Combatting Rendering Latency

Transcription

1 Comtting Rendering Lteny Mr Olno, Jon Cohen, Mrk Mine, Gry Bishop Deprtment of Computer Siene, UNC Chpel Hill ABSTRACT Lteny or lg in n intertive grphis system is the dely etween user input nd displyed output. We hve found lteny nd the pprent oing nd swimming of ojets tht it produes to e serious prolem for hed-mounted disply (HMD) nd ugmented relity pplitions. At UNC, we hve een investigting numer of wys to redue lteny; we present two of these. Slts is n experimentl rendering system for our Pixel-Plnes 5 grphis mhine gurnteeing onstnt single NTSC field of lteny. This gurnteed response is espeilly importnt for preditive trking. Just-in-time pixels is n ttempt to ompenste for rendering lteny y rendering the pixels in snned disply sed on their position in the sn. 1 INTRODUCTION 1.1 Wht is lteny? Performne of urrent grphis systems is ommonly mesured in terms of the numer of tringles rendered per seond or in terms of the numer of omplete frmes rendered per seond. While these mesures re useful, they don t tell the whole story. Lteny, whih mesures the strt to finish time of n opertion suh s drwing single, is n often negleted mesure of grphis performne. For some urrent modes of intertion, like mnipulting 3D ojet with joystik, this mesure of responsiveness my not e importnt. But for emerging modes of nturl intertion, lteny is ritil mesure. 1.2 Why is it there? All grphis systems must hve some lteny simply euse it tkes some time to ompute n. In ddition, system tht n produe new every frme my (nd often will) hve more thn one frme of lteny. This is used y the pipelining used to inrese grphis performne. The lssi prolem with pipelining is tht it provides inresed throughput t ost in lteny. The omputtions required for single frme re divided into stges nd their exeution is overlpped. This n expnd the effetive time ville to work on tht single frme sine severl frmes re eing omputed t one. However, the lteny is s long s the full time spent omputing the frme in ll of its stges. 1.3 Why is it d? Lteny is prolem for hed-mounted disply (HMD) pplitions. The higher the totl lteny, the more the world seems to lg ehind the user s hed motions. The effet of this lg is high visosity world. The effet of lteny is even more notiele with see-through HMDs. Suh displys superimpose omputer generted ojets on the user s view of the physil world. The lg eomes ovious in this sitution euse the rel world moves without lg, while the virtul ojets shift in position during the lg time, thing up to their proper positions when the user stops moving. This swimming of the virtul ojets not only detrts from the desired illusion of the ojets physil presene, ut lso hinders ny effort to use this tehnology for rel pplitions. Most see-through HMD pplitions require world without these swimming effets. If we hope to hve pplitions present 3D instrutions to guide the performne of omplex 3D tsks [9], suh s repirs to photoopy mhine or even jet engine, the instrutions must sty fixed to the mhine in question. Current reserh into the use of see-through HMDs y ostetriins to visulize 3D ultrsound dt indites the need for lower lteny visuliztion systems [3]. The use of see-through HMDs for ssisting surgil proedures is unthinkle until we mke signifint dvnes in the re of low lteny grphis systems. 2 COMBATTING LATENCY 2.1 Mthing A possile solution to this lg prolem is to use video tehniques to use the user s view of the rel world to lg in synhroniztion with the virtul world. However, this only works while the lteny is reltively smll. 2.2 Predition Another solution to the lteny prolem is to predit where the user s hed will e when the is finlly displyed [10, 1, 2]. This tehnique, lled preditive trking, involves using oth reent trking dt nd urte knowledge of the system s totl lteny to mke est guess t the position nd orienttion of the user s hed when the is displyed inside the HMD. Azum sttes tht for predition to work effetively, the lg must e smll nd onsistent. In ft he uses the single field-time lteny rendering system (Slts), whih we will disuss shortly, to hieve urte predition.

2 A/D onverter output Pendulum Worksttion Disply Grphis Proessors LED/Photodiode Trking sensor Digitl Osillosope Photodiode Figure 1: Apprtus for externl mesurement of trking nd disply lteny. 2.3 Rendering lteny: ompenstion nd redution Rnge of solutions There re wide spetrum of pprohes tht n e used to redue lg in genertion or ompenste for it. One wy to ompenste for genertion lteny is to offset the disply of the omputed sed upon the ltest ville trking dt. This tehnique is used, for exmple, y the Visul Disply Reserh Tool (VDRT), flight simultor developed t the Nvl Trining Systems Center in Orlndo, Florid [5, 6]. VDRT is helmet-mounted lser projetion system whih projets s onto retro-refletive dome (insted of using the onventionl mosi of high resolution displys found in most flight simultors). In the VDRT system, s re first omputed sed upon the predited position of the user's hed t the time of disply. Immeditely prior to redout, the most reently ville trking dt is used to ompute the errors in the predited hed position used to generte the. These errors re then used to offset the rster of the lser projetor in pith nd yw so tht the is projeted t the ngle for whih it ws omputed. Rte signls re lso lulted nd re used to develop time dependent orretion signl whih helps keep the projeted t the orret sptil orienttion s the projetor moves during the disply field period. Similrly, Regn nd Pose re uilding the prototype for hrdwre rhiteture lled the ddress relultion pipeline[15]. This system hieves very smll lteny for hed rottions y rendering sene on the six fes of ue. As pixel is needed for disply, pproprite memory lotions from the rendered ue fes re red. A hed rottion simply lters whih memory is essed, nd thus ontriutes nothing to the lteny. Hed trnsltion is hndled y ojet-spe sudivision nd omposition. Ojets re prioritized nd re-rendered s neessry to ommodte trnsltions of the user s hed. The my not lwys e orret if the rendering hrdwre nnot keep up, ut the most importnt ojets, whih inlude the losest ones, should e rendered in time to keep their positions urte. Sine pipelining n e huge soure of lg, lteny n e redued y reduing pipelining or sing it on smller units of time like polygons or pixels insted of frmes. Most ommeril grphis systems re t lest polygon pipelined. Whtever level the pipelining, system tht omputes s frme y frme is y neessity sddled with t lest frme time of lteny. Other methods overome this y divoring the genertion from the disply updte rte. Frmeless rendering[4] n e used to redue lteny in this wy. In this tehnique pixels re updted ontinuously in rndom pttern. This removes the dependene on frmes nd fields. Ring Network Renderers Frme Buffer Worksttion Figure 2: Pixel-Plnes 5 system rhiteture Pixels my e trnsformed t whtever rte is most onvenient. This redues lteny t the ost of lrity sine only portion of the pixels re updted. The trnsform rte n remin loked to the trker updte rte or seprted on pixel-y-pixel sis s with the just-in-time pixels method, disussed next Just-in-time pixels (JITP) We will present tehnique lled just-in-time pixels, whih dels with the plement of pixels on sn-line disply s prolem of temporl lising [14]. Although the disply my tke mny milliseonds to refresh, the we see on the disply typilly represents only single instnt in time. When we see n ojet in motion on the disply, it ppers distorted euse we see the higher sn lines efore we see the lower ones, mking it seem s if the lower prt of the ojet lgs ehind the upper prt. Avoidne of this distortion entils generting every pixel the wy it should pper t the ext time of its disply. This n led to redution in lteny sine neither the hed position dt, nor the output pixels re limited to inrements of n entire frme time. This ide is of limited usefulness on urrent LCD HMDs with their sluggish response. However, it works quite well on the miniture CRT HMDs urrently ville nd is lso pplile to non-intertive video pplitions Slts As more onventionl ttk on lteny, we hve designed rendering pipeline lled Slts s tested for exploring fixed nd low lteny rendering [7]. Unlike just-in-time pixels, Slts still uses the single trnsform per frme prdigm. The rendering lteny of Slts is extly one field time (16.7 ms). This is perfet for preditive trking whih requires low nd preditle lteny. We mesure this rendering lteny from the time Slts egins trnsforming the dt set into sreen oordintes to the time the disply devies egin to sn the pixel olors from the frme uffers onto the sreens. 