Subversion Repositories Games.Chess Giants

Rev

Rev 33 | Rev 154 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed

Rev Author Line No. Line
33 pmbaty 1
#include "chess.h"
2
#include "data.h"
3
#include "epdglue.h"
108 pmbaty 4
#if (CPUS > 1)
5
/* modified 11/04/15 */
33 pmbaty 6
/*
7
 *******************************************************************************
8
 *                                                                             *
108 pmbaty 9
 *   Split() is the driver for the threaded parallel search in Crafty.  The    *
33 pmbaty 10
 *   basic idea is that whenever we notice that one (or more) threads are in   *
108 pmbaty 11
 *   their idle loop, we drop into Split(), from Search(), and begin a new     *
12
 *   parallel search from this node.  This is simply a problem of establishing *
13
 *   a new split point, and then letting each thread join this split point and *
14
 *   copy whatever data they need.                                             *
33 pmbaty 15
 *                                                                             *
108 pmbaty 16
 *   This is generation II of Split().  The primary differences address two    *
17
 *   specific performance-robbing issues.  (1) Excessive waiting for a split   *
18
 *   to be done, and (b) excessive waiting on specific locks.  Generation II   *
19
 *   addresses both of these to significantly improve performance.             *
20
 *                                                                             *
21
 *   The main difference between Gen I and Gen II is the effort required to    *
22
 *   split the tree and which thread(s) expend this effort.  In generation I,  *
23
 *   the parent thread was responsible for allocating a split block for each   *
24
 *   child thread, and then copying the necessary data from the parent split   *
25
 *   block to these child split blocks.  When all of this was completed, the   *
26
 *   child processes were released to start the parallel search after being    *
27
 *   held while the split / copy operations were done.  In the generation II   *
28
 *   Split() we now simply allocate a new split block for THIS thread, flag    *
29
 *   the parent split block as joinable, and then go directly to ThreadWait()  *
30
 *   which will drop us back in to the search immediately.  The idle threads   *
31
 *   are continually looping on Join() which will jump them right into this    *
32
 *   split block letting them do ALL the work of allocating a split block,     *
33
 *   filling it in, and then copying the data to their local split block.      *
34
 *   This distributes the split overhead among all the threads that split,     *
35
 *   rather than this thread having to do all the work while the other threads *
36
 *   sit idle.                                                                 *
37
 *                                                                             *
38
 *   Generation II is also much more lightweight, in that it copies only the   *
39
 *   bare minimum from parent to child.  Generation I safely copied far too    *
40
 *   much since this code was being changed regularly, but that is no longer   *
41
 *   necessary overhead.                                                       *
42
 *                                                                             *
43
 *   Generation II has a zero contention split algorithm.  In the past, when a *
44
 *   thread became idle, it posted a global split request and any thread that  *
45
 *   was at an appropriate point would try to split.  But while it was doing   *
46
 *   the splitting, other threads that were also willing to split would "get   *
47
 *   in line" because Crafty used a global lock to prevent two threads from    *
48
 *   attempting to split at the same instant in time.  They got in line, and   *
49
 *   waited for the original splitter to release the lock, but now they have   *
50
 *   no idle threads to split with.  A waste of time.  Now we allow ANY thread *
51
 *   to attempt to split at the current ply.  When we do what might be called  *
52
 *   a "gratuitous split" the only restriction is that if we have existing     *
53
 *   "gratuitous split points" (split blocks that are joinable but have not    *
54
 *   yet been joined), then we limit the number of such splits (per thread) to *
55
 *   avoid excessive overhead.                                                 *
56
 *                                                                             *
57
 *   Generation II takes another idea from DTS, the idea of "late-join".  The  *
58
 *   idea is fairly simple.  If, when a thread becomes idle, there are already *
59
 *   other split points being searched in parallel, then we will try to join   *
60
 *   one of them rather than waiting for someone to ask us to help.  We use    *
61
 *   some simple criteria:  (1) The split point must be joinable, which simply *
62
 *   means that no processor has exited the split point yet (which would let   *
63
 *   us know there is no more work here and a join would be futile);  (2) We   *
64
 *   compute an "interest" value which is a simple formula based on depth at   *
65
 *   the split point, and the number of moves already searched.  It seems less *
66
 *   risky to choose a split point with max depth AND minimum moves already    *
67
 *   searched so that there is plenty to do.  This was quite simple to add     *
68
 *   after the rest of the Generation II rewrite.  In fact, this is now THE    *
69
 *   way threads join a split point, period, which further simplifies the code *
70
 *   and improves efficiency.  IE, a thread can split when idle threads are    *
71
 *   noticed, or if it is far enough away from the tips to make the cost       *
72
 *   negligible.  At that point any idle thread(s) can join immediately, those *
73
 *   that become idle later can join when they are ready.                      *
74
 *                                                                             *
33 pmbaty 75
 *   There are a number of settable options via the command-line or .craftyrc  *
76
 *   initialization file.  Here's a concise explanation for each option and an *
77
 *   occasional suggestion for testing/tuning.                                 *
78
 *                                                                             *
108 pmbaty 79
 *   smp_affinity (command = smpaffinity=<n> is used to enable or disable      *
80
 *      processor affinity.  -1 disables affinity and lets threads run on any  *
81
 *      available core.  If you use an integer <n> then thread zero will bind  *
82
 *      itself to cpu <n> and each additional thread will bind to the next     *
83
 *      higher cpu number.  This is useful if you try to run two copies of     *
84
 *      crafty on the same machine, now you can cause one to bind to the first *
85
 *      <n> cores, and the second to the last <n> cores.  For the first        *
86
 *      instance of Crafty, you would use smpaffinity=0, and for the second    *
87
 *      smpaffinity=8, assuming you are running 8 threads per copy on a 16 cpu *
88
 *      machine.  If you get this wrong, you can have more than one thread on  *
89
 *      the same cpu which will significantly impact performance.              *
90
 *                                                                             *
33 pmbaty 91
 *   smp_max_threads (command = smpmt=n) sets the total number of allowable    *
92
 *      threads for the search.  The default is one (1) as Crafty does not     *
93
 *      assume it should use all available resources.  For optimal performance *
94
 *      this should be set to the number of physical cores your machine has,   *
95
 *      which does NOT include hyperthreaded cores.                            *
96
 *                                                                             *
97
 *   smp_split_group (command = smpgroup=n) sets the maximum number of threads *
108 pmbaty 98
 *      at any single split point, with the exception of split points fairly   *
99
 *      close to the root where ALL threads are allowed to split together,     *
100
 *      ignoring this limit.  Note that this is ignored in the first 1/2 of    *
101
 *      the tree (the nodes closer to the root).  There it is actually good to *
102
 *      split and get all active threads involved.                             *
33 pmbaty 103
 *                                                                             *
104
 *   smp_min_split_depth (command = smpmin=n) avoids splitting when remaining  *
105
 *      depth < n.  This is used to balance splitting overhead cost against    *
108 pmbaty 106
 *      the speed gain the parallel search produces.  The default is currently *
33 pmbaty 107
 *      5 (which could change with future generations of Intel hardware) but   *
108
 *      values between 4 and 8 will work.  Larger values allow somewhat fewer  *
109
 *      splits, which reduces overhead, but it also increases the percentage   *
110
 *      of the time where a thread is waiting on work.                         *
111
 *                                                                             *
108 pmbaty 112
 *   smp_split_at_root (command = smproot=0 or 1) enables (1) or disables (0)  *
33 pmbaty 113
 *      splitting the tree at the root.  This defaults to 1 which produces the *
114
 *      best performance by a signficiant margin.  But it can be disabled if   *
115
 *      you are playing with code changes.                                     *
116
 *                                                                             *
108 pmbaty 117
 *   smp_gratuitous_depth (command = smpgd=<n>) controls " gratuitous splits"  *
118
 *      which are splits that are done without any idle threads.  This sets a  *
119
 *      depth limit (remaining depth) that must be present before such a split *
120
 *      can be done.  Making this number larger will reduce the number of      *
121
 *      these splits.  Making it too small will increase overhead slightly and *
122
 *      increase split block usage significantly.                              *
33 pmbaty 123
 *                                                                             *
108 pmbaty 124
 *   smp_gratuitous_limit (command = smpgl=<n>) limits the number of these     *
125
 *      splits that a thread can do.  Once a thread has this number of         *
126
 *      unjoined split points, it will not be allowed to split any more until  *
127
 *      one or more threads join at least one of the existing split points.    *
128
 *      In the smp search statistics, where you see output that looks like     *
129
 *      this:                                                                  *
130
 *                                                                             *
131
 *        splits=m(n) ...                                                      *
132
 *                                                                             *
133
 *      m is the total splits done, n is the number of "wasted splits" which   *
134
 *      are basically gratuitous splits where no thread joined before this     *
135
 *      split point was completed and deallocated.                             *
136
 *                                                                             *
137
 *   The best way to tune all of these paramenters is to use the "autotune"    *
138
 *   command (see autotune.c and help autotune) which will automatically run   *
139
 *   tests and optimize the parameters.  More details are in the autotune.c    *
140
 *   source file.                                                              *
141
 *                                                                             *
33 pmbaty 142
 *   A few basic "rules of the road" for anyone interested in changing or      *
143
 *   adding to any of this code.                                               *
144
 *                                                                             *
145
 *   1.  If, at any time, you want to modify your private split block, no lock *
146
 *       is required.                                                          *
147
 *                                                                             *
148
 *   2.  If, at any time, you want to modify another split block, such as the  *
149
 *       parent split block shared move list, you must acquire the lock in the *
108 pmbaty 150
 *       split block first.  IE tree->parent->lock to lock the parent split    *
151
 *       block during NextMove() and such.                                     *
33 pmbaty 152
 *                                                                             *
108 pmbaty 153
 *   3.  If you want to modify any SMP-related data that spans multiple split  *
154
 *       blocks, such as telling sibling processes to stop, etc, then you must *
155
 *       acquire the global "lock_smp" lock first.  This prevents a deadlock   *
156
 *       caused by two different threads walking the split block chain from    *
157
 *       different directions, and acquiring the split block locks in          *
158
 *       different orders, which could cause a catastrophic deadlock to occur. *
159
 *       This is an infrequent event so the overhead is not significant.       *
33 pmbaty 160
 *                                                                             *
161
 *   4.  If you want to do any sort of I/O operation, you must acquire the     *
162
 *       "lock_io" lock first.  Since threads share descriptors, there are     *
163
 *       lots of potential race conditions, from the simple tty-output getting *
164
 *       interlaced from different threads, to outright data corruption in the *
165
 *       book or log files.                                                    *
166
 *                                                                             *
167
 *   Some of the bugs caused by failing to acquire the correct lock will only  *
168
 *   occur infrequently, and they are extremely difficult to find.  Some only  *
169
 *   show up in a public game where everyone is watching, to cause maximum     *
170
 *   embarassment and causes the program to do something extremely stupid.     *
171
 *                                                                             *
172
 *******************************************************************************
173
 */
108 pmbaty 174
int Split(TREE * RESTRICT tree) {
175
  TREE *child;
176
  int tid, tstart, tend;
33 pmbaty 177
 
