如何通过 cuda 中的索引将数组元素设置为零？

Question

I am trying with cuda to set some elements in array by index to zero. My array size has about 7,000,000 elements. The index length is about 1,000. So I want to write the kernel code efficiently. The only technique I know is to set the block size by cudaOccupancyMaxPotentialBlockSize . Could any one give me some suggestion to speed up?

eg The pointer of the array a is double *a , with size n . The index's pointer is int * index , with length n1 .

__global__ void setZero(int n, double * a,int n1, const int* index)
{
  int i = threadIdx.x + blockIdx.x * blockDim.x;
  if (i<n)
  {
    for(int ii=0; ii<n1; ii++) 
      if(i==index[ii]-1)
        a[i] = 0;
  }
}

void main() 
{
    int blockSize;      
    int minGridSize;    
    int gridSize; 
    cudaOccupancyMaxPotentialBlockSize(&minGridSize, &blockSize, setZero, 0, n); 
    gridSize = (n + blockSize - 1) / blockSize;
    setZero<<<gridSize, blockSize>>>(n, d_a, n1, d_index);
}

As a mini sample, a = {1,2,3,4,5}, index = [2,4] . The output is a = {1,0,3,0,5} .

Answer 1

Given your constrains I think the following would already be good enough:

__global__ void setZero(int n, double *a, int n1, const int* index, const int* index_size)
{
  int id = threadIdx.x + blockIdx.x * blockDim.x;
  if (id < index_size)
     a[index[id]]=0
}