Merge branch 'main' into eliasj42/windows-ci

.github/workflows/build_linux_packages.yml

@@ -75,39 +75,32 @@ jobs:
           cmake -B build -GNinja . \
             -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
             -DTHEROCK_AMDGPU_FAMILIES=gfx110X-dgpu \
-            -DTHEROCK_PACKAGE_VERSION="${package_version}"
+            -DTHEROCK_PACKAGE_VERSION="${package_version}" \
+            -DTHEROCK_VERBOSE=ON
           ./build_tools/watch_top_processes.sh &
-          cmake --build build --target therock-archives
+          cmake --build build --target therock-archives therock-dist
           kill %1
 
-      # - name: Build Tarballs
-      #   run: |
-      #     (cd build && cpack -G TXZ)
+      - name: Test Packaging
+        run: |
+          ctest --test-dir build --output-on-failure
 
       - name: Report
         if: ${{ !cancelled() }}
         run: |
-          ls -lh build
+          echo "Full SDK du:"
+          echo "------------"
+          du -h -d 1 build/dist/rocm
+          echo "Artifact Archives:"
+          echo "------------------"
+          ls -lh build/artifacts/*.tar.xz
+          echo "Artifacts:"
+          echo "----------"
+          du -h -d 1 build/artifacts
+          echo "CCache Stats:"
+          echo "-------------"
           ccache -s
 
-      # - name: Upload runtime artifacts
-      #   uses: actions/upload-artifact@v4
-      #   if: ${{ !cancelled() }}
-      #   with:
-      #     name: TheRock-runtime-linux-x86_64-tarball
-      #     path: |
-      #       build/TheRock-amdgpu-runtime-*.tar.xz
-      #     if-no-files-found: warn
-
-      # - name: Upload compiler artifacts
-      #   uses: actions/upload-artifact@v4
-      #   if: ${{ !cancelled() }}
-      #   with:
-      #     name: TheRock-compiler-linux-x86_64-tarball
-      #     path: |
-      #       build/TheRock-amdgpu-compiler*.tar.xz
-      #     if-no-files-found: warn
-
       - name: Save cache
         uses: actions/cache/save@v4
         if: always()

CMakeLists.txt

@@ -12,12 +12,14 @@ endif()
 project(THEROCK)
 
 cmake_policy(SET CMP0135 NEW)
+enable_testing()
 
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 include(CMakeDependentOption)
 include(ExternalProject)
 include(therock_amdgpu_targets)
 include(therock_artifacts)
+include(therock_features)
 include(therock_subproject)
 include(therock_job_pools)
 
@@ -30,15 +32,15 @@ set(THEROCK_BACKGROUND_BUILD_JOBS "0" CACHE STRING "Number of jobs to reserve fo
 
 set(THEROCK_PACKAGE_VERSION "git" CACHE STRING "Sets the package version string")
 set(ROCM_GIT_DIR "${THEROCK_SOURCE_DIR}/sources" CACHE PATH "Directory of rocm-org repo checkout")
+# Disable compatibility symlinks created in default ROCM cmake project wide
+set(ROCM_SYMLINK_LIBS OFF)
 message(STATUS "ROCM_GIT_DIR is set to: ${ROCM_GIT_DIR}")
 
 set(THEROCK_ARTIFACT_ARCHIVE_SUFFIX "" CACHE STRING "Suffix to add to artifact archive file stem names")
 set(THEROCK_ARTIFACT_ARCHIVE_TYPES "tar.xz" CACHE STRING "List of artifact archive types to generate")
 
-# Library specific enable flags.
-option(THEROCK_ENABLE_RCCL "Enable the comm_libs/rccl sub-project" ON)
-option(THEROCK_ENABLE_MATH_LIBS "Enable building of math libraries" ON)
-cmake_dependent_option(THEROCK_ENABLE_ML_LIBS "Enables building of ML frameworks" ON THEROCK_ENABLE_MATH_LIBS OFF)
+# Overall build settings.
+option(THEROCK_VERBOSE "Enables verbose CMake statuses" OFF)
 
 # Initialize the install directory.
 if(CMAKE_INSTALL_PREFIX_INITIALIZED_TO_DEFAULT)
@@ -50,6 +52,91 @@ set(ROCM_MAJOR_VERSION 6)
 set(ROCM_MINOR_VERSION 3)
 set(ROCM_PATCH_VERSION 1)
 