178
/*
179
 ************************************************************
180
 *                                                          *
108 pmbaty 181
 *  Here we prepare to split the tree.  All we really do in *
182
 *  the Generation II threading is grab a split block for   *
183
 *  this thread, then flag the parent as "joinable" and     *
184
 *  then jump right to ThreadWait() to resume where we left *
185
 *  off, with the expectation (but not a requirement) that  *
186
 *  other threads will join us to help.                     *
33 pmbaty 187
 *                                                          *
108 pmbaty 188
 *  Idle threads are sitting in ThreadWait() repeatedly     *
189
 *  calling Join() to find them a split point, which we are *
190
 *  fixing to provide.  They will then join as quickly as   *
191
 *  they can, and other threads that become idle later can  *
192
 *  also join without any further splitting needed.         *
193
 *                                                          *
194
 *  If we are unable to allocate a split block, we simply   *
195
 *  abort this attempted split and return to the search     *
196
 *  since other threads will also split quickly.            *
197
 *                                                          *
33 pmbaty 198
 ************************************************************
199
 */
108 pmbaty 200
  tstart = ReadClock();
201
  tree->nprocs = 0;
202
  for (tid = 0; tid < (int) smp_max_threads; tid++) // Pierre-Marie Baty -- added type cast
203
    tree->siblings[tid] = 0;
204
  child = GetBlock(tree, tree->thread_id);
205
  if (!child)
33 pmbaty 206
    return 0;
108 pmbaty 207
  CopyFromParent(child);
208
  thread[tree->thread_id].tree = child;
209
  tree->joined = 0;
210
  tree->joinable = 1;
211
  parallel_splits++;
212
  smp_split = 0;
213
  tend = ReadClock();
214
  thread[tree->thread_id].idle += tend - tstart;
33 pmbaty 215
/*
216
 ************************************************************
217
 *                                                          *
108 pmbaty 218
 *  We have successfully created a split point, which means *
219
 *  we are done.  The instant we set the "joinable" flag,   *
220
 *  idle threads may begin to join in at this split point   *
221
 *  to help.  Since this thread may finish before any or    *
222
 *  all of the other parallel threads, this thread is sent  *
223
 *  to ThreadWait() which will immediately send it to       *
224
 *  SearchMoveList() like the other threads; however, going *
225
 *  to ThreadWait() allows this thread to join others if it *
226
 *  runs out of work to do.  We do pass ThreadWait() the    *
227
 *  address of the parent split block, so that if this      *
228
 *  thread becomes idle, and this thread block shows no     *
229
 *  threads are still busy, then this thread can return to  *
230
 *  here and then back up into the previous ply as it       *
231
 *  should.  Note that no other thread can back up to the   *
232
 *  previous ply since their recursive call stacks are not  *
233
 *  set for that, while this call stack will bring us back  *
234
 *  to this point where we return to the normal search,     *
235
 *  which we just completed.                                *
33 pmbaty 236
 *                                                          *
108 pmbaty 237
 ************************************************************
238
 */
239
  ThreadWait(tree->thread_id, tree);
240
  if (!tree->joined)
241
    parallel_splits_wasted++;
242
  return 1;
243
}
244
 
