Home My Page Projects Code Snippets Project Openings diderot
Summary Activity Tracker Tasks SCM

SCM Repository

[diderot] Annotation of /trunk/src/lib/parallel-target/main.c
ViewVC logotype

Annotation of /trunk/src/lib/parallel-target/main.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1671 - (view) (download) (as text)

1 : jhr 1232 /*! \file main.c
2 :     *
3 :     * \author John Reppy
4 :     */
5 :    
6 :     /*
7 :     * COPYRIGHT (c) 2011 The Diderot Project (http://diderot-language.cs.uchicago.edu)
8 :     * All rights reserved.
9 :     */
10 :    
11 :     #include <Diderot/diderot.h>
12 :     #include <pthread.h>
13 : jhr 1640 #include <teem/nrrd.h>
14 : jhr 1232
15 :     // #ifdef HAVE_BUILTIN_ATOMIC_OPS
16 :     // STATIC_INLINE uint32_t AtomicInc (uint32_t *x)
17 :     // {
18 :     // return __sync_add_and_fetch(x, 1);
19 :     // }
20 :     // STATIC_INLINE uint32_t AtomicDec (uint32_t *x)
21 :     // {
22 :     // return __sync_sub_and_fetch(x, 1);
23 :     // }
24 :     // #else
25 :     // # error atomic operations not supported
26 :     // #endif
27 :    
28 :     // The number of strands a worker will take for processing at one time
29 : jhr 1640 #define BLOCK_SIZE 4096
30 : jhr 1232
31 : jhr 1640 // cache-line alignment (Xeon has 64-byte lines)
32 :     #define CACHE_ALIGN __attribute__((aligned(64)))
33 :    
34 : jhr 1232 struct struct_world {
35 : jhr 1640 STRUCT_WORLD_PREFIX
36 : jhr 1301 void **inState;
37 :     void **outState;
38 :     uint32_t numWorkers; // number of worker threads
39 :     uint32_t nSteps; // number of super steps
40 : jhr 1232 // synchronization state
41 : jhr 1640 uint32_t nextStrand CACHE_ALIGN; // index of next strand to update
42 :     uint32_t numActive CACHE_ALIGN; // # active strands
43 :     uint32_t numAvail CACHE_ALIGN; // # unevaluated strands
44 :     uint32_t numIdle CACHE_ALIGN; // # idle workers
45 : jhr 1301 pthread_mutex_t lock; // big lock to protect wrld from multiple accesses
46 :     pthread_cond_t barrier; // workers wait on this when they have no work
47 :     pthread_cond_t mainWait; // used to signal main when the workers have finished
48 : jhr 1232 };
49 :    
50 :     typedef struct {
51 : jhr 1301 int id;
52 :     Diderot_World_t *wrld;
53 : jhr 1640 } WorkerArg_t CACHE_ALIGN;
54 : jhr 1232
55 : jhr 1671 static void *Worker (void *arg);
56 : jhr 1301
57 : jhr 1671 int main (int argc, const char **argv)
58 :     {
59 :     CPUInfo_t cpuInfo;
60 :     if (! GetNumCPUs (&cpuInfo)) {
61 :     fprintf(stderr, "unable to get number of processors\n");
62 :     exit (1);
63 :     }
64 :    
65 :     Diderot_int_t np = cpuInfo.numHWCores;
66 :    
67 :     Diderot_Options_t *opts = Diderot_OptNew ();
68 :    
69 :     Diderot_OptAddInt (opts, "np", "specify number of worker threads", &np, true);
70 :     Diderot_RegisterGlobalOpts (opts);
71 :     Diderot_OptProcess (opts, argc, argv);
72 :     Diderot_OptFree (opts);
73 :    
74 :     if (VerboseFlg) fprintf (stderr, "initializing globals ...\n");
75 :     Diderot_InitGlobals ();
76 :    
77 :     if (VerboseFlg) fprintf (stderr, "initializing strands ...\n");
78 :     Diderot_World_t *wrld = Diderot_Initially ();
79 :    
80 :     for (int i = 0; i < wrld->numStrands; i++) {
81 :     // hack to make the invariant part of the state the same in both copies
82 :     memcpy (wrld->outState[i], wrld->inState[i], Diderot_Strands[0]->stateSzb);
83 :     }
84 :    
85 :     // Start worker threads
86 :     int nWorkers = np;
87 :     WorkerArg_t *args = (WorkerArg_t *) malloc (nWorkers * sizeof(WorkerArg_t));
88 :     if (VerboseFlg) printf ("initializing %d workers ...\n", nWorkers);
89 :     double t0 = airTime();
90 :     wrld->numWorkers = nWorkers;
91 :     wrld->numIdle = 0;
92 :     for (int i = 0; i < nWorkers; i++) {
93 :     pthread_t pid;
94 :     args[i].wrld = wrld;
95 :     args[i].id = i;
96 :     if (pthread_create (&pid, NULL, Worker, (void *)&(args[i])) != 0) {
97 :     fprintf (stderr, "unable to create worker thread\n");
98 :     exit (1);
99 :     }
100 :     pthread_detach (pid);
