "-DHIP_COMPILER=${clr}/bin/amdclang++"
+      "-DCMAKE_CXX_COMPILER=${clr}/bin/amdclang++"
+      "-DROCM_PATCH_VERSION=${rocm-core.ROCM_LIBPATCH_VERSION}"
+      "-DROCM_VERSION=${rocm-core.ROCM_LIBPATCH_VERSION}"
       "-DBUILD_BFD=OFF" # Can't get it to detect bfd.h
+      "-DENABLE_MSCCL_KERNEL=ON"
+      "-DENABLE_MSCCLPP=ON"
+      "-DMSCCLPP_ROOT=${mscclpp}"
       # Manually define CMAKE_INSTALL_<DIR>
       # See: https://github.com/NixOS/nixpkgs/pull/197838
       "-DCMAKE_INSTALL_BINDIR=bin"
@@ -62,32 +98,37 @@ stdenv.mkDerivation (finalAttrs: {
       "-DCMAKE_INSTALL_INCLUDEDIR=include"
     ]
     ++ lib.optionals (gpuTargets != [ ]) [
+      # AMD can't make up their minds and keep changing which one is used in different projects.
       "-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
+      "-DGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
     ]
     ++ lib.optionals buildTests [
       "-DBUILD_TESTS=ON"
     ];
 
+  # -O2 and -fno-strict-aliasing due to UB issues in RCCL :c
+  # Reported upstream
+  env.CFLAGS = "-I${clr}/include -O2 -fno-strict-aliasing ${san}-fno-omit-frame-pointer -momit-leaf-frame-pointer";
+  env.CXXFLAGS = "-I${clr}/include -O2 -fno-strict-aliasing ${san}-fno-omit-frame-pointer -momit-leaf-frame-pointer";
+  env.LDFLAGS = "${san}";
   postPatch = ''
     patchShebangs src tools
-
-    # Really strange behavior, `#!/usr/bin/env perl` should work...
-    substituteInPlace CMakeLists.txt \
-      --replace "\''$ \''${hipify-perl_executable}" "${perl}/bin/perl ${hipify}/bin/hipify-perl" \
-      --replace-warn "-parallel-jobs=12" "-parallel-jobs=1" \
-      --replace-warn "-parallel-jobs=16" "-parallel-jobs=1"
   '';
 
-  postInstall = lib.optionalString buildTests ''
-    mkdir -p $test/bin
-    mv $out/bin/* $test/bin
-    rmdir $out/bin
-  '';
+  postInstall =
+    lib.optionalString useAsan ''
+      patchelf --add-needed ${clr}/llvm/lib/linux/libclang_rt.asan-${stdenv.hostPlatform.parsed.cpu.name}.so $out/lib/librccl.so
+    ''
+    + lib.optionalString buildTests ''
+      mkdir -p $test/bin
+      mv $out/bin/* $test/bin
+      rmdir $out/bin
+    '';
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -99,8 +140,5 @@ stdenv.mkDerivation (finalAttrs: {
     ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

pkgs/development/rocm-modules/6/rccl/enable-mscclpp-on-all-gfx9.diff

@@ -0,0 +1,13 @@
+diff --git a/src/init.cc b/src/init.cc
+index 738f756..1b0e4fc 100644
+--- a/src/init.cc
++++ b/src/init.cc
+@@ -2049,7 +2049,7 @@ static ncclResult_t ncclCommInitRankFunc(struct ncclAsyncJob* job_) {
+     if (mscclEnabled() && (comm->topo->mscclEnabled || mscclForceEnabled()) && mscclppCommCompatible(comm)) {
+       hipDeviceProp_t devProp;
+       CUDACHECK(hipGetDeviceProperties(&devProp, cudaDev));
+-      comm->mscclppCompatible = IsArchMatch(devProp.gcnArchName, "gfx94");
++      comm->mscclppCompatible = IsArchMatch(devProp.gcnArchName, "gfx9");
+       if (comm->mscclppCompatible) {
+         bool mapContainsId = (mscclpp_uniqueIdMap.count(job->commId) > 0);
+         auto& mscclppUniqueId = mscclpp_uniqueIdMap[job->commId];

pkgs/development/rocm-modules/6/rccl/fix-mainline-support-and-ub.diff

@@ -0,0 +1,178 @@
+diff --git a/src/include/bootstrap.h b/src/include/bootstrap.h
+index 8c5f081..9922b79 100644
+--- a/src/include/bootstrap.h
++++ b/src/include/bootstrap.h
+@@ -10,11 +10,13 @@
+ #include "nccl.h"
+ #include "comm.h"
+ 
++// this is accessed through unaligned ptrs because ncclUniqueId is a typedef of char[128]
+ struct ncclBootstrapHandle {
+   uint64_t magic;
+   union ncclSocketAddress addr;
+ };
+ static_assert(sizeof(struct ncclBootstrapHandle) <= sizeof(ncclUniqueId), "Bootstrap handle is too large to fit inside NCCL unique ID");
++static_assert(alignof(struct ncclBootstrapHandle) == alignof(ncclUniqueId), "Bootstrap handle must have same alignment as NCCL unique ID to avoid UB");
+ 
+ ncclResult_t bootstrapNetInit();
+ ncclResult_t bootstrapCreateRoot(struct ncclBootstrapHandle* handle, bool idFromEnv);
+diff --git a/src/misc/rocmwrap.cc b/src/misc/rocmwrap.cc
+index b3063d5..464b80d 100644
+--- a/src/misc/rocmwrap.cc
++++ b/src/misc/rocmwrap.cc
+@@ -131,9 +131,12 @@ static void initOnceFunc() {
+       //format and store the kernel conf file location
+       snprintf(kernel_conf_file, sizeof(kernel_conf_file), "/boot/config-%s", utsname.release);
+       fp = fopen(kernel_conf_file, "r");
+-      if (fp == NULL) INFO(NCCL_INIT,"Could not open kernel conf file");
++      if (fp == NULL) {
++        INFO(NCCL_INIT,"Could not open kernel conf file, will assume CONFIG_DMABUF_MOVE_NOTIFY and CONFIG_PCI_P2PDMA are enabled");
++      }
+       //look for kernel_opt1 and kernel_opt2 in the conf file and check
+-      while (fgets(buf, sizeof(buf), fp) != NULL) {
++      // FIXME: This check is broken, CONFIG_DMABUF_MOVE_NOTIFY could be across a buf boundary.
++      while (fp && fgets(buf, sizeof(buf), fp) != NULL) {
+         if (strstr(buf, kernel_opt1) != NULL) {
+           found_opt1 = 1;
+           INFO(NCCL_INIT,"CONFIG_DMABUF_MOVE_NOTIFY=y in /boot/config-%s", utsname.release);
+@@ -143,11 +146,12 @@ static void initOnceFunc() {
+           INFO(NCCL_INIT,"CONFIG_PCI_P2PDMA=y in /boot/config-%s", utsname.release);
+         }
+       }
+-      if (!found_opt1 || !found_opt2) {
++      if (fp && (!found_opt1 || !found_opt2)) {
+         dmaBufSupport = 0;
+         INFO(NCCL_INIT, "CONFIG_DMABUF_MOVE_NOTIFY and CONFIG_PCI_P2PDMA should be set for DMA_BUF in /boot/config-%s", utsname.release);
+         INFO(NCCL_INIT, "DMA_BUF_SUPPORT Failed due to OS kernel support");
+       }
++      if (fp) fclose(fp);
+ 
+       if(dmaBufSupport) INFO(NCCL_INIT, "DMA_BUF Support Enabled");
+       else goto error;
+diff --git a/src/nccl.h.in b/src/nccl.h.in
+index 1d127b0..6296073 100644
+--- a/src/nccl.h.in
++++ b/src/nccl.h.in
+@@ -39,7 +39,7 @@ typedef struct ncclComm* ncclComm_t;
+ #define NCCL_UNIQUE_ID_BYTES 128
+ /*! @brief      Opaque unique id used to initialize communicators
+     @details    The ncclUniqueId must be passed to all participating ranks */
+-typedef struct { char internal[NCCL_UNIQUE_ID_BYTES]; /*!< Opaque array>*/} ncclUniqueId;
++typedef struct alignas(int64_t) { char internal[NCCL_UNIQUE_ID_BYTES]; /*!< Opaque array>*/} ncclUniqueId;
+ 
+ /*! @defgroup   rccl_result_code Result Codes
+     @details    The various result codes that RCCL API calls may return
+diff --git a/src/proxy.cc b/src/proxy.cc
+index 50e5437..51bb401 100644
+--- a/src/proxy.cc
++++ b/src/proxy.cc
+@@ -965,7 +965,11 @@ struct ncclProxyConnectionPool {
+ 
+ static ncclResult_t ncclProxyNewConnection(struct ncclProxyConnectionPool* pool, int* id) {
+   if (pool->offset == NCCL_PROXY_CONN_POOL_SIZE) {
+-    NCCLCHECK(ncclRealloc(&pool->pools, pool->banks, pool->banks+1));
++    if (pool->pools) {
++      NCCLCHECK(ncclRealloc(&pool->pools, pool->banks, pool->banks+1));
++    } else {
++      NCCLCHECK(ncclCalloc(&pool->pools, pool->banks+1));
++    }
+     NCCLCHECK(ncclCalloc(pool->pools+pool->banks, NCCL_PROXY_CONN_POOL_SIZE));
+     pool->banks++;
+     pool->offset = 0;
+diff --git a/src/transport/net_ib.cc b/src/transport/net_ib.cc
+index 6d77784..49762d3 100644
+--- a/src/transport/net_ib.cc
++++ b/src/transport/net_ib.cc
+@@ -573,7 +573,7 @@ ncclResult_t ncclIbGdrSupport() {
+     // Requires support from NIC driver modules
+     // Use ONLY for debugging!
+     moduleLoaded = 1;
+-    INFO(NCCL_INIT, "RCCL_FORCE_ENABLE_GDRDMA = 1, so explicitly setting moduleLoaded = 1");
++    INFO(NCCL_INIT, "ncclIbGdrSupport: RCCL_FORCE_ENABLE_GDRDMA = 1, so explicitly setting moduleLoaded = 1");
+   }
+ 
+   if (moduleLoaded == -1) {
+@@ -586,13 +586,14 @@ ncclResult_t ncclIbGdrSupport() {
+     // or created under a different path like `/sys/kernel/` or `/sys/` (depending on your ib_peer_mem module)
+     const char* memory_peers_paths[] = {"/sys/kernel/mm/memory_peers/amdkfd/version",
+                                   "/sys/kernel/memory_peers/amdkfd/version",
+-                                  "/sys/memory_peers/amdkfd/version"};
++                                  "/sys/memory_peers/amdkfd/version",
++                                  NULL};
+     int i = 0;
+ 
+     while (memory_peers_paths[i]) {
+       if (access(memory_peers_paths[i], F_OK) == 0) {
+         moduleLoaded = 1;
+-        INFO(NCCL_INIT,"Found %s", memory_peers_paths[i]);
++        INFO(NCCL_INIT,"ncclIbGdrSupport: Found %s", memory_peers_paths[i]);
+         break;
+       } else {
+         moduleLoaded = 0;
+@@ -612,22 +613,23 @@ ncclResult_t ncclIbGdrSupport() {
+     if (moduleLoaded == 0) {
+       // Check for `ib_register_peer_memory_client` symbol in `/proc/kallsyms`
+       // if your system uses native OS ib_peer module
+-      char buf[256];
+-      FILE *fp = NULL;
+-      fp = fopen("/proc/kallsyms", "r");
++      FILE *fp = fopen("/proc/kallsyms", "r");
++      char *line = NULL;
++      size_t len = 0;
+ 
+       if (fp == NULL) {
+-        INFO(NCCL_INIT,"Could not open /proc/kallsyms");
++        INFO(NCCL_INIT,"ncclIbGdrSupport: Could not open /proc/kallsyms to check for ib_register_peer_memory_client");
+       } else {
+-        while (fgets(buf, sizeof(buf), fp) != NULL) {
+-          if (strstr(buf, "t ib_register_peer_memory_client") != NULL ||
+-              strstr(buf, "T ib_register_peer_memory_client") != NULL) {
++        while (getline(&line, &len, fp) > 0) {
++          if (line && strstr(line, "ib_register_peer_memory_client") != NULL) {
+             moduleLoaded = 1;
+-            INFO(NCCL_INIT,"Found ib_register_peer_memory_client in /proc/kallsyms");
++            INFO(NCCL_INIT,"ncclIbGdrSupport: Found ib_register_peer_memory_client in /proc/kallsyms");
+             break;
+           }
+         }
+       }
++      if (line) free(line);
++      if (fp) fclose(fp);
+     }
+ #else
+     // Check for the nv_peer_mem module being loaded
+@@ -637,7 +639,7 @@ ncclResult_t ncclIbGdrSupport() {
+ #endif
+   }
+   if (moduleLoaded == 0) {
+-    INFO(NCCL_INIT,"GDRDMA not enabled. Could not find memory_peers directory or peer_memory symbol");
++    INFO(NCCL_INIT,"ncclIbGdrSupport: GDRDMA not enabled. Could not find memory_peers directory or peer_memory symbol");
+     return ncclSystemError;
+   }
+   return ncclSuccess;
+diff --git a/tools/ib-test/include/nccl.h b/tools/ib-test/include/nccl.h
+index 2c86c33..5801c61 100755
+--- a/tools/ib-test/include/nccl.h
++++ b/tools/ib-test/include/nccl.h
+@@ -31,7 +31,7 @@ extern "C" {
+ typedef struct ncclComm* ncclComm_t;
+ 
+ #define NCCL_UNIQUE_ID_BYTES 128
+-typedef struct { char internal[NCCL_UNIQUE_ID_BYTES]; } ncclUniqueId;
++typedef struct alignas(int64_t) { char internal[NCCL_UNIQUE_ID_BYTES]; } ncclUniqueId;
+ 
+ /* Error type */
+ typedef enum { ncclSuccess                 =  0,
+diff --git a/tools/topo_expl/include/nccl.h b/tools/topo_expl/include/nccl.h
+index 729561b..4e4bdd9 100644
+--- a/tools/topo_expl/include/nccl.h
++++ b/tools/topo_expl/include/nccl.h
+@@ -35,7 +35,7 @@ typedef struct ncclComm* ncclComm_t;
+ #define NCCL_COMM_NULL NULL
+ 
+ #define NCCL_UNIQUE_ID_BYTES 128
+-typedef struct { char internal[NCCL_UNIQUE_ID_BYTES]; } ncclUniqueId;
++typedef struct alignas(int64_t) { char internal[NCCL_UNIQUE_ID_BYTES]; } ncclUniqueId;
+ 
+ /*! @brief Error type */
+ typedef enum { ncclSuccess                 =  0,