245
/* modified 12/16/15 */
246
/*
247
 *******************************************************************************
248
 *                                                                             *
249
 *   Join() is called just when we enter the usual spin-loop waiting for work. *
250
 *   We take a quick look at all active split blocks to see if any look        *
251
 *   "joinable".  If so, we compute an "interest" value, which will be defined *
252
 *   below.  We then join the most interesting split point directly. This      *
253
 *   split point might have been created specifically for this thread to join, *
254
 *   or it might be one that was already active when this thread became idle,  *
255
 *   which allows us to join that existing split point and not request a new   *
256
 *   split operation, saving time.                                             *
257
 *                                                                             *
258
 *******************************************************************************
259
 */
260
int Join(int64_t tid) {
261
  TREE *tree, *join_block, *child;
262
  int interest, best_interest, current, pass = 0;
263
 
264
/*
265
 ************************************************************
33 pmbaty 266
 *                                                          *
108 pmbaty 267
 *  First we pass over ALL split blocks, looking for those  *
268
 *  flagged as "joinable" (which means they are actually    *
269
 *  active split points and that no processor at that split *
270
 *  point has run out of work (there is no point in joining *
271
 *  a split point with no remaining work) and no fail high  *
272
 *  has been found which would raise the "stop" flag.) This *
273
 *  is "racy" because we do not acquire any locks, which    *
274
 *  means that the busy threads continue working, and there *
275
 *  is a small probability that the split point will        *
276
 *  disappear while we are in this loop.  To resolve the    *
277
 *  potential race, after we find the most attractive split *
278
 *  point, we acquire the lock for that split block and     *
279
 *  test again, but this time if the block is joinable, we  *
280
 *  can safely join under control of the lock, which is not *
281
 *  held for very long at all.  If the block is not         *
282
 *  joinable once we acquire the lock, we abort joining     *
283
 *  since it is futile.  Note that if this happens, we will *
284
 *  try to find an existing split point we can join three   *
285
 *  times before we exit, setting split to 1 to ask other   *
286
 *  threads to produce more candidate split points.         *
33 pmbaty 287
 *                                                          *
108 pmbaty 288
 ************************************************************
289
 */
290
  for (pass = 0; pass < 3; pass++) {
291
    best_interest = -999999;
292
    join_block = 0;
293
    for (current = 0; current <= (int) smp_max_threads * 64; current++) { // Pierre-Marie Baty -- added type cast
294
      tree = block[current];
295
      if (tree->joinable && (tree->ply <= tree->depth / 2 ||
296
              tree->nprocs < (int) smp_split_group)) { // Pierre-Marie Baty -- added type cast
297
        interest = tree->depth * 2 - tree->searched[0];
298
        if (interest > best_interest) {
299
          best_interest = interest;
300
          join_block = tree;
301
        }
302
      }
303
    }
304
/*
305
 ************************************************************
33 pmbaty 306
 *                                                          *
108 pmbaty 307
 *  Now we acquire the lock for this split block, and then  *
308
 *  check to see if the block is still flagged as joinable. *
309
 *  If so, we set things up, and then we get pretty tricky  *
310
 *  as we then release the lock, and then copy the data     *
311
 *  from the parent to our split block.  There is a chance  *
312
 *  that while we are copying this data, the split point    *
313
 *  gets completed by other threads.  Which would leave an  *
314
 *  apparent race condition exposed where we start copying  *
315
 *  data here, the split point is completed, the parent     *
316
 *  block is released and then reacquired and we continue   *
317
 *  if nothing has happened here, getting data copied from  *
318
 *  two different positions.                                *
319
 *                                                          *
320
 *  Fortunately, we linked this new split block to the old  *
321
 *  (original parent).  If that split block is released, we *
322
 *  will discover this because that operation will also set *
323
 *  our "stop" flag which will prevent us from using this   *
324
 *  data and breaking things.  We allow threads to copy     *
325
 *  this data without any lock protection to eliminate a    *
326
 *  serialization (each node would copy the data serially,  *
327
 *  rather than all at once) with the only consequence to   *
328
 *  this being the overhead of copying and then throwing    *
329
 *  the data away, which can happen on occasion even if we  *
330
 *  used a lock for protection, since once we release the   *
331
 *  lock it still takes time to get into the search and we  *
332
 *  could STILL find that this split block has already been *
333
 *  completed, once again.  Less contention and serial      *
334
 *  computing improves performance.                         *
335
 *                                                          *
33 pmbaty 336
 ************************************************************
337
 */
108 pmbaty 338
    if (join_block) {
339
      Lock(join_block->lock);
340
      if (join_block->joinable) {
341
        child = GetBlock(join_block, (int) tid); // Pierre-Marie Baty -- added type cast
342
        Unlock(join_block->lock);
343
        if (child) {
344
          CopyFromParent(child);
345
          thread[tid].tree = child;
346
          parallel_joins++;
347
          return 1;
348
        }
349
      } else {
350
        Unlock(join_block->lock);
33 pmbaty 351
        break;
108 pmbaty 352
      }
33 pmbaty 353
    }
354
  }
355
/*
356
 ************************************************************
357
 *                                                          *
108 pmbaty 358
 *  We did not acquire a split point to join, so we set     *
359
 *  smp_split to 1 to ask busy threads to create joinable   *
360
 *  split points.                                           *
33 pmbaty 361
 *                                                          *
362
 ************************************************************
363
 */
108 pmbaty 364
  smp_split = 1;
365
  return 0;
366
}
367
 
