c++ Cuda code not running kernel

Question

I recently setup Ubuntu 14.04.1 LTS. I am compiling a small matrix addition program I wrote, matrixAddition.cu. I compile the code like so: nvcc matrixAddition.cu -o matAdd. The produced matrix should show 7 7 7 ... as I am adding an array of 5 with an array of 2. However, it prints 5 5 5 ... for me. Is the kernel failing to work for some reason? Am I missing some code? Thanks for reading.

matrixAddition.cu

    #include <iostream>
    #include <stdio.h>

  __global__ void matAdd(int * d_arra, int * d_arrb, int * length)
  {
    int id = threadIdx.x;
    if(id<*length)
    {
            d_arra[id]=d_arra[id]+d_arrb[id];
    }
 }


int main () {

    //cpu varriables
    int arra[100];
    int arrb[100];
    int leng = 100;

    //gpu varriables
    int * d_arra;
    int * d_arrb;
    int * length;

    //-std=c++11
    for(int itr=0;itr<100;itr++){arra[itr]=5;arrb[itr]=2;}

    cudaMalloc( (void**)&d_arra, 100*sizeof(int)); 
    cudaMalloc( (void**)&d_arrb, 100*sizeof(int)); 
    cudaMemcpy( d_arra, arra, 100*sizeof(int), cudaMemcpyHostToDevice ); 
    cudaMemcpy( d_arrb, arrb, 100*sizeof(int), cudaMemcpyHostToDevice ); 
    cudaMemcpy( length, &leng, sizeof(int), cudaMemcpyHostToDevice ); 


    //max thread per block 1024
    matAdd<<<1, 1024>>>(d_arra, d_arrb, length);
    cudaMemcpy( arra, d_arra, 100*sizeof(int), cudaMemcpyDeviceToHost ); 
    cudaFree( d_arra );
    cudaFree( d_arrb );
    cudaFree( length );

    std::cout << " our data \n";

    for(int itr=0;itr<100;itr++){std::cout << arra[itr]; if(itr%10==0&&itr!=0){std::cout <<"\n";}}
    std::cout<<std::endl;


 return 0;
}

any time you're having trouble with a CUDA code, it's good practice to use [proper cuda error checking](http://stackoverflow.com/questions/14038589/what-is-the-canonical-way-to-check-for-errors-using-the-cuda-runtime-api). (I believe Yuan's answer is correct, however.) — Robert Crovella, Jan 10 '15 at 11:36

Yuan · Answer 1 · 2015-01-10T05:51:07.863

The bug is common in C. Variable "length" is not initialized. You forgot to malloc memory space for a point variable before copying data to the address.

If using type "int *", you have to do cudaMalloc for it as d_arra/d_arrb. But it's best to use type "int" instead of "int *" as the example vectorAdd in CUDA Samples, since only a single int is passed to your CUDA device.

Another small issue is in your cout part. Since array index starts from 0, you need output a line break when "(itr+1)%10==0 && itr!=0".

//cpu varriables
int arra[100];
int arrb[100];
int leng = 100;

//gpu varriables
int * d_arra;
int * d_arrb;
int * length;

//-std=c++11
for (int itr=0; itr<100; itr++)
{
    arra[itr]=5;
    arrb[itr]=2;
}

cudaMalloc( (void**)&d_arra, 100*sizeof(int)); 
cudaMalloc( (void**)&d_arrb, 100*sizeof(int)); 
cudaMalloc( (void**)&length, 1*sizeof(int)); // Add this line
cudaMemcpy( d_arra, arra, 100*sizeof(int), cudaMemcpyHostToDevice ); 
cudaMemcpy( d_arrb, arrb, 100*sizeof(int), cudaMemcpyHostToDevice ); 
cudaMemcpy( length, &leng, sizeof(int), cudaMemcpyHostToDevice ); 


//max thread per block 1024
matAdd<<<1, 1024>>>(d_arra, d_arrb, length);
cudaMemcpy( arra, d_arra, 100*sizeof(int), cudaMemcpyDeviceToHost ); 
cudaFree( d_arra );
cudaFree( d_arrb );
cudaFree( length );

std::cout << " our data \n";

for (int itr=0; itr<100; itr++)
{
    std::cout << arra[itr]; 
    if((itr+1)%10==0 && itr!=0)
    {
        std::cout <<"\n";
    }
}
std::cout<<std::endl;

c++ Cuda code not running kernel

1 Answers1