Results 1 to 2 of 2

Thread: Barrier Function in while loop

  1. #1

    Barrier Function in while loop

    In the kernel function if i am doing min reduction which involves a while loop; i am getting the correct answer for smaller data sets but not getting the correct results for large datasets.
    I think i have a synchronization problem between the threads, can anyone help me or tell me where do i need to put the barrier function in the kernel function and the reason for it.


    __kernel void add_update_min(__global int *selected, __global float *d, __global float *A, __global float *dcopy, __global int *stVertex,__global int *stVertexcopy,__global int *destVertex, int n)
    {

    // Getting the global id of the thread
    int tid = get_global_id(0);

    int p;
    p = destVertex[0];

    // Updating distances of all the vertices after a new vertex is added
    if(selected[tid] == 0)
    {
    if(d[tid] > A[tid*n + p])
    {
    d[tid] = A[tid*n + p];
    stVertex[tid] = p;
    }
    }

    else
    {
    d[tid] = 9999;
    }

    barrier(CLK_GLOBAL_MEM_FENCE);

    stVertexcopy[tid] = tid;

    dcopy[tid] = d[tid];

    barrier(CLK_GLOBAL_MEM_FENCE);

    //Calculating the minimum among all the d[i]'s
    while(tid < n/2 )
    {
    if(n%2 != 0) // check if number of elements in consideration are suitable for normal min reduction (check out the else part)
    {
    if(tid != (n-1)/2) // if thread id is not the same as the middle element : think 9 and 4 [0,1,2,3,4]
    {

    if(dcopy[tid] > dcopy[tid + (n+1)/2]) // find the smalles
    {
    dcopy[tid] = dcopy[tid + (n+1)/2];
    stVertexcopy[tid] = stVertexcopy[tid + (n+1)/2];
    }
    }


    if(tid == 0) // shall also include the middle element in each iteration of min reduction --- think of a 23 element array [0....22]
    {
    // every thread is going to execute this
    if(dcopy[tid] > dcopy[(n-1)/2])
    {
    dcopy[tid] = dcopy[(n-1)/2];
    stVertexcopy[tid] = stVertexcopy[(n-1)/2];
    }
    }
    n = (n-1)/2;

    }

    else
    {
    if(dcopy[tid] > dcopy[tid + n/2])
    {
    dcopy[tid] = dcopy[tid + n/2];
    stVertexcopy[tid] = stVertexcopy[tid + n/2];
    }
    n = n/2;
    }
    barrier(CLK_GLOBAL_MEM_FENCE);
    }

    barrier(CLK_GLOBAL_MEM_FENCE);

    if(tid == 0)
    {
    destVertex[0] = stVertexcopy[0];
    p = destVertex[0];
    selected[p] = 1;
    }
    }


    Here i have the global item size as n
    local item size i have specified it as NULL.

  2. #2

    Re: Barrier Function in while loop

    Although I can't claim to have read your code well enough to understand precisely what you're trying to do, I think I can guess at what the problem is:

    I think this code is written assuming barrier(CLK_GLOBAL_MEM_FENCE) is a barrier for all work items in a kernel. It is not. It is still only a barrier for the work items in an individual workgroup. The CLK_GLOBAL_MEM_FENCE flag is just a memory fence for global memory. There is no barrier that operates beyond an individual workgroup.

    The tipoff for me is to see the global array dcopy initialized using your global ID as an index, and then later read using some other index. Even though there's a barrier in between, there's no guarantee that all of the dcopy elements have been initialized.

Similar Threads

  1. Replies: 4
    Last Post: 01-25-2013, 05:27 AM
  2. why use barrier?
    By memo in forum OpenCL
    Replies: 2
    Last Post: 11-18-2009, 11:25 AM

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •