|
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667266826692670267126722673267426752676267726782679268026812682268326842685268626872688268926902691269226932694269526962697269826992700270127022703270427052706270727082709271027112712271327142715271627172718271927202721272227232724272527262727272827292730273127322733273427352736273727382739274027412742274327442745274627472748274927502751275227532754275527562757275827592760276127622763276427652766276727682769277027712772277327742775277627772778277927802781278227832784278527862787278827892790279127922793279427952796279727982799280028012802280328042805280628072808280928102811281228132814281528162817281828192820282128222823282428252826282728282829283028312832283328342835283628372838283928402841284228432844284528462847284828492850285128522853285428552856285728582859286028612862286328642865286628672868286928702871287228732874287528762877287828792880288128822883288428852886288728882889289028912892289328942895289628972898289929002901290229032904290529062907290829092910291129122913291429152916291729182919292029212922292329242925292629272928292929302931293229332934293529362937293829392940294129422943294429452946294729482949295029512952295329542955295629572958295929602961296229632964296529662967296829692970297129722973297429752976297729782979298029812982298329842985298629872988298929902991299229932994299529962997299829993000300130023003300430053006300730083009301030113012301330143015301630173018301930203021302230233024302530263027302830293030303130323033303430353036303730383039304030413042304330443045304630473048304930503051305230533054305530563057305830593060306130623063306430653066306730683069307030713072307330743075307630773078307930803081308230833084308530863087308830893090309130923093309430953096309730983099310031013102310331043105310631073108310931103111311231133114311531163117311831193120312131223123312431253126312731283129313031313132313331343135313631373138313931403141314231433144314531463147314831493150315131523153315431553156315731583159316031613162316331643165316631673168316931703171317231733174317531763177317831793180318131823183318431853186318731883189319031913192319331943195319631973198319932003201320232033204320532063207320832093210321132123213321432153216321732183219322032213222322332243225322632273228322932303231323232333234323532363237323832393240324132423243324432453246324732483249325032513252325332543255325632573258325932603261326232633264326532663267326832693270327132723273327432753276327732783279328032813282328332843285328632873288328932903291329232933294329532963297329832993300330133023303330433053306330733083309331033113312331333143315331633173318331933203321332233233324332533263327332833293330333133323333333433353336333733383339334033413342334333443345334633473348334933503351335233533354335533563357335833593360336133623363336433653366336733683369337033713372337333743375337633773378337933803381338233833384338533863387338833893390339133923393339433953396339733983399340034013402340334043405340634073408340934103411341234133414341534163417341834193420342134223423342434253426342734283429343034313432343334343435343634373438343934403441344234433444344534463447344834493450345134523453345434553456345734583459346034613462346334643465346634673468346934703471347234733474347534763477347834793480348134823483348434853486348734883489349034913492349334943495349634973498349935003501350235033504350535063507350835093510351135123513351435153516351735183519352035213522352335243525352635273528352935303531353235333534353535363537353835393540354135423543354435453546354735483549355035513552355335543555355635573558355935603561356235633564356535663567356835693570357135723573357435753576357735783579358035813582358335843585358635873588358935903591359235933594359535963597359835993600360136023603360436053606360736083609361036113612361336143615361636173618361936203621362236233624362536263627362836293630363136323633363436353636363736383639364036413642364336443645364636473648364936503651365236533654365536563657365836593660366136623663366436653666366736683669367036713672367336743675367636773678367936803681368236833684368536863687368836893690369136923693369436953696369736983699370037013702370337043705370637073708370937103711371237133714371537163717371837193720372137223723372437253726372737283729373037313732373337343735373637373738373937403741374237433744374537463747374837493750375137523753375437553756375737583759376037613762376337643765376637673768376937703771377237733774377537763777377837793780378137823783378437853786378737883789379037913792379337943795379637973798379938003801380238033804380538063807380838093810381138123813381438153816381738183819382038213822382338243825382638273828382938303831383238333834383538363837383838393840384138423843384438453846384738483849385038513852385338543855385638573858385938603861386238633864386538663867386838693870387138723873387438753876387738783879388038813882388338843885388638873888388938903891389238933894389538963897389838993900390139023903390439053906390739083909391039113912391339143915391639173918391939203921392239233924392539263927392839293930393139323933393439353936393739383939394039413942394339443945394639473948394939503951395239533954395539563957395839593960396139623963396439653966396739683969397039713972397339743975397639773978397939803981398239833984398539863987398839893990399139923993399439953996399739983999400040014002400340044005400640074008400940104011401240134014401540164017401840194020402140224023402440254026402740284029403040314032403340344035403640374038403940404041404240434044404540464047404840494050405140524053405440554056405740584059406040614062406340644065406640674068406940704071407240734074407540764077407840794080408140824083408440854086408740884089409040914092409340944095409640974098409941004101410241034104410541064107410841094110411141124113411441154116411741184119412041214122412341244125412641274128412941304131413241334134413541364137413841394140414141424143414441454146414741484149415041514152415341544155415641574158415941604161416241634164416541664167416841694170417141724173417441754176417741784179418041814182418341844185418641874188418941904191419241934194419541964197419841994200420142024203420442054206420742084209421042114212421342144215421642174218421942204221422242234224422542264227422842294230423142324233423442354236423742384239424042414242424342444245424642474248424942504251425242534254425542564257425842594260426142624263426442654266426742684269427042714272427342744275427642774278427942804281428242834284428542864287428842894290429142924293429442954296429742984299430043014302430343044305430643074308430943104311431243134314431543164317431843194320432143224323432443254326432743284329433043314332433343344335433643374338433943404341434243434344434543464347434843494350435143524353435443554356435743584359436043614362436343644365436643674368436943704371437243734374437543764377437843794380438143824383438443854386438743884389439043914392439343944395439643974398439944004401440244034404440544064407440844094410441144124413441444154416441744184419442044214422442344244425442644274428442944304431443244334434443544364437443844394440444144424443444444454446444744484449445044514452445344544455445644574458445944604461446244634464446544664467446844694470447144724473447444754476447744784479448044814482448344844485448644874488448944904491449244934494449544964497449844994500450145024503450445054506450745084509451045114512451345144515451645174518451945204521452245234524452545264527452845294530453145324533453445354536453745384539454045414542454345444545454645474548454945504551455245534554455545564557455845594560456145624563456445654566456745684569457045714572457345744575457645774578457945804581458245834584458545864587458845894590459145924593459445954596459745984599460046014602460346044605460646074608460946104611461246134614461546164617461846194620462146224623462446254626462746284629463046314632463346344635463646374638463946404641464246434644464546464647464846494650465146524653465446554656465746584659466046614662466346644665466646674668466946704671467246734674467546764677467846794680468146824683468446854686468746884689469046914692469346944695469646974698469947004701470247034704470547064707470847094710471147124713471447154716471747184719472047214722472347244725472647274728472947304731473247334734473547364737473847394740474147424743474447454746474747484749475047514752475347544755475647574758475947604761476247634764476547664767476847694770477147724773477447754776477747784779478047814782478347844785478647874788478947904791479247934794479547964797479847994800480148024803480448054806480748084809481048114812481348144815481648174818481948204821482248234824482548264827482848294830483148324833483448354836483748384839484048414842484348444845484648474848484948504851485248534854485548564857485848594860486148624863486448654866486748684869487048714872487348744875487648774878487948804881488248834884488548864887488848894890489148924893489448954896489748984899490049014902490349044905490649074908490949104911491249134914491549164917491849194920492149224923492449254926492749284929493049314932493349344935493649374938493949404941494249434944494549464947494849494950495149524953495449554956495749584959496049614962496349644965496649674968496949704971497249734974497549764977497849794980498149824983498449854986498749884989499049914992499349944995499649974998499950005001500250035004500550065007500850095010501150125013501450155016501750185019502050215022502350245025502650275028502950305031503250335034503550365037503850395040504150425043504450455046504750485049505050515052505350545055505650575058505950605061506250635064506550665067506850695070507150725073507450755076507750785079508050815082508350845085508650875088508950905091509250935094509550965097509850995100510151025103510451055106510751085109511051115112511351145115511651175118511951205121512251235124512551265127512851295130513151325133513451355136513751385139514051415142514351445145514651475148514951505151515251535154515551565157515851595160516151625163516451655166516751685169517051715172517351745175517651775178517951805181518251835184518551865187518851895190519151925193519451955196519751985199520052015202520352045205520652075208520952105211521252135214521552165217521852195220522152225223522452255226522752285229523052315232523352345235523652375238523952405241524252435244524552465247524852495250525152525253525452555256525752585259526052615262526352645265526652675268526952705271527252735274527552765277527852795280528152825283528452855286528752885289529052915292529352945295529652975298529953005301530253035304530553065307530853095310531153125313531453155316531753185319532053215322532353245325532653275328532953305331533253335334533553365337533853395340534153425343534453455346534753485349535053515352535353545355535653575358535953605361536253635364536553665367536853695370537153725373537453755376537753785379538053815382538353845385538653875388538953905391539253935394539553965397539853995400540154025403540454055406540754085409541054115412541354145415541654175418541954205421542254235424542554265427542854295430543154325433543454355436543754385439544054415442544354445445544654475448544954505451545254535454545554565457545854595460546154625463546454655466546754685469547054715472547354745475547654775478547954805481548254835484548554865487548854895490549154925493549454955496549754985499550055015502550355045505550655075508550955105511551255135514551555165517551855195520552155225523552455255526552755285529553055315532553355345535553655375538553955405541554255435544554555465547554855495550555155525553555455555556555755585559556055615562556355645565556655675568556955705571557255735574557555765577557855795580558155825583558455855586558755885589559055915592559355945595559655975598559956005601560256035604560556065607560856095610561156125613561456155616561756185619562056215622562356245625562656275628562956305631563256335634563556365637563856395640564156425643564456455646564756485649565056515652565356545655565656575658565956605661566256635664566556665667566856695670567156725673567456755676567756785679568056815682568356845685568656875688568956905691569256935694569556965697569856995700570157025703570457055706570757085709571057115712571357145715571657175718571957205721572257235724572557265727572857295730573157325733573457355736573757385739574057415742574357445745574657475748574957505751575257535754575557565757575857595760576157625763576457655766576757685769577057715772577357745775577657775778577957805781578257835784578557865787578857895790579157925793579457955796579757985799580058015802580358045805580658075808580958105811581258135814581558165817581858195820582158225823582458255826582758285829583058315832583358345835583658375838583958405841584258435844584558465847584858495850585158525853585458555856585758585859586058615862586358645865586658675868586958705871587258735874587558765877587858795880588158825883588458855886588758885889589058915892589358945895589658975898589959005901590259035904590559065907590859095910591159125913591459155916591759185919592059215922592359245925592659275928592959305931593259335934593559365937593859395940594159425943594459455946594759485949595059515952595359545955595659575958595959605961596259635964596559665967596859695970597159725973597459755976597759785979598059815982598359845985598659875988598959905991599259935994599559965997599859996000600160026003600460056006600760086009601060116012601360146015601660176018601960206021602260236024602560266027602860296030603160326033603460356036603760386039604060416042604360446045604660476048604960506051