JOURNAL OF COMPUTERS, VOL. 3, NO. 1, JANUARY 2008 7 Comler ak ed desg for traslatg multrado desrtos to oeratg system-less asyhroous roessor dataaths Daraya Guha Cetre for Hgh Performae Emedded Systems, Nayag Tehologal Uversty, Sgaore Emal: guha@tu.edu.sg Thamlla Srkatha Cetre for Hgh Performae Emedded Systems, Nayag Tehologal Uversty, Sgaore Emal: astsrka@tu.edu.sg Astrat Most asyhroous roessor Istruto Set Arhtetures ISA) are ased o a sgle tye of uderlyg asyhroous rut desg style. Asyhroous roessor ISAs are etrely deedet o the tye of asyhroous desg style hose ad a suort a lmted set of smle alatos oly. Desg reuse s tyally dffult to realze suh ases. I ths aer, we show a ehavoral model of a redtor rut system that ofgures a alato rofle-drve asyhroous roessor ISA omrsg two asyhroous desg styles. The redtor rut system s used to traslate alato rofle ad mult-rado ode to the roessor dataath through a omler ak-ed. The target s a asyhroous roessor that does ot ru a oeratg system ad s used oth as a omlemet ad alterate to software-defed rados wth hgh degrees of desg reuse. Idex Terms asyhroous roessor, ISA, redtor rut, mult-rado, omler ak-ed, desg reuse I. INTRODUCTION Commo asyhroous roessors uderlyg desg styles lude DI uas-delay Isestve) [1], STAPL Sgle Trak Hadshake Asyhroous Pulse Log) [2], ad STFB Sgle Trak Full Buffer) [3] self-tmed rut famles. The asyhroous roessor Istruto Set Arhteture ISA) s etrely deedet o the asyhroous desg style tye ad erformae metrs lke dataath erformae ad rah redto for mult-level strutos rema dffult to e streamled for multle sets of alato rofle tasks. Ths aer targets asyhroous roessors ased o a hyrd of asyhroous desg styles DI ad STAPL), where reofgurato the rut omoets of uderlyg asyhroous desg styles a take lae ased o the alato rofle. Ths aer further desres the ehavoral model of a redtor rut system that a traslate alato rofle ad rado rotool ode to the asyhroous Based o Reofgurale Frame Parser Desg for Mult-Rado Suort o Asyhroous Mroroessor Cores, y Daraya Guha ad Thamlla Srkatha whh aeared the Proeedgs of the IEEE Iteratoal Coferee o Comutg: Theory ad Alatos, ICCTA 2007, Platum Julee of the Ida Statstal Isttute, Calutta, Ida, Marh 2007. 2007 IEEE. roessor dataath dretly through a omler ak-ed. Ths desg methodology serves two uroses: 1. hels maage omlex ode mag oto roessor dataaths the asee of a oeratg system, 2. mroves uo asyhroous desg omler ak-eds geeratg hyrd asyhroous desg style hadshakg ruts y ororatg alato rofle ad rado rotool module fte state mahe exeutos) odes a sgle framework. Ths aer s arraged as follows. Seto II dsusses the desg aroah of a hyrd asyhroous roessor that suorts mult-rados the asee of a oeratg system; Seto III dsusses the arthmet of evet sequees asyhroous dataath models, Seto IV shows the system ehavoral modelg math of the redtor rut, Seto V shows the evaluato of the redtor rut system, Seto VI shows the evaluato of the omler ak-ed desg ad Seto VII rgs aout the oluso. II. DESIGN APPROACH OF A HYBRID ASYNCHRONOUS PROCESSOR SUPPORTING MULTI-RADIOS WITHOUT AN OS Hadsets ad ortale osumer eletro deves tyally eed to suort dverse alatos ad usually rovde multle rado hoes for ommuatos. The mult-rado hoes are usually maaged y the oeratg system through drver swthg amog the dvdual rotool Network Iterfae Cards NIC). I the asee of a oeratg system ad software maagemet of rotool modules, the roessor must e ale to suort these rado rotool modules whe voked y the alato. Tll ow, the oly rados that are suorted o oeratg system-less asyhroous roessor ores are small-szed rados lke TR1000 [4,5]. It s very dffult to ma large omlex rado staks otmally oto the asyhroous ore, rado rotool desrto laguage traslato eg a maor ottleek. The asyhroous ore ad dataath s derved out of a sgle tye of asyhroous desg style ad realzg rogrammalty reofgurale style selet ased o rado-ased alato roflg s very dffult.