3 MEASURING LATENCY We hve mde oth externl nd internl mesurements of the lteny of the Pixel-Plnes 5 PPHIGS grphis lirry [13, 7]. These hve shown the genertion lteny to e etween 54 nd 57 ms for miniml dt sets. The internl mesurement methods re quite speifi to the PPHIGS lirry. However, the externl mesurements n e tken for ny grphis system. The externl lteny mesurement pprtus reords three timing signls on digitl osillosope (see figure 1). A pendulum nd led/photodiode pir provide the referene time for rel-world event the low point of the pendulum s r. A trker on the pendulum is fed into the grphis system. The grphis system

3 ojet ojet time t x time t y time t x time t y snline x mer rottion snline y snline x mer rottion snline y snout t time t x snout t time t y snout t time t x snout t time t y perieved ojet perieved ojet Figure 3: Imge genertion in onventionl omputer grphis nimtion. Snline x is displyed t time t x, snline y is displyed t time t y. strts new frme when it detets the pendulum s low point from the trking dt. An D/A onverter is used to tell the osillosope when the new frme hs strted. Frmes lternte drk nd light nd photodiode tthed to the sreen is used to tell when the hnges. The trking lteny ws the time etween the signl from the pendulum s photodiode nd the rendering strt signl out of the D/A onverter. The rendering lteny ws the time etween the signl out of the D/A onverter nd the signl from the photodiode tthed to the sreen. These time stmps were verged over numer of frmes. The internl mesurements found the sme rnge of rendering ltenies. The test ws set up to e s fir s possile given the Pixel-Plnes 5 rhiteture (figure 2, explined in more detil lter). The test involved one full sreen tringle for eh grphis proessor. This ensured tht every grphis proessor would hve work to do nd would hve rendering instrutions to send to every renderer. The first severl frmes were disrded to mke sure the pipeline ws full. Finlly, lteny determined from time stmps on the grphis proessors ws verged over numer of frmes. 4 JUST-IN-TIME PIXELS 4.1 The ide When using rster disply devie, the pixels tht mke up n re not displyed ll t one ut re spred out over time. In onventionl grphis system generting NTSC video, for exmple, the pixels t the ottom of the sreen re displyed lmost 17 ms fter those t the top. Mtters re further ggrvted when using NTSC video y the ft tht not ll of the lines of n NTSC re displyed in one rster sn ut re in ft interled ross two fields. In the first field only the odd lines in n re displyed, nd in the seond field only the even. Figure 4: Imge genertion using just-in-time pixels Thus, unless nimtion is performed on fields (i.e. generting seprte for eh field), the lst pixel in n is displyed more thn 33 ms fter the first. The prolem with this sequentil redout of dt, is tht it is not refleted in the mnner in whih the is omputed. Typilly, in onventionl omputer grphis nimtion, only single trnsform is used in generting the dt for n entire frme. Eh frme represents point smple in time whih is inonsistent with the wy in whih it is displyed. As result, s shown in figure 3 nd plte 1, the does not truly reflet the position of ojets (reltive to the view point of the mer) t the time of disply of eh pixel. A quik k of the envelope lultion n demonstrte the mgnitude of the errors tht result if this disply system dely is ignored. Assuming, for exmple, mer rottion of 200 degrees/seond ( resonle vlue when ompred with pek veloities of 370 degrees/seond during typil hed motion - see [12]) we find: Assume: 1) 200 degrees/se mer rottion 2) mer generting 60 degree Field of View (FOV) 3) NTSC video 60 fields/se NTSC video ~600 pixels/fov horizontl resolution We otin: 200 degrees 1 se degrees = 3.3 mer rottion se 60 fields field Thus in 60 degree FOV when using NTSC video: 3.3 degrees 1 FOV pixels 600 = 33 pixels error 60 degrees FOV

