/* * Copyright 2010 OpenSourceCodeStewardshipFoundation * * Licensed under BSD */ #include #include #include #include "Queue_impl/PrivateQueue.h" #include "Hash_impl/PrivateHash.h" #include "VOMP.h" #include "VOMP_Counter_Recording.h" //========================================================================== void VOMP__init(); void VOMP__init_Helper(); //========================================================================== //=========================================================================== /*These are the library functions *called in the application* * *There's a pattern for the outside sequential code to interact with the * VMS_HW code. *The VMS_HW system is inside a boundary.. every VOMP system is in its * own directory that contains the functions for each of the processor types. * One of the processor types is the "seed" processor that starts the * cascade of creating all the processors that do the work. *So, in the directory is a file called "EntryPoint.c" that contains the * function, named appropriately to the work performed, that the outside * sequential code calls. This function follows a pattern: *1) it calls VOMP__init() *2) it creates the initial data for the seed processor, which is passed * in to the function *3) it creates the seed VOMP processor, with the data to start it with. *4) it calls startVOMPThenWaitUntilWorkDone *5) it gets the returnValue from the transfer struc and returns that * from the function * *For now, a new VOMP system has to be created via VOMP__init every * time an entry point function is called -- later, might add letting the * VOMP system be created once, and let all the entry points just reuse * it -- want to be as simple as possible now, and see by using what makes * sense for later.. */ //=========================================================================== /*This is the "border crossing" function -- the thing that crosses from the * outside world, into the VMS_HW world. It initializes and starts up the * VMS system, then creates one processor from the specified function and * puts it into the readyQ. From that point, that one function is resp. * for creating all the other processors, that then create others, and so * forth. *When all the processors, including the seed, have dissipated, then this * function returns. The results will have been written by side-effect via * pointers read from, or written into initData. * *NOTE: no Threads should exist in the outside program that might touch * any of the data reachable from initData passed in to here */ void VOMP__create_seed_procr_and_do_work( TopLevelFnPtr fnPtr, void *initData ) { VOMPSemEnv *semEnv; SlaveVP *seedPr; VOMP__init(); //normal multi-thd semEnv = _VMSMasterEnv->semanticEnv; //VOMP starts with one processor, which is put into initial environ, // and which then calls create() to create more, thereby expanding work seedPr = VOMP__create_procr_helper( fnPtr, initData, semEnv, semEnv->nextCoreToGetNewPr++ ); resume_slaveVP( seedPr, semEnv ); VMS_SS__start_the_work_then_wait_until_done(); //normal multi-thd VOMP__cleanup_after_shutdown(); } int32 VOMP__giveMinWorkUnitCycles( float32 percentOverhead ) { return MIN_WORK_UNIT_CYCLES; } int32 VOMP__giveIdealNumWorkUnits() { return NUM_ANIM_SLOTS * NUM_CORES; } int32 VOMP__give_number_of_cores_to_schedule_onto() { return NUM_CORES; } /*For now, use TSC -- later, make these two macros with assembly that first * saves jump point, and second jumps back several times to get reliable time */ void VOMP__start_primitive() { saveLowTimeStampCountInto( ((VOMPSemEnv *)(_VMSMasterEnv->semanticEnv))-> primitiveStartTime ); } /*Just quick and dirty for now -- make reliable later * will want this to jump back several times -- to be sure cache is warm * because don't want comm time included in calc-time measurement -- and * also to throw out