pkgs/development/rocm-modules/6/rccl/rccl-test-missing-iomanip.diff

@@ -0,0 +1,10 @@
+--- a/test/common/TestBed.cpp
++++ b/test/common/TestBed.cpp
+@@ -4,6 +4,7 @@
+  * See LICENSE.txt for license information
+  ************************************************************************/
+ #include <unistd.h>
++#include <iomanip>
+ #include "TestBed.hpp"
+ #include <rccl/rccl.h>
+ 

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

@@ -4,9 +4,11 @@
   fetchFromGitHub,
   rocmUpdateScript,
   cmake,
+  amdsmi,
   rocm-smi,
   rocm-runtime,
   libcap,
+  libdrm,
   grpc,
   protobuf,
   openssl,
@@ -46,7 +48,7 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
   pname = "rdc";
-  version = "6.0.2";
+  version = "6.3.3";
 
   outputs =
     [
@@ -63,7 +65,7 @@ stdenv.mkDerivation (finalAttrs: {
     owner = "ROCm";
     repo = "rdc";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-QugcajxILmDeQiWG5uAUO41Wut45irg2Ynufgn1bmps=";
+    hash = "sha256-s/31b8/Kn5l1QJ941UMSB8SCzpvODsPfOLMmEBUYYmY=";
   };
 
   nativeBuildInputs =
@@ -79,9 +81,11 @@ stdenv.mkDerivation (finalAttrs: {
 
   buildInputs =
     [
+      amdsmi
       rocm-smi
       rocm-runtime
       libcap
+      libdrm
       grpc
       openssl
     ]
@@ -89,6 +93,8 @@ stdenv.mkDerivation (finalAttrs: {
       gtest
     ];
 
+  CXXFLAGS = "-I${libcap.dev}/include";
+
   cmakeFlags =
     [
       "-DCMAKE_VERBOSE_MAKEFILE=OFF"
@@ -126,8 +132,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -136,7 +142,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    # broken = versions.minor finalAttrs.version != versions.minor rocm-smi.version || versionAtLeast finalAttrs.version "7.0.0";
-    broken = true; # Too many errors, unsure how to fix
   };
 })

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

@@ -11,6 +11,7 @@
   rocrand,
   clr,
   git,
+  pkg-config,
   openmp,
   openmpi,
   gtest,
@@ -22,7 +23,7 @@
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "rocalution";
-  version = "6.0.2";
+  version = "6.3.3";
 
   outputs =
     [
@@ -42,7 +43,7 @@ stdenv.mkDerivation (finalAttrs: {
     owner = "ROCm";
     repo = "rocALUTION";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-mrN+CI2mqaMi8oKxui7HAIE2qSn50aNaFipkWwYMtbc=";
+    hash = "sha256-xdZ3HUiRGsreHfJH8RgL/s3jGyC5ABmBKcEfgtqWg8Y=";
   };
 
   nativeBuildInputs = [
@@ -50,6 +51,7 @@ stdenv.mkDerivation (finalAttrs: {
     rocm-cmake
     clr
     git
+    pkg-config
   ];
 
   buildInputs =
@@ -65,9 +67,12 @@ stdenv.mkDerivation (finalAttrs: {
       gtest
     ];
 
+  CXXFLAGS = "-I${openmp.dev}/include";
   cmakeFlags =
     [
-      "-DCMAKE_CXX_COMPILER=hipcc"
+      "-DOpenMP_C_INCLUDE_DIR=${openmp.dev}/include"
+      "-DOpenMP_CXX_INCLUDE_DIR=${openmp.dev}/include"
+      "-DOpenMP_omp_LIBRARY=${openmp}/lib"
       "-DROCM_PATH=${clr}"
       "-DHIP_ROOT_DIR=${clr}"
       "-DSUPPORT_HIP=ON"
@@ -82,6 +87,7 @@ stdenv.mkDerivation (finalAttrs: {
     ]
     ++ lib.optionals (gpuTargets != [ ]) [
       "-DAMDGPU_TARGETS=${lib.strings.concatStringsSep ";" gpuTargets}"
+      "-DGPU_TARGETS=${lib.strings.concatStringsSep ";" gpuTargets}"
     ]
     ++ lib.optionals buildTests [
       "-DBUILD_CLIENTS_TESTS=ON"
@@ -115,8 +121,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -125,8 +131,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -14,21 +14,24 @@
   gtest,
   gfortran,
   openmp,
+  git,
   amd-blis,
+  zstd,
+  hipblas-common,
+  hipblaslt,
   python3Packages,
+  rocm-smi,
   buildTensile ? true,
-  buildTests ? false,
-  buildBenchmarks ? false,
-  tensileLogic ? "asm_full",
-  tensileCOVersion ? "default",
+  buildTests ? true,
+  buildBenchmarks ? true,
   # https://github.com/ROCm/Tensile/issues/1757
   # Allows gfx101* users to use rocBLAS normally.
   # Turn the below two values to `true` after the fix has been cherry-picked
   # into a release. Just backporting that single fix is not enough because it
   # depends on some previous commits.
-  tensileSepArch ? false,
-  tensileLazyLib ? false,
-  tensileLibFormat ? "msgpack",
+  tensileSepArch ? true,
+  tensileLazyLib ? true,
+  withHipBlasLt ? true,
   # `gfx940`, `gfx941` are not present in this list because they are early
   # engineering samples, and all final MI300 hardware are `gfx942`:
   # https://github.com/NixOS/nixpkgs/pull/298388#issuecomment-2032791130
@@ -37,38 +40,47 @@
   # would force all `gfx101*` GPUs to run as `gfx1010`, so `gfx101*` GPUs will
   # always try to use `gfx1010` code objects, hence building for `gfx1012` is
   # useless: https://github.com/NixOS/nixpkgs/pull/298388#issuecomment-2076327152
-  gpuTargets ? [
-    "gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx942;gfx1010;gfx1030;gfx1100;gfx1101;gfx1102"
-  ],
+  gpuTargets ? (
+    clr.localGpuTargets or [
+      "gfx900"
+      "gfx906"
+      "gfx908"
+      "gfx90a"
+      "gfx942"
+      "gfx1010"
+      "gfx1030"
+      "gfx1100"
+      "gfx1101"
+      "gfx1102"
+    ]
+  ),
 }:
 
+let
+  gpuTargets' = lib.concatStringsSep ";" gpuTargets;
+in
 stdenv.mkDerivation (finalAttrs: {
-  pname = "rocblas";
-  version = "6.0.2";
+  pname = "rocblas${clr.gpuArchSuffix}";
+  version = "6.3.3";
 
-  outputs =
-    [
-      "out"
-    ]
-    ++ lib.optionals buildTests [
-      "test"
-    ]
-    ++ lib.optionals buildBenchmarks [
-      "benchmark"
-    ];
+  outputs = [
+    "out"
+  ];
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "rocBLAS";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-G68d/gvBbTdNx8xR3xY+OkBm5Yxq1NFjxby9BbpOcUk=";
+    hash = "sha256-IYcrVcGH4yZDkFZeNOJPfG0qsPS/WiH0fTSUSdo1BH4=";
   };
 
   nativeBuildInputs =
     [
       cmake
+      # no ninja, it buffers console output and nix times out long periods of no output
       rocm-cmake
       clr
+      git
     ]
     ++ lib.optionals buildTensile [
       tensile
@@ -77,12 +89,17 @@ stdenv.mkDerivation (finalAttrs: {
   buildInputs =
     [
       python3
+      hipblas-common
+    ]
+    ++ lib.optionals withHipBlasLt [
+      hipblaslt
     ]
     ++ lib.optionals buildTensile [
+      zstd
       msgpack
       libxml2
       python3Packages.msgpack
-      python3Packages.joblib
+      python3Packages.zstandard
     ]
     ++ lib.optionals buildTests [
       gtest
@@ -91,38 +108,61 @@ stdenv.mkDerivation (finalAttrs: {
       gfortran
       openmp
       amd-blis
+      rocm-smi
     ]
     ++ lib.optionals (buildTensile || buildTests || buildBenchmarks) [
       python3Packages.pyyaml
     ];
 
+  dontStrip = true;
+  env.CXXFLAGS =
+    "-O3 -DNDEBUG -I${hipblas-common}/include"
+    + lib.optionalString (buildTests || buildBenchmarks) " -I${amd-blis}/include/blis";
+  # Fails to link tests if we don't add amd-blis libs
+  env.LDFLAGS = lib.optionalString (
+    buildTests || buildBenchmarks
+  ) "-Wl,--as-needed -L${amd-blis}/lib -lblis-mt -lcblas";
+  env.TENSILE_ROCM_ASSEMBLER_PATH = "${stdenv.cc}/bin/clang++";
+
   cmakeFlags =
     [
-      (lib.cmakeFeature "CMAKE_C_COMPILER" "hipcc")
-      (lib.cmakeFeature "CMAKE_CXX_COMPILER" "hipcc")
+      (lib.cmakeFeature "CMAKE_BUILD_TYPE" "Release")
+      (lib.cmakeBool "CMAKE_VERBOSE_MAKEFILE" true)
+      (lib.cmakeFeature "CMAKE_EXECUTE_PROCESS_COMMAND_ECHO" "STDERR")
+      (lib.cmakeFeature "CMAKE_Fortran_COMPILER" "${lib.getBin gfortran}/bin/gfortran")
+      (lib.cmakeFeature "CMAKE_Fortran_COMPILER_AR" "${lib.getBin gfortran}/bin/ar")
+      (lib.cmakeFeature "CMAKE_Fortran_COMPILER_RANLIB" "${lib.getBin gfortran}/bin/ranlib")
       (lib.cmakeFeature "python" "python3")
-      (lib.cmakeFeature "AMDGPU_TARGETS" (lib.concatStringsSep ";" gpuTargets))
+      (lib.cmakeFeature "SUPPORTED_TARGETS" gpuTargets')
+      (lib.cmakeFeature "AMDGPU_TARGETS" gpuTargets')
+      (lib.cmakeFeature "GPU_TARGETS" gpuTargets')
       (lib.cmakeBool "BUILD_WITH_TENSILE" buildTensile)
       (lib.cmakeBool "ROCM_SYMLINK_LIBS" false)
       (lib.cmakeFeature "ROCBLAS_TENSILE_LIBRARY_DIR" "lib/rocblas")
+      (lib.cmakeBool "BUILD_WITH_HIPBLASLT" withHipBlasLt)
       (lib.cmakeBool "BUILD_CLIENTS_TESTS" buildTests)
       (lib.cmakeBool "BUILD_CLIENTS_BENCHMARKS" buildBenchmarks)
-      # rocblas header files are not installed unless we set this
-      (lib.cmakeFeature "CMAKE_INSTALL_INCLUDEDIR" "include")
+      (lib.cmakeBool "BUILD_CLIENTS_SAMPLES" buildBenchmarks)
+      (lib.cmakeBool "BUILD_OFFLOAD_COMPRESS" true)
+      # Temporarily set variables to work around upstream CMakeLists issue
+      # Can be removed once https://github.com/ROCm/rocm-cmake/issues/121 is fixed
+      "-DCMAKE_INSTALL_BINDIR=bin"
+      "-DCMAKE_INSTALL_INCLUDEDIR=include"
+      "-DCMAKE_INSTALL_LIBDIR=lib"
     ]
     ++ lib.optionals buildTensile [
+      "-DCPACK_SET_DESTDIR=OFF"
+      "-DLINK_BLIS=ON"
+      "-DTensile_CODE_OBJECT_VERSION=default"
+      "-DTensile_LOGIC=asm_full"
+      "-DTensile_LIBRARY_FORMAT=msgpack"
       (lib.cmakeBool "BUILD_WITH_PIP" false)
-      (lib.cmakeFeature "Tensile_LOGIC" tensileLogic)
-      (lib.cmakeFeature "Tensile_CODE_OBJECT_VERSION" tensileCOVersion)
       (lib.cmakeBool "Tensile_SEPARATE_ARCHITECTURES" tensileSepArch)
       (lib.cmakeBool "Tensile_LAZY_LIBRARY_LOADING" tensileLazyLib)
-      (lib.cmakeFeature "Tensile_LIBRARY_FORMAT" tensileLibFormat)
-      (lib.cmakeBool "Tensile_PRINT_DEBUG" true)
-    ]
-    ++ lib.optionals (buildTests || buildBenchmarks) [
-      (lib.cmakeFeature "CMAKE_CXX_FLAGS" "-I${amd-blis}/include/blis")
     ];
 