605260536054605560566057605860596060606160626063606460656066606760686069607060716072607360746075607660776078607960806081608260836084608560866087608860896090609160926093609460956096609760986099610061016102610361046105610661076108610961106111611261136114611561166117611861196120612161226123612461256126612761286129613061316132613361346135613661376138613961406141614261436144614561466147614861496150615161526153615461556156615761586159616061616162616361646165616661676168616961706171617261736174617561766177617861796180618161826183618461856186618761886189619061916192619361946195619661976198619962006201620262036204620562066207620862096210621162126213621462156216621762186219622062216222622362246225622662276228622962306231623262336234623562366237623862396240624162426243624462456246624762486249625062516252625362546255625662576258625962606261626262636264626562666267626862696270627162726273627462756276627762786279628062816282628362846285628662876288628962906291629262936294629562966297629862996300630163026303630463056306630763086309631063116312631363146315631663176318631963206321632263236324632563266327632863296330633163326333633463356336633763386339634063416342634363446345634663476348634963506351635263536354635563566357635863596360636163626363636463656366636763686369637063716372637363746375637663776378637963806381638263836384638563866387638863896390639163926393639463956396639763986399640064016402640364046405640664076408640964106411641264136414641564166417641864196420642164226423642464256426642764286429643064316432643364346435643664376438643964406441644264436444644564466447644864496450645164526453645464556456645764586459646064616462646364646465646664676468646964706471647264736474647564766477647864796480648164826483648464856486648764886489649064916492649364946495649664976498649965006501650265036504650565066507650865096510651165126513651465156516651765186519652065216522652365246525652665276528652965306531653265336534653565366537653865396540654165426543654465456546654765486549655065516552655365546555655665576558655965606561656265636564656565666567656865696570657165726573657465756576657765786579658065816582658365846585658665876588658965906591659265936594659565966597659865996600660166026603660466056606660766086609661066116612661366146615661666176618661966206621662266236624662566266627662866296630663166326633663466356636663766386639664066416642664366446645664666476648664966506651665266536654665566566657665866596660666166626663666466656666666766686669667066716672667366746675667666776678667966806681668266836684668566866687668866896690669166926693669466956696669766986699670067016702670367046705670667076708670967106711671267136714671567166717671867196720672167226723672467256726672767286729673067316732673367346735673667376738673967406741674267436744674567466747674867496750675167526753675467556756675767586759676067616762676367646765676667676768676967706771677267736774677567766777677867796780678167826783678467856786678767886789679067916792679367946795679667976798679968006801680268036804680568066807680868096810681168126813681468156816681768186819682068216822682368246825682668276828682968306831683268336834683568366837683868396840684168426843684468456846684768486849685068516852685368546855685668576858685968606861686268636864686568666867686868696870687168726873687468756876687768786879688068816882688368846885688668876888688968906891689268936894689568966897689868996900690169026903690469056906690769086909691069116912691369146915691669176918691969206921692269236924692569266927692869296930693169326933693469356936693769386939694069416942694369446945694669476948694969506951695269536954695569566957695869596960696169626963696469656966696769686969697069716972697369746975697669776978697969806981698269836984698569866987698869896990699169926993699469956996699769986999700070017002700370047005700670077008700970107011701270137014701570167017701870197020702170227023702470257026702770287029703070317032703370347035703670377038703970407041704270437044704570467047704870497050705170527053705470557056705770587059706070617062706370647065706670677068706970707071707270737074707570767077707870797080708170827083708470857086708770887089709070917092709370947095709670977098709971007101710271037104710571067107710871097110711171127113711471157116711771187119712071217122712371247125712671277128712971307131713271337134713571367137713871397140714171427143714471457146714771487149715071517152715371547155715671577158715971607161716271637164716571667167716871697170717171727173717471757176717771787179718071817182718371847185718671877188718971907191719271937194719571967197719871997200720172027203720472057206720772087209721072117212721372147215721672177218721972207221722272237224722572267227722872297230723172327233723472357236723772387239724072417242724372447245724672477248724972507251725272537254725572567257725872597260726172627263726472657266726772687269727072717272727372747275727672777278727972807281728272837284728572867287728872897290729172927293729472957296729772987299730073017302730373047305730673077308730973107311731273137314731573167317731873197320732173227323732473257326732773287329733073317332733373347335733673377338733973407341734273437344734573467347734873497350735173527353735473557356735773587359736073617362736373647365736673677368736973707371737273737374737573767377737873797380738173827383738473857386738773887389739073917392739373947395739673977398739974007401740274037404740574067407740874097410741174127413741474157416741774187419742074217422742374247425742674277428742974307431743274337434743574367437743874397440744174427443744474457446744774487449745074517452745374547455745674577458745974607461746274637464746574667467746874697470747174727473747474757476747774787479748074817482748374847485748674877488748974907491749274937494749574967497749874997500750175027503750475057506750775087509751075117512751375147515751675177518751975207521752275237524752575267527752875297530753175327533753475357536753775387539754075417542754375447545754675477548754975507551755275537554755575567557755875597560756175627563756475657566756775687569757075717572757375747575757675777578757975807581758275837584758575867587758875897590759175927593759475957596759775987599760076017602760376047605760676077608760976107611761276137614761576167617761876197620762176227623762476257626762776287629763076317632763376347635763676377638763976407641764276437644764576467647764876497650765176527653765476557656765776587659766076617662766376647665766676677668766976707671767276737674767576767677767876797680768176827683768476857686768776887689769076917692769376947695769676977698769977007701770277037704770577067707770877097710771177127713771477157716771777187719772077217722772377247725772677277728772977307731773277337734773577367737773877397740774177427743774477457746774777487749775077517752775377547755775677577758775977607761776277637764776577667767776877697770777177727773777477757776777777787779778077817782778377847785778677877788778977907791779277937794779577967797 |
