Mercurial > hg > Game > Cerium
annotate example/regex_mas/main.cc @ 1634:c0841aa109a8 draft
add number of match position.
| author | Masa <e085726@ie.u-ryukyu.ac.jp> |
|---|---|
| date | Tue, 11 Jun 2013 17:30:40 +0900 |
| parents | d2581892b8ad |
| children | aa967337bda2 |
| rev | line source |
|---|---|
| 1598 | 1 #include <stdio.h> |
| 2 #include <stdlib.h> | |
| 3 #include <string.h> | |
| 4 #include <sys/mman.h> | |
| 5 #include <sys/types.h> | |
| 6 #include <sys/stat.h> | |
| 7 #include <fcntl.h> | |
| 8 #include <unistd.h> | |
| 9 #include <sys/time.h> | |
| 10 #include "TaskManager.h" | |
| 11 #include "SchedTask.h" | |
| 12 #include "Func.h" | |
| 13 #include "WordCount.h" | |
| 14 | |
| 1630 | 15 #define EXTRA_LENGTH 3 |
| 16 | |
| 1598 | 17 /* ;TODO |
| 18 * PS3でCPU数が2以上の時に、あまりが計算されてない | |
| 19 */ | |
| 20 | |
| 21 extern void task_init(); | |
| 22 void TMend(TaskManager *); | |
|
1600
e63ce6aee28e
eraser word_count's tarces in regex_mas
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1598
diff
changeset
|
23 //static double st_time; |
|
e63ce6aee28e
eraser word_count's tarces in regex_mas
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1598
diff
changeset
|
24 //static double ed_time; |
| 1598 | 25 int all = 0; |
| 26 int use_task_array = 1; | |
| 27 int use_task_creater = 0; | |
| 28 int use_compat = 0; | |
| 29 int array_task_num = 8; | |
| 30 int spe_num = 1; | |
| 31 CPU_TYPE spe_cpu = SPE_ANY; | |
| 32 const char *usr_help_str = "Usage: ./word_count [-a -c -s] [-cpu spe_num] [-file filename]\n"; | |
| 33 | |
|
1600
e63ce6aee28e
eraser word_count's tarces in regex_mas
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1598
diff
changeset
|
34 //static double |
|
e63ce6aee28e
eraser word_count's tarces in regex_mas
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1598
diff
changeset
|
35 //getTime() { |
|
e63ce6aee28e
eraser word_count's tarces in regex_mas
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1598
diff
changeset
|
36 // struct timeval tv; |
|
e63ce6aee28e
eraser word_count's tarces in regex_mas
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1598
diff
changeset
|
37 // gettimeofday(&tv, NULL); |
|
e63ce6aee28e
eraser word_count's tarces in regex_mas
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1598
diff
changeset
|
38 // return tv.tv_sec + (double)tv.tv_usec*1e-6; |
|
e63ce6aee28e
eraser word_count's tarces in regex_mas
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1598
diff
changeset
|
39 //} |
| 1598 | 40 |
| 41 typedef struct { | |
| 42 caddr_t file_mmap; | |
| 43 off_t size; | |
| 44 } st_mmap_t; | |
| 45 | |
| 46 static void simple_task_creater(int in_total_size, int out_total_size, | |
| 47 int command, int in_data_size, int out_data_size, | |
| 48 void *in_data, void *out_data, SchedTask *manager, | |
| 49 HTask *wait_i, HTask *wait_me) { | |
| 50 int in_task_size = 0; | |
| 51 int out_task_size = 0; | |
| 52 | |
| 53 if (in_total_size != 0) { | |
| 54 in_task_size = in_total_size / in_data_size; | |
| 55 if (in_total_size != in_task_size * in_data_size) { | |
| 56 printf("mismatch of in_total_size and in_data_size\n"); | |
| 57 } | |
| 58 } | |
| 59 | |
| 60 if (out_total_size != 0) { | |
