code example
DCMF IBM Deep Computing Messaging Framework
one-sided communication
OpenMP multi-threading
Cell Cell Broadband Engine
Note that Cell BE does not offer SSE instructions. Use AltiVec instead.
SSE IA SIMD instructions
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
  float ve[4] = {0, 0, 0, 0};
  for (int i=0; iBLOCK; i+=4) {
   float half =.5;
   _mm_store_ps(&x[i],
     _mm_mul_ps(_mm_load1_ps(&half),
       _mm_add_ps(_mm_loadu_ps(&y[i+1]),
         _mm_loadu_ps(&y[i-1]))));
   _mm_store_ps(&ve[0],
     _mm_add_ps(_mm_load_ps(&ve[0]),
       _mm_mul_ps(_mm_load_ps(&y[i]),
         _mm_load_ps(&y[i]))));
  }
  e += ve[0] + ve[1] + ve[2] + ve[3];
   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 <string.h>
#include <dcmf.h>
#include <dcmf_globalcollectives.h>
#include <omp.h>
#include <libspe2.h>
#include "sub1.h"
#include <xmmintrin.h>

float e;
DCMF_Protocol_t barrier_prot, control0_prot, control1_prot,
  put_prot, reduce_prot;

void cb_decr(void *data) {
  unsigned *val = (unsigned*)data;
  (*val)--;
}

void cb_recv(void *data, const DCMF_Control_t *info, unsigned) {
  memcpy((DCMF_Memregion_t*)data, info, sizeof(DCMF_Memregion_t));
}

void barrier() {
  DCMF_CriticalSection_enter(0);
  volatile unsigned active = 1;
  DCMF_Callback_t cb = { cb_decr, (void *) &active };
  DCMF_Request_t req;
  DCMF_GlobalBarrier(&barrier_prot, &req, cb);
  while (active)
   DCMF_Messager_advance();
  DCMF_CriticalSection_exit(0);
}

extern spe_program_handle_t spe1; // defined in SPU code
int main(int argc, char *argv[]) {
  int n = ...;
  DCMF_Messager_initialize();
  { // init barrier, put, reduce
   DCMF_GlobalBarrier_Configuration_t barrier_conf =
    {DCMF_DEFAULT_GLOBALBARRIER_PROTOCOL};
   DCMF_Put_Configuration_t put_conf =
    {DCMF_DEFAULT_PUT_PROTOCOL};
   DCMF_GlobalAllreduce_Configuration_t reduce_conf =
    {DCMF_TREE_GLOBALALLREDUCE_PROTOCOL};
   DCMF_CriticalSection_enter(0);
   DCMF_GlobalBarrier_register(&barrier_prot, &barrier_conf);
   DCMF_Put_register(&put_prot, &put_conf);
   DCMF_GlobalAllreduce_register(&reduce_prot, &reduce_conf);
   DCMF_CriticalSection_exit(0);
  }
  unsigned me = DCMF_Messager_rank();
  unsigned numproc = DCMF_Messager_size();
  int p_left = -1, p_right = -1;
  if (me > 0)
   p_left = me-1;
  if (me < numproc-1)
   p_right = me+1;
  int n_local0 = 1 + (me * (n-1)) / numproc;
  int n_local1 = 1 + ((me+1) * (n-1)) / numproc;
  // allocate only local part + ghost zone of the arrays x,y
  float *x, *y;
  x = new float[n_local1 - n_local0 + 2];
  y = new float[n_local1 - n_local0 + 2];
  x -= (n_local0 - 1);
  y -= (n_local0 - 1);
  // ghost zones
  DCMF_Memregion_t memregion0, memregion1,
   memregion_left, memregion_right;
  size_t bytes;
  DCMF_CriticalSection_enter(0);
  DCMF_Memregion_create(&memregion0, &bytes,
   2 * sizeof(float), &y[n_local0-1], 0);
  DCMF_Memregion_create(&memregion1, &bytes,
   2 * sizeof(float), &y[n_local1-1], 0);
  // set memregion_left, memregion_right
  DCMF_Control_Configuration_t c0_conf =
   { DCMF_DEFAULT_CONTROL_PROTOCOL, cb_recv, &memregion_right};
  DCMF_Control_Configuration_t c1_conf =
   { DCMF_DEFAULT_CONTROL_PROTOCOL, cb_recv, &memregion_left};
  DCMF_Control_register(&control0_prot, &c0_conf);
  DCMF_Control_register(&control1_prot, &c1_conf);
  barrier();
  if (p_left != -1)
   DCMF_Control(&control0_prot, DCMF_MATCH_CONSISTENCY,
    p_left, (DCMF_Control_t*) &memregion0);
  if (p_right != -1)
   DCMF_Control(&control1_prot, DCMF_MATCH_CONSISTENCY,
    p_right, (DCMF_Control_t*) &memregion1);
  barrier();
  DCMF_CriticalSection_exit(0);

  ... // fill x, y

  { // fill ghost zone
  volatile unsigned active0 = 1, active1 = 1;
  DCMF_Callback_t cb0 = { cb_decr, (void*)&active0 },
   cb1 = { cb_decr, (void*)&active1 };
  DCMF_Request_t req0, req1;
  DCMF_CriticalSection_enter(0);
  if (p_left != -1)
   DCMF_Put(&put_prot, &req0, cb0, DCMF_SEQUENTIAL_CONSISTENCY,
    p_left, sizeof(float), &memregion0, &memregion_left,
    sizeof(float), sizeof(float));
  if (p_right != -1)
   DCMF_Put(&put_prot, &req1, cb1, DCMF_SEQUENTIAL_CONSISTENCY,
    p_right, sizeof(float), &memregion1, &memregion_right,
    sizeof(float), 0);
  if (p_left != -1)
   while (active0)
    DCMF_Messager_advance();
  if (p_right != -1)
   while (active1)
    DCMF_Messager_advance();
  DCMF_CriticalSection_exit(0);
  barrier();
  }

  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;
  }

  { // reduction
  DCMF_CriticalSection_enter(0);
  float e_local = e;
  volatile unsigned active = 1;
  DCMF_Callback_t cb = { cb_decr, (void*)&active };
  DCMF_Request_t req;
  DCMF_GlobalAllreduce(&reduce_prot, &req, cb,
   DCMF_MATCH_CONSISTENCY, -1,
   (char*)&e_local, (char*)&e, 1, DCMF_FLOAT, DCMF_SUM);
  while (active)
   DCMF_Messager_advance();
  DCMF_CriticalSection_exit(0);
  }

  ... // output x, e

  barrier();
  DCMF_Memregion_destroy(&memregion0);
  DCMF_Memregion_destroy(&memregion1);
  x += (n_local0 - 1);
  y += (n_local0 - 1);
  delete[] x, y;
  DCMF_Messager_finalize();
  return 0;
}
[start] [references] [download] [install]