1 #ifndef MULTICORE_RUNTIME
2 #define MULTICORE_RUNTIME
3 #include "structdefs.h"
6 #define INLINE inline __attribute__((always_inline))
15 ////////////////////////////////////////////////////////////////
16 // global variables //
17 ///////////////////////////////////////////////////////////////
19 // record the starting time
20 unsigned long long bamboo_start_time;
28 // shared memory pointer for global thread queue
29 // In MGC version, this block of memory is located at the very bottom of the
30 // shared memory with the base address as BAMBOO_BASE_VA.
31 // The bottom of the shared memory = global thread queue + sbstart tbl
32 // + smemtbl + NUMCORES4GC bamboo_rmsp
33 // This queue is always reside at the bottom of the shared memory. It is
34 // considered as runtime structure, during gc, it is scanned for mark and flush
35 // phase but never been compacted.
37 // This is a loop array and the first 4 int fields of the queue are:
38 // mutex + thread counter + start pointer + end pointer
39 #ifdef GC_SMALLPAGESIZE
40 #define BAMBOO_THREAD_QUEUE_SIZE (1024 * 1024)
41 #define BAMBOO_GLOBAL_DEFS_SIZE (1024 * 1024)
43 #define BAMBOO_THREAD_QUEUE_SIZE (BAMBOO_SMEM_SIZE) // (45 * 16 * 1024)
44 #define BAMBOO_GLOBAL_DEFS_SIZE (BAMBOO_SMEM_SIZE)
46 // data structures for threads
47 INTPTR * bamboo_thread_queue;
48 unsigned int bamboo_max_thread_num_mask;
49 INTPTR bamboo_current_thread;
54 // data structures for msgs
55 #define BAMBOO_OUT_BUF_LENGTH 2048
56 #define BAMBOO_OUT_BUF_MASK (0x7FF)
57 #define BAMBOO_MSG_BUF_LENGTH 2048
58 #define BAMBOO_MSG_BUF_MASK (0x7FF)
59 int msgdata[BAMBOO_MSG_BUF_LENGTH];
60 volatile int msgdataindex;
61 volatile int msgdatalast;
63 volatile bool msgdatafull;
64 int outmsgdata[BAMBOO_OUT_BUF_LENGTH];
68 volatile bool isMsgHanging;
70 #define MSG_INDEXINC_I() \
71 msgdataindex = (msgdataindex + 1) & (BAMBOO_MSG_BUF_MASK)
73 #define MSG_LASTINDEXINC_I() \
74 msgdatalast = (msgdatalast + 1) & (BAMBOO_MSG_BUF_MASK)
76 #define MSG_CACHE_I(n) \
77 msgdata[msgdatalast] = (n); \
80 // NOTE: if msgdataindex == msgdatalast, it always means that the buffer if
81 // full. In the case that the buffer is empty, should never call this
83 #define MSG_REMAINSIZE_I(s) \
84 if(msgdataindex < msgdatalast) { \
85 (*(int*)s) = msgdatalast - msgdataindex; \
86 } else if((msgdataindex == msgdatalast) && (!msgdatafull)) { \
89 (*(int*)s) = (BAMBOO_MSG_BUF_LENGTH) - msgdataindex + msgdatalast; \
92 #define OUTMSG_INDEXINC() \
93 outmsgindex = (outmsgindex + 1) & (BAMBOO_OUT_BUF_MASK)
95 #define OUTMSG_LASTINDEXINC() \
96 outmsglast = (outmsglast + 1) & (BAMBOO_OUT_BUF_MASK); \
97 if(outmsglast == outmsgindex) { \
98 BAMBOO_EXIT(0xd101); \
101 #define OUTMSG_CACHE(n) \
102 outmsgdata[outmsglast] = (n); \
103 OUTMSG_LASTINDEXINC();
105 #define MAX_PACKET_WORDS 5
109 * type: 1 -- transfer object
110 * 2 -- transfer stall msg
115 * // add for profile info
116 * 7 -- transfer profile output msg
117 * 8 -- transfer profile output finish msg
118 * // add for alias lock strategy
119 * 9 -- redirect lock request
120 * a -- lock grant with redirect info
121 * b -- lock deny with redirect info
122 * c -- lock release with redirect info
123 * d -- status confirm request
124 * e -- status report msg
126 * 10 -- requiring for new memory
127 * 11 -- response for new memory request
128 * 12 -- GC init phase start
130 * 14 -- compact phase start
131 * 15 -- flush phase start
132 * 16 -- init phase finish
133 * 17 -- mark phase finish
134 * 18 -- compact phase finish
135 * 19 -- flush phase finish
137 * 1b -- marked phase finish confirm request
138 * 1c -- marked phase finish confirm response
139 * 1d -- markedObj msg
140 * 1e -- start moving objs msg
