libgo/runtime/parfor.c

   1 // Copyright 2012 The Go Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style
   3 // license that can be found in the LICENSE file.
   4
   5 // Parallel for algorithm.
   6
   7 #include "runtime.h"
   8 #include "arch.h"
   9
  10 struct ParForThread
  11 {
  12         // the thread's iteration space [32lsb, 32msb)
  13         uint64 pos;
  14         // stats
  15         uint64 nsteal;
  16         uint64 nstealcnt;
  17         uint64 nprocyield;
  18         uint64 nosyield;
  19         uint64 nsleep;
  20         byte pad[CacheLineSize];
  21 };
  22
  23 ParFor*
  24 runtime_parforalloc(uint32 nthrmax)
  25 {
  26         ParFor *desc;
  27
  28         // The ParFor object is followed by CacheLineSize padding
  29         // and then nthrmax ParForThread.
  30         desc = (ParFor*)runtime_malloc(sizeof(ParFor) + CacheLineSize + nthrmax * sizeof(ParForThread));
  31         desc->thr = (ParForThread*)((byte*)(desc+1) + CacheLineSize);
  32         desc->nthrmax = nthrmax;
  33         return desc;
  34 }
  35
  36 void
  37 runtime_parforsetup(ParFor *desc, uint32 nthr, uint32 n, void *ctx, bool wait, void (*body)(ParFor*, uint32))
  38 {
  39         uint32 i, begin, end;
  40         uint64 *pos;
  41
  42         if(desc == nil || nthr == 0 || nthr > desc->nthrmax || body == nil) {
  43                 runtime_printf("desc=%p nthr=%d count=%d body=%p\n", desc, nthr, n, body);
  44                 runtime_throw("parfor: invalid args");
  45         }
  46
  47         desc->body = body;
  48         desc->done = 0;
  49         desc->nthr = nthr;
  50         desc->thrseq = 0;
  51         desc->cnt = n;
  52         desc->ctx = ctx;
  53         desc->wait = wait;
  54         desc->nsteal = 0;
  55         desc->nstealcnt = 0;
  56         desc->nprocyield = 0;
  57         desc->nosyield = 0;
  58         desc->nsleep = 0;
  59         for(i=0; i<nthr; i++) {
  60                 begin = (uint64)n*i / nthr;
  61                 end = (uint64)n*(i+1) / nthr;
  62                 pos = &desc->thr[i].pos;
  63                 if(((uintptr)pos & 7) != 0)
  64                         runtime_throw("parforsetup: pos is not aligned");
  65                 *pos = (uint64)begin | (((uint64)end)<<32);
  66         }
  67 }
  68
  69 void
  70 runtime_parfordo(ParFor *desc)
  71 {
  72         ParForThread *me;
  73         uint32 tid, begin, end, begin2, try, victim, i;
  74         uint64 *mypos, *victimpos, pos, newpos;
  75         void (*body)(ParFor*, uint32);
  76         bool idle;
  77
  78         // Obtain 0-based thread index.
  79         tid = runtime_xadd(&desc->thrseq, 1) - 1;
  80         if(tid >= desc->nthr) {
  81                 runtime_printf("tid=%d nthr=%d\n", tid, desc->nthr);
  82                 runtime_throw("parfor: invalid tid");
  83         }
  84
  85         // If single-threaded, just execute the for serially.
  86         if(desc->nthr==1) {
  87                 for(i=0; i<desc->cnt; i++)
  88                         desc->body(desc, i);
  89                 return;
  90         }
  91
  92         body = desc->body;
  93         me = &desc->thr[tid];
  94         mypos = &me->pos;
  95         for(;;) {
  96                 for(;;) {
  97                         // While there is local work,
  98                         // bump low index and execute the iteration.
  99                         pos = runtime_xadd64(mypos, 1);
 100                         begin = (uint32)pos-1;
 101                         end = (uint32)(pos>>32);
 102                         if(begin < end) {
 103                                 body(desc, begin);
 104                                 continue;
 105                         }
 106                         break;
 107                 }
 108
 109                 // Out of work, need to steal something.
 110                 idle = false;
 111                 for(try=0;; try++) {
 112                         // If we don't see any work for long enough,
 113                         // increment the done counter...
 114                         if(try > desc->nthr*4 && !idle) {
 115                                 idle = true;
 116                                 runtime_xadd(&desc->done, 1);
 117                         }
 118                         // ...if all threads have incremented the counter,
 119                         // we are done.
 120                         if(desc->done + !idle == desc->nthr) {
 121                                 if(!idle)
 122                                         runtime_xadd(&desc->done, 1);
 123                                 goto exit;
 124                         }
 125                         // Choose a random victim for stealing.
 126                         victim = runtime_fastrand1() % (desc->nthr-1);
 127                         if(victim >= tid)
 128                                 victim++;
 129                         victimpos = &desc->thr[victim].pos;
 130                         for(;;) {
 131                                 // See if it has any work.
 132                                 pos = runtime_atomicload64(victimpos);
 133                                 begin = (uint32)pos;
 134                                 end = (uint32)(pos>>32);
 135                                 if(begin+1 >= end) {
 136                                         begin = end = 0;
 137                                         break;
 138                                 }
 139                                 if(idle) {
 140                                         runtime_xadd(&desc->done, -1);
 141                                         idle = false;
 142                                 }
 143                                 begin2 = begin + (end-begin)/2;
 144                                 newpos = (uint64)begin | (uint64)begin2<<32;
 145                                 if(runtime_cas64(victimpos, pos, newpos)) {
 146                                         begin = begin2;
 147                                         break;
 148                                 }
 149                         }
 150                         if(begin < end) {
 151                                 // Has successfully stolen some work.
 152                                 if(idle)
 153                                         runtime_throw("parfor: should not be idle");
 154                                 runtime_atomicstore64(mypos, (uint64)begin | (uint64)end<<32);
 155                                 me->nsteal++;
 156                                 me->nstealcnt += end-begin;
 157                                 break;
 158                         }
 159                         // Backoff.
 160                         if(try < desc->nthr) {
 161                                 // nothing
 162                         } else if (try < 4*desc->nthr) {
 163                                 me->nprocyield++;
 164                                 runtime_procyield(20);
 165                         // If a caller asked not to wait for the others, exit now
 166                         // (assume that most work is already done at this point).
 167                         } else if (!desc->wait) {
 168                                 if(!idle)
 169                                         runtime_xadd(&desc->done, 1);
 170                                 goto exit;
 171                         } else if (try < 6*desc->nthr) {
 172                                 me->nosyield++;
 173                                 runtime_osyield();
 174                         } else {
 175                                 me->nsleep++;
 176                                 runtime_usleep(1);
 177                         }
 178                 }
 179         }
 180 exit:
 181         runtime_xadd64(&desc->nsteal, me->nsteal);
 182         runtime_xadd64(&desc->nstealcnt, me->nstealcnt);
 183         runtime_xadd64(&desc->nprocyield, me->nprocyield);
 184         runtime_xadd64(&desc->nosyield, me->nosyield);
 185         runtime_xadd64(&desc->nsleep, me->nsleep);
 186         me->nsteal = 0;
 187         me->nstealcnt = 0;
 188         me->nprocyield = 0;
 189         me->nosyield = 0;
 190         me->nsleep = 0;
 191 }
 192
 193 // For testing from Go.
 194 void
 195 runtime_parforiters(ParFor *desc, uintptr tid, uintptr *start, uintptr *end)
 196 {
 197         *start = (uint32)desc->thr[tid].pos;
 198         *end = (uint32)(desc->thr[tid].pos>>32);
 199 }