Can't get simple CUDA program to work

Question

I'm trying the "hello world" program of CUDA programming: adding two vectors together. Here's the program I have tried:

#include <cuda.h>
#include <stdio.h> 
#define  SIZE 10

__global__  void vecAdd(float* A, float* B, float* C) 
{ 
   int i = threadIdx.x; 
   C[i] = A[i] + B[i]; 
} 

int  main() 
{ 
     float A[SIZE], B[SIZE], C[SIZE]; 
     float *devPtrA, *devPtrB, *devPtrC; 
     size_t memsize= SIZE * sizeof(float); 

     for (int i=0; i< SIZE; i++) {
        A[i] = i;
        B[i] = i;
     }

     cudaMalloc(&devPtrA, memsize); 
     cudaMalloc(&devPtrB, memsize); 
     cudaMalloc(&devPtrC, memsize); 
     cudaMemcpy(devPtrA, A, memsize,  cudaMemcpyHostToDevice); 
     cudaMemcpy(devPtrB, B, memsize,  cudaMemcpyHostToDevice); 

     vecAdd<<<1, SIZE>>>(devPtrA,  devPtrB, devPtrC); 
     cudaMemcpy(C, devPtrC, memsize,  cudaMemcpyDeviceToHost); 

     for (int i=0; i<SIZE; i++) 
         printf("C[%d]: %f + %f => %f\n",i,A[i],B[i],C[i]); 

     cudaFree(devPtrA); 
     cudaFree(devPtrB); 
     cudaFree(devPtrC); 
}

Compiled with:

nvcc cuda.cu

Output is this:

C[0]: 0.000000 + 0.000000 => 0.000000
C[1]: 1.000000 + 1.000000 => 0.000000
C[2]: 2.000000 + 2.000000 => 0.000000
C[3]: 3.000000 + 3.000000 => 0.000000
C[4]: 4.000000 + 4.000000 => 0.000000
C[5]: 5.000000 + 5.000000 => 0.000000
C[6]: 6.000000 + 6.000000 => 0.000000
C[7]: 7.000000 + 7.000000 => 0.000000
C[8]: 8.000000 + 8.000000 => 366987238703104.000000
C[9]: 9.000000 + 9.000000 => 0.000000

Every time I run it, I get a different answer for C[8], but the results for all the other elements are always 0.000000.

The Ubuntu 11.04 system a 64-bit Xeon server with 4 cores running the latest NVIDIA drivers (downloaded on Oct 4, 2012). The card is an EVGA GeForce GT 430 with 96 cores and 1GB of RAM.

What should I do to figure out what's going on?

Answer 1

It seems that your drivers are not initialized, but not checking the cuda return codes is always a bad practice, you should avoid that. Here is simple function + Macro that you can use for cuda calls(quoted from Cuda by Example):

static void HandleError( cudaError_t err,
                         const char *file,
                         int line ) {
    if (err != cudaSuccess) {
        printf( "%s in %s at line %d\n", cudaGetErrorString( err ),
                file, line );
        exit( EXIT_FAILURE );
    }
}
#define HANDLE_ERROR( err ) (HandleError( err, __FILE__, __LINE__ ))

Now start calling your functions like:

HANDLE_ERROR(cudaMemcpy(...));

Answer 2

Most likely cause: the NVIDIA drivers weren't loaded. On a headless Linux system, X Windows isn't running, so the drivers aren't loaded at boot time.

Run nvidia-smi -a as root to load them and get a confirmation in the form of a report.

Although the drivers are now loaded, they still need to be initialized every time a CUDA program is run. Put the drivers into persistent mode with nvidia-smi -pm 1 so they remain initialized all the time. Add this to a boot script (such as rc.local) so it happens at every boot.