add CUDA provider

bratpiorka · bratpiorka · commit 7eb436465e6b · 2024-08-07T10:20:40.000+02:00
diff --git a/.github/workflows/basic.yml b/.github/workflows/basic.yml
@@ -133,6 +133,7 @@ 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=ON
         -DUMF_FORMAT_CODE_STYLE=OFF
         -DUMF_DEVELOPER_MODE=ON
         -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
@@ -287,6 +288,7 @@ 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 +331,7 @@ 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/gpu_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 +173,10 @@ Additionally, required for tests:
 5) Required packages:
    - liblevel-zero-dev (Linux) or level-zero-sdk (Windows)
 
+#### CUDA memory provider
+
+TODO
+
 ### 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 gpu_shared_memory/level_zero_shared_memory.c
         LIBS disjoint_pool ze_loader umf)
 
     target_include_directories(
diff --git a/examples/gpu_shared_memory/CMakeLists.txt b/examples/gpu_shared_memory/CMakeLists.txt
@@ -3,7 +3,7 @@
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
 cmake_minimum_required(VERSION 3.14.0 FATAL_ERROR)
-project(umf_example_gpu_shared_memory LANGUAGES C)
+project(umf_example_level_zero_shared_memory LANGUAGES C)
 enable_testing()
 
 set(UMF_EXAMPLE_DIR "${CMAKE_SOURCE_DIR}/..")
@@ -48,8 +48,8 @@ set(LEVEL_ZERO_INCLUDE_DIRS
 message(STATUS "Level Zero include directory: ${LEVEL_ZERO_INCLUDE_DIRS}")
 
 # build the example
-set(EXAMPLE_NAME umf_example_gpu_shared_memory)
-add_executable(${EXAMPLE_NAME} gpu_shared_memory.c)
+set(EXAMPLE_NAME umf_example_level_zero_shared_memory)
+add_executable(${EXAMPLE_NAME} level_zero_shared_memory.c)
 target_include_directories(${EXAMPLE_NAME} PRIVATE ${LIBUMF_INCLUDE_DIRS}
                                                    ${UMF_EXAMPLE_DIR}/common)
 target_link_directories(${EXAMPLE_NAME} PRIVATE ${LIBUMF_LIBRARY_DIRS}
diff --git a/examples/gpu_shared_memory/level_zero_shared_memory.c b/examples/gpu_shared_memory/level_zero_shared_memory.c
diff --git a/include/umf/memory_provider.h b/include/umf/memory_provider.h
@@ -17,6 +17,14 @@
 extern "C" {
 #endif
 
+/// @brief USM memory allocation type
+typedef enum umf_usm_memory_type_t {
+    UMF_MEMORY_TYPE_UNKNOWN = 0, ///< The memory pointed to is of unknown type
+    UMF_MEMORY_TYPE_HOST,        ///< The memory pointed to is a host allocation
+    UMF_MEMORY_TYPE_DEVICE, ///< The memory pointed to is a device allocation
+    UMF_MEMORY_TYPE_SHARED, ///< The memory pointed to is a shared ownership allocation
+} umf_usm_memory_type_t;
+
 /// @brief A struct containing memory provider specific set of functions
 typedef struct umf_memory_provider_t *umf_memory_provider_handle_t;
 
diff --git a/include/umf/providers/provider_cuda.h b/include/umf/providers/provider_cuda.h
@@ -0,0 +1,30 @@
+/*
+ * Copyright (C) 2024 Intel Corporation
+ *
+ * Under the Apache License v2.0 with LLVM Exceptions. See LICENSE.TXT.
+ * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+*/
+
+#ifndef UMF_PROVIDER_CUDA_H
+#define UMF_PROVIDER_CUDA_H
+
+#include "umf/memory_provider.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/// @brief CUDA Memory Provider settings struct
+typedef struct cuda_memory_provider_params_t {
+    void *cuda_context_handle;         ///< Handle to the CUDA context
+    int cuda_device_handle;            ///< Handle to the CUDA device
+    umf_usm_memory_type_t memory_type; ///< Allocation memory type
+} cuda_memory_provider_params_t;
+
+umf_memory_provider_ops_t *umfCUDAMemoryProviderOps(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* UMF_PROVIDER_CUDA_H */
diff --git a/include/umf/providers/provider_level_zero.h b/include/umf/providers/provider_level_zero.h
@@ -14,14 +14,6 @@
 extern "C" {
 #endif
 
