/* * Copyright 2010 OpenSourceStewardshipFoundation * * Licensed under BSD */ #include #include #include "PR.h" #include "VSs_impl/VSs.h" inline void replaceWithNewSlotSlv( SlaveVP *requestingSlv, PRProcessEnv *processEnv ); /*The animationMaster embodies most of the animator of the language. The * animator is what emodies the behavior of language constructs. * As such, it is the animationMaster, in combination with the plugin * functions, that make the language constructs do their behavior. * *Within the code, this is the top-level-function of the masterVPs, and * runs when the coreController has no more slave VPs. It's job is to * refill the animation slots with slaves that have work. * *There are multiple versions of the master, each tuned to a specific * combination of modes. This keeps the master simple, with reduced overhead, * when the application is not using the extra complexity. * *As of Sept 2012, the versions available will be: * 1) Single langauge, which only exposes slaves (such as SSR or Vthread) * 2) Single language, which only exposes tasks (such as pure dataflow) * 3) Single language, which exposes both (like Cilk, StarSs, and OpenMP) * 4) Multi-language, which always assumes both tasks and slaves * 5) Multi-language and multi-process, which also assumes both tasks and slaves * * * */ //===================== The versions of the Animation Master ================= // //============================================================================== /* 1) This version is for a single language, that has only slaves, no tasks, * such as Vthread or SSR. *This version is for when an application has only a single language, and * that language exposes slaves explicitly (as opposed to a task based * language like pure dataflow). * * *It scans the animation slots for just-completed slaves. * Each completed slave has a request in it. So, the master hands each to * the plugin's request handler (there is only one plugin, because only one * lang). *Each request represents a language construct that has been encountered * by the application code in the slave. Passing the request to the * request handler is how that language construct's behavior gets invoked. * The request handler then performs the actions of the construct's * behavior. So, the request handler encodes the behavior of the * language's parallelism constructs, and performs that when the master * hands it a slave containing a request to perform that construct. * *On a shared-memory machine, the behavior of parallelism constructs * equals control, over order of execution of code. Hence, the behavior * of the language constructs performed by the request handler is to * choose the order that slaves get animated, and thereby control the * order that application code in the slaves executes. * *To control order of animation of slaves, the request handler has a * semantic environment that holds data structures used to hold slaves * and choose when they're ready to be animated. * *Once a slave is marked as ready to be animated by the request handler, * it is the second plugin function, the Assigner, which chooses the core * the slave gets assigned to for animation. Hence, the Assigner doesn't * perform any of the semantic behavior of language constructs, rather * it gives the language a chance to improve performance. The performance * of application code is strongly related to communication between * cores. On shared-memory machines, communication is caused during * execution of code, by memory accesses, and how much depends on contents * of caches connected to the core executing the code. So, the placement * of slaves determines the communication caused during execution of the * slave's code. *The point of the Assigner, then, is to use application information during * execution of the program, to make choices about slave placement onto * cores, with the aim to put slaves close to caches containing the data * used by the slave's code. * *========================================================================== *In summary, the animationMaster scans the slots, finds slaves * just-finished, which hold requests, pass those to the request handler, * along with the semantic environment, and the request handler then manages * the structures in the semantic env, which controls the order of * animation of slaves, and so embodies the behavior of the language * constructs. *The animationMaster then rescans the slots, offering each empty one to * the Assigner, along with the semantic environment. The Assigner chooses * among the ready slaves in the semantic Env, finding the one best suited * to be animated by that slot's associated core. * *========================================================================== *Implementation Details: * *There is a separate masterVP for each core, but a single semantic * environment shared by all cores. Each core also has its own scheduling * slots, which are used to communicate slaves between animationMaster and * coreController. There is only one global variable, _PRTopEnv, which * holds the semantic env and other things shared by