101 :     }
102 :    
103 :     // wait for the computation to finish
104 :     pthread_mutex_lock (&wrld->lock);
105 :     pthread_cond_wait (&wrld->mainWait, &wrld->lock);
106 :     pthread_mutex_unlock (&wrld->lock);
107 :    
108 :     double totalTime = airTime() - t0;
109 :    
110 :     if (VerboseFlg)
111 :     fprintf (stderr, "done: %d steps, in %f seconds\n", wrld->nSteps, totalTime);
112 :     else if (TimingFlg)
113 :     printf ("np=%d usr=%f\n", nWorkers, totalTime);
114 :    
115 :     // output the final strand states
116 :     if (NrrdOutputFlg)
117 :     Diderot_Output (wrld, Diderot_Strands[0]->outputSzb);
118 :     else
119 :     Diderot_Print (wrld);
120 :    
121 :     Diderot_Shutdown (wrld);
122 :    
123 :     return 0;
124 :    
125 :     }
126 :    
127 : jhr 1232 /* Function which processes active strands. */
128 :     static void *Worker (void *arg)
129 :     {
130 : jhr 1301 WorkerArg_t *myArg = (WorkerArg_t *)arg;
131 :     Diderot_World_t *wrld = myArg->wrld;
132 : jhr 1232
133 : jhr 1301 int nStrandsPerWorker = wrld->numStrands / wrld->numWorkers;
134 :     int start = myArg->id * nStrandsPerWorker;
135 :     int limit;
136 :     if (wrld->numWorkers-1 == myArg->id)
137 :     limit = wrld->numStrands;
138 :     else
139 :     limit = start + nStrandsPerWorker;
140 :    
141 : jhr 1232 while (true) {
142 :     // barrier synchronization at start of super step
143 : jhr 1301 pthread_mutex_lock (&wrld->lock);
144 :     if (wrld->numIdle+1 < wrld->numWorkers) {
145 :     wrld->numIdle++;
146 :     pthread_cond_wait (&wrld->barrier, &wrld->lock);
147 :     }
148 :     else {
149 :     // all other workers are idle, so we can proceed after some initialization
150 :     wrld->numIdle = 0;
151 :     wrld->numAvail = wrld->numStrands; // includes inactive strands
152 :     wrld->nextStrand = 0;
153 :     // swap in and out
154 :     void **tmp = wrld->inState;
155 :     wrld->inState = wrld->outState;
156 :     wrld->outState = tmp;
157 :     pthread_cond_broadcast (&wrld->barrier);
158 :     }
159 :     pthread_mutex_unlock (&wrld->lock);
160 : jhr 1232
161 :     // if there are no active strands left, then we're done
162 : jhr 1301 if (wrld->numActive == 0) {
163 :     pthread_cond_signal (&wrld->mainWait);
164 :     pthread_exit (0);
165 :     }
166 : jhr 1232
167 :     // iterate until there is no more work to do
168 : jhr 1301 int blkStart, blkSize;
169 :     int numDead = 0;
170 :     do {
171 :     // grab some work
172 :     pthread_mutex_lock (&wrld->lock);
173 :     blkStart = wrld->nextStrand;
174 :     blkSize = (wrld->numAvail >= BLOCK_SIZE) ? BLOCK_SIZE : wrld->numAvail;
175 :     wrld->numAvail -= blkSize;
176 :     wrld->nextStrand += blkSize;
177 :     pthread_mutex_unlock (&wrld->lock);
178 :     // update the strands
179 :     for (int i = blkStart; i < blkStart+blkSize; i++) {
180 :     if (! wrld->status[i]) {
181 :     StrandStatus_t sts = Diderot_Strands[0]->update(wrld->inState[i], wrld->outState[i]);
182 :     switch (sts) {
183 :     case DIDEROT_STABILIZE:
184 :     wrld->status[i] = DIDEROT_STABILIZE;
185 :     break;
186 :     case DIDEROT_DIE:
187 :     wrld->status[i] = DIDEROT_DIE;
188 :     numDead++;
189 :     break;
190 :     default:
191 :     break;
192 :     }
193 :     }
194 :     else {
195 :     assert ((wrld->status[i] == DIDEROT_STABLE) || (wrld->status[i] == DIDEROT_DIE));
196 :     }
197 :     }
198 :     } while (blkSize > 0);
199 : jhr 1232
200 :     // barrier synchronization
201 : jhr 1301 pthread_mutex_lock (&wrld->lock);
202 :     wrld->numActive -= numDead;
203 :     if (wrld->numIdle+1 < wrld->numWorkers) {
204 :     wrld->numIdle++;
205 :     pthread_cond_wait (&wrld->barrier, &wrld->lock);
206 :     }
207 :     else {
208 :     // all other workers are idle, so we can proceed
209 :     wrld->numIdle = 0;
210 :     pthread_cond_broadcast (&wrld->barrier);
211 :     wrld->nSteps++;
212 :     }
213 :     pthread_mutex_unlock (&wrld->lock);
214 : jhr 1232
215 :     /**** If there is a global computation phase, it goes here ****/