any "weird" values due to OS interrupt or TSC rollover */ int32 VOMP__end_primitive_and_give_cycles() { int32 endTime, startTime; //TODO: fix by repeating time-measurement saveLowTimeStampCountInto( endTime ); startTime =((VOMPSemEnv*)(_VMSMasterEnv->semanticEnv))->primitiveStartTime; return (endTime - startTime); } //=========================================================================== /*Initializes all the data-structures for a VOMP system -- but doesn't * start it running yet! * *This runs in the main thread -- before VMS starts up * *This sets up the semantic layer over the VMS system * *First, calls VMS_Setup, then creates own environment, making it ready * for creating the seed processor and then starting the work. */ void VOMP__init() { VMS_SS__init(); //masterEnv, a global var, now is partially set up by init_VMS // after this, have VMS_int__malloc and VMS_int__free available VOMP__init_Helper(); } void idle_fn(void* data, SlaveVP *animatingSlv){ while(1){ VMS_int__suspend_slaveVP_and_send_req(animatingSlv); } } void VOMP__init_Helper() { VOMPSemEnv *semanticEnv; PrivQueueStruc **readyVPQs; int coreIdx, i, j; //Hook up the semantic layer's plug-ins to the Master virt procr _VMSMasterEnv->requestHandler = &VOMP__Request_Handler; _VMSMasterEnv->slaveAssigner = &VOMP__assign_slaveVP_to_slot; #ifdef HOLISTIC__TURN_ON_PERF_COUNTERS _VMSMasterEnv->counterHandler = &VOMP__counter_handler; #endif //create the semantic layer's environment (all its data) and add to // the master environment semanticEnv = VMS_int__malloc( sizeof( VOMPSemEnv ) ); _VMSMasterEnv->semanticEnv = semanticEnv; #ifdef HOLISTIC__TURN_ON_PERF_COUNTERS VOMP__init_counter_data_structs(); #endif semanticEnv->shutdownInitiated = FALSE; for(i=0;iidlePr[i][j] = VMS_int__create_slaveVP(&idle_fn,NULL); semanticEnv->idlePr[i][j]->coreAnimatedBy = i; } } #ifdef HOLISTIC__TURN_ON_OBSERVE_UCC semanticEnv->unitList = makeListOfArrays(sizeof(Unit),128); semanticEnv->ctlDependenciesList = makeListOfArrays(sizeof(Dependency),128); semanticEnv->commDependenciesList = makeListOfArrays(sizeof(Dependency),128); semanticEnv->dynDependenciesList = makeListOfArrays(sizeof(Dependency),128); semanticEnv->ntonGroupsInfo = makePrivDynArrayOfSize((void***)&(semanticEnv->ntonGroups),8); semanticEnv->hwArcs = makeListOfArrays(sizeof(Dependency),128); memset(semanticEnv->last_in_slot,0,sizeof(NUM_CORES * NUM_ANIM_SLOTS * sizeof(Unit))); #endif //create the ready queue, hash tables used for pairing send to receive // and so forth //TODO: add hash tables for pairing sends with receives, and // initialize the data ownership system readyVPQs = VMS_int__malloc( NUM_CORES * sizeof(PrivQueueStruc *) ); for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ ) { readyVPQs[ coreIdx ] = makeVMSQ(); } semanticEnv->readyVPQs = readyVPQs; semanticEnv->nextCoreToGetNewPr = 0; semanticEnv->numSlaveVP = 0; semanticEnv->commHashTbl = makeHashTable( 1<<16, &VMS_int__free );//start big //TODO: bug -- turn these arrays into dyn arrays to eliminate limit //semanticEnv->singletonHasBeenExecutedFlags = makeDynArrayInfo( ); //semanticEnv->transactionStrucs = makeDynArrayInfo( ); for( i = 0; i < NUM_STRUCS_IN_SEM_ENV; i++ ) { semanticEnv->fnSingletons[i].endInstrAddr = NULL; semanticEnv->fnSingletons[i].hasBeenStarted = FALSE; semanticEnv->fnSingletons[i].hasFinished = FALSE; semanticEnv->fnSingletons[i].waitQ = makeVMSQ(); semanticEnv->transactionStrucs[i].waitingVPQ = makeVMSQ(); } } /*Frees any memory allocated by VOMP__init() then calls VMS_int__shutdown */ void VOMP__cleanup_after_shutdown() { VOMPSemEnv *semanticEnv; semanticEnv = _VMSMasterEnv->semanticEnv; #ifdef HOLISTIC__TURN_ON_OBSERVE_UCC //UCC FILE* output; int n; char filename[255]; for(n=0;n<255;n++) { sprintf(filename, "./counters/UCC.%d",n); output = fopen(filename,"r"); if(output) { fclose(output); }else{ break; } } if(n<255){ printf("Saving UCC to File: %s ...