+  passthru.amdgpu_targets = gpuTargets';
+
   patches = [
     (fetchpatch {
       name = "Extend-rocBLAS-HIP-ISA-compatibility.patch";
@@ -135,14 +175,17 @@ stdenv.mkDerivation (finalAttrs: {
   postPatch = ''
     substituteInPlace cmake/build-options.cmake \
       --replace-fail 'Tensile_CPU_THREADS ""' 'Tensile_CPU_THREADS "$ENV{NIX_BUILD_CORES}"'
+    substituteInPlace CMakeLists.txt \
+      --replace-fail "4.42.0" "4.43.0"
   '';
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
+  enableParallelBuilding = true;
   requiredSystemFeatures = [ "big-parallel" ];
 
   meta = with lib; {
@@ -151,8 +194,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -1,66 +1,76 @@
-{ lib
-, stdenv
-, fetchFromGitHub
-, rocmUpdateScript
-, cmake
-, rocm-cmake
-, git
-, rocm-comgr
-, rocm-runtime
-, hwdata
-, texliveSmall
-, doxygen
-, graphviz
-, buildDocs ? true
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  rocmUpdateScript,
+  cmake,
+  rocm-cmake,
+  git,
+  rocm-comgr,
+  rocm-runtime,
+  hwdata,
+  texliveSmall,
+  doxygen,
+  graphviz,
+  buildDocs ? true,
 }:
 
 let
-  latex = lib.optionalAttrs buildDocs (texliveSmall.withPackages (ps: with ps; [
-    changepage
-    latexmk
-    varwidth
-    multirow
-    hanging
-    adjustbox
-    collectbox
-    stackengine
-    enumitem
-    alphalph
-    wasysym
-    sectsty
-    tocloft
-    newunicodechar
-    etoc
-    helvetic
-    wasy
-    courier
-  ]));
-in stdenv.mkDerivation (finalAttrs: {
+  latex = lib.optionalAttrs buildDocs (
+    texliveSmall.withPackages (
+      ps: with ps; [
+        changepage
+        latexmk
+        varwidth
+        multirow
+        hanging
+        adjustbox
+        collectbox
+        stackengine
+        enumitem
+        alphalph
+        wasysym
+        sectsty
+        tocloft
+        newunicodechar
+        etoc
+        helvetic
+        wasy
+        courier
+      ]
+    )
+  );
+in
+stdenv.mkDerivation (finalAttrs: {
   pname = "rocdbgapi";
-  version = "6.0.2";
+  version = "6.3.3";
 
-  outputs = [
-    "out"
-  ] ++ lib.optionals buildDocs [
-    "doc"
-  ];
+  outputs =
+    [
+      "out"
+    ]
+    ++ lib.optionals buildDocs [
+      "doc"
+    ];
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "ROCdbgapi";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-+CxaTmxRt/RicqQddqIEHs8vvAPCMKXkWg7kbZvnUsQ=";
+    hash = "sha256-6itfBrWVspobU47aiJAOQoxT8chwrq9scRn0or3bXto=";
   };
 
-  nativeBuildInputs = [
-    cmake
-    rocm-cmake
-    git
-  ] ++ lib.optionals buildDocs [
-    latex
-    doxygen
-    graphviz
-  ];
+  nativeBuildInputs =
+    [
+      cmake
+      rocm-cmake
+      git
+    ]
+    ++ lib.optionals buildDocs [
+      latex
+      doxygen
+      graphviz
+    ];
 
   buildInputs = [
     rocm-comgr
@@ -83,21 +93,15 @@ in stdenv.mkDerivation (finalAttrs: {
     make -j$NIX_BUILD_CORES doc
   '';
 
-  postInstall = ''
-    substituteInPlace $out/lib/cmake/amd-dbgapi/amd-dbgapi-config.cmake \
-      --replace "/build/source/build/" ""
-
-    substituteInPlace $out/lib/cmake/amd-dbgapi/amd-dbgapi-targets.cmake \
-      --replace "/build/source/build" "$out"
-  '' + lib.optionalString buildDocs ''
+  postInstall = lib.optionalString buildDocs ''
     mv $out/share/html/amd-dbgapi $doc/share/doc/amd-dbgapi/html
     rmdir $out/share/html
   '';
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -106,6 +110,5 @@ in stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken = versions.minor finalAttrs.version != versions.minor stdenv.cc.version || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -1,5 +1,4 @@
 {
-  rocfft,
   lib,
   stdenv,
   fetchFromGitHub,
@@ -15,18 +14,18 @@
   gtest,
   openmp,
   rocrand,
-  gpuTargets ? [ ],
+  gpuTargets ? clr.localGpuTargets or clr.gpuTargets,
 }:
 
 stdenv.mkDerivation (finalAttrs: {
-  pname = "rocfft";
-  version = "6.0.2";
+  pname = "rocfft${clr.gpuArchSuffix}";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "rocFFT";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-6Gjsy14GeR08VqnNmFhu8EyYDnQ+VZRlg+u9MAAWfHc=";
+    hash = "sha256-RrxdwZ64uC7lQzyJI1eGHX2dmRnW8TfNThnuvuz5XWo=";
   };
 
   nativeBuildInputs = [
@@ -36,6 +35,8 @@ stdenv.mkDerivation (finalAttrs: {
     rocm-cmake
   ];
 
+  # FIXME: rocfft_aot_helper runs at the end of the build and has a risk of timing it out
+  # due to a long period with no terminal output
   buildInputs = [ sqlite ];
 
   cmakeFlags =
@@ -156,8 +157,8 @@ stdenv.mkDerivation (finalAttrs: {
 
     updateScript = rocmUpdateScript {
       name = finalAttrs.pname;
-      owner = finalAttrs.src.owner;
-      repo = finalAttrs.src.repo;
+      inherit (finalAttrs.src) owner;
+      inherit (finalAttrs.src) repo;
     };
   };
 
@@ -169,8 +170,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -9,11 +9,13 @@
   flex,
   glibc,
   zlib,
+  zstd,
   gmp,
   mpfr,
   ncurses,
   expat,
   rocdbgapi,
+  perl,
   python3,
   babeltrace,
   sourceHighlight,
@@ -21,13 +23,13 @@
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "rocgdb";
-  version = "6.0.2";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "ROCgdb";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-XeX/k8gfo9HgcUSIjs35C7IqCmFhvBOqQJSOoPF6HK4=";
+    hash = "sha256-Z+uk+ViLXgk5hXrIhVHRY0Kly7mktYms7M3o9Tmxv8s=";
   };
 
   nativeBuildInputs = [
@@ -35,10 +37,13 @@ stdenv.mkDerivation (finalAttrs: {
     texinfo # For makeinfo
     bison
     flex
+    perl # used in mkinstalldirs script during installPhase
+    python3
   ];
 
   buildInputs = [
     zlib
+    zstd
     gmp
     mpfr
     ncurses
@@ -56,14 +61,14 @@ stdenv.mkDerivation (finalAttrs: {
 
     "--with-iconv-path=${glibc.bin}"
     "--enable-tui"
-    "--with-babeltrace"
+    "--with-babeltrace=${babeltrace}"
     "--with-python=python3"
     "--with-system-zlib"
+    "--with-system-zstd"
     "--enable-64-bit-bfd"
     "--with-gmp=${gmp.dev}"
     "--with-mpfr=${mpfr.dev}"
-    "--with-expat"
-    "--with-libexpat-prefix=${expat.dev}"
+    "--with-expat=${expat}"
 
     # So the installed binary is called "rocgdb" instead on plain "gdb"
     "--program-prefix=roc"
@@ -74,25 +79,32 @@ stdenv.mkDerivation (finalAttrs: {
     "--disable-ld"
     "--disable-gas"
     "--disable-gdbserver"
-    "--disable-sim"
     "--disable-gdbtk"
     "--disable-gprofng"
     "--disable-shared"
   ];
 
+  postPatch = ''
+    for file in *; do
+      if [ -f "$file" ]; then
+        patchShebangs "$file"
+      fi
+    done
+  '';
+
   # The source directory for ROCgdb (based on upstream GDB) contains multiple project
   # of GNU’s toolchain (binutils and onther), we only need to install the GDB part.
   installPhase = ''
     make install-gdb
   '';
 
-  # `-Wno-format-nonliteral` doesn't work
-  env.NIX_CFLAGS_COMPILE = "-Wno-error=format-security";
+  env.CFLAGS = "-Wno-switch -Wno-format-nonliteral -I${zstd.dev}/include -I${zlib.dev}/include -I${expat.dev}/include -I${ncurses.dev}/include";
+  env.CXXFLAGS = finalAttrs.env.CFLAGS;
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -101,6 +113,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = licenses.gpl3Plus;
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken = versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -1,27 +1,31 @@
-{ lib
-, stdenv
-, fetchFromGitHub
-, rocmUpdateScript
-, cmake
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  rocmUpdateScript,
+  rocm-core,
+  cmake,
 }:
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "rocm-cmake";
-  version = "6.0.2";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "rocm-cmake";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-qSjWT0KOQ5oDV06tfnKN+H/JzdoOnR9KY0c+SjvDepM=";
+    hash = "sha256-U4vGkH2iUlNJkqiNmVuFianD4WR9yuGvZsYG58smg0k=";
   };
 
   nativeBuildInputs = [ cmake ];
 
+  buildInputs = [ rocm-core ];
+
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -30,6 +34,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = licenses.mit;
     maintainers = teams.rocm.members;
     platforms = platforms.unix;
-    broken = versions.minor finalAttrs.version != versions.minor stdenv.cc.version || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -1,11 +1,13 @@
 {
   lib,
   stdenv,
-  fetchFromGitHub,
-  rocmUpdateScript,
+  fetchpatch,
   cmake,
-  rocm-cmake,
+  python3,
+  rocm-merged-llvm,
   rocm-device-libs,
+  zlib,
+  zstd,
   libxml2,
 }:
 
@@ -20,34 +22,45 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
   pname = "rocm-comgr";
-  version = "6.0.2";
+  # In-tree with ROCm LLVM
+  inherit (rocm-merged-llvm) version;
+  src = rocm-merged-llvm.llvm-src;
 
-  src = fetchFromGitHub {
-    owner = "ROCm";
-    repo = "ROCm-CompilerSupport";
-    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-9HuNU/k+kPJMlzqOTM20gm6SAOWJe9tpAZXEj4erdmI=";
-  };
+  sourceRoot = "${finalAttrs.src.name}/amd/comgr";
 
-  sourceRoot = "${finalAttrs.src.name}/lib/comgr";
+  patches = [
+    # [Comgr] Extend ISA compatibility
+    (fetchpatch {
+      sha256 = "sha256-dgow0kwSWM1TnkqWOZDRQrh5nuF8p5jbYyOLCpQsH4k=";
+      url = "https://github.com/GZGavinZhao/rocm-llvm-project/commit/a439e4f37ce71de48d4a979594276e3be0e6278f.patch";
+      relative = "amd/comgr";
+    })
+    #[Comgr] Extend ISA compatibility for CCOB
+    (fetchpatch {
+      sha256 = "sha256-6Rwz12Lk4R2JK3olii3cr2Zd0ZLYe7VSpK1YRCOsJWY=";
+      url = "https://github.com/GZGavinZhao/rocm-llvm-project/commit/2d8c459a4d4c0567a7a275b4b54560d88e5c6919.patch";
+      relative = "amd/comgr";
+    })
+  ];
 
   nativeBuildInputs = [
     cmake
-    rocm-cmake
+    python3
   ];
 
   buildInputs = [
     rocm-device-libs
     libxml2
+    zlib
+    zstd
+    rocm-merged-llvm
   ];
 