the different * masterVPs. The request handler and Assigner are registered with * the animationMaster by the language's init function, and a pointer to * each is in the _PRTopEnv. (There are also some pthread related global * vars, but they're only used during init of PR). *PR gains control over the cores by essentially "turning off" the OS's * scheduler, using pthread pin-to-core commands. * *The masterVPs are created during init, with this animationMaster as their * top level function. The masterVPs use the same SlaveVP data structure, * even though they're not slave VPs. *A "seed slave" is also created during init -- this is equivalent to the * "main" function in C, and acts as the entry-point to the PR-language- * based application. *The masterVPs share a single system-wide master-lock, so only one * masterVP may be animated at a time. *The core controllers access _PRTopEnv to get the masterVP, and when * they start, the slots are all empty, so they run their associated core's * masterVP. The first of those to get the master lock sees the seed slave * in the shared semantic environment, so when it runs the Assigner, that * returns the seed slave, which the animationMaster puts into a scheduling * slot then switches to the core controller. That then switches the core * over to the seed slave, which then proceeds to execute language * constructs to create more slaves, and so on. Each of those constructs * causes the seed slave to suspend, switching over to the core controller, * which eventually switches to the masterVP, which executes the * request handler, which uses PR primitives to carry out the creation of * new slave VPs, which are marked as ready for the Assigner, and so on.. * *On animation slots, and system behavior: * A request may linger in an animation slot for a long time while * the slaves in the other slots are animated. This only becomes a problem * when such a request is a choke-point in the constraints, and is needed * to free work for *other* cores. To reduce this occurrence, the number * of animation slots should be kept low. In balance, having multiple * animation slots amortizes the overhead of switching to the masterVP and * executing the animationMaster code, which drives for more than one. In * practice, the best balance should be discovered by profiling. */ void animationMaster( void *initData, SlaveVP *masterVP ) { //Used while scanning and filling animation slots int32 slotIdx, numSlotsFilled; AnimSlot *currSlot, **animSlots; SlaveVP *assignedSlaveVP; //the slave chosen by the assigner //Local copies, for performance MasterEnv *masterEnv; SlaveAssigner slaveAssigner; RequestHandler requestHandler; void *semanticEnv; int32 thisCoresIdx; //======================== Initializations ======================== masterEnv = (MasterEnv*)_PRTopEnv; thisCoresIdx = masterVP->coreAnimatedBy; animSlots = masterEnv->allAnimSlots[thisCoresIdx]; requestHandler = masterEnv->requestHandler; slaveAssigner = masterEnv->slaveAssigner; semanticEnv = masterEnv->semanticEnv; HOLISTIC__Insert_Master_Global_Vars; //======================== animationMaster ======================== while(1){ MEAS__Capture_Pre_Master_Point //Scan the animation slots numSlotsFilled = 0; for( slotIdx = 0; slotIdx < NUM_ANIM_SLOTS; slotIdx++) { currSlot = animSlots[ slotIdx ]; //Check if newly-done slave in slot, which will need request handled if( currSlot->workIsDone ) { currSlot->workIsDone = FALSE; currSlot->needsSlaveAssigned = TRUE; HOLISTIC__Record_AppResponder_start; MEAS__startReqHdlr; currSlot->workIsDone = FALSE; currSlot->needsSlaveAssigned = TRUE; SlaveVP *currSlave = currSlot->slaveAssignedToSlot; justAddedReqHdlrChg(); //handle the request, either by PR or by the language if( currSlave->requests->reqType != LangReq ) { //The request is a standard PR one, not one defined by the // language, so PR handles it, then queues slave to be assigned handleReqInPR( currSlave ); writePrivQ( currSlave, PRReadyQ ); //Q slave to be assigned below } else { MEAS__startReqHdlr; //Language handles request, which is held inside slave struc (*requestHandler)( currSlave, semanticEnv ); MEAS__endReqHdlr; } } //process the requests made by the slave (held inside slave struc) (*requestHandler)( currSlot->slaveAssignedToSlot, semanticEnv ); HOLISTIC__Record_AppResponder_end; MEAS__endReqHdlr; } //If slot empty, hand to Assigner to fill with a slave if( currSlot->needsSlaveAssigned ) { //Call plugin's Assigner to give slot a new slave HOLISTIC__Record_Assigner_start; assignedSlaveVP = (*slaveAssigner)( semanticEnv, currSlot ); //put the chosen slave into slot, and adjust flags and state if( assignedSlaveVP != NULL ) { currSlot->slaveAssignedToSlot = assignedSlaveVP; assignedSlaveVP->animSlotAssignedTo = currSlot; currSlot->needsSlaveAssigned = FALSE; numSlotsFilled += 1; HOLISTIC__Record_Assigner_end; } } } MEAS__Capture_Post_Master_Point; masterSwitchToCoreCtlr( masterVP ); flushRegisters(); DEBUG__printf(FALSE,"came back after switch to core -- so lock released!"); }//while(1) } /* 2) This version is for a single language that has only tasks, which * cannot be suspended. */ void animationMaster( void *initData, SlaveVP *masterVP ) { //Used while scanning and filling animation slots int32 slotIdx, numSlotsFilled; AnimSlot *currSlot, **animSlots; SlaveVP *assignedSlaveVP; //the slave chosen by the assigner //Local copies, for performance MasterEnv *masterEnv; SlaveAssigner slaveAssigner; RequestHandler requestHandler; PRSemEnv *semanticEnv; int32 thisCoresIdx; //#ifdef MODE__MULTI_LANG SlaveVP *slave; PRProcess *process; PRConstrEnvHolder *constrEnvHolder; int32 langMagicNumber; //#endif //======================== Initializations ======================== masterEnv = (MasterEnv*)_PRTopEnv; thisCoresIdx = masterVP->coreAnimatedBy; animSlots = masterEnv->allAnimSlots[thisCoresIdx]; requestHandler = masterEnv->requestHandler; slaveAssigner = masterEnv->slaveAssigner; semanticEnv = masterEnv->semanticEnv; //initialize, for non-multi-lang, non multi-proc case // default handler gets put into master env by a registration call by lang endTaskHandler = masterEnv->defaultTaskHandler; HOLISTIC__Insert_Master_Global_Vars; //======================== animationMaster ======================== //Do loop gets requests handled and work assigned to slots.. // work can either be a task or a resumed slave //Having two cases makes this logic complex.. can be finishing either, and // then the next available work may be either.. so really have two distinct // loops that are inter-twined.. while(1){ MEAS__Capture_Pre_Master_Point //Scan the animation slots numSlotsFilled = 0; for( slotIdx = 0; slotIdx < NUM_ANIM_SLOTS; slotIdx++) { currSlot = animSlots[ slotIdx ]; //Check if newly-done slave in slot, which will need request handled if( currSlot->workIsDone ) { currSlot->workIsDone = FALSE; HOLISTIC__Record_AppResponder_start; //TODO: update to check which process for each slot MEAS__startReqHdlr; //process the request made by the slave (held inside slave struc) slave = currSlot->slaveAssignedToSlot; //check if the completed work was a task.. if( slave->taskMetaInfo->isATask ) { if( slave->reqst->type == TaskEnd ) { //do task end handler, which is registered separately //note, end hdlr may use semantic data from reqst.. //#ifdef MODE__MULTI_LANG //get end-task handler //taskEndHandler = lookup( slave->reqst->langMagicNumber, processEnv ); taskEndHandler = slave->taskMetaInfo->endTaskHandler; //#endif (*taskEndHandler)( slave, semanticEnv ); goto AssignWork; } else //is a task, and just suspended { //turn slot slave into free task slave & make replacement if( slave->typeOfVP == SlotTaskSlv ) changeSlvType(); //goto normal slave request handling goto SlaveReqHandling; } } else //is a slave that suspended { SlaveReqHandling: (*requestHandler)( slave, semanticEnv ); //(note: indirect Fn call more efficient when use fewer params, instead re-fetch from slave) HOLISTIC__Record_AppResponder_end; MEAS__endReqHdlr; goto AssignWork; } } //if has suspended slave that needs handling //if slot empty, hand to Assigner to fill with a slave if( currSlot->needsSlaveAssigned ) { //Call plugin's Assigner to give slot a new slave HOLISTIC__Record_Assigner_start; AssignWork: assignedSlaveVP = assignWork( semanticEnv, currSlot ); //put the chosen slave into slot, and adjust flags and state if( assignedSlaveVP != NULL ) { currSlot->slaveAssignedToSlot = assignedSlaveVP; assignedSlaveVP->animSlotAssignedTo = currSlot; currSlot->needsSlaveAssigned = FALSE; numSlotsFilled += 1; } else { currSlot->needsSlaveAssigned = TRUE; //local write } HOLISTIC__Record_Assigner_end; }//if slot needs slave assigned }//for( slotIdx.. MEAS__Capture_Post_Master_Point; masterSwitchToCoreCtlr( masterVP ); //returns when ctlr switches back to master flushRegisters(); }//while(1) } /*This is the master when just multi-lang, but not multi-process mode is on. * This version has to handle both tasks and slaves, and do extra work of * looking up the semantic env and handlers to use, for each completed bit of * work. *It also has to search through the semantic envs to find one with work, * then ask that env's assigner to return a unit of that work. * *The language is written to startup in the same way as if it were the only * language in the app, and it operates in the same way, * the only difference between single language and multi-lang is here, in the * master. *This invisibility to mode is why the language has to use registration calls * for everything during startup -- those calls do different things depending * on whether it's single-language or multi-language mode. * *In this version of the master, work can either be a task or a resumed slave *Having two cases makes this logic complex.. can be finishing either, and * then the next available work may be either.. so really have two distinct * loops that are inter-twined.. * *Some special cases: * A task-end is a special case for a few reasons (below). * A task-end can't block a slave (can't cause it to "logically suspend") * A task available for work can only be assigned to a special slave, which * has been set aside for doing tasks, one such task-slave is always * assigned to each