216 :    
217 :     // stabilize any threads that need stabilization. Each worker is responsible for
218 :     // a contiguous region of the strands
219 : jhr 1640 // FIXME: once we switch to dynamic lists of strand blocks, then we can use finer-grain tracking
220 : jhr 1301 int numStabilized = 0;
221 :     for (int i = start; i < limit; i++) {
222 :     if (wrld->status[i] == DIDEROT_STABILIZE) {
223 : jhr 1640 // stabilize the strand's state. Note that the outState has been set by
224 :     // the last call to update, so we make the inState be the target of the
225 :     // stabilize method.
226 :     Diderot_Strands[0]->stabilize(wrld->outState[i], wrld->inState[i]);
227 :     memcpy (wrld->outState[i], wrld->inState[i], Diderot_Strands[0]->stateSzb);
228 : jhr 1301 wrld->status[i] = DIDEROT_STABLE;
229 :     numStabilized++;
230 :     }
231 :     }
232 :     // adjust the numActive count
233 : jhr 1232 #if defined(HAVE_BUILTIN_ATOMIC_OPS)
234 : jhr 1301 __sync_fetch_and_sub(&wrld->numActive, numStabilized);
235 : jhr 1232 #else
236 : jhr 1301 pthread_mutex_lock (&wrld->lock);
237 :     wrld->numActive -= numStabilized;
238 :     pthread_mutex_unlock (&wrld->lock);
239 : jhr 1232 #endif
240 :     } // end while(true)
241 :    
242 :     }
243 :    
244 :     // block allocation of an initial collection of strands
245 :     Diderot_World_t *Diderot_AllocInitially (
246 : jhr 1301 const char *name, // the name of the program
247 :     Strand_t *strand, // the type of strands being allocated
248 :     bool isArray, // is the initialization an array or collection?
249 :     uint32_t nDims, // depth of iteration nesting
250 :     int32_t *base, // nDims array of base indices
251 :     uint32_t *size) // nDims array of iteration sizes
252 : jhr 1232 {
253 : jhr 1640 Diderot_World_t *wrld = NEW(Diderot_World_t);
254 : jhr 1232 if (wrld == 0) {
255 : jhr 1301 fprintf (stderr, "unable to allocate world\n");
256 :     exit (1);
257 : jhr 1232 }
258 :    
259 : jhr 1301 wrld->name = name; /* NOTE: we are assuming that name is statically allocated! */
260 : jhr 1232 wrld->isArray = isArray;
261 :     wrld->nDims = nDims;
262 : jhr 1640 wrld->base = NEWVEC(int32_t, nDims);
263 :     wrld->size = NEWVEC(uint32_t, nDims);
264 : jhr 1232 size_t numStrands = 1;
265 :     for (int i = 0; i < wrld->nDims; i++) {
266 : jhr 1301 numStrands *= size[i];
267 :     wrld->base[i] = base[i];
268 :     wrld->size[i] = size[i];
269 : jhr 1232 }
270 :    
271 : jhr 1301 if (VerboseFlg) {
272 :     fprintf(stderr, "AllocInitially: %d", size[0]);
273 :     for (int i = 1; i < nDims; i++) fprintf(stderr, " x %d", size[i]);
274 :     fprintf(stderr, "\n");
275 :     }
276 : jhr 1232
277 :     // allocate the strand state pointers
278 :     wrld->numStrands = numStrands;
279 : jhr 1640 wrld->inState = NEWVEC(void *, numStrands);
280 :     wrld->outState = NEWVEC(void *, numStrands);
281 :     wrld->status = NEWVEC(uint8_t, numStrands);
282 : jhr 1232 wrld->numActive = wrld->numStrands;
283 :     wrld->nSteps = 0;
284 :     wrld->numWorkers = 0;
285 :    
286 :     // initialize strand state pointers etc.
287 : jhr 1301 for (size_t i = 0; i < numStrands; i++) {
288 : jhr 1640 wrld->inState[i] = CheckedAlloc (strand->stateSzb);
289 :     wrld->outState[i] = CheckedAlloc (strand->stateSzb);
290 : jhr 1301 wrld->status[i] = DIDEROT_ACTIVE;
291 : jhr 1232 }
292 :    
293 :     pthread_mutex_init (&wrld->lock, NULL);
294 :     pthread_cond_init (&wrld->barrier, NULL);
295 :     pthread_cond_init (&wrld->mainWait, NULL);
296 :    
297 :     return wrld;
298 :    
299 :     }
300 :    
301 :     // get strand state pointers
302 :     void *Diderot_InState (Diderot_World_t *wrld, uint32_t i)
303 :     {
304 :     assert (i < wrld->numStrands);
305 :     return wrld->inState[i];
306 :     }
307 :    
308 :     void *Diderot_OutState (Diderot_World_t *wrld, uint32_t i)
309 :     {
310 :     assert (i < wrld->numStrands);
311 :     return wrld->outState[i];
312 :     }

root@smlnj-gforge.cs.uchicago.edu
ViewVC Help
Powered by ViewVC 1.0.0