\n", filename); output = fopen(filename,"w+"); if(output!=NULL){ set_dependency_file(output); //fprintf(output,"digraph Dependencies {\n"); //set_dot_file(output); //FIXME: first line still depends on counters being enabled, replace w/ unit struct! //forAllInDynArrayDo(_VMSMasterEnv->counter_history_array_info, &print_dot_node_info ); forAllInListOfArraysDo(semanticEnv->unitList, &print_unit_to_file); forAllInListOfArraysDo( semanticEnv->commDependenciesList, &print_comm_dependency_to_file ); forAllInListOfArraysDo( semanticEnv->ctlDependenciesList, &print_ctl_dependency_to_file ); forAllInDynArrayDo(semanticEnv->ntonGroupsInfo,&print_nton_to_file); //fprintf(output,"}\n"); fflush(output); } else printf("Opening UCC file failed. Please check that folder \"counters\" exists in run directory and has write permission.\n"); } else { printf("Could not open UCC file, please clean \"counters\" folder. (Must contain less than 255 files.)\n"); } //Loop Graph for(n=0;n<255;n++) { sprintf(filename, "./counters/LoopGraph.%d",n); output = fopen(filename,"r"); if(output) { fclose(output); }else{ break; } } if(n<255){ printf("Saving LoopGraph to File: %s ...\n", filename); output = fopen(filename,"w+"); if(output!=NULL){ set_dependency_file(output); //fprintf(output,"digraph Dependencies {\n"); //set_dot_file(output); //FIXME: first line still depends on counters being enabled, replace w/ unit struct! //forAllInDynArrayDo(_VMSMasterEnv->counter_history_array_info, &print_dot_node_info ); forAllInListOfArraysDo( semanticEnv->unitList, &print_unit_to_file ); forAllInListOfArraysDo( semanticEnv->commDependenciesList, &print_comm_dependency_to_file ); forAllInListOfArraysDo( semanticEnv->ctlDependenciesList, &print_ctl_dependency_to_file ); forAllInListOfArraysDo( semanticEnv->dynDependenciesList, &print_dyn_dependency_to_file ); forAllInListOfArraysDo( semanticEnv->hwArcs, &print_hw_dependency_to_file ); //fprintf(output,"}\n"); fflush(output); } else printf("Opening LoopGraph file failed. Please check that folder \"counters\" exists in run directory and has write permission.\n"); } else { printf("Could not open LoopGraph file, please clean \"counters\" folder. (Must contain less than 255 files.)\n"); } freeListOfArrays(semanticEnv->unitList); freeListOfArrays(semanticEnv->commDependenciesList); freeListOfArrays(semanticEnv->ctlDependenciesList); freeListOfArrays(semanticEnv->dynDependenciesList); #endif #ifdef HOLISTIC__TURN_ON_PERF_COUNTERS for(n=0;n<255;n++) { sprintf(filename, "./counters/Counters.%d.csv",n); output = fopen(filename,"r"); if(output) { fclose(output); }else{ break; } } if(n<255){ printf("Saving Counter measurements to File: %s ...\n", filename); output = fopen(filename,"w+"); if(output!=NULL){ set_counter_file(output); int i; for(i=0;icounterList[i], &print_counter_events_to_file ); fflush(output); } } else printf("Opening UCC file failed. Please check that folder \"counters\" exists in run directory and has write permission.\n"); } else { printf("Could not open UCC file, please clean \"counters\" folder. (Must contain less than 255 files.)\n"); } #endif /* It's all allocated inside VMS's big chunk -- that's about to be freed, so * nothing to do here for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ ) { VMS_int__free( semanticEnv->readyVPQs[coreIdx]->startOfData ); VMS_int__free( semanticEnv->readyVPQs[coreIdx] ); } VMS_int__free( semanticEnv->readyVPQs ); freeHashTable( semanticEnv->commHashTbl ); VMS_int__free( _VMSMasterEnv->semanticEnv ); */ VMS_SS__cleanup_at_end_of_shutdown(); } //=========================================================================== /* */ SlaveVP * VOMP__create_procr_with( TopLevelFnPtr fnPtr, void *initData, SlaveVP *creatingPr ) { VOMPSemReq reqData; //the semantic request data is on the stack