-  cmakeFlags = [ "-DLLVM_TARGETS_TO_BUILD=AMDGPU;X86" ];
-
-  passthru.updateScript = rocmUpdateScript {
-    name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
-  };
+  cmakeFlags = [
+    "-DCMAKE_VERBOSE_MAKEFILE=ON"
+    "-DCMAKE_BUILD_TYPE=Release"
+    "-DLLVM_TARGETS_TO_BUILD=AMDGPU;${llvmNativeTarget}"
+  ];
 
   meta = with lib; {
     description = "APIs for compiling and inspecting AMDGPU code objects";
@@ -55,8 +68,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = licenses.ncsa;
     maintainers = with maintainers; [ lovesegfault ] ++ teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -4,28 +4,43 @@
   fetchFromGitHub,
   rocmUpdateScript,
   cmake,
+  writeText,
 }:
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "rocm-core";
-  version = "6.0.2";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "rocm-core";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-sgL1UMt3o01zA8v41dyCG1fAsK/PkTRsHQJOvlNatZ4=";
+    hash = "sha256-UDnPGvgwzwv49CzF+Kt0v95CsxS33BZeqNcKw1K6jRI=";
   };
 
   nativeBuildInputs = [ cmake ];
-  cmakeFlags = [ "-DROCM_VERSION=${finalAttrs.version}" ];
+  # FIXME: What's the correct way to set this?
+  env.ROCM_LIBPATCH_VERSION = "${lib.versions.major finalAttrs.version}0${lib.versions.minor finalAttrs.version}0${lib.versions.patch finalAttrs.version}";
+  env.BUILD_ID = "nixos-${finalAttrs.env.ROCM_LIBPATCH_VERSION}";
+  env.ROCM_BUILD_ID = "release-${finalAttrs.env.BUILD_ID}";
+  cmakeFlags = [
+    "-DROCM_LIBPATCH_VERSION=${finalAttrs.env.ROCM_LIBPATCH_VERSION}"
+    "-DROCM_VERSION=${finalAttrs.version}"
+    "-DBUILD_ID=${finalAttrs.env.BUILD_ID}"
+  ];
 
+  setupHook = writeText "setupHook.sh" ''
+    export ROCM_LIBPATCH_VERSION="${finalAttrs.env.ROCM_LIBPATCH_VERSION}"
+    export BUILD_ID="${finalAttrs.env.BUILD_ID}"
+    export ROCM_BUILD_ID="${finalAttrs.env.ROCM_BUILD_ID}"
+  '';
+
+  passthru.ROCM_LIBPATCH_VERSION = finalAttrs.env.ROCM_LIBPATCH_VERSION;
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
-    page = "tags?per_page=1";
-    filter = ".[0].name | split(\"-\") | .[1]";
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
+    page = "tags?per_page=4";
   };
 
   meta = with lib; {
@@ -34,8 +49,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

pkgs/development/rocm-modules/6/rocm-device-libs/cmake.patch

@@ -1,7 +1,7 @@
 diff --git a/cmake/Packages.cmake b/cmake/Packages.cmake
 index 07c60eb..c736b3e 100644
---- a/cmake/Packages.cmake
-+++ b/cmake/Packages.cmake
+--- a/amd/device-libs/cmake/Packages.cmake
++++ b/amd/device-libs/cmake/Packages.cmake
 @@ -12,24 +12,29 @@ set_target_properties(${target} PROPERTIES
    IMPORTED_LOCATION \"${target_path}\")")
  endforeach()

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

@@ -1,11 +1,14 @@
 {
   lib,
   stdenv,
-  fetchFromGitHub,
-  rocmUpdateScript,
   cmake,
-  rocm-cmake,
+  ninja,
   libxml2,
+  zlib,
+  zstd,
+  ncurses,
+  rocm-merged-llvm,
+  python3,
 }:
 
 let
@@ -17,32 +20,36 @@ let
     else
       throw "Unsupported ROCm LLVM platform";
 in
-stdenv.mkDerivation (finalAttrs: {
+stdenv.mkDerivation {
   pname = "rocm-device-libs";
-  version = "6.0.2";
+  # In-tree with ROCm LLVM
+  inherit (rocm-merged-llvm) version;
+  src = rocm-merged-llvm.llvm-src;
 
-  src = fetchFromGitHub {
-    owner = "ROCm";
-    repo = "ROCm-Device-Libs";
-    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-7XG7oSkJ3EPWTYGea0I50eB1/DPMD5agmjctxZYTbLQ=";
-  };
+  postPatch = ''
+    cd amd/device-libs
+  '';
 
   patches = [ ./cmake.patch ];
 
   nativeBuildInputs = [
     cmake
-    rocm-cmake
+    ninja
+    python3
   ];
 
-  buildInputs = [ libxml2 ];
-  cmakeFlags = [ "-DLLVM_TARGETS_TO_BUILD=AMDGPU;${llvmNativeTarget}" ];
+  buildInputs = [
+    libxml2
+    zlib
+    zstd
+    ncurses
+    rocm-merged-llvm
+  ];
 
-  passthru.updateScript = rocmUpdateScript {
-    name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
-  };
+  cmakeFlags = [
+    "-DCMAKE_RELEASE_TYPE=Release"
+    "-DLLVM_TARGETS_TO_BUILD=AMDGPU;${llvmNativeTarget}"
+  ];
 
   meta = with lib; {
     description = "Set of AMD-specific device-side language runtime libraries";
@@ -50,8 +57,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = licenses.ncsa;
     maintainers = with maintainers; [ lovesegfault ] ++ teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
-})
+}

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

@@ -1,36 +1,37 @@
-{ lib
-, stdenv
-, fetchFromGitHub
-, gitUpdater
-, buildPythonPackage
-, setuptools
-, beautifulsoup4
-, gitpython
-, pydata-sphinx-theme
-, pygithub
-, sphinx
-, breathe
-, myst-parser
-, sphinx-book-theme
-, sphinx-copybutton
-, sphinx-design
-, sphinx-external-toc
-, sphinx-notfound-page
-, pyyaml
-, fastjsonschema
+{
+  lib,
+  fetchFromGitHub,
+  gitUpdater,
+  buildPythonPackage,
+  setuptools,
+  beautifulsoup4,
+  gitpython,
+  pydata-sphinx-theme,
+  pygithub,
+  sphinx,
+  breathe,
+  myst-nb,
+  myst-parser,
+  sphinx-book-theme,
+  sphinx-copybutton,
+  sphinx-design,
+  sphinx-external-toc,
+  sphinx-notfound-page,
+  pyyaml,
+  fastjsonschema,
 }:
 
 # FIXME: Move to rocmPackages_common
 buildPythonPackage rec {
   pname = "rocm-docs-core";
-  version = "1.12.0";
+  version = "1.17.0";
   format = "pyproject";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "rocm-docs-core";
     rev = "v${version}";
-    hash = "sha256-++Vi0jZLtHWsGy5IUohgF3P+Q6Jg/d0xWyDA6urbHUA=";
+    hash = "sha256-fGRJyQq0Eook1Dc9Qy+dehQ5BVNX+6pkkFN9adb21Eo=";
   };
 
   buildInputs = [ setuptools ];
@@ -42,6 +43,7 @@ buildPythonPackage rec {
     pygithub
     sphinx
     breathe
+    myst-nb
     myst-parser
     sphinx-book-theme
     sphinx-copybutton
@@ -59,7 +61,10 @@ buildPythonPackage rec {
   meta = with lib; {
     description = "ROCm Documentation Python package for ReadTheDocs build standardization";
     homepage = "https://github.com/ROCm/rocm-docs-core";
-    license = with licenses; [ mit cc-by-40 ];
+    license = with licenses; [
+      mit
+      cc-by-40
+    ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
   };

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

@@ -0,0 +1,27 @@
+{
+  symlinkJoin,
+  linkFarm,
+  clr,
+  hipblas,
+  hipblas-common,
+  rocblas,
+  rocsolver,
+  rocsparse,
+  rocm-device-libs,
+  rocm-smi,
+  llvm,
+}:
+symlinkJoin {
+  name = "rocm-path-${clr.version}";
+  paths = [
+    clr
+    hipblas-common
+    hipblas
+    rocblas
+    rocsolver
+    rocsparse
+    rocm-device-libs
+    rocm-smi
+    (linkFarm "rocm-llvm-subdir" { llvm = llvm.clang; })
+  ];
+}

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

@@ -6,37 +6,40 @@
   rocmUpdateScript,
   pkg-config,
   cmake,
+  ninja,
   xxd,
   rocm-device-libs,
-  rocm-thunk,
   elfutils,
   libdrm,
   numactl,
   valgrind,
   libxml2,
+  rocm-merged-llvm,
 }:
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "rocm-runtime";
-  version = "6.0.2";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "ROCR-Runtime";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-xNMG954HI9SOfvYYB/62fhmm9mmR4I10uHP2nqn9EgI=";
+    hash = "sha256-du20+5VNYgwchGO7W7FIVebBqLPtfSBnmPVbPpgEZjo=";
   };
 
-  sourceRoot = "${finalAttrs.src.name}/src";
+  env.CFLAGS = "-I${numactl.dev}/include -I${elfutils.dev}/include -w";
+  env.CXXFLAGS = "-I${numactl.dev}/include -I${elfutils.dev}/include -w";
 
   nativeBuildInputs = [
     pkg-config
     cmake
+    ninja
     xxd
+    rocm-merged-llvm
   ];
 
   buildInputs = [
-    rocm-thunk
     elfutils
     libdrm
     numactl
@@ -44,34 +47,56 @@ stdenv.mkDerivation (finalAttrs: {
     libxml2
   ];
 
+  cmakeFlags = [
+    "-DBUILD_SHARED_LIBS=ON"
+    "-DCMAKE_INSTALL_BINDIR=bin"
+    "-DCMAKE_INSTALL_LIBDIR=lib"
+    "-DCMAKE_INSTALL_INCLUDEDIR=include"
+  ];
+
   patches = [
+    # Patches for UB at runtime https://github.com/ROCm/ROCR-Runtime/issues/272
     (fetchpatch {
-      name = "extend-isa-compatibility-check.patch";
-      url = "https://salsa.debian.org/rocm-team/rocr-runtime/-/raw/076026d43bbee7f816b81fea72f984213a9ff961/debian/patches/0004-extend-isa-compatibility-check.patch";
-      hash = "sha256-cC030zVGS4kNXwaztv5cwfXfVwOldpLGV9iYgEfPEnY=";
-      stripLen = 1;
+      # [PATCH] hsa-runtime: set underlying type of hsa_region_info_t and hsa_amd_region_info_t to int
+      url = "https://github.com/ROCm/ROCR-Runtime/commit/39a6a168fa07e289a10f6e20e6ead4e303e99ba0.patch";
+      hash = "sha256-CshJJDvII1nNyNmt+YjwMwfBHUTlrdsxkhwfgBwO+WE=";
     })
+    (fetchpatch {
+      # [PATCH] rocr: refactor of runtime.cpp based on Coverity
+      url = "https://github.com/ROCm/ROCR-Runtime/commit/441bd9fe6c7bdb5c4c31f71524ed642786bc923e.patch";
+      hash = "sha256-7bQXxGkipzgT2aXRxCuh3Sfmo/zc/IOmA0x1zB+fMb0=";
+    })
+    (fetchpatch {
+      # [PATCH] queues: fix UB due to 1 << 31
+      url = "https://github.com/ROCm/ROCR-Runtime/commit/9b8a0f5dbee1903fa990a7d8accc1c5fbc549636.patch";
+      hash = "sha256-KlZWjfngH8yKly08iwC+Bzpvp/4dkaTpRIKdFYwRI+U=";
+    })
+    (fetchpatch {
+      # [PATCH] topology: fix UB due to 1 << 31
+      url = "https://github.com/ROCm/ROCR-Runtime/commit/d1d00bfee386d263e13c2b64fb6ffd1156deda7c.patch";
+      hash = "sha256-u70WEZaphQ7qTfgQPFATwdKWtHytu7CFH7Pzv1rOM8w=";
+    })
+    (fetchpatch {
+      # [PATCH] kfd_ioctl: fix UB due to 1 << 31
+      url = "https://github.com/ROCm/ROCR-Runtime/commit/41bfc66aef437a5b349f71105fa4b907cc7e17d5.patch";
+      hash = "sha256-A7VhPR3eSsmjq2cTBSjBIz9i//WiNjoXm0EsRKtF+ns=";
+    })
+    ./remove-hsa-aqlprofile-dep.patch
   ];
 
   postPatch = ''
-    patchShebangs image/blit_src/create_hsaco_ascii_file.sh
-    patchShebangs core/runtime/trap_handler/create_trap_handler_header.sh
-    patchShebangs core/runtime/blit_shaders/create_blit_shader_header.sh
+    patchShebangs --host image core runtime
 
     substituteInPlace CMakeLists.txt \
       --replace 'hsa/include/hsa' 'include/hsa'
 
