Skip to content

cmake : pass CPU architecture flags to nvcc #5146

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Jan 26, 2024
Merged

cmake : pass CPU architecture flags to nvcc #5146

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
7105287
cmake : pass CPU architecture flags to nvcc
cebtenzzre Jan 23, 2024
bc0192f
fix appending to CUDA_FLAGS
cebtenzzre Jan 26, 2024
a5d7765
cmake : pass ARCH_FLAGS to -Xcompiler on MSVC
cebtenzzre Jan 26, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
74 changes: 39 additions & 35 deletions CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -466,17 +466,17 @@ function(get_flags CCID CCVER)
(CCID STREQUAL "Clang" AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
(CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
)
set(C_FLAGS ${C_FLAGS} -Wdouble-promotion)
list(APPEND C_FLAGS -Wdouble-promotion)
endif()
elseif (CCID STREQUAL "GNU")
set(C_FLAGS -Wdouble-promotion)
set(CXX_FLAGS -Wno-array-bounds)

if (CCVER VERSION_GREATER_EQUAL 7.1.0)
set(CXX_FLAGS ${CXX_FLAGS} -Wno-format-truncation)
list(APPEND CXX_FLAGS -Wno-format-truncation)
endif()
if (CCVER VERSION_GREATER_EQUAL 8.1.0)
set(CXX_FLAGS ${CXX_FLAGS} -Wextra-semi)
list(APPEND CXX_FLAGS -Wextra-semi)
endif()
elseif (CCID MATCHES "Intel")
# enable max optimization level when using Intel compiler
Expand Down Expand Up @@ -510,16 +510,18 @@ if (LLAMA_ALL_WARNINGS)
endif()
endif()

set(CUDA_CXX_FLAGS "")

if (LLAMA_CUBLAS)
set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
if (NOT MSVC)
set(CUDA_FLAGS ${CUDA_FLAGS} -Wno-pedantic)
list(APPEND CUDA_FLAGS -Wno-pedantic)
endif()

if (LLAMA_ALL_WARNINGS AND NOT MSVC)
set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
set(NVCC_CMD ${NVCC_CMD} -ccbin ${CMAKE_CUDA_HOST_COMPILER})
list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
endif()

execute_process(
Expand Down Expand Up @@ -547,13 +549,8 @@ if (LLAMA_CUBLAS)
message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")

get_flags(${CUDA_CCID} ${CUDA_CCVER})
list(JOIN GF_CXX_FLAGS " " CUDA_CXX_FLAGS) # pass host compiler flags as a single argument
if (NOT CUDA_CXX_FLAGS STREQUAL "")
set(CUDA_FLAGS ${CUDA_FLAGS} -Xcompiler ${CUDA_CXX_FLAGS})
endif()
list(APPEND CUDA_CXX_FLAGS ${GF_CXX_FLAGS}) # This is passed to -Xcompiler later
endif()

add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
endif()

if (WIN32)
Expand Down Expand Up @@ -618,12 +615,7 @@ if (NOT MSVC)
endif()
endif()

function(add_compile_option_cpp ARG)
# Adds a compile option to C/C++ only, but not for Cuda.
# Use, e.g., for CPU-architecture flags.
add_compile_options($<$<COMPILE_LANGUAGE:CXX>:${ARG}>)
add_compile_options($<$<COMPILE_LANGUAGE:C>:${ARG}>)
endfunction()
set(ARCH_FLAGS "")

if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64"))
message(STATUS "ARM detected")
Expand All @@ -636,19 +628,19 @@ if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATC
else()
check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
add_compile_options(-mfp16-format=ieee)
list(APPEND ARCH_FLAGS -mfp16-format=ieee)
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
# Raspberry Pi 1, Zero
add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access)
list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
# Raspberry Pi 2
add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
# Raspberry Pi 3, 4, Zero 2 (32-bit)
add_compile_options(-mno-unaligned-access)
list(APPEND ARCH_FLAGS -mno-unaligned-access)
endif()
endif()
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" )
Expand All @@ -659,7 +651,7 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE
include(cmake/FindSIMD.cmake)
endif ()
if (LLAMA_AVX512)
add_compile_option_cpp(/arch:AVX512)
list(APPEND ARCH_FLAGS /arch:AVX512)
# MSVC has no compile-time flags enabling specific
# AVX512 extensions, neither it defines the
# macros corresponding to the extensions.
Expand All @@ -673,49 +665,61 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
endif()
elseif (LLAMA_AVX2)
add_compile_option_cpp(/arch:AVX2)
list(APPEND ARCH_FLAGS /arch:AVX2)
elseif (LLAMA_AVX)
add_compile_option_cpp(/arch:AVX)
list(APPEND ARCH_FLAGS /arch:AVX)
endif()
else()
if (LLAMA_NATIVE)
add_compile_option_cpp(-march=native)
list(APPEND ARCH_FLAGS -march=native)
endif()
if (LLAMA_F16C)
add_compile_option_cpp(-mf16c)
list(APPEND ARCH_FLAGS -mf16c)
endif()
if (LLAMA_FMA)
add_compile_option_cpp(-mfma)
list(APPEND ARCH_FLAGS -mfma)
endif()
if (LLAMA_AVX)
add_compile_option_cpp(-mavx)
list(APPEND ARCH_FLAGS -mavx)
endif()
if (LLAMA_AVX2)
add_compile_option_cpp(-mavx2)
list(APPEND ARCH_FLAGS -mavx2)
endif()
if (LLAMA_AVX512)
add_compile_option_cpp(-mavx512f)
add_compile_option_cpp(-mavx512bw)
list(APPEND ARCH_FLAGS -mavx512f)
list(APPEND ARCH_FLAGS -mavx512bw)
endif()
if (LLAMA_AVX512_VBMI)
add_compile_option_cpp(-mavx512vbmi)
list(APPEND ARCH_FLAGS -mavx512vbmi)
endif()
if (LLAMA_AVX512_VNNI)
add_compile_option_cpp(-mavx512vnni)
list(APPEND ARCH_FLAGS -mavx512vnni)
endif()
endif()
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
message(STATUS "PowerPC detected")
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
add_compile_options(-mcpu=powerpc64le)
list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
else()
add_compile_options(-mcpu=native -mtune=native)
list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
#TODO: Add targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
endif()
else()
message(STATUS "Unknown architecture")
endif()

add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
add_compile_options("$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")

if (LLAMA_CUBLAS)
list(APPEND CUDA_CXX_FLAGS ${ARCH_FLAGS})
list(JOIN CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED) # pass host compiler flags as a single argument
if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
endif()
add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
endif()

if (MINGW)
# Target Windows 8 for PrefetchVirtualMemory
add_compile_definitions(_WIN32_WINNT=${LLAMA_WIN_VER})
Expand Down