mirror of
https://github.com/pjreddie/darknet.git
synced 2023-08-10 21:13:14 +03:00
automatically detect CUDA and use it, without any possible manual intervention
This commit is contained in:
Stefano Sinigardi
@ -1,5 +1,4 @@
|
||||
# To use your libs, launch cmake with -DUSE_INTEGRATED_LIBS:BOOL=FALSE
|
||||
# To enable CUDA, launch cmake with -DENABLE_CUDA. Also it is advisable to set the highest compute model with -DCOMPUTE_MODEL=..., depending on your specific card capabilities. If CUDA is not required, this setting is completely ignored.
|
||||
|
||||
cmake_minimum_required(VERSION 3.8 FATAL_ERROR)
|
||||
|
||||
@ -17,12 +16,13 @@ else()
|
||||
set(USE_INTEGRATED_LIBS "FALSE" CACHE BOOL "Use libs distributed with this repo")
|
||||
endif()
|
||||
|
||||
set(ENABLE_CUDA "FALSE" CACHE BOOL "Enable building with CUDA")
|
||||
|
||||
enable_language(C)
|
||||
enable_language(CXX)
|
||||
if(ENABLE_CUDA)
|
||||
enable_language(CUDA) # do not enable after searching for OpenCV package!
|
||||
|
||||
include(CheckLanguage)
|
||||
check_language(CUDA)
|
||||
if(CMAKE_CUDA_COMPILER)
|
||||
enable_language(CUDA)
|
||||
endif()
|
||||
|
||||
set(CMAKE_CXX_STANDARD 11)
|
||||
@ -30,24 +30,24 @@ set(CMAKE_MODULE_PATH "${CMAKE_CURRENT_LIST_DIR}/cmake/Modules/" ${CMAKE_MODULE_
|
||||
include_directories(${CMAKE_CURRENT_LIST_DIR}/include)
|
||||
include_directories(${CMAKE_CURRENT_LIST_DIR}/src)
|
||||
|
||||
set(COMPUTE_MODEL_75 "75") #Compute capability for Tesla T4, RTX 2080
|
||||
set(COMPUTE_MODEL_72 "72") #Compute capability for Jetson Xavier
|
||||
set(COMPUTE_MODEL_70 "70") #Compute capability for Tesla V100
|
||||
set(COMPUTE_MODEL_62 "62") #Compute capability for Jetson TX2
|
||||
set(COMPUTE_MODEL_61 "61") #Compute capability for Tesla P40
|
||||
set(COMPUTE_MODEL_60 "60") #Compute capability for Tesla P100
|
||||
set(COMPUTE_MODEL_53 "53") #Compute capability for Jetson TX1
|
||||
set(COMPUTE_MODEL_52 "52") #Compute capability for Tesla M40/M60
|
||||
set(COMPUTE_MODEL_37 "37") #Compute capability for Tesla K80
|
||||
set(COMPUTE_MODEL_35 "35") #Compute capability for Tesla K20/K40
|
||||
set(COMPUTE_MODEL_30 "30") #Compute capability for Tesla K10, Quadro K4000
|
||||
set(CUDA_COMPUTE_MODEL_75 "75") #Compute capability for Tesla T4, RTX 2080
|
||||
set(CUDA_COMPUTE_MODEL_72 "72") #Compute capability for Jetson Xavier
|
||||
set(CUDA_COMPUTE_MODEL_70 "70") #Compute capability for Tesla V100
|
||||
set(CUDA_COMPUTE_MODEL_62 "62") #Compute capability for Jetson TX2
|
||||
set(CUDA_COMPUTE_MODEL_61 "61") #Compute capability for Tesla P40
|
||||
set(CUDA_COMPUTE_MODEL_60 "60") #Compute capability for Tesla P100
|
||||
set(CUDA_COMPUTE_MODEL_53 "53") #Compute capability for Jetson TX1
|
||||
set(CUDA_COMPUTE_MODEL_52 "52") #Compute capability for Tesla M40/M60
|
||||
set(CUDA_COMPUTE_MODEL_37 "37") #Compute capability for Tesla K80
|
||||
set(CUDA_COMPUTE_MODEL_35 "35") #Compute capability for Tesla K20/K40
|
||||
set(CUDA_COMPUTE_MODEL_30 "30") #Compute capability for Tesla K10, Quadro K4000
|
||||
|
||||
set(default_compute_model ${COMPUTE_MODEL_30})
|
||||
if(NOT COMPUTE_MODEL)
|
||||
message(STATUS "Setting the compute model to '${default_compute_model}' as none was specified.")
