Qt Creator cuda动态并行工程配置

动态并行允许从GPU端启动核函数，它要求显卡计算能力大于3.5.

动态并行Qt Creator的工程配置与不使用动态并行时有较大差别，以linux系统为例

QT    += core
greaterThan(QT_MAJOR_VERSION, 4): QT += widgets
TARGET = test
TEMPLATE = app
DEFINES += QT_DEPRECATED_WARNINGS

INCLUDEPATH +=/usr/local/cuda-10.1/include 

SOURCES += main.cpp 
HEADERAS +=algorithm.h

LIBS +=-L/usr/local/cuda-10.1/lib64 \       
       -lcublas -lcuda -lcudadevrt \      
       -lcudart -lcudart_static -lcufft \       
       -lcufftw -lcurand -lcusolver -lcusparse

OTHER_FILES +=./algorithm.cu
CUDA_SOURCES+=./algorithm.cu
CUDA_SDK ="/usr/local/cuda-10.1"   
CUDA_DIR ="/usr/local/cuda-10.1"
QMAKE_LIBDIR += $$CUDA_DIR/lib64
SYSTEM_TYPE = 64
CUDA_ARCH = sm_60
NVCCFLAGS     = --use_fast_math
CUDA_INC = $$join(INCLUDEPATH,'" -I"','-I"','"')
CUDA_LIBS= -L/usr/local/cuda-10.1/lib64 \       
           -lcublas -lcuda -lcudadevrt \       
           -lcudart -lcudart_static -lcufft \       
           -lcufftw -lcurand -lcusolver -lcusparse

CUDA_OBJECTS_DIR = ./
#重点 为了支持动态并行化，使用两步分离编译和链接的过程
cudaIntr.input = CUDA_SOURCES
cudaIntr.output = $$CUDA_OBJECTS_DIR/${QMAKE_FILE_BASE}.o
cudaIntr.commands = $$CUDA_DIR/bin/nvcc $$NVCC_OPTIONS $$CUDA_INC $$CUDA_LIBS --machine $$SYSTEM_TYPE \                  
                    -arch=$$CUDA_ARCH -std=c++11 -rdc=true -c -o ${QMAKE_FILE_OUT} ${QMAKE_FILE_NAME}cudaIntr.dependency_type = TYPE_C
cudaIntr.variable_out = CUDA_OBJ
cudaIntr.variable_out += OBJECTS
cudaIntr.clean = cudaIntrObj/*.o
QMAKE_EXTRA_COMPILERS += cudaIntr

cuda.input = CUDA_OBJ
cuda.output = $$CUDA_OBJECTS_DIR/${QMAKE_FILE_BASE}_link.o
cuda.commands = $$CUDA_DIR/bin/nvcc -arch=$$CUDA_ARCH -std=c++11 -dlink ${QMAKE_FILE_NAME} -o ${QMAKE_FILE_OUT}
cuda.dependency_type = TYPE_C
cuda.depend_command = $$CUDA_DIR/bin/nvcc -g -M $$CUDA_INC $$NVCCFLAGS   ${QMAKE_FILE_NAME}
QMAKE_EXTRA_COMPILERS += cuda

上述目前我也不知道具体原理，是通过网上很多资料搜集整理而成，自己测试是可以运行的，以后懂了再加。