rocmPackages: 6.3.3 -> 6.4.3 (#427944)

This commit is contained in:
Emily
2025-09-15 16:05:16 +01:00
committed by GitHub
75 changed files with 1446 additions and 1676 deletions
+2 -2
View File
@@ -12,7 +12,7 @@
rocmSupport ? config.rocmSupport,
rocmPackages ? { },
rocmGpuTargets ? builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets,
rocmGpuTargets ? rocmPackages.clr.localGpuTargets or rocmPackages.clr.gpuTargets,
openclSupport ? false,
clblast,
@@ -150,7 +150,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
]
++ optionals rocmSupport [
(cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.clr.hipClangPath}/clang++")
(cmakeFeature "CMAKE_HIP_ARCHITECTURES" rocmGpuTargets)
(cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmGpuTargets))
]
++ optionals metalSupport [
(cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
+6 -5
View File
@@ -170,6 +170,10 @@ stdenv.mkDerivation (finalAttrs: {
]
);
env.CFLAGS = "-DADD_" + lib.optionalString rocmSupport " -fopenmp";
env.CXXFLAGS = finalAttrs.env.CFLAGS;
env.FFLAGS = "-DADD_";
cmakeFlags = [
(strings.cmakeFeature "GPU_TARGET" gpuTargetString)
(strings.cmakeBool "MAGMA_ENABLE_CUDA" cudaSupport)
@@ -179,9 +183,6 @@ stdenv.mkDerivation (finalAttrs: {
# otherwise not be set in NVCC_FLAGS or DEVCCFLAGS (which we cannot modify).
# See https://github.com/NixOS/nixpkgs/issues/281656#issuecomment-1902931289
(strings.cmakeBool "USE_FORTRAN" true)
(strings.cmakeFeature "CMAKE_C_FLAGS" "-DADD_")
(strings.cmakeFeature "CMAKE_CXX_FLAGS" "-DADD_")
(strings.cmakeFeature "FORTRAN_CONVENTION" "-DADD_")
]
++ lists.optionals cudaSupport [
(strings.cmakeFeature "CMAKE_CUDA_ARCHITECTURES" cudaArchitecturesString)
@@ -192,8 +193,8 @@ stdenv.mkDerivation (finalAttrs: {
# Can't easily apply the PR as a patch because we rely on the tarball with pregenerated
# hipified files ∴ fetchpatch of the PR will apply cleanly but fail to build
(strings.cmakeFeature "ROCM_CORE" "${rocmPackages.clr}")
(strings.cmakeFeature "CMAKE_C_COMPILER" "${rocmPackages.clr}/bin/hipcc")
(strings.cmakeFeature "CMAKE_CXX_COMPILER" "${rocmPackages.clr}/bin/hipcc")
(strings.cmakeFeature "CMAKE_C_COMPILER" "${rocmPackages.clr}/bin/clang")
(strings.cmakeFeature "CMAKE_CXX_COMPILER" "${rocmPackages.clr}/bin/clang++")
];
# Magma doesn't have a test suite we can easily run, just loose executables, all of which require a GPU.
+1 -4
View File
@@ -32,11 +32,8 @@ let
name = "rocm";
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 (finalAttrs: {
stdenv.mkDerivation (finalAttrs: {
pname = "ucx";
version = "1.19.0";
@@ -187,49 +187,52 @@ let
# TODO: Retest after ROCm 6.4 or torch 2.8
"gfx1010"
"gfx1012"
# Strix Halo seems to be broken as well, see
# https://github.com/NixOS/nixpkgs/pull/440359.
"gfx1151"
] (rocmPackages.clr.localGpuTargets or rocmPackages.clr.gpuTargets)
] rocmPackages.clr.localGpuTargets or rocmPackages.clr.gpuTargets
else
throw "No GPU targets specified"
);
# Use vendored CK as header only dep if rocmPackages' CK doesn't properly support targets
vendorComposableKernel = rocmSupport && !rocmPackages.composable_kernel.anyMfmaTarget;
rocmtoolkit_joined = symlinkJoin {
name = "rocm-merged";
paths = with rocmPackages; [
rocm-core
clr
rccl
miopen
aotriton
composable_kernel
rocrand
rocblas
rocsparse
hipsparse
rocthrust
rocprim
hipcub
roctracer
rocfft
rocsolver
hipfft
hiprand
hipsolver
hipblas-common
hipblas
hipblaslt
rocminfo
rocm-comgr
rocm-device-libs
rocm-runtime
rocm-smi
clr.icd
hipify
];
paths =
with rocmPackages;
[
rocm-core
clr
rccl
miopen
aotriton
rocrand
rocblas
rocsparse
hipsparse
rocthrust
rocprim
hipcub
roctracer
rocfft
rocsolver
hipfft
hiprand
hipsolver
hipblas-common
hipblas
hipblaslt
rocminfo
rocm-comgr
rocm-device-libs
rocm-runtime
rocm-smi
clr.icd
hipify
]
++ lib.optionals (!vendorComposableKernel) [
composable_kernel
];
# Fix `setuptools` not being found
postBuild = ''
@@ -373,8 +376,9 @@ buildPythonPackage rec {
# Doesn't pick up the environment variable?
substituteInPlace third_party/kineto/libkineto/CMakeLists.txt \
--replace-fail "\''$ENV{ROCM_SOURCE_DIR}" "${rocmtoolkit_joined}"
# Use composable kernel as dependency, rather than built-in third-party
''
# When possible, composable kernel as dependency, rather than built-in third-party
+ lib.optionalString (rocmSupport && !vendorComposableKernel) ''
substituteInPlace aten/src/ATen/CMakeLists.txt \
--replace-fail "list(APPEND ATen_HIP_INCLUDE \''${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/composable_kernel/include)" "" \
--replace-fail "list(APPEND ATen_HIP_INCLUDE \''${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/composable_kernel/library/include)" ""
@@ -412,7 +416,7 @@ buildPythonPackage rec {
export ROCM_PATH=${rocmtoolkit_joined}
export ROCM_SOURCE_DIR=${rocmtoolkit_joined}
export PYTORCH_ROCM_ARCH="${gpuTargetString}"
export CMAKE_CXX_FLAGS="-I${rocmtoolkit_joined}/include -I${rocmtoolkit_joined}/include/rocblas"
export CMAKE_CXX_FLAGS="-I${rocmtoolkit_joined}/include"
python tools/amd_build/build_amd.py
'';
@@ -15,24 +15,26 @@ let
esmi_ib_src = fetchFromGitHub {
owner = "amd";
repo = "esmi_ib_library";
rev = "esmi_pkg_ver-3.0.3";
hash = "sha256-q0w5c5c+CpXkklmSyfzc+sbkt4cHNxscGJA3AXwvHxQ=";
rev = "esmi_pkg_ver-4.2";
hash = "sha256-czF9ezkAO0PuDkXh8y639AcOZH+KVcWiXPX74H5W/nw=";
};
in
stdenv.mkDerivation (finalAttrs: {
pname = "amdsmi";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "rocm";
repo = "amdsmi";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-hrPqd4ZWqzTG7JRfVwc1SZx6TNS0Q/LFg8yDxrM3mPo=";
hash = "sha256-9O29O4HGkQxFDglAhHKv5KWA7p97RwMGG2x/fkOS2jE=";
};
postPatch = ''
substituteInPlace goamdsmi_shim/CMakeLists.txt \
--replace-fail "amd_smi)" ${"'"}''${AMD_SMI_TARGET})' \
--replace-fail 'target_link_libraries(''${GOAMDSMI_SHIM_TARGET} -L' '#'
substituteInPlace CMakeLists.txt \
--replace-fail "if(NOT latest_esmi_tag STREQUAL current_esmi_tag)" "if(OFF)"
# Manually unpack esmi_ib_src and add amd_hsmp.h so execute-process git clone doesn't run
cp -rf --no-preserve=mode ${esmi_ib_src} ./esmi_ib_library
@@ -41,10 +43,10 @@ stdenv.mkDerivation (finalAttrs: {
'';
patches = [
# Fix ld.lld undefined reference: drmGetVersion
(fetchpatch {
url = "https://github.com/ROCm/amdsmi/commit/c3864bf6171970d86dc50fd23f06377736823997.patch";
hash = "sha256-zRG1tBD8sIQCWdKfCbXC/Z/6d6NTrRYvRpddPWdM4j8=";
name = "esmi-to-tag-4.2.patch";
url = "https://github.com/ROCm/amdsmi/commit/49aa2af045a4bc688e6f3ee0545f12afc45c1efe.patch";
hash = "sha256-5dH9N4m+2mJIGVEB86SvdK3uAYyGFTfbCBJ8e09iQ3w=";
})
];
@@ -23,17 +23,21 @@ let
"gfx90a"
"gfx942"
"gfx950"
# some gfx1100 kernels fail with error: branch size exceeds simm16
# but build proceeds and those ops will fallback so it's ok
"gfx1100"
"gfx1151"
"gfx1150"
"gfx1201"
"gfx1200"
] gpuTargets;
supportedTargets' = lib.concatStringsSep ";" supportedTargets;
anySupportedTargets = supportedTargets != [ ];
# Pick a single arbitrary target to speed up shim build when we can't support our target
supportedTargets' =
if anySupportedTargets then lib.concatStringsSep ";" supportedTargets else "gfx1200";
in
stdenv.mkDerivation (finalAttrs: {
pname = "aotriton";
pname = "aotriton${lib.optionalString (!anySupportedTargets) "-shim"}";
version = "0.10b";
src = fetchFromGitHub {
@@ -132,13 +136,16 @@ stdenv.mkDerivation (finalAttrs: {
(lib.cmakeBool "AOTRITON_NOIMAGE_MODE" (!anySupportedTargets))
# Use preinstalled triton from our python's site-packages
(lib.cmakeBool "AOTRITON_INHERIT_SYSTEM_SITE_TRITON" true)
# FP32 kernels are optional, turn them off to speed up builds and save space
# Perf sensitive code should be using BF16 or F16
(lib.cmakeBool "AOTRITON_ENABLE_FP32_INPUTS" false)
# Avoid kernels being skipped if build host is overloaded
(lib.cmakeFeature "AOTRITON_GPU_BUILD_TIMEOUT" "0")
# Manually define CMAKE_INSTALL_<DIR>
# See: https://github.com/NixOS/nixpkgs/pull/197838
(lib.cmakeFeature "CMAKE_INSTALL_BINDIR" "bin")
(lib.cmakeFeature "CMAKE_INSTALL_LIBDIR" "lib")
(lib.cmakeFeature "CMAKE_INSTALL_INCLUDEDIR" "include")
]
++ lib.optionals anySupportedTargets [
# Note: build will warn "AMDGPU_TARGETS was not set, and system GPU detection was unsuccsesful."
# but this can safely be ignored, aotriton uses a different approach to pass targets
(lib.cmakeFeature "AOTRITON_TARGET_ARCH" supportedTargets')
@@ -0,0 +1,36 @@
{
lib,
stdenv,
clr,
cmake,
fetchFromGitHub,
}:
stdenv.mkDerivation {
pname = "aqlprofile";
version = "6.4.3";
src = fetchFromGitHub {
# TODO: Will move to rocm-systems repo and have proper tags in 7.x
# pinned to oddly named tag for now
owner = "ROCm";
repo = "aqlprofile";
tag = "rocm-42";
hash = "sha256-avL78ZfB+rJ1TYaejSUzU6i5L9JeMawMwIxaTQINQdE=";
};
env.CXXFLAGS = "-DROCP_LD_AQLPROFILE=1";
nativeBuildInputs = [
cmake
clr
];
meta = with lib; {
description = "AQLPROFILE library for AMD HSA runtime API extension support";
homepage = "https://github.com/ROCm/aqlprofile/";
license = with licenses; [ mit ];
teams = [ teams.rocm ];
platforms = platforms.linux;
};
}
+40 -35
View File
@@ -47,26 +47,29 @@ let
"--set ROCM_PATH $out"
];
amdclang = writeShellScriptBin "amdclang" ''
exec clang "$@"
exec ${hipClang}/bin/clang "$@"
'';
amdclangxx = writeShellScriptBin "amdclang++" ''
exec clang++ "$@"
exec ${hipClang}/bin/clang++ "$@"
'';
in
stdenv.mkDerivation (finalAttrs: {
pname = "clr";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
"icd"
];
__structuredAttrs = true;
strictDeps = true;
src = fetchFromGitHub {
owner = "ROCm";
repo = "clr";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-4qjfnn0kto2sNaSumXxHRHFrf3a3RZILOdhVSxkEs1I=";
hash = "sha256-DOAAuC9TN1//v56GXyUMJwQHgOuctC+WsC5agrgL+QM=";
};
nativeBuildInputs = [
@@ -99,8 +102,10 @@ stdenv.mkDerivation (finalAttrs: {
rocminfo
];
cmakeBuildType = "RelWithDebInfo";
separateDebugInfo = true;
cmakeFlags = [
"-DCMAKE_BUILD_TYPE=Release"
"-DCMAKE_POLICY_DEFAULT_CMP0072=NEW" # Prefer newer OpenGL libraries
"-DCLR_BUILD_HIP=ON"
"-DCLR_BUILD_OCL=ON"
@@ -124,33 +129,27 @@ stdenv.mkDerivation (finalAttrs: {
patches = [
./cmake-find-x11-libgl.patch
(fetchpatch {
# 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";
# [PATCH] improve rocclr isa compatibility check
sha256 = "sha256-oj1loBEuqzuMihOKoN0wR92Wo25AshN5MpBuTq/9TMw=";
url = "https://github.com/GZGavinZhao/clr/commit/f675b9b46d9f7bb8e003f4f47f616fa86a0b7a5e.patch";
})
(fetchpatch {
# improve rocclr isa compatibility check
sha256 = "sha256-wUrhpYN68AbEXeFU5f366C6peqHyq25kujJXY/bBJMs=";
url = "https://github.com/GZGavinZhao/clr/commit/22c17a0ac09c6b77866febf366591f669a1ed133.patch";
})
(fetchpatch {
# [PATCH] Improve hipamd compat check
sha256 = "sha256-uZQ8rMrWH61CCbxwLqQGggDmXFmYTi6x8OcgYPrZRC8=";
url = "https://github.com/GZGavinZhao/clr/commit/63c6ee630966744d4199fdfb854e98d2da9e1122.patch";
})
(fetchpatch {
# [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";
# [PATCH] improve hipamd isa compatibility check
sha256 = "sha256-E3ERoVjUVWCiYHuE1GaVY5jMrAVx3B1cAVHM4/HPuaQ=";
url = "https://github.com/GZGavinZhao/clr/commit/aec0fc56ee2d10a2bc269c418fa847da2ee9969a.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";
})
(fetchpatch {
# CMake 4 compat
# [PATCH] SWDEV-509213 - make cmake_minimum_required consistent across clr
url = "https://github.com/ROCm/clr/commit/fcaefe97b862afe12aaac0147f1004e6dc595fce.patch";
hash = "sha256-hRZXbASbIOOETe+T4mDyyiRWLXd6RDKRieN2ns1w/rs=";
})
];
postPatch = ''
@@ -207,6 +206,7 @@ stdenv.mkDerivation (finalAttrs: {
ln -s ${rocm-core}/.info/ $out/.info
ln -s ${hipClang} $out/llvm
ln -s ${hipClang}/bin/{ld.lld,lld,clang-offload-bundler,llvm-objcopy,clang,clang++} $out/bin/
'';
disallowedRequisites = [
@@ -217,28 +217,33 @@ stdenv.mkDerivation (finalAttrs: {
# All known and valid general GPU targets
# We cannot use this for each ROCm library, as each defines their own supported targets
# See: https://github.com/ROCm/ROCm/blob/77cbac4abab13046ee93d8b5bf410684caf91145/README.md#library-target-matrix
# Generic targets are not yet available in rocm-6.3.1 llvm
gpuTargets = lib.forEach [
# "9-generic"
# "9-generic" # can handle all Vega variants
"900" # MI25, Vega 56/64
"906" # MI50/60, Radeon VII
"908" # MI100
"90a" # MI210 / MI250
# "9-4-generic"
# "902" # Vega 8
# "909" # Renoir Vega APU
# "90c" # Renoir Vega APU
# Past this point cards need their own kernels for perf despite gfx9-generic compat
"906" # MI50/60, Radeon VII - adds dot product & mixed precision FMA ops
"908" # MI100 - adds MFMA (matrix fused multiply-add) ops
"90a" # MI210/MI250 - additional MFMA variants
# "9-4-generic" - since only 942 is valid for 6.4 target it directly
# 940/1 - never released publicly, maybe HPE cray specific MI3xx?
"942" # MI300
# "10-1-generic"
"942" # MI300A/X, MI325X
# "950" # MI350X TODO: Expected in ROCm 7.x
# "10-1-generic" # fine for all RDNA1 cards
"1010"
"1012"
# "10-3-generic"
"1030" # W6800, various Radeon cards
# "11-generic"
# "11-generic" # will handle 7600, hopefully ryzen AI series iGPUs
"1100"
"1101"
"1102"
# 7.x "1150"
"1151" # Strix Halo
"1200" # RX 9070
"1201" # RX 9070 XT
# "12-generic"
"1200" # RX 9060
"1201" # RX 9070 + XT
] (target: "gfx${target}");
inherit hipClangPath;
@@ -0,0 +1,40 @@
From 6ac72ec84269737626b1f5e43e64729f0922d182 Mon Sep 17 00:00:00 2001
From: "Ding, Yi" <yi.ding@amd.com>
Date: Wed, 9 Jul 2025 03:12:39 +0000
Subject: [PATCH] Avoid compile kernel in host pass
---
include/ck_tile/host/kernel_launch.hpp | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/include/ck_tile/host/kernel_launch.hpp b/include/ck_tile/host/kernel_launch.hpp
index 9770e99738..f6ccb6968b 100644
--- a/include/ck_tile/host/kernel_launch.hpp
+++ b/include/ck_tile/host/kernel_launch.hpp
@@ -4,11 +4,12 @@
#pragma once
#include "ck_tile/core/config.hpp"
-#include "ck_tile/host/stream_config.hpp"
+#include "ck_tile/core/utility/ignore.hpp"
#include "ck_tile/host/hip_check_error.hpp"
+#include "ck_tile/host/stream_config.hpp"
#include "ck_tile/host/timer.hpp"
-#include <hip/hip_runtime.h>
#include <cstddef>
+#include <hip/hip_runtime.h>
namespace ck_tile {
@@ -24,7 +25,11 @@ __launch_bounds__(MaxThreadPerBlock, MinBlockPerCu)
#endif
__global__ void kentry(Args... args)
{
+#if defined(__HIP_DEVICE_COMPILE__)
Kernel{}(args...);
+#else
+ (..., (ignore = args, 0));
+#endif
}
//
@@ -32,6 +32,10 @@
),
}:
# TODO: in 7.x CK is likely to gain support for
# a) miopen kernel only build (MIOPEN_REQ_LIBS_ONLY)
# b) header only build (useful for torch) https://github.com/ROCm/composable_kernel/issues/2030
# that will likely allow us to get rid of this complicated split part build!
stdenv.mkDerivation (finalAttrs: {
preBuild = ''
echo "This derivation isn't intended to be built directly and only exists to be overridden and built in chunks";
@@ -39,9 +43,7 @@ stdenv.mkDerivation (finalAttrs: {
'';
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-2024-12-20";
version = "6.4-unstable-2025-05-22";
outputs = [
"out"
@@ -56,8 +58,9 @@ stdenv.mkDerivation (finalAttrs: {
src = fetchFromGitHub {
owner = "ROCm";
repo = "composable_kernel";
rev = "07339c738396ebeae57374771ded4dcf11bddf1e";
hash = "sha256-EvEBxlOpQ71BF57VW79WBo/cdxAwTKFXFMiYKyGyyEs=";
# Using a dev snapshot, trying to get MIOpen to work
rev = "bc2551ac3b27edc31f20863e3a873508fb73aad2";
hash = "sha256-bfmwbgR1ya+zkME3wOyaZX/e+1+ie0sSlugK/kozLsI=";
};
nativeBuildInputs = [
@@ -100,7 +103,9 @@ stdenv.mkDerivation (finalAttrs: {
"-DCMAKE_INSTALL_LIBDIR=lib"
"-DCMAKE_INSTALL_INCLUDEDIR=include"
"-DBUILD_DEV=OFF"
"-DBUILD_MHA_LIB=ON"
"-DROCM_PATH=${clr}"
"-DENABLE_CLANG_CPP_CHECKS=OFF"
"-DCMAKE_HIP_COMPILER_ROCM_ROOT=${clr}"
# FP8 can build for 908/90a but very slow build
@@ -118,6 +123,11 @@ stdenv.mkDerivation (finalAttrs: {
"-DGOOGLETEST_DIR=${gtest.src}" # Custom linker names
];
patches = [
# Significant build performance improvement
./avoid-extra-host-compile.patch
];
# No flags to build selectively it seems...
postPatch =
# Reduce configure time by preventing thousands of clang-tidy targets being added
@@ -128,6 +138,14 @@ stdenv.mkDerivation (finalAttrs: {
--replace-fail clang_tidy_check '#clang_tidy_check'
substituteInPlace CMakeLists.txt \
--replace-fail "add_subdirectory(profiler)" ""
substituteInPlace cmake/EnableCompilerWarnings.cmake \
--replace-fail "-Werror" ""
# Apply equivalent change to https://github.com/ROCm/composable_kernel/pull/2564
# TODO: Remove after ROCm 7.1
find include/ck/tensor_operation/ -type f -name "*.hpp" -exec sed -i \
-e 's/!defined(__HIP_DEVICE_COMPILE__)/false/g' \
{} +
''
# Optionally remove tests
+ lib.optionalString (!buildTests) ''
@@ -152,20 +170,29 @@ stdenv.mkDerivation (finalAttrs: {
mv $out/bin/example_* $example/bin
'';
passthru.updateScript = rocmUpdateScript {
name = finalAttrs.pname;
inherit (finalAttrs.src) owner;
inherit (finalAttrs.src) repo;
passthru = {
inherit gpuTargets;
updateScript = rocmUpdateScript {
name = finalAttrs.pname;
inherit (finalAttrs.src) owner;
inherit (finalAttrs.src) repo;
};
anyGfx9Target = lib.lists.any (lib.strings.hasPrefix "gfx9") gpuTargets;
anyMfmaTarget =
(lib.lists.intersectLists gpuTargets [
"gfx908"
"gfx90a"
"gfx942"
"gfx950"
]) != [ ];
};
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 ];
teams = [ teams.rocm ];
platforms = platforms.linux;
broken = true;
broken = true; # this base package shouldn't be built directly
};
})
@@ -14,7 +14,7 @@ buildPythonPackage {
setuptools
setuptools-scm
];
version = "6.4.0";
version = "6.4.3";
inherit (composable_kernel) src;
pythonImportsCheck = [
"ck4inductor"
@@ -17,61 +17,90 @@ let
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" ];
onlyFor = [
"gfx942"
"gfx950"
];
};
grouped_conv = {
gemm_multiply_multiply_wp = {
targets = [
"device_gemm_multiply_multiply_wp_instance"
];
extraCmakeFlags = [ "-DHIP_CLANG_NUM_PARALLEL_JOBS=2" ];
onlyFor = [
"gfx942"
"gfx950"
];
};
grouped_conv_bwd = {
targets = [
"device_grouped_conv1d_bwd_weight_instance"
"device_grouped_conv2d_bwd_data_instance"
"device_grouped_conv2d_bwd_weight_instance"
];
};
grouped_conv_fwd = {
targets = [
"device_grouped_conv1d_fwd_instance"
"device_grouped_conv2d_fwd_instance"
"device_grouped_conv2d_fwd_bias_relu_instance"
"device_grouped_conv2d_fwd_dynamic_op_instance"
];
requiredSystemFeatures = [ "big-parallel" ];
};
grouped_conv_bwd_3d = {
grouped_conv_bwd_3d1 = {
targets = [
"device_grouped_conv3d_bwd_data_instance"
"device_grouped_conv3d_bwd_data_bilinear_instance"
"device_grouped_conv3d_bwd_data_scale_instance"
];
};
grouped_conv_bwd_3d2 = {
targets = [
"device_grouped_conv3d_bwd_weight_instance"
"device_grouped_conv3d_bwd_weight_bilinear_instance"
"device_grouped_conv3d_bwd_weight_scale_instance"
];
requiredSystemFeatures = [ "big-parallel" ];
};
grouped_conv_fwd_3d = {
grouped_conv_fwd_3d1 = {
targets = [
"device_grouped_conv3d_fwd_instance"
"device_grouped_conv3d_fwd_bias_relu_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"
];
};
grouped_conv_fwd_3d2 = {
targets = [
"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"
];
requiredSystemFeatures = [ "big-parallel" ];
};
batched_gemm = {
batched_gemm1 = {
targets = [
"device_batched_gemm_instance"
"device_batched_gemm_b_scale_instance"
"device_batched_gemm_multi_d_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"
];
};
batched_gemm2 = {
targets = [
"device_batched_gemm_softmax_gemm_permute_instance"
"device_grouped_gemm_instance"
"device_grouped_gemm_bias_instance"
@@ -80,27 +109,33 @@ let
"device_grouped_gemm_fixed_nk_multi_abd_instance"
"device_grouped_gemm_tile_loop_instance"
];
requiredSystemFeatures = [ "big-parallel" ];
};
gemm_universal = {
gemm_universal1 = {
targets = [
"device_gemm_universal_instance"
"device_gemm_universal_batched_instance"
];
};
gemm_universal2 = {
targets = [
"device_gemm_universal_reduce_instance"
"device_gemm_universal_streamk_instance"
];
requiredSystemFeatures = [ "big-parallel" ];
extraCmakeFlags = [ "-DHIP_CLANG_NUM_PARALLEL_JOBS=2" ];
};
gemm_other = {
gemm_other1 = {
targets = [
"device_gemm_instance"
"device_gemm_b_scale_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"
];
};
gemm_other2 = {
targets = [
"device_gemm_add_relu_add_layernorm_instance"
"device_gemm_add_silu_instance"
"device_gemm_bias_add_reduce_instance"
@@ -112,7 +147,6 @@ let
"device_gemm_splitk_instance"
"device_gemm_streamk_instance"
];
requiredSystemFeatures = [ "big-parallel" ];
};
conv = {
targets = [
@@ -123,7 +157,6 @@ let
"device_conv2d_fwd_bias_relu_add_instance"
"device_conv3d_bwd_data_instance"
];
requiredSystemFeatures = [ "big-parallel" ];
};
pool = {
targets = [
@@ -141,11 +174,7 @@ let
"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"
];
requiredSystemFeatures = [ "big-parallel" ];
};
other2 = {
targets = [
@@ -154,10 +183,16 @@ let
"device_permute_scale_instance"
"device_quantization_instance"
"device_reduce_instance"
];
};
other3 = {
targets = [
"device_normalization_bwd_data_instance"
"device_normalization_bwd_gamma_beta_instance"
"device_normalization_fwd_instance"
"device_softmax_instance"
"device_transpose_instance"
];
requiredSystemFeatures = [ "big-parallel" ];
};
};
tensorOpBuilder =
@@ -165,51 +200,59 @@ let
part,
targets,
extraCmakeFlags ? [ ],
requiredSystemFeatures ? [ ],
requiredSystemFeatures ? [ "big-parallel" ],
onlyFor ? [ ],
}:
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:' ""
'';
let
supported =
onlyFor == [ ] || (lib.lists.intersectLists composable_kernel_base.gpuTargets onlyFor) != [ ];
in
if supported then
(composable_kernel_base.overrideAttrs (old: {
inherit requiredSystemFeatures;
pname = "composable_kernel${clr.gpuArchSuffix}-${part}";
makeTargets = targets;
preBuild = ''
echo "Building ${part}"
makeFlagsArray+=($makeTargets)
substituteInPlace $(find ./ -name "Makefile" -type f) \
--replace-fail '.NOTPARALLEL:' '.UNUSED_NOTPARALLEL:'
'';
# Compile parallelism adjusted based on available RAM
# Never uses less than NIX_BUILD_CORES/4, never uses more than NIX_BUILD_CORES
# CK uses an unusually high amount of memory per core in the build step
# Nix/nixpkgs doesn't really have any infra to tell it that this build is unusually memory hungry
# So, bodge. Otherwise you end up having to build all of ROCm with a low core limit when
# it's only this package that has trouble.
preConfigure = old.preConfigure or "" + ''
MEM_GB_TOTAL=$(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;
};
});
# 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;
};
}))
else
null;
composable_kernel_parts = builtins.mapAttrs (
part: targets: tensorOpBuilder (targets // { inherit part; })
) parts;
@@ -218,8 +261,8 @@ in
composable_kernel_base.overrideAttrs (
finalAttrs: old: {
pname = "composable_kernel${clr.gpuArchSuffix}";
parts_dirs = builtins.attrValues composable_kernel_parts;
disallowedReferences = builtins.attrValues composable_kernel_parts;
parts_dirs = builtins.filter (x: x != null) (builtins.attrValues composable_kernel_parts);
disallowedReferences = builtins.filter (x: x != null) (builtins.attrValues composable_kernel_parts);
preBuild = ''
for dir in $parts_dirs; do
find "$dir" -type f -name "*.o" | while read -r file; do
@@ -240,7 +283,7 @@ composable_kernel_base.overrideAttrs (
parts = composable_kernel_parts;
};
meta = old.meta // {
# Builds which don't don't target any gfx9 cause cmake errors in dependent projects
# Builds without any gfx9 fail
broken = !finalAttrs.passthru.anyGfx9Target;
};
}
+39 -43
View File
@@ -6,15 +6,12 @@
recurseIntoAttrs,
symlinkJoin,
fetchFromGitHub,
ffmpeg_4,
boost179,
opencv,
libjpeg_turbo,
python3Packages,
triton-llvm,
openmpi,
stdenv,
rocmGpuArches ? [ ],
}:
let
@@ -22,14 +19,16 @@ let
self:
let
inherit (self) llvm;
origStdenv = stdenv;
pyPackages = python3Packages;
openmpi-orig = openmpi;
rocmClangStdenv = llvm.rocmClangStdenv;
in
{
inherit rocmGpuArches;
inherit rocmClangStdenv;
stdenv = rocmClangStdenv;
buildTests = false;
buildBenchmarks = false;
stdenv = llvm.rocmClangStdenv;
rocmPath = self.callPackage ./rocm-path { };
rocmUpdateScript = self.callPackage ./update.nix { };
@@ -37,38 +36,43 @@ let
## ROCm ##
llvm = recurseIntoAttrs (
callPackage ./llvm/default.nix {
inherit (self) rocm-device-libs rocm-runtime;
# rocm-device-libs is used for .src only
# otherwise would cause infinite recursion
inherit (self) rocm-device-libs;
}
);
inherit (self.llvm) rocm-merged-llvm clang openmp;
rocm-core = self.callPackage ./rocm-core { };
rocm-core = self.callPackage ./rocm-core { stdenv = origStdenv; };
rocm-cmake = self.callPackage ./rocm-cmake { stdenv = origStdenv; };
rocm-device-libs = self.callPackage ./rocm-device-libs {
stdenv = origStdenv;
inherit (llvm) rocm-merged-llvm;
};
rocm-runtime = self.callPackage ./rocm-runtime {
stdenv = origStdenv;
inherit (llvm) rocm-merged-llvm;
};
rocm-comgr = self.callPackage ./rocm-comgr {
stdenv = origStdenv;
inherit (llvm) rocm-merged-llvm;
};
rocminfo = self.callPackage ./rocminfo { stdenv = origStdenv; };
amdsmi = pyPackages.callPackage ./amdsmi {
inherit (self) rocmUpdateScript;
};
rocm-cmake = self.callPackage ./rocm-cmake { };
rocm-smi = pyPackages.callPackage ./rocm-smi {
inherit (self) rocmUpdateScript;
};
rocm-device-libs = self.callPackage ./rocm-device-libs {
inherit (llvm) rocm-merged-llvm;
};
rocm-runtime = self.callPackage ./rocm-runtime {
inherit (llvm) rocm-merged-llvm;
};
rocm-comgr = self.callPackage ./rocm-comgr {
inherit (llvm) rocm-merged-llvm;
};
rocminfo = self.callPackage ./rocminfo { };
# Unfree
hsa-amd-aqlprofile-bin = self.callPackage ./hsa-amd-aqlprofile-bin { };
aqlprofile = self.callPackage ./aqlprofile { };
rdc = self.callPackage ./rdc { };
@@ -76,17 +80,18 @@ let
hip-common = self.callPackage ./hip-common { };
# Eventually will be in the LLVM repo
hipcc = self.callPackage ./hipcc {
stdenv = origStdenv;
inherit (llvm) rocm-merged-llvm;
};
# Replaces hip, opencl-runtime, and rocclr
clr = self.callPackage ./clr { };
aotriton = self.callPackage ./aotriton { inherit stdenv; };
aotriton = self.callPackage ./aotriton { stdenv = origStdenv; };
hipify = self.callPackage ./hipify {
stdenv = origStdenv;
inherit (llvm)
clang
rocm-merged-llvm
@@ -103,7 +108,6 @@ let
inherit (llvm) clang;
};
# Needs GCC
roctracer = self.callPackage ./roctracer { };
rocgdb = self.callPackage ./rocgdb { };
@@ -128,13 +132,6 @@ let
rccl = self.callPackage ./rccl { };
# 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 { };
@@ -194,7 +191,7 @@ let
miopen-hip = self.miopen;
migraphx = self.callPackage ./migraphx { };
migraphx = self.callPackage ./migraphx { stdenv = origStdenv; };
rpp = self.callPackage ./rpp { };
@@ -214,10 +211,8 @@ let
};
mivisionx = self.callPackage ./mivisionx {
stdenv = origStdenv;
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 {
@@ -275,7 +270,7 @@ let
rocm-developer-tools = symlinkJoin {
name = "rocm-developer-tools-meta";
paths = [
hsa-amd-aqlprofile-bin
aqlprofile
rocm-core
rocr-debug-agent
roctracer
@@ -298,7 +293,6 @@ let
name = "rocm-ml-libraries-meta";
paths = [
llvm.clang
llvm.mlir
llvm.openmp
rocm-core
miopen-hip
@@ -357,7 +351,6 @@ let
paths = [
rocm-core
llvm.clang
llvm.mlir
llvm.openmp # openmp-extras-devel (https://github.com/ROCm/aomp)
rocm-language-runtime
];
@@ -389,7 +382,6 @@ let
hipify
rocm-cmake
llvm.clang
llvm.mlir
llvm.openmp
rocm-runtime
rocm-hip-runtime
@@ -440,6 +432,10 @@ let
};
}
// lib.optionalAttrs config.allowAliases {
hsa-amd-aqlprofile-bin = lib.warn ''
'hsa-amd-aqlprofile-bin' has been replaced by 'aqlprofile'.