| 61 out_task_size = out_total_size / out_data_size; | |
| 62 if (out_total_size != out_task_size * out_data_size) { | |
| 63 printf("mismatch of out_total_size and out_data_size\n"); | |
| 64 } | |
| 65 } | |
| 66 | |
| 67 /*in, out の大きい方に合わせるのがいいかな？ Taskの数は1Task分に使うデータの大きいほうを取るような仕様がいいかな*/ | |
| 68 int task_num = (in_task_size > out_task_size) ? in_task_size : out_task_size; | |
| 69 | |
| 70 if (task_num == 0) task_num = 1; | |
| 71 | |
| 72 /*spe分あればいいのかな？*/ | |
| 73 | |
| 74 int array_num = spe_num; | |
| 75 if (task_num < array_num) { | |
| 76 array_num = task_num; | |
| 77 } | |
| 78 | |
| 79 | |
| 80 HTaskPtr *task_array = (HTask**)manager->allocate(sizeof(HTask*)*array_num); | |
| 81 TaskPtr *t_exec = (Task**)manager->allocate(sizeof(Task*)*array_num); | |
| 82 | |
| 83 int array_length = task_num / array_num; | |
| 84 int rest = task_num % array_num; | |
| 85 | |
| 86 int index = 0; | |
| 87 | |
| 88 for (int k = 0; k < array_num; k++) { | |
| 89 | |
| 90 task_array[k] = manager->create_task_array(command,array_length,0,1,1); | |
| 91 t_exec[k] = 0; | |
| 92 | |
| 93 if (wait_me != 0) { | |
| 94 wait_me->wait_for(task_array[k]); | |
| 95 } | |
| 96 if (wait_i != 0) { | |
| 97 task_array[k]->wait_for(wait_i); | |
| 98 } | |
| 99 | |
| 100 } | |
| 101 | |
| 102 int length = in_data_size/sizeof(char); | |
| 103 for (int j = 0; j < array_length; j++) { | |
| 104 for (int k = 0; k < array_num; k++) { | |
| 105 | |
| 106 t_exec[k] = task_array[k]->next_task_array(command,t_exec[k]); | |
| 107 t_exec[k]->set_param(0,(memaddr)length); | |
| 108 t_exec[k]->set_inData(0,(char*)in_data + index*in_data_size, in_data_size); | |
| 109 t_exec[k]->set_outData(0,(char*)out_data + index*out_data_size, out_data_size); | |
| 110 index++; | |
| 111 | |
| 112 } | |
| 113 } | |
| 114 | |
| 115 for (int k = 0; k < array_num; k++) { | |
| 116 task_array[k]->spawn_task_array(t_exec[k]->next()); | |
| 117 task_array[k]->set_cpu(spe_cpu); | |
| 118 task_array[k]->spawn(); | |
| 119 } | |
| 120 | |
| 121 for (int k = 0; k < rest; k++) { | |
| 122 HTaskPtr t_exec = manager->create_task(command); | |
| 123 t_exec->set_param(0,(memaddr)length); | |
| 1610 | 124 int offset = length*k; |
| 125 t_exec->set_param(1,(memaddr)offset); | |
| 1598 | 126 t_exec->set_inData(0,(char*)in_data + index*in_data_size, in_data_size); |
| 127 t_exec->set_outData(0,(char*)out_data + index*out_data_size, out_data_size); | |
| 128 | |
| 129 index++; | |
| 130 | |
| 131 if (wait_me != 0) { | |
| 132 wait_me->wait_for(t_exec); | |
| 133 } | |
| 134 if (wait_i != 0) { | |
| 135 t_exec->wait_for(wait_i); | |
| 136 } | |
| 137 | |
| 138 t_exec->set_cpu(spe_cpu); | |
| 139 t_exec->spawn(); | |
| 140 | |
| 141 } | |
| 142 | |
| 143 | |
| 144 } | |
| 145 | |
| 146 | |
| 147 /*与えられたsizeをfix_byte_sizeの倍数にする(丸め込むっていうのかな？)*/ | |
| 148 static int | |
| 149 fix_byte(int size,int fix_byte_size) | |
| 150 { | |
| 151 size = (size/fix_byte_size)*fix_byte_size + ((size%fix_byte_size)!= 0)*fix_byte_size; | |
| 152 | |
| 153 return size; | |
| 154 } | |
| 155 | |
| 156 | |
| 157 static st_mmap_t | |
| 158 my_mmap(char *filename) | |
| 159 { | |
| 160 | |
| 161 /*マッピングだよ!*/ | |
| 162 int fd = -1; | |
| 163 int map = MAP_PRIVATE; | |