+################################################################################
+# Feature selection
+# Each feature added via therock_add_feature produces a boolean cache variable
+# like `THEROCK_ENABLE_${feature_name}` that takes its default value from the
+# `THEROCK_ENABLE_${group_name}` if a GROUP is present.
+################################################################################
+option(THEROCK_ENABLE_ALL "Enables building of all feature groups" ON)
+cmake_dependent_option(THEROCK_ENABLE_CORE "Enable building of core libraries" ON "THEROCK_ENABLE_ALL" OFF)
+cmake_dependent_option(THEROCK_ENABLE_COMM_LIBS "Enable building of comm libraries" ON "THEROCK_ENABLE_ALL" OFF)
+cmake_dependent_option(THEROCK_ENABLE_MATH_LIBS "Enable building of math libraries" ON "THEROCK_ENABLE_ALL" OFF)
+cmake_dependent_option(THEROCK_ENABLE_ML_LIBS "Enable building of ML libraries" ON "THEROCK_ENABLE_ALL" OFF)
+option(THEROCK_RESET_FEATURES "One-shot flag which forces all feature flags to their default state for this configuration run" OFF)
+
+# Base Features.
+therock_add_feature(COMPILER
+  GROUP ALL
+  DESCRIPTION "Enables the AMDGPU+host compiler toolchain"
+)
+therock_add_feature(HIPIFY
+  GROUP ALL
+  DESCRIPTION "Enables the hipify tool"
+  REQUIRES COMPILER
+)
+
+# Core Features.
+therock_add_feature(CORE_RUNTIME
+  GROUP CORE
+  DESCRIPTION "Enables the core device runtime and tools"
+)
+therock_add_feature(HIP_RUNTIME
+  GROUP CORE
+  DESCRIPTION "Enables the HIP runtime"
+  REQUIRES COMPILER CORE_RUNTIME
+)
+
+# Comm-libs Features.
+therock_add_feature(RCCL
+  GROUP COMM_LIBS
+  DESCRIPTION "Enables rccl"
+  REQUIRES COMPILER HIP_RUNTIME HIPIFY
+)
+
+# Math-libs Features.
+therock_add_feature(PRIM
+  GROUP MATH_LIBS
+  DESCRIPTION "Enables prim libraries (rocprim)"
+  REQUIRES COMPILER HIP_RUNTIME
+)
+therock_add_feature(BLAS
+  GROUP MATH_LIBS
+  DESCRIPTION "Enables blas libraries (hipblaslt, rocblas, hipblas)"
+  REQUIRES COMPILER HIP_RUNTIME
+)
+therock_add_feature(RAND
+  GROUP MATH_LIBS
+  DESCRIPTION "Enables rand libraries (hiprand, rocrand)"
+  REQUIRES COMPILER HIP_RUNTIME
+)
+therock_add_feature(FFT
+  GROUP MATH_LIBS
+  DESCRIPTION "Enables fft libraries"
+  REQUIRES COMPILER HIP_RUNTIME RAND
+)
+therock_add_feature(SOLVER
+  GROUP MATH_LIBS
+  DESCRIPTION "Enables solver libraries (hipsolver, rocsolver)"
+  REQUIRES COMPILER HIP_RUNTIME BLAS PRIM
+)
+therock_add_feature(SPARSE
+  GROUP MATH_LIBS
+  DESCRIPTION "Enables sparse libraries (hipsparse, rocsparse)"
+  REQUIRES COMPILER HIP_RUNTIME BLAS PRIM
+)
+
+# ML-Libs Features.
+therock_add_feature(MIOPEN
+  GROUP ML_LIBS
+  DESCRIPTION "Enables the MIOpen project (with minimal deps defaults to ON)"
+  REQUIRES COMPILER HIP_RUNTIME BLAS RAND
+)
+
+# Finalize all feature flags.
+therock_finalize_features()
+therock_report_features()
+
 ################################################################################
 # GPU target selection
 #
@@ -122,6 +209,7 @@ else()
   # TODO(#36): Enable more project builds on Windows and/or make the full list configurable.
 endif()
 
+add_subdirectory(examples)
 
 # ################################################################################
 # # Testing
@@ -144,3 +232,4 @@ endif()
 # Finalization
 ################################################################################
 therock_subproject_merge_compile_commands()
+therock_create_dist()