'' self.aqlprofile; # Added 2025-08-27
triton = throw ''
'rocmPackages.triton' has been removed. Please use python3Packages.triton
''; # Added 2025-08-24
@@ -9,7 +9,7 @@
stdenv.mkDerivation (finalAttrs: {
pname = "half";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
@@ -7,13 +7,13 @@
stdenv.mkDerivation (finalAttrs: {
pname = "hip-common";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "HIP";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-eKDbuG57KCg/nZpy+RcDsjyw0pP68qO9K/3ZpSPuMNw=";
hash = "sha256-B4Gc119iff3ak9tmpz3rUJBtCk5T1AA8z67K9PshTLQ=";
};
dontConfigure = true;
@@ -8,13 +8,13 @@
}:
stdenv.mkDerivation (finalAttrs: {
pname = "hipblas-common";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "hipBLAS-common";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-tvNz4ymQ1y3YSUQxAtNu2who79QzSKR+3JEevr+GDWo=";
hash = "sha256-eTwoAXH2HGdSAOLTZHJUFHF+c2wWHixqeMqr60KxJrc=";
};
nativeBuildInputs = [
@@ -18,12 +18,14 @@
buildTests ? false,
buildBenchmarks ? false,
buildSamples ? false,
# for passthru.tests
hipblas,
}:
# Can also use cuBLAS
stdenv.mkDerivation (finalAttrs: {
pname = "hipblas";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -42,15 +44,16 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "hipBLAS";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-Rz1KAhBUbvErHTF2PM1AkVhqo4OHldfSNMSpp5Tx9yk=";
hash = "sha256-lQv8Ik6+0ldqyeJ05CSGB0309nIpzlRL3CRYeQxVfd0=";
};
patches = [
# https://github.com/ROCm/hipBLAS/pull/952
(fetchpatch {
name = "transitively-depend-hipblas-common.patch";
url = "https://github.com/ROCm/hipBLAS/commit/54220fdaebf0fb4fd0921ee9e418ace5b143ec8f.patch";
hash = "sha256-MFEhv8Bkrd2zD0FFIDg9oJzO7ztdyMAF+R9oYA0rmwQ=";
# Subject: [PATCH] Add gfx1150, gfx1150, gfx1200, gfx1201 support (#1055)
# This was merged to release/rocm-rel-6.4 but AMD forgot to tag it for 6.4.3
name = "release-6.4-arch-extra.patch";
url = "https://github.com/ROCm/hipBLAS/commit/0100b32ccff9a0f12134694315b4e44884e25a8e.patch";
hash = "sha256-BmktlLJpYaTcogHzEKpZdCnksIIysEO47WMezXoxvCs=";
})
];
@@ -82,15 +85,13 @@ stdenv.mkDerivation (finalAttrs: {
];
cmakeFlags = [
"-DCMAKE_BUILD_TYPE=Release"
"-DCMAKE_CXX_COMPILER=${lib.getExe' clr "hipcc"}"
# Upstream is migrating to amdclang++, it is likely this will be correct in next version bump
#"-DCMAKE_CXX_COMPILER=${lib.getBin clr}/bin/amdclang++"
"-DCMAKE_CXX_COMPILER=${lib.getExe' clr "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"
"-DBUILD_WITH_SOLVER=ON"
"-DAMDGPU_TARGETS=${rocblas.amdgpu_targets}"
]
++ lib.optionals buildTests [
@@ -125,6 +126,11 @@ stdenv.mkDerivation (finalAttrs: {
inherit (finalAttrs.src) owner;
inherit (finalAttrs.src) repo;
};
passthru.tests.hipblas-tested = hipblas.override {
buildTests = true;
buildBenchmarks = true;
buildSamples = true;
};
meta = with lib; {
description = "ROCm BLAS marshalling library";
@@ -1,232 +1,229 @@
{
lib,
stdenv,
fetchpatch,
fetchFromGitHub,
writableTmpDirAsHomeHook,
cmake,
rocm-cmake,
rocm-smi,
pkg-config,
clr,
gfortran,
gtest,
msgpack,
boost,
msgpack-cxx,
amd-blis,
libxml2,
python3,
python3Packages,
openmp,
hipblas-common,
tensile,
lapack-reference,
ncurses,
ninja,
libffi,
zlib,
zstd,
rocmUpdateScript,
buildTests ? false,
buildBenchmarks ? false,
buildSamples ? false,
# hipblaslt supports only devices with MFMA or WMMA
# WMMA on gfx1100 may be broken
# MFMA on MI100 may be broken
# MI200/MI300 known to work
gpuTargets ? (
clr.localGpuTargets or [
# "gfx908" FIXME: confirm MFMA on MI100 works
"gfx90a"
"gfx942"
# "gfx1100" FIXME: confirm WMMA targets work
]
),
gpuTargets ? (clr.localGpuTargets or clr.gpuTargets),
}:
stdenv.mkDerivation (
finalAttrs:
let
supportsTargetArches =
(builtins.any (lib.strings.hasPrefix "gfx9") gpuTargets)
|| (builtins.any (lib.strings.hasPrefix "gfx11") gpuTargets);
tensile' = (tensile.override { isTensileLite = true; }).overrideAttrs {
inherit (finalAttrs) src;
sourceRoot = "${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"
let
# hipblaslt is extremely particular about what it will build with
# so intersect with a known supported list and use only those
supportedTargets = (
lib.lists.intersectLists gpuTargets [
"gfx908"
"gfx90a"
"gfx942"
"gfx950"
"gfx1100"
"gfx1101"
# 7.x "gfx1150"
"gfx1151"
"gfx1200"
"gfx1201"
]
++ lib.optionals buildTests [
"test"
]
++ lib.optionals buildBenchmarks [
"benchmark"
]
++ lib.optionals buildSamples [
"sample"
];
);
supportsTargetArches = supportedTargets != [ ];
py = python3.withPackages (ps: [
ps.pyyaml
ps.setuptools
ps.packaging
ps.nanobind
ps.joblib
ps.msgpack
]);
# workaround: build for one working target if no targets are supported
# a few CXX files are still build for the device
gpuTargets' =
if supportsTargetArches then (lib.concatStringsSep ";" supportedTargets) else "gfx1200";
compiler = "amdclang++";
# no-switch due to spammy warnings on some cases with fixme messages
# FIXME(LunNova@): cmake files need patched to include this properly or
# maybe we improve the toolchain to use config files + assemble a sysroot
# so system wide include assumptions work
cFlags = "-Wno-switch -fopenmp -I${lib.getDev zstd}/include -I${amd-blis}/include/blis/ -I${lib.getDev msgpack-cxx}/include";
in
stdenv.mkDerivation (finalAttrs: {
pname = "hipblaslt${clr.gpuArchSuffix}";
version = "6.5-unstable-2025-08-21";
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 '${"\${rocm_path}"}/bin/' ""
'';
src = fetchFromGitHub {
owner = "ROCm";
repo = "rocm-libraries";
rev = "a676499add42941ff6af1e8d3f0504416dac7429";
hash = "sha256-zIYdHFbHyP2V6dkx6Ueb6NBqWu8tJji2hSWF9zWEJa4=";
sparseCheckout = [ "projects/hipblaslt" ];
};
sourceRoot = "${finalAttrs.src.name}/projects/hipblaslt";
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++";
requiredSystemFeatures = [ "big-parallel" ];
# 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";
})
];
__structuredAttrs = true;
strictDeps = true;
doCheck = false;
doInstallCheck = false;
outputs = [
"out"
# benchmarks are non-optional
"benchmark"
]
++ lib.optionals buildTests [
"test"
]
++ lib.optionals buildSamples [
"sample"
];
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
];
patches = [
# Upstream issue requesting properly specifying
# parallel-jobs for these invocations
# https://github.com/ROCm/rocm-libraries/issues/1242
./parallel-buildSourceCodeObjectFile.diff
# Support loading zstd compressed .dat files, required to keep output under
# hydra size limit
./messagepack-compression-support.patch
];
buildInputs = [
hipblas-common
tensile'
openmp
libffi
ncurses
postPatch = ''
# git isn't needed and we have no .git
substituteInPlace cmake/dependencies.cmake \
--replace-fail "find_package(Git REQUIRED)" ""
substituteInPlace CMakeLists.txt \
--replace-fail " LANGUAGES CXX" " LANGUAGES CXX C ASM"
'';
# 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
];
doCheck = false;
doInstallCheck = false;
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"
];
nativeBuildInputs = [
cmake
rocm-cmake
py
clr
gfortran
pkg-config
ninja
rocm-smi
];
postInstall =
lib.optionalString buildTests ''
buildInputs = [
clr
rocm-cmake
hipblas-common
amd-blis
rocm-smi
openmp
libffi
ncurses
lapack-reference
# Tensile deps - not optional, building without tensile isn't actually supported
msgpack-cxx
libxml2
python3Packages.msgpack
python3Packages.joblib
zlib
zstd
]
++ lib.optionals buildTests [
gtest
];
cmakeFlags = [
(lib.cmakeFeature "Boost_INCLUDE_DIR" "${lib.getDev boost}/include") # msgpack FindBoost fails to find boost
(lib.cmakeFeature "GPU_TARGETS" gpuTargets')
(lib.cmakeBool "BUILD_TESTING" buildTests)
(lib.cmakeBool "HIPBLASLT_ENABLE_BLIS" true)
(lib.cmakeBool "HIPBLASLT_BUILD_TESTING" buildTests)
(lib.cmakeBool "HIPBLASLT_ENABLE_SAMPLES" buildSamples)
(lib.cmakeBool "HIPBLASLT_ENABLE_DEVICE" supportsTargetArches)
# FIXME: Enable for ROCm 7.x
(lib.cmakeBool "HIPBLASLT_ENABLE_ROCROLLER" false)
"-DCMAKE_C_COMPILER=amdclang"
"-DCMAKE_HIP_COMPILER=${compiler}"
"-DCMAKE_CXX_COMPILER=${compiler}"
"-DROCM_FOUND=ON" # hipblaslt tries to download rocm-cmake if this isn't set
"-DBLIS_ROOT=${amd-blis}"
"-DBLIS_LIB=${amd-blis}/lib/libblis-mt.so"
"-DBLIS_INCLUDE_DIR=${amd-blis}/include/blis/"
"-DBLA_PREFER_PKGCONFIG=ON"
"-DFETCHCONTENT_SOURCE_DIR_NANOBIND=${python3Packages.nanobind.src}"
# 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"
];
postInstall =
# Compress msgpack .dat files to stay under hydra output size limit
# Relies on messagepack-compression-support.patch
''
for file in $out/lib/hipblaslt/library/*.dat; do
zstd -19 --long -f "$file" -o "$file.tmp" && mv "$file.tmp" "$file"
done
''
# Move binaries to appropriate outputs and delete leftover /bin
+ ''
mkdir -p $benchmark/bin
mv $out/bin/hipblaslt-{api-overhead,sequence,bench*} $out/bin/*.yaml $out/bin/*.py $benchmark/bin
${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 ''
''}
${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 = "Library that provides general matrix-matrix operations with a flexible API";
homepage = "https://github.com/ROCm/hipBLASlt";
license = with licenses; [ mit ];
teams = [ teams.rocm ];
platforms = platforms.linux;
};
}
)
''}
rmdir $out/bin
'';
# If this is false there are no kernels in the output lib
# supporting the target device
# 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;
};
meta = with lib; {
description = "Library that provides general matrix-matrix operations with a flexible API";
homepage = "https://github.com/ROCm/hipBLASlt";
license = with licenses; [ mit ];
teams = [ teams.rocm ];
platforms = platforms.linux;
};
})
@@ -0,0 +1,56 @@
diff --git a/Tensile/Source/lib/source/msgpack/MessagePack.cpp b/Tensile/Source/lib/source/msgpack/MessagePack.cpp
index de97929c..dbc397e0 100644
--- a/tensilelite/src/msgpack/MessagePack.cpp
+++ b/tensilelite/src/msgpack/MessagePack.cpp
@@ -28,6 +28,8 @@
#include <Tensile/msgpack/Loading.hpp>
+#include <zstd.h>
+
#include <fstream>
namespace Tensile
@@ -86,6 +88,34 @@ namespace Tensile
return nullptr;
}
+ // Check if the file is zstd compressed
+ char magic[4];
+ in.read(magic, 4);
+ bool isCompressed = (in.gcount() == 4 && magic[0] == '\x28' && magic[1] == '\xB5' && magic[2] == '\x2F' && magic[3] == '\xFD');
+ // Reset file pointer to the beginning
+ in.seekg(0, std::ios::beg);
+
+ if (isCompressed) {
+ // Decompress zstd file
+ std::vector<char> compressedData((std::istreambuf_iterator<char>(in)), std::istreambuf_iterator<char>());
+
+ size_t decompressedSize = ZSTD_getFrameContentSize(compressedData.data(), compressedData.size());
+ if (decompressedSize == ZSTD_CONTENTSIZE_ERROR || decompressedSize == ZSTD_CONTENTSIZE_UNKNOWN) {
+ if(Debug::Instance().printDataInit())
+ std::cout << "Error: Unable to determine decompressed size for " << filename << std::endl;
+ return false;
+ }
+
+ std::vector<char> decompressedData(decompressedSize);
+ size_t dSize = ZSTD_decompress(decompressedData.data(), decompressedSize, compressedData.data(), compressedData.size());
+ if (ZSTD_isError(dSize)) {
+ if(Debug::Instance().printDataInit())
+ std::cout << "Error: ZSTD decompression failed for " << filename << std::endl;
+ return false;
+ }
+
+ msgpack::unpack(result, decompressedData.data(), dSize);
+ } else {
msgpack::unpacker unp;
bool finished_parsing;
constexpr size_t buffer_size = 1 << 19;
@@ -109,6 +139,7 @@ namespace Tensile
return nullptr;
}
+ }
}
catch(std::runtime_error const& exc)
{
@@ -0,0 +1,12 @@
diff --git a/tensilelite/Tensile/Toolchain/Source.py b/tensilelite/Tensile/Toolchain/Source.py
index c9862e6c..dfa5ba40 100644
--- a/tensilelite/Tensile/Toolchain/Source.py
+++ b/tensilelite/Tensile/Toolchain/Source.py
@@ -102,6 +102,7 @@ def buildSourceCodeObjectFiles(
coPaths= []
objPath = str(tmpObjDir / objFilename)
+ compiler.default_args += ["-parallel-jobs=8"]
compiler(str(includeDir), cmdlineArchs, str(kernelPath), objPath)
for target in bundler.targets(objPath):
@@ -17,7 +17,7 @@
# CUB can also be used as a backend instead of rocPRIM.
stdenv.mkDerivation (finalAttrs: {
pname = "hipcub";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -33,7 +33,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "hipCUB";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-uECOQWG9C64tg5YZdm9/3+fZXaZVGslu8vElK3m23GY=";
hash = "sha256-yRJxhYFZYiKNR2xrn5fif/+vjHKKcKdn0JKPi972g+0=";
};
nativeBuildInputs = [
@@ -22,7 +22,7 @@
# Can also use cuFFT
stdenv.mkDerivation (finalAttrs: {
pname = "hipfft";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -41,7 +41,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "hipFFT";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-Jq/YHEtOo7a0/Ki7gxZATKmSqPU6cyLf5gx3A4MAZNw=";
hash = "sha256-4W93OOKTqNteoQ4GKycr06cjvGy5NF7RR08F+rfn+0o=";
fetchSubmodules = true;
};
@@ -10,13 +10,13 @@
stdenv.mkDerivation (finalAttrs: {
pname = "hipfort";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "hipfort";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-V5XDNM0bYHKnpkcnaDyxIS1zwsgaByJj+znFxJ6VxR0=";
hash = "sha256-Nks1+0X8bLtZ9HqZXJOtrAWQlJquMH+feuu1stf/9Vo=";
};
nativeBuildInputs = [
@@ -14,13 +14,13 @@
stdenv.mkDerivation (finalAttrs: {
pname = "hipify";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "HIPIFY";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-o/1LNsNtAyQcSug1gf7ujGNRRbvC33kwldrJKZi2LA0=";
hash = "sha256-uj25WmGCpwouS1yzW9Oil5Vyrbyj5yRITvWF9WaGozM=";
};
nativeBuildInputs = [
@@ -14,7 +14,7 @@
stdenv.mkDerivation (finalAttrs: {
pname = "hiprand";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -27,7 +27,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "hipRAND";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-TVc+qFwRiS5tAo1OKI1Wu5hadlwPZmSVZ9SvVvH1w7Y=";
hash = "sha256-f/AWDV7vkjt8BnjhLwT2m1Y6dlZy054+z6J0UW9Glg8=";
};
nativeBuildInputs = [
@@ -21,7 +21,7 @@
# Can also use cuSOLVER
stdenv.mkDerivation (finalAttrs: {
pname = "hipsolver";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -40,7 +40,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "hipSOLVER";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-ZQUKU3L4DgZ5zM7pCYEix0ulRkl78x/5wJnyCndTAwk=";
hash = "sha256-YP88fSM27Do0/tZ676Pvv2tr+lnlQa4vr3UnVNaVpLA=";
};
nativeBuildInputs = [
@@ -20,7 +20,7 @@
# This can also use cuSPARSE as a backend instead of rocSPARSE
stdenv.mkDerivation (finalAttrs: {
pname = "hipsparse";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -36,7 +36,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "hipSPARSE";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-3a7fKpYyiqG3aGOg7YrTHmKoH4rgTVLD16DvrZ3YY1g=";
hash = "sha256-fbh9fKlzxuIBTeCV/bEQbUS3lO6O3KoGF7/tTqRaCpE=";
};
nativeBuildInputs = [
@@ -1,53 +0,0 @@
{
lib,
stdenv,
fetchurl,
callPackage,
dpkg,
rocm-core,
}:
stdenv.mkDerivation (finalAttrs: {
pname = "hsa-amd-aqlprofile-bin";
version = "6.3.3";
src =
let
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/${poolVersion}/pool/main/h/hsa-amd-aqlprofile/hsa-amd-aqlprofile_1.0.0.${patch}-${incremental}~${osRelease}_amd64.deb";
hash = "sha256-Lo6gU9ywkujtsKvnOAwL3L8qQNPwjjm0Pm4OyzoUYao=";
};
nativeBuildInputs = [ dpkg ];
dontPatch = true;
dontConfigure = true;
dontBuild = true;
installPhase = ''
runHook preInstall
mkdir -p $out
cp -a opt/rocm-${finalAttrs.version}*/* $out
chmod +x $out/lib/libhsa-amd-aqlprofile64.so.1.*
chmod +x $out/lib/hsa-amd-aqlprofile/librocprofv2_att.so
runHook postInstall
'';
passthru.updateScript = (callPackage ./update.nix { }) { inherit (finalAttrs) version; };
meta = with lib; {
description = "AQLPROFILE library for AMD HSA runtime API extension support";
homepage = "https://rocm.docs.amd.com/en/latest/";
license = with licenses; [ unfree ];
teams = [ teams.rocm ];
platforms = platforms.linux;
};
})
@@ -1,61 +0,0 @@
{
lib,
writeScript,
}:
{ version }:
let
prefix = "hsa-amd-aqlprofile";
extVersion = lib.strings.concatStrings (
lib.strings.intersperse "0" (lib.versions.splitVersion version)
);
major = lib.versions.major version;
minor = lib.versions.minor version;
patch = lib.versions.patch version;
updateScript = writeScript "update.sh" ''
#!/usr/bin/env nix-shell
#!nix-shell -i bash -p curl common-updater-scripts
apt="https://repo.radeon.com/rocm/apt"
pool="pool/main/h/${prefix}/"
url="$apt/latest/$pool"
res="$(curl -sL "$url")"
deb="${prefix}$(echo "$res" | grep -o -P "(?<=href=\"${prefix}).*(?=\">)" | tail -1)"
patch="${patch}"
# Try up to 10 patch versions
for i in {1..10}; do
((patch++))
extVersion="$(echo "$deb" | grep -o -P "(?<=\.....).*(?=\..*-)")"
if (( ''${#extVersion} == 6 )) && (( $extVersion <= ${extVersion} )); then
url="https://repo.radeon.com/rocm/apt/${major}.${minor}.$patch/pool/main/h/${prefix}/"
res="$(curl -sL "$url")"
deb="${prefix}$(echo "$res" | grep -o -P "(?<=href=\"${prefix}).*(?=\">)" | tail -1)"
else
break
fi
done
extVersion="$(echo $deb | grep -o -P "(?<=\.....).*(?=\..*-)")"
version="$(echo $extVersion | sed "s/0/./1" | sed "s/0/./1")"
IFS='.' read -a version_arr <<< "$version"
if (( ''${version_arr[0]} > 6 )); then
echo "'rocmPackages_6.${prefix}-bin' is already at it's maximum allowed version.''\nAny further upgrades should go into 'rocmPackages_X.${prefix}-bin'." 1>&2
exit 1
fi
if (( ''${#extVersion} == 6 )); then
repoVersion="$version"
if (( ''${version:4:1} == 0 )); then
repoVersion=''${version:0:3}
fi
update-source-version rocmPackages_6.${prefix}-bin "$version" "" "$apt/$repoVersion/$pool$deb" --ignore-same-hash
fi
'';
in
[ updateScript ]
+220 -207
View File
@@ -1,16 +1,17 @@
{
lib,
stdenv,
llvmPackages_18,
# default LLVM version is used as stdenv to build our toolchain
llvmPackages,
# LLVM version closest to ROCm fork to override
llvmPackages_20,
overrideCC,
rocm-device-libs,
rocm-runtime,
fetchFromGitHub,
runCommand,
symlinkJoin,
rdfind,
wrapBintoolsWith,
emptyDirectory,
zstd,
zlib,
gcc-unwrapped,
@@ -21,33 +22,49 @@
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
# leaving compressed line tables (-g1 -gz) unstripped
# TODO: Should also apply to downstream packages which use rocmClangStdenv?
profilableStdenv ? false,
}:
# Whether to use LTO when building the ROCm toolchain
# Slows down this toolchain's build, for typical ROCm usecase
# time saved building composable_kernel and other heavy packages
# will outweight that. ~3-4% speedup multiplied by thousands
# of corehours.
withLto ? true,
# whether rocm stdenv uses libcxx (clang c++ stdlib) instead of gcc stdlibc++
withLibcxx ? false,
}@args:
let
llvmPackagesNoBintools = llvmPackages_18.override {
version = "6.4.3";
# major version of this should be the clang version ROCm forked from
rocmLlvmVersion = "19.0.0-rocm";
# llvmPackages_base version should be close to rocmLlvmVersion,
# may be one off because AMD backports a lot and the +1 patches
# may be easier to get to build
llvmPackages_base = llvmPackages_20;
llvmPackagesNoBintools = llvmPackages.override {
bootBintools = null;
bootBintoolsNoLibc = null;
};
useLibcxx = false; # whether rocm stdenv uses libcxx (clang c++ stdlib) instead of gcc stdlibc++
llvmStdenv = overrideCC llvmPackagesNoBintools.libcxxStdenv llvmPackagesNoBintools.clangUseLLVM;
llvmLibstdcxxStdenv = overrideCC llvmPackagesNoBintools.stdenv (
llvmPackagesNoBintools.libstdcxxClang.override {
inherit (llvmPackages_18) bintools;
}
);
stdenvToBuildRocmLlvm = if useLibcxx then llvmStdenv else llvmLibstdcxxStdenv;
stdenvToBuildRocmLlvm =
if withLibcxx then
overrideCC llvmPackagesNoBintools.libcxxStdenv llvmPackagesNoBintools.clangUseLLVM
else
# oddly fuse-ld=lld fails without this override
overrideCC llvmPackagesNoBintools.stdenv (
llvmPackagesNoBintools.libstdcxxClang.override {
inherit (llvmPackages) bintools;
}
);
gcc-include = runCommand "gcc-include" { } ''
mkdir -p $out
ln -s ${gcc-unwrapped}/include/ $out/
ln -s ${gcc-unwrapped}/lib/ $out/
'';
# A prefix for use as the GCC prefix when building rocmcxx
disallowedRefsForToolchain = [
stdenv.cc
stdenv.cc.cc
@@ -57,40 +74,47 @@ let
stdenvToBuildRocmLlvm.cc
stdenvToBuildRocmLlvm.cc.cc
];
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
# A prefix for use as the GCC prefix when building rocmcxx
gcc-prefix-headers = symlinkJoin {
name = "gcc-prefix-headers";
paths = [
glibc.dev
gcc-unwrapped.out
];
disallowedRequisites = [
glibc.dev
gcc-unwrapped.out
];
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}";
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/
'';
};
gcc-prefix = symlinkJoin {
name = "gcc-prefix";
paths = [
gcc-prefix-headers
glibc
gcc-unwrapped.lib
];
disallowedRequisites = [
glibc.dev
gcc-unwrapped.out
];
postBuild = ''
rm -rf $out/{bin,libexec,nix-support,lib64,share,etc}
rm $out/lib/ld-linux-x86-64.so.2
ln -s $out $out/x86_64-unknown-linux-gnu
'';
};
usefulOutputs =
drv:
builtins.filter (x: x != null) [
@@ -100,27 +124,24 @@ let
];
listUsefulOutputs = builtins.concatMap usefulOutputs;
llvmSrc = fetchFromGitHub {
# Performance improvements cherry-picked on top of rocm-6.3.x
# most importantly, amdgpu-early-alwaysinline memory usage fix
owner = "LunNova";
repo = "llvm-project-rocm";
rev = "4182046534deb851753f0d962146e5176f648893";
hash = "sha256-sPmYi1WiiAqnRnHVNba2nPUxGflBC01FWCTNLPlYF9c=";
owner = "ROCm";
repo = "llvm-project";
rev = "rocm-${version}";
hash = "sha256-12ftH5fMPAsbcEBmhADwW1YY/Yxo/MAK1FafKczITMg=";
};
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;
llvmPackagesRocm = llvmPackages_base.override (_old: {
stdenv = stdenvToBuildRocmLlvm;
# not setting gitRelease = because that causes patch selection logic to use git patches
# ROCm LLVM is closer to 18 official
# ROCm LLVM is closer to 20 official
# gitRelease = {}; officialRelease = null;
officialRelease = { }; # Set but empty because we're overriding everything from it.
