Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.5.2 (#72

)

CMakeLists.txt

@@ -48,10 +48,13 @@ option(BUILD_EXAMPLES "Build examples" TRUE)
 option(BUILD_FT "Build functional tests" TRUE)
 option(BUILD_REG_TESTS "Build regression tests" TRUE)
 option(BUILD_CONFIG "Build cmake configs" TRUE)
-option(ENABLE_MPI "Enable MPI for library" TRUE)
-option(ENABLE_MPI_TESTS "Enable MPI for tests" TRUE)
-option(ENABLE_SYCL_INTEROP_EVENT "Enable support for interop event functionality" TRUE)
-option(ENABLE_OFI_HMEM "Enable support for OFI HMEM" FALSE)
+option(ENABLE_MPI "Enable MPI support" TRUE)
+option(ENABLE_MPI_TESTS "Enable MPI tests support" TRUE)
+option(ENABLE_SYCL_INTEROP_EVENT "Enable SYCL interop event support" TRUE)
+option(ENABLE_OFI_HMEM "Enable support OFI HMEM support" FALSE)
+option(ENABLE_OFI_OOT_PROV "Enable OFI out-of-tree providers support" FALSE)
+option(ENABLE_ITT "Enable ITT profiling support" TRUE)
+option(ENABLE_STUB_BACKEND "Enable stub backend" TRUE)
 
 option(USE_CODECOV_FLAGS "Calculate code coverage" FALSE)
 option(WITH_ASAN "Use address sanitizer, can only be used in Debug build" FALSE)
@@ -75,6 +78,9 @@ message(STATUS "Enable MPI support: ${ENABLE_MPI}")
 message(STATUS "Enable MPI tests support: ${ENABLE_MPI_TESTS}")
 message(STATUS "Enable SYCL interop event support: ${ENABLE_SYCL_INTEROP_EVENT}")
 message(STATUS "Enable OFI HMEM support: ${ENABLE_OFI_HMEM}")
+message(STATUS "Enable OFI out-of-tree providers support: ${ENABLE_OFI_OOT_PROV}")
+message(STATUS "Enable ITT profiling support: ${ENABLE_ITT}")
+message(STATUS "Enable stub backend" ${ENABLE_STUB_BACKEND})
 
 add_definitions(-DCCL_C_COMPILER="${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
 add_definitions(-DCCL_CXX_COMPILER="${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION}")
@@ -118,6 +124,13 @@ set(HWLOC_LIB_DIR "${DEPS_DIR}/hwloc/lib/")
 message(STATUS "HWLOC_INCLUDE_DIR: ${HWLOC_INCLUDE_DIR}")
 message(STATUS "HWLOC_LIB_DIR: ${HWLOC_LIB_DIR}")
 
+set(ITT_INCLUDE_DIR "${DEPS_DIR}/itt/include")
+set(ITT_LIB_DIR "${DEPS_DIR}/itt/lib64")
+message(STATUS "ITT_INCLUDE_DIR: ${ITT_INCLUDE_DIR}")
+message(STATUS "ITT_LIB_DIR: ${ITT_LIB_DIR}")
+
+set(LEVEL_ZERO_INCLUDE_DIR "${DEPS_DIR}/level_zero/include/")
+message(STATUS "LEVEL_ZERO_INCLUDE_DIR: ${LEVEL_ZERO_INCLUDE_DIR}")
 
 set(CMAKE_SKIP_INSTALL_RPATH TRUE)
 set(CMAKE_SKIP_RPATH TRUE)
@@ -130,30 +143,51 @@ if (${CMAKE_VERSION} VERSION_LESS 3.1)
 endif()
 