-    # We compile clang before rocm-device-libs, so patch it in afterwards
-    # Replace object version: https://github.com/ROCm/ROCR-Runtime/issues/166 (TODO: Remove on LLVM update?)
-    substituteInPlace image/blit_src/CMakeLists.txt \
-      --replace '-cl-denorms-are-zero' '-cl-denorms-are-zero --rocm-device-lib-path=${rocm-device-libs}/amdgcn/bitcode' \
-      --replace '-mcode-object-version=4' '-mcode-object-version=5'
+    export HIP_DEVICE_LIB_PATH="${rocm-device-libs}/amdgcn/bitcode"
   '';
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -80,8 +105,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ ncsa ];
     maintainers = with maintainers; [ lovesegfault ] ++ teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

pkgs/development/rocm-modules/6/rocm-runtime/remove-hsa-aqlprofile-dep.patch

@@ -0,0 +1,27 @@
+libhsa-amd-aqlprofile64 library is unfree
+Bug: https://github.com/ROCm/ROCm/issues/1781
+--- a/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
++++ b/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
+@@ -1333,11 +1333,6 @@ hsa_status_t GpuAgent::GetInfo(hsa_agent_info_t attribute, void* value) const {
+         setFlag(HSA_EXTENSION_AMD_PC_SAMPLING);
+       }
+ 
+-      if (os::LibHandle lib = os::LoadLib(kAqlProfileLib)) {
+-        os::CloseLib(lib);
+-        setFlag(HSA_EXTENSION_AMD_AQLPROFILE);
+-      }
+-
+       setFlag(HSA_EXTENSION_AMD_PROFILER);
+ 
+       break;
+--- a/runtime/hsa-runtime/core/runtime/hsa.cpp
++++ b/runtime/hsa-runtime/core/runtime/hsa.cpp
+@@ -490,7 +490,7 @@ hsa_status_t hsa_system_get_major_extension_table(uint16_t extension, uint16_t v
+     return HSA_STATUS_SUCCESS;
+   }
+ 
+-  if (extension == HSA_EXTENSION_AMD_AQLPROFILE) {
++  if (0) {
+     if (version_major != hsa_ven_amd_aqlprofile_VERSION_MAJOR) {
+       debug_print("aqlprofile API incompatible ver %d, current ver %d\n",
+         version_major, hsa_ven_amd_aqlprofile_VERSION_MAJOR);

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

@@ -1,20 +1,21 @@
-{ lib
-, stdenv
-, fetchFromGitHub
-, rocmUpdateScript
-, cmake
-, wrapPython
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  rocmUpdateScript,
+  cmake,
+  wrapPython,
 }:
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "rocm-smi";
-  version = "6.0.2";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "rocm_smi_lib";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-fS52hpTv1WEycwkGZLXjz383WJWzyk8RvJRshEQSG/A=";
+    hash = "sha256-j9pkyUt+p6IkhawIhiTymqDBydxXZunxmdyCyRN0RxE=";
   };
 
   patches = [ ./cmake.patch ];
@@ -34,13 +35,14 @@ stdenv.mkDerivation (finalAttrs: {
 
   postInstall = ''
     wrapPythonProgramsIn $out
+    mv $out/libexec/rocm_smi/.rsmiBindingsInit.py-wrapped $out/libexec/rocm_smi/rsmiBindingsInit.py
     mv $out/libexec/rocm_smi/.rsmiBindings.py-wrapped $out/libexec/rocm_smi/rsmiBindings.py
   '';
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -49,6 +51,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = with maintainers; [ lovesegfault ] ++ teams.rocm.members;
     platforms = [ "x86_64-linux" ];
-    broken = versions.minor finalAttrs.version != versions.minor stdenv.cc.version || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -0,0 +1,32 @@
+{
+  clr,
+  ollama,
+  python3Packages,
+  rocmPackages,
+  magma-hip,
+  emptyDirectory,
+  stdenv,
+}:
+# This package exists purely to have a bunch of passthru.tests attrs
+stdenv.mkDerivation {
+  name = "rocm-tests";
+  nativeBuildInputs = [
+    clr
+  ];
+  src = emptyDirectory;
+  postInstall = "mkdir -p $out";
+  passthru.tests = {
+    ollama = ollama.override {
+      inherit rocmPackages;
+      acceleration = "rocm";
+    };
+    torch = python3Packages.torch.override {
+      inherit rocmPackages;
+      rocmSupport = true;
+      cudaSupport = false;
+      magma-hip = magma-hip.override {
+        inherit rocmPackages;
+      };
+    };
+  };
+}

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

@@ -1,54 +0,0 @@
-{ lib
-, stdenv
-, fetchFromGitHub
-, rocmUpdateScript
-, pkg-config
-, cmake
-, libdrm
-, numactl
-}:
-
-stdenv.mkDerivation (finalAttrs: {
-  pname = "rocm-thunk";
-  version = "6.0.2";
-
-  src = fetchFromGitHub {
-    owner = "ROCm";
-    repo = "ROCT-Thunk-Interface";
-    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-F6Qi+A9DuSx2e4WSfp4cnniKr0CkCZcZqsKwQmmZHhk=";
-  };
-
-  nativeBuildInputs = [
-    pkg-config
-    cmake
-  ];
-
-  buildInputs = [
-    libdrm
-    numactl
-  ];
-
-  cmakeFlags = [
-    # Manually define CMAKE_INSTALL_<DIR>
-    # See: https://github.com/NixOS/nixpkgs/pull/197838
-    "-DCMAKE_INSTALL_BINDIR=bin"
-    "-DCMAKE_INSTALL_LIBDIR=lib"
-    "-DCMAKE_INSTALL_INCLUDEDIR=include"
-  ];
-
-  passthru.updateScript = rocmUpdateScript {
-    name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
-  };
-
-  meta = with lib; {
-    description = "Radeon open compute thunk interface";
-    homepage = "https://github.com/ROCm/ROCT-Thunk-Interface";
-    license = with licenses; [ bsd2 mit ];
-    maintainers = with maintainers; [ lovesegfault ] ++ teams.rocm.members;
-    platforms = platforms.linux;
-    broken = versions.minor finalAttrs.version != versions.minor stdenv.cc.version || versionAtLeast finalAttrs.version "7.0.0";
-  };
-})

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

@@ -9,24 +9,25 @@
   busybox,
   python3,
   gnugrep,
+  clr, # Only for localGpuTargets
   # rocminfo requires that the calling user have a password and be in
   # the video group. If we let rocm_agent_enumerator rely upon
   # rocminfo's output, then it, too, has those requirements. Instead,
   # we can specify the GPU targets for this system (e.g. "gfx803" for
   # Polaris) such that no system call is needed for downstream
   # compilers to determine the desired target.
-  defaultTargets ? [ ],
+  defaultTargets ? (clr.localGpuTargets or [ ]),
 }:
 
 stdenv.mkDerivation (finalAttrs: {
-  version = "6.0.2";
+  version = "6.3.3";
   pname = "rocminfo";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "rocminfo";
     rev = "rocm-${finalAttrs.version}";
-    sha256 = "sha256-k0QeCyQcarGbAh4ft8Y7JBK6l2nWxDUc20XoYmtrMMs=";
+    sha256 = "sha256-fQPtO5TNbCbaZZ7VtGkkqng5QZ+FcScdh1opWr5YkLU=";
   };
 
   nativeBuildInputs = [
@@ -49,8 +50,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -59,9 +60,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = licenses.ncsa;
     maintainers = with maintainers; [ lovesegfault ] ++ teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      stdenv.hostPlatform.isAarch64
-      || versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -2,12 +2,10 @@
   lib,
   stdenv,
   fetchFromGitHub,
-  fetchpatch,
   rocmUpdateScript,
   cmake,
   rocm-cmake,
   rocminfo,
-  ninja,
   clr,
   git,
   libxml2,
@@ -20,6 +18,13 @@
   buildTests ? false, # `argument of type 'NoneType' is not iterable`
 }:
 
+# FIXME: rocmlir has an entire separate LLVM build in a subdirectory this is silly
+# It seems to be forked from AMD's own LLVM
+# If possible reusing the rocmPackages.llvm build would be better
+# Would have to confirm it is compatible with ROCm's tagged LLVM.
+# Fairly likely it's not given AMD's track record with forking their own software in incompatible ways
+# in subdirs
+
 # Theoretically, we could have our MLIR have an output
 # with the source and built objects so that we can just
 # use it as the external LLVM repo for this
@@ -36,7 +41,7 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
   pname = "rocmlir${suffix}";
-  version = "6.0.2";
+  version = "6.3.3";
 
   outputs =
     [
@@ -50,13 +55,12 @@ stdenv.mkDerivation (finalAttrs: {
     owner = "ROCm";
     repo = "rocMLIR";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-AypY0vL8Ij1zLycwpG2EPWWl4utp4ejXpAK0Jj/UvrA=";
+    hash = "sha256-0SQ6uLDRfVfdCX+8a7D6pu6dYlFvX0HFzCDEvlKYfak=";
   };
 
   nativeBuildInputs = [
     cmake
     rocm-cmake
-    ninja
     clr
     python3Packages.python
     python3Packages.tomli
@@ -75,23 +79,17 @@ stdenv.mkDerivation (finalAttrs: {
   ];
 
   patches = [
-    (fetchpatch {
-      name = "fix-TosaToRock-missing-includes.patch";
-      url = "https://github.com/ROCm/rocMLIR/commit/80b8c94a5dd6ab832733116fe0339c1d6011ab57.patch";
-      hash = "sha256-przg1AQZTiVbVd/4wA+KlGXu/RISO5n11FBkmUFKRSA=";
-    })
-    (fetchpatch {
-      name = "fix-cmake-depedency-on-transforms.patch";
-      url = "https://github.com/ROCm/rocMLIR/commit/b85ca4855e0f0214c2fd695e493c884cf08a3472.patch";
-      hash = "sha256-m108PnwvDAN3xWko+gZMgvCNFl4LXTvC67JHXhFHeBc=";
-    })
+    ./initparamdata-sort-const.patch
   ];
 
   cmakeFlags =
     [
       "-DLLVM_TARGETS_TO_BUILD=AMDGPU;${llvmNativeTarget}"
-      "-DLLVM_ENABLE_ZSTD=ON"
-      "-DLLVM_ENABLE_ZLIB=ON"
+      "-DCMAKE_BUILD_TYPE=Release"
+      "-DLLVM_USE_LINKER=lld"
+      "-DLLVM_ENABLE_ZSTD=FORCE_ON"
+      "-DLLVM_ENABLE_ZLIB=FORCE_ON"
+      "-DLLVM_ENABLE_LIBCXX=ON"
       "-DLLVM_ENABLE_TERMINFO=ON"
       "-DROCM_PATH=${clr}"
       # Manually define CMAKE_INSTALL_<DIR>
@@ -99,9 +97,7 @@ stdenv.mkDerivation (finalAttrs: {
       "-DCMAKE_INSTALL_BINDIR=bin"
       "-DCMAKE_INSTALL_LIBDIR=lib"
       "-DCMAKE_INSTALL_INCLUDEDIR=include"
-    ]
-    ++ lib.optionals buildRockCompiler [
-      "-DBUILD_FAT_LIBROCKCOMPILER=ON"
+      (lib.cmakeBool "BUILD_FAT_LIBROCKCOMPILER" buildRockCompiler)
     ]
     ++ lib.optionals (!buildRockCompiler) [
       "-DROCM_TEST_CHIPSET=gfx000"
@@ -111,6 +107,10 @@ stdenv.mkDerivation (finalAttrs: {
     patchShebangs mlir
     patchShebangs external/llvm-project/mlir/lib/Dialect/GPU/AmdDeviceLibsIncGen.py
 