4 Thus with mer rottion of pproximtely 200 degrees/seond, registrtion errors of more thn 30 pixels (for NTSC video) n our in one field time. The term registrtion is eing used here to desrie the orrespondene etween the displyed nd the plement of ojets in the omputer generted world. Note tht even though the ove disussion onentrtes on mer rottion, the rgument is vlid for ny reltive motion etween the mer nd virtul ojets. Thus, even if the mer's view point is unhnging, ojets moving reltive to the mer will exhiit the sme registrtion errors s ove. The mount of error is dependent upon the veloity of the ojet reltive to the mer s view diretion. If ojet motion is omined with rottion the resulting errors re orrespondingly worse. The idel wy to generte n, therefore, would e to relulte for eh pixel the position nd orienttion of the mer nd the position nd orienttion of the sene s ojets, sed upon the time of disply of tht pixel. The resulting olor nd intensity generted for the pixel will e onsistent with the pixel s time of disply. Though ojets moving reltive to the mer would pper distorted when the frme is shown sttilly, the distorted JITP ojets will tully pper undistorted when viewed on the rster disply. As shown in figure 4 nd plte 2, eh pixel in n idel just-in-time pixels renderer represents smple of the virtul world tht is onsistent with the time of the pixel s disply. Computtion of oth the mtrix nd ojet positions for eh pixel is quite expensive. Aeptle pproximtions to justin-time pixels n e otined, however, with onsiderly less omputtion. One option is to use single trnsformtion per sn line. This relies on the hnges eing smll during the short (pproximtely 65 µs) time for the line. Clultions show this to e resonle ssumption, llowing on the order of 0.13 pixels error. Another pproximtion is to use only two trnsformtions per field, one for the first pixel nd one for the lst pixel. Ojet positions re linerly interpolted etween these two. 4.3 JITP pplied to lteny A prtil test implementtion hs een onstruted tht renders s using the just-in-time pixels prdigm. This system is intended to e used in see-through HMD to help redue genertion lteny. In rel-time JITP system, insted of omputing pixel vlues sed upon the predited position nd veloity of the virtul mer, eh pixel is omputed sed upon the position nd orienttion of the user s hed t the time of disply of tht pixel. Genertion of just-in-time pixel in rel time, therefore, requires knowledge of when pixel is going to e displyed nd where the user is going to e looking t the time. This implies the ontinuous nd prllel exeution of the following two entrl funtions: 1) Synhroniztion of genertion nd snout 2) Determintion of the position nd orienttion of the user s hed t the time of disply of eh pixel By synhronizing genertion nd snout, the JITP renderer n mke use of the detils of how the pixels in n re snned out to determine when prtiulr pixel is to e displyed. By knowing wht snline the pixel is on, for exmple, nd how fst the snlines in n re displyed, the JITP renderer n esily lulte the time of disply of tht pixel. Determintion of where the user is looking n e omplished through use of onventionl hed trking system (mgneti or optil for exmple). Determintion of where the user is looking t the time of disply of pixel requires the use of preditive trking sheme. This is due to the presene of delys etween the smpling of the position