スレッド数などの負荷調整関連を実装。
branches/mty-makai/util.h (diff)
branches/mty-makai/mty.c (diff)| @@ -54,6 +54,16 @@ | ||
| 54 | 54 | return popcnt32(x) + popcnt32(x >> 32); |
| 55 | 55 | } |
| 56 | 56 | |
| 57 | +#ifdef DEBUG | |
| 58 | + #define DPRINT( ... ) \ | |
| 59 | + printf( "[DBG] %s:%d ", __FILE__, __LINE__ ); \ | |
| 60 | + printf( __VA_ARGS__ ) | |
| 61 | +#else | |
| 62 | + #define DPRINT( ... ) | |
| 63 | +#endif | |
| 64 | + | |
| 65 | +#define OLDPRINT( ... ) | |
| 66 | + | |
| 57 | 67 | #endif /* UTIL_H__ */ |
| 58 | 68 | |
| 59 | 69 | /* |
| @@ -34,6 +34,21 @@ | ||
| 34 | 34 | #include "synth.h" |
| 35 | 35 | #include "util.h" |
| 36 | 36 | |
| 37 | +#ifdef WIN32 | |
| 38 | +/* 優先度 */ | |
| 39 | +#define PRIO_MIN 0 | |
| 40 | +#define PRIO_NORM 0 | |
| 41 | +#define PRIO_BELO 1 | |
| 42 | +#define PRIO_IDLE 2 | |
| 43 | +#define PRIO_MAX 2 | |
| 44 | +#define PRIO_DEF 2 | |
| 45 | +#endif /* WIN32 */ | |
| 46 | + | |
| 47 | +/* スレッド数と CPU 数 */ | |
| 48 | +#define THREAD_MIN 1 | |
| 49 | +#define THREAD_MAX 32 | |
| 50 | +int availCPU; /* usage() で使うので大域変数 */ | |
| 51 | + | |
| 37 | 52 | static HANDLE mutex_key; |
| 38 | 53 | |
| 39 | 54 | /* CRYPT64 記述子 */ |
| @@ -331,6 +346,7 @@ | ||
| 331 | 346 | (unsigned)mask, |
| 332 | 347 | (unsigned)mask_s); |
| 333 | 348 | #endif |
| 349 | +#if 0 | |
| 334 | 350 | if (popcnt64(mask_s) == 1) |
| 335 | 351 | /* 何も言うまい */; |
| 336 | 352 | else if (mask == mask_s) |
| @@ -341,6 +357,7 @@ | ||
| 341 | 357 | fprintf(stderr, |
| 342 | 358 | "最高速力の%g倍の力でてきとうにがんばるよ。\n", |
| 343 | 359 | (double)popcnt64(mask) / popcnt64(mask_s)); |
| 360 | +#endif /* 0 */ | |
| 344 | 361 | return mask; |
| 345 | 362 | |
| 346 | 363 | #elif defined(__linux__) /* sched.h 拡張 */ |
| @@ -408,6 +425,7 @@ | ||
| 408 | 425 | *ploop += N_ALU * ALU_BITS; |
| 409 | 426 | } |
| 410 | 427 | while (key_inc(&key, 6, 8) || key_inc(&key, KEY_SHUFFLE_POS, 8)); |
| 428 | + DPRINT( "saaaaalt chaaaaange\n" ); | |
| 411 | 429 | |
| 412 | 430 | WaitForSingleObject(mutex_key, INFINITE); |
| 413 | 431 | key_reset(&key, 0); |
| @@ -417,6 +435,54 @@ | ||
| 417 | 435 | /* notreached */ |
| 418 | 436 | } |
| 419 | 437 | |
| 438 | +char | |
| 439 | +*bname( char *path ) | |
| 440 | +{ | |
| 441 | + char *p; | |
| 442 | + | |
| 443 | + OLDPRINT( "path <%s>\n", path ); | |
| 444 | + | |
| 445 | + for ( p = path; *p != '\0'; p++ ) | |
| 446 | + ; | |
| 447 | + while ( *p != '/' && *p != '\\' && p != path ) { | |
| 448 | + if ( *p == '.' ) { | |
| 449 | + *p = '\0'; | |
| 450 | + } | |
| 451 | + p--; | |
| 452 | + } | |
| 453 | + if ( p != path ) { | |
| 454 | + p++; | |
| 455 | + } | |
| 456 | + | |
| 457 | + OLDPRINT( "p <%s>\n", p ); | |
| 458 | + | |
| 459 | + return( p ); | |
| 460 | +} | |
| 461 | + | |
| 462 | +void | |
| 463 | +usage( path ) | |
| 464 | +char *path; | |
| 465 | +{ | |
| 466 | + char *myName; | |