+    # Fixes mlir/lib/Analysis/BufferDependencyAnalysis.cpp:41:19: error: redefinition of 'read'
+    substituteInPlace mlir/lib/Analysis/BufferDependencyAnalysis.cpp \
+      --replace-fail "enum EffectType { read, write, unknown };" "enum class EffectType { read, write, unknown };"
+
     # remove when no longer required
     substituteInPlace mlir/test/{e2e/generateE2ETest.py,fusion/e2e/generate-fusion-tests.py} \
       --replace-fail "\"/opt/rocm/bin" "\"${rocminfo}/bin"
@@ -150,10 +150,9 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
-    page = "tags?per_page=2";
-    filter = ".[1].name | split(\"-\") | .[1]";
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
+    page = "tags?per_page=4";
   };
 
   meta = with lib; {
@@ -162,8 +161,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ asl20 ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

pkgs/development/rocm-modules/6/rocmlir/initparamdata-sort-const.patch

@@ -0,0 +1,13 @@
+diff --git a/mlir/include/mlir/Dialect/Rock/Tuning/GridwiseGemmParams.h b/mlir/include/mlir/Dialect/Rock/Tuning/GridwiseGemmParams.h
+index 3f5ee596819a..590d53788822 100644
+--- a/mlir/include/mlir/Dialect/Rock/Tuning/GridwiseGemmParams.h
++++ b/mlir/include/mlir/Dialect/Rock/Tuning/GridwiseGemmParams.h
+@@ -209,7 +209,7 @@ private:
+     size_t original_pos;
+     int64_t padding_amount;
+ 
+-    bool operator<(const InitParamData &rhs) {
++    bool operator<(const InitParamData &rhs) const {
+       if (this->padding_amount < rhs.padding_amount) {
+         return true;
+       } else if (this->padding_amount == rhs.padding_amount) {

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

@@ -1,87 +1,91 @@
-{ lib
-, fetchpatch
-, stdenv
-, fetchFromGitHub
-, rocmUpdateScript
-, cmake
-, rocm-cmake
-, clr
-, gtest
-, gbenchmark
-, buildTests ? false
-, buildBenchmarks ? false
-, gpuTargets ? [ ]
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  rocmUpdateScript,
+  cmake,
+  rocm-cmake,
+  clr,
+  gtest,
+  gbenchmark,
+  buildTests ? false,
+  buildBenchmarks ? false,
+  gpuTargets ? [ ],
 }:
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "rocprim";
-  version = "6.0.2";
+  version = "6.3.3";
 
-  outputs = [
-    "out"
-  ] ++ lib.optionals buildTests [
-    "test"
-  ] ++ lib.optionals buildBenchmarks [
-    "benchmark"
-  ];
+  outputs =
+    [
+      "out"
+    ]
+    ++ lib.optionals buildTests [
+      "test"
+    ]
+    ++ lib.optionals buildBenchmarks [
+      "benchmark"
+    ];
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "rocPRIM";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-nWvq26qRPZ6Au1rc5cR74TKArcdUFg7O9djFi8SvMeM=";
+    hash = "sha256-0aHxpBuYIYhI2UER45YhHHL5YcxA+XeXoihcUs2AmCo=";
   };
 
-  patches = [
-    (fetchpatch {
-      name = "arch-conversion-marco.patch";
-      url = "https://salsa.debian.org/rocm-team/rocprim/-/raw/70c8aaee3cf545d92685f4ed9bf8f41e3d4d570c/debian/patches/arch-conversion-macro.patch";
-      hash = "sha256-oXdmbCArOB5bKE8ozDFrSh4opbO+c4VI6PNhljeUSms=";
-    })
-  ];
-
   nativeBuildInputs = [
     cmake
     rocm-cmake
     clr
   ];
 
-  buildInputs = lib.optionals buildTests [
-    gtest
-  ] ++ lib.optionals buildBenchmarks [
-    gbenchmark
-  ];
+  buildInputs =
+    lib.optionals buildTests [
+      gtest
+    ]
+    ++ lib.optionals buildBenchmarks [
+      gbenchmark
+    ];
 
-  cmakeFlags = [
-    "-DCMAKE_CXX_COMPILER=hipcc"
-    # Manually define CMAKE_INSTALL_<DIR>
-    # See: https://github.com/NixOS/nixpkgs/pull/197838
-    "-DCMAKE_INSTALL_BINDIR=bin"
-    "-DCMAKE_INSTALL_LIBDIR=lib"
-    "-DCMAKE_INSTALL_INCLUDEDIR=include"
-  ] ++ lib.optionals (gpuTargets != [ ]) [
-    "-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
-  ] ++ lib.optionals buildTests [
-    "-DBUILD_TEST=ON"
-  ] ++ lib.optionals buildBenchmarks [
-    "-DBUILD_BENCHMARK=ON"
-  ];
+  cmakeFlags =
+    [
+      "-DCMAKE_BUILD_TYPE=Release"
+      # Manually define CMAKE_INSTALL_<DIR>
+      # See: https://github.com/NixOS/nixpkgs/pull/197838
+      "-DCMAKE_INSTALL_BINDIR=bin"
+      "-DCMAKE_INSTALL_LIBDIR=lib"
+      "-DCMAKE_INSTALL_INCLUDEDIR=include"
+    ]
+    ++ lib.optionals (gpuTargets != [ ]) [
+      "-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
+    ]
+    ++ lib.optionals buildTests [
+      "-DBUILD_TEST=ON"
+    ]
+    ++ lib.optionals buildBenchmarks [
+      "-DBUILD_BENCHMARK=ON"
+    ];
 
-  postInstall = lib.optionalString buildTests ''
-    mkdir -p $test/bin
-    mv $out/bin/test_* $test/bin
-    mv $out/bin/rocprim $test/bin
-  '' + lib.optionalString buildBenchmarks ''
-    mkdir -p $benchmark/bin
-    mv $out/bin/benchmark_* $benchmark/bin
-  '' + lib.optionalString (buildTests || buildBenchmarks) ''
-    rmdir $out/bin
-  '';
+  postInstall =
+    lib.optionalString buildTests ''
+      mkdir -p $test/bin
+      mv $out/bin/test_* $test/bin
+      mv $out/bin/rocprim $test/bin
+    ''
+    + lib.optionalString buildBenchmarks ''
+      mkdir -p $benchmark/bin
+      mv $out/bin/benchmark_* $benchmark/bin
+    ''
+    + lib.optionalString (buildTests || buildBenchmarks) ''
+      rmdir $out/bin
+    '';
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -90,6 +94,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken = versions.minor finalAttrs.version != versions.minor stdenv.cc.version || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -0,0 +1,79 @@
+{
+  lib,
+  stdenv,
+  rocm-runtime,
+  rocprofiler,
+  numactl,
+  libpciaccess,
+  libxml2,
+  elfutils,
+  fetchFromGitHub,
+  rocmUpdateScript,
+  cmake,
+  clang,
+  clr,
+  python3Packages,
+  gpuTargets ? clr.gpuTargets,
+}:
+
+stdenv.mkDerivation (finalAttrs: {
+  pname = "rocprofiler-register";
+  version = "6.3.3";
+
+  src = fetchFromGitHub {
+    owner = "ROCm";
+    repo = "rocprofiler-register";
+    rev = "rocm-${finalAttrs.version}";
+    hash = "sha256-UZsCiGnudsbL1v5lKBx7Vz3/HRnGn4f86Pd+qu3ryh0=";
+    fetchSubmodules = true;
+  };
+
+  nativeBuildInputs = [
+    cmake
+    clang
+    clr
+  ];
+
+  buildInputs = [
+    numactl
+    libpciaccess
+    libxml2
+    elfutils
+    rocm-runtime
+
+    rocprofiler.rocmtoolkit-merged
+
+    python3Packages.lxml
+    python3Packages.cppheaderparser
+    python3Packages.pyyaml
+    python3Packages.barectf
+    python3Packages.pandas
+  ];
+  cmakeFlags = [
+    "-DCMAKE_MODULE_PATH=${clr}/lib/cmake/hip"
+    "-DHIP_ROOT_DIR=${clr}"
+    "-DGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
+    "-DBUILD_TEST=OFF"
+    "-DROCPROFILER_BUILD_TESTS=0"
+    "-DROCPROFILER_BUILD_SAMPLES=0"
+    # Manually define CMAKE_INSTALL_<DIR>
+    # See: https://github.com/NixOS/nixpkgs/pull/197838
+    "-DCMAKE_INSTALL_BINDIR=bin"
+    "-DCMAKE_INSTALL_LIBDIR=lib"
+    "-DCMAKE_INSTALL_INCLUDEDIR=include"
+  ];
+
+  passthru.updateScript = rocmUpdateScript {
+    name = "rocprofiler-register";
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
+  };
+
+  meta = with lib; {
+    description = "Profiling with perf-counters and derived metrics";
+    homepage = "https://github.com/ROCm/rocprofiler";
+    license = with licenses; [ mit ]; # mitx11
+    maintainers = teams.rocm.members;
+    platforms = platforms.linux;
+  };
+})

pkgs/development/rocm-modules/6/rocprofiler/0001-fix-shell-scripts.patch

@@ -1,109 +0,0 @@
-diff --git a/bin/rocprofv2 b/bin/rocprofv2
-index d0445e7..6079af8 100755
---- a/bin/rocprofv2
-+++ b/bin/rocprofv2
-@@ -7,7 +7,8 @@ if [ -n "${ROCP_PRELOAD}" ]; then LD_PRELOAD="${ROCP_PRELOAD}"; fi
- 
- CURRENT_DIR="$( dirname -- "$0"; )";
- ROCPROFV2_DIR=$(dirname -- $(realpath ${BASH_SOURCE[0]}));
--ROCM_DIR=$( dirname -- "$ROCPROFV2_DIR"; )
-+ROCPROFILER_DIR=$( dirname -- "$ROCPROFV2_DIR"; )
-+ROCM_DIR=@rocmtoolkit_merged@
- PLUGIN_LIST=("ctf" "perfetto" "file" "att" "cli")
- RUN_FROM_BUILD=0
- if [[ $ROCPROFV2_DIR == *"/build"* ]]; then
-@@ -15,7 +16,7 @@ if [[ $ROCPROFV2_DIR == *"/build"* ]]; then
-   ROCM_DIR=$ROCPROFV2_DIR
- fi
- 
--export ROCPROFILER_METRICS_PATH=$ROCM_DIR/libexec/rocprofiler/counters/derived_counters.xml
-+export ROCPROFILER_METRICS_PATH=$ROCPROFILER_DIR/libexec/rocprofiler/counters/derived_counters.xml
- export LD_LIBRARY_PATH=$ROCM_DIR/lib:$LD_LIBRARY_PATH
- 
- # Define color code
-@@ -83,7 +84,7 @@ while [ 1 ]; do
-       exit 1
-     fi
-   elif [[ "$1" == "--list-counters" ]]; then
--    export LD_PRELOAD=$LD_PRELOAD:$ROCM_DIR/lib/rocprofiler/librocprofiler_tool.so
-+    export LD_PRELOAD=$LD_PRELOAD:$ROC_DIR/lib/rocprofiler/librocprofiler_tool.so
-     eval $ROCM_DIR/libexec/rocprofiler/ctrl
-     exit 1
-   elif [[ "$1" == "-i" || "$1" == "--input" ]]; then
-@@ -221,7 +222,7 @@ while [ 1 ]; do
-       if [ $RUN_FROM_BUILD == 1 ]; then
-         ATT_PATH=$ROCM_DIR/plugin/att/att/att.py
-       else
--        ATT_PATH=$ROCM_DIR/libexec/rocprofiler/att/att.py
-+        ATT_PATH=$ROCPROFILER_DIR/libexec/rocprofiler/att/att.py
-         export ROCPROFV2_ATT_LIB_PATH=$ROCM_DIR/lib/hsa-amd-aqlprofile/librocprofv2_att.so
-       fi
-       ATT_ARGV=$3
-@@ -294,13 +295,13 @@ if [ -n "$PMC_LINES" ] && [ ! -n "$ATT_ARGV" ]; then
-       export OUTPUT_PATH=$FINAL_PATH
-     fi
-     let COUNTER=COUNTER+1
--    LD_PRELOAD=$LD_PRELOAD:$ROCM_DIR/lib/rocprofiler/librocprofiler_tool.so $*
-+    LD_PRELOAD=$LD_PRELOAD:$ROCPROFILER_DIR/lib/rocprofiler/librocprofiler_tool.so $*
-     if [ -n "$OUTPUT_PATH" ]; then
-       echo -e "\nThe output path for the following counters: $OUTPUT_PATH"
-     fi
-   done
- else
--  LD_PRELOAD=$LD_PRELOAD:$ROCM_DIR/lib/rocprofiler/librocprofiler_tool.so $*
-+  LD_PRELOAD=$LD_PRELOAD:$ROCPROFILER_DIR/lib/rocprofiler/librocprofiler_tool.so $*
- fi
- 
- get_pmc_results_txt_path() {
-diff --git a/bin/rpl_run.sh b/bin/rpl_run.sh
-index 6b236ed..a9c233c 100755
---- a/bin/rpl_run.sh
-+++ b/bin/rpl_run.sh
-@@ -25,16 +25,17 @@
- ROCPROF_ARGS="$*"
- time_stamp=`date +%y%m%d_%H%M%S`
- BIN_DIR=$(dirname $(realpath ${BASH_SOURCE[0]}))
--ROOT_DIR=$(dirname $BIN_DIR)
-+ROCPROFILER_DIR=$(dirname $BIN_DIR)
-+ROOT_DIR=@rocmtoolkit_merged@
- RUN_DIR=`pwd`
- TMP_DIR="/tmp"
- DATA_DIR="rpl_data_${time_stamp}_$$"
- 
--RPL_PATH=$ROOT_DIR/lib
-+RPL_PATH=$ROCPROFILER_DIR/lib
- TLIB_PATH=$RPL_PATH/rocprofiler
- TTLIB_PATH=$ROOT_DIR/lib/roctracer
- ROCM_LIB_PATH=$ROOT_DIR/lib
--PROF_BIN_DIR=$ROOT_DIR/libexec/rocprofiler
-+PROF_BIN_DIR=$ROCPROFILER_DIR/libexec/rocprofiler
- 
- # check if rocprof is supportd on this gpu arch
- V1_SUPPORTED_GPU_ARCHS=("gfx80x","gfx90x","gfx10xx","gfx94x")
-@@ -80,7 +81,7 @@ unset ROCP_PROXY_QUEUE
- # Disable AQL-profile read API
- export AQLPROFILE_READ_API=0
- # ROC Profiler package path
--export ROCP_PACKAGE_DIR=$ROOT_DIR
-+export ROCP_PACKAGE_DIR=$ROCPROFILER_DIR
- # enabled SPM KFD mode
- export ROCP_SPM_KFD_MODE=1
- 
-@@ -400,7 +401,7 @@ unset_v1_envs() {
- 
- ################################################################################################
- # main
--echo "RPL: on '$time_stamp' from '$ROOT_DIR' in '$RUN_DIR'"
-+echo "RPL: on '$time_stamp' from '$ROCPROFILER_DIR' in '$RUN_DIR'"
- # Parsing arguments
- if [ -z "$1" ] ; then
-   usage
-@@ -633,7 +634,7 @@ elif [ "$input_type" = "txt" -o "$input_type" = "none" ] ; then
-   else
-     echo "<metric></metric>" > $RES_DIR/input.xml
-   fi
--  input_list=`/bin/ls $RES_DIR/input*.xml`
-+  input_list=`ls $RES_DIR/input*.xml`
-   export ROCPROFILER_SESS=$RES_DIR
- else
-   fatal "Bad input file type '$INPUT_FILE'"

