add CUDA provider

oneapi-src · Aug 9, 2024 · 12c6784 · 12c6784
1 parent 3040c31
commit 12c6784
Show file tree

Hide file tree

Showing 27 changed files with 935 additions and 18 deletions.
diff --git a/.github/workflows/basic.yml b/.github/workflows/basic.yml
@@ -22,6 +22,7 @@ jobs:
         compiler: [{c: gcc, cxx: g++}]
         shared_library: ['OFF']
         level_zero_provider: ['ON']
+        cuda_provider: ['ON']
         install_tbb: ['ON']
         disable_hwloc: ['OFF']
         link_hwloc_statically: ['OFF']
@@ -31,58 +32,67 @@ jobs:
             compiler: {c: gcc-7, cxx: g++-7}
             shared_library: 'OFF'
             level_zero_provider: 'ON'
+            cuda_provider: 'ON'
             install_tbb: 'ON'
           - os: 'ubuntu-22.04'
             build_type: Release
             compiler: {c: clang, cxx: clang++}
             shared_library: 'OFF'
             level_zero_provider: 'ON'
+            cuda_provider: 'ON'
             install_tbb: 'ON'
           - os: 'ubuntu-22.04'
             build_type: Release
             compiler: {c: gcc, cxx: g++}
             shared_library: 'ON'
             level_zero_provider: 'ON'
+            cuda_provider: 'ON'
             install_tbb: 'ON'
           - os: 'ubuntu-22.04'
             build_type: Debug
             compiler: {c: gcc, cxx: g++}
             shared_library: 'ON'
             level_zero_provider: 'ON'
+            cuda_provider: 'ON'
             install_tbb: 'ON'
           # test level_zero_provider='OFF'
           - os: 'ubuntu-22.04'
             build_type: Release
             compiler: {c: gcc, cxx: g++}
             shared_library: 'OFF'
             level_zero_provider: 'OFF'
+            cuda_provider: 'OFF'
             install_tbb: 'ON'
           # test icx compiler
           - os: 'ubuntu-22.04'
             build_type: Release
             compiler: {c: icx, cxx: icpx}
             shared_library: 'ON'
             level_zero_provider: 'ON'
+            cuda_provider: 'ON'
             install_tbb: 'ON' 
           # test without installing TBB
           - os: 'ubuntu-22.04'
             build_type: Release
             compiler: {c: gcc, cxx: g++}
             shared_library: 'ON'
             level_zero_provider: 'ON'
+            cuda_provider: 'ON'
             install_tbb: 'OFF'
           - os: 'ubuntu-22.04'
             build_type: Debug
             compiler: {c: gcc, cxx: g++}
             shared_library: 'ON'
             level_zero_provider: 'ON'
+            cuda_provider: 'ON'
             install_tbb: 'ON'
             disable_hwloc: 'ON'
           - os: 'ubuntu-22.04'
             build_type: Release
             compiler: {c: gcc, cxx: g++}
             shared_library: 'ON'
             level_zero_provider: 'ON'
+            cuda_provider: 'ON'
             install_tbb: 'ON'
             link_hwloc_statically: 'ON'
     runs-on: ${{matrix.os}}
@@ -133,6 +143,7 @@ jobs:
         -DCMAKE_C_COMPILER=${{matrix.compiler.c}}
         -DCMAKE_CXX_COMPILER=${{matrix.compiler.cxx}}
         -DUMF_BUILD_LEVEL_ZERO_PROVIDER=${{matrix.level_zero_provider}}
+        -DUMF_BUILD_CUDA_PROVIDER=${{matrix.cuda_provider}}
         -DUMF_FORMAT_CODE_STYLE=OFF
         -DUMF_DEVELOPER_MODE=ON
         -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
@@ -177,23 +188,27 @@ jobs:
         compiler: [{c: cl, cxx: cl}]
         shared_library: ['ON', 'OFF']
         level_zero_provider: ['ON']
+        cuda_provider: ['ON']
         include:
           - os: 'windows-2022'
             build_type: Release
             compiler: {c: clang-cl, cxx: clang-cl}
             shared_library: 'ON'
             level_zero_provider: 'ON'
+            cuda_provider: 'ON'
             toolset: "-T ClangCL"
           - os: 'windows-2022'
             build_type: Release
             compiler: {c: cl, cxx: cl}
             shared_library: 'ON'
             level_zero_provider: 'ON'
+            cuda_provider: 'ON'
           - os: 'windows-2022'
             build_type: Release
             compiler: {c: cl, cxx: cl}
             shared_library: 'ON'
             level_zero_provider: 'OFF'
+            cuda_provider: 'OFF'
 
