code example
shmem SGI shmem
one-sided communication
OpenMP multi-threading
Cell Cell Broadband Engine
file sub1.h: common interface

class data1 {
public:
   data1() {}
   int n0, n1;
   float *x, *y;
   float e;
   int pad[27];// pad to 128 byte length
};

file spe1.cc: SPU code

#include <spu_mfcio.h>
#include "sub1.h"
#define BLOCK (2048)
int main(unsigned long long id,
     unsigned long long argp, unsigned long long envp) {
  data1 vec __attribute__ (aligned(128));
  // read parameter
  mfc_get(&vec, (unsigned int)argp, sizeof(data1), 1, 0, 0);
  mfc_write_tag_mask(0xffffffff);
  mfc_read_tag_status_all();// wait for data transfer
  float e = 0;
  // allocate double buffer in local SPU memory
  float *x = (float*)malloc(BLOCK*sizeof(float));
  float *x0 = (float*)malloc(BLOCK*sizeof(float));
  float *y = (float*)malloc((BLOCK+32)*sizeof(float)) + 1;
  float *y0 = (float*)malloc((BLOCK+32)*sizeof(float)) + 1;
  // get first buffer
  mfc_get(x, &vec.x[vec.n0], BLOCK*sizeof(float), 2, 0, 0);
  mfc_get(y-1, &vec.y[vec.n0], (BLOCK+32)*sizeof(float), 3, 0, 0);
  mfc_read_tag_status_all();
  for (int ib=vec.n0; ib<vec.n1; ib += BLOCK) {
   if (ib+BLOCK<vec.n1) {
    // get next buffer, fence after put x
    mfc_getf(x0, &vec.x[ib+BLOCK], BLOCK*sizeof(float), 2, 0, 0);
    mfc_get(y0-1, &vec.y[ib+BLOCK], (BLOCK+32)*sizeof(float), 3, 0, 0);
   }
  // do computation
  for (int i=0; iBLOCK; ++i) {
   x[i] += ( y[i+1] + y[i-1] )*.5;
   e += y[i] * y[i];
  }
   mfc_read_tag_status_all(); // wait for data transfer
   // put current buffer
   mfc_put(x, &vec.x[ib], BLOCK*sizeof(float), 2, 0, 0);
   float *t = x; x = x0; x0 = t;
   t = y; y = y0; y0 = t; // swap buffers
  }
  // put reduction value
  vec.e = e;
  mfc_put(&vec, argp, sizeof(data1), 1, 0, 0);
  mfc_read_tag_status_all(); // wait for data transfer
  return 0;
}

main file
#include <mpp/shmem.h>
#include <omp.h>
#include <libspe2.h>
#include "sub1.h"

float e;

extern spe_program_handle_t spe1; // defined in SPU code
int main(int argc, char *argv[]) {
  int n = ...;
  start_pes(0);
  int nn = (n-1) / _num_pes();
  int n_local0 = 1 + _my_pe() * nn;
  int n_local1 = 1 + (_my_pe()+1) * nn;
  // allocate only local part + ghost zone of the arrays x,y
  float *x, *y;
  x = (float*)shmalloc((n_local1 - n_local0 + 2)*sizeof(float));
  y = (float*)shmalloc((n_local1 - n_local0 + 2)*sizeof(float));
  x -= (n_local0 - 1);
  y -= (n_local0 - 1);
  shmem_barrier_all();

  ... // fill x, y

  // fill ghost zone
  if (_my_pe() > 0)
   shmem_float_put(&y[n_local1], &y[n_local0], 1, _my_pe()-1);
  if (_my_pe() < _num_pes()-1)
   shmem_float_put(&y[n_local0-1], &y[n_local1-1], 1, _my_pe()+1);
  shmem_barrier_all();

  e = 0;
  #pragma omp parallel
  {
  int p = omp_get_thread_num();
  int num = omp_get_num_threads();
  data1 block __attribute__ (aligned(128));
  spe_context_ptr_t ctxs;
  spe_stop_info_t st;
  ctxs = spe_context_create(0, NULL);
  spe_program_load (ctxs, &spe1);
  block.n0 = n_local0;
  block.n1 = n_local1;
  block.x = &x[n_local0];
  block.y = &y[n_local0-1];
  unsigned int entry = SPE_DEFAULT_ENTRY;
  int n0 = 1+((n_local1-n_local0)*p)/num;
  int n1 = 1+((n_local1-n_local0)*(p+1))/num;
  // execute code on a single SPU and wait for termination
  spe_context_run(ctxs[p], &entry, 0, &block, NULL, st);
  spe_context_destroy(ctxs);
  #pragma omp atomic
  e += block.e;
  }

  static float work[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
  static long sync[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
  static float el, es;
  el = e;
  shmem_float_sum_to_all(&es, &el, 1,
   0, 0, _num_pes(), work, sync);
  e = es;

  ... // output x, e

  x += (n_local0 - 1);
  y += (n_local0 - 1);
  shfree(x);
  shfree(y);
  return 0;
}

[start] [references] [download] [install]