CUTLASS 2.1 (#83)
CUTLASS 2.1 contributes: - BLAS-style host-side API added to CUTLASS Library - Planar Complex GEMM kernels targeting Volta and Turing Tensor Cores - Minor enhancements and bug fixes
This commit is contained in:
Andrew Kerr
Regular → Executable
+134
-101
@@ -32,15 +32,14 @@ endif()
|
||||
|
||||
message(STATUS "CMake Version: ${CMAKE_VERSION}")
|
||||
|
||||
project(CUTLASS VERSION 2.0.0 LANGUAGES CXX)
|
||||
project(CUTLASS VERSION 2.1.0 LANGUAGES CXX)
|
||||
include(${CMAKE_CURRENT_SOURCE_DIR}/CUDA.cmake)
|
||||
|
||||
find_package(Doxygen QUIET)
|
||||
|
||||
#
|
||||
# CUTLASS 2.0 requires C++11
|
||||
# CUTLASS 2.x requires C++11
|
||||
#
|
||||
|
||||
set(CMAKE_CXX_STANDARD 11)
|
||||
set(CMAKE_CXX_STANDARD_REQUIRED ON)
|
||||
set(CMAKE_CXX_EXTENSIONS OFF)
|
||||
@@ -49,7 +48,7 @@ if(CUTLASS_NATIVE_CUDA)
|
||||
set(CMAKE_CUDA_STANDARD 11)
|
||||
set(CMAKE_CUDA_STANDARD_REQUIRED ON)
|
||||
else()
|
||||
string(APPEND NVCC_FLAGS " --std=c++11")
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS --std=c++11)
|
||||
endif()
|
||||
|
||||
if(CMAKE_INSTALL_PREFIX_INITIALIZED_TO_DEFAULT)
|
||||
@@ -58,13 +57,26 @@ endif()
|
||||
|
||||
message(STATUS "Default Install Location: ${CMAKE_INSTALL_PREFIX}")
|
||||
|
||||
if(${CMAKE_PROJECT_NAME} MATCHES ${PROJECT_NAME})
|
||||
set(_CUTLASS_ENABLE_TESTS ON)
|
||||
set(CUTLASS_ENABLE_HEADERS_ONLY OFF CACHE BOOL "Enable only the header library")
|
||||
|
||||
if(CUTLASS_ENABLE_HEADERS_ONLY)
|
||||
set(CUTLASS_ENABLE_EXAMPLES_INIT OFF)
|
||||
set(CUTLASS_ENABLE_TOOLS_INIT OFF)
|
||||
else()
|
||||
set(_CUTLASS_ENABLE_TESTS OFF)
|
||||
set(CUTLASS_ENABLE_EXAMPLES_INIT ON)
|
||||
set(CUTLASS_ENABLE_TOOLS_INIT ON)
|
||||
endif()
|
||||
|
||||
set(CUTLASS_ENABLE_TESTS ${_CUTLASS_ENABLE_TESTS} CACHE BOOL "Enable CUTLASS Tests")
|
||||
set(CUTLASS_ENABLE_EXAMPLES ${CUTLASS_ENABLE_EXAMPLES_INIT} CACHE BOOL "Enable CUTLASS Examples")
|
||||
set(CUTLASS_ENABLE_TOOLS ${CUTLASS_ENABLE_TOOLS_INIT} CACHE BOOL "Enable CUTLASS Tools")
|
||||
|
||||
if(${CMAKE_PROJECT_NAME} STREQUAL ${PROJECT_NAME})
|
||||
set(CUTLASS_ENABLE_TESTS_INIT ${CUTLASS_ENABLE_TOOLS_INIT})
|
||||
else()
|
||||
set(CUTLASS_ENABLE_TESTS_INIT OFF)
|
||||
endif()
|
||||
|
||||
set(CUTLASS_ENABLE_TESTS ${CUTLASS_ENABLE_TESTS_INIT} CACHE BOOL "Enable CUTLASS Tests")
|
||||
|
||||
if (CUTLASS_ENABLE_TESTS)
|
||||
include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/googletest.cmake)
|
||||
@@ -86,31 +98,22 @@ endif()
|
||||
if (NOT CUDA_VERSION VERSION_LESS 10.0)
|
||||
list(APPEND CUTLASS_NVCC_ARCHS_SUPPORTED 75)
|
||||
endif()
|
||||
|
||||
if(CUDA_COMPILER MATCHES "[Cc]lang")
|
||||
if(NOT CMAKE_CXX_COMPILER_ID MATCHES "Clang" )
|
||||
message(FATAL_ERROR "Clang CUDA compilation requires Clang CXX compilation. Currently CMAKE_CXX_COMPILER is ${CMAKE_CXX_COMPILER_ID}" )
|
||||
endif()
|
||||
if (CMAKE_CXX_COMPILER_VERSION VERSION_LESS 7.0)
|
||||
message(FATAL_ERROR "Clang 7.0+ required for GPU compilation")
|
||||
endif()
|
||||
endif()
|
||||
|
||||
set(CUTLASS_NVCC_ARCHS ${CUTLASS_NVCC_ARCHS_SUPPORTED} CACHE STRING "The SM architectures requested.")