- #define REALNAME ASMNAME
- #define ASSEMBLER
- #include "common.h"
-
- #define FETCH ld
- #define STACKSIZE 160
- #define gsLQC1(base,fq,ft,offset) .word(0x32<<26|base<<21|ft<<16|0x1<<15|offset<<6|0x1<<5|fq)
- #define gsSQC1(base,fq,ft,offset) .word(0x3A<<26|base<<21|ft<<16|0x1<<15|offset<<6|0x1<<5|fq)
-
-
- ##### Parameter registers ####
-
- #define M $4
- #define N $5
- #define K $6
- #define A $8
- #define B $9
- #define C $10
- #define LDC $11
-
- #### Pointer A, B, C ####
- #define AO $12
- #define BO $13
-
- #define CO1 $14
- #define CO2 $15
- #define CO3 $16
- #define CO4 $17
-
- #define PREA $18
- #define PREB $19
-
- #### Used registers ####
- #define A1 $f0
- #define A2 $f1
- #define A3 $f2
- #define A4 $f3
- #define A5 $f4
- #define A6 $f5
- #define A7 $f6
- #define A8 $f7
-
- #define B1 $f8
- #define B2 $f9
- #define B3 $f10
- #define B4 $f11
- #define B5 $f12
- #define B6 $f13
- #define B7 $f14
- #define B8 $f15
-
- #define C11 $f16
- #define C12 $f17
- #define C21 $f18
- #define C22 $f19
- #define C31 $f20
- #define C32 $f21
- #define C41 $f22
- #define C42 $f23
- #define C13 $f24
- #define C14 $f25
- #define C23 $f26
- #define C24 $f27
- #define C33 $f28
- #define C34 $f29
- #define C43 $f30
- #define C44 $f31
-
- #define I $2
- #define J $3
- #define L $7
-
- #### Alpha register ####
- #define ALPHA $f15
-
- #define F31 31
- #define F30 30
- #define F29 29
- #define F28 28
- #define F27 27
- #define F26 26
- #define F25 25
- #define F24 24
- #define F23 23
- #define F22 22
- #define F21 21
- #define F20 20
- #define F19 19
- #define F18 18
- #define F17 17
- #define F16 16
- #define F15 15
- #define F14 14
- #define F13 13
- #define F12 12
- #define F11 11
- #define F10 10
- #define F9 9
- #define F8 8
- #define F7 7
- #define F6 6
- #define F5 5
- #define F4 4
- #define F3 3
- #define F2 2
- #define F1 1
- #define F0 0
-
- #define R12 12
- #define R13 13
-
- #define R14 14
- #define R15 15
- #define R16 16
- #define R17 17
-
- #if defined(TRMMKERNEL)
- #define OFFSET $23
- #define KK $24
- #define TEMP $25
- #endif
-
- # .text
- # .align 2
- ## .globl gemm
- # .set nomips16
- # .ent gemm
- # .type gemm, @function
- #gemm:
- # .frame $sp,STACKSIZE,$31 # vars= 48, regs= 1/0, args= 0, gp= 0
- # .mask 0x40000000,-8
- # .fmask 0x00000000,0
- # .set noreorder
- # .set nomacro
-
-
- PROLOGUE
-
- daddiu $sp,$sp,-STACKSIZE
-
- sd $16, 0($sp)
- sd $17, 8($sp)
- sd $18, 16($sp)
- sd $19, 24($sp)
- sd $20, 32($sp)
- sd $21, 40($sp)
- sd $22, 48($sp)
-
- sdc1 $f24, 56($sp)
- sdc1 $f25, 64($sp)
- sdc1 $f26, 72($sp)
- sdc1 $f27, 80($sp)
- sdc1 $f28, 88($sp)
-
- #if defined(TRMMKERNEL)
- sd $23, 96($sp)
- sd $24, 104($sp)
- sd $25, 112($sp)
-
- LDARG OFFSET, 160($sp)
- #endif
-
- #ifndef __64BIT__
- sdc1 $f20,120($sp)
- sdc1 $f21,128($sp)
- sdc1 $f22,136($sp)
- sdc1 $f23,144($sp)
- #endif
-
- .align 4
- .L4:
- dsra J, N, 2 # NR=4
- dsll LDC, LDC, BASE_SHIFT# LDC*SIZE
-
- #if defined(TRMMKERNEL) && !defined(LEFT)
- neg KK, OFFSET
- #endif
-
- blez J, .L2
- ST ALPHA, 152($sp)
-
- .L48:
- dsra I, M, 3 # MR=8
- dsll PREA, K, BASE_SHIFT
-
- move AO, A # Reset A
- move CO1, C
-
- daddu CO2, C, LDC
- daddu CO3, CO2, LDC
-
- daddu CO4, CO3, LDC
- daddu PREA, A, PREA
-
- #if defined(TRMMKERNEL) && defined(LEFT)
- move KK, OFFSET
- #endif
-
- blez I, .L44
- daddu C, CO4, LDC
-
- .align 4
- .L481:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) ||\
- (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, 3 + BASE_SHIFT # kk*8mr*datasize
- dsll TEMP, KK, 2 + BASE_SHIFT
-
- daddu AO, AO, L # AO point to the data addr
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- dsll PREB, K, BASE_SHIFT
- MOV C21, C11
- MOV C22, C11
-
- MOV C31, C11
- MOV C32, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MOV C41, C11
- MOV C42, C11
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MOV C13, C11
- MOV C14, C11
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MOV C23, C11
- FETCH $0, 0 * SIZE(CO1)
- MOV C24, C11
- FETCH $0, 4 * SIZE(CO1)
-
- MOV C33, C11
- FETCH $0, 0 * SIZE(CO2)
- MOV C34, C11
- FETCH $0, 4 * SIZE(CO2)
-
- daddu PREB, B, PREB
- MOV C43, C11
- FETCH $0, 0 * SIZE(CO3)
-
- MOV C44, C11
- FETCH $0, 4 * SIZE(CO3)
-
- PLU B3, B1, B1
- FETCH $0, 0 * SIZE(CO4)
-
- PLU B4, B2, B2
- FETCH $0, 4 * SIZE(CO4)
-
- #if (defined(LEFT) && !defined(TRANSA)) ||\
- (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK # TEMP is the length of the data part
- #elif defined(LEFT)
- daddiu TEMP, KK, 8
- #else
- daddiu TEMP, KK, 4
- #endif
- dsra L, TEMP, 6
- blez L, .L482
- NOP
- #else
- # GEMM PART
- move BO, B # Reset B
- dsra L, K, 6 # UnRoll K=64
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- dsll PREB, K, BASE_SHIFT
- MOV C21, C11
- MOV C22, C11
-
- MOV C31, C11
- MOV C32, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MOV C41, C11
- MOV C42, C11
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MOV C13, C11
- MOV C14, C11
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MOV C23, C11
- FETCH $0, 0 * SIZE(CO1)
- MOV C24, C11
- FETCH $0, 4 * SIZE(CO1)
-
- MOV C33, C11
- FETCH $0, 0 * SIZE(CO2)
- MOV C34, C11
- FETCH $0, 4 * SIZE(CO2)
-
- daddu PREB, B, PREB
- MOV C43, C11
- FETCH $0, 0 * SIZE(CO3)
-
- MOV C44, C11
- FETCH $0, 4 * SIZE(CO3)
-
- PLU B3, B1, B1
- FETCH $0, 0 * SIZE(CO4)
-
- PLU B4, B2, B2
- blez L, .L482
- FETCH $0, 4 * SIZE(CO4)
- #endif
-
- .L4810:
- daddiu L, L, -1
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- bgtz L, .L4810
- MADPS C44, C44, A8, B8
-
- .align 4
- .L482:
- #ifndef TRMMKERNEL
- andi L, K, 32
- #else
- andi L, TEMP, 32
- #endif
- blez L, .L483
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
-
- .align 4
- .L483:
- #ifndef TRMMKERNEL
- andi L, K, 16
- #else
- andi L, TEMP, 16
- #endif
- blez L, .L484
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
-
- .align 4
- .L484:
- #ifndef TRMMKERNEL
- andi L, K, 8
- #else
- andi L, TEMP, 8
- #endif
- blez L, .L485
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
-
- .align 4
- .L485:
- #ifndef TRMMKERNEL
- andi L, K, 4
- #else
- andi L, TEMP, 4
- #endif
- blez L, .L486
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 4) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 5) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 7) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 8 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 32 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 16 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 20 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 24 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 28 * SIZE(PREA)