and disappears when this // call returns -- it's guaranteed to remain in the VP's stack for as // long as the VP is suspended. reqData.reqType = 0; //know type because in a VMS create req reqData.coreToAssignOnto = -1; //means round-robin assign reqData.fnPtr = fnPtr; reqData.initData = initData; reqData.sendPr = creatingPr; VMS_WL__send_create_slaveVP_req( &reqData, creatingPr ); return creatingPr->dataRetFromReq; } SlaveVP * VOMP__create_procr_with_affinity( TopLevelFnPtr fnPtr, void *initData, SlaveVP *creatingPr, int32 coreToAssignOnto ) { VOMPSemReq reqData; //the semantic request data is on the stack and disappears when this // call returns -- it's guaranteed to remain in the VP's stack for as // long as the VP is suspended. reqData.reqType = 0; //know type because in a VMS create req reqData.coreToAssignOnto = coreToAssignOnto; reqData.fnPtr = fnPtr; reqData.initData = initData; reqData.sendPr = creatingPr; VMS_WL__send_create_slaveVP_req( &reqData, creatingPr ); return creatingPr->dataRetFromReq; } void VOMP__dissipate_procr( SlaveVP *procrToDissipate ) { VMS_WL__send_dissipate_req( procrToDissipate ); } //=========================================================================== void * VOMP__malloc_to( int32 sizeToMalloc, SlaveVP *owningPr ) { VOMPSemReq reqData; reqData.reqType = malloc_req; reqData.sendPr = owningPr; reqData.sizeToMalloc = sizeToMalloc; VMS_WL__send_sem_request( &reqData, owningPr ); return owningPr->dataRetFromReq; } /*Sends request to Master, which does the work of freeing */ void VOMP__free( void *ptrToFree, SlaveVP *owningPr ) { VOMPSemReq reqData; reqData.reqType = free_req; reqData.sendPr = owningPr; reqData.ptrToFree = ptrToFree; VMS_WL__send_sem_request( &reqData, owningPr ); } void VOMP__transfer_ownership_of_from_to( void *data, SlaveVP *oldOwnerSlv, SlaveVP *newOwnerPr ) { //TODO: put in the ownership system that automatically frees when no // owners of data left -- will need keeper for keeping data around when // future created processors might need it but don't exist yet } void VOMP__add_ownership_by_to( SlaveVP *newOwnerSlv, void *data ) { } void VOMP__remove_ownership_by_from( SlaveVP *loserSlv, void *dataLosing ) { } /*Causes the VOMP system to remove internal ownership, so data won't be * freed when VOMP shuts down, and will persist in the external program. * *Must be called from the processor that currently owns the data. * *IMPL: Transferring ownership touches two different virtual processor's * state -- which means it has to be done carefully -- the VMS rules for * semantic layers say that a work-unit is only allowed to touch the * virtual processor it is part of, and that only a single work-unit per * virtual processor be assigned to a slave at a time. So, this has to * modify the virtual processor that owns the work-unit that called this * function, then create a request to have the other processor modified. *However, in this case, the TO processor is the outside, and transfers * are only allowed to be called by the giver-upper, so can mark caller of * this function as no longer owner, and return -- done. */ void VOMP__transfer_ownership_to_outside( void *data ) { //TODO: removeAllOwnersFrom( data ); } //=========================================================================== void VOMP__send_of_type_to( SlaveVP *sendPr, void *msg, const int type, SlaveVP *receivePr) { VOMPSemReq reqData; reqData.receivePr = receivePr; reqData.sendPr = sendPr; reqData.reqType = send_type; reqData.msgType = type; reqData.msg = msg; reqData.nextReqInHashEntry = NULL; //On ownership -- remove inside the send and let ownership sit in limbo // as a potential in an entry in the hash table, when this receive msg // gets paired to a send, the ownership gets added to the receivePr -- // the next work-unit in the receivePr's trace will have