Optimized Data Transfers Based on the OpenCL Event Management Mechanism

Listing 4

A Himeno benchmark code with the proposed OpenCL extension.

() cl_command_queue cmd1, cmd2;

() cl_mem p_new, p_old, p_tmp;

() cl_event e[2];

()

() for(int i(0);i<N;++i){

()  p_tmp = p_new; p_new = p_old; p_old = p_tmp;

()  if( rank%2 == 0) {

()    jacobi_kernel_even_A(cmd1,p_new…0,NULL,&e[0]);

()    clEnqueueSendrecvBuffer(cmd2,p_old,…0,NULL,&e[1]);

()  jacobi_kernel_even_B(cmd2,p_new…1,&e[1],NULL);

()  clEnqueueSendrecvBuffer(cmd1,p_new,…1,&e[0],NULL);

() }

() else {

()  jacobi_kernel_odd_B(cmd2,p_new…0,NULL,&e[0]);

()  clEnqueueSendrecvBuffer(cmd1,p_old,…0,NULL,&e[1]);

()  jacobi_kernel_odd_A(cmd1,p_new…1,&e[1],NULL);

()  clEnqueueSendrecvBuffer(cmd2,p_new,…1,&e[0],NULL);

() }

() clFinish(cmd1);clFinish(cmd2);

() /* error calculation */

() }