| 467 | + | |
| 468 | + myName = bname( path ); | |
| 469 | + printf( "まあ、待て屋。魔改造版 (%s)\n", KIND ); | |
| 470 | + printf( "%s [-h] [-v] [-p num] [-t num|-m mask]\n", myName ); | |
| 471 | + printf( " -h : これを表示\n" ); | |
| 472 | + printf( " -v : 冗長メッセージ\n" ); | |
| 473 | + printf( " -p num : 優先度の設定 ( %d ≦ num ≦ %d, デフォルトは %d )\n", | |
| 474 | + PRIO_MIN, PRIO_MAX, PRIO_DEF ); | |
| 475 | + printf( " %d : 通常\n", PRIO_NORM ); | |
| 476 | + printf( " %d : 通常以下\n", PRIO_BELO ); | |
| 477 | + printf( " %d : 低\n", PRIO_IDLE ); | |
| 478 | + printf( " ※ -t と -m は排他 (あとから指定したほうが有効)\n" ); | |
| 479 | + printf( " ※ -m で指定できるのは有効な論理 CPU 数分まで\n" ); | |
| 480 | + printf( " -t num : 検索スレッド数 ( %d ≦ num ≦ %d, デフォルトは %d )\n", | |
| 481 | + THREAD_MIN, availCPU, availCPU ); | |
| 482 | + printf( " -m mask : 実行する CPU を指定するマスク ( 1 ビット ≦ mask のビット数 ≦ %d ビット )\n", | |
| 483 | + THREAD_MAX ); | |
| 484 | +} | |
| 485 | + | |
| 420 | 486 | /*************************************************************** |
| 421 | 487 | * |
| 422 | 488 | * メインループとか |
| @@ -464,6 +530,135 @@ | ||
| 464 | 530 | exit(1); |
| 465 | 531 | } |
| 466 | 532 | |
| 533 | +{ | |
| 534 | + int optChar; | |
| 535 | + extern char *optarg; | |
| 536 | + extern int optind; | |
| 537 | + char *chPtr; | |
| 538 | + uint64_t availMask; | |
| 539 | + int nThread; | |
| 540 | + uint64_t pmask; | |
| 541 | + DWORD priority; | |
| 542 | + int verbose; | |
| 543 | + | |
| 544 | + availMask = thread_avail(); /* 有効なマスク */ | |
| 545 | + availCPU = popcnt64( availMask ); /* 有効な CPU 数 */ | |
| 546 | + | |
| 547 | + priority = PRIO_DEF; | |
| 548 | + nThread = 0; | |
| 549 | + pmask = 0; | |
| 550 | + verbose = 0; | |
| 551 | + | |
| 552 | + /* abcdefg ijkl no qrs u wxyz 未使用 */ | |
| 553 | + /* h m p t v 使用済み */ | |
| 554 | + while ( (optChar = getopt(argc, argv, "hm:p:t:v")) != EOF ) { | |
| 555 | + switch ( optChar ) { | |
| 556 | + case 'h': | |
| 557 | + usage( argv[0] ); | |
| 558 | + exit( 0 ); | |
| 559 | + break; | |
| 560 | + case 'm': | |
| 561 | + nThread = 0; /* スレッド数とは排他 */ | |
| 562 | + if ( strlen( optarg ) > THREAD_MAX ) { | |
| 563 | + usage( argv[0] ); | |
| 564 | + exit( 1 ); | |
| 565 | + } | |
| 566 | + for ( chPtr = optarg; *chPtr != '\0'; chPtr++ ) { | |
| 567 | + pmask <<= 1; | |
| 568 | + switch ( *chPtr ) { | |
| 569 | + case '0': /* なにもしない */ break; | |
| 570 | + case '1': pmask |= 1; break; | |
| 571 | + default: | |
| 572 | + usage( argv[0] ); | |
| 573 | + exit( 1 ); | |
| 574 | + break; | |
| 575 | + } | |
| 576 | + } | |
| 577 | + if ( pmask == 0 ) { | |
| 578 | + usage( argv[0] ); | |
| 579 | + exit( 1 ); | |
| 580 | + } | |
| 581 | + break; | |
| 582 | + case 'p': | |
| 583 | + priority = atoi( optarg ); | |
| 584 | + if ( priority < PRIO_MIN || priority > PRIO_MAX ) { | |
| 585 | + usage( argv[0] ); | |
| 586 | + exit( 1 ); | |