version = rocmLlvmVersion;
src = llvmSrcFixed;
monorepoSrc = llvmSrcFixed;
src = llvmSrc;
monorepoSrc = llvmSrc;
doCheck = false;
});
sysrootCompiler =
@@ -158,82 +179,112 @@ let
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 $out/ # create links *within* the sysroot to save space
'';
findClangNostdlibincPatch =
# Removes patches which either aren't desired, or don't apply against ROCm LLVM
removeInapplicablePatches =
x:
(
(lib.strings.hasSuffix "add-nostdlibinc-flag.patch" (builtins.baseNameOf x))
|| (lib.strings.hasSuffix "clang-at-least-16-LLVMgold-path.patch" (builtins.baseNameOf x))
);
tablegenUsage = x: !(lib.strings.hasInfix "llvm-tblgen" x);
llvmTargetsFlag = "-DLLVM_TARGETS_TO_BUILD=AMDGPU;${
{
"x86_64" = "X86";
"aarch64" = "AArch64";
}
.${llvmStdenv.targetPlatform.parsed.cpu.name}
or (throw "Unsupported CPU architecture: ${llvmStdenv.targetPlatform.parsed.cpu.name}")
.${stdenv.targetPlatform.parsed.cpu.name}
or (throw "Unsupported CPU architecture: ${stdenv.targetPlatform.parsed.cpu.name}")
}";
# -ffat-lto-objects = emit LTO object files that are compatible with non-LTO-supporting builds too
# FatLTO objects are a special type of fat object file that contain LTO compatible IR in addition to generated object code,
# instead of containing object code for multiple target architectures. This allows users to defer the choice of whether to
# use LTO or not to link-time, and has been a feature available in other compilers, like GCC, for some time.
llvmMeta = {
# TODO(@LunNova): it would be nice to support aarch64 for rocmPackages
platforms = [ "x86_64-linux" ];
};
# TODO(@LunNova): Some of this might be worth supporting in llvmPackages, dropping from here
commonCmakeFlags = [
llvmTargetsFlag
# Compression support is required for compressed offload kernels
# Set FORCE_ON so that failure to find the compression libs will be a build error
(lib.cmakeFeature "LLVM_ENABLE_ZSTD" "FORCE_ON")
(lib.cmakeFeature "LLVM_ENABLE_ZLIB" "FORCE_ON")
# required for threaded ThinLTO to work
(lib.cmakeBool "LLVM_ENABLE_THREADS" true)
# LLVM tries to call git to embed VCS info if FORCE_VC_ aren't set
(lib.cmakeFeature "LLVM_FORCE_VC_REVISION" "rocm-${version}")
(lib.cmakeFeature "LLVM_FORCE_VC_REPOSITORY" "https://github.com/ROCm/llvm-project")
(lib.cmakeFeature "LLVM_VERSION_SUFFIX" "")
(lib.cmakeBool "LLVM_ENABLE_LIBCXX" withLibcxx)
(lib.cmakeFeature "CLANG_DEFAULT_CXX_STDLIB" (if withLibcxx then "libc++" else "libstdc++"))
(lib.cmakeFeature "CLANG_VENDOR" "nixpkgs-AMD")
(lib.cmakeFeature "CLANG_REPOSITORY_STRING" "https://github.com/ROCm/llvm-project/tree/rocm-${version}")
]
++ lib.optionals withLibcxx [
(lib.cmakeFeature "CLANG_DEFAULT_RTLIB" "compiler-rt")
]
++ lib.optionals withLto [
(lib.cmakeBool "CMAKE_INTERPROCEDURAL_OPTIMIZATION" true)
(lib.cmakeBool "LLVM_ENABLE_FATLTO" false)
]
++ lib.optionals (withLto && stdenvToBuildRocmLlvm.cc.isClang) [
(lib.cmakeFeature "LLVM_ENABLE_LTO" "FULL")
(lib.cmakeFeature "LLVM_USE_LINKER" "lld")
];
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");
llvmExtraCflags = lib.concatStringsSep " " (
lib.optionals (stdenv.hostPlatform.system == "x86_64-linux") [
# Unprincipled decision to build x86_64 ROCm clang for at least skylake and tune for zen3+
# In practice building the ROCm package set with anything earlier than zen3 is annoying
# and earlier than skylake is implausible due to too few cores and too little RAM
# Speeds up composable_kernel builds by ~4%
# If this causes trouble in practice we can drop this. Set since 2025-03-24.
"-march=skylake"
"-mtune=znver3"
]
++ lib.optionals profilableStdenv [
# compressed line only debug info for profiling
"-gz"
"-g1"
]
);
in
rec {
inherit (llvmPackagesRocm) libunwind;
inherit (llvmPackagesRocm) libcxx;
inherit args;
# 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: {
llvm = llvmPackagesRocm.llvm.overrideAttrs (old: {
patches = old.patches ++ [
(fetchpatch {
# fix compile error in tools/gold/gold-plugin.cpp
name = "gold-plugin-fix.patch";
url = "https://github.com/llvm/llvm-project/commit/b0baa1d8bd68a2ce2f7c5f2b62333e410e9122a1.patch";
hash = "sha256-yly93PvGIXOnFeDGZ2W+W6SyhdWFM6iwA+qOeaptrh0=";
relative = "llvm";
})
(fetchpatch {
# fix tools/llvm-exegesis/X86/latency/ failing with glibc 2.4+
name = "exegesis-latency-glibc-fix.patch";
sha256 = "sha256-CjKxQlYwHXTM0mVnv8k/ssg5OXuKpJxRvBZGXjrFZAg=";
url = "https://github.com/llvm/llvm-project/commit/1e8df9e85a1ff213e5868bd822877695f27504ad.patch";
relative = "llvm";
})
./perf-increase-namestring-size.patch
# TODO: consider reapplying "Don't include aliases in RegisterClassInfo::IgnoreCSRForAllocOrder"
# it was reverted as it's a pessimization for non-GPU archs, but this compiler
# is used mostly for amdgpu
];
dontStrip = profilableStdenv;
hardeningDisable = [ "all" ];
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=${llvmExtraCflags}'
)
'';
env = (old.env or { }) // {
NIX_CFLAGS_COMPILE = "${(old.env or { }).NIX_CFLAGS_COMPILE or ""} ${llvmExtraCflags}";
};
cmakeFlags = (builtins.filter tablegenUsage old.cmakeFlags) ++ commonCmakeFlags;
# Ensure we don't leak refs to compiler that was used to bootstrap this LLVM
disallowedReferences = (old.disallowedReferences or [ ]) ++ disallowedRefsForToolchain;
postFixup = ''
@@ -243,48 +294,26 @@ rec {
find $lib -type f -exec remove-references-to -t ${stdenvToBuildRocmLlvm.cc} {} +
find $lib -type f -exec remove-references-to -t ${stdenv.cc.bintools} {} +
'';
meta = old.meta // llvmMeta;
});
lld =
(llvmPackagesRocm.lld.override {
libllvm = llvm;
ninja = emptyDirectory;
}).overrideAttrs
(old: {
patches = builtins.filter (
x: !(lib.strings.hasSuffix "more-openbsd-program-headers.patch" (builtins.baseNameOf x))
) old.patches;
dontStrip = profilableStdenv;
hardeningDisable = [ "all" ];
nativeBuildInputs = old.nativeBuildInputs ++ [
llvmPackagesNoBintools.lld
removeReferencesTo
];
buildInputs = old.buildInputs ++ [
zstd
zlib
];
env.NIX_BUILD_ID_STYLE = "fast";
LDFLAGS = "-Wl,--build-id=sha1,--icf=all,--compress-debug-sections=zlib";
cmakeFlags =
(builtins.filter tablegenUsage old.cmakeFlags)
++ [
llvmTargetsFlag
"-DCMAKE_BUILD_TYPE=Release"
"-DLLVM_ENABLE_ZSTD=FORCE_ON"
"-DLLVM_ENABLE_ZLIB=FORCE_ON"
"-DLLVM_ENABLE_THREADS=ON"
"-DLLVM_ENABLE_LTO=Thin"
"-DLLVM_USE_LINKER=lld"
(lib.cmakeBool "LLVM_ENABLE_LIBCXX" useLibcxx)
"-DCLANG_DEFAULT_CXX_STDLIB=${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=${gcc-unwrapped}/bin/gcc-nm"
]
++ lib.optionals useLibcxx [
"-DLLVM_ENABLE_LIBCXX=ON"
];
env = (old.env or { }) // {
NIX_CFLAGS_COMPILE = "${(old.env or { }).NIX_CFLAGS_COMPILE or ""} ${llvmExtraCflags}";
};
cmakeFlags = (builtins.filter tablegenUsage old.cmakeFlags) ++ commonCmakeFlags;
# Ensure we don't leak refs to compiler that was used to bootstrap this LLVM
disallowedReferences = (old.disallowedReferences or [ ]) ++ disallowedRefsForToolchain;
postFixup = ''
@@ -292,36 +321,40 @@ rec {
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=${llvmExtraCflags}'
)
'';
meta = old.meta // llvmMeta;
});
clang-unwrapped =
(
(llvmPackagesRocm.clang-unwrapped.override {
libllvm = llvm;
ninja = emptyDirectory;
}).overrideAttrs
(
old:
let
filteredPatches = builtins.filter (x: !(findClangNostdlibincPatch x)) old.patches;
filteredPatches = builtins.filter (x: !(removeInapplicablePatches x)) old.patches;
in
{
meta.platforms = [
"x86_64-linux"
];
pname = "${old.pname}-rocm";
patches = filteredPatches ++ [
passthru = old.passthru // {
inherit gcc-prefix;
};
patches = [
(fetchpatch {
# [PATCH] [clang] Install scan-build-py into plain "lib" directory
# Backported so 19/clang/gnu-install-dirs patch applies to AMD's LLVM fork
hash = "sha256-bOqAjBwRKcERpQkiBpuojGs6ddd5Ht3zL5l3TuJK2w8=";
url = "https://github.com/llvm/llvm-project/commit/816fde1cbb700ebcc8b3df81fb93d675c04c12cd.patch";
relative = "clang";
})
]
++ 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
./opt-offload-compress-on-by-default.patch
./perf-shorten-gcclib-include-paths.patch
(fetchpatch {
# [ClangOffloadBundler]: Add GetBundleIDsInFile to OffloadBundler
sha256 = "sha256-G/mzUdFfrJ2bLJgo4+mBcR6Ox7xGhWu5X+XxT4kH2c8=";
@@ -334,65 +367,41 @@ rec {
libllvmLibdir = "${llvm.lib}/lib";
})
];
hardeningDisable = [ "all" ];
nativeBuildInputs = old.nativeBuildInputs ++ [
llvmPackagesNoBintools.lld
removeReferencesTo
];
buildInputs = old.buildInputs ++ [
zstd
zlib
];
dontStrip = profilableStdenv;
LDFLAGS = "-Wl,--build-id=sha1,--icf=all,--compress-debug-sections=zlib";
env = (old.env or { }) // {
NIX_BUILD_ID_STYLE = "fast";
NIX_CFLAGS_COMPILE = "${(old.env or { }).NIX_CFLAGS_COMPILE or ""} ${llvmExtraCflags}";
};
dontStrip = profilableStdenv;
# Ensure we don't leak refs to compiler that was used to bootstrap this LLVM
disallowedReferences = (old.disallowedReferences or [ ]) ++ disallowedRefsForToolchain;
# Enable structured attrs for separateDebugInfo, because it is required with disallowedReferences set
__structuredAttrs = true;
requiredSystemFeatures = (old.requiredSystemFeatures or [ ]) ++ [ "big-parallel" ];
# https://github.com/llvm/llvm-project/blob/6976deebafa8e7de993ce159aa6b82c0e7089313/clang/cmake/caches/DistributionExample-stage2.cmake#L9-L11
cmakeFlags =
(builtins.filter tablegenUsage old.cmakeFlags)
++ [
llvmTargetsFlag
"-DCMAKE_BUILD_TYPE=Release"
"-DLLVM_ENABLE_ZSTD=FORCE_ON"
"-DLLVM_ENABLE_ZLIB=FORCE_ON"
"-DLLVM_ENABLE_THREADS=ON"
"-DLLVM_ENABLE_LTO=Thin"
"-DLLVM_USE_LINKER=lld"
(lib.cmakeBool "LLVM_ENABLE_LIBCXX" useLibcxx)
"-DCLANG_DEFAULT_CXX_STDLIB=${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=${gcc-unwrapped}/bin/gcc-nm"
]
++ lib.optionals useLibcxx [
"-DLLVM_ENABLE_LTO=Thin"
"-DLLVM_ENABLE_LIBCXX=ON"
"-DLLVM_USE_LINKER=lld"
"-DCLANG_DEFAULT_RTLIB=compiler-rt"
]
++ lib.optionals (!useLibcxx) [
++ commonCmakeFlags
++ lib.optionals (!withLibcxx) [
# FIXME: Config file in rocmcxx instead of GCC_INSTALL_PREFIX?
# Expected to be fully removed eventually
"-DUSE_DEPRECATED_GCC_INSTALL_PREFIX=ON"
"-DGCC_INSTALL_PREFIX=${gcc-prefix}"
];
postFixup = (old.postFixup or "") + ''
find $lib -type f -exec remove-references-to -t ${stdenvToBuildRocmLlvm} {} +
find $lib -type f -exec remove-references-to -t ${stdenvToBuildRocmLlvm.cc} {} +
find $lib -type f -exec remove-references-to -t ${stdenvToBuildRocmLlvm.cc.cc} {} +
find $lib -type f -exec remove-references-to -t ${stdenv.cc} {} +
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=${llvmExtraCflags}'
)
'';
meta = old.meta // llvmMeta;
}
)
)
@@ -409,11 +418,12 @@ rec {
clang-unwrapped
bintools
compiler-rt
openmp
]
++ (lib.optionals useLibcxx [
++ (lib.optionals withLibcxx [
libcxx
])
++ (lib.optionals (!useLibcxx) [
++ (lib.optionals (!withLibcxx) [
gcc-include
glibc
glibc.dev
@@ -438,7 +448,14 @@ rec {
hash = "sha256-Db+L1HFMWVj4CrofsGbn5lnMoCzEcU+7q12KKFb17/g=";
relative = "compiler-rt";
})
(fetchpatch {
# Fixes fortify hardening compile error related to openat usage
hash = "sha256-pgpN1q1vIQrPXHPxNSZ6zfgV2EflHO5Amzl+2BDjXbs=";
url = "https://github.com/llvm/llvm-project/commit/155b7a12820ec45095988b6aa6e057afaf2bc892.patch";
relative = "compiler-rt";
})
];
meta = old.meta // llvmMeta;
});
compiler-rt = compiler-rt-libc;
bintools = wrapBintoolsWith {
@@ -450,6 +467,7 @@ rec {
clang = rocmcxx;
# Emulate a monolithic ROCm LLVM build to support building ROCm's in-tree LLVM projects
# TODO(@LunNova): destroy this
rocm-merged-llvm = symlinkJoin {
name = "rocm-llvm-merge";
paths = [
@@ -464,7 +482,7 @@ rec {
compiler-rt.dev
rocmcxx
]
++ lib.optionals useLibcxx [
++ lib.optionals withLibcxx [
libcxx
libcxx.out
libcxx.dev
@@ -489,37 +507,32 @@ rec {
};
rocmClangStdenv = overrideCC (
if useLibcxx then llvmPackagesRocm.libcxxStdenv else llvmPackagesRocm.stdenv
if withLibcxx then llvmPackagesRocm.libcxxStdenv else llvmPackagesRocm.stdenv
) clang;
# Projects
openmp =
(llvmPackagesRocm.openmp.override {
stdenv = rocmClangStdenv;
llvm = rocm-merged-llvm;
targetLlvm = rocm-merged-llvm;
clang-unwrapped = clang;
llvm = llvm;
targetLlvm = llvm;
clang-unwrapped = clang-unwrapped;
}).overrideAttrs
(old: {
disallowedReferences = (old.disallowedReferences or [ ]) ++ disallowedRefsForToolchain;
nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [ removeReferencesTo ];
nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
removeReferencesTo
];
cmakeFlags =
old.cmakeFlags
++ commonCmakeFlags
++ [
"-DDEVICELIBS_ROOT=${rocm-device-libs.src}"
# OMPD support is broken in ROCm 6.3. Haven't investigated why.
# 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
clang-unwrapped
zlib
zstd
libxml2
@@ -0,0 +1,22 @@
From 762511ad9c9cac7769a795cc4df888a1559dca02 Mon Sep 17 00:00:00 2001
From: Luna Nova <git@lunnova.dev>
Date: Mon, 16 Dec 2024 14:56:57 -0800
Subject: [PATCH] [AMDGPU] CommonArgs: Enable offload compression by default
---
clang/lib/Driver/ToolChains/CommonArgs.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/clang/lib/Driver/ToolChains/CommonArgs.cpp b/clang/lib/Driver/ToolChains/CommonArgs.cpp
index 33f12fb2d075d..ba35f65210c7a 100644
--- a/lib/Driver/ToolChains/CommonArgs.cpp
+++ b/lib/Driver/ToolChains/CommonArgs.cpp
@@ -2914,7 +2914,7 @@ void tools::addHIPRuntimeLibArgs(const ToolChain &TC, Compilation &C,
void tools::addOffloadCompressArgs(const llvm::opt::ArgList &TCArgs,
llvm::opt::ArgStringList &CmdArgs) {
if (TCArgs.hasFlag(options::OPT_offload_compress,
- options::OPT_no_offload_compress, false))
+ options::OPT_no_offload_compress, true))
CmdArgs.push_back("-compress");
if (TCArgs.hasArg(options::OPT_v))
CmdArgs.push_back("-verbose");
@@ -0,0 +1,28 @@
From a57eb548d46a37e4aed33eefccc9e6374b579ff9 Mon Sep 17 00:00:00 2001
From: Luna Nova <git@lunnova.dev>
Date: Mon, 16 Dec 2024 13:58:54 -0800
Subject: [PATCH] [AMDGPU] TargetMachine: increase NameString size to 256
128 is too small a default for compilers used to build
AMD's rocBLAS, composable_kernel and similar libs because
they use very long symbol names such as
device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_default_instance
This fixes spending ~10% of compile time in malloc inside this function.