ownership. VMS_WL__send_sem_request( &reqData, sendPr ); //When come back from suspend, no longer own data reachable from msg //TODO: release ownership here } void VOMP__send_from_to( void *msg, SlaveVP *sendPr, SlaveVP *receivePr ) { VOMPSemReq reqData; //hash on the receiver, 'cause always know it, but sometimes want to // receive from anonymous sender reqData.receivePr = receivePr; reqData.sendPr = sendPr; reqData.reqType = send_from_to; reqData.msg = msg; reqData.nextReqInHashEntry = NULL; VMS_WL__send_sem_request( &reqData, sendPr ); } //=========================================================================== void * VOMP__receive_any_to( SlaveVP *receivePr ) { } void * VOMP__receive_type_to( const int type, SlaveVP *receivePr ) { DEBUG__printf1(dbgRqstHdlr,"WL: receive type to: %d", receivePr->slaveID); VOMPSemReq reqData; reqData.receivePr = receivePr; reqData.reqType = receive_type; reqData.msgType = type; reqData.nextReqInHashEntry = NULL; VMS_WL__send_sem_request( &reqData, receivePr ); return receivePr->dataRetFromReq; } /*Call this at point receiving virt pr wants in-coming data. * *The reason receivePr must call this is that it modifies the receivPr * loc structure directly -- and the VMS rules state a virtual processor * loc structure can only be modified by itself. */ void * VOMP__receive_from_to( SlaveVP *sendPr, SlaveVP *receivePr ) { DEBUG__printf2(dbgRqstHdlr,"WL: receive from %d to: %d", sendPr->slaveID, receivePr->slaveID); VOMPSemReq reqData; //hash on the receiver, 'cause always know it, but sometimes want to // receive from anonymous sender reqData.receivePr = receivePr; reqData.sendPr = sendPr; reqData.reqType = receive_from_to; reqData.nextReqInHashEntry = NULL; VMS_WL__send_sem_request( &reqData, receivePr ); return receivePr->dataRetFromReq; } //=========================================================================== // /*A function singleton is a function whose body executes exactly once, on a * single core, no matter how many times the fuction is called and no * matter how many cores or the timing of cores calling it. * *A data singleton is a ticket attached to data. That ticket can be used * to get the data through the function exactly once, no matter how many * times the data is given to the function, and no matter the timing of * trying to get the data through from different cores. */ /*asm function declarations*/ void asm_save_ret_to_singleton(VOMPSingleton *singletonPtrAddr); void asm_write_ret_from_singleton(VOMPSingleton *singletonPtrAddr); /*Fn singleton uses ID as index into array of singleton structs held in the * semantic environment. */ void VOMP__start_fn_singleton( int32 singletonID, SlaveVP *animPr ) { VOMPSemReq reqData; // reqData.reqType = singleton_fn_start; reqData.singletonID = singletonID; VMS_WL__send_sem_request( &reqData, animPr ); if( animPr->dataRetFromReq ) //will be 0 or addr of label in end singleton { VOMPSemEnv *semEnv = VMS_int__give_sem_env_for( animPr ); asm_write_ret_from_singleton(&(semEnv->fnSingletons[ singletonID])); } } /*Data singleton hands addr of loc holding a pointer to a singleton struct. * The start_data_singleton makes the structure and puts its addr into the * location. */ void VOMP__start_data_singleton( VOMPSingleton **singletonAddr, SlaveVP *animPr ) { VOMPSemReq reqData; if( *singletonAddr && (*singletonAddr)->hasFinished ) goto JmpToEndSingleton; reqData.reqType = singleton_data_start; reqData.singletonPtrAddr = singletonAddr; VMS_WL__send_sem_request( &reqData, animPr ); if( animPr->dataRetFromReq ) //either 0 or end singleton's return addr { //Assembly code changes the return addr on the stack to the one // saved into the singleton by the end-singleton-fn //The return addr is at 0x4(%%ebp) JmpToEndSingleton: asm_write_ret_from_singleton(*singletonAddr); } //now, simply return //will exit either from the start singleton call or