| 587 | + } | |
| 588 | + break; | |
| 589 | + case 't': | |
| 590 | + pmask = 0; /* マスクとは排他 */ | |
| 591 | + nThread = atoi( optarg ); | |
| 592 | + if ( nThread < THREAD_MIN || nThread > availCPU ) { | |
| 593 | + usage( argv[0] ); | |
| 594 | + exit( 1 ); | |
| 595 | + } | |
| 596 | + break; | |
| 597 | + case 'v': | |
| 598 | + verbose = 1; | |
| 599 | + break; | |
| 600 | + default: | |
| 601 | + usage( argv[0] ); | |
| 602 | + exit( 1 ); | |
| 603 | + } | |
| 604 | + } | |
| 605 | + | |
| 606 | + if ( pmask == 0 && nThread == 0 ) { | |
| 607 | + /* 指定がなければ、使える CPU 全部 */ | |
| 608 | + proc_mask = availMask; | |
| 609 | + } | |
| 610 | + | |
| 611 | + if ( pmask != 0 ) { | |
| 612 | + /* マスクで指定の場合、そのまんま */ | |
| 613 | + if ( (availMask & pmask) != pmask ) { | |
| 614 | + printf( "そんな CPU はねぇ！\n" ); | |
| 615 | + exit( 1 ); | |
| 616 | + } | |
| 617 | + proc_mask = pmask; | |
| 618 | + } | |
| 619 | + | |
| 620 | + if ( nThread != 0 ) { | |
| 621 | + /* スレッド数の場合、マスクに変換 */ | |
| 622 | + proc_mask = 0; /* 念のため */ | |
| 623 | + for ( i = 0; i < THREAD_MAX; i++ ) { | |
| 624 | + if ( availMask & 1ULL<<i ) { | |
| 625 | + if ( nThread > 0 ) { | |
| 626 | + proc_mask |= 1ULL<<i; | |
| 627 | + } | |
| 628 | + nThread--; | |
| 629 | + } | |
| 630 | + } | |
| 631 | + } | |
| 632 | + | |
| 633 | + printf( "%d 個の検索スレッドを起動\n", popcnt64( proc_mask ) ); | |
| 634 | + if ( verbose ) { | |
| 635 | + int i; | |
| 636 | + printf( "優先度を" ); | |
| 637 | + switch ( priority ) { | |
| 638 | + case PRIO_NORM : printf( "通常" ); break; | |
| 639 | + case PRIO_BELO : printf( "通常以下" ); break; | |
| 640 | + case PRIO_IDLE : printf( "低" ); break; | |
| 641 | + } | |
| 642 | + printf( "に設定\n" ); | |
| 643 | + printf( "CPU : " ); | |
| 644 | + for ( i = 0; i < THREAD_MAX; i++ ) { | |
| 645 | + if ( proc_mask & 1ULL<<i ) { | |
| 646 | + printf( "%d ", i ); | |
| 647 | + } | |
| 648 | + } | |
| 649 | + printf( "を使用\n" ); | |
| 650 | + } | |
| 651 | + | |
| 652 | +#ifdef WIN32 | |
| 653 | + switch ( priority ) { | |
| 654 | + case PRIO_NORM : priority = NORMAL_PRIORITY_CLASS; break; | |
| 655 | + case PRIO_BELO : priority = BELOW_NORMAL_PRIORITY_CLASS; break; | |
| 656 | + case PRIO_IDLE : priority = IDLE_PRIORITY_CLASS; break; | |
| 657 | + } | |
| 658 | + SetPriorityClass( GetCurrentProcess(), priority ); | |
| 659 | +#endif | |
| 660 | +} | |
| 661 | + | |
| 467 | 662 | assert((1 << N_STRIDE) == N_ALU * ALU_BITS); |
| 468 | 663 | |
| 469 | 664 | mutex_key = CreateMutex(NULL, FALSE, NULL); |
| @@ -476,7 +671,6 @@ | ||
| 476 | 671 | 生成するコードを変える */ |
| 477 | 672 | sfp = scoreboard_open(); |
| 478 | 673 | fwrite(crypt64_descs[0]->pro, 1, crypt64_descs[0]->cmp_pro - crypt64_descs[0]->pro, sfp); /* prologue & コアループ */ |
| 479 | - proc_mask = thread_avail(); | |
| 480 | 674 | |
| 481 | 675 | #if 0 |
| 482 | 676 | if (0&&proc_mask == 1U) |