| 164 st_mmap_t st_mmap; | |
| 165 struct stat sb; | |
| 166 | |
| 167 if ((fd=open(filename,O_RDONLY,0666))==0) { | |
| 168 fprintf(stderr,"can't open %s\n",filename); | |
| 169 } | |
| 170 | |
| 171 if (fstat(fd,&sb)) { | |
| 172 fprintf(stderr,"can't fstat %s\n",filename); | |
| 173 } | |
| 174 | |
| 175 /*sizeをページングサイズの倍数にあわせる*/ | |
| 1630 | 176 st_mmap.size = fix_byte(sb.st_size,4096 + EXTRA_LENGTH); //セグフォ防止 |
| 1598 | 177 |
| 178 st_mmap.file_mmap = (char*)mmap(NULL,st_mmap.size,PROT_READ,map,fd,(off_t)0); | |
| 179 if (st_mmap.file_mmap == (caddr_t)-1) { | |
| 180 fprintf(stderr,"Can't mmap file\n"); | |
| 181 perror(NULL); | |
| 182 exit(0); | |
| 183 } | |
| 184 | |
| 185 return st_mmap; | |
| 186 | |
| 187 } | |
| 188 | |
| 189 static void | |
| 190 run_tasks(SchedTask *manager, WordCount *w, int task_count, HTaskPtr t_next, int size) | |
| 191 { | |
| 192 | |
| 193 if (task_count < array_task_num) { | |
| 194 array_task_num = task_count; | |
| 195 if (task_count<=0) return; | |
| 196 } | |
| 197 | |
| 198 if (use_task_creater) { | |
| 199 simple_task_creater(w->file_size, w->division_out_size * w->task_num, TASK_EXEC, w->division_size, w->division_out_size, | |
| 200 w->file_mmap, w->o_data, manager, w->t_print, 0); | |
| 201 } | |
| 202 | |
| 203 if (use_task_array) { | |
| 204 | |
| 205 int spl = spe_num * array_task_num; | |
| 206 int loop = (task_count + spl - 1) / spl; | |
| 207 | |
| 208 for (int i = 0; i < loop; i += 1) { | |
| 209 | |
| 210 if (spl > w->task_num) { | |
| 211 if (w->task_num >= spe_num) { | |
| 212 array_task_num = w->task_num / spe_num; | |
| 213 } else { | |
| 214 | |
| 215 int task_num = w->task_num; | |
| 216 | |
| 217 for (int j = 0; j < task_num; j++) { | |
| 218 HTask *h_exec = 0; | |
| 219 int i = w->task_spwaned++; | |
| 220 | |
| 221 if (w->size < size) size = w->size; | |
| 222 | |
| 223 h_exec = manager->create_task(TASK_EXEC, | |
| 224 (memaddr)(w->file_mmap + i*w->division_size), size, | |
| 225 (memaddr)(w->o_data + i*w->out_size), w->division_out_size); | |
| 226 | |
| 227 if (all) { | |
| 228 w->t_print->wait_for(h_exec); | |
| 229 } else { | |
| 230 t_next->wait_for(h_exec); | |
| 231 } | |
| 232 | |
| 233 h_exec->set_cpu(spe_cpu); | |
| 234 h_exec->spawn(); | |
| 235 | |
| 236 w->size -= size; | |
| 237 if (w->size == 0) break; | |
| 238 w->task_num--; | |
| 239 | |
| 240 } | |
| 241 | |
| 242 return; | |
| 243 } | |
| 244 } | |
| 245 | |
| 246 // ここから | |
| 247 HTask **task_array = (HTask**)manager->allocate(sizeof(HTask*)*spe_num); | |
| 248 Task **t_exec = (Task**)manager->allocate(sizeof(Task*)*spe_num); | |
| 249 | |
| 250 for (int k = 0; k < spe_num; k++) { | |
| 251 task_array[k] = manager->create_task_array(TASK_EXEC,array_task_num,1,1,1); | |
| 252 t_exec[k] = 0; | |
| 253 if (all) { | |
| 254 w->t_print->wait_for(task_array[k]); | |
| 255 } else { | |
| 256 t_next->wait_for(task_array[k]); | |
| 257 } | |
| 258 } | |
| 259 | |
| 260 for (int j = 0; j < array_task_num; j++) { | |
| 261 for (int k = 0; k < spe_num; k++) { | |
| 262 | |
| 263 int a = w->task_spwaned++; | |
| 264 | |
| 265 if (w->size < size) size = w->size; | |
| 266 int length = size/sizeof(char); | |
|
1634
c0841aa109a8
add number of match position.
