How To Start writing CUDA Program

1. Who can start CUDA programming? – Any one who has some basicknowledge on C programming. No prior knowledge of graphics programming.

– CUDA is C with minimal extension and some restrictions.( CUDA is C for GPU)

2. If you have a algo that you want to write in CUDA then follow the below steps. 3. Start writing your algo in C/C++. 4. Changing your implemented C/C++ program according to CUDA requirements. 5. The basic diagram for CUDA code implementation as: Allocate CPU memory  >> Allocate same amount of GPU memory >> take data input in CPU memory >> copy data into GPU memory >> doing processing in GPU memory >> copying final data in CPU memory. 6. Explainig  simple example that calculate cosine of range 0 to 1024.

7.  C implementation of algo.

void vcos( int n, float* x, float* y )
        {
           for ( int i = 0; i < n; ++i)
               y[i] = cos( x[i] );
         }

    int main()
       {
   float *host_x, *host_y;

    int n = 1024;
           host_x = (float*)malloc( n*sizeof(float) );
           host_y = (float*)malloc( n*sizeof(float) );
 
    for(  int i = 0; i < 1024; ++i )
        host_x[i] = (float)i;

           vcos( n, host_x, host_y );
                  free(host_x);
                  free(host_y);
 
      return( 0 );
 }
 8 .   CUDA implementation of Same algo( what should be minimal changes in C code. )
 
          __global__ void vcos( int n, float* x, float* y )
        {
           int  i = blockIdx.x * blockDim.x + threadIdx.x;
           y[i] = cos( x[i] );
         }

    int main()
       {
   float *host_x, *host_y;
   float *dev_x, *dev_y;
   int n = 1024;
 
       host_x = (float*)malloc( n*sizeof(float) );
          host_y = (float*)malloc( n*sizeof(float) );
   cudaMalloc( &dev_x, n*sizeof(float) );
          cudaMalloc( &dev_y, n*sizeof(float) );
   for(  int i = 0; i < 1024; ++i )
          host_x[i] = (float) i;
  /* fill host_x[i] with data here */
          cudaMemcpy( dev_x, host_x, n*sizeof(float), cudaMemcpyHostToDevice );
        /* launch 1 thread per vector-element, 256 threads per block */
          bk = (int)( n / 256 );
          vcos( n, dev_x, dev_y );
          cudaMemcpy( host_y, dev_y, n*sizeof(float), cudaMemcpyDeviceToHost );
       /* host_y now contains cos(x) data */

                free(host_x);  
                free(host_y);
                cudaFree(dev_x);
                cudaFree(dev_y);
 
      return( 0 );
 }