mirror of
https://github.com/deepseek-ai/DeepGEMM
synced 2025-05-06 12:14:22 +00:00
45 lines
1.8 KiB
CMake
45 lines
1.8 KiB
CMake
# NOTES: current just for CMake-based IDE (e.g. CLion) indexing, the real compilation is done via JIT
|
|
# TODO: add CUDA utils' library via CMake
|
|
cmake_minimum_required(VERSION 3.10)
|
|
project(deep_gemm LANGUAGES CXX CUDA)
|
|
|
|
set(CMAKE_CXX_STANDARD 20)
|
|
set(CMAKE_CUDA_STANDARD 20)
|
|
set(CMAKE_VERBOSE_MAKEFILE ON)
|
|
|
|
find_package(CUDAToolkit REQUIRED)
|
|
find_package(pybind11 REQUIRED)
|
|
|
|
file(WRITE ${CMAKE_BINARY_DIR}/test_cuda.cu "extern \"C\" __global__ void testKernel() { }")
|
|
execute_process(
|
|
COMMAND ${CUDA_NVCC_EXECUTABLE} ${CMAKE_CUDA_FLAGS} -gencode arch=compute_90a,code=sm_90a -o ${CMAKE_BINARY_DIR}/test_cuda.o -c ${CMAKE_BINARY_DIR}/test_cuda.cu
|
|
RESULT_VARIABLE NVCC_RESULT
|
|
OUTPUT_VARIABLE NVCC_OUTPUT
|
|
ERROR_VARIABLE NVCC_ERROR_OUTPUT
|
|
WORKING_DIRECTORY ${CMAKE_BINARY_DIR}
|
|
)
|
|
|
|
if (NVCC_RESULT EQUAL "0")
|
|
set(NVCC_SUPPORTS_SM90 TRUE)
|
|
message(STATUS "NVCC supports SM90")
|
|
else()
|
|
message(STATUS "NVCC does not support SM90")
|
|
endif()
|
|
|
|
if (NVCC_SUPPORTS_SM90)
|
|
set(TORCH_CUDA_ARCH_LIST "8.6" CACHE STRING "Add arch tag 90a to NVCC" FORCE)
|
|
list(APPEND CUDA_NVCC_FLAGS "-gencode;arch=compute_90a,code=sm_90a")
|
|
endif()
|
|
find_package(Torch REQUIRED)
|
|
|
|
include_directories(deep_gemm/include third-party/cutlass/include third-party/cutlass/tools/util/include)
|
|
include_directories(${CUDA_TOOLKIT_ROOT_DIR}/include ${TORCH_INCLUDE_DIRS} ${PYTHON_INCLUDE_DIRS})
|
|
link_directories(${TORCH_INSTALL_PREFIX}/lib ${CUDA_TOOLKIT_ROOT_DIR}/lib)
|
|
|
|
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -O3 -fPIC")
|
|
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3 -fPIC")
|
|
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -O3 -fPIC -DNDEBUG")
|
|
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -O3 -std=c++17 -DNDEBUG --ptxas-options=--register-usage-level=10")
|
|
|
|
cuda_add_library(example_gemm STATIC indexing/main.cu)
|