- daddiu PREA, PREA, 32 * SIZE
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
-
- .align 4
- .L486:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L487
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 8 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 16 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- PLU B7, B5, B5
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C24, C24, A2, B4
- PLU B8, B6, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddiu PREB, PREB, 8 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- PLU B3, B1, B1
- FETCH $0, 8 * SIZE(PREA)
-
- MADPS C24, C24, A6, B8
- PLU B4, B2, B2
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
- daddiu PREA, PREA, 16 * SIZE
-
-
- .align 4
- .L487:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L480
- LD ALPHA, 152($sp)
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
-
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 4 * SIZE # 4KR*4NR
- MADPS C23, C23, A2, B3
- daddiu AO, AO, 8 * SIZE # 4KR*8MR
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
- MADPS C14, C14, A1, B4
- MADPS C24, C24, A2, B4
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
-
-
- .align 4
- .L480: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C13 # A1=C13.upper=c12
- CVTU A2, C11 # A2=C11.upper=c22
-
- CVTU A3, C23 # A3=C23.upper=c14
- LD B1, 1 * SIZE(CO1)
-
- CVTU A4, C21 # A4=C21.upper=c24
- LD B2, 1 * SIZE(CO2)
-
- CVTU A5, C33 # A5=C33.upper=c16
- LD B3, 3 * SIZE(CO1)
-
- CVTU A6, C31 # A6=C31.upper=c26
- LD B4, 3 * SIZE(CO2)
-
- CVTU A7, C43 # A7=C43.upper=c18
- LD B5, 5 * SIZE(CO1)
-
- CVTU A8, C41 # A8=C41.upper=c28
- LD B6, 5 * SIZE(CO2)
-
- MADD A1, B1, A1, ALPHA # c12
- LD B7, 7 * SIZE(CO1)
-
- MADD A2, B2, A2, ALPHA # c22
- LD B1, 7 * SIZE(CO2)
-
- MADD A3, B3, A3, ALPHA # c14
- LD B2, 0 * SIZE(CO1)
-
- MADD A4, B4, A4, ALPHA # c24
- LD B3, 0 * SIZE(CO2)
-
- MADD A5, B5, A5, ALPHA # c16
- LD B4, 2 * SIZE(CO1)
-
- MADD A6, B6, A6, ALPHA # c26
- LD B5, 2 * SIZE(CO2)
-
- MADD A7, B7, A7, ALPHA # c18
- LD B6, 4 * SIZE(CO1)
-
- MADD A8, B1, A8, ALPHA # c28
- ST A1, 1 * SIZE(CO1)
-
- MADD C11, B2, C11, ALPHA # c12
- LD B7, 4 * SIZE(CO2)
-
- MADD C13, B3, C13, ALPHA # c22
- ST A2, 1 * SIZE(CO2)
-
- MADD C21, B4, C21, ALPHA # c14
- LD A1, 6 * SIZE(CO1)
-
- MADD C23, B5, C23, ALPHA # c24
- ST A3, 3 * SIZE(CO1)
-
- MADD C31, B6, C31, ALPHA # c16
- LD A2, 6 * SIZE(CO2)
-
- MADD C33, B7, C33, ALPHA # c26
- ST A4, 3 * SIZE(CO2)
-
- ST A5, 5 * SIZE(CO1)
- ST A6, 5 * SIZE(CO2)
- ST A7, 7 * SIZE(CO1)
- ST A8, 7 * SIZE(CO2)
-
- MADD C41, A1, C41, ALPHA # c18
- ST C11, 0 * SIZE(CO1)
-
- MADD C43, A2, C43, ALPHA # c28
- ST C13, 0 * SIZE(CO2)
-
- ST C21, 2 * SIZE(CO1)
- ST C23, 2 * SIZE(CO2)
- ST C31, 4 * SIZE(CO1)
- ST C33, 4 * SIZE(CO2)
- ST C41, 6 * SIZE(CO1)
-
- CVTU A1, C14 # B1=C12.upper=c42
- ST C43, 6 * SIZE(CO2)
-
- CVTU A2, C12 # B2=C14.upper=c32
- LD B1, 1 * SIZE(CO3)
-
- CVTU A3, C24 # B3=C22.upper=c44
- LD B2, 1 * SIZE(CO4)
-
- CVTU A4, C22 # B4=C24.upper=c34
- LD B3, 3 * SIZE(CO3)
-
- CVTU A5, C34 # B5=C32.upper=c46
- LD B4, 3 * SIZE(CO4)
-
- CVTU A6, C32 # B6=C24.upper=c36
- LD B5, 5 * SIZE(CO3)
-
- CVTU A7, C44 # B7=C42.upper=c48
- LD B6, 5 * SIZE(CO4)
-
- CVTU A8, C42 # A1=C44.upper=c38
- LD B7, 7 * SIZE(CO3)
-
- MADD A1, B1, A1, ALPHA # c31
- LD C11, 7 * SIZE(CO4)
-
- MADD A2, B2, A2, ALPHA
- LD C13, 0 * SIZE(CO3)
-
- MADD A3, B3, A3, ALPHA
- LD C21, 0 * SIZE(CO4)
-
- MADD A4, B4, A4, ALPHA
- LD C23, 2 * SIZE(CO3)
-
- MADD A5, B5, A5, ALPHA
- LD C31, 2 * SIZE(CO4)
-
- MADD A6, B6, A6, ALPHA
- LD C33, 4 * SIZE(CO3)
-
- MADD A7, B7, A7, ALPHA
- LD C41, 4 * SIZE(CO4)
-
- MADD A8, C11, A8, ALPHA
- ST A1, 1 * SIZE(CO3)
-
- MADD C12, C13, C12, ALPHA
- LD C43, 6 * SIZE(CO3)
-
- MADD C14, C21, C14, ALPHA
- ST A2, 1 * SIZE(CO4)
-
- MADD C22, C23, C22, ALPHA
- LD B1, 6 * SIZE(CO4)
-
- MADD C24, C31, C24, ALPHA
- ST A3, 3 * SIZE(CO3)
-
- MADD C32, C33, C32, ALPHA
- ST A4, 3 * SIZE(CO4)
-
- MADD C34, C41, C34, ALPHA
- ST A5, 5 * SIZE(CO3)
-
- MADD C42, C43, C42, ALPHA
- ST A6, 5 * SIZE(CO4)
-
- ST A7, 7 * SIZE(CO3)
- NOP
-
- MADD C44, B1, C44, ALPHA
- ST A8, 7 * SIZE(CO4)
-
- ST C12, 0 * SIZE(CO3)
- ST C14, 0 * SIZE(CO4)
- ST C22, 2 * SIZE(CO3)
- ST C24, 2 * SIZE(CO4)
- ST C32, 4 * SIZE(CO3)
- ST C34, 4 * SIZE(CO4)
- ST C42, 6 * SIZE(CO3)
- ST C44, 6 * SIZE(CO4)
-
- daddiu CO1, CO1, 8 * SIZE
- daddiu CO2, CO2, 8 * SIZE
- daddiu CO3, CO3, 8 * SIZE
- bgtz I, .L481
- daddiu CO4, CO4, 8 * SIZE
- #else
- daddiu I, I, -1
- CVTU A1, C13 # A1=C13.upper=c12
- CVTU A2, C11 # A2=C11.upper=c22
- CVTU A3, C23 # A3=C23.upper=c14
- CVTU A4, C21 # A4=C21.upper=c24
- CVTU A5, C33 # A5=C33.upper=c16
- CVTU A6, C31 # A6=C31.upper=c26
- CVTU A7, C43 # A7=C43.upper=c18
- CVTU A8, C41 # A8=C41.upper=c28
-
- MUL A1, A1, ALPHA # c12
- MUL A2, A2, ALPHA # c22
- MUL A3, A3, ALPHA # c14
- MUL A4, A4, ALPHA # c24
- MUL A5, A5, ALPHA # c16
- MUL A6, A6, ALPHA # c26
- MUL A7, A7, ALPHA # c18
- MUL A8, A8, ALPHA # c28
-
- MUL C11, C11, ALPHA # c12
- ST A1, 1 * SIZE(CO1)
-
- MUL C13, C13, ALPHA # c22
- ST A2, 1 * SIZE(CO2)
-
- MUL C21, C21, ALPHA # c14
- ST A3, 3 * SIZE(CO1)
-
- MUL C23, C23, ALPHA # c24
- ST A4, 3 * SIZE(CO2)
-
- MUL C31, C31, ALPHA # c16
- ST A5, 5 * SIZE(CO1)
-
- MUL C33, C33, ALPHA # c26
- ST A6, 5 * SIZE(CO2)
-
- MUL C41, C41, ALPHA # c18
- ST A7, 7 * SIZE(CO1)
-
- MUL C43, C43, ALPHA # c28
- ST A8, 7 * SIZE(CO2)
-
- CVTU A1, C14 # B1=C12.upper=c42
- ST C11, 0 * SIZE(CO1)
-
- CVTU A2, C12 # B2=C14.upper=c32
- ST C13, 0 * SIZE(CO2)
-
- CVTU A3, C24 # B3=C22.upper=c44
- ST C21, 2 * SIZE(CO1)
-
- CVTU A4, C22 # B4=C24.upper=c34
- ST C23, 2 * SIZE(CO2)
-
- CVTU A5, C34 # B5=C32.upper=c46
- ST C31, 4 * SIZE(CO1)
-
- CVTU A6, C32 # B6=C24.upper=c36
- ST C33, 4 * SIZE(CO2)
-
- CVTU A7, C44 # B7=C42.upper=c48
- ST C41, 6 * SIZE(CO1)
-
- CVTU A8, C42 # A1=C44.upper=c38
- ST C43, 6 * SIZE(CO2)
-
- MUL A1, A1, ALPHA # c31
- MUL A2, A2, ALPHA
- MUL A3, A3, ALPHA
- MUL A4, A4, ALPHA
- MUL A5, A5, ALPHA
- MUL A6, A6, ALPHA
- MUL A7, A7, ALPHA
- MUL A8, A8, ALPHA
-
- MUL C12, C12, ALPHA
- ST A1, 1 * SIZE(CO3)
-
- MUL C14, C14, ALPHA
- ST A2, 1 * SIZE(CO4)
-
- MUL C22, C22, ALPHA
- ST A3, 3 * SIZE(CO3)
-
- MUL C24, C24, ALPHA
- ST A4, 3 * SIZE(CO4)
-
- MUL C32, C32, ALPHA
- ST A5, 5 * SIZE(CO3)
-
- MUL C34, C34, ALPHA
- ST A6, 5 * SIZE(CO4)
-
- MUL C42, C42, ALPHA
- ST A7, 7 * SIZE(CO3)
-
- MUL C44, C44, ALPHA
- ST A8, 7 * SIZE(CO4)
-
- ST C12, 0 * SIZE(CO3)
- ST C14, 0 * SIZE(CO4)
- ST C22, 2 * SIZE(CO3)
- ST C24, 2 * SIZE(CO4)
- ST C32, 4 * SIZE(CO3)
- ST C34, 4 * SIZE(CO4)
- ST C42, 6 * SIZE(CO3)
- ST C44, 6 * SIZE(CO4)
-
- daddiu CO1, CO1, 8 * SIZE
- daddiu CO2, CO2, 8 * SIZE
- daddiu CO3, CO3, 8 * SIZE
- daddiu CO4, CO4, 8 * SIZE
-
- #if ( defined(LEFT) && defined(TRANSA)) ||\
- (!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
- #ifdef LEFT
- daddiu TEMP, TEMP, -8
- #else
- daddiu TEMP, TEMP, -4
- #endif
- dsll L, TEMP, 3 + BASE_SHIFT
- dsll TEMP, TEMP, 2 + BASE_SHIFT
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 8
- #endif
-
- bgtz I, .L481
- NOP
- #endif
-
- .align 4
- .L44:
- andi I, M, 4 # MR=4
- blez I, .L42
- NOP
-
- .align 4
- .L441:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) ||\
- (!defined(LEFT) && !defined(TRANSA))
- move BO, B # Reset B
- #else
- dsll L, KK, 2 + BASE_SHIFT
- dsll TEMP, KK, 2 + BASE_SHIFT
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- dsll PREB, K, BASE_SHIFT
- MOV C21, C11
- MOV C22, C11
-
- MOV C31, C11
- MOV C32, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MOV C41, C11
- MOV C42, C11
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MOV C13, C11
- MOV C14, C11
-
- MOV C23, C11
- FETCH $0, 0 * SIZE(CO1)
- MOV C24, C11
-
- MOV C33, C11
- FETCH $0, 0 * SIZE(CO2)
- MOV C34, C11
-
- daddu PREB, B, PREB
- MOV C43, C11
- FETCH $0, 0 * SIZE(CO3)
-
- MOV C44, C11
- PLU B3, B1, B1
-
- FETCH $0, 0 * SIZE(CO4)
- PLU B4, B2, B2
-
- #if (defined(LEFT) && !defined(TRANSA)) ||\
- (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddu TEMP, KK, 4
- #else
- daddu TEMP, KK, 4
- #endif
- dsra L, TEMP, 2
- blez L, .L442
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=4
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- dsll PREB, K, BASE_SHIFT
- MOV C21, C11
- MOV C22, C11
-
- MOV C31, C11
- MOV C32, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MOV C41, C11
- MOV C42, C11
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MOV C13, C11
- MOV C14, C11
-
- MOV C23, C11
- FETCH $0, 0 * SIZE(CO1)
- MOV C24, C11
-
- MOV C33, C11
- FETCH $0, 0 * SIZE(CO2)
- MOV C34, C11
-
- daddu PREB, B, PREB
- MOV C43, C11
- FETCH $0, 0 * SIZE(CO3)
-
- MOV C44, C11
- PLU B3, B1, B1
-
- FETCH $0, 0 * SIZE(CO4)
- blez L, .L442
- PLU B4, B2, B2
- #endif
-
- .L4410: #
- daddiu L, L, -1
- MADPS C11, C11, A1, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C21, C21, A2, B1
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C12, C12, A1, B2
- FETCH $0, 0 * SIZE(PREB)
-
- MADPS C22, C22, A2, B2
- FETCH $0, 0 * SIZE(PREA)
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C14, C14, A1, B4
- MADPS C24, C24, A2, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- MADPS C11, C11, A3, B5
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C21, C21, A4, B5
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- MADPS C12, C12, A3, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C22, C22, A4, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C13, C13, A3, B7
- MADPS C23, C23, A4, B7
-
- MADPS C14, C14, A3, B8
- MADPS C24, C24, A4, B8
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- MADPS C11, C11, A5, B1
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- MADPS C21, C21, A6, B1
- gsLQC1(R12, F7, F6, 3) # A7 A8
-
- MADPS C12, C12, A5, B2
- FETCH $0, 8 * SIZE(PREB)
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
-
- MADPS C22, C22, A6, B2
- FETCH $0, 8 * SIZE(PREA)
- daddiu AO, AO, 16 * SIZE # 4KR*4MR
-
- MADPS C13, C13, A5, B3
- MADPS C23, C23, A6, B3
-
- MADPS C14, C14, A5, B4
- MADPS C24, C24, A6, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- MADPS C11, C11, A7, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C21, C21, A8, B5
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MADPS C12, C12, A7, B6
- FETCH $0, 12 * SIZE(PREB)
-
- MADPS C22, C22, A8, B6
- FETCH $0, 12 * SIZE(PREA)
-
- MADPS C13, C13, A7, B7
- daddiu PREA, PREA, 16 * SIZE
- MADPS C23, C23, A8, B7
- daddiu PREB, PREB, 16 * SIZE
-
- MADPS C14, C14, A7, B8
- MADPS C24, C24, A8, B8
-
- PLU B3, B1, B1
- bgtz L, .L4410
- PLU B4, B2, B2
-
- .align 4
- .L442:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L443
- NOP
-
- MADPS C11, C11, A1, B1
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C21, C21, A2, B1
- gsLQC1(R12, F3, F2, 1) # A3 A4
-
- MADPS C12, C12, A1, B2
- FETCH $0, 0 * SIZE(PREB)
- daddiu BO, BO, 8 * SIZE # 2KR*4NR
-
- MADPS C22, C22, A2, B2
- FETCH $0, 0 * SIZE(PREA)
- daddiu AO, AO, 8 * SIZE # 2KR*4MR
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C14, C14, A1, B4
- MADPS C24, C24, A2, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- MADPS C11, C11, A3, B5
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C21, C21, A4, B5
- gsLQC1(R12, F1, F0, 0) # A5 A6
-
- MADPS C12, C12, A3, B6
- FETCH $0, 4 * SIZE(PREB)
-
- MADPS C22, C22, A4, B6
- FETCH $0, 4 * SIZE(PREA)
-
- MADPS C13, C13, A3, B7
- daddiu PREB, PREB, 8
- MADPS C23, C23, A4, B7
- daddiu PREA, PREA, 8
-
- MADPS C14, C14, A3, B8
- MADPS C24, C24, A4, B8
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
-
- .align 4
- .L443:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L440
- LD ALPHA, 152($sp)
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- MADPS C12, C12, A1, B2
- daddiu BO, BO, 4 * SIZE # 1KR*4NR
- MADPS C22, C22, A2, B2
- daddiu AO, AO, 4 * SIZE # 1KR*4MR
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C14, C14, A1, B4
- MADPS C24, C24, A2, B4
-
-
- .align 4
- .L440:
- #ifndef TRMMKERNEL
- CVTU A1, C13 # A1=C13.upper=c12
- LD B1, 1 * SIZE(CO1)
-
- CVTU A2, C11 # A2=C11.upper=c22
- LD B2, 1 * SIZE(CO2)
-
- CVTU A3, C23 # A3=C23.upper=c14
- LD B3, 3 * SIZE(CO1)
-
- CVTU A4, C21 # A4=C21.upper=c24
- LD B4, 3 * SIZE(CO2)
-
-
- MADD A1, B1, A1, ALPHA # c12
- LD B5, 0 * SIZE(CO1)
-
- MADD A2, B2, A2, ALPHA # c22
- LD B6, 0 * SIZE(CO2)
-
- MADD A3, B3, A3, ALPHA # c14
- LD B7, 2 * SIZE(CO1)
-
- MADD A4, B4, A4, ALPHA # c24
- LD B1, 2 * SIZE(CO2)
-
- MADD C11, B5, C11, ALPHA # c12
- ST A1, 1 * SIZE(CO1)
-
- MADD C13, B6, C13, ALPHA # c22
- ST A2, 1 * SIZE(CO2)
-
- MADD C21, B7, C21, ALPHA # c14
- ST A3, 3 * SIZE(CO1)
-
- MADD C23, B1, C23, ALPHA # c24
- ST A4, 3 * SIZE(CO2)
-
- ST C11, 0 * SIZE(CO1)
- ST C13, 0 * SIZE(CO2)
- ST C21, 2 * SIZE(CO1)
- ST C23, 2 * SIZE(CO2)
-
- CVTU A1, C14 # B1=C12.upper=c42
- LD B1, 1 * SIZE(CO3)
-
- CVTU A2, C12 # B2=C14.upper=c32
- LD B2, 1 * SIZE(CO4)
-
- CVTU A3, C24 # B3=C22.upper=c44
- LD B3, 3 * SIZE(CO3)
-
- CVTU A4, C22 # B4=C24.upper=c34
- LD B4, 3 * SIZE(CO4)
-
- MADD A1, B1, A1, ALPHA # c31
- LD A5, 0 * SIZE(CO3)
-
- MADD A2, B2, A2, ALPHA
- LD A6, 0 * SIZE(CO4)
-
- MADD A3, B3, A3, ALPHA
- LD A7, 2 * SIZE(CO3)
-
- MADD A4, B4, A4, ALPHA
- LD A8, 2 * SIZE(CO4)
-
- MADD C12, A5, C12, ALPHA
- ST A1, 1 * SIZE(CO3)
-
- MADD C14, A6, C14, ALPHA
- ST A2, 1 * SIZE(CO4)
-
- MADD C22, A7, C22, ALPHA
- ST A3, 3 * SIZE(CO3)
-
- MADD C24, A8, C24, ALPHA
- ST A4, 3 * SIZE(CO4)
-
- ST C12, 0 * SIZE(CO3)
- ST C14, 0 * SIZE(CO4)
- ST C22, 2 * SIZE(CO3)
- ST C24, 2 * SIZE(CO4)
-
- daddiu CO1, CO1, 4 * SIZE
- daddiu CO2, CO2, 4 * SIZE
- daddiu CO3, CO3, 4 * SIZE
- daddiu CO4, CO4, 4 * SIZE
-
- #else
- CVTU A1, C13 # A1=C13.upper=c12
- CVTU A2, C11 # A2=C11.upper=c22
- CVTU A3, C23 # A3=C23.upper=c14
- CVTU A4, C21 # A4=C21.upper=c24
-
- MUL A1, A1, ALPHA # c12
- MUL A2, A2, ALPHA # c22
- MUL A3, A3, ALPHA # c14
- MUL A4, A4, ALPHA # c24
-
- MUL C11, C11, ALPHA # c12
- ST A1, 1 * SIZE(CO1)
-
- MUL C13, C13, ALPHA # c22
- ST A2, 1 * SIZE(CO2)
-
- MUL C21, C21, ALPHA # c14
- ST A3, 3 * SIZE(CO1)
-
- MUL C23, C23, ALPHA # c24
- ST A4, 3 * SIZE(CO2)
-
- CVTU A5, C14 # B1=C12.upper=c42
- ST C11, 0 * SIZE(CO1)
-
- CVTU A6, C12 # B2=C14.upper=c32
- ST C13, 0 * SIZE(CO2)
-
- CVTU A7, C24 # B3=C22.upper=c44
- ST C21, 2 * SIZE(CO1)
-
- CVTU A8, C22 # B4=C24.upper=c34
- ST C23, 2 * SIZE(CO2)
-
- MUL A5, A5, ALPHA # c31
- MUL A6, A6, ALPHA
- MUL A7, A7, ALPHA
- MUL A8, A8, ALPHA
-
- MUL C12, C12, ALPHA
- ST A5, 1 * SIZE(CO3)
-
- MUL C14, C14, ALPHA
- ST A6, 1 * SIZE(CO4)
-
- MUL C22, C22, ALPHA
- ST A7, 3 * SIZE(CO3)
-
- MUL C24, C24, ALPHA
- ST A8, 3 * SIZE(CO4)
-
- ST C12, 0 * SIZE(CO3)
- ST C14, 0 * SIZE(CO4)
- ST C22, 2 * SIZE(CO3)
- ST C24, 2 * SIZE(CO4)
-
- daddiu CO1, CO1, 4 * SIZE
- daddiu CO2, CO2, 4 * SIZE
- daddiu CO3, CO3, 4 * SIZE
- daddiu CO4, CO4, 4 * SIZE
-
- #if ( defined(LEFT) && defined(TRANSA))||\
- (!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
- #ifdef LEFT
- daddiu TEMP, TEMP, -4
- #else
- daddiu TEMP, TEMP, -4
- #endif
- dsll L, TEMP, 2 + BASE_SHIFT
- dsll TEMP, TEMP, 2 + BASE_SHIFT
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 4
- #endif
- #endif
-
- .align 4
- .L42:
- andi I, M, 2
- blez I, .L41
- NOP
-
- .align 4
- .L421:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) ||\
- (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, 1 + BASE_SHIFT
- dsll TEMP, KK, 2 + BASE_SHIFT
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- MOV C21, C11
- MOV C22, C11
-
- MOV C31, C11
- MOV C32, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MOV C41, C11
- MOV C42, C11
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MOV C13, C11
- MOV C14, C11
-
- MOV C23, C11
- FETCH $0, 0 * SIZE(CO1)
- MOV C24, C11
-
- MOV C33, C11
- FETCH $0, 0 * SIZE(CO2)
- MOV C34, C11
-
- MOV C43, C11
- FETCH $0, 0 * SIZE(CO3)
-
- MOV C44, C11
- PLU B3, B1, B1
-
- FETCH $0, 0 * SIZE(CO4)
- PLU B4, B2, B2
- #if (defined(LEFT) && !defined(TRANSA)) ||\
- (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 2
- #else
- daddiu TEMP, KK, 4
- #endif
- dsra L, TEMP, 2
- blez L, .L422
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=4
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- MOV C21, C11
- MOV C22, C11
-
- MOV C31, C11
- MOV C32, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MOV C41, C11
- MOV C42, C11
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- MOV C13, C11
- MOV C14, C11
-
- MOV C23, C11
- FETCH $0, 0 * SIZE(CO1)
- MOV C24, C11
-
- MOV C33, C11
- FETCH $0, 0 * SIZE(CO2)
- MOV C34, C11
-
- MOV C43, C11
- FETCH $0, 0 * SIZE(CO3)
-
- MOV C44, C11
- PLU B3, B1, B1
-
- FETCH $0, 0 * SIZE(CO4)
- blez L, .L422
- PLU B4, B2, B2
- #endif
-
- .L4210:
- daddiu L, L, -1
- MADPS C11, C11, A1, B1
- MADPS C12, C12, A1, B2
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C13, C13, A1, B3
- MADPS C14, C14, A1, B4
- gsLQC1(R12, F3, F2, 1) # B1 B2
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- MADPS C11, C11, A2, B5
- MADPS C12, C12, A2, B6
- daddiu AO, AO, 8 * SIZE # 4KR*2MR
- gsLQC1(R13, F9, F8, 2) # B1 B2
-
- MADPS C13, C13, A2, B7
- MADPS C14, C14, A2, B8
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- MADPS C11, C11, A3, B1
- gsLQC1(R12, F1, F0, 0) # B3 B4
-
- MADPS C12, C12, A3, B2
- gsLQC1(R13, F13, F12, 3) # B3 B4
-
- daddiu BO, BO, 16 * SIZE # 4KR*4NR
- MADPS C13, C13, A3, B3
- MADPS C14, C14, A3, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- MADPS C11, C11, A4, B5
- MADPS C12, C12, A4, B6
- gsLQC1(R13, F9, F8, 0) # B3 B4
-
- MADPS C13, C13, A4, B7
- MADPS C14, C14, A4, B8
-
- PLU B3, B1, B1
- bgtz L, .L4210
- PLU B4, B2, B2
-
- .align 4
- .L422:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L423
- NOP
-
- daddiu AO, AO, 4 * SIZE # 2KR*2MR
- MADPS C11, C11, A1, B1
- MADPS C12, C12, A1, B2
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- MADPS C13, C13, A1, B3
- MADPS C14, C14, A1, B4
- daddiu BO, BO, 8 * SIZE # 2KR*2MR
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- MADPS C11, C11, A2, B5
- MADPS C12, C12, A2, B6
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- MADPS C13, C13, A2, B7
- MADPS C14, C14, A2, B8
- gsLQC1(R12, F1, F0, 0)
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- .L423:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L420
- LD ALPHA, 152($sp)
-
- MADPS C11, C11, A1, B1
- MADPS C12, C12, A1, B2
- daddiu BO, BO, 4 * SIZE # 2KR*4NR
- daddiu AO, AO, 2 * SIZE # 2KR*4MR
-
- MADPS C13, C13, A1, B3
- MADPS C14, C14, A1, B4
-
- .align 4
- .L420:
- #ifndef TRMMKERNEL
- CVTU A1, C13 # A1=C13.upper=c12
- LD B1, 1 * SIZE(CO1)
-
- CVTU A2, C11 # A2=C11.upper=c22
- LD B2, 1 * SIZE(CO2)
-
- MADD A1, B1, A1, ALPHA # c12
- LD B5, 0 * SIZE(CO1)
-
- MADD A2, B2, A2, ALPHA # c22
- LD B6, 0 * SIZE(CO2)
-
- MADD C11, B5, C11, ALPHA # c12
- ST A1, 1 * SIZE(CO1)
-
- MADD C13, B6, C13, ALPHA # c22
- ST A2, 1 * SIZE(CO2)
-
- ST C11, 0 * SIZE(CO1)
- ST C13, 0 * SIZE(CO2)
-
- CVTU A1, C14 # B1=C12.upper=c42
- LD B1, 1 * SIZE(CO3)
-
- CVTU A2, C12 # B2=C14.upper=c32
- LD B2, 1 * SIZE(CO4)
-
- MADD A1, B1, A1, ALPHA # c31
- LD A5, 0 * SIZE(CO3)
-
- MADD A2, B2, A2, ALPHA
- LD A6, 0 * SIZE(CO4)
-
- MADD C12, A5, C12, ALPHA
- ST A1, 1 * SIZE(CO3)
-
- MADD C14, A6, C14, ALPHA
- ST A2, 1 * SIZE(CO4)
-
- ST C12, 0 * SIZE(CO3)
- ST C14, 0 * SIZE(CO4)
-
- daddiu CO1, CO1, 2 * SIZE
- daddiu CO2, CO2, 2 * SIZE
- daddiu CO3, CO3, 2 * SIZE
- daddiu CO4, CO4, 2 * SIZE
- #else
- CVTU A1, C13 # A1=C13.upper=c12
- CVTU A2, C11 # A2=C11.upper=c22
-
- MUL A1, A1, ALPHA # c12
- MUL A2, A2, ALPHA # c22
-
- MUL C11, C11, ALPHA # c12
- MUL C13, C13, ALPHA # c22
-
- CVTU A3, C14 # B1=C12.upper=c42
- CVTU A4, C12 # B2=C14.upper=c32
-
- MUL A3, A3, ALPHA # c31
- ST A1, 1 * SIZE(CO1)
-
- MUL A4, A4, ALPHA
- ST A2, 1 * SIZE(CO2)
-
- MUL C12, C12, ALPHA
- ST C11, 0 * SIZE(CO1)
-
- MUL C14, C14, ALPHA
- ST C13, 0 * SIZE(CO2)
-
- ST A3, 1 * SIZE(CO3)
- ST A4, 1 * SIZE(CO4)
-
- ST C12, 0 * SIZE(CO3)
- ST C14, 0 * SIZE(CO4)
-
- daddiu CO1, CO1, 2 * SIZE
- daddiu CO2, CO2, 2 * SIZE
- daddiu CO3, CO3, 2 * SIZE
- daddiu CO4, CO4, 2 * SIZE
- #if ( defined(LEFT) && defined(TRANSA))||\
- (!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
- #ifdef LEFT
- daddiu TEMP, TEMP, -2
- #else
- daddiu TEMP, TEMP, -4
- #endif
- dsll L, TEMP, 1 + BASE_SHIFT
- dsll TEMP, TEMP, 2 + BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 2
- #endif
- #endif
-
-
- .align 4
- .L41:
- andi I, M, 1
- blez I, .L40