141 * 1f -- ask for mapping info of a markedObj
142 * 20 -- mapping info of a markedObj
143 * 21 -- large objs info request
144 * 22 -- large objs info response
145 * 23 -- large objs mapping info
147 * ObjMsg: 1 + size of msg + obj's address + (task index + param index)+
148 * StallMsg: 2 + corenum + sendobjs + receiveobjs
149 * (size is always 4 * sizeof(int))
150 * LockMsg: 3 + lock type + obj pointer + lock + request core
151 * (size is always 5 * sizeof(int))
152 * 4/5/6 + lock type + obj pointer + lock
153 * (size is always 4 * sizeof(int))
154 * 9 + lock type + obj pointer + redirect lock + root request core
156 * (size is always 6 * sizeof(int))
157 * a/b + lock type + obj pointer + redirect lock
158 * (size is always 4 * sizeof(int))
159 * c + lock type + lock + redirect lock
160 * (size is always 4 * sizeof(int))
161 * lock type: 0 -- read; 1 -- write
162 * ProfileMsg: 7 + totalexetime
163 * (size is always 2 * sizeof(int))
165 * (size is always 2 * sizeof(int))
166 * StatusMsg: d (size is always 1 * sizeof(int))
167 * e + status + corenum + sendobjs + receiveobjs
168 * (size is always 5 * sizeof(int))
169 * status: 0 -- stall; 1 -- busy
170 * TerminateMsg: f (size is always 1 * sizeof(int)
171 * MemoryMsg: 10 + size + corenum
172 * (size is always 3 * sizeof(int))
173 * 11 + base_va + size
174 * (size is always 3 * sizeof(int))
175 * GCMsg: 12/13 (size is always 1 * sizeof(int))
176 * 14 + size of msg + (num of objs to move + (start address
177 * + end address + dst core + start dst)+)?
178 * + (num of incoming objs + (start dst + orig core)+)?
179 * + (num of large obj lists + (start address + lenght
181 * 15 (size is always 1 * sizeof(int))
183 * (size is always 2 * sizeof(int))
184 * 17 + corenum + gcsendobjs + gcreceiveobjs
185 * (size if always 4 * sizeof(int))
186 * 18 + corenum + fulfilled blocks num + (finish compact(1) + current
187 * heap top)/(need mem(0) + mem need)
188 * size is always 5 * sizeof(int))
190 * (size is always 2 * sizeof(int))
191 * 1a (size is always 1 * sizeof(int))
192 * 1b (size if always 1 * sizeof(int))
193 * 1c + size of msg + corenum + gcsendobjs + gcreceiveobjs
194 * (size is always 5 * sizeof(int))
195 * 1d + obj's address + request core
196 * (size is always 3 * sizeof(int))
197 * 1e + corenum + start addr + end addr
198 * (size if always 4 * sizeof(int))
199 * 1f + obj's address + corenum
200 * (size is always 3 * sizeof(int))
201 * 20 + obj's address + dst address
202 * (size if always 3 * sizeof(int))
203 * 21 (size is always 1 * sizeof(int))
204 * 22 + size of msg + corenum + current heap size
205 * + (num of large obj lists + (start address + length)+)?
206 * 23 + orig large obj ptr + new large obj ptr
207 * (size is always 3 * sizeof(int))
210 MSGSTART = 0xD0, // 0xD0
217 PROFILEOUTPUT, // 0xD7
218 PROFILEFINISH, // 0xD8
219 REDIRECTLOCK, // 0xD9
220 REDIRECTGROUNT, // 0xDa
221 REDIRECTDENY, // 0xDb
222 REDIRECTRELEASE, // 0xDc
223 STATUSCONFIRM, // 0xDd
224 STATUSREPORT, // 0xDe
232 GCSTARTCOMPACT, // 0xE5
233 GCSTARTMAPINFO, // 0xE6
234 GCSTARTFLUSH, // 0xE7
236 GCFINISHINIT, // 0xE9
237 GCFINISHMARK, // 0xEa
238 GCFINISHCOMPACT, // 0xEb
239 GCFINISHMAPINFO, // 0xEc
240 GCFINISHFLUSH, // 0xEd
242 GCMARKCONFIRM, // 0xEf
243 GCMARKREPORT, // 0xF0
246 GCMAPREQUEST, // 0xF3
249 GCLOBJREQUEST, // 0xF6
251 GCLOBJMAPPING, // 0xF8
255 #ifdef GC_CACHE_ADAPT
256 GCSTARTPOSTINIT, // 0xFa
258 GCFINISHPOSTINIT, // 0xFc
259 GCFINISHPREF, // 0xFd
260 #endif // GC_CACHE_ADAPT
261 #endif // MULTICORE_GC
265 /////////////////////////////////////////////////////////////////////////////////
266 // NOTE: BAMBOO_TOTALCORE -- number of the available cores in the processor.