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1630
diff
changeset
|
267 const int ONE_TASK_LENGTH = w->division_size; |
| 1598 | 268 |
| 269 t_exec[k] = task_array[k]->next_task_array(TASK_EXEC,t_exec[k]); | |
|
1634
c0841aa109a8
add number of match position.
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1630
diff
changeset
|
270 t_exec[k]->set_param(0,(memaddr)length); |
|
c0841aa109a8
add number of match position.
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1630
diff
changeset
|
271 const int ONE_LOOP_LENGTH = array_task_num*spe_num*ONE_TASK_LENGTH; |
|
c0841aa109a8
add number of match position.
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1630
diff
changeset
|
272 const int ARRAY_LENGTH_SIZE = spe_num * ONE_TASK_LENGTH; |
|
c0841aa109a8
add number of match position.
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1630
diff
changeset
|
273 int offset = ONE_LOOP_LENGTH * i + ARRAY_LENGTH_SIZE*j + ONE_TASK_LENGTH*k; |
|
c0841aa109a8
add number of match position.
Masa <e085726@ie.u-ryukyu.ac.jp>
parents:
1630
diff
changeset
|
274 t_exec[k]->set_param(1,(memaddr)offset); |
| 1630 | 275 t_exec[k]->set_inData(0,w->file_mmap + a*w->division_size, size + EXTRA_LENGTH); //それとも、ここで指定？？ここっぽい |
| 1598 | 276 t_exec[k]->set_outData(0,w->o_data + a*w->out_size, w->division_out_size); |
| 277 | |
| 278 w->size -= size; | |
| 279 w->task_num--; | |
| 280 } | |
| 281 } | |
| 282 for (int k = 0; k < spe_num; k++) { | |
| 283 task_array[k]->spawn_task_array(t_exec[k]->next()); | |
| 284 task_array[k]->set_cpu(spe_cpu); | |
| 285 task_array[k]->spawn(); | |
| 286 } | |
| 287 | |
| 288 } | |
| 289 | |
| 290 return; | |
| 291 | |
| 292 } | |
| 293 | |
| 294 for (int i = 0; i < task_count; i += array_task_num) { | |
| 295 | |
| 296 HTask *h_exec = 0; | |
| 297 for (int j = 0; j < array_task_num; j++) { | |
| 298 int i = w->task_spwaned++; | |
| 299 if (w->size < size) size = w->size; | |
| 300 int length = size/sizeof(char); | |
| 301 if (size==0) break; | |
| 302 | |
| 303 if (use_compat) { | |
| 304 h_exec = manager->create_task(TASK_EXEC); | |
| 305 h_exec->set_param(0,(memaddr)length); | |
| 306 h_exec->set_inData(0,w->file_mmap + i*w->division_size, size); | |
| 307 h_exec->set_outData(0,w->o_data + i*w->out_size, w->division_out_size); | |
| 308 | |
| 309 | |
| 310 if (all) { | |
| 311 w->t_print->wait_for(h_exec); | |
| 312 } else { | |
| 313 t_next->wait_for(h_exec); | |
| 314 } | |
| 315 | |
| 316 h_exec->set_cpu(spe_cpu); | |
| 317 h_exec->spawn(); | |
| 318 | |
| 319 } else { | |
| 320 h_exec = manager->create_task(TASK_EXEC, | |
| 321 (memaddr)(w->file_mmap + i*w->division_size), size, | |
| 322 (memaddr)(w->o_data + i*w->out_size), w->division_out_size); | |
