Yes, you can.
The key is to use separate compilation with device code linking, which is available with nvcc. Since this is already required for usage of dynamic parallelism, there's really nothing new here.
Here's a simple example:
ch_kernel.cu:
#include <stdio.h>
__global__ void ch_kernel(){
printf("hello from child kernel
");
}
main.cu:
#include <stdio.h>
extern __global__ void ch_kernel();
__global__ void kernel(){
ch_kernel<<<1,1>>>();
}
int main(){
kernel<<<1,1>>>();
cudaDeviceSynchronize();
}
compile with:
nvcc -arch=sm_35 -rdc=true -o test ch_kernel.cu main.cu -lcudadevrt
与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…