nd orienttion of the user s hed nd the orresponding disply of pixel. Inluded in the end-to-end delys is the time to ollet trking dt, genertion time nd the delys due to snout. In the urrent implementtion, the lultions for eh snline re pushed s lte s possile. Idelly dt for eh snline is trnsferred to the frme uffer just efore it is red out y the rster sn. This tehnique, known s em ring, ws first used in erly flight simultors. By pushing the grphis lultion s lte s possile, em ring llows genertion delys to e omined with disply system delys. The result is lower overll end-to-end dely whih simplifies the tsk of prediting the future position nd orienttion of the user s hed. Predition lso enefits from the ft tht the delyed omputtion mkes it possile to use the ltest ville trking dt in the genertion of the predited user view point. 5 SLATS 5.1 Brief Pixel-Plnes 5 desription To understnd how Slts works requires some knowledge of Pixel-Plnes 5 [11]. Using Pixel-Plnes 5 gve us totl ontrol over the grphis softwre, whih ws ll developed in-house. Beuse our gol ws to hieve lower lteny y modifying the rendering pipeline, suh low-level ontrol ws neessry. Referring to figure 2, Pixel-Plnes 5 uses prllelism t oth the trnsformtion nd rsteriztion stges of the rendering proess. Primitives re typilly generted on host worksttion nd sent vi ring network to set of grphis proessors (GPs), where they re stored in lol disply lists. The grphis proessors trverse these disply lists, trnsforming the primitives from ojet oordintes to sreen oordintes nd generting pproprite rendering ommnds. The grphis proessors then send these ommnds over the ring to the renderers, whih perform rsteriztion nd shding. Eh of whih hndles 128x128 region of the sreen. Finlly, the renderers send the resulting pixel vlues to frme uffer, whih is synhronized with video disply for output. 5.2 PPHIGS pipeline PPHIGS is the stndrd rendering lirry for Pixel-Plnes 5. It is ontrolled y softwre lyer lled Rendering Control [8]. The rendering proess is roken into three min stges. In the trnsform stge, the GPs trnsform the primitives. In the render stge, the renderers sn onvert nd shde the primitives. If there re more regions on the sreen thn there re renderers, the first renderer to finish strts on the next sreen region. Finlly, in the opy stge, the pixel dt is opied into the frme uffer. This is illustrted in figure 5. In this timing digrm nd the ones tht follow, eh line shows use of n independent hrdwre resoure. So the GPs, renderers, nd frme uffer n ll e used simultneously. However one stge on the GPs must e finished efore the next n egin. Arrows show, for one frme of interest, the dependenies etween the different resoures. All other timings n (nd proly will) hnge depending on the ontents of the sene.

5 GPs Renderers Frme Buffer = one frme GPs Figure 5: Bsi PPHIGS timing for frme pssing through the pipeline.,, nd re the trnsform, render, nd opy stges respetively for single frme. The rrow etween the middle of nd the strt of indites tht n egin s soon s the first region is finished in. For stereo opertion, PPHIGS hndles first the left eye nd then the right eye. However, oth re onsidered prt of single unit. When the pplition softwre sys to drw frme, s for oth eyes re drwn. This is illustrted in figure 6. Renderers Frme Buffers R L R = left eye = right eye GP Renderer L R e d f Figure 6: PPHIGS timing for stereo pir pssing through the pipeline.,, nd e re the trnsform, render, nd opy stges of the left eye., d, nd f re the right eye. As ws mentioned erlier, the timings, other thn those expliitly shown, n vry quite it. The lowest lteny possile with PPHIGS ours when the trnsform nd render stges re smll nd the opy time is the limiting ftor. In this se, the synhroniztion etween the stges fores three fields of lteny etween the time the trnsformtion egins nd the time oth eyes re omplete nd the s re displyed. This is illustrted in figure 7. Frme Buffer = left eye = right eye d e f = vertil retre Figure 7: PPHIGS timing for stereo pir with miniml lteny. Render stge to opy stge dependenies re not shown for lrity. 