368
/* modified 11/04/15 */
33 pmbaty 369
/*
108 pmbaty 370
 *******************************************************************************
371
 *                                                                             *
372
 *   ThreadInit() is called after a process is created.  Its main task is to   *
373
 *   initialize the process local memory so that it will fault in and be       *
374
 *   allocated on the local node rather than the node where the original       *
375
 *   (first) process was running.  All threads will hang here via a custom     *
376
 *   WaitForALlThreadsInitialized() procedure so that all the local thread     *
377
 *   blocks are usable before the search actually begins.                      *
378
 *                                                                             *
379
 *******************************************************************************
380
 */
381
void *STDCALL ThreadInit(void *t) {
382
  int tid = (int64_t) t;
383
 
384
  ThreadAffinity(tid);
385
#  if !defined(UNIX)
386
  ThreadMalloc((uint64_t) tid);
387
#  endif
388
  thread[tid].blocks = 0xffffffffffffffffull;
389
  Lock(lock_smp);
390
  initialized_threads++;
391
  Unlock(lock_smp);
392
  WaitForAllThreadsInitialized();
393
  ThreadWait(tid, (TREE *) 0);
394
  Lock(lock_smp);
395
  smp_threads--;
396
  Unlock(lock_smp);
397
  return 0;
398
}
399
 
400
/* modified 01/08/15 */
401
/*
402
 *******************************************************************************
403
 *                                                                             *
404
 *   ThreadAffinity() is called to "pin" a thread to a specific processor.  It *
405
 *   is a "noop" (no-operation) if Crafty was not compiled with -DAFFINITY, or *
406
 *   if smp_affinity is negative (smpaffinity=-1 disables affinity).  It       *
407
 *   simply sets the affinity for the current thread to the requested CPU and  *
408
 *   returns.  NOTE:  If hyperthreading is enabled, there is no guarantee that *
409
 *   this will work as expected and pin one thread per physical core.  It      *
410
 *   depends on how the O/S numbers the SMT cores.                             *
411
 *                                                                             *
412
 *******************************************************************************
413
 */
414
void ThreadAffinity(int cpu) {
415
#  if defined(AFFINITY)
416
  cpu_set_t cpuset;
417
  pthread_t current_thread = pthread_self();
418
 
419
  if (smp_affinity >= 0) {
420
    CPU_ZERO(&cpuset);
421
    CPU_SET(cpu + smp_affinity, &cpuset);
422
    pthread_setaffinity_np(current_thread, sizeof(cpu_set_t), &cpuset);
423
  }
424
#  endif
425
}
426
 
427
/* modified 11/04/15 */
428
/*
429
 *******************************************************************************
430
 *                                                                             *
431
 *   ThreadSplit() is used to determine if we should split at the current ply. *
432
 *   There are some basic constraints on when splits can be done, such as the  *
433
 *   depth remaining in the search (don't split to near the tips), and have we *
434
 *   searched at least one move to get a score or bound (YBW condition).       *
435
 *                                                                             *
436
 *   If those conditions are satisfied, AND either a thread has requested a    *
437
 *   split OR we are far enough away from the tips of the tree to justify a    *
438
 *   "gratuitout split" then we return "success."  A "gratuitout split" is a   *
439
 *   split done without any idle threads.  Since splits are not free, we only  *
440
 *   do this well away from tips to limit overhead.  We do this so that when a *
441
 *   thread becomes idle, it will find these split points immediately and not  *
442
 *   have to wait for a split after the fact.                                  *
443
 *                                                                             *
444
 *******************************************************************************
445
 */
446
int ThreadSplit(TREE *RESTRICT tree, int ply, int depth, int alpha, int o_alpha, // Pierre-Marie Baty -- missing keyword
447
    int done) {
448
  TREE *used;
449
  int64_t tblocks;
450
  int temp, unused = 0;
451
 