8 JOURNAL OF COMPUTERS, VOL. 3, NO. 1, JANUARY 2008 I the asee of a oeratg system, the maor hallege eomes to desg effet mro-arhtetures that resod to exteral terruts geerated y hgh-level desrto laguages. Alato odes ad rado desrtos are frst rofled to detfy ommo futoal modules a mro-arhteture aware maer, so that the roessor dataath a e ofgured exatly the maer as t would e f a software terrut was voked y a oeratg system. A redtor rut system forms a key omoet of ths mro-arhteture aware traslato of rotool ad hgh level desrto laguages to the asyhroous roessor dataath. Most of the urret emodmets of asyhroous roessors usg reshuffled ommuato roesses eled asyhroous ruts are ased o DI rut models [1]. A dffulty wth DI desg suortg artrary frequet hages the alato rofle s the dffulty satsfyg the requremet of orret workg of the desg resee of uouded oerator delay varaes. Esurg the roerty that the asyhroous roessg etty the roessor struto set) does ot get loked to a artular rado rotool, t s mortat to esure that there s a mehasm to swth over the rado mode after a self-ferred tme terval. For DI, f ths s to e aheved, heker ruts to detet ourree of trastos as rado rotool staks that are hadled f a oeratg system s reset) are eeded eve though they would hae ayway. Ths a e overome y STAPL ruts that elmate the akowledgemet ad data reset hases of the four-hase hadshakg rotool used DI ruts. I ths aer, a hyrd DI-STAPL desg style roessor dataath has ee osdered. The redtor rut system s also omrsed of ths hyrd desg style, ad ehavoral modelg has ee arred out to oform to the arthmet of evet sequees asyhroous dataath models. III. ARITHMETIC OF EVENT SEUENCES IN ASYNCHRONOUS DATAPATH MODELS LAD Legth Adatve Data) [4, 5] s ommoly-used arthmet modelg evet sequees asyhroous roessor dataaths ad struto set arhtetures. Ths s a artularly useful arthmet log where t-seral struto set arhtetures s tued le wth the legth adatve data words. Ths roosto s extremely useful for mlemetg desg reuse asyhroous roessor ased latforms. Curretly, LAD arthmet axoms suort oly oe task at a tme, ad do ot allow data terleavg or delmter re-assgmet. A few extesos of the LAD arthmet axomat foudatos to lude smultaeous ourret mult-tasks y ollaoratvely re-assgg delmter LAD dgts A delmter t s a t uo whose ourree, all the leftmost t ostos are take to e of the value of the delmter t) has ee show [6, 7]. Earler fds [6] have suggested ehag LAD arthmet the otext of odetermst data-flow omutato for ollaoratvely assgg delmter LAD dgts. Ths s a system desg hallege eause t volves o-determst dataflow omutato ased o the fat that data arrvg should e roessed mmedately) ad omutatos wth artrato for qualty-of-serve rortes), whh makes the usual LAD arhteture emodmets related asyhroous roessor desgs dffult for arallel re-overso ad otmzed LAD arthmet maagemet. Delmters a our aywhere wth a data word. A examle of ths stuato s that a adder may eouter 0111 + 1111, where 1 s a delmter t. The result should e a sgle 1, ad ot 1111, whh s wastage of eergy for roessor dataath omresso. Ths rolem s eve more rooued for odetermst data-flow omutato mas that are a evtalty of mult-rado suort asyhroous roessors, rmarly due to artrary rado voatos ad data terleavg. Ths work s urretly eg further studed ad we reset some of our tal fds o addressg ths ssue through a system ehavor model of a redtor rut that mas alato tasks to the roessor dataath through multle rado lass lkage assoatos. IV. SYSTEM BEHAVIORAL MODELING MATH OF THE PREDICTOR CIRCUIT The system ehavoral model of the redtor rut s ased o the followg assumtos: 1. Alato tasks are radomly dstruted ad a our at ay tme ay tye of alato a request a serve at ay tme). 2. A rado s voked y a alato the alato that wats serve wll eed to e ommuated). 3. A rado a e voked y multle alatos at the same tme the same rado may e used ommuatg multle alatos wth a requested qualty of serve smultaeously). 