README.md

@@ -17,7 +17,8 @@ pip install CppHeaderParser
 ```
 sudo apt install \
   repo git-lfs libnuma-dev ninja-build cmake g++ pkg-config libdrm-dev \
-  libelf-dev xxd libgl1-mesa-dev
+  libelf-dev xxd libgl1-mesa-dev libbz2-dev libsqlite3-dev libgtest-dev
+
 ```
 
 # Checkout Sources
@@ -78,45 +79,50 @@ cmake -B build -GNinja . -DTHEROCK_AMDGPU_FAMILIES=gfx110X-dgpu
 # Or if iterating and wishing to cache:
 #   cmake -Bbuild -GNinja -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache .
 cmake --build build
-
-# Install per component.
-cmake --install build --component amdgpu-runtime
-cmake --install build --component amdgpu-runtime-dev
-cmake --install build --component amdgpu-compiler
-
-# Create archives.
-(cd build && cpack -G TBZ2)
 ```
 
-## Sanity Checks
+## Feature Flags
 
-The following does not replace a robust test suite. However, it will tell you
-whether the toolchain you have just crafted is viable at all (as in can load
-and enumerate devices).
-
-```
-./build/dlopen-hip install/lib/libamdhip64.so
-```
+By default, the project builds everything available. The following group flags
+allow enable/disable of selected subsets:
 
-HIP enabled IREE can also be used to enumerate:
+- `-DTHEROCK_ENABLE_ALL=OFF`: Disables all optional components.
+- `-DTHEROCK_ENABLE_CORE=OFF`: Disables all core components.
+- `-DTHEROCK_ENABLE_COMM_LIBS=OFF`: Disables all communication libraries.
+- `-DTHEROCK_ENABLE_MATH_LIBS=OFF`: Disables all math libraries.
+- `-DTHEROCK_ENABLE_ML_LIBS=OFF`: Disables all ML libraries.
 
-```
-LD_LIBRARY_PATH=install/lib iree-run-module --dump_devices=hip
-```
+Individual features can be controlled separately (typically in combination with
+`-DTHEROCK_ENABLE_ALL=OFF` or `-DTHEROCK_RESET_FEATURES=ON` to force a
+minimal build):
 
-# Development Notes
+- `-DTHEROCK_ENABLE_COMPILER=ON`: Enables the GPU+host compiler toolchain.
+- `-DTHEROCK_ENABLE_HIPIFY=ON`: Enables the hipify tool.
+- `-DTHEROCK_ENABLE_CORE_RUNTIME=ON`: Enables the core runtime components and tools.
+- `-DTHEROCK_ENABLE_HIP_RUNTIME=ON`: Enables the HIP runtime components.
+- `-DTHEROCK_ENABLE_RCCL=ON`: Enables RCCL.
+- `-DTHEROCK_ENABLE_PRIM=ON`: Enables the PRIM library.
+- `-DTHEROCK_ENABLE_BLAS=ON`: Enables the BLAS libraries.
+- `-DTHEROCK_ENABLE_RAND=ON`: Enables the RAND libraries.
+- `-DTHEROCK_ENABLE_SOLVER=ON`: Enables the SOLVER libraries.
+- `-DTHEROCK_ENABLE_SPARSE=ON`: Enables the SPARSE libraries.
+- `-DTHEROCK_ENABLE_MIOPEN=ON`: Enables MIOpen.
 
-## Building in a manylinux container
+Enabling any features will implicitly enable its *minimum* dependencies. Some
+libraries (like MIOpen) have a number of *optional* dependencies, which must
+be enabled manually if enabling/disabling individual features.
 
-Our CI builds run in such a container, and it can be useful to run locally.
+A report of enabled/disabled features and flags will be printed on every
+CMake configure.
 
-```
-docker run --rm -it -v $PWD:$PWD --entrypoint /bin/bash ghcr.io/nod-ai/manylinux_ghr_x86_64:main
-```
+## Sanity Checks
 
-Packages needed:
+Tests of the integrity of the build are enabled by default and can be run
+with ctest:
 
