cuda_driver_dlopen.c

// OK:
// gcc test_dlopen.c -ldl -lcuda -lcudart -L/usr/local/cuda-12.2.1/lib64 -I/usr/local/cuda-12.2.1/include -o test

// FAIL with "Error getting global symbol: 201" when compiled with "-DDLOPEN":
// gcc test_dlopen.c -ldl -lcuda -lcudart -L/usr/local/cuda-12.2.1/lib64 -I/usr/local/cuda-12.2.1/include -o test -DDLOPEN

#include <stdio.h>
#include <dlfcn.h>
#include <cuda.h>
#include <cuda_runtime_api.h>


int main() {

#ifdef DLOPEN
    void* libHandle = dlopen("libcuda.so", RTLD_NOW);
    if (!libHandle) {
        fprintf(stderr, "Error loading libcuda.so: %s\n", dlerror());
        return -1;
    }
    typedef CUresult (*cuInit_t)(unsigned int);
    typedef CUresult (*cuDeviceGet_t)(CUdevice*, int);
    typedef CUresult (*cuCtxCreate_t)(CUcontext*, unsigned int, CUdevice);
    typedef CUresult (*cuModuleLoad_t)(CUmodule*, const char*);
    typedef CUresult (*cuModuleGetGlobal_t)(CUdeviceptr*, size_t*, CUmodule, const char*);
    typedef CUresult (*cuModuleUnload_t)(CUmodule);
    typedef CUresult (*cuCtxDestroy_t)(CUcontext);
    cuInit_t cuInit = (cuInit_t)dlsym(libHandle, "cuInit");
    cuDeviceGet_t cuDeviceGet = (cuDeviceGet_t)dlsym(libHandle, "cuDeviceGet");
    cuCtxCreate_t cuCtxCreate = (cuCtxCreate_t)dlsym(libHandle, "cuCtxCreate");
    cuModuleLoad_t cuModuleLoad = (cuModuleLoad_t)dlsym(libHandle, "cuModuleLoad");
    cuModuleGetGlobal_t cuModuleGetGlobal = (cuModuleGetGlobal_t)dlsym(libHandle, "cuModuleGetGlobal");
    cuModuleUnload_t cuModuleUnload = (cuModuleUnload_t)dlsym(libHandle, "cuModuleUnload");
    cuCtxDestroy_t cuCtxDestroy = (cuCtxDestroy_t)dlsym(libHandle, "cuCtxDestroy");
#endif

    CUresult res;
    CUdevice cuDevice;
    CUcontext cuContext;
    CUmodule cuModule;
    CUdeviceptr dptr;
    size_t bytes;

    // Create an implicit context.
    cudaError_t err = cudaSetDevice(0);
    if (err != CUDA_SUCCESS) {
        fprintf(stderr, "Error setting device: %d\n", err);
        return -1;
    }

    // Load the module from a PTX file
    res = cuModuleLoad(&cuModule, "test.ptx");
    if (res != CUDA_SUCCESS) {
        printf("Error loading module: %d\n", res);
        return -1;
    }

    // Retrieve a pointer to the global symbol
    res = cuModuleGetGlobal(&dptr, &bytes, cuModule, "some_array");
    if (res != CUDA_SUCCESS) {
        printf("Error getting global symbol: %d\n", res);
        cuModuleUnload(cuModule);
        return -1;
    }

    return 0;
}