pkgs/development/rocm-modules/6/rocprofiler/0002-include-stdint-in-version.patch

@@ -1,12 +0,0 @@
-diff --git a/src/tools/versioning/version.cpp b/src/tools/versioning/version.cpp
-index 11bdd00..339743c 100644
---- a/src/tools/versioning/version.cpp
-+++ b/src/tools/versioning/version.cpp
-@@ -1,6 +1,7 @@
- #include <rocm-core/rocm_version.h>
- #include <iostream>
- #include <sstream>
-+#include <stdint.h>
- 
- 
- int main() {

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

@@ -4,17 +4,14 @@
   fetchFromGitHub,
   rocmUpdateScript,
   symlinkJoin,
-  replaceVars,
   cmake,
   clang,
   clr,
   rocm-core,
-  rocm-thunk,
+  rocm-runtime,
   rocm-device-libs,
   roctracer,
   rocdbgapi,
-  rocm-smi,
-  hsa-amd-aqlprofile-bin,
   numactl,
   libpciaccess,
   libxml2,
@@ -22,6 +19,7 @@
   mpi,
   systemd,
   gtest,
+  git,
   python3Packages,
   gpuTargets ? clr.gpuTargets,
 }:
@@ -32,12 +30,10 @@ let
 
     paths = [
       rocm-core
-      rocm-thunk
+      rocm-runtime
       rocm-device-libs
       roctracer
       rocdbgapi
-      rocm-smi
-      hsa-amd-aqlprofile-bin
       clr
     ];
 
@@ -48,32 +44,27 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
   pname = "rocprofiler";
-  version = "6.0.2";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "rocprofiler";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-yzgw9g5cHAZpdbU44+1ScZyUcZ2I4GGfjbm9GSqCClk=";
+    hash = "sha256-x6DVt1logBE8aNnuwukQhsv/vRqkJALcfAF+6yEQuIk=";
+    fetchSubmodules = true;
   };
 
   patches = [
     # These just simply won't build
     ./0000-dont-install-tests-hsaco.patch
-
-    # Fix bad paths
-    (replaceVars ./0001-fix-shell-scripts.patch {
-      rocmtoolkit_merged = rocmtoolkit-merged;
-    })
-
-    # Fix for missing uint32_t not defined
-    ./0002-include-stdint-in-version.patch
+    ./optional-aql-in-cmake.patch
   ];
 
   nativeBuildInputs = [
     cmake
     clang
     clr
+    git
     python3Packages.lxml
     python3Packages.cppheaderparser
     python3Packages.pyyaml
@@ -93,12 +84,20 @@ stdenv.mkDerivation (finalAttrs: {
 
   propagatedBuildInputs = [ rocmtoolkit-merged ];
 
+  # HACK: allow building without aqlprofile, probably explodes at runtime if use profiling
+  env.LDFLAGS = "-z nodefs -Wl,-undefined,dynamic_lookup,--unresolved-symbols=ignore-all";
+  #HACK: rocprofiler's cmake doesn't add these deps properly
+  env.CXXFLAGS = "-I${libpciaccess}/include -I${numactl.dev}/include -I${rocmtoolkit-merged}/include -I${elfutils.dev}/include -w";
+
   cmakeFlags = [
     "-DCMAKE_MODULE_PATH=${clr}/lib/cmake/hip"
     "-DHIP_ROOT_DIR=${clr}"
     "-DGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
     # Manually define CMAKE_INSTALL_<DIR>
     # See: https://github.com/NixOS/nixpkgs/pull/197838
+    "-DBUILD_TEST=OFF"
+    "-DROCPROFILER_BUILD_TESTS=0"
+    "-DROCPROFILER_BUILD_SAMPLES=0"
     "-DCMAKE_INSTALL_BINDIR=bin"
     "-DCMAKE_INSTALL_LIBDIR=lib"
     "-DCMAKE_INSTALL_INCLUDEDIR=include"
@@ -107,6 +106,13 @@ stdenv.mkDerivation (finalAttrs: {
   postPatch = ''
     patchShebangs .
 
+    substituteInPlace cmake_modules/rocprofiler_utils.cmake \
+      --replace-fail 'function(ROCPROFILER_CHECKOUT_GIT_SUBMODULE)' 'function(ROCPROFILER_CHECKOUT_GIT_SUBMODULE)
+      return()'
+
+    substituteInPlace CMakeLists.txt \
+      --replace-fail 'set(ROCPROFILER_BUILD_TESTS ON)' ""
+
     substituteInPlace tests-v2/featuretests/profiler/CMakeLists.txt \
       --replace "--build-id=sha1" "--build-id=sha1 --rocm-path=${clr} --rocm-device-lib-path=${rocm-device-libs}/amdgcn/bitcode"
 
@@ -115,20 +121,16 @@ stdenv.mkDerivation (finalAttrs: {
   '';
 
   postInstall = ''
-    # Why do these not already have the executable bit set?
-    chmod +x $out/lib/rocprofiler/librocprof-tool.so
-    chmod +x $out/share/rocprofiler/tests-v1/test/ocl/SimpleConvolution
-
     # Why do these have the executable bit set?
-    chmod -x $out/libexec/rocprofiler/counters/basic_counters.xml
-    chmod -x $out/libexec/rocprofiler/counters/derived_counters.xml
+    chmod -x $out/libexec/rocprofiler/counters/*.xml
   '';
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
+  passthru.rocmtoolkit-merged = rocmtoolkit-merged;
 
   meta = with lib; {
     description = "Profiling with perf-counters and derived metrics";
@@ -136,8 +138,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ]; # mitx11
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor clr.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

pkgs/development/rocm-modules/6/rocprofiler/optional-aql-in-cmake.patch

@@ -0,0 +1,147 @@
+From https://raw.githubusercontent.com/AphidGit/rocm_compile/refs/heads/main/rocprofiler.patch
+diff --git a/cmake_modules/rocprofiler_env.cmake b/cmake_modules/rocprofiler_env.cmake
+index 7b7c472..0aba3ed 100644
+--- a/cmake_modules/rocprofiler_env.cmake
++++ b/cmake_modules/rocprofiler_env.cmake
+@@ -36,6 +36,7 @@ if(ROCPROFILER_DEBUG_TRACE)
+     target_compile_definitions(rocprofiler-build-flags INTERFACE DEBUG_TRACE=1)
+ endif()
+ 
++set(ROCPROFILER_LD_AQLPROFILE false)
+ # Enable direct loading of AQL-profile HSA extension
+ if(ROCPROFILER_LD_AQLPROFILE)
+     target_compile_definitions(rocprofiler-build-flags INTERFACE ROCP_LD_AQLPROFILE=1)
+@@ -80,9 +81,3 @@ if("${ROCM_ROOT_DIR}" STREQUAL "")
+     message(FATAL_ERROR "ROCM_ROOT_DIR is not found.")
+ endif()
+ 
+-find_library(
+-    HSA_AMD_AQLPROFILE_LIBRARY
+-    NAMES hsa-amd-aqlprofile64
+-    HINTS ${CMAKE_PREFIX_PATH}
+-    PATHS ${ROCM_ROOT_DIR}
+-    PATH_SUFFIXES lib REQUIRED)
+diff --git a/src/api/CMakeLists.txt b/src/api/CMakeLists.txt
+index 61782f0..16c83bf 100644
+--- a/src/api/CMakeLists.txt
++++ b/src/api/CMakeLists.txt
+@@ -51,15 +51,6 @@ find_file(
+     NO_DEFAULT_PATH REQUIRED)
+ get_filename_component(HSA_RUNTIME_INC_PATH ${HSA_H} DIRECTORY)
+ 
+-find_library(
+-    AQLPROFILE_LIB "libhsa-amd-aqlprofile64.so"
+-    HINTS ${CMAKE_PREFIX_PATH}
+-    PATHS ${ROCM_PATH}
+-    PATH_SUFFIXES lib)
+-
+-if(NOT AQLPROFILE_LIB)
+-    message(FATAL_ERROR "AQL_PROFILE not installed. Please install hsa-amd-aqlprofile!")
+-endif()
+ 
+ # ########################################################################################
+ # Adding Old Library Files
+@@ -247,7 +238,7 @@ target_include_directories(
+     PUBLIC $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include/rocprofiler>
+     PRIVATE ${LIB_DIR} ${ROOT_DIR} ${PROJECT_SOURCE_DIR}/include/rocprofiler)
+ target_link_libraries(
+-    ${ROCPROFILER_TARGET} PRIVATE ${AQLPROFILE_LIB} hsa-runtime64::hsa-runtime64 c stdc++
++    ${ROCPROFILER_TARGET} PRIVATE hsa-runtime64::hsa-runtime64 c stdc++
+                                   dl rocprofiler::build-flags rocprofiler::memcheck)
+ 
+ get_target_property(ROCPROFILER_LIBRARY_V1_NAME ${ROCPROFILER_TARGET} NAME)
+@@ -325,8 +316,7 @@ target_link_options(
+     -Wl,--no-undefined)
+ target_link_libraries(
+     rocprofiler-v2
+-    PRIVATE ${AQLPROFILE_LIB}
+-            hsa-runtime64::hsa-runtime64
++    PRIVATE hsa-runtime64::hsa-runtime64
+             Threads::Threads
+             atomic
+             numa
+diff --git a/src/util/hsa_rsrc_factory.cpp b/src/util/hsa_rsrc_factory.cpp
+index 2c47186..6b39634 100644
+--- a/src/util/hsa_rsrc_factory.cpp
++++ b/src/util/hsa_rsrc_factory.cpp
+@@ -155,17 +155,6 @@ HsaRsrcFactory::HsaRsrcFactory(bool initialize_hsa) : initialize_hsa_(initialize
+   if (kern_arg_pool_ == nullptr)
+     CHECK_STATUS("Kern-arg memory pool is not found", HSA_STATUS_ERROR);
+ 
+-  // Get AqlProfile API table
+-  aqlprofile_api_ = {};
+-#ifdef ROCP_LD_AQLPROFILE
+-  status = LoadAqlProfileLib(&aqlprofile_api_);
+-#else
+-  status = hsa_api_.hsa_system_get_major_extension_table(HSA_EXTENSION_AMD_AQLPROFILE,
+-                                                         hsa_ven_amd_aqlprofile_VERSION_MAJOR,
+-                                                         sizeof(aqlprofile_api_), &aqlprofile_api_);
+-#endif
+-  CHECK_STATUS("aqlprofile API table load failed", status);
+-
+   // Get Loader API table
+   loader_api_ = {};
+   status = hsa_api_.hsa_system_get_major_extension_table(HSA_EXTENSION_AMD_LOADER, 1,
+diff --git a/test/util/hsa_rsrc_factory.cpp b/test/util/hsa_rsrc_factory.cpp
+index 0a44d18..fab5b75 100644
+--- a/test/util/hsa_rsrc_factory.cpp
++++ b/test/util/hsa_rsrc_factory.cpp
+@@ -137,17 +137,6 @@ HsaRsrcFactory::HsaRsrcFactory(bool initialize_hsa) : initialize_hsa_(initialize
+   if (cpu_pool_ == NULL) CHECK_STATUS("CPU memory pool is not found", HSA_STATUS_ERROR);
+   if (kern_arg_pool_ == NULL) CHECK_STATUS("Kern-arg memory pool is not found", HSA_STATUS_ERROR);
+ 
+-  // Get AqlProfile API table
+-  aqlprofile_api_ = {0};
+-#ifdef ROCP_LD_AQLPROFILE
+-  status = LoadAqlProfileLib(&aqlprofile_api_);
+-#else
+-  status = hsa_api_.hsa_system_get_major_extension_table(HSA_EXTENSION_AMD_AQLPROFILE,
+-                                                         hsa_ven_amd_aqlprofile_VERSION_MAJOR,
+-                                                         sizeof(aqlprofile_api_), &aqlprofile_api_);
+-#endif
+-  CHECK_STATUS("aqlprofile API table load failed", status);
+-
+   // Get Loader API table
+   loader_api_ = {0};
+   status = hsa_api_.hsa_system_get_major_extension_table(HSA_EXTENSION_AMD_LOADER, 1,
+diff --git a/tests-v2/unittests/core/CMakeLists.txt b/tests-v2/unittests/core/CMakeLists.txt
+index 107cb51..0f6d4bf 100644
+--- a/tests-v2/unittests/core/CMakeLists.txt
++++ b/tests-v2/unittests/core/CMakeLists.txt
+@@ -235,8 +235,7 @@ set_target_properties(runCoreUnitTests PROPERTIES
+  INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_DATAROOTDIR}/${PROJECT_NAME}/tests")
+ target_link_libraries(
+     runCoreUnitTests
+-    PRIVATE ${AQLPROFILE_LIB}
+-            test_hsatool_library
++    PRIVATE test_hsatool_library
+             hsa-runtime64::hsa-runtime64
+             Threads::Threads
+             GTest::gtest GTest::gtest_main
+@@ -285,4 +284,4 @@ endif()
+ # for the *_FilePlugin tests
+ if(NOT EXISTS "${PROJECT_BINARY_DIR}/test-output")
+      file(MAKE_DIRECTORY "${PROJECT_BINARY_DIR}/test-output")
+-endif()
+\ No newline at end of file
++endif()
+diff --git a/tests-v2/unittests/profiler/CMakeLists.txt b/tests-v2/unittests/profiler/CMakeLists.txt
+index 53180d5..0c4d4a7 100644
+--- a/tests-v2/unittests/profiler/CMakeLists.txt
++++ b/tests-v2/unittests/profiler/CMakeLists.txt
+@@ -122,7 +122,7 @@ target_compile_definitions(
+      PRIVATE PROF_API_IMPL HIP_PROF_HIP_API_STRING=1 __HIP_PLATFORM_AMD__=1)
+ 
+ target_link_libraries(
+-     runUnitTests PRIVATE rocprofiler-v2 ${AQLPROFILE_LIB} hsa-runtime64::hsa-runtime64
++     runUnitTests PRIVATE rocprofiler-v2 hsa-runtime64::hsa-runtime64
+                          GTest::gtest GTest::gtest_main stdc++fs ${PCIACCESS_LIBRARIES} dw elf c dl)
+ 
+ add_dependencies(tests runUnitTests)
+@@ -158,4 +158,4 @@ endif()
+ # for the *_FilePlugin tests
+ if(NOT EXISTS "${PROJECT_BINARY_DIR}/test-output")
+     file(MAKE_DIRECTORY "${PROJECT_BINARY_DIR}/test-output")
+-endif()
+\ No newline at end of file
++endif()