5.3 Slts pipeline Slts hieves its gurnteed lteny y insisting tht ll the work for one field e finished during the field immeditely efore it. Sine it is uilt with lteny sensitive HMD pplitions in mind, it lwys genertes stereo s. The pipelining in Slts is t polygon level. As soon s set of polygons re trnsformed (in lumps of 30 for ring network effiieny), they re sent to the renderers. Eh renderer hndles four sreen regions so the entire sreen for oth eyes n e overed y the ville renderers. GP Renderer Frme Buffer = oth eyes = vertil retre Figure 8: Slts timing for stereo pir.,, nd re the trnsform, render, nd opy stges respetively. Stge strts fter the first th of tringles re trnsformed in. The first hlf of must finish efore the vertil retre. Sine field is two regions high, the opy stge hppens in two prts. The opy of the seond hlf of the sreen, whih only tkes 3.9 ms, doesn t our until fter the field is lredy eing displyed. The opying of the first hlf of the sreen must e done efore the vertil retre sine those pixels re immeditely displyed. This is illustrted in figure 8. In mny wys, Slts flls short of generl grphis lirry like PPHIGS. For the ske of simpliity, it uses only single GP insted of the mny (up to 50) ville to PPHIGS. This severely limits the numer of tringles tht Slts n hndle. The use of four regions per renderer mkes polygon level pipelining esier, ut lso limits the shding model to simple Gourud olor interpoltion. All of the tringles must e trnsformed nd rendered efore the first opy egins, period of out 12.8 ms. If there re too mny primitives to mke this dedline, Slts fils to generte orret. In the urrent implementtion, this trnsltes to out 100 tringles (or 12,000 tringles per seond). Even if we optimized the ode nd PPHIGS hieved out ftor of three performne inrese fter the tringle ode ws optimized to fit in the GP instrution he the ommunition ndwidth out of one GP nd the speed of the renderers limits the mximum performne to out 250 tringles. We estimte tht using multiple GPs nd more renderers we might e le to push this to few thousnd, ut urrently don t hve plns to follow this pth. These limittions re not flws, Slts exels t wht it is uilt for: experiments requiring low lteny, fixed lteny, or oth. Azum s work on preditive trking [1] used Slts for just this reson. Beuse it onsiders oth eyes simultneously, it n shre more of the work thn PPHIGS, whih hndles them sequentilly ut grouped. In ft, oth eyes n e opied t the sme time. Beuse it only renders the lines of the visile in eh field the even lines re rendered while the odd field is visile, nd the odd lines re rendered while the even lines re visile it hs hlf the rendering nd hlf the opying. As omprison of the performne of oth, figure 9 shows the pixel error for the setup used in our video. There is 33 ms of lteny for the optil eiling trker[2], mking totl of 90 ms for PPHIGS nd 50 ms for Slts. Other trkers my hve lower lteny, ut this will only inrese the importne of genertion lteny sine the error is liner with respet to lteny[1]. The error ws lulted off-line with ptured trker dt from typil demo with nive user under the optil eiling trker. The pixel error shown is omputed y tking point in the enter of the field of view for eh frme nd