---
llvm/lib/Target/TargetMachine.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/lib/Target/TargetMachine.cpp b/llvm/lib/Target/TargetMachine.cpp
index abd0fdf2390c0..86ce468b75643 100644
--- a/lib/Target/TargetMachine.cpp
+++ b/lib/Target/TargetMachine.cpp
@@ -270,7 +270,7 @@ MCSymbol *TargetMachine::getSymbol(const GlobalValue *GV) const {
if (MCSymbol *TargetSymbol = TLOF->getTargetSymbol(GV, *this))
return TargetSymbol;
- SmallString<128> NameStr;
+ SmallString<256> NameStr;
getNameWithPrefix(NameStr, GV, TLOF->getMangler());
return TLOF->getContext().getOrCreateSymbol(NameStr);
}
@@ -0,0 +1,57 @@
From ef6c5b353861be727c98f1319c81d0c6b609d644 Mon Sep 17 00:00:00 2001
From: Luna Nova <git@lunnova.dev>
Date: Tue, 17 Dec 2024 04:29:11 -0800
Subject: [PATCH] HACK: Get canonical GCC include path so doesn't have
../../../../
This allows more of the strings used in compilation to fit inside
fixed size stack allocated buffers instead of spilling into the heap
---
clang/lib/Driver/ToolChains/Gnu.cpp | 16 ++++++++++------
1 file changed, 10 insertions(+), 6 deletions(-)
diff --git a/clang/lib/Driver/ToolChains/Gnu.cpp b/clang/lib/Driver/ToolChains/Gnu.cpp
index af9fd46f0ce7b..a63a7a93f6a78 100644
--- a/lib/Driver/ToolChains/Gnu.cpp
+++ b/lib/Driver/ToolChains/Gnu.cpp
@@ -3394,29 +3394,33 @@ bool Generic_GCC::addLibStdCXXIncludePaths(Twine IncludeDir, StringRef Triple,
if (!getVFS().exists(IncludeDir))
return false;
+ SmallString<260> CanonicalIncludeDir;
+ if (getVFS().getRealPath(IncludeDir, CanonicalIncludeDir))
+ return false;
+
// Debian native gcc uses g++-multiarch-incdir.diff which uses
// include/x86_64-linux-gnu/c++/10$IncludeSuffix instead of
// include/c++/10/x86_64-linux-gnu$IncludeSuffix.
- std::string Dir = IncludeDir.str();
StringRef Include =
- llvm::sys::path::parent_path(llvm::sys::path::parent_path(Dir));
+ llvm::sys::path::parent_path(llvm::sys::path::parent_path(CanonicalIncludeDir));
std::string Path =
- (Include + "/" + Triple + Dir.substr(Include.size()) + IncludeSuffix)
+ (Include + "/" + Triple + CanonicalIncludeDir.substr(Include.size()) + IncludeSuffix)
.str();
if (DetectDebian && !getVFS().exists(Path))
return false;
// GPLUSPLUS_INCLUDE_DIR
- addSystemInclude(DriverArgs, CC1Args, IncludeDir);
+ addSystemInclude(DriverArgs, CC1Args, CanonicalIncludeDir);
// GPLUSPLUS_TOOL_INCLUDE_DIR. If Triple is not empty, add a target-dependent
// include directory.
if (DetectDebian)
addSystemInclude(DriverArgs, CC1Args, Path);
else if (!Triple.empty())
addSystemInclude(DriverArgs, CC1Args,
- IncludeDir + "/" + Triple + IncludeSuffix);
+ CanonicalIncludeDir + "/" + Triple + IncludeSuffix);
// GPLUSPLUS_BACKWARD_INCLUDE_DIR
- addSystemInclude(DriverArgs, CC1Args, IncludeDir + "/backward");
+ if (getVFS().exists(CanonicalIncludeDir + "/backward"))
+ addSystemInclude(DriverArgs, CC1Args, CanonicalIncludeDir + "/backward");
return true;
}
@@ -18,7 +18,8 @@
abseil-cpp,
half,
nlohmann_json,
msgpack,
boost,
msgpack-cxx,
sqlite,
oneDNN,
blaze,
@@ -28,6 +29,7 @@
docutils,
ghostscript,
python3Packages,
writableTmpDirAsHomeHook,
buildDocs ? false,
buildTests ? false,
gpuTargets ? clr.gpuTargets,
@@ -63,7 +65,7 @@ let
in
stdenv.mkDerivation (finalAttrs: {
pname = "migraphx";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -79,9 +81,13 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "AMDMIGraphX";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-h9cTbrMwHeRGVJS/uHQnCXplNcrBqxbhwz2AcAEso0M=";
hash = "sha256-8iOBoRBygTvn9eX5f9cG0kBHKgKSeflqHkV6Qwh/ruA=";
};
patches = [
./msgpack-6-compat.patch
];
nativeBuildInputs = [
pkg-config
cmake
@@ -97,6 +103,7 @@ stdenv.mkDerivation (finalAttrs: {
ghostscript
python3Packages.sphinx-rtd-theme
python3Packages.breathe
writableTmpDirAsHomeHook
];
buildInputs = [
@@ -110,7 +117,8 @@ stdenv.mkDerivation (finalAttrs: {
protobuf
half
nlohmann_json
msgpack
boost
msgpack-cxx
sqlite
oneDNN'
blaze
@@ -136,9 +144,6 @@ stdenv.mkDerivation (finalAttrs: {
# migraphxs relies on miopen which relies on current composable_kernel
# impossible to build with this ON; we can't link both of them even if we package both
"-DMIGRAPHX_USE_COMPOSABLEKERNEL=OFF"
"-DOpenMP_C_INCLUDE_DIR=${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"
@@ -166,7 +171,6 @@ stdenv.mkDerivation (finalAttrs: {
# Unfortunately, it seems like we have to call make on this manually
preInstall = lib.optionalString buildDocs ''
export HOME=$(mktemp -d)
make -j$NIX_BUILD_CORES doc
cd ../doc/pdf
make -j$NIX_BUILD_CORES
@@ -0,0 +1,29 @@
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 1dba2c8d..7795ba78 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -358,13 +358,19 @@ target_link_libraries(migraphx PRIVATE nlohmann_json::nlohmann_json)
find_package(SQLite3 REQUIRED)
target_link_libraries(migraphx PRIVATE SQLite::SQLite3)
-find_package(msgpackc-cxx QUIET)
-if(NOT msgpackc-cxx_FOUND)
- find_package(msgpack REQUIRED)
+# See: https://github.com/msgpack/msgpack-c/wiki/Q%26A#how-to-support-both-msgpack-c-c-version-5x-and-6x-
+# Prefer 6.x (msgpack-cxx)
+find_package(msgpack-cxx)
+if(msgpack-cxx_FOUND)
+ message(STATUS "Found msgpack-cxx (>=6.x)")
+else()
+ find_package(msgpackc-cxx REQUIRED NAMES msgpackc-cxx msgpack)
+ message(STATUS "Found msgpackc-cxx (<=5.x)")
+ add_library(msgpack-cxx ALIAS msgpackc-cxx)
endif()
-target_link_libraries(migraphx PRIVATE msgpackc-cxx)
+target_link_libraries(migraphx PRIVATE msgpack-cxx)
# Make this available to the tests
-target_link_libraries(migraphx INTERFACE $<BUILD_INTERFACE:msgpackc-cxx>)
+target_link_libraries(migraphx INTERFACE $<BUILD_INTERFACE:msgpack-cxx>)
add_library(migraphx_all_targets INTERFACE)
@@ -34,41 +34,50 @@
rocm-comgr,
roctracer,
python3Packages,
# FIXME: should be able to use all clr targets
gpuTargets ? [
gpuTargets ? clr.localGpuTargets or clr.gpuTargets,
buildDocs ? false, # Needs internet because of rocm-docs-core
buildTests ? false,
withComposableKernel ? composable_kernel.anyMfmaTarget,
}:
let
# FIXME: cmake files need patched to include this properly
cFlags = "-Wno-documentation-pedantic --offload-compress -I${hipblas-common}/include -I${hipblas}/include -I${roctracer}/include -I${nlohmann_json}/include -I${sqlite.dev}/include -I${rocrand}/include";
version = "6.4.3";
# Targets outside this list will get
# error: use of undeclared identifier 'CK_BUFFER_RESOURCE_3RD_DWORD'
supportedTargets = lib.intersectLists [
"gfx900"
"gfx906"
"gfx908"
"gfx90a"
"gfx942"
"gfx950"
"gfx1030"
"gfx1031"
"gfx1100"
"gfx1101"
"gfx1102"
], # clr.gpuTargets
buildDocs ? false, # Needs internet because of rocm-docs-core
buildTests ? false,
withComposableKernel ? composable_kernel.anyGfx9Target,
}:
let
# FIXME: cmake files need patched to include this properly
cFlags = "-O3 -DNDEBUG -Wno-documentation-pedantic --offload-compress -I${hipblas-common}/include -I${hipblas}/include -I${roctracer}/include -I${nlohmann_json}/include -I${sqlite.dev}/include -I${rocrand}/include";
version = "6.3.3";
"gfx1150"
"gfx1151"
"gfx1200"
"gfx1201"
] gpuTargets;
src = fetchFromGitHub {
owner = "ROCm";
repo = "MIOpen";
rev = "rocm-${version}";
hash = "sha256-rX+BE6wBDMnLyc6eai3bDVvmfahomDO0s10n6HhWu7c=";
hash = "sha256-DEcVj2vOwIYYyNKEKFqZ0fb9o+/QRpwiSksxwnmgEMc=";
fetchLFS = true;
fetchSubmodules = true;
# WORKAROUND: .lfsconfig is incorrectly set to exclude everything upstream
leaveDotGit = true;
# FIXME: if someone can reduce the level of awful here that would be really nice
postFetch = ''
export HOME=$(mktemp -d)
cd $out
set -x
git remote add origin $url
git fetch origin +refs/tags/rocm-${version}:refs/tags/rocm-${version}
git clean -fdx
@@ -77,8 +86,8 @@ let
rm .lfsconfig
git lfs install
git lfs track "*.kdb.bz2"
GIT_TRACE=1 git lfs fetch --include="src/kernels/**"
GIT_TRACE=1 git lfs pull --include="src/kernels/**"
git lfs fetch --include="src/kernels/**"
git lfs pull --include="src/kernels/**"
git lfs checkout
rm -rf .git
@@ -132,7 +141,6 @@ stdenv.mkDerivation (finalAttrs: {
# Find zstd and add to target. Mainly for torch.
patches = [
./skip-preexisting-dbs.patch
./fix-isnan.patch # https://github.com/ROCm/MIOpen/pull/3448
(fetchpatch {
url = "https://github.com/ROCm/MIOpen/commit/e608b4325646afeabb5e52846997b926d2019d19.patch";
hash = "sha256-oxa3qlIC2bzbwGxrQOZXoY/S7CpLsMrnWRB7Og0tk0M=";
@@ -205,9 +213,9 @@ stdenv.mkDerivation (finalAttrs: {
];
cmakeFlags = [
"-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
"-DGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
"-DGPU_ARCHS=${lib.concatStringsSep ";" gpuTargets}"
"-DAMDGPU_TARGETS=${lib.concatStringsSep ";" supportedTargets}"
"-DGPU_TARGETS=${lib.concatStringsSep ";" supportedTargets}"
"-DGPU_ARCHS=${lib.concatStringsSep ";" supportedTargets}"
"-DMIOPEN_USE_SQLITE_PERFDB=ON"
"-DCMAKE_VERBOSE_MAKEFILE=ON"
"-DCMAKE_MODULE_PATH=${clr}/hip/cmake"
@@ -271,7 +279,7 @@ stdenv.mkDerivation (finalAttrs: {
'';
postInstall = ''
rm $out/bin/install_precompiled_kernels.sh
rm $out/libexec/miopen/install_precompiled_kernels.sh
ln -sf ${gfx900} $out/share/miopen/db/gfx900.kdb
ln -sf ${gfx906} $out/share/miopen/db/gfx906.kdb
ln -sf ${gfx908} $out/share/miopen/db/gfx908.kdb
@@ -1,31 +0,0 @@
From 17f67e0aa31cd2f1c1cb012d3858abf6956acc72 Mon Sep 17 00:00:00 2001
From: "Sv. Lockal" <lockalsash@gmail.com>
Date: Tue, 24 Dec 2024 14:43:10 +0000
Subject: [PATCH] Fix missing isnan definition on libstdc++ >=14 systems
Closes #3441
---
driver/reducecalculation_driver.hpp | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/driver/reducecalculation_driver.hpp b/driver/reducecalculation_driver.hpp
index 8226b3c953..2001969509 100644
--- a/driver/reducecalculation_driver.hpp
+++ b/driver/reducecalculation_driver.hpp
@@ -33,6 +33,7 @@
#include "random.hpp"
#include <algorithm>
#include <cfloat>
+#include <cmath>
#include <cstdlib>
#include <memory>
#include <miopen/miopen.h>
@@ -77,7 +78,7 @@ int32_t mloReduceCalculationForwardRunHost(miopenTensorDescriptor_t inputDesc,
for(size_t i = 0; i < reduce_size; ++i)
{
Tcheck val = static_cast<Tcheck>(input[input_idx]);
- if(nanPropagation && isnan(val))
+ if(nanPropagation && std::isnan(val))
{
val = 0.0f;
}
@@ -5,7 +5,6 @@
rocmUpdateScript,
cmake,
rocm-cmake,
rocm-device-libs,
clr,
pkg-config,
rpp,
@@ -42,13 +41,13 @@ stdenv.mkDerivation (finalAttrs: {
"cpu"
);
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "MIVisionX";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-SisCbUDCAiWQ1Ue7qrtoT6vO/1ztzqji+3cJD6MXUNw=";
hash = "sha256-07MivgCYmKLnhGDjOYsFBfwIxEoQLYNoRbOo3MPpVzE=";
};
patches = [
@@ -58,19 +57,18 @@ stdenv.mkDerivation (finalAttrs: {
nativeBuildInputs = [
cmake
rocm-cmake
clr
pkg-config
]
++ lib.optionals (!useOpenCL && !useCPU) [
clr
]
++ lib.optionals buildDocs [
rocm-docs-core
python3Packages.python
];
buildInputs = [
miopen
migraphx
rpp
rocblas
openmp
half
protobuf
@@ -83,28 +81,31 @@ stdenv.mkDerivation (finalAttrs: {
rapidjson
python3Packages.pybind11
python3Packages.numpy
python3Packages.torchWithRocm
]
++ lib.optionals (!useOpenCL && !useCPU) [
miopen
rocblas
migraphx
];
cmakeFlags = [
"-DROCM_PATH=${clr}"
"-DAMDRPP_PATH=${rpp}"
# 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_INSTALL_PREFIX_PYTHON=lib"
"-DOpenMP_C_INCLUDE_DIR=${openmp.dev}/include"
"-DOpenMP_CXX_INCLUDE_DIR=${openmp.dev}/include"
"-DOpenMP_omp_LIBRARY=${openmp}/lib"
# "-DAMD_FP16_SUPPORT=ON" `error: typedef redefinition with different types ('__half' vs 'half_float::half')`
]
++ lib.optionals (gpuTargets != [ ]) [
"-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
]
++ lib.optionals (!useOpenCL && !useCPU) [
"-DROCM_PATH=${clr}"
"-DAMDRPP_PATH=${rpp}"
"-DBACKEND=HIP"
"-DCMAKE_C_COMPILER=hipcc"
"-DCMAKE_CXX_COMPILER=hipcc"
]
++ lib.optionals (useOpenCL && !useCPU) [
"-DBACKEND=OCL"
@@ -114,17 +115,17 @@ 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"
# Properly find miopen, fix ffmpeg version detection
substituteInPlace amd_openvx_extensions/CMakeLists.txt \
--replace-fail "miopen PATHS \''${ROCM_PATH} QUIET" "miopen PATHS ${miopen} QUIET" \
--replace-fail "\''${ROCM_PATH}/include/miopen/config.h" "${miopen}/include/miopen/config.h"
# Properly find turbojpeg
substituteInPlace cmake/FindTurboJpeg.cmake \
--replace-fail "\''${TURBO_JPEG_PATH}/include" "${libjpeg_turbo.dev}/include" \
--replace-fail "\''${TURBO_JPEG_PATH}/lib" "${libjpeg_turbo.out}/lib"
${lib.optionalString (!useOpenCL && !useCPU) ''
# Properly find miopen
substituteInPlace amd_openvx_extensions/CMakeLists.txt \
--replace-fail "miopen PATHS \''${ROCM_PATH} QUIET" "miopen PATHS ${miopen} QUIET" \
--replace-fail "\''${ROCM_PATH}/include/miopen/config.h" "${miopen}/include/miopen/config.h"
''}
'';
postBuild = lib.optionalString buildDocs ''
@@ -7,6 +7,7 @@
rocm-cmake,
rocm-smi,
rocm-core,
pkg-config,
clr,
mscclpp,
perl,
@@ -14,11 +15,14 @@
python3,
gtest,
chrpath,
roctracer,
rocprofiler,
rocprofiler-register,
autoPatchelfHook,
buildTests ? false,
gpuTargets ? (clr.localGpuTargets or [ ]),
# for passthru.tests
rccl,
}:
let
@@ -35,7 +39,7 @@ in
# infiniband ib_peer_mem support isn't in the mainline kernel but is carried by some distros
stdenv.mkDerivation (finalAttrs: {
pname = "rccl${clr.gpuArchSuffix}";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -48,13 +52,14 @@ stdenv.mkDerivation (finalAttrs: {
./fix-mainline-support-and-ub.diff
./enable-mscclpp-on-all-gfx9.diff
./rccl-test-missing-iomanip.diff
./fix_hw_reg_hw_id_gt_gfx10.patch
];
src = fetchFromGitHub {
owner = "ROCm";
repo = "rccl";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-998tDiC0Qp9hhcXtFpiCWqwdKPVT2vNp0GU/rng03Bw=";
hash = "sha256-XpD+UjgdbAoGYK5UvvTX3f8rny4tiEDH/vYoCdZhtjo=";
};
nativeBuildInputs = [
@@ -64,12 +69,14 @@ stdenv.mkDerivation (finalAttrs: {
perl
hipify
python3
pkg-config
autoPatchelfHook # ASAN doesn't add rpath without this
];
buildInputs = [
rocm-smi
gtest
roctracer
rocprofiler
rocprofiler-register
mscclpp
@@ -88,8 +95,9 @@ stdenv.mkDerivation (finalAttrs: {
"-DROCM_VERSION=${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=${mscclpp}"
# FIXME: this is still running a download because if(NOT mscclpp_nccl_FOUND) is commented out T_T
"-DENABLE_MSCCLPP=OFF"
#"-DMSCCLPP_ROOT=${mscclpp}"
# Manually define CMAKE_INSTALL_<DIR>
# See: https://github.com/NixOS/nixpkgs/pull/197838
"-DCMAKE_INSTALL_BINDIR=bin"
@@ -107,11 +115,15 @@ stdenv.mkDerivation (finalAttrs: {
# -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 ${san}-fno-omit-frame-pointer -momit-leaf-frame-pointer";
env.CFLAGS = "-I${clr}/include -I${roctracer}/include -O2 -fno-strict-aliasing ${san}-fno-omit-frame-pointer -momit-leaf-frame-pointer";
env.CXXFLAGS = "-I${clr}/include -I${roctracer}/include -O2 -fno-strict-aliasing ${san}-fno-omit-frame-pointer -momit-leaf-frame-pointer";
env.LDFLAGS = "${san}";
postPatch = ''
patchShebangs src tools
substituteInPlace CMakeLists.txt \
--replace-fail '${"\${HOST_OS_ID}"}' '"ubuntu"' \
--replace-fail 'target_include_directories(rccl PRIVATE ''${ROCM_SMI_INCLUDE_DIR})' \
'target_include_directories(rccl PRIVATE ''${ROCM_SMI_INCLUDE_DIRS})'
'';
postInstall =
@@ -130,6 +142,12 @@ stdenv.mkDerivation (finalAttrs: {
inherit (finalAttrs.src) repo;
};
# This package with sanitizers + manual integration test binaries built
# must be ran manually
passthru.tests.rccl = rccl.override {
buildTests = true;
};
meta = with lib; {
description = "ROCm communication collectives library";
homepage = "https://github.com/ROCm/rccl";
@@ -92,13 +92,7 @@ index 6d77784..49762d3 100644
}
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};
@@ -586,9 +586,9 @@ ncclResult_t ncclIbGdrSupport() {
int i = 0;
while (memory_peers_paths[i]) {
@@ -0,0 +1,33 @@
fix: __trace_hwreg should use HW_REG_HW_ID1 for all gfx10/11/12
diff --git a/src/device/common.h b/src/device/common.h
index c6c61021..742885f4 100644
--- a/src/device/common.h
+++ b/src/device/common.h
@@ -26,8 +26,9 @@
{ __atomic_store_n((DST), (SRC), __ATOMIC_SEQ_CST); }
#endif
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1200__) || defined(__gfx1201__)
-#define __trace_hwreg()
+#if defined(__GFX10__) || defined(__GFX11__) || defined(__GFX12__)
+#define __trace_hwreg() \
+ asm volatile ("s_getreg_b32 %0, hwreg(HW_REG_HW_ID1)" : "=s" (collTrace->data_0));
#else
#define __trace_hwreg() \
asm volatile ("s_getreg_b32 %0, hwreg(HW_REG_HW_ID)" : "=s" (collTrace->data_0));
diff --git a/tools/JitterBench/Common.hpp b/tools/JitterBench/Common.hpp
index bad12a1b..b59856a8 100644
--- a/tools/JitterBench/Common.hpp
+++ b/tools/JitterBench/Common.hpp
@@ -43,9 +43,9 @@ THE SOFTWARE.
