|
| 1 | +option(FLANG_EXPERIMENTAL_CUDA_RUNTIME |
| 2 | + "Compile Fortran runtime as CUDA sources (experimental)" OFF |
| 3 | + ) |
| 4 | + |
| 5 | +set(FLANG_LIBCUDACXX_PATH "" CACHE PATH "Path to libcu++ package installation") |
| 6 | + |
| 7 | +set(FLANG_EXPERIMENTAL_OMP_OFFLOAD_BUILD "off" CACHE STRING |
| 8 | + "Compile Fortran runtime as OpenMP target offload sources (experimental). Valid options are 'off', 'host_device', 'nohost'") |
| 9 | + |
| 10 | +set(FLANG_OMP_DEVICE_ARCHITECTURES "all" CACHE STRING |
| 11 | + "List of OpenMP device architectures to be used to compile the Fortran runtime (e.g. 'gfx1103;sm_90')") |
| 12 | + |
| 13 | +macro(enable_cuda_compilation files) |
| 14 | + if (FLANG_EXPERIMENTAL_CUDA_RUNTIME) |
| 15 | + if (BUILD_SHARED_LIBS) |
| 16 | + message(FATAL_ERROR |
| 17 | + "BUILD_SHARED_LIBS is not supported for CUDA build of Fortran runtime" |
| 18 | + ) |
| 19 | + endif() |
| 20 | + |
| 21 | + enable_language(CUDA) |
| 22 | + |
| 23 | + # TODO: figure out how to make target property CUDA_SEPARABLE_COMPILATION |
| 24 | + # work, and avoid setting CMAKE_CUDA_SEPARABLE_COMPILATION. |
| 25 | + set(CMAKE_CUDA_SEPARABLE_COMPILATION ON) |
| 26 | + |
| 27 | + # Treat all supported sources as CUDA files. |
| 28 | + set_source_files_properties(${files} PROPERTIES LANGUAGE CUDA) |
| 29 | + set(CUDA_COMPILE_OPTIONS) |
| 30 | + if ("${CMAKE_CUDA_COMPILER_ID}" MATCHES "Clang") |
| 31 | + # Allow varargs. |
| 32 | + set(CUDA_COMPILE_OPTIONS |
| 33 | + -Xclang -fcuda-allow-variadic-functions |
| 34 | + ) |
| 35 | + endif() |
| 36 | + if ("${CMAKE_CUDA_COMPILER_ID}" MATCHES "NVIDIA") |
| 37 | + set(CUDA_COMPILE_OPTIONS |
| 38 | + --expt-relaxed-constexpr |
| 39 | + # Disable these warnings: |
| 40 | + # 'long double' is treated as 'double' in device code |
| 41 | + -Xcudafe --diag_suppress=20208 |
| 42 | + -Xcudafe --display_error_number |
| 43 | + ) |
| 44 | + endif() |
| 45 | + set_source_files_properties(${files} PROPERTIES COMPILE_OPTIONS |
| 46 | + "${CUDA_COMPILE_OPTIONS}" |
| 47 | + ) |
| 48 | + |
| 49 | + if (EXISTS "${FLANG_LIBCUDACXX_PATH}/include") |
| 50 | + # When using libcudacxx headers files, we have to use them |
| 51 | + # for all files of F18 runtime. |
| 52 | + include_directories(AFTER ${FLANG_LIBCUDACXX_PATH}/include) |
| 53 | + add_compile_definitions(RT_USE_LIBCUDACXX=1) |
| 54 | + endif() |
| 55 | + endif() |
| 56 | +endmacro() |
| 57 | + |
| 58 | +macro(enable_omp_offload_compilation files) |
| 59 | + if (NOT FLANG_EXPERIMENTAL_OMP_OFFLOAD_BUILD STREQUAL "off") |
| 60 | + # 'host_device' build only works with Clang compiler currently. |
| 61 | + # The build is done with the CMAKE_C/CXX_COMPILER, i.e. it does not use |
| 62 | + # the in-tree built Clang. We may have a mode that would use the in-tree |
| 63 | + # built Clang. |
| 64 | + # |
| 65 | + # 'nohost' is supposed to produce an LLVM Bitcode library, |