slot. So, when a task ends, a new task is assigned to * that slot's task-slave right away. * But if no tasks are available, then have to switch over to looking at * slaves to find one ready to resume, to find work for the slot. * If a task just suspends, not ends, then its task-slave is no longer * available to take new tasks, so a new task-slave has to be assigned to * that slot. Then the slave of the suspended task is turned into a free * task-slave and request handling is done on it as if it were a slave * that suspended. * After request handling, do the same sequence of looking for a task to be * work, and if none, look for a slave ready to resume, as work for the slot. * If a slave suspends, handle its request, then look for work.. first for a * task to assign, and if none, slaves ready to resume. * Another special case is when task-end is done on a free task-slave.. in * that case, the slave has no more work and no way to get more.. so place * it into a recycle queue. * If no work is found of either type, then do a special thing to prune down * the extra slaves in the recycle queue, just so don't get too many.. * *The multi-lang thing complicates matters.. * *For request handling, it means have to first fetch the semantic environment * of the language, and then do the request handler pointed to by that * semantic env. *For assigning, things get more complex because of competing goals.. One * goal is for language specific stuff to be used during assignment, so * assigner can make higher quality decisions.. but with multiple languages, * which only get mixed in the application, the assigners can't be written * with knowledge of each other. So, they can only make localized decisions, * and so different language's assigners may interfere with each other.. * *So, have some possibilities available: *1) can have a fixed scheduler in the proto-runtime, that all the * languages give their work to.. (but then lose language-specific info, * there is a standard PR format for assignment info, and the langauge * attaches this to the work-unit when it gives it to PR.. also have issue * with HWSim, which uses a priority Q instead of FIFO, and requests can * "undo" previous work put in, so request handlers need way to manipulate * the work-holding Q..) (this might be fudgeable with * HWSim, if the master did a lang-supplied callback each time it assigns a * unit to a slot.. then HWSim can keep exactly one unit of work in PR's * queue at a time.. but this is quite hack-like.. or perhaps HWSim supplies * a task-end handler that kicks the next unit of work from HWSim internal * priority queue, over to PR readyQ) *2) can have each language have its own semantic env, that holds its own * work, which is assigned by its own assigner.. then the master searches * through all the semantic envs to find one with work and asks it give work.. * (this has downside of blinding assigners to each other.. but does work * for HWSim case) *3) could make PR have a different readyQ for each core, and ask the lang * to put work to the core it prefers.. but the work may be moved by PR if * needed, say if one core idles for too long. This is a hybrid approach, * letting the language decide which core, but PR keeps the work and does it * FIFO style.. (this might als be fudgeable with HWSim, in similar fashion, * but it would be complicated by having to track cores separately) * *Choosing 2, to keep compatibility with single-lang mode.. it allows the same * assigner to be used for single-lang as for multi-lang.. the overhead of * the extra master search for work is part of the price of the flexibility, * but should be fairly small.. takes the first env that has work available, * and whatever it returns is assigned to the slot.. * *As a hybrid, giving an option for a unified override assigner to be registered * and used.. This allows something like a static analysis to detect * which languages are grouped together, and then analyze the pattern of * construct calls, and generate a custom assigner that uses info from all * the languages in a unified way.. Don't really expect this to happen, * but making it possible. */ #ifdef MODE__MULTI_LANG void animationMaster( void *initData, SlaveVP *masterVP ) { //Used while scanning and filling animation slots int32 slotIdx, numSlotsFilled; AnimSlot *currSlot, **animSlots; SlaveVP *assignedSlaveVP; //the slave chosen by the assigner //Local copies, for performance MasterEnv *masterEnv; SlaveAssigner slaveAssigner; RequestHandler requestHandler; PRSemEnv *semanticEnv; int32 thisCoresIdx; //#ifdef MODE__MULTI_LANG SlaveVP *slave; PRProcess *process; PRConstrEnvHolder *constrEnvHolder; int32 langMagicNumber; //#endif //======================== Initializations ======================== masterEnv = (MasterEnv*)_PRTopEnv; thisCoresIdx = masterVP->coreAnimatedBy; animSlots = masterEnv->allAnimSlots[thisCoresIdx]; requestHandler = masterEnv->requestHandler; slaveAssigner = masterEnv->slaveAssigner; semanticEnv = masterEnv->semanticEnv; //initialize, for non-multi-lang, non multi-proc case // default handler gets put into master