cuda cross-compilation with sdk

[aurelien-enchanted-tools]

Hello,

I decided to start cuda programming to better understand build issues.
I found an example here (https://developer.nvidia.com/blog/even-easier-introduction-cuda/) that I modified (with https://docs.nvidia.com/cuda/archive/11.4.0/cuda-c-programming-guide/index.html#kernels):

cat > add.cu << "EOF"
#include <iostream>
#include <math.h>
// Kernel function to add the elements of two arrays
__global__ void add(float* x, float* y, float* z) {
    int i = threadIdx.x;
    z[i] = x[i] + y[i];
}

int main(void) {
    int N = 1<<8;
    float *x, *y, *z;

    // Allocate Unified Memory – accessible from CPU or GPU
    cudaMallocManaged(&x, N*sizeof(float));
    cudaMallocManaged(&y, N*sizeof(float));
    cudaMallocManaged(&z, N*sizeof(float));

    // initialize x , y and z arrays on the host
    for (int i = 0; i < N; i++) {
        x[i] = 1.0f;
        y[i] = 2.0f;
        z[i] = 0.0f;
    }

    // Run kernel on 256 elements on the GPU
    add<<<1, N>>>(x, y, z);

    // Wait for GPU to finish before accessing on host
    cudaDeviceSynchronize();

    // all values should be 3.0f
    for (int i = 0; i < N; i++)
        std::cout << i << ":" << z[i] << std::endl;

    // Free memory
    cudaFree(x);
    cudaFree(y);
    cudaFree(z);

    return 0;
}
EOF

I build both the example and my customization, with ${CUDACXX} add.cu ${CUDAFLAGS} -ccbin=${CUDAHOSTCXX} -o add from an sdk env, and export it and run it to a orin NX target.

But for both, I don't obtain correct results. For the example, I obtain Max error: 1, and for my customization, I obtain a zero value for all index of z.

Why these do not give the expected results? Thank you in advance.

Edit:

It seems that the __global__ is never taken into account. I suspect there’s something missing in the build command line.

if I use from https://stackoverflow.com/questions/14038589/what-is-the-canonical-way-to-check-for-errors-using-the-cuda-runtime-api#answer-14038590:

#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, char *file, int line, bool abort=true)
{
   if (code != cudaSuccess)
   {
      fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
      if (abort) exit(code);
   }
}

and if I add it after the cuda kernel call

    add<<<1, N>>>(x, y, z);
    gpuErrchk( cudaPeekAtLastError() );

I obtain at runtime:

GPUassert: no kernel image is available for execution on the device add.cu 36

EDIT: I found my mistake, incompatibility between the architecture of cuda of the sdk and that of the image.