chenchenghao/DeepGEMM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add more GPU architectures support (#112)

Browse Source 
* Add more GPU architectures support

* Update layout.py

* Optimize performance, Add SM90 support, Add 1D2D SM100 support

* Add fmtlib submodule at commit 553ec11

---------

Co-authored-by: fzyzcjy <5236035+fzyzcjy@users.noreply.github.com>
This commit is contained in:
Ray Wang
2025-07-18 11:32:22 +08:00
committed by GitHub
co-authored by fzyzcjy
parent 03d0be3d2d
commit 9da4a23561
67 changed files with 5586 additions and 2965 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CMakeLists.txt
+22 -33
View File
@@ -1,44 +1,33 @@
# 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)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -O3 -fPIC -Wno-psabi")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3 -fPIC -Wno-psabi")
set(CUDA_SEPARABLE_COMPILATION ON)
list(APPEND CUDA_NVCC_FLAGS "-DENABLE_FAST_DEBUG")
list(APPEND CUDA_NVCC_FLAGS "-O3")
list(APPEND CUDA_NVCC_FLAGS "--ptxas-options=--verbose,--register-usage-level=10,--warn-on-local-memory-usage")
set(USE_SYSTEM_NVTX on)
set(CUDA_ARCH_LIST "9.0" CACHE STRING "List of CUDA architectures to compile")
set(TORCH_CUDA_ARCH_LIST "${CUDA_ARCH_LIST}")
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 ${CUDA_TOOLKIT_ROOT_DIR}/lib)
set(CMAKE_CXX_STANDARD 20)
set(CMAKE_CUDA_STANDARD 20)
set(CMAKE_C_FLAGS "${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 "${CUDA_NVCC_FLAGS} -O3 -std=c++17 -DNDEBUG --ptxas-options=--register-usage-level=10")
include_directories(deep_gemm/include third-party/cutlass/include third-party/cutlass/tools/util/include third-party/fmt/include)
include_directories(${CUDA_TOOLKIT_ROOT_DIR}/targets/x86_64-linux/include ${TORCH_INCLUDE_DIRS} ${PYTHON_INCLUDE_DIRS})
link_directories(${TORCH_INSTALL_PREFIX}/lib ${CUDA_TOOLKIT_ROOT_DIR}/lib64 ${CUDA_TOOLKIT_ROOT_DIR}/lib64/stubs)
cuda_add_library(example_gemm STATIC indexing/main.cu)
# The main Python API entrance
pybind11_add_module(deep_gemm_cpp csrc/python_api.cpp)
target_link_libraries(deep_gemm_cpp PRIVATE ${TORCH_LIBRARIES} torch_python cuda)
# Enable kernel code indexing with CMake-based IDEs
cuda_add_library(deep_gemm_indexing_cuda STATIC csrc/indexing/main.cu)