|
||||
set(COMPUTE_MODEL "${default_compute_model}" CACHE
|
||||
set(default_cuda_compute_model ${CUDA_COMPUTE_MODEL_30})
|
||||
if(CMAKE_CUDA_COMPILER AND NOT CUDA_COMPUTE_MODEL)
|
||||
message(STATUS "Setting the compute model to ${default_cuda_compute_model} as none was specified.")
|
||||
set(CUDA_COMPUTE_MODEL "${default_cuda_compute_model}" CACHE
|
||||
STRING "Choose the best compute model supported by your GPU" FORCE)
|
||||
set_property(CACHE COMPUTE_MODEL PROPERTY STRINGS ${COMPUTE_MODEL_75} ${COMPUTE_MODEL_72} ${COMPUTE_MODEL_70} ${COMPUTE_MODEL_62} ${COMPUTE_MODEL_61} ${COMPUTE_MODEL_60} ${COMPUTE_MODEL_53} ${COMPUTE_MODEL_52} ${COMPUTE_MODEL_37} ${COMPUTE_MODEL_35} ${COMPUTE_MODEL_30})
|
||||
set_property(CACHE CUDA_COMPUTE_MODEL PROPERTY STRINGS ${CUDA_COMPUTE_MODEL_75} ${CUDA_COMPUTE_MODEL_72} ${CUDA_COMPUTE_MODEL_70} ${CUDA_COMPUTE_MODEL_62} ${CUDA_COMPUTE_MODEL_61} ${CUDA_COMPUTE_MODEL_60} ${CUDA_COMPUTE_MODEL_53} ${CUDA_COMPUTE_MODEL_52} ${CUDA_COMPUTE_MODEL_37} ${CUDA_COMPUTE_MODEL_35} ${CUDA_COMPUTE_MODEL_30})
|
||||
endif()
|
||||
|
||||
if(USE_INTEGRATED_LIBS)
|
||||
@ -108,17 +108,17 @@ if(OpenCV_FOUND)
|
||||
include_directories(${OpenCV_INCLUDE_DIRS})
|
||||
endif()
|
||||
|
||||
if(ENABLE_CUDA)
|
||||
if(CMAKE_CUDA_COMPILER)
|
||||
add_definitions(-DGPU)
|
||||
if(COMPUTE_MODEL VERSION_GREATER_EQUAL 75 AND CMAKE_CUDA_COMPILER_VERSION VERSION_LESS_EQUAL 9)
|
||||
if(CUDA_COMPUTE_MODEL VERSION_GREATER_EQUAL 75 AND CMAKE_CUDA_COMPILER_VERSION VERSION_LESS_EQUAL 9)
|
||||
message(STATUS "CUDA 10 is required for Compute Model 7.5 - Lowering settings to 7.0")
|
||||
set(COMPUTE_MODEL "70")
|
||||
set(CUDA_COMPUTE_MODEL ${CUDA_COMPUTE_MODEL_70})
|
||||
endif()
|
||||
find_package(CUDNN)
|
||||
if(CUDNN_FOUND)
|
||||
add_definitions(-DCUDNN)
|
||||
include_directories(${CUDNN_INCLUDE_DIR})
|
||||
if(COMPUTE_MODEL VERSION_GREATER_EQUAL 53)
|
||||
if(CUDA_COMPUTE_MODEL VERSION_GREATER_EQUAL 53)
|
||||
add_definitions(-DCUDNN_HALF)
|
||||
message(STATUS "Enabled CUDNN_HALF")
|
||||
endif()
|
||||
@ -126,29 +126,29 @@ if(ENABLE_CUDA)
|
||||
if (MSVC)
|
||||
if(CUDNN_FOUND)
|
||||
if(OpenCV_FOUND)
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${COMPUTE_MODEL},code=[sm_${COMPUTE_MODEL},compute_${COMPUTE_MODEL}] -Wno-deprecated-declarations -Xcompiler=\"/wd4028,/wd4190,/wd4244,/wd4267,/wd4305,/wd4477,/wd4996,/DGPU,/DCUDNN,/DOPENCV\" ${CMAKE_CUDA_FLAGS}")
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${CUDA_COMPUTE_MODEL},code=[sm_${CUDA_COMPUTE_MODEL},compute_${CUDA_COMPUTE_MODEL}] -Wno-deprecated-declarations -Xcompiler=\"/wd4028,/wd4190,/wd4244,/wd4267,/wd4305,/wd4477,/wd4996,/DGPU,/DCUDNN,/DOPENCV\" ${CMAKE_CUDA_FLAGS}")
|
||||
else()