|
||||
set(CUTLASS_NVCC_ARCHS_ENABLED ${CUTLASS_NVCC_ARCHS} CACHE STRING "The SM architectures to build code for.")
|
||||
|
||||
# Special policy introduced in CMake 3.13
|
||||
if (POLICY CMP0076)
|
||||
cmake_policy(SET CMP0076 NEW)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
# check if the configuration is supported
|
||||
if(NOT CMAKE_SIZEOF_VOID_P EQUAL 8)
|
||||
if( NOT CMAKE_SIZEOF_VOID_P EQUAL 8 )
|
||||
message(FATAL_ERROR "CUTLASS requires a 64-bit compiler!")
|
||||
endif()
|
||||
|
||||
include(GNUInstallDirs)
|
||||
|
||||
link_directories(${CUDA_TOOLKIT_ROOT_DIR}/lib64/stubs)
|
||||
|
||||
###################################################################################################
|
||||
#
|
||||
# Configure CMake variables
|
||||
@@ -120,11 +123,14 @@ include(GNUInstallDirs)
|
||||
message(STATUS "CUDA Compilation Architectures: ${CUTLASS_NVCC_ARCHS_ENABLED}")
|
||||
|
||||
if (NOT (CMAKE_BUILD_TYPE OR CONFIGURATION_TYPES))
|
||||
# By default we want to build in Release mode to ensure that we're getting best performance.
|
||||
# By default we want to build in Release mode to ensure that we're getting best performance.
|
||||
set(CMAKE_BUILD_TYPE Release CACHE STRING "Choose build level" FORCE)
|
||||
set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "RelWithDebInfo" "Release")
|
||||
endif()
|
||||
|
||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
||||
set(CUTLASS_LIBRARY_DEBUG_POSTFIX ".debug" CACHE STRING "Default postfix value for debug libraries")
|
||||
|
||||
if(WIN32)
|
||||
# On Windows we link against the shared (DLL) runtime. Change gtest settings to match this.
|
||||
set(gtest_force_shared_crt ON CACHE BOOL "Use shared (DLL) run-time lib even when Google Test is built as static lib" FORCE)
|
||||
@@ -132,29 +138,35 @@ endif()
|
||||
|
||||
if (WIN32)
|
||||
# Enable more warnings and treat as errors
|
||||
string(APPEND NVCC_FLAGS " -Xcompiler /W3 -Xcompiler /WX")
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS -Xcompiler=/W3 -Xcompiler=/WX)
|
||||
|
||||
# Disable warning on Unicode characters
|
||||
string(APPEND NVCC_FLAGS " -Xcompiler /wd4819")
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS -Xcompiler=/wd4819)
|
||||
|
||||
# Disable excess x86 floating point precision that can lead to results being labeled incorrectly
|
||||
string(APPEND NVCC_FLAGS " -Xcompiler /fp:strict")
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS -Xcompiler=/fp:strict)
|
||||
endif(WIN32)
|
||||
|
||||
if (${CUTLASS_NVCC_VERBOSE})
|
||||
string(APPEND NVCC_FLAGS " -v")
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS -v)
|
||||
endif()
|
||||
|
||||
set(CUTLASS_NVCC_EMBED_CUBIN ON CACHE BOOL "Embed compiled CUDA kernel binaries into executables.")