     runs-on: ${{matrix.os}}
 
@@ -229,6 +244,7 @@ jobs:
         -DUMF_BUILD_LIBUMF_POOL_DISJOINT=ON
         -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
         -DUMF_BUILD_LEVEL_ZERO_PROVIDER=${{matrix.level_zero_provider}}
+        -DUMF_BUILD_CUDA_PROVIDER=${{matrix.cuda_provider}}
         -DUMF_TESTS_FAIL_ON_SKIP=ON
 
     - name: Build UMF
@@ -287,6 +303,7 @@ jobs:
         -DUMF_BUILD_LIBUMF_POOL_DISJOINT=ON
         -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=OFF
         -DUMF_BUILD_LEVEL_ZERO_PROVIDER=ON
+        -DUMF_BUILD_CUDA_PROVIDER=ON
         -DUMF_TESTS_FAIL_ON_SKIP=ON
         -DUMF_LINK_HWLOC_STATICALLY=ON
 
@@ -329,6 +346,7 @@ jobs:
         -DUMF_BUILD_LIBUMF_POOL_DISJOINT=ON
         -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=OFF
         -DUMF_BUILD_LEVEL_ZERO_PROVIDER=ON
+        -DUMF_BUILD_CUDA_PROVIDER=ON
         -DUMF_TESTS_FAIL_ON_SKIP=ON
         -DUMF_LINK_HWLOC_STATICALLY=ON
 

diff --git a/.github/workflows/benchmarks.yml b/.github/workflows/benchmarks.yml
@@ -63,6 +63,7 @@ jobs:
           -DUMF_FORMAT_CODE_STYLE=OFF
           -DUMF_DEVELOPER_MODE=OFF
           -DUMF_BUILD_LEVEL_ZERO_PROVIDER=ON
+          -DUMF_BUILD_CUDA_PROVIDER=ON
           -DUMF_BUILD_LIBUMF_POOL_DISJOINT=ON
           -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
 

diff --git a/.github/workflows/codeql.yml b/.github/workflows/codeql.yml
@@ -84,6 +84,7 @@ jobs:
         -DUMF_DEVELOPER_MODE=ON
         -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
         -DUMF_BUILD_LEVEL_ZERO_PROVIDER=ON
+        -DUMF_BUILD_CUDA_PROVIDER=ON
         -DUMF_TESTS_FAIL_ON_SKIP=ON
 
     - name: Build

diff --git a/.github/workflows/fast.yml b/.github/workflows/fast.yml
@@ -106,6 +106,7 @@ jobs:
         -DUMF_BUILD_TESTS=${{matrix.build_tests}}
         -DUMF_BUILD_EXAMPLES=ON
         -DUMF_BUILD_LEVEL_ZERO_PROVIDER=ON
+        -DUMF_BUILD_CUDA_PROVIDER=ON
         -DUMF_TESTS_FAIL_ON_SKIP=ON
         -DUMF_BUILD_SHARED_LIBRARY=ON
         ${{matrix.extra_build_options}}

diff --git a/.github/workflows/gpu.yml b/.github/workflows/gpu.yml
@@ -13,7 +13,7 @@ env:
   INSTL_DIR : "${{github.workspace}}/../install-dir"
 
 jobs:
-  gpu:
+  gpu-L0:
     name: Build
     env:
       BUILD_TYPE: Release
@@ -63,6 +63,7 @@ jobs:
           -DUMF_BUILD_LIBUMF_POOL_DISJOINT=ON
           -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
           -DUMF_BUILD_LEVEL_ZERO_PROVIDER=ON
+          -DUMF_BUILD_CUDA_PROVIDER=OFF
           -DUMF_TESTS_FAIL_ON_SKIP=ON
       