#endif
// Macro for collecting HW_REG_HW_ID
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__NVCC__)
+#if defined(__GFX10__) || defined(__GFX11__) || defined(__GFX12__)
#define GetHwId(val) \
- val = 0
+ asm volatile ("s_getreg_b32 %0, hwreg(HW_REG_HW_ID1)" : "=s" (val));
#else
#define GetHwId(val) \
asm volatile ("s_getreg_b32 %0, hwreg(HW_REG_HW_ID)" : "=s" (val));
@@ -48,7 +48,7 @@ let
in
stdenv.mkDerivation (finalAttrs: {
pname = "rdc";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -64,7 +64,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "rdc";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-s/31b8/Kn5l1QJ941UMSB8SCzpvODsPfOLMmEBUYYmY=";
hash = "sha256-zILZPW9Lx5T+cMDqTg/zWy3ro+Nypzc9bDNTupZjt4s=";
};
nativeBuildInputs = [
@@ -95,6 +95,7 @@ stdenv.mkDerivation (finalAttrs: {
cmakeFlags = [
"-DCMAKE_VERBOSE_MAKEFILE=OFF"
"-DRDC_INSTALL_PREFIX=${placeholder "out"}"
"-DBUILD_RVS=OFF" # TODO: Needs RVS package
"-DBUILD_ROCRTEST=ON"
"-DRSMI_INC_DIR=${rocm-smi}/include"
"-DRSMI_LIB_DIR=${rocm-smi}/lib"
@@ -23,7 +23,7 @@
stdenv.mkDerivation (finalAttrs: {
pname = "rocalution";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -42,7 +42,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "rocALUTION";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-xdZ3HUiRGsreHfJH8RgL/s3jGyC5ABmBKcEfgtqWg8Y=";
hash = "sha256-bZx1Cc2jcIfysohKCKzj5mowM3IeCelRhVaBU73KnTo=";
};
nativeBuildInputs = [
@@ -65,11 +65,7 @@ stdenv.mkDerivation (finalAttrs: {
gtest
];
CXXFLAGS = "-I${openmp.dev}/include";
cmakeFlags = [
"-DOpenMP_C_INCLUDE_DIR=${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"
@@ -4,12 +4,14 @@
fetchFromGitHub,
fetchpatch,
rocmUpdateScript,
writableTmpDirAsHomeHook,
cmake,
rocm-cmake,
clr,
python3,
tensile,
msgpack,
boost,
msgpack-cxx,
libxml2,
gtest,
gfortran,
@@ -17,46 +19,19 @@
git,
amd-blis,
zstd,
roctracer,
hipblas-common,
hipblaslt,
python3Packages,
rocm-smi,
pkg-config,
buildTensile ? true,
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 ? 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
#
# `gfx1012` is not present in this list because the ISA compatibility patches
# 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 ? (
clr.localGpuTargets or [
"gfx900"
"gfx906"
"gfx908"
"gfx90a"
"gfx942"
"gfx1010"
"gfx1030"
"gfx1100"
"gfx1101"
"gfx1102"
"gfx1151"
"gfx1200"
"gfx1201"
]
),
gpuTargets ? (clr.localGpuTargets or clr.gpuTargets),
}:
let
@@ -64,17 +39,13 @@ let
in
stdenv.mkDerivation (finalAttrs: {
pname = "rocblas${clr.gpuArchSuffix}";
version = "6.3.3";
outputs = [
"out"
];
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "rocBLAS";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-IYcrVcGH4yZDkFZeNOJPfG0qsPS/WiH0fTSUSdo1BH4=";
hash = "sha256-FCzo/BOk4xLEFkdOdqcCXh4a9t3/OIIBEy8oz6oOMWg=";
};
nativeBuildInputs = [
@@ -83,6 +54,7 @@ stdenv.mkDerivation (finalAttrs: {
rocm-cmake
clr
git
pkg-config
]
++ lib.optionals buildTensile [
tensile
@@ -91,13 +63,16 @@ stdenv.mkDerivation (finalAttrs: {
buildInputs = [
python3
hipblas-common
roctracer
openmp
amd-blis
]
++ lib.optionals withHipBlasLt [
hipblaslt
]
++ lib.optionals buildTensile [
zstd
msgpack
msgpack-cxx
libxml2
python3Packages.msgpack
python3Packages.zstandard
@@ -107,27 +82,19 @@ stdenv.mkDerivation (finalAttrs: {
]
++ lib.optionals (buildTests || buildBenchmarks) [
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${amd-blis}/lib -lblis-mt -lcblas";
env.CXXFLAGS = "-fopenmp -I${lib.getDev boost}/include -I${hipblas-common}/include -I${roctracer}/include";
# Fails to link tests with undefined symbol: cblas_*
env.LDFLAGS = lib.optionalString (buildTests || buildBenchmarks) "-Wl,--as-needed -lcblas";
env.TENSILE_ROCM_ASSEMBLER_PATH = "${stdenv.cc}/bin/clang++";
cmakeFlags = [
(lib.cmakeFeature "CMAKE_BUILD_TYPE" "Release")
(lib.cmakeBool "CMAKE_VERBOSE_MAKEFILE" true)
(lib.cmakeFeature "Boost_INCLUDE_DIR" "${lib.getDev boost}/include") # msgpack FindBoost fails to find boost
(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")
@@ -144,8 +111,8 @@ stdenv.mkDerivation (finalAttrs: {
(lib.cmakeBool "BUILD_CLIENTS_BENCHMARKS" buildBenchmarks)
(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
# # 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"
@@ -153,6 +120,9 @@ stdenv.mkDerivation (finalAttrs: {
++ lib.optionals buildTensile [
"-DCPACK_SET_DESTDIR=OFF"
"-DLINK_BLIS=ON"
"-DBLIS_LIB=${amd-blis}/lib/libblis-mt.so"
"-DBLIS_INCLUDE_DIR=${amd-blis}/include/blis/"
"-DBLA_PREFER_PKGCONFIG=ON"
"-DTensile_CODE_OBJECT_VERSION=default"
"-DTensile_LOGIC=asm_full"
"-DTensile_LIBRARY_FORMAT=msgpack"
@@ -169,6 +139,7 @@ stdenv.mkDerivation (finalAttrs: {
url = "https://github.com/GZGavinZhao/rocBLAS/commit/89b75ff9cc731f71f370fad90517395e117b03bb.patch";
hash = "sha256-W/ohOOyNCcYYLOiQlPzsrTlNtCBdJpKVxO8s+4G7sjo=";
})
./hiplaslt-unstable-compat.patch
];
# Pass $NIX_BUILD_CORES to Tensile
@@ -176,7 +147,8 @@ stdenv.mkDerivation (finalAttrs: {
substituteInPlace cmake/build-options.cmake \
--replace-fail 'Tensile_CPU_THREADS ""' 'Tensile_CPU_THREADS "$ENV{NIX_BUILD_CORES}"'
substituteInPlace CMakeLists.txt \
--replace-fail "4.42.0" "4.43.0"
--replace-fail "4.43.0" "4.44.0" \
--replace-fail '0.10' '1.0'
'';
passthru.updateScript = rocmUpdateScript {
@@ -0,0 +1,99 @@
diff --git a/library/src/hipblaslt_host.cpp b/library/src/hipblaslt_host.cpp
index 8080070c..97d5216e 100644
--- a/library/src/hipblaslt_host.cpp
+++ b/library/src/hipblaslt_host.cpp
@@ -155,22 +155,22 @@ namespace
hipblaslt_compute_type<Tc>);
hipblaslt_ext::GemmProblemType problemType;
- problemType.op_a = (hipblasOperation_t)prob.trans_a;
- problemType.op_b = (hipblasOperation_t)prob.trans_b;
- problemType.type_a = hipblaslt_datatype<TiA>;
- problemType.type_b = hipblaslt_datatype<TiB>;
- problemType.type_c = hipblaslt_datatype<To>;
- problemType.type_d = hipblaslt_datatype<To>;
- problemType.type_compute = hipblaslt_compute_type<Tc>;
+ problemType.setOpA((hipblasOperation_t)prob.trans_a);
+ problemType.setOpB((hipblasOperation_t)prob.trans_b);
+ problemType.setTypeA(hipblaslt_datatype<TiA>);
+ problemType.setTypeB(hipblaslt_datatype<TiB>);
+ problemType.setTypeC(hipblaslt_datatype<To>);
+ problemType.setTypeD(hipblaslt_datatype<To>);
+ problemType.setTypeCompute(hipblaslt_compute_type<Tc>);
hipblaslt_ext::GemmEpilogue epilogue;
hipblaslt_ext::GemmInputs inputs;
- inputs.a = (void*)(prob.A + prob.buffer_offset_a);
- inputs.b = (void*)(prob.B + prob.buffer_offset_b);
- inputs.c = (void*)(prob.C + prob.buffer_offset_c);
- inputs.d = (void*)(prob.D + prob.buffer_offset_d);
- inputs.alpha = (void*)prob.alpha;
- inputs.beta = (void*)prob.beta;
+ inputs.setA((void*)(prob.A + prob.buffer_offset_a));
+ inputs.setB((void*)(prob.B + prob.buffer_offset_b));
+ inputs.setC((void*)(prob.C + prob.buffer_offset_c));
+ inputs.setD((void*)(prob.D + prob.buffer_offset_d));
+ inputs.setAlpha((void*)prob.alpha);
+ inputs.setBeta((void*)prob.beta);
gemm.setProblem(prob.m,
prob.n,
@@ -214,13 +214,13 @@ namespace
hipblaslt_compute_type<Tc>);
hipblaslt_ext::GemmProblemType problemType;
- problemType.op_a = (hipblasOperation_t)prob.trans_a;
- problemType.op_b = (hipblasOperation_t)prob.trans_b;
- problemType.type_a = hipblaslt_datatype<TiA>;
- problemType.type_b = hipblaslt_datatype<TiB>;
- problemType.type_c = hipblaslt_datatype<To>;
- problemType.type_d = hipblaslt_datatype<To>;
- problemType.type_compute = hipblaslt_compute_type<Tc>;
+ problemType.setOpA((hipblasOperation_t)prob.trans_a);
+ problemType.setOpB((hipblasOperation_t)prob.trans_b);
+ problemType.setTypeA(hipblaslt_datatype<TiA>);
+ problemType.setTypeB(hipblaslt_datatype<TiB>);
+ problemType.setTypeC(hipblaslt_datatype<To>);
+ problemType.setTypeD(hipblaslt_datatype<To>);
+ problemType.setTypeCompute(hipblaslt_compute_type<Tc>);
std::vector<int64_t> Ms(prob.batch_count);
std::vector<int64_t> Ns(prob.batch_count);
@@ -251,12 +251,12 @@ namespace
stridecs[batch] = prob.batch_stride_c;
strideds[batch] = prob.batch_stride_d;
batch_counts[batch] = 1;
- inputs[batch].a = (void*)(prob.batch_A[batch] + prob.buffer_offset_a);
- inputs[batch].b = (void*)(prob.batch_B[batch] + prob.buffer_offset_b);
- inputs[batch].c = (void*)(prob.batch_C[batch] + prob.buffer_offset_c);
- inputs[batch].d = (void*)(prob.batch_D[batch] + prob.buffer_offset_d);
- inputs[batch].alpha = (void*)prob.alpha;
- inputs[batch].beta = (void*)prob.beta;
+ inputs[batch].setA((void*)(prob.batch_A[batch] + prob.buffer_offset_a));
+ inputs[batch].setB((void*)(prob.batch_B[batch] + prob.buffer_offset_b));
+ inputs[batch].setC((void*)(prob.batch_C[batch] + prob.buffer_offset_c));
+ inputs[batch].setD((void*)(prob.batch_D[batch] + prob.buffer_offset_d));
+ inputs[batch].setAlpha((void*)prob.alpha);
+ inputs[batch].setBeta((void*)prob.beta);
}
gemm.setProblem(Ms,
diff --git a/library/src/tensile_host.cpp b/library/src/tensile_host.cpp
index 1b1289f3..ed463725 100644
--- a/library/src/tensile_host.cpp
+++ b/library/src/tensile_host.cpp
@@ -271,14 +271,6 @@ namespace
{
return Tensile::LazyLoadingInit::gfx90a;
}
- else if(deviceString.find("gfx940") != std::string::npos)
- {
- return Tensile::LazyLoadingInit::gfx940;
- }
- else if(deviceString.find("gfx941") != std::string::npos)
- {
- return Tensile::LazyLoadingInit::gfx941;
- }
else if(deviceString.find("gfx942") != std::string::npos)
{
return Tensile::LazyLoadingInit::gfx942;
@@ -12,6 +12,7 @@
texliveSmall,
doxygen,
graphviz,
writableTmpDirAsHomeHook,
buildDocs ? true,
}:
@@ -52,7 +53,7 @@ let
in
stdenv.mkDerivation (finalAttrs: {
pname = "rocdbgapi";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -65,7 +66,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "ROCdbgapi";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-6itfBrWVspobU47aiJAOQoxT8chwrq9scRn0or3bXto=";
hash = "sha256-Rr8+SNeFps0rjk4Jn2+rFmtRJfL42l0tNOz13oZQy+I=";
};
# FIXME: remove once https://github.com/doxygen/doxygen/issues/11634 is resolved
@@ -84,6 +85,7 @@ stdenv.mkDerivation (finalAttrs: {
git
]
++ lib.optionals buildDocs [
writableTmpDirAsHomeHook
latex
doxygen
graphviz
@@ -106,7 +108,6 @@ stdenv.mkDerivation (finalAttrs: {
# Unfortunately, it seems like we have to call make on this manually
postBuild = lib.optionalString buildDocs ''
export HOME=$(mktemp -d)
make -j$NIX_BUILD_CORES doc
'';
@@ -14,18 +14,19 @@
gtest,
openmp,
rocrand,
hiprand,
gpuTargets ? clr.localGpuTargets or clr.gpuTargets,
}:
stdenv.mkDerivation (finalAttrs: {
pname = "rocfft${clr.gpuArchSuffix}";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "rocFFT";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-RrxdwZ64uC7lQzyJI1eGHX2dmRnW8TfNThnuvuz5XWo=";
hash = "sha256-yaOjBF2aJkCBlxkydyOsrfT4lNZ0BVkS2jJC0fEiBug=";
};
nativeBuildInputs = [
@@ -35,14 +36,23 @@ stdenv.mkDerivation (finalAttrs: {
rocm-cmake
];
# FIXME: rocfft_aot_helper runs at the end of the build and has a risk of timing it out
# due to a long period with no terminal output
buildInputs = [ sqlite ];
buildInputs = [
sqlite
hiprand
];
patches = [
# Fixes build timeout due to no log output during rocfft_aot step
./log-every-n-aot-jobs.patch
];
cmakeFlags = [
"-DCMAKE_C_COMPILER=hipcc"
"-DCMAKE_CXX_COMPILER=hipcc"
"-DSQLITE_USE_SYSTEM_PACKAGE=ON"
"-DHIP_PLATFORM=amd"
"-DBUILD_CLIENTS=OFF"
"-DBUILD_SHARED_LIBS=ON"
"-DUSE_HIPRAND=ON"
"-DROCFFT_KERNEL_CACHE_ENABLE=ON"
# Manually define CMAKE_INSTALL_<DIR>
# See: https://github.com/NixOS/nixpkgs/pull/197838
"-DCMAKE_INSTALL_BINDIR=bin"
@@ -50,7 +60,7 @@ stdenv.mkDerivation (finalAttrs: {
"-DCMAKE_INSTALL_INCLUDEDIR=include"
]
++ lib.optionals (gpuTargets != [ ]) [
"-DAMDGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
"-DGPU_TARGETS=${lib.concatStringsSep ";" gpuTargets}"
];
passthru = {
@@ -74,11 +84,7 @@ stdenv.mkDerivation (finalAttrs: {
gtest
openmp
rocrand
];
cmakeFlags = [
"-DCMAKE_C_COMPILER=hipcc"
"-DCMAKE_CXX_COMPILER=hipcc"
hiprand
];
postInstall = ''
@@ -112,11 +118,6 @@ stdenv.mkDerivation (finalAttrs: {
rocrand
];
cmakeFlags = [
"-DCMAKE_C_COMPILER=hipcc"
"-DCMAKE_CXX_COMPILER=hipcc"
];
postInstall = ''
cp -a ../../../scripts/perf "$out/bin"
'';
@@ -141,11 +142,6 @@ stdenv.mkDerivation (finalAttrs: {
rocrand
];
cmakeFlags = [
"-DCMAKE_C_COMPILER=hipcc"
"-DCMAKE_CXX_COMPILER=hipcc"
];
installPhase = ''
runHook preInstall
mkdir "$out"
@@ -0,0 +1,35 @@
diff --git a/library/src/rocfft_aot_helper.cpp b/library/src/rocfft_aot_helper.cpp
index f0a889f4..452eb37f 100644
--- a/library/src/rocfft_aot_helper.cpp
+++ b/library/src/rocfft_aot_helper.cpp
@@ -771,26 +771,22 @@ int main(int argc, char** argv)
for(size_t i = 0; i < NUM_THREADS; ++i)
{
threads.emplace_back([&queue, &gpu_archs]() {
+ int compile_count = 0;
while(true)
{
auto item = queue.pop();
if(item.kernel_name.empty())
break;
+ if(++compile_count % 16 == 0)
+ std::cerr << "rocfft_aot_helper processing " << item.kernel_name << std::endl << std::flush;
for(const auto& gpu_arch : gpu_archs)
{
- if(item.sol_arch_name.empty())
+ if(item.sol_arch_name.empty() || gpu_arch.find(item.sol_arch_name) != std::string::npos)
{
RTCCache::cached_compile(
item.kernel_name, gpu_arch, item.generate_src, generator_sum());
}
- else if(gpu_arch.find(item.sol_arch_name) != std::string::npos)
- {
- // std::cout << "arch: " << gpu_arch
- // << ", solution-kernel: " << item.kernel_name << std::endl;
- RTCCache::cached_compile(
- item.kernel_name, gpu_arch, item.generate_src, generator_sum());
- }
}
}
});
@@ -23,13 +23,13 @@
stdenv.mkDerivation (finalAttrs: {
pname = "rocgdb";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "ROCgdb";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-Z+uk+ViLXgk5hXrIhVHRY0Kly7mktYms7M3o9Tmxv8s=";
hash = "sha256-evDWg2w2FHv6OU5BQOCAXTlDm7JpwdJ3Wh5a2i5r1gQ=";
};
nativeBuildInputs = [
@@ -9,13 +9,13 @@
stdenv.mkDerivation (finalAttrs: {
pname = "rocm-cmake";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "rocm-cmake";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-U4vGkH2iUlNJkqiNmVuFianD4WR9yuGvZsYG58smg0k=";
hash = "sha256-wAipNWAB66YNf7exLSNPAzg3NgkGD9LPKfKiulL5yak=";
};
nativeBuildInputs = [ cmake ];
@@ -37,8 +37,8 @@ stdenv.mkDerivation (finalAttrs: {
})
#[Comgr] Extend ISA compatibility for CCOB
(fetchpatch {
sha256 = "sha256-6Rwz12Lk4R2JK3olii3cr2Zd0ZLYe7VSpK1YRCOsJWY=";
url = "https://github.com/GZGavinZhao/rocm-llvm-project/commit/2d8c459a4d4c0567a7a275b4b54560d88e5c6919.patch";
sha256 = "sha256-PCi0QHLiEQCTIYRtSSbhOjXANJ3zC3VLdMED1BEfQeg=";
url = "https://github.com/GZGavinZhao/rocm-llvm-project/commit/fa80abb77d5ae6f8d89ab956e7ebda9c802a804f.patch";
relative = "amd/comgr";
})
];
@@ -17,13 +17,13 @@ let
in
stdenv.mkDerivation (finalAttrs: {
pname = "rocm-core";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "rocm-core";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-UDnPGvgwzwv49CzF+Kt0v95CsxS33BZeqNcKw1K6jRI=";
hash = "sha256-Z21k+0vi/P35WTHGAdfUAX7/jVv+d9g5YBo+HTi1Mpk=";
};
patches = [
@@ -24,14 +24,14 @@
# FIXME: Move to rocmPackages_common
buildPythonPackage rec {
pname = "rocm-docs-core";
version = "1.17.0";
version = "1.23.0";
format = "pyproject";
src = fetchFromGitHub {
owner = "ROCm";
repo = "rocm-docs-core";
rev = "v${version}";
hash = "sha256-fGRJyQq0Eook1Dc9Qy+dehQ5BVNX+6pkkFN9adb21Eo=";
hash = "sha256-5Qh83eJ9vju/uUb9gGA4B5Bh1WZCygIRbSnaEZzIdbw=";
};
buildInputs = [ setuptools ];
@@ -19,17 +19,18 @@
stdenv.mkDerivation (finalAttrs: {
pname = "rocm-runtime";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "ROCR-Runtime";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-du20+5VNYgwchGO7W7FIVebBqLPtfSBnmPVbPpgEZjo=";
hash = "sha256-zs0nydwYUY+8uiPyJxgTfAiV7spUMbESb0jUUvFf+AU=";
};
env.CFLAGS = "-I${numactl.dev}/include -I${elfutils.dev}/include -w";
env.CXXFLAGS = "-I${numactl.dev}/include -I${elfutils.dev}/include -w";
cmakeBuildType = "RelWithDebInfo";
separateDebugInfo = true;
__structuredAttrs = true;
nativeBuildInputs = [
pkg-config
@@ -56,17 +57,17 @@ stdenv.mkDerivation (finalAttrs: {
];
patches = [
(fetchpatch {
# rocr: Extend HIP ISA compatibility check
sha256 = "sha256-8r2Lb5lBfFaZC3knCxfXGcnkzNv6JxOKyJn2rD5gus4=";
url = "https://github.com/GZGavinZhao/ROCR-Runtime/commit/7c63e7185d8fcf08537a278908946145f6231121.patch";
})
# Patches for UB at runtime https://github.com/ROCm/ROCR-Runtime/issues/272
(fetchpatch {
# [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";
@@ -82,6 +83,9 @@ stdenv.mkDerivation (finalAttrs: {
url = "https://github.com/ROCm/ROCR-Runtime/commit/41bfc66aef437a5b349f71105fa4b907cc7e17d5.patch";
hash = "sha256-A7VhPR3eSsmjq2cTBSjBIz9i//WiNjoXm0EsRKtF+ns=";
})
# This causes a circular dependency, aqlprofile relies on hsa-runtime64
# which is part of rocm-runtime
# Worked around by having rocprofiler load aqlprofile directly
./remove-hsa-aqlprofile-dep.patch
];
@@ -3,26 +3,35 @@
stdenv,
fetchFromGitHub,
rocmUpdateScript,
pkg-config,
libdrm,
cmake,
wrapPython,
}:
stdenv.mkDerivation (finalAttrs: {
pname = "rocm-smi";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "rocm_smi_lib";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-j9pkyUt+p6IkhawIhiTymqDBydxXZunxmdyCyRN0RxE=";
hash = "sha256-yJ3Bf+tM39JWbY+A0NlpHNkvythdAdz6ZVp1AvLcXhk=";
};
patches = [ ./cmake.patch ];
patches = [
./cmake.patch
];
propagatedBuildInputs = [
libdrm
];
nativeBuildInputs = [
cmake
wrapPython
pkg-config
];
cmakeFlags = [
@@ -33,11 +42,21 @@ stdenv.mkDerivation (finalAttrs: {
"-DCMAKE_INSTALL_INCLUDEDIR=include"
];
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
'';
postInstall =
# wrap python programs, but undo two that need to be importable at that path
''
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
''
# workaround: propagate libdrm/ manually
# rocmcxx doesn't automatically add buildInputs to isystem include path like
# wrapper based toolchains, cmake files often don't find_package(rocm-smi) so
# can't rely on cmake propagated interface
# upstream have been shipping libdrm copied into /opt/rocm
+ ''
ln -s ${libdrm.dev}/include/libdrm/ $out/include/
'';
passthru.updateScript = rocmUpdateScript {
name = finalAttrs.pname;
