Question

I have a vector of size n (unknown before the main program is launched), and I want to experiment with the work-group sizes. However, unless I set the local_work_size to a number that exactly divides n, I only get 0 values in fprop below.

Kernel:

__kernel void palt(__global double *fprop, __global const double *fcoll,
                   __global const int *nn, const uint max_size)
{
    size_t l = get_global_id(0);

    if( l > max_size ) return;

    fprop[l] = fcoll[nn[l]];

}

Host code:

int block_sz_p = 128;
const int max_size = ns*imax;

// set the parameters for the propagation operator
errNum = clSetKernelArg(propagation_kernel, 0, sizeof(cl_mem), &fpd);
errNum |= clSetKernelArg(propagation_kernel, 1, sizeof(cl_mem), &fcd);
errNum |= clSetKernelArg(propagation_kernel, 2, sizeof(cl_mem), &nnd);
errNum |= clSetKernelArg(propagation_kernel, 3, sizeof(int), (void *) &max_sz);
checkErr(errNum, "clSetKernelArg(propagation)");

// specify the work group size/dim
const size_t work_dim = 3;  
const size_t global_work_size_propagation[] = {imax*ns, 1, 1};
const size_t local_work_size_propagation[] = {block_sz_p, 1, 1};

// propagation
clEnqueueNDRangeKernel(queue, propagation_kernel, work_dim, NULL, 
                       global_work_size_propagation, local_work_size_propagation,
                       0, NULL, &event);

clWaitForEvents(1, &event); 
clGetEventProfilingInfo(event, CL_PROFILING_COMMAND_START, 
                        sizeof(cl_ulong), &start, NULL);
clGetEventProfilingInfo(event, CL_PROFILING_COMMAND_END, 
                        sizeof(cl_ulong), &end, NULL);
tker2 = (end-start);

What’s going on here?

Answer 1

You should check for CL errors, clEnqueueNDRangeKernel and other calls return a error code (others return the error code by reference).

I assume the problem is when the global workgroup size isn’t divisible by the local workgroup size, which is not supported and generates a CL error:

CL_INVALID_WORK_GROUP_SIZE if local_work_size is specified and number of work-items specified by global_work_size is not evenly divisable by size of work-group given by local_work_size or does not match the work-group size specified for kernel using the attribute((reqd_work_group_size(X, Y, Z))) qualifier in program source.

From the clEnqueueNDRangeKernel man page.