easybuilders · smoors · Sep 21, 2024 · Sep 4, 2024
diff --git a/easybuild/easyconfigs/c/Clang/Clang-17.0.0_20230515-GCCcore-12.3.0-CUDA-12.1.1.eb b/easybuild/easyconfigs/c/Clang/Clang-17.0.0_20230515-GCCcore-12.3.0-CUDA-12.1.1.eb
@@ -0,0 +1,61 @@
+##
+# This file is an EasyBuild reciPY as per https://github.com/easybuilders/easybuild
+#
+# Copyright:: Copyright 2013-2015 Dmitri Gribenko, Ward Poelmans
+# Authors:: Dmitri Gribenko <[email protected]>
+# Authors:: Ward Poelmans <[email protected]>
+# License:: GPLv2 or later, MIT, three-clause BSD.
+# $Id$
+##
+
+name = 'Clang'
+version = '17.0.0_20230515'
+versionsuffix = '-CUDA-%(cudaver)s'
+_commit = 'c5dede880d17'
+
+homepage = 'https://clang.llvm.org/'
+description = """C, C++, Objective-C compiler, based on LLVM.  Does not
+ include C++ standard library -- use libstdc++ from GCC."""
+
+# Clang also depends on libstdc++ during runtime, but this dependency is
+# already specified as the toolchain.
+toolchain = {'name': 'GCCcore', 'version': '12.3.0'}
+
+sources = [{
+    'source_urls': ["https://github.com/llvm/llvm-project/archive"],
+    'download_filename': '%s.tar.gz' % _commit,
+    'filename': 'llvm-project-%s.tar.gz' % version,
+}]
+checksums = ['6f371f9ac208b8e9dc57fc117b1a9c8565d7ea2bbb49a2768cb9c3c0fee0291d']
+
+builddependencies = [
+    ('CMake', '3.26.3'),
+    ('Perl', '5.36.1'),
+    # Including Python bindings would require this as a runtime dep
+    ('Python', '3.11.3'),
+]
+dependencies = [
+    # since Clang is a compiler, binutils is a runtime dependency too
+    ('binutils', '2.40'),
+    ('hwloc', '2.9.1'),
+    ('libxml2', '2.11.4'),
+    ('ncurses', '6.4'),
+    ('GMP', '6.2.1'),
+    ('Z3', '4.12.2'),
+    ('CUDA', '12.1.1', '', SYSTEM),
+]
+
+# enabling RTTI makes the flang compiler need to link to libc++ so instead of
+#   flang-new -flang-experimental-exec -fopenmp hello_openmp.f90
+# you would need
+#   flang-new -flang-experimental-exec -fopenmp hello_openmp.f90 -l c++
+enable_rtti = False
+
+assertions = True
+python_bindings = False
+skip_all_tests = True
+
+llvm_runtimes = ['libunwind', 'libcxx', 'libcxxabi']
+llvm_projects = ['polly', 'lld', 'lldb', 'clang-tools-extra', 'flang']
+
+moduleclass = 'compiler'
diff --git a/easybuild/easyconfigs/t/Triton/Triton-2.1.0-disable_rocm_support.patch b/easybuild/easyconfigs/t/Triton/Triton-2.1.0-disable_rocm_support.patch
@@ -0,0 +1,88 @@
+Disable experimental support for AMD GPUs
+author: Alex Domingo (Vrije Universiteit Brussel)
+author: Samuel Moors (Vrije Universiteit Brussel)
+diff -ur triton-2.1.0.orig/CMakeLists.txt triton-2.1.0/CMakeLists.txt
+--- triton-2.1.0.orig/CMakeLists.txt	2023-09-01 08:28:27.000000000 +0200
++++ triton-2.1.0/CMakeLists.txt	2024-09-01 10:40:17.863374000 +0200
+@@ -70,7 +70,7 @@
+         AMDGPUInfo AMDGPUcodegen
+       )
+     else()
+-      find_package(LLVM 11 REQUIRED COMPONENTS "nvptx;amdgpu")