4. Multle rados are eessarly voked y multle alatos at the same tme multle rado terfaes are reset the ommuatg deve). 5. The assoato of alato task o roesses to rado lass lkage sets s stataeous. Rado voato s doe stataeously the deve as soo as a alato arrves for ommuatos. Ths s doe loally ased o guaratee of qualty of serve ad to suort the roessg ad ommuato aaltes of that alato). 6. Alato task o roesses a shft aross rado lass lkage sets. Ths meas that the tasks a e arttoed ad loally assged to a omutg grou ad-ho ased o the ature of omutato smlarty). 7. Prorty of the task o roesses geerated y multle alatos may hage durg the roessg stage. Ths s to aommodate alatos that may eed more tme-rtal resoure-ased roessg. Followg ot 6, ths also meas that there s rovso to loally
JOURNAL OF COMPUTERS, VOL. 3, NO. 1, JANUARY 2008 9 hage the task o roess rorty assoated wth a artular rado lass lkage set. To guaratee qualty of serve, f a alato s advertsed usg a rado tye for whh the hael s degraded, the alato may start to e advertsed usg a dfferet rado o the same hysal hael ut dfferet erformae haratersts) 8. Tasks searated to sets voked uder dfferet rado lkage sets may e shfted terally wth the set grous. Ths s to esure that eve f the otrol hael odto hages, data are stll ommuated wth the qualty of serve requested aross dfferet rado meda avalale. Ths s a loal oly ad teral to the redtor rut for helg hoose the rado selet for ommuatg) 9. A alato has to e mmedately roessed ad aot e uffered Wthout a oeratg system ad wthout sgfat terrut hadlg suort, the tasks would eed to e exeuted as soo as they arrve to revet rado lok- ad degradato of qualty of serve). A. Mathematal assumtos 1. The rado lkage lass sets that serves alato task o roesses s osdered to form a dmesoal vetor sae. 2. The mathematal reresetatve values of the redtor rut system resoures are ostraed to the rage of 0, 1). Ths meas that the redtor rut system hadlg a artular tye of alato vokg oly oe rado lass lkage set s rereseted y 0,0,0 ) o load ostrat) ad the odto that t mas multle alatos vokg all the rado lass lkage sets s rereseted y 1,1,1,.), sgfyg a fully loaded ostrat. 3. The otmal ost futo for task admttae aross dfferet rado lass lkage sets sas over ths dmesoal vetor sae. 4. The dmesoal vetor a e used to desre the urret state of the redtor system, ad the orm of ths vetor forms the ass of the rakg of the redtor rut load term of rado lass lkage set assoatos. 5. For tryg to make the rolem of mult-alato task o roess alloato aross all rado lass lkage sets ad su-arttos of the vetor sae NP hard, os are desred ad raked usg the same omutato resoure requremet dmesos. Jos a e added ad sutrated from the redtor system state vetor. 6. The system ehavor of the redtor rut s modeled terms of the ut futo that s a dual varale of rado lass lkage sets ad resoure demad from a alato task o roess. B. Tehques for alato s serve request task o ft aross the rado lass lkage sets Frst Ft: Partto the vetor of rado lass lkage sets equally amog all the alato tasks that requests serve Best Ft: Partto the vetor of rado lass lkage sets otmally ad shft task o roesses to that rado lass lkage set whh most losely fulfls the alato s serve request wth the least umer of os volved. Worst Ft: Partto the vetor of rado lass lkage sets ad shft task o roesses to that rado lass lkage set whh most losely fulfls the alato s serve request wth the maxmum umer of os volved. C. Predtor rut system load estmato terms of alato os roessed through rado lass lkage set assoatos) ad roess o shft The load dstruto mehasm the redtor rut system reeves gloal state udates terms of alatos roessed ad rados voked) artrarly tme. Ths resets a losure rolem of the redto of the degree of assoatg a task to rado lass lkage sets as o mmedate feedak to a o laemet s rovded. Therefore the redto susequet to the frst eah tme frame are made