Robust/src/Runtime/multicoreruntime.h

   1 #ifndef MULTICORE_RUNTIME
   2 #define MULTICORE_RUNTIME
   3
   4 #ifndef INLINE
   5 #define INLINE    inline __attribute__((always_inline))
   6 #endif
   7
   8 ////////////////////////////////////////////////////////////////
   9 // global variables                                          //
  10 ///////////////////////////////////////////////////////////////
  11
  12 // data structures for msgs
  13 #define BAMBOO_OUT_BUF_LENGTH 3000
  14 #define BAMBOO_MSG_BUF_LENGTH 3000
  15 int msgdata[BAMBOO_MSG_BUF_LENGTH];
  16 int msgdataindex;
  17 int msgdatalast;
  18 int msglength;
  19 int outmsgdata[BAMBOO_OUT_BUF_LENGTH];
  20 int outmsgindex;
  21 int outmsglast;
  22 int outmsgleft;
  23 bool isMsgHanging;
  24 volatile bool isMsgSending;
  25
  26 #define MSG_INDEXINC_I() \
  27         msgdataindex = (msgdataindex + 1) % (BAMBOO_MSG_BUF_LENGTH)
  28
  29 #define MSG_LASTINDEXINC_I() \
  30         msgdatalast = (msgdatalast + 1) % (BAMBOO_MSG_BUF_LENGTH)
  31
  32 #define MSG_CACHE_I(n) \
  33         msgdata[msgdatalast] = (n); \
  34   MSG_LASTINDEXINC_I()
  35
  36 // NOTE: if msgdataindex == msgdatalast, it always means that the buffer if
  37 //       full. In the case that the buffer is empty, should never call this
  38 //       MACRO
  39 #define MSG_REMAINSIZE_I(s) \
  40         if(msgdataindex < msgdatalast) { \
  41                 (*(int*)s) = msgdatalast - msgdataindex; \
  42         } else { \
  43                 (*(int*)s) = (BAMBOO_MSG_BUF_LENGTH) - msgdataindex + msgdatalast; \
  44         }
  45
  46 #define OUTMSG_INDEXINC() \
  47         outmsgindex = (outmsgindex + 1) % (BAMBOO_OUT_BUF_LENGTH)
  48
  49 #define OUTMSG_LASTINDEXINC() \
  50         outmsglast = (outmsglast + 1) % (BAMBOO_OUT_BUF_LENGTH); \
  51         if(outmsglast == outmsgindex) { \
  52                 BAMBOO_EXIT(0xdd01); \
  53         }
  54
  55 #define OUTMSG_CACHE(n) \
  56         outmsgdata[outmsglast] = (n); \
  57   OUTMSG_LASTINDEXINC();
  58
  59 /* Message format:
  60  *      type + Msgbody
  61  * type: 1 -- transfer object
  62  *       2 -- transfer stall msg
  63  *       3 -- lock request
  64  *       4 -- lock grount
  65  *       5 -- lock deny
  66  *       6 -- lock release
  67  *       // add for profile info
  68  *       7 -- transfer profile output msg
  69  *       8 -- transfer profile output finish msg
  70  *       // add for alias lock strategy
  71  *       9 -- redirect lock request
  72  *       a -- lock grant with redirect info
  73  *       b -- lock deny with redirect info
  74  *       c -- lock release with redirect info
  75  *       d -- status confirm request
  76  *       e -- status report msg
  77  *       f -- terminate
  78  *      10 -- requiring for new memory
  79  *      11 -- response for new memory request
  80  *      12 -- GC init phase start
  81  *      13 -- GC start
  82  *      14 -- compact phase start
  83  *      15 -- flush phase start
  84  *      16 -- init phase finish
  85  *      17 -- mark phase finish
  86  *      18 -- compact phase finish
  87  *      19 -- flush phase finish
  88  *      1a -- GC finish
  89  *      1b -- marked phase finish confirm request
  90  *      1c -- marked phase finish confirm response
  91  *      1d -- markedObj msg
  92  *      1e -- start moving objs msg