452
/*
33 pmbaty 453
 ************************************************************
454
 *                                                          *
108 pmbaty 455
 *  First, we see if we meet the basic criteria to create a *
456
 *  split point, that being that we must not be too far     *
457
 *  from the root (smp_min_split_depth).                    *
33 pmbaty 458
 *                                                          *
459
 ************************************************************
460
 */
108 pmbaty 461
  if (depth < (int) smp_min_split_depth) // Pierre-Marie Baty -- added type cast
462
    return 0;
33 pmbaty 463
/*
464
 ************************************************************
465
 *                                                          *
108 pmbaty 466
 *  If smp_split is NOT set, we are checking to see if it   *
467
 *  is acceptable to do a gratuitous split here.            *
33 pmbaty 468
 *                                                          *
108 pmbaty 469
 *  (1) if we are too far from the root we do not do        *
470
 *      gratuitous splits to avoid the overhead.            *
471
 *                                                          *
472
 *  (2) if we have searched more than one move at this ply, *
473
 *      we don't do any further tests to see if a           *
474
 *      gratuitous split is acceptable, since we have       *
475
 *      previously done this test at this ply and decided   *
476
 *      one should not be done.  That condition has likely  *
477
 *      not changed.                                        *
478
 *                                                          *
479
 *  (3) if we have pre-existing gratuitous split points for *
480
 *      this thread, we make certain we don't create more   *
481
 *      than the gratuitous split limit as excessive splits *
482
 *      just add to the overhead with no benefit.           *
483
 *                                                          *
33 pmbaty 484
 ************************************************************
485
 */
108 pmbaty 486
  if (!smp_split) {
487
    if (depth < (int) smp_gratuitous_depth || done > 1) // Pierre-Marie Baty -- added type cast
488
      return 0;
489
    tblocks = ~thread[tree->thread_id].blocks;
490
    while (tblocks) {
491
      temp = LSB(tblocks);
492
      used = block[temp + tree->thread_id * 64 + 1];
493
      if (used->joinable && !used->joined)
494
        unused++;
495
      Clear(temp, tblocks);
496
    }
497
    if (unused > (int) smp_gratuitous_limit) // Pierre-Marie Baty -- added type cast
498
      return 0;
499
  }
500
/*
501
 ************************************************************
502
 *                                                          *
503
 *  If smp_split IS set, we are checking to see if it is    *
504
 *  acceptable to do a split because there are idle threads *
505
 *  that need work to do.                                   *
506
 *                                                          *
507
 *  The only reason this would be false is if we have a     *
508
 *  pre-existing split point that is joinable but has not   *
509
 *  been joined. If one exists, there is no need to split   *
510
 *  again as there is already an accessible split point.    *
511
 *  Otherwise, if we are at the root and we are either not  *
512
 *  allowed to split at the root, or we have additional     *
513
 *  root moves that have to be searched one at a time using *
514
 *  all available threads we also can not split here.       *
515
 *                                                          *
516
 ************************************************************
517
 */
518
  else {
519
    if (ply == 1 && (!smp_split_at_root || !NextRootMoveParallel() ||
520
            alpha == o_alpha))
521
      return 0;
522
    tblocks = ~thread[tree->thread_id].blocks;
523
    while (tblocks) {
524
      temp = LSB(tblocks);
525
      used = block[temp + tree->thread_id * 64 + 1];
526
      if (used->joinable && !used->joined)
527
        unused++;
528
      Clear(temp, tblocks);
529
    }
530
    if (unused > (int) smp_gratuitous_limit) // Pierre-Marie Baty -- added type cast
531
      return 0;
532
  }
33 pmbaty 533
  return 1;
534
}
535
 
108 pmbaty 536
/* modified 11/04/15 */
33 pmbaty 537
/*
538
 *******************************************************************************
539
 *                                                                             *
108 pmbaty 540
 *   ThreadStop() is called from SearchMoveList() when it detects a beta       *
541
 *   cutoff (fail high) at a node that is being searched in parallel.  We need *
542
 *   to stop all threads here, and since this threading algorithm is recursive *
543
 *   it may be necessary to stop other threads that are helping search this    *
544
 *   branch further down into the tree.  This function simply sets appropriate *
545
 *   tree->stop variables to 1, which will stop those particular threads       *
546
 *   instantly and return them to the idle loop in ThreadWait().               *
33 pmbaty 547
 *                                                                             *
548
 *******************************************************************************
549
 */
108 pmbaty 550
void ThreadStop(TREE * RESTRICT tree) {
551
  int proc;
33 pmbaty 552
 
108 pmbaty 553
  Lock(tree->lock);
554
  tree->stop = 1;
555
  tree->joinable = 0;
556
  for (proc = 0; proc < (int) smp_max_threads; proc++) // Pierre-Marie Baty -- added type cast
557
    if (tree->siblings[proc])
558
      ThreadStop(tree->siblings[proc]);
559
  Unlock(tree->lock);
560
}
561
 
562
/* modified 11/04/15 */
33 pmbaty 563
/*
108 pmbaty 564
 *******************************************************************************
565
 *                                                                             *
566
 *   ThreadTrace() is a debugging tool that simply walks the split block tree  *
567
 *   and displays interesting data to help debug the parallel search whenever  *
568
 *   changes break things.                                                     *
569
 *                                                                             *
570
 *******************************************************************************
571
 */
572
void ThreadTrace(TREE * RESTRICT tree, int depth, int brief) {
573
  int proc, i;
574
 
575
  Lock(tree->lock);
576
  Lock(lock_io);
577
  if (!brief) {
578
    for (i = 0; i < 4 * depth; i++)
579
      Print(4095, " ");
580
    depth++;
581
    Print(4095, "block[%d]  thread=%d  ply=%d  nprocs=%d  ",
582
        FindBlockID(tree), tree->thread_id, tree->ply, tree->nprocs);
583
    Print(4095, "joined=%d  joinable=%d  stop=%d  nodes=%d", tree->joined,
584
        tree->joinable, tree->stop, tree->nodes_searched);
585
    Print(4095, "  parent=%d\n", FindBlockID(tree->parent));
586
  } else {
587
    if (tree->nprocs > 1) {
588
      for (i = 0; i < 4 * depth; i++)
589
        Print(4095, " ");
590
      depth++;
591
      Print(4095, "(ply %d)", tree->ply);
592
    }
593
  }
594
  if (tree->nprocs) {
595
    if (!brief) {
596
      for (i = 0; i < 4 * depth; i++)
597
        Print(4095, " ");
598
      Print(4095, "          parent=%d  sibling threads=",
599
          FindBlockID(tree->parent));
600
      for (proc = 0; proc < (int) smp_max_threads; proc++) // Pierre-Marie Baty -- added type cast
601
        if (tree->siblings[proc])
602
          Print(4095, " %d(%d)", proc, FindBlockID(tree->siblings[proc]));
603
      Print(4095, "\n");
604
    } else {
605
      if (tree->nprocs > 1) {
606
        Print(4095, " helping= ");
607
        for (proc = 0; proc < (int) smp_max_threads; proc++) // Pierre-Marie Baty -- added type cast
608
          if (tree->siblings[proc]) {
609
            if (proc == tree->thread_id)
610
              Print(4095, "[");
611
            Print(4095, "%d", proc);
612
            if (proc == tree->thread_id)
613
              Print(4095, "]");
614
            Print(4095, " ");
615
          }
616
        Print(4095, "\n");
617
      }
618
    }
619
  }
620
  Unlock(lock_io);
621
  for (proc = 0; proc < (int) smp_max_threads; proc++) // Pierre-Marie Baty -- added type cast
622
    if (tree->siblings[proc])
623
      ThreadTrace(tree->siblings[proc], depth, brief);
624
  Unlock(tree->lock);
625
}
626
 