wth data that s kow to e out of date, ad rolems smlar to those eoutered y the least loaded tal ft model may arse. Oe advatage of usg redted resoure requremets terms of roessed alato o tasks for dfferet rado lkage lass sets s that the same tral fttg tehque that s used to selet the dataath ma a e used to rovde a estmate of the system load after o shft ad laemet. Thus the result of the fttg omutato s used to dretly udate the load value stored for the dataath ma. Charatersts of the arttoed vetor lass model the system ehavor desrto of the redtor rut a e modeled uder ths assumto. A key ssue of the redtor rut system load estmato s the atual o shft of alato task roesses order of the rado lkage lass set assoato rorty volvg hage ogog task o exeuto rorty status. The other ssue s the resoluto of rorty oflts for tme-rtal ad o-tme rtal tasks where the level of rorty s suet to hage. Mathematally, ths meas that the vetor sae arttog wll hage ad dmesoal reresetato mght ot rema the same. D. Mathematal modelg of the redtor system target mahe omlato ut) Let = The alato task orresodg to rado lkage lass set ; roess orresodg to lass. E = Set of elgle roesses,.e. the umer of tasks that a e shfted amog rado lkage lass sets = Exeted roess lfetme,.e. the tme tll a roess fshes exeutg C = Cost of shftg the roess amog rado lkage lass sets k = Normalzato ostat. e E
10 JOURNAL OF COMPUTERS, VOL. 3, NO. 1, JANUARY 2008 The,,. *... 1) Let E ff, = Codto of redtor system at tme t t ) k, terms of the rado lass lkage ouay vetors., for = Jo the redtor system at tme t t ) shftg the task roess amog rado lkage lasses. Fudametally, the otmzato rolem for omutg resoure alloato the redtor system for mag oto the roessor dataath would e: m ).. 2),.e. mmze the orm aross all the trasto roess vetors. The geeral seleto oly for o shftg aross the rado lkage lasses s thus gve y: max * ) e C. 3) There exsts a lesser tha the least feasle soluto terms of geeral dsrete GD) ost futos to a lass arttog rolem f the otmal ost threshold value s at least.durg the shft of alato tasks C MAX / stataeously over the rado lkage lasses, there remas to e show ad rove that there exst a feasle ost that s always attaed whh s lesser tha a least threshold value. Ths meas that the ost of trasferrg alato tasks of a artular rado lkage lass terally durg rado rotool laguage mag to the roessor dataath does ot overall redue the ealty for ufferg t wthout roessg mmedately. I order that the redtor system modelg for dvdual rado lkage lasses o a oe-to-oe ass holds, we shall show that the ost of trasferrg alato tasks aross dfferet rado lkage lasses suets the ehavoral modelg to e the same as whe the tasks are roessed dvdually y the redtor o a er-rado lkage ass. Let vetor sae x e a otmal artto aross E for some. The hysal terretato of ths s that the rado lkage lasses form a ovex futo set ad has a otmal artto aross the set of vetor saes saed y the rado lkage lasses. The dfferee s that these lasses are set statally. The ssue of NP-hardess omes to osderato f the sae saed y ad ehaves quas-statally at every tme whe there s a trasto amog the s. The questos to e aswered ths rolem formulato are thus: 1. Does there exst a otmal artto of a quasstat vetor sae V 1 for a elemet trasto defed from a ovex futo set F1 to aother ovex futo set F2 defed aother quas-stat vetor sae, V 2? 2. Is ths rolem NP-hard? Let us frst fous o the defto of the ovex futos aross quas-stat vetor saes. A futo s alled strogly ovex f for three ots x 1 ad x 3, the followg result holds: x2 x1 ) * x1) x3 x2 ) * x3) x2 ) x1 x2 x3 x3 x1.. 4) From 1), Let us ostrut a set P } { over some, the exeted roess lfetme. For a ouded, P wll e a ouded set MAX wth some least uer oud P u. 5) For the sets P ad P, thus we a hoose the trlets 1 2 3 P P P x1 P P P x, x, x ) ad ) suh that equato 4) holds for eah of these trads o ther resetve doma sets.. 6) A futo F P ) a thus e defed to e ovex for P P P the trlets x1 ) aross P.Smlarly, a futo F P ) a e defed to e ovex for the trlets P P P x1 ) aross P 7) From the defto of equato 2), the vetor sae V ) sag aross the set a e defed as P P P the ass sa of x 1 ), essetally otg the fat that the odto of the redtor system at tme t t ) where t MAX ) s a ermutato of the elemets of P. From 5), eause of the oudedess of P, there exsts a futo F P ) whh a e C ermuted exatly to a ovex futo F suh that C P P P C F { x1 } ad V ) V F ). 8) C Smlarly from 8), V ) V F ) 9) From the defto of vetor saes,