pkgs/development/rocm-modules/6/rocr-debug-agent/default.nix

@@ -12,13 +12,13 @@
 
 stdenv.mkDerivation (finalAttrs: {
   pname = "rocr-debug-agent";
-  version = "6.0.2";
+  version = "6.3.3";
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "rocr_debug_agent";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-8Q800T7mwBy8/rujVNyCQ0ZpZ9uPKKk+Sv9ibpWou/8=";
+    hash = "sha256-HYag5/E72hopDhS9EVcdyGgSvzbCMzKqLC+SIS28Y9M=";
   };
 
   nativeBuildInputs = [
@@ -45,8 +45,8 @@ stdenv.mkDerivation (finalAttrs: {
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -55,8 +55,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ ncsa ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -1,34 +1,38 @@
-{ lib
-, stdenv
-, fetchFromGitHub
-, rocmUpdateScript
-, cmake
-, rocm-cmake
-, clr
-, gtest
-, gbenchmark
-, buildTests ? false
-, buildBenchmarks ? false
-, gpuTargets ? [ ]
+{
+  lib,
+  stdenv,
+  fetchFromGitHub,
+  rocmUpdateScript,
+  cmake,
+  rocm-cmake,
+  clr,
+  gtest,
+  gbenchmark,
+  buildTests ? false,
+  buildBenchmarks ? false,
+  gpuTargets ? clr.localGpuTargets or [ ],
 }:
 
 stdenv.mkDerivation (finalAttrs: {
-  pname = "rocrand";
-  version = "6.0.2";
+  pname = "rocrand${clr.gpuArchSuffix}";
+  version = "6.3.3";
 
-  outputs = [
-    "out"
-  ] ++ lib.optionals buildTests [
-    "test"
-  ] ++ lib.optionals buildBenchmarks [
-    "benchmark"
-  ];
+  outputs =
+    [
+      "out"
+    ]
+    ++ lib.optionals buildTests [
+      "test"
+    ]
+    ++ lib.optionals buildBenchmarks [
+      "benchmark"
+    ];
 
   src = fetchFromGitHub {
     owner = "ROCm";
     repo = "rocRAND";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-BBkcYOP+zh3OQTxuSkeiJizwnE9Gr5Jbhx0e8SU/mmU=";
+    hash = "sha256-rrRLPqEw39M+6dtPW8DcnQiSZNwxWNINJ1wjU098Vkk=";
   };
 
   nativeBuildInputs = [
@@ -37,45 +41,52 @@ stdenv.mkDerivation (finalAttrs: {
     clr
   ];
 
-  buildInputs = lib.optionals buildTests [
-    gtest
-  ] ++ lib.optionals buildBenchmarks [
-    gbenchmark
-  ];
+  buildInputs =
+    lib.optionals buildTests [
+      gtest
+    ]
+    ++ lib.optionals buildBenchmarks [
+      gbenchmark
+    ];
 
-  cmakeFlags = [
-    "-DCMAKE_C_COMPILER=hipcc"
-    "-DCMAKE_CXX_COMPILER=hipcc"
-    "-DHIP_ROOT_DIR=${clr}"
-    # Manually define CMAKE_INSTALL_<DIR>
-    # See: https://github.com/NixOS/nixpkgs/pull/197838
-    "-DCMAKE_INSTALL_BINDIR=bin"
-    "-DCMAKE_INSTALL_LIBDIR=lib"
-    "-DCMAKE_INSTALL_INCLUDEDIR=include"
-  ] ++ lib.optionals (gpuTargets != [ ]) [
-    "-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
-  ] ++ lib.optionals buildTests [
-    "-DBUILD_TEST=ON"
-  ] ++ lib.optionals buildBenchmarks [
-    "-DBUILD_BENCHMARK=ON"
-  ];
+  cmakeFlags =
+    [
+      "-DHIP_ROOT_DIR=${clr}"
+      # Manually define CMAKE_INSTALL_<DIR>
+      # See: https://github.com/NixOS/nixpkgs/pull/197838
+      "-DCMAKE_INSTALL_BINDIR=bin"
+      "-DCMAKE_INSTALL_LIBDIR=lib"
+      "-DCMAKE_INSTALL_INCLUDEDIR=include"
+    ]
+    ++ lib.optionals (gpuTargets != [ ]) [
+      "-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
+    ]
+    ++ lib.optionals buildTests [
+      "-DBUILD_TEST=ON"
+    ]
+    ++ lib.optionals buildBenchmarks [
+      "-DBUILD_BENCHMARK=ON"
+    ];
 
-  postInstall = lib.optionalString buildTests ''
-    mkdir -p $test/bin
-    mv $out/bin/test_* $test/bin
-  '' + lib.optionalString buildBenchmarks ''
-    mkdir -p $benchmark/bin
-    mv $out/bin/benchmark_* $benchmark/bin
-  '' + lib.optionalString (buildTests || buildBenchmarks) ''
-    rm -r $out/bin/rocRAND
-    # Fail if bin/ isn't actually empty
-    rmdir $out/bin
-  '';
+  postInstall =
+    lib.optionalString buildTests ''
+      mkdir -p $test/bin
+      mv $out/bin/test_* $test/bin
+    ''
+    + lib.optionalString buildBenchmarks ''
+      mkdir -p $benchmark/bin
+      mv $out/bin/benchmark_* $benchmark/bin
+    ''
+    + lib.optionalString (buildTests || buildBenchmarks) ''
+      rm -r $out/bin/rocRAND
+      # Fail if bin/ isn't actually empty
+      rmdir $out/bin
+    '';
 
   passthru.updateScript = rocmUpdateScript {
     name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    inherit (finalAttrs.src) owner;
+    inherit (finalAttrs.src) repo;
   };
 
   meta = with lib; {
@@ -84,6 +95,5 @@ stdenv.mkDerivation (finalAttrs: {
     license = with licenses; [ mit ];
     maintainers = teams.rocm.members;
     platforms = platforms.linux;
-    broken = versions.minor finalAttrs.version != versions.minor stdenv.cc.version || versionAtLeast finalAttrs.version "7.0.0";
   };
 })

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

@@ -6,6 +6,7 @@
   cmake,
   rocm-cmake,
   rocblas,
+  rocprim,
   rocsparse,
   clr,
   fmt,
@@ -14,12 +15,25 @@
   lapack-reference,
   buildTests ? false,
   buildBenchmarks ? false,
-  gpuTargets ? [ ], # gpuTargets = [ "gfx803" "gfx900" "gfx906:xnack-" ]
+  gpuTargets ? (
+    clr.localGpuTargets or [
+      "gfx900"
+      "gfx906"
+      "gfx908"
+      "gfx90a"
+      "gfx942"
+      "gfx1010"
+      "gfx1030"
+      "gfx1100"
+      "gfx1101"
+      "gfx1102"
+    ]
+  ),
 }:
 
 stdenv.mkDerivation (finalAttrs: {
-  pname = "rocsolver";
-  version = "6.0.2";
+  pname = "rocsolver${clr.gpuArchSuffix}";
+  version = "6.3.3";
 
   outputs =
     [
@@ -36,12 +50,13 @@ stdenv.mkDerivation (finalAttrs: {
     owner = "ROCm";
     repo = "rocSOLVER";
     rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-tglQpwCSFABRuEDiJrzQVFIdx9p85E2MiUYN0aoTAXo=";
+    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
     ]
@@ -51,7 +66,11 @@ stdenv.mkDerivation (finalAttrs: {
 
   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
     ]
@@ -64,8 +83,9 @@ stdenv.mkDerivation (finalAttrs: {
 
   cmakeFlags =
     [
-      "-DCMAKE_CXX_COMPILER=hipcc"
-      "-DCMAKE_CXX_FLAGS=-Wno-switch" # Way too many warnings
+      "-DHIP_CLANG_NUM_PARALLEL_JOBS=4"
+      "-DCMAKE_BUILD_TYPE=Release"
+      "-DCMAKE_VERBOSE_MAKEFILE=ON"
       # Manually define CMAKE_INSTALL_<DIR>
       # See: https://github.com/NixOS/nixpkgs/pull/197838
       "-DCMAKE_INSTALL_BINDIR=bin"
@@ -96,9 +116,8 @@ stdenv.mkDerivation (finalAttrs: {
     '';
 
   passthru.updateScript = rocmUpdateScript {
-    name = finalAttrs.pname;
-    owner = finalAttrs.src.owner;
-    repo = finalAttrs.src.repo;
+    name = "rocsolver";
+    inherit (finalAttrs.src) owner repo;
   };
 
   requiredSystemFeatures = [ "big-parallel" ];
@@ -111,8 +130,5 @@ stdenv.mkDerivation (finalAttrs: {
     platforms = platforms.linux;
     timeout = 14400; # 4 hours
     maxSilent = 14400; # 4 hours
-    broken =
-      versions.minor finalAttrs.version != versions.minor stdenv.cc.version
-      || versionAtLeast finalAttrs.version "7.0.0";
   };
 })