       - name: Configure build for Ubuntu
@@ -84,6 +85,71 @@ jobs:
           -DUMF_BUILD_LIBUMF_POOL_DISJOINT=ON
           -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
           -DUMF_BUILD_LEVEL_ZERO_PROVIDER=ON
+          -DUMF_BUILD_CUDA_PROVIDER=OFF
+          -DUMF_TESTS_FAIL_ON_SKIP=ON
+
+      - name: Build UMF
+        run: cmake --build ${{env.BUILD_DIR}} --config ${{env.BUILD_TYPE}} -j ${{matrix.number_of_processors}}
+
+      - name: Run tests
+        working-directory: ${{env.BUILD_DIR}}
+        run: ctest -C ${{env.BUILD_TYPE}} --output-on-failure --test-dir test
+
+      - name: Run examples
+        working-directory: ${{env.BUILD_DIR}}
+        run: ctest --output-on-failure --test-dir examples -C ${{env.BUILD_TYPE}}
+
+      - name: Run benchmarks
+        working-directory: ${{env.BUILD_DIR}}
+        run: ctest --output-on-failure --test-dir benchmark -C ${{env.BUILD_TYPE}} --exclude-regex umf-bench-multithreaded
+
+  gpu-CUDA:
+    name: Build
+    env:
+      BUILD_TYPE: Release
+    # run only on upstream; forks will not have the HW
+    if: github.repository == 'oneapi-src/unified-memory-framework'
+    strategy:
+      matrix:
+        shared_library: ['ON', 'OFF']
+        # TODO add windows
+        os: ['Ubuntu']
+        include:
+          - os: 'Ubuntu'
+            compiler: {c: gcc, cxx: g++}
+            number_of_processors: '$(nproc)'
+
+    runs-on: ["DSS-CUDA", "DSS-${{matrix.os}}"]
+    steps:
+      - name: Checkout
+        uses: actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11 # v4.1.1
+        with:
+          fetch-depth: 0
+
+      - name: Get information about platform
+        if: matrix.os == 'Ubuntu'
+        run: .github/scripts/get_system_info.sh
+
+      - name: Configure build for Ubuntu
+        if: matrix.os == 'Ubuntu'
+        run: >
+          cmake
+          -B ${{env.BUILD_DIR}}
+          -DCMAKE_INSTALL_PREFIX="${{env.INSTL_DIR}}"
+          -DCMAKE_BUILD_TYPE=${{env.BUILD_TYPE}}
+          -DCMAKE_C_COMPILER=${{matrix.compiler.c}}
+          -DCMAKE_CXX_COMPILER=${{matrix.compiler.cxx}}
+          -DUMF_BUILD_SHARED_LIBRARY=${{matrix.shared_library}}
+          -DUMF_BUILD_BENCHMARKS=ON
+          -DUMF_BUILD_TESTS=ON
+          -DUMF_BUILD_GPU_TESTS=ON
+          -DUMF_BUILD_GPU_EXAMPLES=ON
+          -DUMF_FORMAT_CODE_STYLE=OFF
+          -DUMF_DEVELOPER_MODE=ON
+          -DUMF_BUILD_LIBUMF_POOL_DISJOINT=ON
+          -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
+          -DUMF_BUILD_LEVEL_ZERO_PROVIDER=OFF
+          -DUMF_BUILD_CUDA_PROVIDER=ON
           -DUMF_TESTS_FAIL_ON_SKIP=ON
 
