Loading module containig printf seems to fail

[pieterverstraete]

I have extended the example that is described in the readme to print out the thread id. The code then looks like this:

#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 then compile this code to ptx using:

nvcc -ptx -arch=sm_20 vadd.cu

The -arch=sm_20 is added to make the printf work. However, now I cannot load the module into Julia anymore. It keeps complaining about an invalid kernel image:

ERROR: 200
 in error at error.jl:22
 in include_from_node1 at loading.jl:120
while loading /var/cache/workdir/pfverstr/test2.jl, in expression starting on line 16

However, the code does work when I execute the same calls to libcuda (cuInit, cuDeviceGet, cuCtxCreate and cuModuleLoad) from a program written in C++.

[pieterverstraete]

It seems like it has something to do with dlopen, because when I use dlopen in cpp code, it also fails.

[lindahua]

Would you please let me know the OS you are using?

[pieterverstraete]

The result of uname -a is:
Linux phoenix 3.13-1-amd64 #1 SMP Debian 3.13.5-1 (2014-03-04) x86_64 GNU/Linux

[maleadt]

This is because cuCtxCreate is used, while for more recent API versions cuCtxCreate_v2 should be used. Also see my response at http://stackoverflow.com/questions/22612879/cuda-debug-invalid-kernel-image-error/22634798

cuda.h contains API-conditional code managing this. I guess CUDA.jl should mirror (part of) this, selecting proper API calls based on the requested version?

[lindahua]

I just updated the package, with several bug fixes (e.g. corrected function names and pointer sizes). Please checkout the latest and try again.