++      find_package(LLVM 11 REQUIRED COMPONENTS "nvptx")
+     endif()
+
+     message(STATUS "Found LLVM ${LLVM_PACKAGE_VERSION}")
+@@ -89,9 +89,7 @@
+       LLVMNVPTXCodeGen
+       LLVMNVPTXDesc
+       LLVMNVPTXInfo
+-      LLVMAMDGPUDisassembler
+       LLVMMCDisassembler
+-      LLVMAMDGPUCodeGen
+       LLVMMIRParser
+       LLVMGlobalISel
+       LLVMSelectionDAG
+@@ -116,10 +114,7 @@
+       LLVMObject
+       LLVMTextAPI
+       LLVMBitReader
+-      LLVMAMDGPUAsmParser
+       LLVMMCParser
+-      LLVMAMDGPUDesc
+-      LLVMAMDGPUUtils
+       LLVMMC
+       LLVMDebugInfoCodeView
+       LLVMDebugInfoMSF
+@@ -127,7 +122,6 @@
+       LLVMRemarks
+       LLVMBitstreamReader
+       LLVMBinaryFormat
+-      LLVMAMDGPUInfo
+       LLVMSupport
+       LLVMDemangle
+       LLVMPasses
+diff -ur triton-2.1.0.orig/lib/Target/HSACO/HSACOTranslation.cpp triton-2.1.0/lib/Target/HSACO/HSACOTranslation.cpp
+--- triton-2.1.0.orig/lib/Target/HSACO/HSACOTranslation.cpp	2023-09-01 08:28:27.000000000 +0200
++++ triton-2.1.0/lib/Target/HSACO/HSACOTranslation.cpp	2024-09-02 21:27:09.233712000 +0200
+@@ -37,13 +37,6 @@
+
+ namespace {
+
+-void init_llvm() {
+-  LLVMInitializeAMDGPUTarget();
+-  LLVMInitializeAMDGPUTargetInfo();
+-  LLVMInitializeAMDGPUTargetMC();
+-  LLVMInitializeAMDGPUAsmParser();
+-  LLVMInitializeAMDGPUAsmPrinter();
+-}
+
+ std::unique_ptr<llvm::TargetMachine>
+ initialize_module(llvm::Module *module, const std::string &triple,
+@@ -155,7 +148,6 @@
+ llir_to_amdgcn_and_hsaco(llvm::Module *module, std::string gfx_arch,
+                          std::string gfx_triple, std::string gfx_features) {
+
+-  init_llvm();
+
+   // verify and store llvm
+   auto module_obj = llvm::CloneModule(*module);
+diff -ur triton-2.1.0.orig/python/src/triton.cc triton-2.1.0/python/src/triton.cc
+--- triton-2.1.0.orig/python/src/triton.cc	2023-09-01 08:28:27.000000000 +0200
++++ triton-2.1.0/python/src/triton.cc	2024-09-01 10:41:29.146862440 +0200
+@@ -60,7 +60,6 @@
+ enum backend_t {
+   HOST,
+   CUDA,
+-  ROCM,
+ };
+
+ void init_triton_runtime(py::module &&m) {
+@@ -68,7 +67,6 @@
+   py::enum_<backend_t>(m, "backend", py::module_local())
+       .value("HOST", HOST)
+       .value("CUDA", CUDA)
+-      .value("ROCM", ROCM)
+       .export_values();
+ }
+
diff --git a/easybuild/easyconfigs/t/Triton/Triton-2.1.0-foss-2023a-CUDA-12.1.1.eb b/easybuild/easyconfigs/t/Triton/Triton-2.1.0-foss-2023a-CUDA-12.1.1.eb
@@ -0,0 +1,56 @@
+easyblock = 'PythonPackage'
+
+name = 'Triton'
+version = '2.1.0'
+versionsuffix = '-CUDA-%(cudaver)s'
+
+homepage = 'https://triton-lang.org/'
+
+description = """Triton is a language and compiler for parallel programming. It aims to provide a
+Python-based programming environment for productively writing custom DNN compute
+kernels capable of running at maximal throughput on modern GPU hardware."""
+
+toolchain = {'name': 'foss', 'version': '2023a'}
+
+github_account = 'openai'
+source_urls = [GITHUB_LOWER_SOURCE]
+sources = ['v%(version)s.tar.gz']
+patches = [
+    '%(name)s-%(version)s-disable_rocm_support.patch',
+    '%(name)s-%(version)s-use_eb_env_python_build.patch',
+]
+checksums = [
+    {'v2.1.0.tar.gz': '4338ca0e80a059aec2671f02bfc9320119b051f378449cf5f56a1273597a3d99'},
+    {'Triton-2.1.0-disable_rocm_support.patch': 'e4d7c0947c3287b3f0871a004e8b483963f637c9fa3ef6212ac3a34660de2a7c'},
+    {'Triton-2.1.0-use_eb_env_python_build.patch': 'd68bf766c699ad6a778d9449d3bccdadc2f20f1f86ba13e1359ad297b12fbf7c'},
+]
+
+builddependencies = [
+    ('Clang', '17.0.0_20230515', versionsuffix),  # this is the exact commit that would be downloaded by Triton
+    ('CMake', '3.26.3'),
+]
+
+dependencies = [
+    ('CUDA', '12.1.1', '', SYSTEM),
+    ('Python', '3.11.3'),
+    ('PyTorch', '2.1.2', versionsuffix),
+]
+
+use_pip = True
+download_dep_fail = True
+
+start_dir = 'python'
+
+preinstallopts = 'export LLVM_INCLUDE_DIRS=$EBROOTCLANG/include && '
+preinstallopts += 'export LLVM_LIBRARY_DIR=$EBROOTCLANG/lib && '
+preinstallopts += 'export LLVM_SYSPATH=$EBROOTCLANG && '
+preinstallopts += 'export TRITON_BUILD_WITH_CLANG_LLD=1 && '
+
+# make pip print output of cmake
+installopts = "-v "
+
+sanity_pip_check = True
+
+modluafooter = 'setenv("TRITON_PTXAS_PATH", pathJoin(os.getenv("CUDA_HOME"), "bin", "ptxas"))'
+
+moduleclass = 'devel'
diff --git a/easybuild/easyconfigs/t/Triton/Triton-2.1.0-use_eb_env_python_build.patch b/easybuild/easyconfigs/t/Triton/Triton-2.1.0-use_eb_env_python_build.patch
@@ -0,0 +1,93 @@
+Fix Triton cmake build to use dependencies from EasyBuild
+author: Alex Domingo (Vrije Universiteit Brussel)
+author: Samuel Moors (Vrije Universiteit Brussel)
+diff -ur triton-2.1.0.orig/lib/Target/LLVMIR/LLVMIRTranslation.cpp triton-2.1.0/lib/Target/LLVMIR/LLVMIRTranslation.cpp
+--- triton-2.1.0.orig/lib/Target/LLVMIR/LLVMIRTranslation.cpp	2023-09-01 08:28:27.000000000 +0200
++++ triton-2.1.0/lib/Target/LLVMIR/LLVMIRTranslation.cpp	2024-09-02 19:46:06.574421829 +0200
+@@ -35,6 +35,7 @@
+ #endif
+ #include <filesystem>
+ #include <iterator>
++#include <cstdlib>  // for getenv
+
+ namespace fs = std::filesystem;
+
+@@ -177,6 +178,7 @@
+   }
+
+   if (!funcs.empty()) {
++    std::filesystem::path cuda_home = std::getenv("CUDA_HOME");
+     static const std::string libdevice = "libdevice";
+     // first search for environmental path
+     std::string env_path = ::triton::tools::getenv("TRITON_LIBDEVICE_PATH");
+@@ -189,8 +191,7 @@
+     // `triton/third_party/cuda/lib/libdevice.10.bc`
+     static const auto this_library_path = getThisLibraryPath();
+     static const auto runtime_path =
+-        this_library_path.parent_path().parent_path() / "third_party" / "cuda" /