  93  *      1f -- ask for mapping info of a markedObj
  94  *      20 -- mapping info of a markedObj
  95  *      21 -- large objs info request
  96  *      22 -- large objs info response
  97  *      23 -- large objs mapping info
  98  *
  99  * ObjMsg: 1 + size of msg + obj's address + (task index + param index)+
 100  * StallMsg: 2 + corenum + sendobjs + receiveobjs
 101  *             (size is always 4 * sizeof(int))
 102  * LockMsg: 3 + lock type + obj pointer + lock + request core
 103  *            (size is always 5 * sizeof(int))
 104  *          4/5/6 + lock type + obj pointer + lock
 105  *            (size is always 4 * sizeof(int))
 106  *          9 + lock type + obj pointer +  redirect lock + root request core
 107  *            + request core
 108  *            (size is always 6 * sizeof(int))
 109  *          a/b + lock type + obj pointer + redirect lock
 110  *              (size is always 4 * sizeof(int))
 111  *          c + lock type + lock + redirect lock
 112  *            (size is always 4 * sizeof(int))
 113  *          lock type: 0 -- read; 1 -- write
 114  * ProfileMsg: 7 + totalexetime
 115  *               (size is always 2 * sizeof(int))
 116  *             8 + corenum
 117  *               (size is always 2 * sizeof(int))
 118  * StatusMsg: d (size is always 1 * sizeof(int))
 119  *            e + status + corenum + sendobjs + receiveobjs
 120  *              (size is always 5 * sizeof(int))
 121  *            status: 0 -- stall; 1 -- busy
 122  * TerminateMsg: f (size is always 1 * sizeof(int)
 123  * MemoryMsg: 10 + size + corenum
 124  *              (size is always 3 * sizeof(int))
 125  *           11 + base_va + size
 126  *              (size is always 3 * sizeof(int))
 127  * GCMsg: 12/13 (size is always 1 * sizeof(int))
 128  *        14 + size of msg + (num of objs to move + (start address
 129  *           + end address + dst core + start dst)+)?
 130  *           + (num of incoming objs + (start dst + orig core)+)?
 131  *           + (num of large obj lists + (start address + lenght
 132  *           + start dst)+)?
 133  *        15 (size is always 1 * sizeof(int))
 134  *        16 + corenum
 135  *           (size is always 2 * sizeof(int))
 136  *        17 + corenum + gcsendobjs + gcreceiveobjs
 137  *           (size if always 4 * sizeof(int))
 138  *        18 + corenum + fulfilled blocks num + (finish compact(1) + current
 139  *           heap top)/(need mem(0) + mem need)
 140  *           size is always 5 * sizeof(int))
 141  *        19 + corenum
 142  *              (size is always 2 * sizeof(int))
 143  *        1a (size is always 1 * sizeof(int))
 144  *        1b (size if always 1 * sizeof(int))
 145  *        1c + size of msg + corenum + gcsendobjs + gcreceiveobjs
 146  *           (size is always 5 * sizeof(int))
 147  *        1d + obj's address
 148  *           (size is always 2 * sizeof(int))
 149  *        1e + corenum + start addr + end addr
 150  *           (size if always 4 * sizeof(int))
 151  *        1f + obj's address + corenum
 152  *           (size is always 3 * sizeof(int))
 153  *        20 + obj's address + dst address
 154  *           (size if always 3 * sizeof(int))
 155  *        21 (size is always 1 * sizeof(int))
 156  *        22 + size of msg + corenum + current heap size
 157  *           + (num of large obj lists + (start address + length)+)?
 158  *        23 + orig large obj ptr + new large obj ptr
 159  *            (size is always 3 * sizeof(int))
 160  */
 161 typedef enum {
 162         MSGSTART = 0xD0, // 0xD0
 163         TRANSOBJ,        // 0xD1
 164         TRANSTALL,       // 0xD2
 165         LOCKREQUEST,     // 0xD3
 166         LOCKGROUNT,      // 0xD4
 167         LOCKDENY,        // 0xD5
 168         LOCKRELEASE,     // 0xD6
 169         PROFILEOUTPUT,   // 0xD7
 170         PROFILEFINISH,   // 0xD8
 171         REDIRECTLOCK,    // 0xD9
 172         REDIRECTGROUNT,  // 0xDa
 173         REDIRECTDENY,    // 0xDb
 174         REDIRECTRELEASE, // 0xDc
 175         STATUSCONFIRM,   // 0xDd
 176         STATUSREPORT,    // 0xDe
 177         TERMINATE,       // 0xDf
 178         MEMREQUEST,      // 0xE0
 179         MEMRESPONSE,     // 0xE1
 180 #ifdef MULTICORE_GC
 181         GCSTARTINIT,     // 0xE2
 182         GCSTART,         // 0xE3
 183         GCSTARTCOMPACT,  // 0xE4
 184         GCSTARTFLUSH,    // 0xE5
 185         GCFINISHINIT,    // 0xE6
 186         GCFINISHMARK,    // 0xE7
 187         GCFINISHCOMPACT, // 0xE8
 188         GCFINISHFLUSH,   // 0xE9
 189         GCFINISH,        // 0xEa
 190         GCMARKCONFIRM,   // 0xEb
 191         GCMARKREPORT,    // 0xEc
 192         GCMARKEDOBJ,     // 0xEd
 193         GCMOVESTART,     // 0xEe
 194         GCMAPREQUEST,    // 0xEf
 195         GCMAPINFO,       // 0xF0
 196         GCLOBJREQUEST,   // 0xF1
 197         GCLOBJINFO,      // 0xF2
 198         GCLOBJMAPPING,   // 0xF3
 199 #endif
 200         MSGEND
 201 } MSGTYPE;
 202
 203 /////////////////////////////////////////////////////////////////////////////////
 204 // NOTE: BAMBOO_TOTALCORE -- number of the available cores in the processor.
 205 //                           No greater than the number of all the cores in
 206 //                           the processor
 207 //       NUMCORES -- number of cores chosen to deploy the application. It can
 208 //                   be greater than that required to fully parallelize the
 209 //                   application. The same as NUMCORES.
 210 //       NUMCORESACTIVE -- number of cores that really execute the
 211 //                         application. No greater than NUMCORES
 212 //       NUMCORES4GC -- number of cores for gc. No greater than NUMCORES.
 213 //                      NOTE: currently only support ontinuous cores as gc
 214 //                            cores, i.e. 0~NUMCORES4GC-1
 215 ////////////////////////////////////////////////////////////////////////////////
 216 // data structures of status for termination
 217 // only check working cores
 218 volatile int corestatus[NUMCORESACTIVE]; // records status of each core
 219                                          // 1: running tasks
 220                                          // 0: stall
 221 volatile int numsendobjs[NUMCORESACTIVE]; // records how many objects a core
 222                                           // has sent out
 223 volatile int numreceiveobjs[NUMCORESACTIVE]; // records how many objects a
 224                                              // core has received
 225 volatile int numconfirm;
 226 volatile bool waitconfirm;
 227 bool busystatus;
 228 int self_numsendobjs;
 229 int self_numreceiveobjs;
 230
 231 // get rid of lock msgs for GC version
 232 #ifndef MULTICORE_GC
 233 // data structures for locking
 234 struct RuntimeHash locktable;
 235 static struct RuntimeHash* locktbl = &locktable;
 236 struct RuntimeHash * lockRedirectTbl;
 237 struct RuntimeHash * objRedirectLockTbl;
 238 #endif
 239 struct LockValue {
 240         int redirectlock;
 241         int value;
 242 };
 243 int lockobj;
 244 int lock2require;
 245 int lockresult;
 246 bool lockflag;
 247
 248 // data structures for waiting objs
 249 struct Queue objqueue;
 250 struct Queue * totransobjqueue; // queue to hold objs to be transferred
 251                                 // should be cleared whenever enter a task
 252
 253 // data structures for shared memory allocation
 254 #define BAMBOO_BASE_VA 0xd000000
 255 #ifdef GC_DEBUG
 256 #include "structdefs.h"
 257 #define BAMBOO_NUM_PAGES (NUMCORES4GC*(2+1)+3)
 258 #define BAMBOO_PAGE_SIZE (64 * 64)
 259 #define BAMBOO_SMEM_SIZE (64 * 64) // (BAMBOO_PAGE_SIZE)
 260 #define BAMBOO_SHARED_MEM_SIZE ((BAMBOO_PAGE_SIZE) * (BAMBOO_NUM_PAGES))
 261 #else
