CUDA debug invalid kernel image error

Question

I wrote the following CUDA kernel and am trying to load it into a module:

#include <stdio.h>

extern "C"   // ensure function name to be exactly "vadd"
{
    __global__ void vadd(const float *a, const float *b, float *c)
    {
        int i = threadIdx.x + blockIdx.x * blockDim.x;
        printf("Thread id %d\n", i);
        c[i] = a[i] + b[i];
    }
}

I compile it to ptx code using the following command:

nvcc -ptx -arch=sm_20 vadd.cu

When trying to load this file into a module using cuModuleLoad I get a CUDA 200 error (invalid kernel image). How can I find out what is wrong with the kernel image? I have tried ptxas , but according to that, the generated ptx code is fine.

Edit : This is the code I am using to load the module:

#include "cuda.h"
#include <cassert>
#include <dlfcn.h>
#include <stdio.h>

void check(CUresult err) {
  if (err != CUDA_SUCCESS) {
    printf("Error %i\n", err);
  }
  assert(err == CUDA_SUCCESS);
}

int main(int argc, char **argv) {
    void *cuda = dlopen("libcuda.so", RTLD_NOW | RTLD_DEEPBIND | RTLD_GLOBAL);
    assert(cuda != NULL);

    printf("cuInit\n");
    CUresult (*Init)() = (CUresult (*)()) dlsym(cuda, "cuInit");
    check(Init());

    printf("cuDeviceGet\n");
    CUresult (*DeviceGet)(CUdevice *, int) = (CUresult (*)(CUdevice *, int)) dlsym(cuda, "cuDeviceGet");
    CUdevice device;
    check(DeviceGet(&device, 0));

    printf("cuCtxCreate\n");
    CUresult (*CtxCreate)(CUcontext * , unsigned int, CUdevice) = (CUresult (*)(CUcontext * , unsigned int, CUdevice)) dlsym(cuda, "cuCtxCreate");
    CUcontext context;
    check(CtxCreate(&context, 0, device));

    printf("cuModuleLoad\n");
    CUresult (*ModuleLoad)(CUmodule *, const char*) = (CUresult (*)(CUmodule *, const char*)) dlsym(cuda, "cuModuleLoad");
    CUmodule mod;
    check(ModuleLoad(&mod, "vadd.ptx"));

    return 0;
}

Answer 1

This is related to Why cuCtxCreate creates old context? : you are using cuCtxCreate directly, which gives you an old API context (v3.1) incompatible with your usage of printf . You can check the API version with cuCtxGetApiVersion . If you switch to cuCtxCreate_v2 , which is normally used through some #define 's in cuda.h , you'll get a more recent API context.

In order to spot this discrepancy, I've run your sample with LD_DEBUG=symbols , and compared it to using the CUDA API directly (since it properly runs your sample PTX). Comparing symbol resolutions, the big difference was the call to cuCtxCreate :

cuCtxCreate(...)
    symbol=cuCtxCreate_v2;  lookup in file=./test [0]
    symbol=cuCtxCreate_v2;  lookup in file=/usr/lib/x86_64-linux-gnu/libcuda.so.1 [0]

... which in your original code, using dlsym(..., "cuCtxCreate") mapped directly to cuCtxCreate .