       - name: Build UMF

diff --git a/.github/workflows/nightly.yml b/.github/workflows/nightly.yml
@@ -79,6 +79,7 @@ jobs:
         -DUMF_BUILD_LIBUMF_POOL_DISJOINT=ON
         -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
         -DUMF_BUILD_LEVEL_ZERO_PROVIDER=OFF
+        -DUMF_BUILD_CUDA_PROVIDER=OFF
         -DUMF_USE_VALGRIND=1
         -DUMF_TESTS_FAIL_ON_SKIP=ON
 

diff --git a/.github/workflows/pr_push.yml b/.github/workflows/pr_push.yml
@@ -37,6 +37,7 @@ jobs:
         -DUMF_FORMAT_CODE_STYLE=ON
         -DUMF_BUILD_TESTS=OFF
         -DUMF_BUILD_LEVEL_ZERO_PROVIDER=OFF
+        -DUMF_BUILD_CUDA_PROVIDER=OFF
         -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=OFF
 
     - name: Check C/C++ formatting

diff --git a/.github/workflows/sanitizers.yml b/.github/workflows/sanitizers.yml
@@ -55,6 +55,7 @@ jobs:
         -DCMAKE_C_COMPILER=${{matrix.compiler.c}}
         -DCMAKE_CXX_COMPILER=${{matrix.compiler.cxx}}
         -DUMF_BUILD_LEVEL_ZERO_PROVIDER=ON
+        -DUMF_BUILD_CUDA_PROVIDER=ON
         -DUMF_FORMAT_CODE_STYLE=OFF
         -DUMF_DEVELOPER_MODE=ON
         -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
@@ -130,6 +131,7 @@ jobs:
         -DUMF_USE_ASAN=${{matrix.sanitizers.asan}}
         -DUMF_BUILD_EXAMPLES=ON
         -DUMF_BUILD_LEVEL_ZERO_PROVIDER=OFF
+        -DUMF_BUILD_CUDA_PROVIDER=OFF
         -DUMF_TESTS_FAIL_ON_SKIP=ON
 
     - name: Build UMF

diff --git a/.github/workflows/valgrind.yml b/.github/workflows/valgrind.yml
@@ -35,6 +35,7 @@ jobs:
         -DUMF_BUILD_LIBUMF_POOL_DISJOINT=ON
         -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
         -DUMF_BUILD_LEVEL_ZERO_PROVIDER=OFF
+        -DUMF_BUILD_CUDA_PROVIDER=OFF
         -DUMF_USE_VALGRIND=1
         -DUMF_TESTS_FAIL_ON_SKIP=ON
 

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -36,6 +36,7 @@ find_package(PkgConfig)
 # Build Options
 option(UMF_BUILD_SHARED_LIBRARY "Build UMF as shared library" OFF)
 option(UMF_BUILD_LEVEL_ZERO_PROVIDER "Build Level Zero memory provider" ON)
+option(UMF_BUILD_CUDA_PROVIDER "Build CUDA memory provider" ON)
 option(UMF_BUILD_LIBUMF_POOL_DISJOINT
        "Build the libumf_pool_disjoint static library" OFF)
 option(UMF_BUILD_LIBUMF_POOL_JEMALLOC
@@ -399,6 +400,11 @@ if(UMF_BUILD_LEVEL_ZERO_PROVIDER)
     add_optional_symbol(umfLevelZeroMemoryProviderOps)
 endif()
 
+# Conditional configuration for CUDA provider
+if(UMF_BUILD_CUDA_PROVIDER)
+    add_optional_symbol(umfCUDAMemoryProviderOps)
+endif()
+
 if(NOT UMF_DISABLE_HWLOC)
     add_optional_symbol(umfOsMemoryProviderOps)
 endif()

diff --git a/README.md b/README.md
@@ -21,7 +21,7 @@ For a quick introduction to UMF usage, please see
 documentation, which includes the code of the
 [basic example](https://github.com/oneapi-src/unified-memory-framework/blob/main/examples/basic/basic.c)
 and the more advanced one that allocates
-[USM memory from the GPU device](https://github.com/oneapi-src/unified-memory-framework/blob/main/examples/basic/gpu_shared_memory.c)
+[USM memory from the Level Zero device](https://github.com/oneapi-src/unified-memory-framework/blob/main/examples/level_zero_shared_memory/level_zero_shared_memory.c)
 using the Level Zero API and UMF Level Zero memory provider.
 