 # common release/debug compilation settings
-set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${C_COMPILER_FLAGS} -Wall -Wextra -Wno-unused-parameter -Wno-implicit-fallthrough -Werror -D_GNU_SOURCE -fvisibility=internal")
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${C_COMPILER_FLAGS} -Wall -Wextra -Wno-unused-parameter -Werror -D_GNU_SOURCE -fvisibility=internal")
 set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} ${C_COMPILER_FLAGS} -O0 -g -DENABLE_DEBUG")
 set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} ${C_COMPILER_FLAGS} -O3")
 set(CMAKE_C_FLAGS_RELWITHDEBINFO "${CMAKE_C_FLAGS_RELWITHDEBINFO} ${C_COMPILER_FLAGS} -O2 -g")
 set(CMAKE_C_STANDARD 99)
 set(CMAKE_C_STANDARD_REQUIRED ON)
 
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${CXX_COMPILER_FLAGS} -Wall -Wextra -Wno-unused-parameter -Wno-implicit-fallthrough -Werror -D_GNU_SOURCE -fvisibility=internal")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${CXX_COMPILER_FLAGS} -Wall -Wextra -Wno-unused-parameter -Werror -D_GNU_SOURCE -fvisibility=internal")
 set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} ${CXX_COMPILER_FLAGS} -O0 -g -DENABLE_DEBUG")
 set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} ${CXX_COMPILER_FLAGS} -O3")
 set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} ${CXX_COMPILER_FLAGS} -O2 -g")
 set(CMAKE_CXX_STANDARD 11)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
+if (NOT ${CMAKE_CXX_COMPILER_ID} STREQUAL "Intel")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Wno-implicit-fallthrough")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-implicit-fallthrough")
+endif()
+
 set(COMMON_CMAKE_DIR ${PROJECT_SOURCE_DIR}/cmake)
 if (COMPUTE_BACKEND)
+    precheck_compute_backend()
     message(STATUS "COMPUTE_BACKEND: ${COMPUTE_BACKEND}")
     set_compute_backend(${COMMON_CMAKE_DIR})
-    if (${COMPUTE_BACKEND} STREQUAL "dpcpp_level_zero" AND ENABLE_OFI_HMEM)
+    if (${COMPUTE_BACKEND} STREQUAL "dpcpp" AND ENABLE_OFI_HMEM)
         set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DCCL_ENABLE_OFI_HMEM=1")
         message(STATUS "Enable OFI HMEM support for compute backend ${COMPUTE_BACKEND}")
     endif()
 endif()
 
+if (ENABLE_OFI_OOT_PROV)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DCCL_ENABLE_OFI_OOT_PROV=1")
+    message(STATUS "Enable OFI out-of-tree providers support")
+endif()
+
+if (ENABLE_ITT)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DCCL_ENABLE_ITT=1")
+    message(STATUS "Enable ITT profiling support")
+endif()
+
+if (ENABLE_STUB_BACKEND)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DCCL_ENABLE_STUB_BACKEND=1")
+    message(STATUS "Enable stub backend")
+endif()
+
 if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU" AND ${CMAKE_CXX_COMPILER_ID} STREQUAL "GNU")
     if(NOT CMAKE_CXX_COMPILER_VERSION VERSION_LESS 7.0)
         #c++17 introduces algined new operator, use it
@@ -192,7 +226,7 @@ enable_testing()
 
 set(EXTERNAL_LIBS "")
 
-set(EXAMPLES_INC_DIRS ${PROJECT_SOURCE_DIR}/include ${PROJECT_SOURCE_DIR}/examples/include ${MPI_INCLUDE_DIR})
+set(EXAMPLES_INC_DIRS ${PROJECT_SOURCE_DIR}/include ${PROJECT_SOURCE_DIR}/examples/include ${MPI_INCLUDE_DIR} ${LEVEL_ZERO_INCLUDE_DIR})
 set(EXAMPLES_LIB_DIRS ${MPI_LIB_DIR} ${LIBFABRIC_LIB_DIR})
 
 # allow `deprecated`
@@ -221,7 +255,7 @@ endif()
 
 set(CCL_MAJOR_VERSION     "2021")
 set(CCL_MINOR_VERSION     "5")
-set(CCL_UPDATE_VERSION    "0")
+set(CCL_UPDATE_VERSION    "2")
 set(CCL_PRODUCT_STATUS    "Gold")
 string(TIMESTAMP CCL_PRODUCT_BUILD_DATE "%Y-%m-%dT %H:%M:%SZ")
 get_vcs_properties("git")

INSTALL.md

@@ -55,7 +55,7 @@ If your CXX compiler requires SYCL, it is possible to specify it (DPC++ is suppo
 Modify `cmake` command as follows:
 
 ```
-cmake .. -DCMAKE_C_COMPILER=your_c_compiler -DCMAKE_CXX_COMPILER=dpcpp -DCOMPUTE_BACKEND=dpcpp_level_zero
+cmake .. -DCMAKE_C_COMPILER=your_c_compiler -DCMAKE_CXX_COMPILER=dpcpp -DCOMPUTE_BACKEND=dpcpp
 ```
 
 ## Specify the build type

README.md

@@ -37,6 +37,12 @@ Intel(R) oneAPI DPC++/C++ Compiler with Level Zero v1.0 support.
 
 To install Level Zero, refer to the instructions in [Intel(R) Graphics Compute Runtime repository](https://github.com/intel/compute-runtime/releases) or to the [installation guide](https://dgpu-docs.intel.com/installation-guides/index.html) for oneAPI users.
 
+### BF16 support <!-- omit in toc -->
+
+- AVX512F-based implementation requires GCC 4.9 or higher.
+- AVX512_BF16-based implementation requires GCC 10.0 or higher and GNU binutils 2.33 or higher.
+
+
 ## Installation
 
 General installation scenario:

cmake/FindIntelSYCL_level_zero.cmake

@@ -26,14 +26,10 @@ else()
     message("DPCPP_ROOT prefix path hint is not defiend")
 endif()
 
-find_package(level_zero REQUIRED)
-set(COMPUTE_BACKEND_NAME ze_loader)
-
 if (NOT COMPUTE_BACKEND_NAME)
     message("Not OpenCL or L0")
 endif()
 
-
 include(CheckCXXCompilerFlag)
 include(FindPackageHandleStandardArgs)
 
@@ -70,12 +66,12 @@ find_package_handle_standard_args(IntelSYCL_level_zero
         INTEL_SYCL_SUPPORTED)
 
 if(IntelSYCL_level_zero_FOUND AND NOT TARGET Intel::SYCL_level_zero)
-	get_target_property(LEVEL_ZERO_INCLUDE_DIRECTORIES ze_loader INTERFACE_INCLUDE_DIRECTORIES)
-
     add_library(Intel::SYCL_level_zero UNKNOWN IMPORTED)
-	list(APPEND SYCL_LEVEL_ZERO_INCLUDE_DIRS "${LEVEL_ZERO_INCLUDE_DIRECTORIES}")
-	list(APPEND SYCL_LEVEL_ZERO_INCLUDE_DIRS "${INTEL_SYCL_INCLUDE_DIRS}")
+    message(STATUS "IntelSYCL_level_zero_FOUND: ${LEVEL_ZERO_INCLUDE_DIR}")
+    list(APPEND SYCL_LEVEL_ZERO_INCLUDE_DIRS "${LEVEL_ZERO_INCLUDE_DIR}")
+    list(APPEND SYCL_LEVEL_ZERO_INCLUDE_DIRS "${INTEL_SYCL_INCLUDE_DIRS}")
 
+    message(STATUS "SYCL_LEVEL_ZERO_INCLUDE_DIRS: ${SYCL_LEVEL_ZERO_INCLUDE_DIRS}")
     set(imp_libs
         $<$<STREQUAL:$<TARGET_PROPERTY:TYPE>,SHARED_LIBRARY>:-fsycl>
         ${COMPUTE_BACKEND_NAME})