|
||||
set(CUTLASS_NVCC_EMBED_PTX ON CACHE BOOL "Embed compiled PTX into executables.")
|
||||
set(CUTLASS_NVCC_KEEP OFF CACHE BOOL "Keep intermediate files generated by NVCC.")
|
||||
set(CUTLASS_ENABLE_F16C ON CACHE BOOL "Enable F16C x86 extensions in host code.")
|
||||
set(CUTLASS_LIBRARY_KERNELS "128x128" CACHE STRING "Comma delimited list of kernel name filters. Default '' means all kernels are enabled.")
|
||||
|
||||
#
|
||||
# CUTLASS generator cmake configuration
|
||||
#
|
||||
set(CUTLASS_LIBRARY_OPERATIONS "all" CACHE STRING "Comma delimited list of operation name filters. Default '' means all operations are enabled.")
|
||||
set(CUTLASS_LIBRARY_KERNELS "" CACHE STRING "Comma delimited list of kernel name filters. If unspecified, only the largest tile size is enabled. If 'all' is specified, all kernels are enabled.")
|
||||
|
||||
|
||||
# Test Levels L0, L1, L2
|
||||
set(CUTLASS_TEST_LEVEL "0" CACHE STRING "Level of tests to compile.")
|
||||
set_property(CACHE CUTLASS_TEST_LEVEL PROPERTY STRINGS 0 1 2)
|
||||
string(APPEND NVCC_FLAGS " -DCUTLASS_TEST_LEVEL=${CUTLASS_TEST_LEVEL}")
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS -DCUTLASS_TEST_LEVEL=${CUTLASS_TEST_LEVEL})
|
||||
|
||||
#
|
||||
# CUDA 10.1 introduces "mma" in PTX performing collective matrix multiply operations.
|
||||
@@ -166,7 +178,7 @@ else()
|
||||
set(CUTLASS_ENABLE_TENSOR_CORE_MMA_DEFAULT ON)
|
||||
endif()
|
||||
|
||||
set(CUTLASS_ENABLE_TENSOR_CORE_MMA ${CUTLASS_ENABLE_TENSOR_CORE_MMA_DEFAULT} CACHE BOOL
|
||||
set(CUTLASS_ENABLE_TENSOR_CORE_MMA ${CUTLASS_ENABLE_TENSOR_CORE_MMA_DEFAULT} CACHE BOOL
|
||||
"Enable PTX mma instruction for collective matrix multiply operations.")
|
||||
|
||||
#
|
||||
@@ -182,7 +194,7 @@ set(CUTLASS_ENABLE_TENSOR_CORE_MMA ${CUTLASS_ENABLE_TENSOR_CORE_MMA_DEFAULT} CAC
|
||||
# ...
|
||||
#
|
||||
if(ENABLE_ASAN) # https://github.com/google/sanitizers/wiki/AddressSanitizer
|
||||
string(APPEND NVCC_FLAGS " --compiler-options -fsanitize=address --compiler-options -fno-omit-frame-pointer")
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS --compiler-options=-fsanitize=address --compiler-options=-fno-omit-frame-pointer)
|
||||
string(APPEND CMAKE_EXE_LINKER_FLAGS " -fsanitize=address")
|
||||
endif()
|
||||
|
||||
@@ -192,85 +204,127 @@ endif()
|
||||
#
|
||||
###################################################################################################
|
||||
|
||||
foreach(ARCH ${CUTLASS_NVCC_ARCHS_ENABLED})
|
||||
if(CUTLASS_NVCC_EMBED_CUBIN)
|
||||
string(APPEND NVCC_GENCODE_FLAGS " -gencode=arch=compute_${ARCH},code=sm_${ARCH}")
|
||||
endif()
|
||||
if(CUTLASS_NVCC_EMBED_PTX)
|
||||
string(APPEND NVCC_GENCODE_FLAGS " -gencode=arch=compute_${ARCH},code=compute_${ARCH}")
|
||||
endif()
|
||||
string(APPEND CLANG_FLAGS " --cuda-gpu-arch=sm_${ARCH}")
|
||||
endforeach()
|
||||
|
||||
if(CUTLASS_NVCC_EMBED_PTX)
|
||||
string(APPEND CLANG_FLAGS " --cuda-include-ptx=all")
|
||||
list(APPEND CUTLASS_CUDA_CLANG_FLAGS --cuda-include-ptx=all)
|
||||
endif()
|
||||
|
||||
if (CUTLASS_ENABLE_TENSOR_CORE_MMA)
|
||||
string(APPEND COMMON_FLAGS " -DCUTLASS_ENABLE_TENSOR_CORE_MMA=1")
|
||||
list(APPEND CUTLASS_CUDA_FLAGS -DCUTLASS_ENABLE_TENSOR_CORE_MMA=1)
|
||||
endif()
|
||||
|
||||
if (NOT MSVC AND CUTLASS_NVCC_KEEP)