env by a registration call by lang endTaskHandler = masterEnv->defaultTaskHandler; HOLISTIC__Insert_Master_Global_Vars; //======================== animationMaster ======================== //Do loop gets requests handled and work assigned to slots.. // work can either be a task or a resumed slave //Having two cases makes this logic complex.. can be finishing either, and // then the next available work may be either.. so really have two distinct // loops that are inter-twined.. while(1){ MEAS__Capture_Pre_Master_Point //Scan the animation slots numSlotsFilled = 0; for( slotIdx = 0; slotIdx < NUM_ANIM_SLOTS; slotIdx++) { currSlot = animSlots[ slotIdx ]; //Check if newly-done slave in slot, which will need request handled if( currSlot->workIsDone ) { currSlot->workIsDone = FALSE; HOLISTIC__Record_AppResponder_start; //TODO: update to check which process for each slot MEAS__startReqHdlr; //process the request made by the slave (held inside slave struc) slave = currSlot->slaveAssignedToSlot; //check if the completed work was a task.. if( slave->taskMetaInfo->isATask ) { if( slave->reqst->type == TaskEnd ) { //do task end handler, which is registered separately //note, end hdlr may use semantic data from reqst.. //#ifdef MODE__MULTI_LANG //get end-task handler //taskEndHandler = lookup( slave->reqst->langMagicNumber, processEnv ); taskEndHandler = slave->taskMetaInfo->endTaskHandler; //#endif (*taskEndHandler)( slave, semanticEnv ); goto AssignWork; } else //is a task, and just suspended { //turn slot slave into free task slave & make replacement if( slave->typeOfVP == SlotTaskSlv ) changeSlvType(); //goto normal slave request handling goto SlaveReqHandling; } } else //is a slave that suspended { SlaveReqHandling: (*requestHandler)( slave, semanticEnv ); //(note: indirect Fn call more efficient when use fewer params, instead re-fetch from slave) HOLISTIC__Record_AppResponder_end; MEAS__endReqHdlr; goto AssignWork; } } //if has suspended slave that needs handling //if slot empty, hand to Assigner to fill with a slave if( currSlot->needsSlaveAssigned ) { //Call plugin's Assigner to give slot a new slave HOLISTIC__Record_Assigner_start; AssignWork: assignedSlaveVP = assignWork( semanticEnv, currSlot ); //put the chosen slave into slot, and adjust flags and state if( assignedSlaveVP != NULL ) { currSlot->slaveAssignedToSlot = assignedSlaveVP; assignedSlaveVP->animSlotAssignedTo = currSlot; currSlot->needsSlaveAssigned = FALSE; numSlotsFilled += 1; } else { currSlot->needsSlaveAssigned = TRUE; //local write } HOLISTIC__Record_Assigner_end; }//if slot needs slave assigned }//for( slotIdx.. MEAS__Capture_Post_Master_Point; masterSwitchToCoreCtlr( masterVP ); //returns when ctlr switches back to master flushRegisters(); }//while(1) } #endif //MODE__MULTI_LANG //This is the master when both multi-lang and multi-process modes are turned on //#ifdef MODE__MULTI_LANG //#ifdef MODE__MULTI_PROCESS void animationMaster( void *initData, SlaveVP *masterVP ) { int32 slotIdx; // int32 numSlotsFilled; AnimSlot *currSlot; //Used while scanning and filling animation slots AnimSlot **animSlots; //Local copies, for performance MasterEnv *masterEnv; int32 thisCoresIdx; //======================== Initializations ======================== masterEnv = (MasterEnv*)_PRTopEnv; thisCoresIdx = masterVP->coreAnimatedBy; animSlots = masterEnv->allAnimSlots[thisCoresIdx]; HOLISTIC__Insert_Master_Global_Vars; //======================== animationMaster ======================== //Do loop gets requests handled and work assigned to slots.. // work can either be a task or a resumed slave //Having two cases makes this logic complex.. can be finishing either, and // then the next available work may be either.. so really have two distinct // loops that are inter-twined.. while(1) { MEAS__Capture_Pre_Master_Point for( slotIdx = 0; slotIdx < NUM_ANIM_SLOTS; slotIdx++) { currSlot = animSlots[ slotIdx ]; masterFunction_multiLang( currSlot ); } MEAS__Capture_Post_Master_Point; masterSwitchToCoreCtlr( masterVP ); //returns when ctlr switches back to master flushRegisters(); } } #endif //MODE__MULTI_LANG #endif //MODE__MULTI_PROCESS inline void masterFunction_multiLang( AnimSlot *currSlot ) { //Scan the animation slots int32 magicNumber; SlaveVP *slave; SlaveVP *assignedSlaveVP; PRSemEnv *semanticEnv; PRReqst *req; RequestHandler requestHandler; //Check if newly-done slave in slot, which will need request handled if( currSlot->workIsDone ) { currSlot->workIsDone = FALSE; HOLISTIC__Record_AppResponder_start; //TODO: update to check which process for each slot MEAS__startReqHdlr; //process the request made by the slave (held inside slave struc) slave = currSlot->slaveAssignedToSlot; //check if the slave was doing a task.. //Action depends on both on the request type, and whether it's on // a generic slave vs a suspended task if( slave->metaTask->taskType == AtomicTask || slave->metaTask->taskType == SuspendedTask ) { switch( slave->request->reqType ) { case TaskEnd: { PRHandle_EndTask( slave ); //if free task slave, update count, put into recycle Q -- do handler before lang's