Can't use constants in Cuda

Question

Example:

#include <cuda.h>
#include <stdint.h>
#include <assert.h>

__constant__ int32_t m;

int main(int argc, char* argv[])
{
    void* s;
    int r = cudaGetSymbolAddress( &s, m);
    assert( r == cudaSuccess );
    return 0;
}

Compile:

$ nvcc test.cu -o test -arch compute_20 -code sm_20

Run:

$ ./test

Get:

test: test.cu:15: int main(int, char**): Assertion `r == cudaSuccess' failed.
Aborted (core dumped)

(If that makes any difference, I tested this in two different cards in two different computers. Cuda 6 in both cases.)

What's wrong there?

Answer 1

As @sgar91 pointed out, the problem is the compilation target mismatching the actual GPU.

To be specific: you have -code sm_20 in your options which will make the compiler build a binary for sm_20 and there will be no PTX in your binary - that means it cannot be JIT compiled for your devices (compute capability > 2.0) and hence your GPU operations will fail. You should have -code compute_20 or one or more -gencode arguments (see nvcc manual for more examples).

Some examples:

$ nvcc test.cu -o test -arch compute_20 -code compute_20
$ nvcc test.cu -o test -gencode="arch=compute_20,code=\"compute_20,sm_20,sm_30\""
$ nvcc test.cu -o test -gencode="arch=compute_20,code=\"sm_20,sm_21\"" -gencode="arch=compute_30,code=\"compute_30,sm_30\""

Rather than doing an assert on your CUDA API call, you should report the actual error since that would have been helpful here.