JOURNAL OF COMPUTERS, VOL. 3, NO. 1, JANUARY 2008 11 V ) V ) V ) 10) V From the defto of equato 2), a sae ) sag aross the set a e thus defed o P P P 1 2 3 the ass sa of x, x, x ), essetally otg that o of shftg a roess the set E aross the sae V ) for defed s a futo F P ) that mas the { } s to some elemets { } s for t [ t, t ] defed ad ouded y MAX.. 11) As { } are ermutato elemets of s, the result C from 11) ad 8) s that F P ) mas to..12) V ) s the a susae of V ) From equato 10) t thus follows that. 14) F..13) ad from 10) aga, m, ).. 14) Ths shows that a otmal artto of the quasstat vetor saes saed y the rado lkage lasses does exst. 15).E.D. We ow show that deed ths s a NP-hard rolem. Let us defe as the otmal artto of the quasstat vetor saes saed y the rado lkage lasses, ased o equatos 1) through 15). From 11) ad 13), t s easy to see that m ). 16) Let us defe a Vetor Sae V ow sag over ), draw from the elemets V ) ad V ) 17) From 13), we a say that, where s a salar. 18) We defe as { q }, where q 1, f mas exatly to, whh meas, there s erfet -order roessg of dvdual alato tasks wth all the dfferet rado lkage lasses. q 0, otherwse. 19) We a thus solve a suset rolem of set, derved fromv, over the elemets { q } ad a oud, suh q..20) that From the ost futo defto 1) ad the oly of o shftg aross dfferet rado lkage lasses equato 3)), we a say from Equatos 16), 17), 18) ad 19) as the fat that exst ff for whh q. 21) The hysal terretato of ths statemet s that the seleto oly hooses the alato tasks that artular rado lkage lass to shft for whh there s a maxmum the ost futo for shftg,.e. the math etwee the umer of tasks to shft mathes resely the umer of tasks assoated that rado lkage lass whh are eg maed oto the dataath y the redtor system at that stat of tme. Ths roves that the rolem for quas-stat vetor saes saed y the mult-rado lkage lasses s NP-omlete..E.D) Susequet to ths, we shall ow rove a more owerful relato reset to determg the otmal rado lkage assoato ost futo whe mag multle alato tasks to the asyhroous roessor dataath through dvdual rado lkage assoatos. Suose the set has elemets all. Let us also assume that the total set of all ossle rado lkage lasses s K. Ths s to geeralze the rolem more, the ratal osequee of whh s that more rados a e suorted the same latform through frmware ugrades, athes ad etrely ew software sutes for ewer rado stadards that may ome the future. Let the ost futo, as defed from Equato 3), o e " C. 22) The, the ost er eah rado lkage lass " C C K..23) / I the usual ase, the ost futo for alato task admttae to a artular rado lkage lass determes whh rado s voked for whh alato) s deoted yc. At ay ot of tme t MAX, there exsts some { q } for whh equato 21) s satsfed. Ths aks o the assumto that the roessor s ot dle as the tasks are maed oto the dataath through dvdual rado lkage lasses. Mathematally, ths s equvalet to the fat that there s a otmal kerel ma amog quasstat vetor saes. For K, thus there exsts a mmal threshold for eah arttoed rado lkage su-lass, the value of whh s gve y Equato 14). We wll rove ow that C 1 ) C, whh sgfes that task admttae ost futo for a set of multle alatos voked aross dfferet tyes of rado lkage lass assoatos overges towards the task admttae ost futo where a artular rado lkage lass s assoated wth a artular alato.