@@ -9,25 +9,17 @@
busybox,
python3,
gnugrep,
clr, # Only for localGpuTargets
# rocminfo requires that the calling user have a password and be in
# the video group. If we let rocm_agent_enumerator rely upon
# rocminfo's output, then it, too, has those requirements. Instead,
# we can specify the GPU targets for this system (e.g. "gfx803" for
# Polaris) such that no system call is needed for downstream
# compilers to determine the desired target.
defaultTargets ? (clr.localGpuTargets or [ ]),
}:
stdenv.mkDerivation (finalAttrs: {
version = "6.3.3";
version = "6.4.3";
pname = "rocminfo";
src = fetchFromGitHub {
owner = "ROCm";
repo = "rocminfo";
rev = "rocm-${finalAttrs.version}";
sha256 = "sha256-fQPtO5TNbCbaZZ7VtGkkqng5QZ+FcScdh1opWr5YkLU=";
sha256 = "sha256-YscZ5sFsLOVBg98w2X6vTzniTvl9NfCkIE+HAH6vv5Y=";
};
strictDeps = true;
@@ -46,10 +38,6 @@ stdenv.mkDerivation (finalAttrs: {
sed 's,lsmod | grep ,${busybox}/bin/lsmod | ${gnugrep}/bin/grep ,' -i rocminfo.cc
'';
postInstall = lib.optionalString (defaultTargets != [ ]) ''
echo '${lib.concatStringsSep "\n" defaultTargets}' > $out/bin/target.lst
'';
passthru.updateScript = rocmUpdateScript {
name = finalAttrs.pname;
inherit (finalAttrs.src) owner;
@@ -5,7 +5,6 @@
rocmUpdateScript,
cmake,
rocm-cmake,
rocminfo,
clr,
git,
libxml2,
@@ -41,7 +40,7 @@ let
in
stdenv.mkDerivation (finalAttrs: {
pname = "rocmlir${suffix}";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -54,13 +53,12 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "rocMLIR";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-0SQ6uLDRfVfdCX+8a7D6pu6dYlFvX0HFzCDEvlKYfak=";
hash = "sha256-p/gvr1Z6yZtO5N+ecSouXiCrf520jt1HMOy/tohUHfI=";
};
nativeBuildInputs = [
cmake
rocm-cmake
clr
python3Packages.python
python3Packages.tomli
];
@@ -77,10 +75,6 @@ stdenv.mkDerivation (finalAttrs: {
ncurses
];
patches = [
./initparamdata-sort-const.patch
];
cmakeFlags = [
"-DLLVM_TARGETS_TO_BUILD=AMDGPU;${llvmNativeTarget}"
"-DCMAKE_BUILD_TYPE=Release"
@@ -109,12 +103,8 @@ stdenv.mkDerivation (finalAttrs: {
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" "\"${rocminfo}/bin"
substituteInPlace mlir/utils/performance/common/CMakeLists.txt \
--replace-fail "/opt/rocm" "${clr}"
--replace-fail " PATHS /opt/rocm" ""
'';
dontBuild = true;
@@ -1,13 +0,0 @@
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) {
@@ -15,7 +15,7 @@
stdenv.mkDerivation (finalAttrs: {
pname = "rocprim";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -31,7 +31,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "rocPRIM";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-0aHxpBuYIYhI2UER45YhHHL5YcxA+XeXoihcUs2AmCo=";
hash = "sha256-lH4MlBEkVJocq1VliGBtb7VvWfS6p/uIKWR239fSjRY=";
};
nativeBuildInputs = [
@@ -49,7 +49,7 @@ stdenv.mkDerivation (finalAttrs: {
];
cmakeFlags = [
"-DCMAKE_BUILD_TYPE=Release"
"-DHIP_CXX_COMPILER=amdclang++"
# Manually define CMAKE_INSTALL_<DIR>
# See: https://github.com/NixOS/nixpkgs/pull/197838
"-DCMAKE_INSTALL_BINDIR=bin"
@@ -18,13 +18,13 @@
stdenv.mkDerivation (finalAttrs: {
pname = "rocprofiler-register";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "rocprofiler-register";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-UZsCiGnudsbL1v5lKBx7Vz3/HRnGn4f86Pd+qu3ryh0=";
hash = "sha256-HaN4XMHuCRDfKOpfuZ2SkOEQfAZKouh6luqbtATUYm0=";
fetchSubmodules = true;
};
@@ -1,15 +0,0 @@
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index 46efbd5..ca2cc3b 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -127,10 +127,6 @@ function(generate_hsaco TARGET_ID INPUT_FILE OUTPUT_FILE)
DEPENDS ${INPUT_FILE} clang
COMMENT "Building ${OUTPUT_FILE}..."
VERBATIM)
- install(
- FILES ${PROJECT_BINARY_DIR}/${OUTPUT_FILE}
- DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/${PROJECT_NAME}/tests-v1
- COMPONENT tests)
set(HSACO_TARGET_LIST
${HSACO_TARGET_LIST} ${PROJECT_BINARY_DIR}/${OUTPUT_FILE}
PARENT_SCOPE)
@@ -7,6 +7,7 @@
cmake,
clang,
clr,
aqlprofile,
rocm-core,
rocm-runtime,
rocm-device-libs,
@@ -44,22 +45,16 @@ let
in
stdenv.mkDerivation (finalAttrs: {
pname = "rocprofiler";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "rocprofiler";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-x6DVt1logBE8aNnuwukQhsv/vRqkJALcfAF+6yEQuIk=";
hash = "sha256-CgW8foM4W3K19kUK/l8IsH2Q9DHi/z88viXTxhNqlHQ=";
fetchSubmodules = true;
};
patches = [
# These just simply won't build
./0000-dont-install-tests-hsaco.patch
./optional-aql-in-cmake.patch
];
nativeBuildInputs = [
cmake
clang
@@ -80,12 +75,11 @@ stdenv.mkDerivation (finalAttrs: {
mpi
systemd
gtest
aqlprofile
];
propagatedBuildInputs = [ rocmtoolkit-merged ];
# HACK: allow building without aqlprofile, probably explodes at runtime if use profiling
env.LDFLAGS = "-z nodefs -Wl,-undefined,dynamic_lookup,--unresolved-symbols=ignore-all";
#HACK: rocprofiler's cmake doesn't add these deps properly
env.CXXFLAGS = "-I${libpciaccess}/include -I${numactl.dev}/include -I${rocmtoolkit-merged}/include -I${elfutils.dev}/include -w";
@@ -95,9 +89,6 @@ stdenv.mkDerivation (finalAttrs: {
"-DGPU_TARGETS=${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"
@@ -123,6 +114,14 @@ stdenv.mkDerivation (finalAttrs: {
postInstall = ''
# Why do these have the executable bit set?
chmod -x $out/libexec/rocprofiler/counters/*.xml
# rocprof shell script wants to find it in the same bin dir, easiest to symlink in
ln -s ${clr}/bin/rocm_agent_enumerator $out/bin/rocm_agent_enumerator
'';
postFixup = ''
patchelf $out/lib/*.so \
--add-rpath ${aqlprofile}/lib \
--add-needed libhsa-amd-aqlprofile64.so
'';
passthru.updateScript = rocmUpdateScript {
@@ -1,147 +0,0 @@
From https://raw.githubusercontent.com/AphidGit/rocm_compile/refs/heads/main/rocprofiler.patch
diff --git a/cmake_modules/rocprofiler_env.cmake b/cmake_modules/rocprofiler_env.cmake
index 7b7c472..0aba3ed 100644
--- a/cmake_modules/rocprofiler_env.cmake
+++ b/cmake_modules/rocprofiler_env.cmake
@@ -36,6 +36,7 @@ if(ROCPROFILER_DEBUG_TRACE)
target_compile_definitions(rocprofiler-build-flags INTERFACE DEBUG_TRACE=1)
endif()
+set(ROCPROFILER_LD_AQLPROFILE false)
# Enable direct loading of AQL-profile HSA extension
if(ROCPROFILER_LD_AQLPROFILE)
target_compile_definitions(rocprofiler-build-flags INTERFACE ROCP_LD_AQLPROFILE=1)
@@ -80,9 +81,3 @@ if("${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 "${PROJECT_BINARY_DIR}/test-output")
-endif()
\ No newline at end of file
+endif()
@@ -12,13 +12,13 @@
stdenv.mkDerivation (finalAttrs: {
pname = "rocr-debug-agent";
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "rocr_debug_agent";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-HYag5/E72hopDhS9EVcdyGgSvzbCMzKqLC+SIS28Y9M=";
hash = "sha256-otoxZ2NHkPDIFhvn4/nvaQ/W4LF38Nx9MZ9IYEf1DyY=";
};
nativeBuildInputs = [
@@ -15,7 +15,7 @@
stdenv.mkDerivation (finalAttrs: {
pname = "rocrand${clr.gpuArchSuffix}";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -31,7 +31,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "rocRAND";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-rrRLPqEw39M+6dtPW8DcnQiSZNwxWNINJ1wjU098Vkk=";
hash = "sha256-pZAwmsWup0byGxJ5ADbztco8svBpIjBWITjs+OgKvQc=";
};
nativeBuildInputs = [
@@ -36,7 +36,7 @@
stdenv.mkDerivation (finalAttrs: {
pname = "rocsolver${clr.gpuArchSuffix}";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -52,7 +52,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "rocSOLVER";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-+sGU+0CB48iolJSyYo+xH36q5LCUp+nKtOYbguzMuhg=";
hash = "sha256-JthNY5rkrrm9bf1fgkO9HnARdX6H0kiF9EW1jMbBmj4=";
};
nativeBuildInputs = [
@@ -20,7 +20,7 @@
stdenv.mkDerivation (finalAttrs: {
pname = "rocsparse${clr.gpuArchSuffix}";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -36,7 +36,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "rocSPARSE";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-6Cut5rbyqKFzHaXfJZGApyY9Mj1Zq/+U8MkXgy4X4Pw=";
hash = "sha256-PrLyqHMAJYACkBMz97iBfwCaQ9Kf1IBL7IEf1IF1/m0=";
};
nativeBuildInputs = [
@@ -59,7 +59,7 @@ stdenv.mkDerivation (finalAttrs: {
];
cmakeFlags = [
"-DCMAKE_BUILD_TYPE=Release"
"-DCMAKE_CXX_COMPILER=amdclang++"
# Manually define CMAKE_INSTALL_<DIR>
# See: https://github.com/NixOS/nixpkgs/pull/197838
"-DCMAKE_INSTALL_BINDIR=bin"
@@ -15,7 +15,7 @@
stdenv.mkDerivation (finalAttrs: {
pname = "rocthrust";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -31,7 +31,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "rocThrust";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-c1+hqP/LipaQ2/lPJo79YBd9H0n0Y7yHkxe0/INE14s=";
hash = "sha256-IfMBVISClD1dk7FnAakP2GIpyZFrCnAloFRTaNdSKyw=";
};
nativeBuildInputs = [
@@ -19,7 +19,7 @@
stdenv.mkDerivation (finalAttrs: {
pname = "roctracer";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -35,7 +35,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "roctracer";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-GhnF7rqNLQLLB7nzIp0xNqyqBOwj9ZJ+hzzj1EAaXWU=";
hash = "sha256-Dwk5cBZLysmsVA2kwpQM0FQt2KXOGcaZcAw/d8VUaXw=";
};
nativeBuildInputs = [
@@ -19,7 +19,7 @@
stdenv.mkDerivation (finalAttrs: {
pname = "rocwmma";
version = "6.3.3";
version = "6.4.3";
outputs = [
"out"
@@ -38,7 +38,7 @@ stdenv.mkDerivation (finalAttrs: {
owner = "ROCm";
repo = "rocWMMA";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-kih3hn6QhcMmyj9n8f8eO+RIgKQgWKIuzg8fb0eoRPE=";
hash = "sha256-fjyxMrzt74rE7Gf4v4WawYltuw1fvahwZUpauMIE3qc=";
};
patches = lib.optionals (buildTests || buildBenchmarks) [
@@ -61,9 +61,6 @@ stdenv.mkDerivation (finalAttrs: {
];
cmakeFlags = [
"-DOpenMP_C_INCLUDE_DIR=${openmp.dev}/include"
"-DOpenMP_CXX_INCLUDE_DIR=${openmp.dev}/include"
"-DOpenMP_omp_LIBRARY=${openmp}/lib"
"-DROCWMMA_BUILD_TESTS=${if buildTests || buildBenchmarks then "ON" else "OFF"}"
"-DROCWMMA_BUILD_SAMPLES=${if buildSamples then "ON" else "OFF"}"
# Manually define CMAKE_INSTALL_<DIR>
@@ -29,13 +29,13 @@ stdenv.mkDerivation (finalAttrs: {
"cpu"
);
version = "6.3.3";
version = "6.4.3";
src = fetchFromGitHub {
owner = "ROCm";
repo = "rpp";
rev = "rocm-${finalAttrs.version}";
hash = "sha256-METwagek17/DdZGaOTQqvyU6xGt7OBMLHk4YM4KmgtA=";
hash = "sha256-rccVjSrOVIe4ZDtloCoCCI3u9UIcUqdirHIzS7ffAas=";
};
nativeBuildInputs = [
@@ -54,12 +54,7 @@ stdenv.mkDerivation (finalAttrs: {
boost
];
CFLAGS = "-I${openmp.dev}/include";
CXXFLAGS = "-I${openmp.dev}/include";
cmakeFlags = [
"-DOpenMP_C_INCLUDE_DIR=${openmp.dev}/include"
"-DOpenMP_CXX_INCLUDE_DIR=${openmp.dev}/include"
"-DOpenMP_omp_LIBRARY=${openmp}/lib"
"-DROCM_PATH=${clr}"
]
++ lib.optionals (gpuTargets != [ ]) [
@@ -6,55 +6,43 @@
buildPythonPackage,
pytestCheckHook,
setuptools,
distro,
pyyaml,
msgpack,
simplejson,
ujson,
orjson,
pandas,
joblib,
filelock,
clr,
rich,
isTensileLite ? false,
}:
buildPythonPackage rec {
pname = if isTensileLite then "tensilelite" else "tensile";
# Using a specific commit which has code object compression support from after the 6.3 release
pname = "tensile";
# Using a specific commit which has compression support from after the 6.4 release
# Without compression packages are too large for hydra
version = "6.3-unstable-2024-12-10";
version = "6.4-unstable-2025-06-12";
format = "pyproject";
src = fetchFromGitHub {
owner = "ROCm";
repo = "Tensile";
rev = "1752af518190500891a865379a4569b8abf6ba01";
hash = "sha256-Wvz4PVs//3Ox7ykZHpjPzOVwlyATyc+MmVVenfTzWK4=";
rev = "1ce87a9fe73610ffb962082f0a882360cd39b103";
hash = "sha256-qIuoIbmridy1HQVV10qPTzbccuxNJPsOvePaQQnClZc=";
};
# TODO: It should be possible to run asm caps test ONCE for all supported arches
# We currently disable the test because it's slow and runs each time tensile launches
postPatch =
lib.optionalString (!isTensileLite) ''
if grep -F .SafeLoader Tensile/LibraryIO.py; then
substituteInPlace Tensile/LibraryIO.py \
--replace-fail "yaml.SafeLoader" "yaml.CSafeLoader"
fi
# See TODO above about asm caps test
substituteInPlace Tensile/Common.py \
--replace-fail 'if globalParameters["AssemblerPath"] is not None:' "if False:"
''
+ ''
# Add an assert that the fallback 9,0,0 is supported before setting the kernel to it
# If it's not detected as supported we have an issue with compiler paths or the compiler is broken
# and it's better to stop immediately
substituteInPlace Tensile/KernelWriter.py \
--replace-fail '= (9,0,0)' '= (9,0,0);assert(globalParameters["AsmCaps"][(9,0,0)]["SupportedISA"])'
find . -type f -iname "*.sh" -exec chmod +x {} \;
patchShebangs Tensile
'';
postPatch = ''
substituteInPlace Tensile/Common.py \
--replace-fail 'if globalParameters["AssemblerPath"] is not None:' "if False:"
# Add an assert that the fallback 9,0,0 is supported before setting the kernel to it
# If it's not detected as supported we have an issue with compiler paths or the compiler is broken
# and it's better to stop immediately
substituteInPlace Tensile/KernelWriter.py \
--replace-fail '= (9,0,0)' '= (9,0,0);assert(globalParameters["AsmCaps"][(9,0,0)]["SupportedISA"])'
find . -type f -iname "*.sh" -exec chmod +x {} \;
patchShebangs Tensile
'';
buildInputs = [ setuptools ];
@@ -63,27 +51,19 @@ buildPythonPackage rec {
msgpack
pandas
joblib
]
++ lib.optionals (!isTensileLite) [
distro
rich
]
++ lib.optionals isTensileLite [
simplejson
ujson
orjson
];
patches =
lib.optional (!isTensileLite) ./tensile-solutionstructs-perf-fix.diff
++ lib.optional (!isTensileLite) ./tensile-create-library-dont-copy-twice.diff
++ lib.optional (!isTensileLite) (fetchpatch {
patches = [
./tensile-solutionstructs-perf-fix.diff
./tensile-create-library-dont-copy-twice.diff
(fetchpatch {
# [PATCH] Extend Tensile HIP ISA compatibility
sha256 = "sha256-d+fVf/vz+sxGqJ96vuxe0jRMgbC5K6j5FQ5SJ1e3Sl8=";
url = "https://github.com/GZGavinZhao/Tensile/commit/855cb15839849addb0816a6dde45772034a3e41f.patch";
})
++ lib.optional isTensileLite ./tensilelite-create-library-dont-copy-twice.diff
++ lib.optional isTensileLite ./tensilelite-gen_assembly-venv-err-handling.diff
++ lib.optional isTensileLite ./tensilelite-compression.diff;
];
doCheck = false; # Too many errors, not sure how to set this up properly
@@ -1,345 +0,0 @@
diff --git a/Tensile/TensileCreateLibrary.py b/Tensile/TensileCreateLibrary.py
index b8cea84558..1bc24bd1dd 100644
--- a/Tensile/TensileCreateLibrary.py
+++ b/Tensile/TensileCreateLibrary.py
@@ -41,6 +41,7 @@
from .SolutionLibrary import MasterSolutionLibrary
from .SolutionStructs import Solution
from .CustomYamlLoader import load_logic_gfx_arch
+from .Utilities.Profile import profile
import argparse
import collections
@@ -1233,7 +1234,7 @@ def validateLibrary(masterLibraries: MasterSolutionLibrary,
################################################################################
# Tensile Create Library
################################################################################
-@timing
+@profile
def TensileCreateLibrary():
print1("")
print1(HR)
@@ -1558,7 +1559,6 @@ def param(key, value):
print1("# Check if generated files exists.")