|
||||
# MSVC flow handles caching already, but for other generators we handle it here.
|
||||
set(CUTLASS_NVCC_KEEP_DIR ${CMAKE_CURRENT_BINARY_DIR}/tmp CACHE PATH "Location to store NVCC scratch files")
|
||||
file(MAKE_DIRECTORY ${CUTLASS_NVCC_KEEP_DIR})
|
||||
string(APPEND NVCC_FLAGS " --keep") # --keep-dir may not work with nvcc for some directories.
|
||||
string(APPEND CLANG_FLAGS " -save-temps=${CUTLASS_NVCC_KEEP_DIR}")
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS --keep) # --keep-dir may not work with nvcc for some directories.
|
||||
list(APPEND CUTLASS_CUDA_CLANG_FLAGS -save-temps=${CUTLASS_NVCC_KEEP_DIR})
|
||||
endif()
|
||||
|
||||
if (CUTLASS_ENABLE_F16C)
|
||||
string(APPEND COMPILER_FLAGS " -DCUTLASS_ENABLE_F16C=1")
|
||||
if (CUTLASS_ENABLE_F16C AND NOT CMAKE_CROSSCOMPILING)
|
||||
list(APPEND CUTLASS_CUDA_FLAGS -DCUTLASS_ENABLE_F16C=1)
|
||||
if ((CMAKE_CXX_COMPILER_ID MATCHES "GNU") OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang"))
|
||||
string(APPEND NVCC_FLAGS " -Xcompiler -mf16c")
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS -Xcompiler=-mf16c)
|
||||
elseif((CMAKE_CXX_COMPILER_ID MATCHES "MSVC"))
|
||||
string(APPEND NVCC_FLAGS " -Xcompiler /arch:AVX2")
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS -Xcompiler=/arch:AVX2)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
string(APPEND NVCC_FLAGS " -lineinfo")
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS $<$<BOOL:${UNIX}>:-Xcompiler=-Wconversion>)
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS $<$<BOOL:${UNIX}>:-Xcompiler=-fno-strict-aliasing>)
|
||||
|
||||
string(APPEND CLANG_FLAGS " -gmlt")
|
||||
|
||||
if (UNIX)
|
||||
string(APPEND NVCC_FLAGS " -Xcompiler -Wconversion")
|
||||
string(APPEND NVCC_FLAGS " -Xcompiler -fno-strict-aliasing")
|
||||
# Don't leak lineinfo in release builds
|
||||
if (NOT CMAKE_BUILD_TYPE MATCHES "Release")
|
||||
list(APPEND CUTLASS_CUDA_CLANG_FLAGS -gmlt)
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS -lineinfo)
|
||||
endif()
|
||||
|
||||
if(CUDA_COMPILER MATCHES "[Cc]lang")
|
||||
string(APPEND CLANG_FLAGS " --cuda-path=${CUDA_TOOLKIT_ROOT_DIR}")
|
||||
string(APPEND CLANG_FLAGS " -mllvm -pragma-unroll-threshold=100000")
|
||||
string(APPEND CLANG_FLAGS " -mllvm -unroll-threshold=5000")
|
||||
string(APPEND CLANG_FLAGS " -Wno-unused-command-line-argument")
|