627
/* modified 11/04/15 */
628
/*
629
 *******************************************************************************
630
 *                                                                             *
631
 *   ThreadWait() is the idle loop for the N threads that are created at the   *
632
 *   beginning when Crafty searches.  Threads are "parked" here waiting on a   *
633
 *   pointer to something they should search (a parameter block built in the   *
634
 *   function Split() in this case.  When this pointer becomes non-zero, each  *
635
 *   thread "parked" here will immediately call SearchMoveList() and begin the *
636
 *   parallel search as directed.                                              *
637
 *                                                                             *
638
 *   Upon entry, all threads except for the "master" will arrive here with a   *
639
 *   value of zero (0) in the waiting parameter below.  This indicates that    *
640
 *   they will search and them be done.  The "master" will arrive here with a  *
641
 *   pointer to the parent split block in "waiting" which says I will sit here *
642
 *   waiting for work OR when the waiting split block has no threads working   *
643
 *   on it, at which point I should return which will let me "unsplit" here    *
644
 *   and clean things up.  The call to here in Split() passes this block       *
645
 *   address while threads that are helping get here with a zero.              *
646
 *                                                                             *
647
 *******************************************************************************
648
 */
649
int ThreadWait(int tid, TREE * RESTRICT waiting) {
650
  int value, tstart, tend;
651
 
652
/*
33 pmbaty 653
 ************************************************************
654
 *                                                          *
108 pmbaty 655
 *  When we reach this point, one of three possible         *
656
 *  conditions is true (1) we already have work to do, as   *
657
 *  we are the "master thread" and we have already split    *
658
 *  the tree, we are coming here to join in;  (2) we are    *
659
 *  the master, and we are waiting on our split point to    *
660
 *  complete, so we come here to join and help currently    *
661
 *  active threads;  (3) we have no work to do, so we will  *
662
 *  spin until Join() locates a split pont we can join to   *
663
 *  help out.                                               *
33 pmbaty 664
 *                                                          *
108 pmbaty 665
 *  Note that when we get here, the parent already has a    *
666
 *  split block and does not need to call Join(), it simply *
667
 *  falls through the while spin loop below because its     *
668
 *  "tree" pointer is already non-zero.                     *
669
 *                                                          *
33 pmbaty 670
 ************************************************************
671
 */
108 pmbaty 672
  while (1) {
673
    tstart = ReadClock();
674
    while (!thread[tid].tree && (!waiting || waiting->nprocs) && !Join(tid) &&
675
        !thread[tid].terminate)
676
      Pause();
677
    tend = ReadClock();
678
    if (!thread[tid].tree)
679
      thread[tid].tree = waiting;
680
    thread[tid].idle += tend - tstart;
681
    if (thread[tid].tree == waiting || thread[tid].terminate)
682
      return 0;
683
/*
684
 ************************************************************
685
 *                                                          *
686
 *  Once we get here, we have a good split point, so we are *
687
 *  ready to participate in a parallel search.  Once we     *
688
 *  return from SearchMoveList() we copy our results back   *
689
 *  to the parent via CopyToParent() before we look for a   *
690
 *  new split point.  If we are a parent, we will slip out  *
691
 *  of the spin loop at the top and return to the normal    *
692
 *  serial search to finish up here.                        *
693
 *                                                          *
694
 *  When we return from SearchMoveList(), we need to        *
695
 *  decrement the "nprocs" value since there is now one     *
696
 *  less thread working at this split point.                *
697
 *                                                          *
698
 *  Note:  CopyToParent() marks the current split block as  *
699
 *  unused once the copy is completed, so we don't have to  *
700
 *  do anything about that here.                            *
701
 *                                                          *
702
 ************************************************************
703
 */
704
    value =
705
        SearchMoveList(thread[tid].tree, thread[tid].tree->ply,
706
        thread[tid].tree->depth, thread[tid].tree->wtm,
707
        thread[tid].tree->alpha, thread[tid].tree->beta,
708
        thread[tid].tree->searched, thread[tid].tree->in_check, 0, parallel);
709
    tstart = ReadClock();
710
    Lock(thread[tid].tree->parent->lock);
711
    thread[tid].tree->parent->joinable = 0;
712
    CopyToParent((TREE *) thread[tid].tree->parent, thread[tid].tree, value);
713
    thread[tid].tree->parent->nprocs--;
714
    thread[tid].tree->parent->siblings[tid] = 0;
715
    Unlock(thread[tid].tree->parent->lock);
716
    thread[tid].tree = 0;
717
    tend = ReadClock();
718
    thread[tid].idle += tend - tstart;
33 pmbaty 719
  }
108 pmbaty 720
}
721
 
722
/* modified 11/04/15 */
33 pmbaty 723
/*
108 pmbaty 724
 *******************************************************************************
725
 *                                                                             *
726
 *   CopyFromParent() is used to copy data from a parent thread to a child     *
727
 *   thread.  This only copies the appropriate parts of the TREE structure to  *
728
 *   avoid burning memory bandwidth by copying everything.                     *
729
 *                                                                             *
730
 *******************************************************************************
731
 */
732
void CopyFromParent(TREE * RESTRICT child) {
733
  TREE *parent = child->parent;
734
  int i, ply;
735
 