 262 #define BAMBOO_NUM_PAGES (15 * 1024) //(64 * 4 * 0.75) //(1024 * 1024 * 3.5)  3G
 263 #define BAMBOO_PAGE_SIZE (16 * 1024)// * 1024)  // (4096)
 264 #define BAMBOO_SMEM_SIZE (16 * 1024)
 265 #define BAMBOO_SHARED_MEM_SIZE (1024 * 1024 * 240) //(1024 * 1024 * 1024)
 266 //(3.0 * 1024 * 1024 * 1024) // 3G// ((BAMBOO_PAGE_SIZE) * (BAMBOO_NUM_PAGES))
 267 #endif
 268
 269 #ifdef MULTICORE_GC
 270 #include "multicoregarbage.h"
 271
 272 typedef enum {
 273         SMEMLOCAL = 0x0, // 0x0, using local mem only
 274         SMEMFIXED,       // 0x1, use local mem in lower address space(1 block only)
 275                          //      and global mem in higher address space
 276         SMEMMIXED,       // 0x2, like FIXED mode but use a threshold to control
 277         SMEMGLOBAL,      // 0x3, using global mem only
 278         SMEMEND
 279 } SMEMSTRATEGY;
 280
 281 SMEMSTRATEGY bamboo_smem_mode; //-DSMEML: LOCAL; -DSMEMF: FIXED;
 282                               //-DSMEMM: MIXED; -DSMEMG: GLOBAL;
 283
 284 struct freeMemItem {
 285         INTPTR ptr;
 286         int size;
 287         int startblock;
 288         int endblock;
 289         struct freeMemItem * next;
 290 };
 291
 292 struct freeMemList {
 293         struct freeMemItem * head;
 294         struct freeMemItem * backuplist;  // hold removed freeMemItem for reuse;
 295                                           // only maintain 1 fremmMemItem
 296 };
 297
 298 // table recording the number of allocated bytes on each block
 299 // Note: this table resides on the bottom of the shared heap for all cores
 300 //       to access
 301 volatile int * bamboo_smemtbl;
 302 volatile int bamboo_free_block;
 303 //bool bamboo_smem_flushed;
 304 //struct freeMemList * bamboo_free_mem_list;
 305 int bamboo_reserved_smem; // reserved blocks on the top of the shared heap
 306                           // e.g. 20% of the heap and should not be allocated
 307                                                                                                         // otherwise gc is invoked
 308 #else
 309 volatile mspace bamboo_free_msp;
 310 #endif
 311 volatile bool smemflag;
 312 volatile INTPTR bamboo_cur_msp;
 313 volatile int bamboo_smem_size;
 314
 315 // for test TODO
 316 int total_num_t6;
 317
 318 // data structures for profile mode
 319 #ifdef PROFILE
 320
 321 #define TASKINFOLENGTH 30000
 322 //#define INTERRUPTINFOLENGTH 500
 323
 324 bool stall;
 325 //bool isInterrupt;
 326 int totalexetime;
 327
 328 typedef struct task_info {
 329   char* taskName;
 330   unsigned long long startTime;
 331   unsigned long long endTime;
 332   unsigned long long exitIndex;
 333   struct Queue * newObjs;
 334 } TaskInfo;
 335
 336 /*typedef struct interrupt_info {
 337    int startTime;
 338    int endTime;
 339    } InterruptInfo;*/
 340
 341 TaskInfo * taskInfoArray[TASKINFOLENGTH];
 342 int taskInfoIndex;
 343 bool taskInfoOverflow;
 344 /*InterruptInfo * interruptInfoArray[INTERRUPTINFOLENGTH];
 345    int interruptInfoIndex;
 346    bool interruptInfoOverflow;*/
 347 volatile int profilestatus[NUMCORESACTIVE]; // records status of each core
 348                                             // 1: running tasks
 349                                             // 0: stall
 350 #endif // #ifdef PROFILE
 351
 352 #ifndef INTERRUPT
 353 bool reside;
 354 #endif
 355 /////////////////////////////////////////////////////////////
 356
 357 ////////////////////////////////////////////////////////////
 358 // these are functions should be implemented in           //
 359 // multicore runtime for any multicore processors         //
 360 ////////////////////////////////////////////////////////////
 361 #ifdef TASK
 362 #ifdef MULTICORE
 363 INLINE void initialization(void);
 364 INLINE void initCommunication(void);
 365 INLINE void fakeExecution(void);
 366 INLINE void terminate(void);
 367 INLINE void initlock(struct ___Object___ * v);
 368
 369 // lock related functions
 370 bool getreadlock(void* ptr);
 371 void releasereadlock(void* ptr);
 372 bool getwritelock(void* ptr);
 373 void releasewritelock(void* ptr);
 374 bool getwritelock_I(void* ptr);
 375 void releasewritelock_I(void * ptr);
 376 #ifndef MULTICORE_GC
 377 void releasewritelock_r(void * lock, void * redirectlock);
 378 #endif
 379 /* this function is to process lock requests.