267 // No greater than the number of all the cores in
269 // NUMCORES -- number of cores chosen to deploy the application. It can
270 // be greater than that required to fully parallelize the
271 // application. The same as NUMCORES.
272 // NUMCORESACTIVE -- number of cores that really execute the
273 // application. No greater than NUMCORES
274 // NUMCORES4GC -- number of cores for gc. No greater than NUMCORES.
275 // NOTE: currently only support ontinuous cores as gc
276 // cores, i.e. 0~NUMCORES4GC-1
277 ////////////////////////////////////////////////////////////////////////////////
278 // data structures of status for termination
279 // only check working cores
280 volatile int corestatus[NUMCORESACTIVE]; // records status of each core
283 volatile int numsendobjs[NUMCORESACTIVE]; // records how many objects a core
285 volatile int numreceiveobjs[NUMCORESACTIVE]; // records how many objects a
287 volatile int numconfirm;
288 volatile bool waitconfirm;
290 int self_numsendobjs;
291 int self_numreceiveobjs;
293 // TASK specific data structures
295 // get rid of lock msgs for GC version
297 // data structures for locking
298 struct RuntimeHash locktable;
299 static struct RuntimeHash* locktbl = &locktable;
300 struct RuntimeHash * lockRedirectTbl;
301 struct RuntimeHash * objRedirectLockTbl;
302 #endif // ifndef MULTICORE_GC
312 // data structures for waiting objs
313 struct Queue objqueue;
314 struct Queue * totransobjqueue; // queue to hold objs to be transferred
315 // should be cleared whenever enter a task
320 // data structures for profile mode
322 #define TASKINFOLENGTH 3000 // 0
323 #ifdef PROFILE_INTERRUPT
324 #define INTERRUPTINFOLENGTH 50 //0
325 #endif // PROFILE_INTERRUPT
327 typedef struct task_info {
329 unsigned long long startTime;
330 unsigned long long endTime;
331 unsigned long long exitIndex;
332 struct Queue * newObjs;
335 TaskInfo * taskInfoArray[TASKINFOLENGTH];
337 bool taskInfoOverflow;
338 #ifdef PROFILE_INTERRUPT
339 typedef struct interrupt_info {
340 unsigned long long startTime;
341 unsigned long long endTime;
344 InterruptInfo * interruptInfoArray[INTERRUPTINFOLENGTH];
345 int interruptInfoIndex;
346 bool interruptInfoOverflow;
347 #endif // PROFILE_INTERUPT
348 volatile int profilestatus[NUMCORESACTIVE]; // records status of each core
351 #endif // #ifdef PROFILE
354 #include "multicoremem.h"
356 /////////////////////////////////////////////////////////////
358 ////////////////////////////////////////////////////////////
359 // these are functions should be implemented in //
360 // multicore runtime for any multicore processors //
361 ////////////////////////////////////////////////////////////
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 #ifdef BAMBOO_MEMPROF
369 INLINE void terminatememprof(void);
370 #endif // BAMBOO_MEMPROF
372 // msg related functions
373 INLINE void send_hanging_msg(bool isInterrupt);
374 INLINE void send_msg_1(int targetcore,
377 INLINE void send_msg_2(int targetcore,
381 INLINE void send_msg_3(int targetcore,
386 INLINE void send_msg_4(int targetcore,
392 INLINE void send_msg_5(int targetcore,
399 INLINE void send_msg_6(int targetcore,
407 INLINE void cache_msg_1(int targetcore,
409 INLINE void cache_msg_2(int targetcore,
412 INLINE void cache_msg_3(int targetcore,
416 INLINE void cache_msg_4(int targetcore,
421 INLINE void cache_msg_5(int targetcore,
427 INLINE void cache_msg_6(int targetcore,
434 INLINE int receiveMsg(unsigned int send_port_pending);
437 INLINE void transferMarkResults();
438 #endif // MULTICORE_GC
441 // lock related functions
442 bool getreadlock(void* ptr);
443 void releasereadlock(void* ptr);
444 bool getwritelock(void* ptr);
445 void releasewritelock(void* ptr);
446 bool getwritelock_I(void* ptr);
447 void releasewritelock_I(void * ptr);
449 void releasewritelock_r(void * lock, void * redirectlock);
450 #endif // ifndef MULTICORE_GC
451 /* this function is to process lock requests.
