mycuda_scan.h

Go to the documentation of this file.

namespace mycuda_scan
{



template <typename T>
__device__ void scan_dev( T   * r,     
                          T     x,     
                          int   i,     
                          int   n)     
{
    r[i] = 0.0f;
    i   += n;
    r[i] = x;

    for( uint offset=1; offset<n; offset<<=1){
        __syncthreads();
        T t = r[i] + r[i - offset];
        __syncthreads();
        r[i] = t;
    }
    __syncthreads();
}





 
template <typename T>
__global__ 
void scan_blocks( T*  w,           
                  T*  x,           
                  T*  blocksum,    
                  int inclusive )  
{
    int n    = blockDim.x;
    int j    = blockIdx.x;
    int tid  = threadIdx.x;
    int i    = j*n + tid;

    extern __shared__ T r[];    

    scan_dev( r, x[i], tid, n );
    if ( inclusive ) {
        w[i] = r[tid+n];
    } 
    else {
        w[i] = ( tid==0 ? 0.0f : r[tid+n-1] );
    }
    if (tid==0) blocksum[j] = r[2*n-1]; 
}
}  // end namespace mycuda_scan