 380  * can only be invoked in receiveObject() */
 381 // if return -1: the lock request is redirected
 382 //            0: the lock request is approved
 383 //            1: the lock request is denied
 384 INLINE int processlockrequest(int locktype,
 385                                           int lock,
 386                                                                                                                         int obj,
 387                                                                                                                         int requestcore,
 388                                                                                                                         int rootrequestcore,
 389                                                                                                                         bool cache);
 390 INLINE void processlockrelease(int locktype,
 391                                            int lock,
 392                                                                                                                          int redirectlock,
 393                                                                                                                          bool redirect);
 394
 395 // msg related functions
 396 INLINE void send_hanging_msg();
 397 INLINE void send_msg_1(int targetcore,
 398                                    unsigned long n0,
 399                                                                                          bool isinterrupton);
 400 INLINE void send_msg_2(int targetcore,
 401                                    unsigned long n0,
 402                                                                                          unsigned long n1,
 403                                                                                          bool isinterrupton);
 404 INLINE void send_msg_3(int targetcore,
 405                                    unsigned long n0,
 406                                                                                          unsigned long n1,
 407                                                                                          unsigned long n2,
 408                                                                                          bool isinterrupton);
 409 INLINE void send_msg_4(int targetcore,
 410                                    unsigned long n0,
 411                                                                                          unsigned long n1,
 412                                                                                          unsigned long n2,
 413                                                                                          unsigned long n3,
 414                                                                                          bool isinterrupton);
 415 INLINE void send_msg_5(int targetcore,
 416                                    unsigned long n0,
 417                                                                                          unsigned long n1,
 418                                                                                          unsigned long n2,
 419                                                                                          unsigned long n3,
 420                                                                                          unsigned long n4,
 421                                                                                          bool isinterrupton);
 422 INLINE void send_msg_6(int targetcore,
 423                                    unsigned long n0,
 424                                                                                          unsigned long n1,
 425                                                                                          unsigned long n2,
 426                                                                                          unsigned long n3,
 427                                                                                          unsigned long n4,
 428                                                                                          unsigned long n5,
 429                                                                                          bool isinterrupton);
 430 INLINE void cache_msg_1(int targetcore,
 431                                                                                                 unsigned long n0);
 432 INLINE void cache_msg_2(int targetcore,
 433                                     unsigned long n0,
 434                                                                                                 unsigned long n1);
 435 INLINE void cache_msg_3(int targetcore,
 436                                     unsigned long n0,
 437                                                                                                 unsigned long n1,
 438                                                                                                 unsigned long n2);
 439 INLINE void cache_msg_4(int targetcore,
 440                                     unsigned long n0,
 441                                                                                                 unsigned long n1,
 442                                                                                                 unsigned long n2,
 443                                                                                                 unsigned long n3);
 444 INLINE void cache_msg_5(int targetcore,
 445                                     unsigned long n0,
 446                                                                                                 unsigned long n1,