- @timing
def checkFileExistence(files):
for filePath in files:
if not os.path.exists(filePath):
diff --git a/Tensile/Utilities/Profile.py b/Tensile/Utilities/Profile.py
new file mode 100644
index 0000000000..cc3c7eb44c
--- /dev/null
+++ b/Tensile/Utilities/Profile.py
@@ -0,0 +1,77 @@
+################################################################################
+#
+# Copyright (C) 2016-2024 Advanced Micro Devices, Inc. All rights reserved.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in
+# all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+#
+################################################################################
+
+import cProfile
+import pstats
+import os
+
+from pathlib import Path
+from datetime import datetime, timezone
+from typing import Callable, Tuple
+
+PROFILE_ENV_VAR = "TENSILE_PROFILE"
+
+def profile(func: Callable) -> Callable:
+ """Profiling decorator.
+
+ Add ``@profile`` to mark a function for profiling; set the environment variable
+ TENSILE_PROFILE=ON to enable profiling decorated functions.
+ """
+ if not envVariableIsSet(PROFILE_ENV_VAR):
+ return func
+ def wrapper(*args, **kwargs):
+ path, filename = initProfileArtifacts(func.__name__)
+
+ prof = cProfile.Profile()
+ output = prof.runcall(func, *args, **kwargs)
+ result = pstats.Stats(prof)
+ result.sort_stats(pstats.SortKey.TIME)
+ result.dump_stats(path/filename)
+
+ return output
+ return wrapper
+
+def envVariableIsSet(varName: str) -> bool:
+ """Checks if the provided environment variable is set to "ON", "TRUE", or "1"
+ Args:
+ varName: Environment variable name.
+ Returns:
+ True if the environment variable is set, otherwise False.
+ """
+ value = os.environ.get(varName, "").upper()
+ return True if value in ["ON", "TRUE", "1"] else False
+
+def initProfileArtifacts(funcName: str) -> Tuple[Path, str]:
+ """Initializes filenames and paths for profiling artifacts based on the current datetime
+ Args:
+ funcName: The name of the function being profiled, nominally passed via func.__name__
+ Returns:
+ A tuple (path, filename) where the path is the artifact directory and filename is
+ a .prof file with the profiling results.
+ """
+ dt = datetime.now(timezone.utc)
+ filename = f"{funcName}-{dt.strftime('%Y-%m-%dT%H-%M-%SZ')}.prof"
+ path = Path().cwd()/f"profiling-results-{dt.strftime('%Y-%m-%d')}"
+ path.mkdir(exist_ok=True)
+ return path, filename
diff --git a/Tensile/TensileCreateLibrary.py b/Tensile/TensileCreateLibrary.py
index e62b0072df..2c843ba936 100644
--- a/Tensile/TensileCreateLibrary.py
+++ b/Tensile/TensileCreateLibrary.py
@@ -56,7 +56,7 @@
import sys
from timeit import default_timer as timer
from pathlib import Path
-from typing import Sequence, List
+from typing import Sequence, List, Union
def timing(func):
def wrapper(*args, **kwargs):
@@ -90,87 +90,142 @@ def processKernelSource(kernel, kernelWriterAssembly, ti):
return (err, src, header, kernelName, filename)
-def getAssemblyCodeObjectFiles(kernels, kernelWriterAssembly, outputPath):
- destDir = ensurePath(os.path.join(outputPath, 'library'))
- asmDir = kernelWriterAssembly.getAssemblyDirectory()
- archs = collections.defaultdict(list)
+def linkIntoCodeObject(
+ objFiles: List[str], coPathDest: Union[Path, str], kernelWriterAssembly: KernelWriterAssembly
+):
+ """Links object files into a code object file.
- for k in filter(lambda k: k['KernelLanguage'] == 'Assembly', kernels):
- archs[tuple(k['ISA'])].append(k)
+ Args:
+ objectFiles: A list of object files to be linked.
+ coPathDest: The destination path for the code object file.
+ kernelWriterAssembly: An instance of KernelWriterAssembly to get link arguments.
- coFiles = []
+ Raises:
+ RuntimeError: If linker invocation fails.
+ """
+ if os.name == "nt":
+ # On Windows, the objectFiles list command line (including spaces)
+ # exceeds the limit of 8191 characters, so using response file
+
+ responseFile = os.path.join('/tmp', 'clangArgs.txt')
+ with open(responseFile, 'wt') as file:
+ file.write(" ".join(objFiles))
+ file.flush()
+
+ args = [globalParameters['AssemblerPath'], '-target', 'amdgcn-amd-amdhsa', '-o', coFileRaw, '@clangArgs.txt']
+ subprocess.check_call(args, cwd=asmDir)
+ else:
+ numObjFiles = len(objFiles)
+ maxObjFiles = 10000
+
+ if numObjFiles > maxObjFiles:
+ batchedObjFiles = [objFiles[i:i+maxObjFiles] for i in range(0, numObjFiles, maxObjFiles)]
+ batchSize = int(math.ceil(numObjFiles / maxObjFiles))
+
+ newObjFiles = [str(coPathDest) + "." + str(i) for i in range(0, batchSize)]
+ newObjFilesOutput = []
+
+ for batch, filename in zip(batchedObjFiles, newObjFiles):
+ if len(batch) > 1:
+ args = [globalParameters["ROCmLdPath"], "-r"] + batch + [ "-o", filename]
+ print2(f"Linking object files into fewer object files: {' '.join(args)}")
+ subprocess.check_call(args)
+ newObjFilesOutput.append(filename)
+ else:
+ newObjFilesOutput.append(batchedObjFiles[0])
+
+ args = kernelWriterAssembly.getLinkCodeObjectArgs(newObjFilesOutput, str(coPathDest))
+ print2(f"Linking object files into code object: {' '.join(args)}")
+ subprocess.check_call(args)
+ else:
+ args = kernelWriterAssembly.getLinkCodeObjectArgs(objFiles, str(coPathDest))
+ print2(f"Linking object files into code object: {' '.join(args)}")
+ subprocess.check_call(args)
+
+
+def compressCodeObject(
+ coPathSrc: Union[Path, str], coPathDest: Union[Path, str], gfx: str, bundler: str
+):
+ """Compresses a code object file using the provided bundler.
+
+ Args:
+ coPathSrc: The source path of the code object file to be compressed.
+ coPathDest: The destination path for the compressed code object file.
+ gfx: The target GPU architecture.
+ bundler: The path to the Clang Offload Bundler executable.
+
+ Raises:
+ RuntimeError: If compressing the code object file fails.
+ """
+ args = [
+ bundler,
+ "--compress",
+ "--type=o",
+ "--bundle-align=4096",
+ f"--targets=host-x86_64-unknown-linux,hipv4-amdgcn-amd-amdhsa--{gfx}",
+ "--input=/dev/null",
+ f"--input={str(coPathSrc)}",
+ f"--output={str(coPathDest)}",
+ ]
+
+ print2(f"Bundling/compressing code objects: {' '.join(args)}")
+ try:
+ out = subprocess.check_output(args, stderr=subprocess.STDOUT)
+ print2(f"Output: {out}")
+ except subprocess.CalledProcessError as err:
+ raise RuntimeError(
+ f"Error compressing code object via bundling: {err.output}\nFailed command: {' '.join(args)}"
+ )
+
+def buildAssemblyCodeObjectFiles(kernels, kernelWriterAssembly, outputPath):
+
+ isAsm = lambda k: k["KernelLanguage"] == "Assembly"
+
+ extObj = ".o"
+ extCo = ".co"
+ extCoRaw = ".co.raw"
- for arch, archKernels in archs.items():
+ destDir = Path(ensurePath(os.path.join(outputPath, 'library')))
+ asmDir = Path(kernelWriterAssembly.getAssemblyDirectory())
+
+ archKernelMap = collections.defaultdict(list)
+ for k in filter(isAsm, kernels):
+ archKernelMap[tuple(k['ISA'])].append(k)
+
+ coFiles = []
+ for arch, archKernels in archKernelMap.items():
if len(archKernels) == 0:
continue
- archName = getGfxName(arch)
+ gfx = getGfxName(arch)
if globalParameters["MergeFiles"] or globalParameters["NumMergedFiles"] > 1 or globalParameters["LazyLibraryLoading"]:
- objectFiles = [kernelWriterAssembly.getKernelFileBase(k) + '.o' for k in archKernels if 'codeObjectFile' not in k]
+ objectFiles = [str(asmDir / (kernelWriterAssembly.getKernelFileBase(k) + extObj)) for k in archKernels if 'codeObjectFile' not in k]
- #Group kernels from placeholder libraries
coFileMap = collections.defaultdict(list)
+
if len(objectFiles):
- coFileMap[os.path.join(destDir, "TensileLibrary_"+archName+".co")] = objectFiles
+ coFileMap[asmDir / ("TensileLibrary_"+ gfx + extCoRaw)] = objectFiles
for kernel in archKernels:
coName = kernel.get("codeObjectFile", None)
if coName:
- coFileMap[os.path.join(destDir, coName+".co")] += [kernelWriterAssembly.getKernelFileBase(kernel) + '.o']
+ coFileMap[asmDir / (coName + extCoRaw)].append(str(asmDir / (kernelWriterAssembly.getKernelFileBase(kernel) + extObj)))
- for coFile, objectFiles in coFileMap.items():
- if os.name == "nt":
- # On Windows, the objectFiles list command line (including spaces)
- # exceeds the limit of 8191 characters, so using response file
+ for coFileRaw, objFiles in coFileMap.items():
- responseArgs = objectFiles
- responseFile = os.path.join(asmDir, 'clangArgs.txt')
- with open(responseFile, 'wt') as file:
- file.write( " ".join(responseArgs) )
- file.flush()
-
- args = [globalParameters['AssemblerPath'], '-target', 'amdgcn-amd-amdhsa', '-o', coFile, '@clangArgs.txt']
- subprocess.check_call(args, cwd=asmDir)
- else:
- numOfObjectFiles = len(objectFiles)
- splitFiles = 10000
- if numOfObjectFiles > splitFiles:
- slicedObjectFilesList = [objectFiles[x:x+splitFiles] for x in range(0, numOfObjectFiles, splitFiles)]
- objectFileBasename = os.path.split(coFile)[-1].split('.')[0]
- numOfOneSliceOfObjectFiles = int(math.ceil(numOfObjectFiles / splitFiles))
- newObjectFiles = [ objectFileBasename + "_" + str(i) + ".o" for i in range(0, numOfOneSliceOfObjectFiles)]
- newObjectFilesOutput = []
- for slicedObjectFiles, objectFile in zip(slicedObjectFilesList, newObjectFiles):
- if len(slicedObjectFiles) > 1:
- args = [globalParameters["ROCmLdPath"], "-r"] + slicedObjectFiles + [ "-o", objectFile ]
- if globalParameters["PrintCodeCommands"]:
- print(asmDir)
- print(' '.join(args))
- subprocess.check_call(args, cwd=asmDir)
- newObjectFilesOutput.append(objectFile)
- else:
- newObjectFilesOutput.append(slicedObjectFiles[0])
- args = kernelWriterAssembly.getLinkCodeObjectArgs(newObjectFilesOutput, coFile)
- if globalParameters["PrintCodeCommands"]:
- print(asmDir)
- print(' '.join(args))
- subprocess.check_call(args, cwd=asmDir)
- else:
- args = kernelWriterAssembly.getLinkCodeObjectArgs(objectFiles, coFile)
- if globalParameters["PrintCodeCommands"]:
- print(asmDir)
- print(' '.join(args))
- subprocess.check_call(args, cwd=asmDir)
+ linkIntoCodeObject(objFiles, coFileRaw, kernelWriterAssembly)
+ coFile = destDir / coFileRaw.name.replace(extCoRaw, extCo)
+ compressCodeObject(coFileRaw, coFile, gfx, globalParameters["ClangOffloadBundlerPath"])
coFiles.append(coFile)
else:
# no mergefiles
def newCoFileName(kName):
if globalParameters["PackageLibrary"]:
- return os.path.join(destDir, archName, kName + '.co')
+ return os.path.join(destDir, gfx, kName + '.co')
else:
- return os.path.join(destDir, kName + '_' + archName + '.co')
+ return os.path.join(destDir, kName + '_' + gfx + '.co')
def orgCoFileName(kName):
return os.path.join(asmDir, kName + '.co')
@@ -179,6 +234,8 @@ def orgCoFileName(kName):
map(lambda k: kernelWriterAssembly.getKernelFileBase(k), archKernels)), "Copying code objects"):
shutil.copyfile(src, dst)
coFiles.append(dst)
+ printWarning("Code object files are not compressed in `--no-merge-files` build mode.")
+
return coFiles
def which(p):
@@ -645,7 +702,7 @@ def success(kernel):
if not globalParameters["GenerateSourcesAndExit"]:
codeObjectFiles += buildSourceCodeObjectFiles(CxxCompiler, kernelFiles, outputPath)
- codeObjectFiles += getAssemblyCodeObjectFiles(kernelsToBuild, kernelWriterAssembly, outputPath)
+ codeObjectFiles += buildAssemblyCodeObjectFiles(kernelsToBuild, kernelWriterAssembly, outputPath)
Common.popWorkingPath() # build_tmp
Common.popWorkingPath() # workingDir
@@ -1,37 +0,0 @@
diff --git a/Tensile/TensileCreateLibrary.py b/Tensile/TensileCreateLibrary.py
index 2b9da394..b001fa7c 100644
--- a/Tensile/TensileCreateLibrary.py
+++ b/Tensile/TensileCreateLibrary.py
@@ -808,10 +808,13 @@ def copyStaticFiles(outputPath=None):
"ReductionTemplate.h",
"memory_gfx.h" ]
+ import filecmp
for fileName in libraryStaticFiles:
- # copy file
- shutil.copy( os.path.join(globalParameters["SourcePath"], fileName), \
- outputPath )
+ src = os.path.join(globalParameters["SourcePath"], fileName)
+ dst = os.path.join(outputPath, os.path.basename(src))
+ # no need to copy twice if it has already been copied
+ if not os.path.isfile(dst) or not filecmp.cmp(src, dst):
+ shutil.copyfile(src, dst)
return libraryStaticFiles
@@ -1417,9 +1420,13 @@ def TensileCreateLibrary():
writeCMake(outputPath, solutionFiles, sourceKernelFiles, staticFiles, masterLibraries)
# Make sure to copy the library static files.
+ import filecmp
for fileName in staticFiles:
- shutil.copy( os.path.join(globalParameters["SourcePath"], fileName), \
- outputPath )
+ src = os.path.join(globalParameters["SourcePath"], fileName)
+ dst = os.path.join(outputPath, os.path.basename(src))
+ # no need to copy twice if it has already been copied
+ if not os.path.isfile(dst) or not filecmp.cmp(src, dst):
+ shutil.copyfile(src, dst)
# write solutions and kernels
codeObjectFiles = writeSolutionsAndKernels(outputPath, CxxCompiler, None, solutions,
@@ -1,36 +0,0 @@
diff --git a/Tensile/Ops/gen_assembly.sh b/Tensile/Ops/gen_assembly.sh
index 0b21b6c6..609f1dd1 100755
--- a/Tensile/Ops/gen_assembly.sh
+++ b/Tensile/Ops/gen_assembly.sh
@@ -23,6 +23,8 @@
#
################################################################################
+set -x
+
archStr=$1
dst=$2
venv=$3
@@ -35,7 +37,13 @@ fi
toolchain=${rocm_path}/llvm/bin/clang++
-. ${venv}/bin/activate
+if ! [ -z ${TENSILE_GEN_ASSEMBLY_TOOLCHAIN+x} ]; then
+ toolchain="${TENSILE_GEN_ASSEMBLY_TOOLCHAIN}"
+fi
+
+if [ -f ${venv}/bin/activate ]; then
+ . ${venv}/bin/activate
+fi
IFS=';' read -r -a archs <<< "$archStr"
@@ -77,4 +85,6 @@ for arch in "${archs[@]}"; do
python3 ./ExtOpCreateLibrary.py --src=$dst --co=$dst/extop_$arch.co --output=$dst --arch=$arch
done
-deactivate
+if [ -f ${venv}/bin/activate ]; then
+ deactivate
+fi