-/// @brief USM memory allocation type
-typedef enum umf_usm_memory_type_t {
-    UMF_MEMORY_TYPE_UNKNOWN = 0, ///< The memory pointed to is of unknown type
-    UMF_MEMORY_TYPE_HOST,        ///< The memory pointed to is a host allocation
-    UMF_MEMORY_TYPE_DEVICE, ///< The memory pointed to is a device allocation
-    UMF_MEMORY_TYPE_SHARED, ///< The memory pointed to is a shared ownership allocation
-} umf_usm_memory_type_t;
-
 /// @brief Level Zero Memory Provider settings struct
 typedef struct level_zero_memory_provider_params_t {
     void *level_zero_context_handle;   ///< Handle to the Level Zero context
diff --git a/scripts/docs_config/examples.rst b/scripts/docs_config/examples.rst
@@ -111,7 +111,7 @@ Freeing memory is as easy as can be::
 GPU shared memory
 ==============================================================================
 
-You can find the full example code in the `examples/gpu_shared_memory/gpu_shared_memory.c`_ file
+You can find the full example code in the `examples/gpu_shared_memory/level_zero_shared_memory.c`_ file
 in the UMF repository.
 
 TODO
@@ -209,7 +209,7 @@ function is called on the consumer side. The memory mappings on the consumer sid
 the :any:`umfCloseIPCHandle` function is called.
 
 .. _examples/basic/basic.c: https://github.com/oneapi-src/unified-memory-framework/blob/main/examples/basic/basic.c
-.. _examples/gpu_shared_memory/gpu_shared_memory.c: https://github.com/oneapi-src/unified-memory-framework/blob/main/examples/gpu_shared_memory/gpu_shared_memory.c
+.. _examples/gpu_shared_memory/level_zero_shared_memory.c: https://github.com/oneapi-src/unified-memory-framework/blob/main/examples/gpu_shared_memory/level_zero_shared_memory.c
 .. _examples/ipc_level_zero/ipc_level_zero.c: https://github.com/oneapi-src/unified-memory-framework/blob/main/examples/ipc_level_zero/ipc_level_zero.c
 .. _examples/custom_provider/file_provider.c: https://github.com/oneapi-src/unified-memory-framework/blob/main/examples/custom_provider/file_provider.c
 .. _examples/memspace: https://github.com/oneapi-src/unified-memory-framework/blob/main/examples/memspace/
diff --git a/scripts/qemu/run-build.sh b/scripts/qemu/run-build.sh
@@ -21,6 +21,7 @@ cd build
 cmake .. \
     -DCMAKE_BUILD_TYPE=Debug \
     -DUMF_BUILD_LEVEL_ZERO_PROVIDER=ON \
+    -DUMF_BUILD_CUDA_PROVIDER=ON \
     -DUMF_FORMAT_CODE_STYLE=OFF \
     -DUMF_DEVELOPER_MODE=ON \
     -DUMF_BUILD_LIBUMF_POOL_DISJOINT=ON \
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
diff --git a/src/provider/provider_cuda.c b/src/provider/provider_cuda.c
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
diff --git a/test/providers/cuda_helpers.cpp b/test/providers/cuda_helpers.cpp
diff --git a/test/providers/cuda_helpers.h b/test/providers/cuda_helpers.h
diff --git a/test/providers/provider_cuda.cpp b/test/providers/provider_cuda.cpp
