An introduction to 
      many-core parallel 
      computing with 
      OpenCL  
     Simon McIntosh-Smith 
      Twitter: @simonmcs 
                       UPMARC summer school 
                                   th
                       Uppsala 28-29  2014 
          Recap 
 5 simple steps in a basic OpenCL program: 
 1.  Define the platform = devices + context + 
  queues 
 2.  Create and Build the program (dynamic 
  library of kernels) 
 3.  Setup memory objects 
 4.  Define the kernels 
 5.  Submit commands … transfer memory 
  objects and execute kernels 
                       We have now covered the basic 
                    platform runtime APIs in OpenCL 
                                                                                                                                       
                                                                          CPU                                                    GPU 
                                                                                               Context 
                              Programs                                           Kernels                            Memory Objects                              Command Queues 
                              Programs 
        __kernel void                                                                dp_mul                     Buffers               Images 
                                                                                      arg [0] 
        dp_mul(global const float *a,                  dp_mul                        arg [0]                                                                          In               Out of 
                                                                                        value                                                                        In               Out of 
                                                                                    arg[0] value 
               global const float *b,           CPU program binary                    value                                                                       Order                Order 
               global float *c)                                                       arg [1]                                                                     Order               Order 
        {                                                                            arg [1] 
                                                       dp_mul                           value                                                                     Queue               Queue 
                                                                                    arg[1] value                                                                                      Queue 
          int id = get_global_id(0);            GPU program binary                    value                                                                      Queue 
          c[id] = a[id] * b[id];                                                      arg [2] 
                                                                                     arg [2] 
        }                                                                               value 
                                                                                    arg[2] value 
                                                                                      value                                                                      GPU 
                                                                                                                                                               Compute Device 
  OPENCL KERNEL PROGRAMMING