| 66 | + # and it has to be done with a C/C++ compiler producing LLVM Bitcode |
| 67 | + # compatible with the LLVM toolchain version distributed with the Flang |
| 68 | + # compiler. |
| 69 | + # In general, the in-tree built Clang should be used for 'nohost' build. |
| 70 | + # Note that 'nohost' build does not produce the host version of Flang |
| 71 | + # runtime library, so there will be two separate distributable objects. |
| 72 | + # 'nohost' build is a TODO. |
| 73 | + |
| 74 | + if (NOT FLANG_EXPERIMENTAL_OMP_OFFLOAD_BUILD STREQUAL "host_device") |
| 75 | + message(FATAL_ERROR "Unsupported OpenMP offload build of Flang runtime") |
| 76 | + endif() |
| 77 | + if (BUILD_SHARED_LIBS) |
| 78 | + message(FATAL_ERROR |
| 79 | + "BUILD_SHARED_LIBS is not supported for OpenMP offload build of Fortran runtime" |
| 80 | + ) |
| 81 | + endif() |
| 82 | + |
| 83 | + if ("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang" AND |
| 84 | + "${CMAKE_C_COMPILER_ID}" MATCHES "Clang") |
| 85 | + |
| 86 | + set(all_amdgpu_architectures |
| 87 | + "gfx700;gfx701;gfx801;gfx803;gfx900;gfx902;gfx906" |
| 88 | + "gfx908;gfx90a;gfx90c;gfx940;gfx1010;gfx1030" |
| 89 | + "gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036" |
| 90 | + "gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151" |
| 91 | + ) |
| 92 | + set(all_nvptx_architectures |
| 93 | + "sm_35;sm_37;sm_50;sm_52;sm_53;sm_60;sm_61;sm_62" |
| 94 | + "sm_70;sm_72;sm_75;sm_80;sm_86;sm_89;sm_90" |
| 95 | + ) |
| 96 | + set(all_gpu_architectures |
| 97 | + "${all_amdgpu_architectures};${all_nvptx_architectures}" |
| 98 | + ) |
| 99 | + # TODO: support auto detection on the build system. |
| 100 | + if (FLANG_OMP_DEVICE_ARCHITECTURES STREQUAL "all") |
| 101 | + set(FLANG_OMP_DEVICE_ARCHITECTURES ${all_gpu_architectures}) |
| 102 | + endif() |
| 103 | + list(REMOVE_DUPLICATES FLANG_OMP_DEVICE_ARCHITECTURES) |
| 104 | + |
| 105 | + string(REPLACE ";" "," compile_for_architectures |
| 106 | + "${FLANG_OMP_DEVICE_ARCHITECTURES}" |
| 107 | + ) |
| 108 | + |
| 109 | + set(OMP_COMPILE_OPTIONS |
| 110 | + -fopenmp |
| 111 | + -fvisibility=hidden |
| 112 | + -fopenmp-cuda-mode |
| 113 | + --offload-arch=${compile_for_architectures} |
| 114 | + # Force LTO for the device part. |
| 115 | + -foffload-lto |
| 116 | + ) |
| 117 | + set_source_files_properties(${files} PROPERTIES COMPILE_OPTIONS |
| 118 | + "${OMP_COMPILE_OPTIONS}" |
| 119 | + ) |
| 120 | + |
| 121 | + # Enable "declare target" in the source code. |
| 122 | + set_source_files_properties(${files} |
| 123 | + PROPERTIES COMPILE_DEFINITIONS OMP_OFFLOAD_BUILD |
| 124 | + ) |
| 125 | + else() |
| 126 | + message(FATAL_ERROR |
| 127 | + "Flang runtime build is not supported for these compilers:\n" |
| 128 | + "CMAKE_CXX_COMPILER_ID: ${CMAKE_CXX_COMPILER_ID}\n" |
| 129 | + "CMAKE_C_COMPILER_ID: ${CMAKE_C_COMPILER_ID}") |
| 130 | + endif() |
| 131 | + endif() |
| 132 | +endmacro() |
0 commit comments