code example
PVM parallel virtual machine
Cell Cell Broadband Engine
Use multiple SPUs via multi-threading.
AltiVec PowerPC 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* yp = &y[i+1], y0 = &y[i], ym = &y[i-1];
   vec_st(vec_madd(
    vec_splats(.5),
    vec_add(
     vec_perm(vec_ld(0,ym), vec_ld(16,ym),
      vec_lvsl(0,ym)),
     vec_perm(vec_ld(0,yp), vec_ld(16,yp),
      vec_lvsl(0,yp))),
    vec_splats(0.)),
    0, &x[i]);
   vec_st(vec_add(
    vec_ld(0,&ve[0]),vec_madd(
     vec_ld(0,y0),
     vec_ld(0,y0),
     vec_splats(0.))),
    0, &ve[0]);
  }
  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 "pvm3.h"
#include <libspe2.h>
#include "sub1.h"
#include <altivec.h>


extern spe_program_handle_t spe1; // defined in SPU code
int main(int argc, char *argv[]) {
  int n = ...;
  if (pvm_parent() == PvmNoParent) {
  #define N (4)
  int tid[N];
  pvm_spawn("program", argv, PvmTaskDefault, (char*)0, N, &tid[0]);
  } else {
  int mytid = pvm_mytid();
  int *tids, me = -1;
  int ntids = pvm_siblings(&tids);
  for (int i=0; i<ntids; ++i)
   if ( tids[i] == mytid) {
    me = i;
    break;
   }
  int p_left = -1, p_right = -1;
  if (me > 0)
    p_left = tids[me-1];
  if (me < ntids-1)
    p_right = tids[me+1];
  int n_local0 = 1 + (me * (n-1)) / ntids;
  int n_local1 = 1 + ((me+1) * (n-1)) / ntids;
  pvm_joingroup("worker");   // 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);

  ... // fill x, y

  // fill ghost zone
  if (p_left != -1) {
   pvm_initsend(PvmDataDefault);
   pvm_pkfloat(&y[n_local0], 1, 1);
   int msgtag = 1;
   pvm_send(p_left, msgtag);
  }
  if (p_right != -1) {
   int msgtag = 1;
   pvm_recv(p_right, msgtag);
   pvm_upkfloat(&y[n_local1], 1, 1);
   pvm_initsend(PvmDataDefault);
   pvm_pkfloat(&y[n_local1-1], 1, 1);
   msgtag = 2;
   pvm_send(p_right, msgtag);
  }
  if (p_left != -1) {
   int msgtag = 2;
   pvm_recv(p_left, msgtag);
   pvm_upkfloat(&y[n_local0-1], 1, 1);
  }

  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;
  float e = 0;
  // execute code on a single SPU and wait for termination
  spe_context_run(ctxs[p], &entry, 0, &block, NULL, st);
  spe_context_destroy(ctxs);
  e += block.e;

  int msgtag = 3;
  pvm_reduce(PvmSum, &e, 1, PVM_FLOAT, msgtag, "worker", tids[0]);
  msgtag = 4;
  if (me==0) {
   pvm_initsend(PvmDataDefault);
   pvm_pkfloat(&e, 1, 1);
   pvm_bcast("worker", msgtag);
  } else {
   pvm_recv(tids[0], msgtag);
   pvm_upkfloat(&e, 1, 1);
  }

  ... // output x, e

  x += (n_local0 - 1);
  y += (n_local0 - 1);
  delete[] x, y;
  }
  pvm_exit();
  return 0;
}
[start] [references] [download] [install]