6 Error in pixels ms 50 ms Error in pixels Figure 9: Pixels of error etween pixel t the enter of the sreen nd the lotion where it should hve een displyed y the time the frme ws visile. For 90 ms, orresponding to PPHIGS 33 ms trker lteny, nd 50 ms, orresponding to Slts 33 ms trker lteny. determining how fr from the enter it would e when the frme is displyed. 6 CONCLUSION We hve presented two methods for reduing genertion lteny. Both, neessrily, t ost in polygon performne. As HMD pplitions eome more prevlent, people will require miniml lteny, muh s they do high polygon rendering performne tody. 7 ACKNOWLEDGMENTS We would like to give speil thnks to Ron Azum. Ron is responsile figure 9, nd ws huge help in the retion of the Slts video. We would lso like to thnk Tony Apod nd Pixr for ess to RenderMn, whih ws used to rete the old well JITP simultion. This projet ws funded in prt y the Ntionl Siene Foundtion, NSF Grnt Numer MIP nd NSF Coopertive Agreement Numer ASC , nd y the Advned Reserh Projets Ageny, ARPA ISTO Order Numer A410 nd ARPA Contrt DABT63-93-C-C REFERENCES 1. Azum, Ronld nd Gry Bishop. Improving Stti nd Dynmi Registrtion in n Optil See-through HMD. Proeedings of SIGGRAPH 94 (Orlndo, Florid, July 24 29, 1994). In Computer Grphis Proeedings, Annul Conferene Series, ACM SIGGRAPH, New York, 1994, pp Azum, Ronld. Preditive Trking for Augmented Relity. UNC Chpel Hill Deprtment of Computer Siene PhD Disserttion, Bjur, Mihel, Henry Fuhs nd Ryutrou Ohuhi. Merging Virtul Ojets with the Rel World: Seeing Ultrsound Imgery within the Ptient. Proeedings of SIGGRAPH 92 (Chigo, Illinois, July 26-31, 1992). In Computer Grphis, 26, 2 (July 1992), ACM SIGGRAPH, New York, 1992, pp Bishop, Gry, Henry Fuhs, Leonrd MMilln nd Ellen Sher Zgier. Frmeless Rendering: Doule Buffering Considered Hrmful. Proeedings of SIGGRAPH 94 (Orlndo, Florid, July 24 29, 1994). In Computer Grphis Proeedings, Annul Conferene Series, ACM SIGGRAPH, New York, 1994, pp Bregli, Denis, Mihel Spooner nd Dn Lo. Helmet Mounted Lser Projetor. Proeedings of the Imge Genertion/Disply Conferene II (Sottsdle, Arizon June 10 12, 1981). pp Buridge, Dik, Pul Murry. Hrdwre Improvements To The Helmet Mounted Projetor On the Visul Disply Reserh Tool (VDRT) At The Nvl Trining Systems Center. Proeedings of the SPIE onferene on Hed-Mounted Displys, Cohen, Jon nd Mr Olno. Low Lteny Rendering on Pixel-Plnes 5. UNC Chpel Hill Deprtment of Computer Siene tehnil report TR94-028, Dvid Ellsworth. Pixel-Plnes 5 Rendering Control. UNC Chpel Hill Deprtment of Computer Siene Softwre Doumenttion, Feiner, Steven, Blir MIntyre nd Dorée Seligmnn. Knowledge-sed Augmented Relity. Communitions of the ACM, 36, 7, July 1993, pp Friedmnn, Mrtin, Thd Strner nd Alex Pentlnd. Devie Synhroniztion Using n Optiml Filter. Proeedings of 1992 Symposium on Intertive 3d Grphis (Cmridge, Msshusetts, Mrh 29 April 1, 1992). Speil issue of Computer Grphis, ACM SIGGRAPH, New York, 1992 pp Fuhs, Henry, John Poulton, John Eyles, et l. Pixel- Plnes 5: A Heterogeneous Multiproessor Grphis System Using Proessor-Enhned Memories. Proeedings of SIGGRAPH 89 (Boston, MA, July 31 August 4, 1989). In Computer Grphis, 23, 3 (July 1989), ACM SIGGRAPH, New York, 1989, pp List, Uwe Nonliner Predition of Hed Movements for Helmet-Mounted Displys. Tehnil Pper AFHRL-TP-83-45, Deemer Mine, Mrk. Chrteriztion of End-to-End Delys in Hed-Mounted Disply Systems. UNC Chpel Hill Deprtment of Computer Siene tehnil report TR93-001, Mine, Mrk nd Gry Bishop. Just-In-Time Pixels. UNC Chpel Hill Deprtment of Computer Siene tehnil report TR93-005, Regn, Mtthew nd Ronld Pose. Priority Rendering with Virtul Relity Address Relultion Pipeline. Proeedings of SIGGRAPH 94 (Orlndo, Florid, July 24 29, 1994). In Computer Grphis Proeedings, Annul Conferene Series, ACM SIGGRAPH, New York, 1994, pp Wrd, Mrk, Ronld Azum, Roert Bennett, Stefn Gottshlk nd Henry Fuhs. A Demonstrted Optil Trker with Slle Work Are for Hed Mounted Disply Systems. Proeedings of 1992 Symposium on Intertive 3d Grphis (Cmridge, Msshusetts, Mrh 29 April 1, 1992). Speil issue of Computer Grphis, ACM SIGGRAPH, New York, 1992, pp

7 Plte 1: Imge from n nimtion of UNC s Old Well, moving from left to right, rendered with onventionl methods. On sn line disply, this ppers slnted. Plte 2: Imge from n nimtion of UNC s Old Well, moving from left to right, rendered with the just-in-time pixels method. On sn line disply, this ppers to e stright.