 ## Build
@@ -101,6 +101,7 @@ List of options provided by CMake:
 | - | - | - | - |
 | UMF_BUILD_SHARED_LIBRARY | Build UMF as shared library | ON/OFF | OFF |
 | UMF_BUILD_LEVEL_ZERO_PROVIDER | Build Level Zero memory provider | ON/OFF | ON |
+| UMF_BUILD_CUDA_PROVIDER | Build CUDA memory provider | ON/OFF | ON |
 | UMF_BUILD_LIBUMF_POOL_DISJOINT | Build the libumf_pool_disjoint static library | ON/OFF | OFF |
 | UMF_BUILD_LIBUMF_POOL_JEMALLOC | Build the libumf_pool_jemalloc static library | ON/OFF | OFF |
 | UMF_BUILD_TESTS | Build UMF tests | ON/OFF | ON |
@@ -172,6 +173,22 @@ Additionally, required for tests:
 5) Required packages:
    - liblevel-zero-dev (Linux) or level-zero-sdk (Windows)
 
+#### CUDA memory provider
+
+A memory provider that provides memory from CUDA device.
+
+##### Requirements
+
+1) Linux or Windows OS
+2) The `UMF_BUILD_CUDA_PROVIDER` option turned `ON` (by default)
+
+Additionally, required for tests:
+
+3) The `UMF_BUILD_GPU_TESTS` option turned `ON`
+4) System with CUDA compatible GPU
+5) Required packages:
+   - nvidia-cuda-dev (Linux) or cuda-sdk (Windows)
+
 ### Memory pool managers
 
 #### Proxy pool (part of libumf)

diff --git a/benchmark/CMakeLists.txt b/benchmark/CMakeLists.txt
@@ -85,6 +85,10 @@ function(add_umf_benchmark)
         target_compile_definitions(${BENCH_NAME}
                                    PRIVATE UMF_BUILD_LEVEL_ZERO_PROVIDER=1)
     endif()
+    if(UMF_BUILD_CUDA_PROVIDER)
+        target_compile_definitions(${BENCH_NAME}
+                                   PRIVATE UMF_BUILD_CUDA_PROVIDER=1)
+    endif()
     if(UMF_BUILD_GPU_TESTS)
         target_compile_definitions(${BENCH_NAME} PRIVATE UMF_BUILD_GPU_TESTS=1)
     endif()

diff --git a/benchmark/ubench.c b/benchmark/ubench.c
@@ -505,6 +505,8 @@ UBENCH_EX(ipc, disjoint_pool_with_level_zero_provider) {
 }
 #endif /* (defined UMF_BUILD_LIBUMF_POOL_DISJOINT && defined UMF_BUILD_LEVEL_ZERO_PROVIDER && defined UMF_BUILD_GPU_TESTS) */
 
+// TODO add IPC benchmark for CUDA
+
 UBENCH_MAIN()
 
 #if defined(_MSC_VER)

diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
@@ -41,14 +41,16 @@ if(UMF_POOL_SCALABLE_ENABLED)
     endif()
 endif()
 
+# TODO add CUDA examples
+
 if(UMF_BUILD_GPU_EXAMPLES
    AND UMF_BUILD_LIBUMF_POOL_DISJOINT
    AND UMF_BUILD_LEVEL_ZERO_PROVIDER)
-    set(EXAMPLE_NAME umf_example_gpu_shared_memory)
+    set(EXAMPLE_NAME umf_example_level_zero_shared_memory)
 
     add_umf_executable(
         NAME ${EXAMPLE_NAME}
-        SRCS gpu_shared_memory/gpu_shared_memory.c
+        SRCS level_zero_shared_memory/level_zero_shared_memory.c
         LIBS disjoint_pool ze_loader umf)
 
     target_include_directories(