736
/*
33 pmbaty 737
 ************************************************************
738
 *                                                          *
739
 *  We have allocated a split block.  Now we copy the tree  *
740
 *  search state from the parent block to the child in      *
741
 *  preparation for starting the parallel search.           *
742
 *                                                          *
743
 ************************************************************
744
 */
108 pmbaty 745
  ply = parent->ply;
33 pmbaty 746
  child->ply = ply;
747
  child->position = parent->position;
108 pmbaty 748
  for (i = 0; i <= rep_index + parent->ply; i++)
33 pmbaty 749
    child->rep_list[i] = parent->rep_list[i];
108 pmbaty 750
  for (i = ply - 1; i < MAXPLY; i++)
751
    child->killers[i] = parent->killers[i];
752
  for (i = ply - 1; i <= ply; i++) {
33 pmbaty 753
    child->curmv[i] = parent->curmv[i];
108 pmbaty 754
    child->pv[i] = parent->pv[i];
33 pmbaty 755
  }
108 pmbaty 756
  child->in_check = parent->in_check;
757
  child->last[ply] = child->move_list;
758
  child->status[ply] = parent->status[ply];
759
  child->status[1] = parent->status[1];
760
  child->save_hash_key[ply] = parent->save_hash_key[ply];
761
  child->save_pawn_hash_key[ply] = parent->save_pawn_hash_key[ply];
33 pmbaty 762
  child->nodes_searched = 0;
763
  child->fail_highs = 0;
764
  child->fail_high_first_move = 0;
765
  child->evaluations = 0;
766
  child->egtb_probes = 0;
108 pmbaty 767
  child->egtb_hits = 0;
33 pmbaty 768
  child->extensions_done = 0;
769
  child->qchecks_done = 0;
770
  child->moves_fpruned = 0;
108 pmbaty 771
  child->moves_mpruned = 0;
772
  for (i = 0; i < 16; i++) {
773
    child->LMR_done[i] = 0;
774
    child->null_done[i] = 0;
775
  }
33 pmbaty 776
  child->alpha = parent->alpha;
777
  child->beta = parent->beta;
778
  child->value = parent->value;
108 pmbaty 779
  child->wtm = parent->wtm;
33 pmbaty 780
  child->depth = parent->depth;
108 pmbaty 781
  child->searched = parent->searched;
782
  strcpy(child->root_move_text, parent->root_move_text);
783
  strcpy(child->remaining_moves_text, parent->remaining_moves_text);
33 pmbaty 784
}
785
 
108 pmbaty 786
/* modified 11/04/15 */
33 pmbaty 787
/*
788
 *******************************************************************************
789
 *                                                                             *
790
 *   CopyToParent() is used to copy data from a child thread to a parent       *
791
 *   thread.  This only copies the appropriate parts of the TREE structure to  *
792
 *   avoid burning memory bandwidth by copying everything.                     *
793
 *                                                                             *
794
 *******************************************************************************
795
 */
796
void CopyToParent(TREE * RESTRICT parent, TREE * RESTRICT child, int value) {
108 pmbaty 797
  int i, ply = parent->ply, which;
33 pmbaty 798
 
799
/*
800
 ************************************************************
801
 *                                                          *
108 pmbaty 802
 *  The only concern here is to make sure that the info is  *
803
 *  only copied to the parent if our score is > than the    *
804
 *  parent value, and that we were not stopped for any      *
805
 *  reason which could produce a partial score that is      *
806
 *  worthless and dangerous to use.                         *
33 pmbaty 807
 *                                                          *
108 pmbaty 808
 *  One important special case.  If we get here with the    *
809
 *  thread->stop flag set, and ply is 1, then we need to    *
810
 *  clear the "this move has been searched" flag in the ply *
811
 *  1 move list since we did not complete the search.  If   *
812
 *  we fail to do this, then a move being searched in       *
813
 *  parallel at the root will be "lost" for this iteration  *
814
 *  and won't be searched again until the next iteration.   *
815
 *                                                          *
33 pmbaty 816
 *  In any case, we add our statistical counters to the     *
817
 *  parent's totals no matter whether we finished or not    *
818
 *  since the total nodes searched and such should consider *
819
 *  everything searched, not just the "useful stuff."       *
820
 *                                                          *
108 pmbaty 821
 *  After we finish copying everything, we mark this split  *
822
 *  block as free in the split block bitmap.                *
823
 *                                                          *
33 pmbaty 824
 ************************************************************
825
 */
826
  if (child->nodes_searched && !child->stop && value > parent->value &&
827
      !abort_search) {
828
    parent->pv[ply] = child->pv[ply];
829
    parent->value = value;
830
    parent->cutmove = child->curmv[ply];
831
  }
108 pmbaty 832
  if (child->stop && ply == 1)
833
    for (which = 0; which < n_root_moves; which++)
834
      if (root_moves[which].move == child->curmv[ply]) {
835
        root_moves[which].status &= 7;
836
        break;
837
      }
33 pmbaty 838
  parent->nodes_searched += child->nodes_searched;
839
  parent->fail_highs += child->fail_highs;
840
  parent->fail_high_first_move += child->fail_high_first_move;
841
  parent->evaluations += child->evaluations;
842
  parent->egtb_probes += child->egtb_probes;
108 pmbaty 843
  parent->egtb_hits += child->egtb_hits;
33 pmbaty 844
  parent->extensions_done += child->extensions_done;
845
  parent->qchecks_done += child->qchecks_done;
846
  parent->moves_fpruned += child->moves_fpruned;
108 pmbaty 847
  parent->moves_mpruned += child->moves_mpruned;
848
  for (i = 1; i < 16; i++) {
849
    parent->LMR_done[i] += child->LMR_done[i];
850
    parent->null_done[i] += child->null_done[i];
851
  }
852
  which = FindBlockID(child) - 64 * child->thread_id - 1;
853
  Set(which, thread[child->thread_id].blocks);
33 pmbaty 854
}
855
 
108 pmbaty 856
/* modified 11/04/15 */
33 pmbaty 857
/*
858
 *******************************************************************************
859
 *                                                                             *
108 pmbaty 860
 *   GetBlock() is used to allocate a split block and fill in only SMP-        *
861
 *   critical information.  The child process will copy the rest of the split  *
862
 *   block information as needed.                                              *
33 pmbaty 863
 *                                                                             *
108 pmbaty 864
 *   When we arrive here, the parent split block must be locked since we are   *
865
 *   going to change data in that block as well as copy data from that block   *
866
 *   the current split block.  The only exception is when this is the original *
867
 *   split operation, since this is done "out of sight" of other threads which *
868
 *   means no locks are needed until after the "joinable" flag is set, which   *
869
 *   exposes this split point to other threads instantly.                      *
870
 *                                                                             *
33 pmbaty 871
 *******************************************************************************
872
 */
108 pmbaty 873
TREE *GetBlock(TREE * /*RESTRICT*/ parent, int tid) { // Pierre-Marie Baty -- keyword not present in declaration
874
  TREE *child;
875
  static int warnings = 0;
876
  int i, unused;
877
/*
878
 ************************************************************
879
 *                                                          *
880
 *  One NUMA-related trick is that we only allocate a split *
881
 *  block in the thread's local memory.  Each thread has a  *
882
 *  group of split blocks that were first touched by the    *
883
 *  correct CPU so that the split blocks page-faulted into  *
884
 *  local memory for that specific processor.  If we can't  *
885
 *  find an optimally-placed block, we return a zero which  *
886
 *  will prevent this thread from joining the split point.  *
887
 *  This is highly unlikely as it would mean the current    *
888
 *  thread has 64 active split blocks where it is waiting   *
889
 *  on other threads to complete the last bit of work at    *
890
 *  each.  This is extremely unlikely.                      *
891
 *                                                          *
892
 *  Here we use a simple 64 bit bit-map per thread that     *
893
 *  indicates which blocks are free (1) and which blocks    *
894
 *  used (0).  We simply use LSB() to find the rightmost    *
895
 *  one-bit set and that is the local block number.  We     *
896
 *  convert that to a global block number by doing the      *
897
 *  simple computation:                                     *
898
 *                                                          *
899
 *     global = local + 64 * tid + 1                        *
900
 *                                                          *
901
 *  Each thread has exactly 64 split blocks, and block 0    *
902
 *  is the "master block" that never gets allocated or      *
903
 *  freed.  Once we find a free block for the current       *
904
 *  thread, we zero that bit so that the block won't be     *
905
 *  used again until it is released.                        *
906
 *                                                          *
907
 ************************************************************
908
 */
909
  if (thread[tid].blocks) {
910
    unused = LSB(thread[tid].blocks);
911
    Clear(unused, thread[tid].blocks);
912
    Set(unused, thread[tid].max_blocks);
913
  } else {
914
    if (++warnings < 6)
915
      Print(2048,
916
          "WARNING.  local SMP block cannot be allocated, thread %d\n", tid);
917
    return 0;
918
  }
919
  child = block[unused + tid * 64 + 1];
920
/*
921
 ************************************************************
922
 *                                                          *
923
 *  Found a split block.  Now we need to fill in only the   *
924
 *  critical information that can't be delayed due to race  *
925
 *  conditions.  When we get here, the parent split block   *
926
 *  must be locked, that lets us safely update the number   *
927
 *  of processors working here, etc, without any ugly race  *
928
 *  conditions that would corrupt this critical data.       *
929
 *                                                          *
930
 ************************************************************
931
 */
932
  for (i = 0; i < (int) smp_max_threads; i++) // Pierre-Marie Baty -- added type cast
933
    child->siblings[i] = 0;
934
  child->nprocs = 0;
935
  child->stop = 0;
936
  child->joinable = 0;
937
  child->joined = 0;
938
  child->parent = parent;
939
  child->thread_id = tid;
940
  parent->nprocs++;
941
  parent->siblings[tid] = child;
942
  parent->joined = 1;
943
  return child;
33 pmbaty 944
}
945
 
946
/*
947
 *******************************************************************************
948
 *                                                                             *
108 pmbaty 949
 *   WaitForAllThreadsInitialized() waits until all smp_max_threads are        *
950
 *   initialized.  We have to initialize each thread and malloc() its split    *
951
 *   blocks before we start the actual parallel search.  Otherwise we will see *
952
 *   invalid memory accesses and crash instantly.                              *
33 pmbaty 953
 *                                                                             *
954
 *******************************************************************************
955
 */
108 pmbaty 956
void WaitForAllThreadsInitialized(void) {
957
  while (initialized_threads < (int) smp_max_threads); /* Do nothing */ // Pierre-Marie Baty -- added type cast
33 pmbaty 958
}
959
 
108 pmbaty 960
#  if !defined (UNIX)
33 pmbaty 961
/* modified 01/17/09 */
962
/*
963
 *******************************************************************************
964
 *                                                                             *
965
 *   ThreadMalloc() is called from the ThreadInit() function.  It malloc's the *
966
 *   split blocks in the local memory for the processor associated with the    *
967
 *   specific thread that is calling this code.                                *
968
 *                                                                             *
969
 *******************************************************************************
970
 */
971
extern void *WinMalloc(size_t, int);
972
void ThreadMalloc(int64_t tid) {
108 pmbaty 973
  int i;
33 pmbaty 974
 
108 pmbaty 975
  for (i = (int) tid * 64 + 1; i < (int) tid * 64 + 65; i++) { // Pierre-Marie Baty -- added type casts
33 pmbaty 976
    if (block[i] == NULL)
977
      block[i] =
978
          (TREE *) ((~(size_t) 127) & (127 + (size_t) WinMalloc(sizeof(TREE) +
979
                  127, (int) tid))); // Pierre-Marie Baty -- added type cast
980
    block[i]->parent = NULL;
981
    LockInit(block[i]->lock);
982
  }
983
}
108 pmbaty 984
#  endif
33 pmbaty 985
#endif