the end-singleton call } /*Uses ID as index into array of flags. If flag already set, resumes from * end-label. Else, sets flag and resumes normally. * *Note, this call cannot be inlined because the instr addr at the label * inside is shared by all invocations of a given singleton ID. */ void VOMP__end_fn_singleton( int32 singletonID, SlaveVP *animPr ) { VOMPSemReq reqData; //don't need this addr until after at least one singleton has reached // this function VOMPSemEnv *semEnv = VMS_int__give_sem_env_for( animPr ); asm_write_ret_from_singleton(&(semEnv->fnSingletons[ singletonID])); reqData.reqType = singleton_fn_end; reqData.singletonID = singletonID; VMS_WL__send_sem_request( &reqData, animPr ); EndSingletonInstrAddr: return; } void VOMP__end_data_singleton( VOMPSingleton **singletonPtrAddr, SlaveVP *animPr ) { VOMPSemReq reqData; //don't need this addr until after singleton struct has reached // this function for first time //do assembly that saves the return addr of this fn call into the // data singleton -- that data-singleton can only be given to exactly // one instance in the code of this function. However, can use this // function in different places for different data-singletons. // (*(singletonAddr))->endInstrAddr = &&EndDataSingletonInstrAddr; asm_save_ret_to_singleton(*singletonPtrAddr); reqData.reqType = singleton_data_end; reqData.singletonPtrAddr = singletonPtrAddr; VMS_WL__send_sem_request( &reqData, animPr ); } /*This executes the function in the masterVP, so it executes in isolation * from any other copies -- only one copy of the function can ever execute * at a time. * *It suspends to the master, and the request handler takes the function * pointer out of the request and calls it, then resumes the VP. *Only very short functions should be called this way -- for longer-running * isolation, use transaction-start and transaction-end, which run the code * between as work-code. */ void VOMP__animate_short_fn_in_isolation( PtrToAtomicFn ptrToFnToExecInMaster, void *data, SlaveVP *animPr ) { VOMPSemReq reqData; // reqData.reqType = atomic; reqData.fnToExecInMaster = ptrToFnToExecInMaster; reqData.dataForFn = data; VMS_WL__send_sem_request( &reqData, animPr ); } /*This suspends to the master. *First, it looks at the VP's data, to see the highest transactionID that VP * already has entered. If the current ID is not larger, it throws an * exception stating a bug in the code. Otherwise it puts the current ID * there, and adds the ID to a linked list of IDs entered -- the list is * used to check that exits are properly ordered. *Next it is uses transactionID as index into an array of transaction * structures. *If the "VP_currently_executing" field is non-null, then put requesting VP * into queue in the struct. (At some point a holder will request * end-transaction, which will take this VP from the queue and resume it.) *If NULL, then write requesting into the field and resume. */ void VOMP__start_transaction( int32 transactionID, SlaveVP *animPr ) { VOMPSemReq reqData; // reqData.sendPr = animPr; reqData.reqType = trans_start; reqData.transID = transactionID; VMS_WL__send_sem_request( &reqData, animPr ); } /*This suspends to the master, then uses transactionID as index into an * array of transaction structures. *It looks at VP_currently_executing to be sure it's same as requesting VP. * If different, throws an exception, stating there's a bug in the code. *Next it looks at the queue in the structure. *If it's empty, it sets VP_currently_executing field to NULL and resumes. *If something in, gets it, sets VP_currently_executing to that VP, then * resumes both. */ void VOMP__end_transaction( int32 transactionID, SlaveVP *animPr ) { VOMPSemReq reqData; // reqData.sendPr = animPr; reqData.reqType = trans_end; reqData.transID = transactionID; VMS_WL__send_sem_request( &reqData, animPr ); }