| 323 | |
| 324 if (all) { | |
| 325 w->t_print->wait_for(h_exec); | |
| 326 } else { | |
| 327 t_next->wait_for(h_exec); | |
| 328 } | |
| 329 | |
| 330 h_exec->set_cpu(spe_cpu); | |
| 331 h_exec->spawn(); | |
| 332 } | |
| 333 w->size -= size; | |
| 334 w->task_num--; | |
| 335 } | |
| 336 | |
| 337 } | |
| 338 | |
| 339 } | |
| 340 | |
| 341 /** | |
| 342 * このTaskは、PPE上で実行されるので、並列に実行されることはない | |
| 343 * 二つ実行されていて、Task が足りなくなることがないようにしている。 | |
| 344 */ | |
| 345 | |
| 346 SchedDefineTask1(RUN_TASK_BLOCKS,run16); | |
| 347 | |
| 348 static int | |
| 349 run16(SchedTask *manager, void *in, void *out) | |
| 350 { | |
| 351 WordCount *w = *(WordCount **)in; | |
| 352 | |
| 353 if (w->task_num < w->task_blocks) { | |
| 354 // last case | |
| 355 while (w->size >= w->division_size) | |
| 356 run_tasks(manager,w,w->task_num, w->t_print, w->division_size); | |
| 357 // remaining data | |
| 358 while (w->size>0) | |
| 359 run_tasks(manager,w,1, w->t_print, w->size); | |
| 360 // printf("run16 last %d\n",w->task_num); | |
| 361 } else { | |
| 362 HTaskPtr t_next = manager->create_task(RUN_TASK_BLOCKS, | |
| 363 (memaddr)&w->self,sizeof(memaddr),0,0); | |
| 364 w->t_print->wait_for(t_next); | |
| 365 | |
| 366 run_tasks(manager,w, w->task_blocks, t_next, w->division_size); | |
| 367 | |
| 368 t_next->spawn(); | |
| 369 // printf("run16 next %d\n",w->task_num); | |
| 370 } | |
| 371 return 0; | |
| 372 } | |
| 373 | |
| 374 | |
| 375 static int blocks = 48; | |
| 376 //static int blocks = 31 * 6 * 24; | |
| 377 static int division = 16; // in Kbyte | |
| 378 | |
| 379 static void | |
| 380 run_start(TaskManager *manager, char *filename) | |
| 381 { | |
| 382 HTaskPtr t_print; | |
| 383 | |
| 384 st_mmap_t st_mmap; | |
| 385 st_mmap = my_mmap(filename); | |
| 386 WordCount *w = (WordCount*)manager->allocate(sizeof(WordCount)); | |
| 387 // bzero(w,sizeof(WordCount)); | |
| 388 | |
| 389 //w->task_blocks = blocks; | |
| 390 w->self = w; | |
| 391 w->task_spwaned = 0; | |
| 392 | |
| 393 /*sizeはdivision_sizeの倍数にしている。*/ | |
| 394 w->size = w->file_size = st_mmap.size; | |
| 395 w->file_mmap = st_mmap.file_mmap; | |
| 396 /* 1task分のデータサイズ(byte) */ | |
| 397 if (w->size >= 1024*division) { | |
| 398 w->division_size = 1024 * division;/*16kbyte*/ | |
| 399 } else { | |
| 400 w->division_size = w->size; | |
| 401 } | |
| 402 | |
| 403 /* "word num" and "line num" */ | |
| 404 w->status_num = 2; | |
| 405 /* taskの数 */ | |
| 406 w->task_num = w->size / w->division_size; | |
| 407 w->task_num = w->task_num + (w->division_size*w->task_num < w->size); | |
| 408 int out_task_num = w->task_num; | |
| 409 | |
| 410 if(!all) { | |
| 411 w->task_blocks = blocks; | |
| 412 } else { | |
| 413 w->task_blocks = w->task_num; | |