452 * can only be invoked in receiveObject() */
453 // if return -1: the lock request is redirected
454 // 0: the lock request is approved
455 // 1: the lock request is denied
456 INLINE int processlockrequest(int locktype,
462 INLINE void processlockrelease(int locktype,
467 // msg related functions
468 INLINE void transferObject(struct transObjInfo * transObj);
471 INLINE void profileTaskStart(char * taskname);
472 INLINE void profileTaskEnd(void);
473 void outputProfileData();
474 #endif // #ifdef PROFILE
475 ///////////////////////////////////////////////////////////
477 /////////////////////////////////////////////////////////////////////////////
478 // For each version of BAMBOO runtime, there should be a header file named //
479 // runtim_arch.h defining following MARCOS: //
480 // BAMBOO_NUM_OF_CORE: the # of current residing core //
481 // BAMBOO_GET_NUM_OF_CORE(): compute the # of current residing core //
482 // BAMBOO_COORDS(c, x, y): convert the cpu # to coords (*x, *y) //
483 // BAMBOO_DEBUGPRINT(x): print out integer x //
484 // BAMBOO_DEBUGPRINT_REG(x): print out value of variable x //
485 // BAMBOO_EXIT_APP(x): exit the whole application //
486 // BAMBOO_EXIT(x): error exit routine with error # //
487 // BAMBOO_DIE(x): error exit routine with error msg //
488 // BAMBOO_GET_EXE_TIME(): rountine to get current clock cycle number //
489 // BAMBOO_MSG_AVAIL(): checking if there are msgs coming in //
490 // BAMBOO_ENTER_RUNTIME_MODE_FROM_CLIENT(): change to runtime mode from //
492 // BAMBOO_ENTER_CLIENT_MODE_FROM_RUNTIME(): change to client mode from //
494 // BAMBOO_ENTER_SEND_MODE_FROM_CLIENT(): change to send mode from //
496 // BAMBOO_ENTER_CLIENT_MODE_FROM_SEND(): change to client mode from //
498 // BAMBOO_ENTER_RUNTIME_MODE_FROM_SEND(): change to runtime mode from //
500 // BAMBOO_ENTER_SEND_MODE_FROM_RUNTIME(): change to send mode from //
502 // BAMBOO_WAITING_FOR_LOCK(): routine executed while waiting for lock //
503 // request response //
504 // BAMBOO_LOCAL_MEM_CALLOC(x, y): allocate an array of x elements each of //
505 // whose size in bytes is y on local memory //
506 // which is given by the hypervisor //
507 // BAMBOO_LOCAL_MEM_FREE(x): free space with ptr x on local memory //
508 // BAMBOO_LOCAL_MEM_CLOSE(): close the local heap //
509 // BAMBOO_LOCAL_MEM_CALLOC_S(x, y): allocate an array of x elements each of//
510 // whose size in bytes is y on local //
511 // memory which is not from the hypervisor//
512 // but is allocated from the free memory //
513 // BAMBOO_LOCAL_MEM_FREE_S(x): free space with ptr x on self-allocated //
515 // BAMBOO_LOCAL_MEM_CLOSE_S(): close the self-allocated local heap //
516 // BAMBOO_SHARE_MEM_CALLOC_I(x, y): allocate an array of x elements each of//
517 // whose size in bytes is y on shared memory//
518 // BAMBOO_SHARE_MEM_CLOSE(): close the shared heap //
519 // BAMBOO_CACHE_LINE_SIZE: the cache line size //
520 // BAMBOO_CACHE_LINE_MASK: mask for a cache line //
521 // BAMBOO_CACHE_FLUSH_RANGE(x, y): flush cache lines started at x with //
523 // BAMBOO_CACHE_FLUSH_ALL(): flush the whole cache of a core if necessary //
524 // BAMBOO_MEMSET_WH(x, y, z): memset the specified region of memory (start //
525 // address x, size z) to value y with write //
526 // hint, the processor will not fetch the //
527 // current content of the memory and directly //
529 // BAMBOO_CLEAN_DTLB(): zero-out all the dtlb entries //
530 // BAMBOO_CACHE_FLUSH_L2(): Flush the contents of this tile's L2 back to //
532 // BAMBOO_CACHE_FLUSH_RANGE_NO_FENCE(x, y): flush a range of mem without //
534 // BAMBOO_CACHE_MEM_FENCE_INCOHERENT(): fence to guarantee visibility of //
535 // stores to incoherent memory //
536 /////////////////////////////////////////////////////////////////////////////
538 #endif // #ifdef TASK
539 #endif // #ifdef MULTICORE
540 #endif // #ifndef MULTICORE_RUNTIME