Optimized Data Transfers Based on the OpenCL Event Management Mechanism

Listing 1

A simple pseudocode combining OpenCL and MPI.

()  cl_command_queue cmd;

()  cl_kernel kern;

()  cl_event evt;

()

()  for(int i(0);i<N;++i){

()   // (1) computation on a device

()   clEnqueueNDRangeKernel(cmd,kern,…,0,NULL,&evt);

()

()   // (2) read the result from device to host

() clEnqueueReadBuffer(cmd,…,1,&evt,NULL);

() clFinish(cmd);  // the host thread is blocked

()

() // (3) exchange data with other nodes

() MPI_Sendrecv(…);  // blocking function call

()

() // (4) write the received data to device memory

() clEnqueueWriteBuffer(cmd,…);

() }