||||
if( NOT CMAKE_CXX_COMPILER_ID MATCHES "Clang" )
|
||||
message(FATAL_ERROR "Clang CUDA compilation requires Clang CXX compilation. Currently CMAKE_CXX_COMPILER is ${CMAKE_CXX_COMPILER_ID}" )
|
||||
endif()
|
||||
|
||||
if (CMAKE_CXX_COMPILER_VERSION VERSION_LESS 7.0)
|
||||
message(FATAL_ERROR "Clang 7.0+ required for GPU compilation")
|
||||
endif()
|
||||
|
||||
list(APPEND CUTLASS_CUDA_CLANG_FLAGS --cuda-path=${CUDA_TOOLKIT_ROOT_DIR})
|
||||
list(APPEND CUTLASS_CUDA_CLANG_FLAGS -mllvm=-pragma-unroll-threshold=100000)
|
||||
list(APPEND CUTLASS_CUDA_CLANG_FLAGS -mllvm=-unroll-threshold=5000)
|
||||
list(APPEND CUTLASS_CUDA_CLANG_FLAGS -Wno-unused-command-line-argument)
|
||||
|
||||
string(REPLACE "." ";" CUDA_VERSION_PARTS ${CMAKE_CUDA_COMPILER_VERSION})
|
||||
list(GET CUDA_VERSION_PARTS 0 CUDA_VERSION_MAJOR)
|
||||
list(GET CUDA_VERSION_PARTS 1 CUDA_VERSION_MINOR)
|
||||
string(APPEND CLANG_FLAGS " -D__CUDACC_VER_MAJOR__=${CUDA_VERSION_MAJOR} -D__CUDACC_VER_MINOR__=${CUDA_VERSION_MINOR}")
|
||||
list(APPEND CUTLASS_CUDA_CLANG_FLAGS -D__CUDACC_VER_MAJOR__=${CUDA_VERSION_MAJOR} -D__CUDACC_VER_MINOR__=${CUDA_VERSION_MINOR})
|
||||
|
||||
|
||||
# needed for libcublasLt.so in case it's installed in the same location as libcudart.so
|
||||
# dynamic linker can find it if linker sets RPATH (forced by --disable-new-tags)
|
||||
# Otherwise linker uses RUNPATH and that does not propagate to loaded libs.
|
||||
string(APPEND CLANG_FLAGS " -Wl,--disable-new-dtags")
|
||||
list(APPEND CUTLASS_CUDA_CLANG_FLAGS -Wl,--disable-new-dtags)
|
||||
|
||||
link_libraries(nvidia::cudart)
|
||||
endif()
|
||||
|
||||
if(CUDA_COMPILER MATCHES "[Cc]lang")
|
||||
string(APPEND CMAKE_CXX_FLAGS "${COMMON_FLAGS} ${CLANG_FLAGS}")
|
||||
string(APPEND CMAKE_CXX_FLAGS_RELEASE "${COMMON_FLAGS_RELEASE} ${CLANG_FLAGS_RELEASE}")
|
||||
string(APPEND CMAKE_CXX_FLAGS_RELWITHDEBINFO "${COMMON_FLAGS_RELWITHDEBINFO} ${CLANG_FLAGS_RELWITHDEBINFO}")
|
||||
string(APPEND CMAKE_CXX_FLAGS_DEBUG "${COMMON_FLAGS_DEBUG} ${CLANG_FLAGS_DEBUG}")
|
||||
elseif (CUTLASS_NATIVE_CUDA)
|
||||
string(APPEND CMAKE_CUDA_FLAGS "${COMMON_FLAGS} ${NVCC_FLAGS} ${NVCC_GENCODE_FLAGS}")
|
||||
string(APPEND CMAKE_CUDA_FLAGS_RELEASE "${COMMON_FLAGS_RELEASE} ${NVCC_FLAGS_RELEASE}")
|
||||
string(APPEND CMAKE_CUDA_FLAGS_RELWITHDEBINFO "${COMMON_FLAGS_RELWITHDEBINFO} ${NVCC_FLAGS_RELWITHDEBINFO}")