12 JOURNAL OF COMPUTERS, VOL. 3, NO. 1, JANUARY 2008 There s thus the rovso of hadlg alato task roesses eve f the reedee level hages the order of rado lass lkage set elemets urretly eg exeuted, as well as ay ew alato tasks that are geerated whh vokes a dfferet rado, reatg a ew assoato to a rado lkage lass. From the roertes of ay set, we have " C K..23) Thus, " " C K) C K K. 24) From equato 20), ovously, ay vetor susae, e e e K KC e, as for Parttog to susaes always reases the overall orm of the susae [8]. Thus from equato 24), we havec, whh redues toc C C, where s a salar exressg the ost futo of task admsso to all the rado lass lkage sets. Thus, we have, C 1 ).. 25) C Now, aga, we a wrte from Equato 23) " C K, whh redues to C / K)..26) Se K KC, 1/ K 1/ KC / K / KC. 27) Hee from 26), we have C 1 1/ KC ) C 1 ) C 28) From Equatos 25) ad 28) t s ow easy to see that C has a ostve overgee towardsc. From Equatos 25), 26) ad 28), the overgee s of the order of log, y exadg omally equatos 25) ad 28). Ths shows a mortat result. Whle the dyam shftg of alato task roesses aross the arttoed rado lkage lasses hels redug the task admttae ost futo, ths s deedet of the overall redtor system ehavor, ad a e roessed statly wthout ufferg, savg o ostly hardware ad atter mathg ruts. The mathematal aalyss also shows that the redtor system rut a work well uder odtos where multle alatos voke multle rados, the hysal terretato of whh s - multle alatos a e suorted o the asyhroous roessor latforms reasg the utlty of suh deves whle extedg attery lfe redug dyam ower osumto y dog away wth the OS) The mathematal aalyss forms the ass of desgg the redtor rut system ad s the key strategy ehd the hadshakg rut desg terms of tme ad ourrey. The system s ehavorally modeled osderg deedet rado ehavor ad lkage lass geerato ased o er-alato task set tye, whereas the oerato of the rut volves teral task roess trastos aross rado lkage lasses. Ths ehavor s the modeled to elmate omutatos wth artrato ad hage reedee levels of the exeuted os. The rolem of detetg whe a artular evet ours a alato task s arttoed ad lked to a artular rado lkage lass) s redued to a trval ase of aommodatg t ay artular lass set as ad whe t omes alog. There s o eed for seal detetor ruts for dog ths ostly oerato. V. EVALUATION OF THE PREDICTOR CIRCUIT SYSTEM Based o the exteded LAD mathemats [6, 7], a termedate rotool desrto ad asyhroous ehavoral model laguage s eg vestgated urretly outo wth the redtor rut system desg. We deded to use the Balsa [14] asyhroous sythess laguage ad omler terfae eause t s oe-soure ad easly ustomzale. Balsa geerates DI hadshake rut models for sythess, ad we frst looked at how to exted the omler ak-ed to geerate hadshakg models other tha DI ased o our exteded LAD arthmet. A. Modelg usg the Balsa toolset Based o the ehavoral modelg of the redtor system, t eomes ossle to exted a oet of rogrammale state level eodg delay-sestve eles where rorty of the os eg exeuted a hage at radom. Ths a hel keeg data ad delmter formato deedet of oe aother, ad hel strter otmal mas of rado-lked alato rofles to the roessor dataath. The redtor system also eeds to e modeled wth the trae results otaed durg the ode geerato ad otmzato hase assoated wth alato roflg ad omler ak-ed ostruto. Ivestgatos related to ths hase of work are urretly eg arred out. Ths desg methodology a hel to overome a dffulty asyhroous roessor ased latform desg that relates to kowg whe a sef omutato s doe for o-determst evet geerato. The rofle-derved ode has to e tegrated wth the modfed Balsa omler framework for ths urose. The ehavoral model of the redtor system s desred the Balsa laguage the modfed omler framework. The ehavoral model a e sytheszed y usual Balsa API tegrato wth stadard EDA sythess tools lke Iarus Verlog [10]. Balsa has dedated APIs for tegrato wth ths tool. The smulato was ru to omleto for our desged redtor system model ad the tme of smulato oted
JOURNAL OF COMPUTERS, VOL. 3, NO. 1, JANUARY 2008 13 tll the etre hadshakg rut grah was traversed. The left sde of the wdow ae shows the values held the uffer orts, uffer les ad ommuatg rals. The red le orresods to the hadshakg rut geerated the frst target mahe geerato. We reset the uffer smulato traes for a 2-DI rut model for hadshakg tmes of 50 µs ad 500 µs. Fg. 1. Hadshakg smulato traes of the redtor system terms of DI ad STAPL rut models 2 DI uffer, hadshakg tme 50 mroseods. Red surge uffer 2 shows the rado task alloato set almost exlusvely oued for oe tye of alato task. values of hadshakg tmes for two dual-ral ofguratos wth 8 DI uffers oeted a arallel toology wth STAPL rals. The tests were desged to kee md o-determst terrut geeratos. As rados ad alato rofles a hage retty artrarly the real-world for ortale deves, a artular rut that s already roessg formato wll eed to e terruted ad ofgured aga for roessg a dfferet set of formato. As there s o oeratg system to hael these frequet sets of terruts, the asyhroous dataath must wholly hael ths to the roessor ore. Wth reset to that ew terrut, a hadshakg rut ofgurato must ow e geerated as the target of omutato. The hadshakg rut ofgurato that s urretly reset eeds to e ated uo ased o the terrut rorty. Potetal deadloks may arse owg to artrato oteto. Wth reset to ths tme of ew target geerato, hadshakg tmes are egatve for the targets already geerated the roess. Ths also hels verfato of the redtor rut system to hael the formato rofles to the dataath removg artrato ssues. Fg.3 shows the FSM trasto of the hadshakg task grah for 8 dual-ral DI uffers wth arallel deomostos for a hadshakg tme of 100 s. Fg. 2. Hadshakg smulato traes of the redtor system terms of DI ad STAPL rut models 2 DI uffer, hadshakg tme 500 mroseods. Dual red surges uffer 2 ad the ma module shows the rado task alloato sets eg swaed for a gve alato rofle, meag that alato task os are shfted terally wth the sets. VI. EVALUATION OF THE COMPILER BACK-END DESIGN Balsa also rovdes a set of wraers that a e used to ma the target mahes geerated dretly oto ovetoal EDA sythess tools. Ths s leveraged uo to geerate the target mahes after the rado rotool desrto otmzato ad modfed LAD arthmet mlemetato the Balsa omler framework s tegrated ad ult to oe. A umer of smulatos were arred out usg the modfed LAD arthmet ased omler ak-ed usg Balsa. The exermet was arred out y settg dfferet Fg.3. FSM trasto of the hadshakg task grah for 8 DI uffer wth STAPL rals wth hadshakg tme = 100 s Ths meas that the extato to the seod target mahe geerato ofgurato was doe 100 s after the frst target mahe was geerated) VII. CONCLUSION Hgh-level desrto laguage omlato o target mahes volvg hyrd asyhroous roessor ISA s oe of the reet hoteds of fous hadset vedors owg to sueror attery lfe rosets. Ths aer shows a desg methodology wherey hgh-level desrto laguages eomassg rut ad system-level ehavor as well as rado fte state mahe FSM) exeutos a e fed to geerate hyrd asyhroous desg style hadshakg ruts. Ths aer dsusses relmary fds otaed durg the frst rototyg of hyrd asyhroous roessor-ased latforms targeted at hadset IPs. There s a oe ssue aout the target mahes hagg durg omlato the ofgurato of DI