| 414 } | |
| 415 | |
| 416 w->out_task_num = out_task_num; | |
| 417 | |
| 418 /* out用のdivision_size. statusが2つなので、あわせて16byteになるように、long long(4byte)を使用 */ | |
| 419 | |
| 420 w->division_out_size = sizeof(unsigned long long)*4; | |
| 421 int out_size = w->division_out_size*out_task_num; | |
| 422 w->o_data = (unsigned long long *)manager->allocate(out_size); | |
| 423 w->out_size = 4; | |
| 424 | |
| 425 /*各SPEの結果を合計して出力するタスク*/ | |
| 426 | |
| 427 t_print = manager->create_task(TASK_PRINT, | |
| 428 (memaddr)&w->self,sizeof(memaddr),0,0); | |
| 429 w->t_print = t_print; | |
| 430 | |
| 431 for(int i = 0;i<20;i++) { | |
| 432 /* Task を task_blocks ずつ起動する Task */ | |
| 433 /* serialize されていると仮定する... */ | |
| 434 HTaskPtr t_exec = manager->create_task(RUN_TASK_BLOCKS, | |
| 435 (memaddr)&w->self,sizeof(memaddr),0,0); | |
| 436 t_print->wait_for(t_exec); | |
| 437 t_exec->spawn(); | |
| 438 } | |
| 439 | |
| 440 t_print->spawn(); | |
| 441 } | |
| 442 | |
| 443 static char* | |
| 444 init(int argc, char **argv) | |
| 445 { | |
| 446 | |
| 447 char *filename = 0; | |
| 448 | |
| 449 for (int i = 1; argv[i]; ++i) { | |
| 450 if (strcmp(argv[i], "-file") == 0) { | |
| 451 filename = argv[i+1]; | |
| 452 } else if (strcmp(argv[i], "-division") == 0) { | |
| 453 division = atoi(argv[i+1]); | |
| 454 } else if (strcmp(argv[i], "-block") == 0) { | |
| 455 blocks = atoi(argv[i+1]); | |
| 456 } else if (strcmp(argv[i], "-a") == 0) { | |
| 457 // create task all at once | |
| 458 all = 1; | |
| 459 } else if (strcmp(argv[i], "-c") == 0) { | |
| 460 use_task_array = 0; | |
| 461 use_compat = 1; | |
| 462 } else if (strcmp(argv[i], "-s") == 0) { | |
| 463 use_task_array = 0; | |
| 464 use_compat = 0; | |
| 465 } else if (strcmp(argv[i], "-t") == 0) { | |
| 466 use_task_creater = 1; | |
| 467 use_task_array = 0; | |
| 468 use_compat = 0; | |
| 469 } else if (strcmp(argv[i], "-anum") == 0) { | |
| 470 array_task_num = atoi(argv[i+1]); | |
| 471 } else if (strcmp(argv[i], "-g") == 0 ) { | |
| 472 spe_cpu = GPU_0; | |
| 473 } else if (strcmp(argv[i], "-cpu") == 0) { | |
| 474 spe_num = atoi(argv[i+1]); | |
| 475 if (spe_num==0) spe_num = 1; | |
| 476 } | |
| 477 } | |
| 478 if (filename==0) { | |
| 479 puts(usr_help_str); | |
| 480 exit(1); | |
| 481 } | |
| 482 | |
| 483 return filename; | |
| 484 } | |
| 485 | |
| 486 | |
| 487 int | |
| 488 TMmain(TaskManager *manager, int argc, char *argv[]) | |
| 489 { | |
| 490 | |
| 491 char *filename = 0; | |
| 492 filename = init(argc, argv); | |
| 493 | |
| 494 if (filename < 0) { | |
| 495 return -1; | |
| 496 } | |
| 497 | |
| 498 task_init(); | |
| 499 run_start(manager, filename); | |
| 500 return 0; | |
| 501 } | |
| 502 | |
| 503 /* end */ |