 ```
-yum install -y numactl-devel elfutils-libelf-devel vim-common git-lfs
-pip install CppHeaderParser
+ctest --test-dir build
 ```
+
+Testing functionality on an actual GPU is in progress and will be documented
+separately.

base/CMakeLists.txt

@@ -1,3 +1,12 @@
+################################################################################
+# aux-overlay
+################################################################################
+
+therock_cmake_subproject_declare(therock-aux-overlay
+  EXTERNAL_SOURCE_DIR "aux-overlay"
+)
+therock_cmake_subproject_activate(therock-aux-overlay)
+
 ################################################################################
 # rocm-cmake
 ################################################################################
@@ -6,7 +15,7 @@ therock_cmake_subproject_declare(rocm-cmake
   EXTERNAL_SOURCE_DIR "rocm-cmake"
 )
 therock_cmake_subproject_provide_package(rocm-cmake
-  ROCmCMakeBuildToolsConfig share/rocmcmakebuildtools/cmake)
+  ROCmCMakeBuildTools share/rocmcmakebuildtools/cmake)
 therock_cmake_subproject_provide_package(rocm-cmake
   ROCM share/rocm/cmake)
 therock_cmake_subproject_activate(rocm-cmake)
@@ -123,4 +132,5 @@ therock_provide_artifact(base
     rocm-cmake
     rocm-core
     rocm-half
+    therock-aux-overlay
 )

base/artifact.toml

@@ -36,3 +36,10 @@ include = "libexec/**"
 include = [
   "share/rocprofiler-register/tests/**",
 ]
+
+# therock-aux-overlay
+[components.lib."base/aux-overlay/stage"]
+default_patterns = false
+include = [
+  "**/*",
+]

base/aux-overlay/CMakeLists.txt

@@ -0,0 +1,25 @@
+# Auxiliary Overlay
+#
+# This project contains some odds and ends needed in order to complete
+# assembly of a full SDK. This includes:
+#
+# * Various redirect files to account for where things live in the tree from
+# a user perspective.
+# * Symlinks to account for distribution-level layout changes.
+# * Other hacks as needed.
+
+cmake_minimum_required(VERSION 3.25)
+
+project(therock-aux-overlay)
+
+if(NOT WIN32)
+  add_custom_target(symlinks ALL
+    COMMAND "${CMAKE_COMMAND}" -E create_symlink "lib/llvm" "${CMAKE_CURRENT_BINARY_DIR}/llvm"
+  )
+  install(FILES
+    "${CMAKE_CURRENT_BINARY_DIR}/llvm"
+    DESTINATION .
+  )
+endif()
+
+install(DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/overlay/" DESTINATION ".")

base/aux-overlay/overlay/lib/cmake/AMDDeviceLibs/AMDDeviceLibsConfig.cmake

@@ -0,0 +1,2 @@
+# Redirect to llvm/lib/cmake
+include("${CMAKE_CURRENT_LIST_DIR}/../../llvm/lib/cmake/AMDDeviceLibs/AMDDeviceLibsConfig.cmake")

base/aux-overlay/overlay/lib/cmake/clang/ClangConfig.cmake

@@ -0,0 +1,2 @@
+# Redirect to llvm/lib/cmake
+include("${CMAKE_CURRENT_LIST_DIR}/../../llvm/lib/cmake/clang/ClangConfig.cmake")

base/aux-overlay/overlay/lib/cmake/clang/ClangConfigVersion.cmake

@@ -0,0 +1,2 @@
+# Redirect to llvm/lib/cmake
+include("${CMAKE_CURRENT_LIST_DIR}/../../llvm/lib/cmake/clang/ClangConfigVersion.cmake")

base/aux-overlay/overlay/lib/cmake/lld/LLDConfig.cmake

@@ -0,0 +1,2 @@
+# Redirect to llvm/lib/cmake
+include("${CMAKE_CURRENT_LIST_DIR}/../../llvm/lib/cmake/lld/LLDConfig.cmake")

base/aux-overlay/overlay/lib/cmake/lld/LLDConfigVersion.cmake

@@ -0,0 +1,2 @@
+# Redirect to llvm/lib/cmake
+include("${CMAKE_CURRENT_LIST_DIR}/../../llvm/lib/cmake/lld/LLDConfigVersion.cmake")

base/aux-overlay/overlay/lib/cmake/llvm/LLVMConfig.cmake

@@ -0,0 +1,2 @@
+# Redirect to llvm/lib/cmake
+include("${CMAKE_CURRENT_LIST_DIR}/../../llvm/lib/cmake/llvm/LLVMConfig.cmake")

base/aux-overlay/overlay/lib/cmake/llvm/LLVMConfigVersion.cmake

@@ -0,0 +1,2 @@
+# Redirect to llvm/lib/cmake
+include("${CMAKE_CURRENT_LIST_DIR}/../../llvm/lib/cmake/llvm/LLVMConfigVersion.cmake")