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${COMPUTE_MODEL},code=[sm_${COMPUTE_MODEL},compute_${COMPUTE_MODEL}] -Wno-deprecated-declarations -Xcompiler=\"/wd4028,/wd4190,/wd4244,/wd4267,/wd4305,/wd4477,/wd4996,/DGPU,/DCUDNN\" ${CMAKE_CUDA_FLAGS}")
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${CUDA_COMPUTE_MODEL},code=[sm_${CUDA_COMPUTE_MODEL},compute_${CUDA_COMPUTE_MODEL}] -Wno-deprecated-declarations -Xcompiler=\"/wd4028,/wd4190,/wd4244,/wd4267,/wd4305,/wd4477,/wd4996,/DGPU,/DCUDNN\" ${CMAKE_CUDA_FLAGS}")
|
||||
endif()
|
||||
else()
|
||||
if(OpenCV_FOUND)
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${COMPUTE_MODEL},code=[sm_${COMPUTE_MODEL},compute_${COMPUTE_MODEL}] -Wno-deprecated-declarations -Xcompiler=\"/wd4028,/wd4190,/wd4244,/wd4267,/wd4305,/wd4477,/wd4996,/DGPU,/DOPENCV\" ${CMAKE_CUDA_FLAGS}")
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${CUDA_COMPUTE_MODEL},code=[sm_${CUDA_COMPUTE_MODEL},compute_${CUDA_COMPUTE_MODEL}] -Wno-deprecated-declarations -Xcompiler=\"/wd4028,/wd4190,/wd4244,/wd4267,/wd4305,/wd4477,/wd4996,/DGPU,/DOPENCV\" ${CMAKE_CUDA_FLAGS}")
|
||||
else()
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${COMPUTE_MODEL},code=[sm_${COMPUTE_MODEL},compute_${COMPUTE_MODEL}] -Wno-deprecated-declarations -Xcompiler=\"/wd4028,/wd4190,/wd4244,/wd4267,/wd4305,/wd4477,/wd4996,/DGPU\" ${CMAKE_CUDA_FLAGS}")
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${CUDA_COMPUTE_MODEL},code=[sm_${CUDA_COMPUTE_MODEL},compute_${CUDA_COMPUTE_MODEL}] -Wno-deprecated-declarations -Xcompiler=\"/wd4028,/wd4190,/wd4244,/wd4267,/wd4305,/wd4477,/wd4996,/DGPU\" ${CMAKE_CUDA_FLAGS}")
|
||||
endif()
|
||||
endif()
|
||||
else()
|
||||
if(CUDNN_FOUND)
|
||||
if(OpenCV_FOUND)
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${COMPUTE_MODEL},code=[sm_${COMPUTE_MODEL},compute_${COMPUTE_MODEL}] --compiler-options \"-Wall -Wno-unused-result -Wno-unknown-pragmas -Wfatal-errors -Wno-deprecated-declarations -fPIC -fopenmp -Ofast -DOPENCV -DGPU -DCUDNN \" ${CMAKE_CUDA_FLAGS}")
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${CUDA_COMPUTE_MODEL},code=[sm_${CUDA_COMPUTE_MODEL},compute_${CUDA_COMPUTE_MODEL}] --compiler-options \"-Wall -Wno-unused-result -Wno-unknown-pragmas -Wfatal-errors -Wno-deprecated-declarations -fPIC -fopenmp -Ofast -DOPENCV -DGPU -DCUDNN \" ${CMAKE_CUDA_FLAGS}")
|
||||
else()
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${COMPUTE_MODEL},code=[sm_${COMPUTE_MODEL},compute_${COMPUTE_MODEL}] --compiler-options \"-Wall -Wno-unused-result -Wno-unknown-pragmas -Wfatal-errors -Wno-deprecated-declarations -fPIC -fopenmp -Ofast -DGPU -DCUDNN \" ${CMAKE_CUDA_FLAGS}")