handler //do task end handler, which is registered separately //note, end hdlr may use semantic data from reqst.. //get end-task handler RequestHandler taskEndHandler = slave->request->handler; semanticEnv = PR_int__give_sem_env_for_slave( slave, slave->request->langMagicNumber ); (*taskEndHandler)( slave, semanticEnv ); goto AssignWork; } case TaskCreate: { PRHandle_CreateTask( slave ); RequestHandler taskCreateHandler = slave->request->handler; semanticEnv = PR_int__give_sem_env_for_slave( slave, slave->request->langMagicNumber ); (*taskCreateHandler)( slave, semanticEnv ); //resumes creating slave goto AssignWork; } default: { //is a task, and just suspended, so tied to a free task slave //First turn slot slave into free task slave & make replacement if( slave->typeOfVP == SlotTaskSlv ) replaceWithNewSlotSlv( slave, slave->processSlaveIsIn->processEnv ); //goto normal slave request handling goto SlaveReqHandling; } } } else //is a slave that suspended { SlaveReqHandling: //Q: put the switch in inline call, to clean up code? req = slave->request; switch( req->reqType ) { case SlvCreate: PRHandle_CreateSlave( slave ); break; case SlvDissipate: PRHandle_Dissipate( slave ); break; case Service: PR_int__handle_PRServiceReq( slave ); break; //resume into PR's own semantic env case Hardware: //for future expansion case IO: //for future expansion case OSCall: //for future expansion case Language: //normal sem request magicNumber = slave->request->langMagicNumber; semanticEnv = PR_PI__give_sem_env_for( slave, magicNumber ); requestHandler = semanticEnv->requestHdlr; (*requestHandler)( slave, semanticEnv ); //(note: indirect Fn call more efficient when use fewer params, instead re-fetch from slave) } HOLISTIC__Record_AppResponder_end; MEAS__endReqHdlr; goto AssignWork; } } //if has suspended slave that needs handling //End up here when the slot did not have ended work in it (no req) //So, here, if slot empty, look for work to fill the slot if( currSlot->needsSlaveAssigned ) { HOLISTIC__Record_Assigner_start; AssignWork: //Scan sem environs, looking for semEnv with ready work. // call the Assigner for that sem Env, to get a slave for the slot assignedSlaveVP = assignWork( semanticEnv, currSlot ); //put the chosen slave into slot, and adjust flags and state if( assignedSlaveVP != NULL ) { currSlot->slaveAssignedToSlot = assignedSlaveVP; assignedSlaveVP->animSlotAssignedTo = currSlot; currSlot->needsSlaveAssigned = FALSE; } else { currSlot->needsSlaveAssigned = TRUE; //local write } HOLISTIC__Record_Assigner_end; }//if slot needs slave assigned } //========================================================================== /*When a task in a slot slave suspends, the slot slave has to be changed to * a free task slave, then the slot slave replaced. The replacement can be * either a recycled free task slave that finished it's task and has been * idle in the recycle queue, or else create a new slave to be the slot slave. *The master only calls this with a slot slave that needs to be replaced. */ inline void replaceWithNewSlotSlv( SlaveVP *requestingSlv, PRProcess *process ) { SlaveVP *newSlotSlv; //get a new slave to be the slot slave newSlotSlv = readPrivQ( process->freeTaskSlvRecycleQ ); if( newSlotSlv == NULL ) { newSlotSlv = PR_int__create_slaveVP( &idle_fn, NULL, process, 0); //just made a new free task slave, so count it process->numLiveFreeTaskSlvs += 1; } //set slave values to make it the slot slave newSlotSlv->metaTask = NULL; newSlotSlv->typeOfVP = SlotTaskSlv; newSlotSlv->needsTaskAssigned = TRUE; //a slot slave is pinned to a particular slot on a particular core //Note, this happens before the request is seen by handler, so nothing // has had a chance to change the coreAnimatedBy or anything else.. newSlotSlv->animSlotAssignedTo = requestingSlv->animSlotAssignedTo; newSlotSlv->coreAnimatedBy = requestingSlv->coreAnimatedBy; //put it into the slot slave matrix int32 slotNum = requestingSlv->animSlotAssignedTo->slotIdx; int32 coreNum = requestingSlv->coreAnimatedBy; process->slotTaskSlvs[coreNum][slotNum] = newSlotSlv; //Fix up requester, to be an extra slave now (but not an ended one) // because it's active, doesn't go into freeTaskSlvRecycleQ requestingSlv->typeOfVP = FreeTaskSlv; check_if_need_to_change_metaTask_type_or_something; } /*This does: * 1) searches the semantic environments for one with work ready * if finds one, asks its assigner to return work * 2) checks what kind of work: new task, resuming task, resuming slave * if new task, gets the slot slave and assigns task to it and returns slave * else, gets the slave attached to the metaTask and returns that. * 3) if no work found, then prune former task slaves waiting to be recycled. * If no work and no slaves to prune, check for shutdown conditions. * * Semantic env keeps its own work in its own structures, and has its own * assigner. It chooses * However, include a switch that switches-in an override assigner, which * sees all the work in all the semantic env's. This is most likely * generated by static tools and included in the executable. That means it * has