|
||||
string(APPEND CMAKE_CUDA_FLAGS_DEBUG "${COMMON_FLAGS_DEBUG} ${NVCC_FLAGS_DEBUG}")
|
||||
else()
|
||||
string(APPEND CUDA_NVCC_FLAGS "${COMMON_FLAGS} ${NVCC_FLAGS} ${NVCC_GENCODE_FLAGS}")
|
||||
string(APPEND CUDA_NVCC_FLAGS_RELEASE "${COMMON_FLAGS_RELEASE} ${NVCC_FLAGS_RELEASE}")
|
||||
string(APPEND CUDA_NVCC_FLAGS_RELWITHDEBINFO "${COMMON_FLAGS_RELWITHDEBINFO} ${NVCC_FLAGS_RELWITHDEBINFO}")
|
||||
string(APPEND CUDA_NVCC_FLAGS_DEBUG "${COMMON_FLAGS_DEBUG} ${NVCC_FLAGS_DEBUG}")
|
||||
endif()
|
||||
function(cutlass_apply_cuda_gencode_flags TARGET)
|
||||
|
||||
set(NVCC_FLAGS)
|
||||
set(CLANG_FLAGS)
|
||||
foreach(ARCH ${CUTLASS_NVCC_ARCHS_ENABLED})
|
||||
list(APPEND CUTLASS_CUDA_CLANG_FLAGS --cuda-gpu-arch=sm_${ARCH})
|
||||
set(CODES)
|
||||
if(CUTLASS_NVCC_EMBED_CUBIN)
|
||||
list(APPEND CODES sm_${ARCH})
|
||||
endif()
|
||||
if(CUTLASS_NVCC_EMBED_PTX)
|
||||
list(APPEND CODES compute_${ARCH})
|
||||
endif()
|
||||
list(JOIN CODES "," CODES_STR)
|
||||
list(APPEND NVCC_FLAGS -gencode=arch=compute_${ARCH},code=[${CODES_STR}])
|
||||
endforeach()
|
||||
|
||||
if (CUDA_COMPILER MATCHES "[Cc]lang")
|
||||
target_compile_options(
|
||||
${TARGET}
|
||||
PRIVATE
|
||||
$<$<COMPILE_LANGUAGE:CXX>:${CLANG_FLAGS}>
|
||||
)
|
||||
else()
|
||||
target_compile_options(
|
||||
${TARGET}
|
||||
PRIVATE
|
||||
$<$<COMPILE_LANGUAGE:CUDA>:${NVCC_FLAGS}>
|
||||
)
|
||||
endif()
|
||||
|
||||
endfunction()
|
||||
|
||||
function(cutlass_apply_standard_compile_options TARGET)
|
||||
|
||||
if(CUDA_COMPILER MATCHES "[Cc]lang")
|
||||
set(CUDA_COMPILE_LANGUAGE CXX)
|
||||
set(_FLAGS ${CUTLASS_CUDA_FLAGS} ${CUTLASS_CUDA_CLANG_FLAGS})
|
||||
set(_FLAGS_RELEASE ${CUTLASS_CUDA_FLAGS_RELEASE} ${CUTLASS_CUDA_CLANG_FLAGS_RELEASE})
|
||||
set(_FLAGS_RELWITHDEBINFO ${CUTLASS_CUDA_FLAGS_RELWITHDEBINFO} ${CUTLASS_CUDA_CLANG_FLAGS_RELWITHDEBINFO})
|
||||
set(_FLAGS_DEBUG ${CUTLASS_CUDA_FLAGS_DEBUG} ${CUTLASS_CUDA_CLANG_FLAGS_DEBUG})
|
||||
else()
|
||||
set(CUDA_COMPILE_LANGUAGE CUDA)
|
||||
set(_FLAGS ${CUTLASS_CUDA_FLAGS} ${CUTLASS_CUDA_NVCC_FLAGS})
|
||||
set(_FLAGS_RELEASE ${CUTLASS_CUDA_FLAGS_RELEASE} ${CUTLASS_CUDA_NVCC_FLAGS_RELEASE})
|
||||
set(_FLAGS_RELWITHDEBINFO ${CUTLASS_CUDA_FLAGS_RELWITHDEBINFO} ${CUTLASS_CUDA_NVCC_FLAGS_RELWITHDEBINFO})
|
||||
set(_FLAGS_DEBUG ${CUTLASS_CUDA_FLAGS_DEBUG} ${CUTLASS_CUDA_NVCC_FLAGS_DEBUG})
|
||||
endif()
|
||||
|
||||
target_compile_options(
|
||||
${TARGET}
|
||||
PRIVATE
|
||||
$<$<COMPILE_LANGUAGE:${CUDA_COMPILE_LANGUAGE}>:${_FLAGS}>
|
||||
$<$<COMPILE_LANGUAGE:${CUDA_COMPILE_LANGUAGE}>:$<$<CONFIG:RELEASE>:${_FLAGS_RELEASE}>>
|
||||
$<$<COMPILE_LANGUAGE:${CUDA_COMPILE_LANGUAGE}>:$<$<CONFIG:RELWITHDEBINFO>:${_FLAGS_RELWITHDEBINFO}>>
|
||||
$<$<COMPILE_LANGUAGE:${CUDA_COMPILE_LANGUAGE}>:$<$<CONFIG:DEBUG>:${_FLAGS_DEBUG}>>
|
||||
)
|
||||
|
||||
endfunction()
|
||||
|
||||
#
|
||||
# The following items should eventually be pushed into cutlass/CMakeLists.txt
|
||||
@@ -324,8 +378,8 @@ if (NOT DEFINED CUTLASS_REVISION)
|
||||
endif()
|
||||
|
||||
configure_file(
|
||||
${CMAKE_CURRENT_SOURCE_DIR}/cmake/version.h.in
|
||||
${CMAKE_CURRENT_BINARY_DIR}/include/cutlass/version.h
|
||||
${CMAKE_CURRENT_SOURCE_DIR}/cmake/version.h.in
|
||||
${CMAKE_CURRENT_BINARY_DIR}/include/cutlass/version.h
|
||||
@ONLY)
|
||||
|
||||
target_include_directories(
|
||||
@@ -338,8 +392,8 @@ target_include_directories(
|
||||
)
|
||||
|
||||
install(
|
||||
DIRECTORY
|
||||
${CUTLASS_INCLUDE_DIR}/
|
||||
DIRECTORY
|
||||
${CUTLASS_INCLUDE_DIR}/
|
||||
${CMAKE_CURRENT_BINARY_DIR}/include/
|
||||
DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
|
||||
)
|
||||
@@ -399,27 +453,6 @@ endif()
|
||||
|
||||
################################################################################
|
||||
|
||||
set(CUTLASS_ENABLE_HEADERS_ONLY OFF CACHE BOOL "Enable only the header library")
|
||||
|
||||
if(CUTLASS_ENABLE_HEADERS_ONLY)
|
||||
set(CUTLASS_ENABLE_EXAMPLES_INIT OFF)
|
||||
set(CUTLASS_ENABLE_TOOLS_INIT OFF)
|
||||
else()
|
||||
set(CUTLASS_ENABLE_EXAMPLES_INIT ON)
|
||||
set(CUTLASS_ENABLE_TOOLS_INIT ON)
|
||||
endif()
|
||||
|
||||
set(CUTLASS_ENABLE_EXAMPLES ${CUTLASS_ENABLE_EXAMPLES_INIT} CACHE BOOL "Enable CUTLASS Examples")
|
||||
set(CUTLASS_ENABLE_TOOLS ${CUTLASS_ENABLE_TOOLS_INIT} CACHE BOOL "Enable CUTLASS Tools")
|
||||
|
||||
if(${CMAKE_PROJECT_NAME} STREQUAL ${PROJECT_NAME})
|
||||
set(CUTLASS_ENABLE_TESTS_INIT ${CUTLASS_ENABLE_TOOLS_INIT})
|
||||
else()
|
||||
set(CUTLASS_ENABLE_TESTS_INIT OFF)
|
||||
endif()
|
||||
|
||||
set(CUTLASS_ENABLE_TESTS ${CUTLASS_ENABLE_TESTS_INIT} CACHE BOOL "Enable CUTLASS Tests")
|
||||
|
||||
if(CUTLASS_ENABLE_TOOLS)
|
||||
add_subdirectory(tools)
|
||||
endif()
|
||||
|
||||
Reference in New Issue
Block a user