+-        "lib" / "libdevice.10.bc";
++        cuda_home / "nvvm" / "libdevice" / "libdevice.10.bc";
+     if (fs::exists(runtime_path)) {
+       externLibs.try_emplace(libdevice, runtime_path.string());
+     } else {
+diff -ur triton-2.1.0.orig/python/setup.py triton-2.1.0/python/setup.py
+--- triton-2.1.0.orig/python/setup.py	2023-09-01 08:28:27.000000000 +0200
++++ triton-2.1.0/python/setup.py	2024-09-02 18:16:58.044760300 +0200
+@@ -202,13 +202,13 @@
+
+     def build_extension(self, ext):
+         lit_dir = shutil.which('lit')
+-        user_home = os.getenv("HOME") or os.getenv("USERPROFILE") or \
+-            os.getenv("HOMEPATH") or None
+-        if not user_home:
+-            raise RuntimeError("Could not find user home directory")
+-        triton_cache_path = os.path.join(user_home, ".triton")
++        # user_home = os.getenv("HOME") or os.getenv("USERPROFILE") or \
++        #     os.getenv("HOMEPATH") or None
++        # if not user_home:
++        #     raise RuntimeError("Could not find user home directory")
++        # triton_cache_path = os.path.join(user_home, ".triton")
+         # lit is used by the test suite
+-        thirdparty_cmake_args = get_thirdparty_packages(triton_cache_path)
++        # thirdparty_cmake_args = get_thirdparty_packages(triton_cache_path)
+         extdir = os.path.abspath(os.path.dirname(self.get_ext_fullpath(ext.path)))
+         # create build directories
+         if not os.path.exists(self.build_temp):
+@@ -216,8 +216,9 @@
+         # python directories
+         python_include_dir = sysconfig.get_path("platinclude")
+         cmake_args = [
++            "-DCMAKE_VERBOSE_MAKEFILE:BOOL=ON",
+             "-DCMAKE_EXPORT_COMPILE_COMMANDS=ON",
+-            "-DLLVM_ENABLE_WERROR=ON",
++            # "-DLLVM_ENABLE_WERROR=ON",
+             "-DCMAKE_LIBRARY_OUTPUT_DIRECTORY=" + extdir,
+             "-DTRITON_BUILD_TUTORIALS=OFF",
+             "-DTRITON_BUILD_PYTHON_MODULE=ON",
+@@ -227,7 +228,7 @@
+         ]
+         if lit_dir is not None:
+             cmake_args.append("-DLLVM_EXTERNAL_LIT=" + lit_dir)
+-        cmake_args.extend(thirdparty_cmake_args)
++        # cmake_args.extend(thirdparty_cmake_args)
+
+         # configuration
+         cfg = get_build_type()
+@@ -245,7 +246,7 @@
+             build_args += ["--", "/m"]
+         else:
+             cmake_args += ["-DCMAKE_BUILD_TYPE=" + cfg]
+-            max_jobs = os.getenv("MAX_JOBS", str(2 * os.cpu_count()))
++            max_jobs = os.getenv("MAX_JOBS", str(len(os.sched_getaffinity(0))))
+             build_args += ['-j' + max_jobs]
+
+         if check_env_flag("TRITON_BUILD_WITH_CLANG_LLD"):
+@@ -262,7 +263,7 @@
+         subprocess.check_call(["cmake", "--build", "."] + build_args, cwd=cmake_dir)
+
+
+-download_and_copy_ptxas()
++# download_and_copy_ptxas()
+
+
+ setup(