Undefined reference to `cuMemGetAccess' and `cuMemSetAccess'

[zhoujinzhoujinzhoujin]

Hi,
I am going to test these two functions cuMemGetAccess' and cuMemSetAccess’ but nvcc reports undefined references.
My cuda version is 10.2 and I do see the declarations of the two functions in the header file “/usr/local/cuda-10.2/targets/aarch64-linux/include/cuda.h”. I also added “#include <cuda.h>” and “#include “/usr/local/cuda-10.2/targets/aarch64-linux/include/cuda.h”” but neither solves the issue.
Have I missed anything? Any help is appreciated.

[AastaLLL]

Hi,

Please use #include <cuda.h>.
And add /usr/local/cuda-10.2/include to include path and -lcudart to the linked library.

Below is an example that also use driver API for your reference:

/usr/local/cuda-10.2/samples/0_Simple/matrixMul

Thanks.

[zhoujinzhoujinzhoujin]

Thanks for the reply.
Now my compilation command is this:
nvcc my_program.cu -O0 -g -lcudart -I/usr/local/cuda-10.2/include/ -o my_program
But the problem still exists. Could you explain more about the solution?

[AastaLLL]

Hi,

Could you attach the source of my_program.cu for us checking?

Thanks.

[zhoujinzhoujinzhoujin]

This is my source code:

#include <stdlib.h>
#include <time.h>
#include <stdio.h>
#include <thread>
#include <assert.h>
#include <sys/mman.h>
#include <errno.h>
#include <string.h>
#include <cuda.h>

__global__ void gpuAccess(char * tmp) {
    tmp[0] = 'A';
}

void cpuAccess(char * tmp) {
    tmp[0] = 'B';
}


int main() {

    char * tmp;
    cudaMallocManaged((void**)&tmp, sizeof(char)*10000*4096);
    for(int i = 0; i < 10000; ++i) {
        cpuAccess(tmp+i*4096);

        unsigned long long flag;
        cuMemGetAccess(&flag, 0, (CUdeviceptr)(tmp+i*4096));
        fprintf(stderr, "flag = %lu\n", flag);

        gpuAccess<<<1, 1>>>(tmp+i*4096);
        cudaDeviceSynchronize();
    }
    return 0;
}

[AastaLLL]

Hi,

Sorry for the mistake.

Since you are using CUDA driver API, it should be -lcuda rather than -lcudart (for runtime API).

$ nvcc my_program.cu -O0 -g -lcuda -I/usr/local/cuda-10.2/include/ -o my_program

Thanks.