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${CUDA_COMPUTE_MODEL},code=[sm_${CUDA_COMPUTE_MODEL},compute_${CUDA_COMPUTE_MODEL}] --compiler-options \"-Wall -Wno-unused-result -Wno-unknown-pragmas -Wfatal-errors -Wno-deprecated-declarations -fPIC -fopenmp -Ofast -DGPU -DCUDNN \" ${CMAKE_CUDA_FLAGS}")
|
||||
endif()
|
||||
else()
|
||||
if(OpenCV_FOUND)
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${COMPUTE_MODEL},code=[sm_${COMPUTE_MODEL},compute_${COMPUTE_MODEL}] --compiler-options \"-Wall -Wno-unused-result -Wno-unknown-pragmas -Wfatal-errors -Wno-deprecated-declarations -fPIC -fopenmp -Ofast -DOPENCV -DGPU \" ${CMAKE_CUDA_FLAGS}")
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${CUDA_COMPUTE_MODEL},code=[sm_${CUDA_COMPUTE_MODEL},compute_${CUDA_COMPUTE_MODEL}] --compiler-options \"-Wall -Wno-unused-result -Wno-unknown-pragmas -Wfatal-errors -Wno-deprecated-declarations -fPIC -fopenmp -Ofast -DOPENCV -DGPU \" ${CMAKE_CUDA_FLAGS}")
|
||||
else()
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${COMPUTE_MODEL},code=[sm_${COMPUTE_MODEL},compute_${COMPUTE_MODEL}] --compiler-options \"-Wall -Wno-unused-result -Wno-unknown-pragmas -Wfatal-errors -Wno-deprecated-declarations -fPIC -fopenmp -Ofast -DGPU \" ${CMAKE_CUDA_FLAGS}")
|
||||
set(CMAKE_CUDA_FLAGS "-gencode arch=compute_${CUDA_COMPUTE_MODEL},code=[sm_${CUDA_COMPUTE_MODEL},compute_${CUDA_COMPUTE_MODEL}] --compiler-options \"-Wall -Wno-unused-result -Wno-unknown-pragmas -Wfatal-errors -Wno-deprecated-declarations -fPIC -fopenmp -Ofast -DGPU \" ${CMAKE_CUDA_FLAGS}")
|
||||
endif()
|
||||
endif()
|
||||
endif()
|
||||
@ -195,7 +195,7 @@ list(REMOVE_ITEM sources
|
||||
${CMAKE_CURRENT_LIST_DIR}/src/getopt.c
|
||||
)
|
||||
|
||||
if(ENABLE_CUDA)
|
||||
if(CMAKE_CUDA_COMPILER)
|
||||
file(GLOB cuda_sources "${CMAKE_CURRENT_LIST_DIR}/src/*.cu")
|
||||
endif()
|
||||
|
||||
@ -215,7 +215,7 @@ set_source_files_properties(${sources} PROPERTIES LANGUAGE CXX)
|
||||
|
||||
add_library(darklib SHARED ${CMAKE_CURRENT_LIST_DIR}/include/yolo_v2_class.hpp ${CMAKE_CURRENT_LIST_DIR}/src/yolo_v2_class.cpp ${sources} ${headers} ${cuda_sources})
|
||||
set_target_properties(darklib PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||
if(ENABLE_CUDA)
|
||||
if(CMAKE_CUDA_COMPILER)
|
||||
set_target_properties(darklib PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
|
||||
endif()
|
||||
target_compile_definitions(darklib PRIVATE LIB_EXPORTS=1)
|
||||
@ -264,7 +264,7 @@ endif()
|
||||
target_link_libraries(darknet Threads::Threads)
|
||||
target_link_libraries(darklib Threads::Threads)
|
||||
|
||||
if(ENABLE_CUDA)
|
||||
if(CMAKE_CUDA_COMPILER)
|
||||
target_link_libraries(darknet curand cublas)
|
||||
target_link_libraries(darklib curand cublas)
|
||||
endif()
|
||||
|
||||
Reference in New Issue
Block a user