 447                                                                                                 unsigned long n2,
 448                                                                                                 unsigned long n3,
 449                                                                                                 unsigned long n4);
 450 INLINE void cache_msg_6(int targetcore,
 451                                     unsigned long n0,
 452                                                                                                 unsigned long n1,
 453                                                                                                 unsigned long n2,
 454                                                                                                 unsigned long n3,
 455                                                                                                 unsigned long n4,
 456                                                                                                 unsigned long n5);
 457 INLINE void transferObject(struct transObjInfo * transObj);
 458 INLINE int receiveMsg(void);
 459
 460 #ifdef MULTICORE_GC
 461 INLINE void transferMarkResults();
 462 #endif
 463
 464 #ifdef PROFILE
 465 INLINE void profileTaskStart(char * taskname);
 466 INLINE void profileTaskEnd(void);
 467 void outputProfileData();
 468 #endif  // #ifdef PROFILE
 469 ///////////////////////////////////////////////////////////
 470
 471 /////////////////////////////////////////////////////////////////////////////
 472 // For each version of BAMBOO runtime, there should be a header file named //
 473 // runtim_arch.h defining following MARCOS:                                //
 474 // BAMBOO_TOTALCORE: the total # of cores in the processor                 //
 475 // BAMBOO_NUM_OF_CORE: the # of current residing core                      //
 476 // BAMBOO_GET_NUM_OF_CORE(): compute the # of current residing core        //
 477 // BAMBOO_DEBUGPRINT(x): print out integer x                               //
 478 // BAMBOO_DEBUGPRINT_REG(x): print out value of variable x                 //
 479 // BAMBOO_LOCAL_MEM_CALLOC(x, y): allocate an array of x elements each of  //
 480 //                                whose size in bytes is y on local memory //
 481 // BAMBOO_LOCAL_MEM_FREE(x): free space with ptr x on local memory         //
 482 // BAMBOO_SHARE_MEM_CALLOC(x, y): allocate an array of x elements each of  //
 483 //                                whose size in bytes is y on shared memory//
 484 // BAMBOO_START_CRITICAL_SECTION_OBJ_QUEUE()                               //
 485 // BAMBOO_CLOSE_CRITICAL_SECTION_OBJ_QUEUE(): locks for global data        //
 486 //                                            structures related to obj    //
 487 //                                            queue                        //
 488 // BAMBOO_START_CRITICAL_SECTION_STATUS()                                  //
 489 // BAMBOO_CLOSE_CRITICAL_SECTION_STATUS(): locks for global data structures//
 490 //                                         related to status data          //
 491 // BAMBOO_START_CRITICAL_SECTION_MSG()                                     //
 492 // BAMBOO_CLOSE_CRITICAL_SECTION_MSG(): locks for global data structures   //
 493 //                                      related to msg data                //
 494 // BAMBOO_START_CRITICAL_SECTION_LOCK()                                    //
 495 // BAMBOO_CLOSE_CRITICAL_SECTION_LOCK(): locks for global data structures  //
 496 //                                       related to lock table             //
 497 // BAMBOO_START_CRITICAL_SECTION_MEM()                                     //
 498 // BAMBOO_CLOSE_CRITICAL_SECTION_MEM(): locks for allocating memory        //
 499 // BAMBOO_START_CRITICAL_SECTION()                                         //
 500 // BAMBOO_CLOSE_CRITICAL_SECTION(): locks for all global data structures   //
 501 // BAMBOO_WAITING_FOR_LOCK(): routine executed while waiting for lock      //
 502 //                            request response                             //
 503 // BAMBOO_CACHE_LINE_SIZE: the cache line size                             //
 504 // BAMBOO_CACHE_LINE_MASK: mask for a cache line                           //
 505 // BAMBOO_CACHE_FLUSH_RANGE(x, y): flush cache lines started at x with     //
 506 //                                 length y                                //
 507 // BAMBOO_CACHE_FLUSH_ALL(): flush the whole cache of a core if necessary  //
 508 // BAMBOO_EXIT(x): exit routine                                            //
 509 // BAMBOO_MSG_AVAIL(): checking if there are msgs coming in                //
 510 // BAMBOO_GCMSG_AVAIL(): checking if there are gcmsgs coming in            //
 511 // BAMBOO_GET_EXE_TIME(): rountine to get current clock cycle number       //
 512 // BAMBOO_MEMSET_WH(x, y, z): memset the specified region of memory (start //
 513 //                            address x, size z) to value y with write     //
 514 //                            hint, the processor will not fetch the       //
 515 //                            current content of the memory and directly   //
 516 //                            write                                        //
 517 //                                                                         //
 518 // runtime_arch.h should also define following global parameters:          //
 519 // bamboo_cpu2coords: map the cpu # to (x,y) coordinates                   //
 520 // bamboo_coords2cpu: map the (x,y) coordinates to cpu #                   //
 521 /////////////////////////////////////////////////////////////////////////////
 522
 523 #endif  // #ifdef MULTICORE
 524 #endif  // #ifdef TASK
 525 #endif  // #ifndef MULTICORE_RUNTIME