14 JOURNAL OF COMPUTERS, VOL. 3, NO. 1, JANUARY 2008 ad STAPL ruts hagg ased o the tye ad umer of alatos hadled y the latform). Traslato of suh evets lke hardware terruts suh ases would eed to e traslated to the asyhroous roessor ISA a re-redted maer. Dyam target hagg the otext of a sgle rutme rogram arttog for asyhroous targets s oe aset that eeds to e ehaed the futoal exteslty of the omler ak-ed framework. The other oe ssue s the oteto dataath tmg losures due to hadshakg whe a large umer of alato tasks are volved. Ths may e dffult to mlemet usg the urret DI rut model theory that Balsa uses. Further vestgatos are eg urretly arred out ths area. REFERENCES [1] M. Reaud, P. Vvet ad F. Ro, ASPRO-16: A stadard ell DI 16-t RISC asyhroous mroroessor, ro. of 4 th. Iteratoal Symosum o Asyhroous Cruts ad Systems, ASYNC 1998, Arl 1998,. 22-32 [2] Mka Nystrom ad Ala J. Mart, Method ad aaratus for a asyhroous ulse log rut, Uted States Patet Alato # 20050007151, Ser. # 10/693543, Ja 13, 2005 [3] USC Asyhroous VLSI Desg, htt://ugfrau.us.edu/ew/researh/urret/asy1 [4] Rat Maohar, Wdth-adatve data word arhtetures, ro. of Advaed Researh VLSI 2001, ARVLSI 2001, 14-16 Marh 2001,. 112-129 [5] V. Ekaayake, Clto Kelly IV ad R. Maohar, A ultra-low ower roessor for sesor etworks, ro. of 11 th. Iteratoal Coferee o Arhtetural Suort for Programmg Laguages ad Oeratg Systems, ACM ASPLOS, Arl 2004,.27-36 [6] Daraya Guha ad Thamlla Srkatha, Reofgurale Frame Parser Desg for Mult-Rado Suort o Asyhroous Mroroessor Cores, ro. of IEEE Iteratoal Coferee o Comutg: Theory ad Alatos, ICCTA 2007, Platum Julee of the Ida Statstal Isttute, Marh 2007,. 122-127 [7] Daraya Guha ad Thamlla Srkatha, Mult- Rado suort o Asyhroous Proessor Cores: A Desg Methodology aroah for Cogtve Rados, ro. of IEEE Iteratoal Coferee o Portale Iformato Deves, PORTABLE 2007, May 2007,. 1-4 [8] Dael Smso, Lear reresetatos of artally ordered sets ad vetor sae ategores, Gordo ad Breah See Pulshers, 1992 [9] The Balsa Asyhroous Sythess System, Uversty of Mahester Advaed Proessor Tehologes Grou, htt://traet.s.ma.a.uk/at/roets/tools/alsa [10] Iarus Verlog, htt://www.arus.om/eda/verlog Daraya Guha reeved hs B.Eg. Frst Class Hoors) degree Eletros ad Teleommuatos Egeerg from Jadavur Uversty, Calutta, Ida, 2000 ad hs M.Eg degree Eletral Egeerg from Corell Uversty, Ithaa, New York, 2004. He urretly holds a teured osto of Researh Assoate wth the Ceter for Hgh Performae Emedded Systems, Nayag Tehologal Uversty, Sgaore. Pror to ths aotmet, he had worked wth Saske, Aglet Tehologes ad Samsug. He has ulshed more tha 15 aers the areas of emedded systems desg methodologes ad data ommuatos. He s also a atve otrutor to teret egeerg stadards ad has o-authored fve teret drafts o ath omutato that were soured to eome RFCs the IETF PCE WG. Mr. Guha wo the Brtsh Cheveg Sholarsh 2003 ad a NSF studet aer award 2004. Dr. Thamlla Srkatha reeved hs B.S. degree Hoors) Comuter ad Cotrol Systems ad hs Ph.D. System Modelg ad Iformato Systems Egeerg from Covetry Uversty, Uted Kgdom, 1980 ad 1986 resetvely. He s the Dretor of Ceter for Hgh Performae Emedded Systems ad teured Professor of Comuter Egeerg, Nayag Tehologal Uversty, Sgaore. He has ulshed more tha 220 aers ad holds three veto dslosures. Hs researh terests lude alato-sef arhtetures for emedded systems ad desg methodologes for hgh erformae emedded systems. Dr. Srkatha s a Cororate Memer of the Isttuto of Eletral Egeers MIEE) ad Chartered Egeer CEg) se 1990. He s also a Seor Memer of the Isttute of Eletral ad Eletros Egeers SMIEEE) se 1994. Dr. Srkatha was awarded the Pul Admstrato Medal Broze) o 2006 Sgaore Natoal Day for outstadg otrutos to eduato Sgaore.