to be called via a registered pointer from here. The idea is that * the static tools know which languages are grouped together.. and the * override enables them to generate a custom assigner that uses info from * all the languages in a unified way.. Don't really expect this to happen, * but am making it possible. */ inline SlaveVP * assignWork( PRProcess *process, AnimSlot *slot ) { SlaveVP *returnSlv; //VSsSemEnv *semEnv; //VSsSemData *semData; int32 coreNum, slotNum; PRMetaTask *newMetaTask, *assignedMetaTask; SlaveVP *freeTaskSlv; coreNum = slot->coreSlotIsOn; if( _PRTopEnv->overrideAssigner != NULL ) { assignedMetaTask = (*_PRTopEnv->overrideAssigner)( process, slot ); if( assignedMetaTask != NULL ) { //have work, so reset Done flag (caused by work generated on other core) if( process->coreIsDone[coreNum] == TRUE ) //reads are higher perf process->coreIsDone[coreNum] = FALSE; //don't just write always switch( assignedMetaTask->taskType ) { case GenericSlave: goto AssignSlave; case SuspendedTask: goto AssignSlave; case AtomicTask: goto AssignNewTask; default: PR_int__throw_exception( "unknown task type ret by assigner" ); } } else goto NoWork; } //If here, then no override assigner, so search semantic envs for work int32 envIdx, numEnvs; PRSemEnv **semEnvs, *semEnv; SlaveAssigner assigner; semEnvs = process->semEnvs; numEnvs = process->numSemEnvs; for( envIdx = 0; envIdx < numEnvs; envIdx++ ) //keep semEnvs in hash AND array { semEnv = semEnvs[envIdx]; if( semEnv->hasWork ) { assigner = semEnv->slaveAssigner; assignedMetaTask = (*assigner)( semEnv, slot ); //have work, so reset Done flag (caused by work generated on other core) if( process->coreIsDone[coreNum] == TRUE ) //reads are higher perf process->coreIsDone[coreNum] = FALSE; //don't just write always switch( assignedMetaTask->taskType ) { case GenericSlave: goto AssignSlave; case SuspendedTask: goto AssignSlave; case AtomicTask: goto AssignNewTask; default: PR_int__throw_exception( "unknown task type ret by assigner" ); } } } NoWork: //No work, if reach here.. { goto ReturnTheSlv; } AssignSlave: //Have a metaTask attached to a slave, so get the slave out { //get slave pointed to by meta task. returnSlv = assignedMetaTask->slaveAssignedTo; returnSlv->coreAnimatedBy = coreNum; goto ReturnTheSlv; } AssignNewTask: { //get the slot slave to assign the task to.. coreNum = slot->coreSlotIsOn; slotNum = slot->slotIdx; returnSlv = process->slotTaskSlvs[coreNum][slotNum]; //point slave to task's function, and mark slave as having task PR_int__reset_slaveVP_to_TopLvlFn( returnSlv, assignedMetaTask->topLevelFn, assignedMetaTask->initData ); returnSlv->metaTask = assignedMetaTask; assignedMetaTask->slaveAssignedTo = returnSlv; returnSlv->needsTaskAssigned = FALSE; //slot slave is a "Task" slave type //have work, so reset Done flag, if was set if( process->coreIsDone[coreNum] == TRUE ) //reads are higher perf process->coreIsDone[coreNum] = FALSE; //don't just write always goto ReturnTheSlv; } ReturnTheSlv: //All paths goto here.. to provide single point for holistic.. #ifdef HOLISTIC__TURN_ON_OBSERVE_UCC if( returnSlv == NULL ) { returnSlv = process->idleSlv[coreNum][slotNum]; //things that would normally happen in resume(), but idle VPs // never go there returnSlv->numTimesAssignedToASlot++; //gives each idle unit a unique ID Unit newU; newU.vp = returnSlv->slaveID; newU.task = returnSlv->numTimesAssignedToASlot; addToListOfArrays(Unit,newU,process->unitList); if (returnSlv->numTimesAssignedToASlot > 1) //make a dependency from prev idle unit { Dependency newD; // to this one newD.from_vp = returnSlv->slaveID; newD.from_task = returnSlv->numTimesAssignedToASlot - 1; newD.to_vp = returnSlv->slaveID; newD.to_task = returnSlv->numTimesAssignedToASlot; addToListOfArrays(Dependency, newD ,process->ctlDependenciesList); } } else //have a slave will be assigned to the slot { //assignSlv->numTimesAssigned++; //get previous occupant of the slot Unit prev_in_slot = process->last_in_slot[coreNum * NUM_ANIM_SLOTS + slotNum]; if(prev_in_slot.vp != 0) //if not first slave in slot, make dependency { Dependency newD; // is a hardware dependency newD.from_vp = prev_in_slot.vp; newD.from_task = prev_in_slot.task; newD.to_vp = returnSlv->slaveID; newD.to_task = returnSlv->numTimesAssignedToASlot; addToListOfArrays(Dependency,newD,process->hwArcs); } prev_in_slot.vp = returnSlv->slaveID; //make new slave the new previous prev_in_slot.task = returnSlv->numTimesAssignedToASlot; process->last_in_slot[coreNum * NUM_ANIM_SLOTS + slotNum] = prev_in_slot; } #endif return( returnSlv ); } /*In creator, only PR related things happen, and things in the langlet whose * creator construct was used. *Other langlet still gets a chance to create semData -- but by registering a * "createSemData" handler in the semEnv. When a construct of the langlet * calls "PR__give_sem_data()", if there is no semData for that langlet, * the PR will call the creator in the langlet's semEnv, place whatever it * makes as the semData in that slave for that langlet, and return that semData * *So, as far as counting things, a langlet is only allowed to count creation * of slaves it creates itself.. may have to change this later.. add a way for * langlet to register a trigger Fn called each time a slave gets created.. * need more experience with what langlets will do at create time.. think Cilk * has interesting create behavior.. not sure how that will differ in light * of true tasks and langlet approach. Look at it after all done and start * modifying the langs to be langlets.. * *PR itself needs to create the slave, then update numLiveSlaves in process, * copy processID from requestor to newly created */ PRHandle_CreateSlave( PRReqst *req, SlaveVP *requestingSlv ) { SlaveVP *newSlv; PRMetaTask metaTask; PRProcess *process; process = requestingSlv->processSlaveIsIn; newSlv = PR_int__create_slaveVP(); newSlv->typeOfVP = GenericSlv; newSlv->processSlaveIsIn = process; process->numLiveGenericSlvs += 1; metaTask = PR_int__create_slave_meta_task(); metaTask->taskID = req->ID; metaTask->taskType = GenericSlave; (*req->handler)(newSlv); } /*The dissipate handler has to update the number of slaves of the type, within * the process, and call the langlet handler linked into the request, * and after that returns, then call the PR function that frees the slave state * (or recycles the slave). * *The PR function that frees the slave state has to also free all of the * semData in the slave.. or else reset all of the semDatas.. by, say, marking * them, then in PR__give_semData( magicNum ) call the langlet registered * "resetSemData" Fn. */ PRHandle_Dissipate( SlaveVP *slave ) { PRProcess *process; void *semEnv; process = slave->processSlaveIsIn; //do the language's dissipate handler semEnv = PR_int__give_sem_env_for( slave, slave->request->langMagicNumber ); (*slave->request->handler)( slave, semEnv ); process->numLiveGenericSlvs -= 1; PR_int__dissipate_slaveVP_multilang( slave ); //recycles and resets semDatas //check End Of Process Condition if( process->numLiveTasks == 0 && process->numLiveGenericSlvs == 0 ) signalEndOfProcess; } /*Create task is a special form, that has PR behavior in addition to plugin * behavior. Master calls this first, and this in turn calls the plugin's * create task handler. */ inline void PRHandle_CreateTask( TopLevelFn topLevelFn, void *initData, PRReqst *req, SlaveVP *requestingSlv ) { PRMetaTask *metaTask; PRProcess *process; void *semEnv, _langMetaTask; PRLangMetaTask *langMetaTask; process = requestingSlv->processSlaveIsIn; metaTask = PR_int__create_meta_task( req ); metaTask->taskID = req->ID; //may be NULL metaTask->topLevelFn = topLevelFn; metaTask->initData = initData; process->numLiveTasks += 1; //plugin tracks tasks ready, and has its own assigner, so task doesn't // come back from lang's handler -- it's consumed and stays in semEnv. //But handler gives back the language-specific meta-task it creates, and // then hook that into the PR meta-task //(Could also do PRMetaTask as a prolog -- make a Fn that takes the size // of the lang's metaTask, and alloc's that plus the prolog and returns // ptr to position just above the prolog) semEnv = PR_int__give_semEnv_of_req( req, requestingSlv ); //magic num in req _langMetaTask = (*requestingSlv->request->handler)(req, semEnv); langMetaTask = (PRLangMetaTask *)_langMetaTask; metaTask->langMetaTask = langMetaTask; langMetaTask->protoMetaTask = metaTask; return; } /*When a task ends, are two scenarios: 1) task ran to completion, or 2) task * suspended at some point in its code. *For 1, just decr count of live tasks (and check for end condition) -- the * master loop will decide what goes into the slot freed up by this task end, * so, here, don't worry about assigning a new task to the slot slave. *For 2, the task's slot slave has been converted to a free task slave, which * now has nothing more to do, so send it to the recycle Q (which includes * freeing all the semData and meta task structs alloc'd for it). Then * decrement the live task count and check end condition. * *PR has to update count of live tasks, and check end of process condition. * There are constructs that wait for a process to end, so when end detected, * have to resume what's waiting.. *Thing is, the wait is used in "main", so it's an OS thread. That means * PR internals have to do OS thread signaling. Want to do that in the * core controller, which has the original stack of an OS thread. * *So here, when detect process end, signal to the core controller, which will * then do the condition variable notify to the OS thread that's waiting. */ inline void PRHandle_EndTask( SlaveVP *requestingSlv ) { void *semEnv; PRReqst *req; PRMetaTask *metaTask; PRProcess *process; req = requestingSlv->request; semEnv = PR_int__give_semEnv_of_req( req, requestingSlv ); //magic num in req metaTask = req->metaTask; //Want to keep PRMetaTask hidden from plugin, so extract semReq.. (*req->handler)( metaTask, req->semReq, semEnv ); recycleFreeTaskSlave( requestingSlv ); process->numLiveTasks -= 1; //check End Of Process Condition if( process->numLiveTasks == 0 && process->numLiveGenericSlvs == 0 ) signalEndOfProcessToCoreCtlr; }