- NOP
-
- .align 4
- .L411:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) ||\
- (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, BASE_SHIFT
- dsll TEMP, KK, 2 + BASE_SHIFT
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C21, C11
- MOV C22, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C31, C11
- MOV C32, C11
- LD B2, 1 * SIZE(BO)
-
- MOV C41, C11
- MOV C42, C11
- LD B3, 2 * SIZE(BO)
-
- MOV C13, C11
- MOV C14, C11
- LD B4, 3 * SIZE(BO)
-
- MOV C23, C11
- MOV C24, C11
-
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- MOV C44, C11
- #if (defined(LEFT) && !defined(TRANSA))||\
- (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 1
- #else
- daddiu TEMP, KK, 4
- #endif
- dsra L, TEMP, 2
- blez L, .L412
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=4
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C21, C11
- MOV C22, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C31, C11
- MOV C32, C11
- LD B2, 1 * SIZE(BO)
-
- MOV C41, C11
- MOV C42, C11
- LD B3, 2 * SIZE(BO)
-
- MOV C13, C11
- MOV C14, C11
- LD B4, 3 * SIZE(BO)
-
- MOV C23, C11
- MOV C24, C11
-
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- blez L, .L412
- MOV C44, C11
- #endif
-
- .L4110:
- daddiu L, L, -1
- LD A2, 1 * SIZE(AO)
-
- MADD C11, C11, A1, B1
- LD B5, 4 * SIZE(BO)
-
- MADD C12, C12, A1, B2
- LD B6, 5 * SIZE(BO)
-
- MADD C13, C13, A1, B3
- LD B7, 6 * SIZE(BO)
-
- MADD C14, C14, A1, B4
- LD B8, 7 * SIZE(BO)
-
- LD A3, 2 * SIZE(AO)
- NOP
-
- MADD C11, C11, A2, B5
- LD B1, 8 * SIZE(BO)
-
- MADD C12, C12, A2, B6
- LD B2, 9 * SIZE(BO)
-
- MADD C13, C13, A2, B7
- LD B3, 10 * SIZE(BO)
-
- MADD C14, C14, A2, B8
- LD B4, 11 * SIZE(BO)
-
- LD A4, 3 * SIZE(AO)
- daddiu AO, AO, 4 * SIZE
-
- MADD C11, C11, A3, B1
- LD B5, 12 * SIZE(BO)
-
- MADD C12, C12, A3, B2
- LD B6, 13 * SIZE(BO)
-
- MADD C13, C13, A3, B3
- LD B7, 14 * SIZE(BO)
-
- MADD C14, C14, A3, B4
- LD B8, 15 * SIZE(BO)
-
- LD A1, 0 * SIZE(AO)
- daddiu BO, BO, 16 * SIZE
-
- MADD C11, C11, A4, B5
- LD B1, 0 * SIZE(BO)
-
- MADD C12, C12, A4, B6
- LD B2, 1 * SIZE(BO)
-
- MADD C13, C13, A4, B7
- LD B3, 2 * SIZE(BO)
-
- MADD C14, C14, A4, B8
- bgtz L, .L4110
- LD B4, 3 * SIZE(BO)
-
- .L412:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L413
- NOP
-
- LD A2, 1 * SIZE(AO)
- daddiu AO, AO, 2 * SIZE
-
- MADD C11, C11, A1, B1
- LD B5, 4 * SIZE(BO)
-
- MADD C12, C12, A1, B2
- LD B6, 5 * SIZE(BO)
-
- MADD C13, C13, A1, B3
- LD B7, 6 * SIZE(BO)
-
- MADD C14, C14, A1, B4
- LD B8, 7 * SIZE(BO)
-
- LD A1, 0 * SIZE(AO)
- daddiu BO, BO, 8 * SIZE
-
- MADD C11, C11, A2, B5
- LD B1, 0 * SIZE(BO)
-
- MADD C12, C12, A2, B6
- LD B2, 1 * SIZE(BO)
-
- MADD C13, C13, A2, B7
- LD B3, 2 * SIZE(BO)
-
- MADD C14, C14, A2, B8
- LD B4, 3 * SIZE(BO)
-
- .L413:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L410
- LD ALPHA, 152($sp)
-
- MADD C11, C11, A1, B1
- MADD C12, C12, A1, B2
- daddiu AO, AO, 1 * SIZE
- MADD C13, C13, A1, B3
- MADD C14, C14, A1, B4
- daddiu BO, BO, 4 * SIZE
-
- .align 4
- .L410:
- #ifndef TRMMKERNEL
- LD A5, 0 * SIZE(CO1)
- LD A6, 0 * SIZE(CO2)
- LD A7, 0 * SIZE(CO3)
- LD A8, 0 * SIZE(CO4)
-
- MADD A5, A5, C11, ALPHA
- MADD A6, A6, C12, ALPHA
- MADD A7, A7, C13, ALPHA
- MADD A8, A8, C14, ALPHA
-
- ST A5, 0 * SIZE(CO1)
- ST A6, 0 * SIZE(CO2)
- ST A7, 0 * SIZE(CO3)
- ST A8, 0 * SIZE(CO4)
-
- daddiu CO1, CO1, 1 * SIZE
- daddiu CO2, CO2, 1 * SIZE
- daddiu CO3, CO3, 1 * SIZE
- daddiu CO4, CO4, 1 * SIZE
- #else
- MUL A5, C11, ALPHA
- MUL A6, C12, ALPHA
- MUL A7, C13, ALPHA
- MUL A8, C14, ALPHA
-
- ST A5, 0 * SIZE(CO1)
- ST A6, 0 * SIZE(CO2)
- ST A7, 0 * SIZE(CO3)
- ST A8, 0 * SIZE(CO4)
-
- daddiu CO1, CO1, 1 * SIZE
- daddiu CO2, CO2, 1 * SIZE
- daddiu CO3, CO3, 1 * SIZE
- daddiu CO4, CO4, 1 * SIZE
-
- #if ( defined(LEFT) && defined(TRANSA))||\
- (!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
- #ifdef LEFT
- daddiu TEMP, TEMP, -1
- #else
- daddiu TEMP, TEMP, -4
- #endif
-
- dsll L, TEMP, BASE_SHIFT
- dsll TEMP, TEMP, 2 + BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
- #ifdef LEFT
- daddiu KK, KK, 1
- #endif
- #endif
-
- .align 4
- .L40:
- #if defined(TRMMKERNEL) && !defined(LEFT)
- daddiu KK, KK, 4
- #endif
- daddiu J, J, -1
- move B, BO
- bgtz J, .L48
- NOP
-
-
-
- .align 4
- .L2: # Nr=2
- andi J, N, 2
- blez J, .L1
- NOP
-
- .L28:
- dsra I, M, 3 # MR=8
-
- move AO, A # Reset A
- move CO1, C
-
- #if defined(TRMMKERNEL) && defined(LEFT)
- move KK, OFFSET
- #endif
- daddu CO2, C, LDC
- blez I, .L24
- daddu C, CO2, LDC
-
- .align 4
- .L281:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, 3 + BASE_SHIFT
- dsll TEMP, KK, 1 + BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- LD A1, 0 * SIZE(AO)
-
- MOV C12, C11
- LD A2, 1 * SIZE(AO)
-
- MOV C21, C11
- LD A3, 2 * SIZE(AO)
-
- MOV C22, C11
- LD A4, 3 * SIZE(AO)
-
- MOV C31, C11
- LD A5, 4 * SIZE(AO)
-
- MOV C32, C11
- LD A6, 5 * SIZE(AO)
-
- MOV C41, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C42, C11
- LD B2, 1 * SIZE(BO)
-
- MOV C13, C11
- LD A7, 6 * SIZE(AO)
-
- MOV C14, C11
- LD A8, 7 * SIZE(AO)
-
- MOV C23, C11
- MOV C24, C11
-
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- MOV C44, C11
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 8
- #else
- daddiu TEMP, KK, 2
- #endif
- dsra L, TEMP, 1
- blez L, .L282
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 1 # UnRoll K=4
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- LD A1, 0 * SIZE(AO)
-
- MOV C12, C11
- LD A2, 1 * SIZE(AO)
-
- MOV C21, C11
- LD A3, 2 * SIZE(AO)
-
- MOV C22, C11
- LD A4, 3 * SIZE(AO)
-
- MOV C31, C11
- LD A5, 4 * SIZE(AO)
-
- MOV C32, C11
- LD A6, 5 * SIZE(AO)
-
- MOV C41, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C42, C11
- LD B2, 1 * SIZE(BO)
-
- MOV C13, C11
- LD A7, 6 * SIZE(AO)
-
- MOV C14, C11
- LD A8, 7 * SIZE(AO)
-
- MOV C23, C11
- MOV C24, C11
-
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- blez L, .L282
- MOV C44, C11
- #endif
-
- .align 4
- .L2810:
- daddiu L, L, -1
- MADD C11, C11, A1, B1
- LD B5, 8 * SIZE(AO)
-
- MADD C21, C21, A2, B1
- LD B6, 9 * SIZE(AO)
-
- MADD C31, C31, A3, B1
- LD B7, 10 * SIZE(AO)
-
- MADD C41, C41, A4, B1
- LD B8, 11 * SIZE(AO)
-
- MADD C12, C12, A1, B2
- MADD C22, C22, A2, B2
- LD B3, 2 * SIZE(BO)
-
- MADD C32, C32, A3, B2
- MADD C42, C42, A4, B2
- LD B4, 3 * SIZE(BO)
- daddiu BO, BO, 4 * SIZE
-
- MADD C13, C13, A5, B1
- MADD C23, C23, A6, B1
- LD A1, 12 * SIZE(AO)
-
- MADD C33, C33, A7, B1
- MADD C43, C43, A8, B1
- LD A2, 13 * SIZE(AO)
-
- MADD C14, C14, A5, B2
- MADD C24, C24, A6, B2
- LD A3, 14 * SIZE(AO)
-
- MADD C34, C34, A7, B2
- MADD C44, C44, A8, B2
- LD A4, 15 * SIZE(AO)
- daddiu AO, AO, 16 * SIZE
-
- MADD C11, C11, B5, B3
- LD A5, 4 * SIZE(AO)
-
- MADD C21, C21, B6, B3
- LD A6, 5 * SIZE(AO)
-
- MADD C13, C13, A1, B3
- MADD C23, C23, A2, B3
- LD A7, 6 * SIZE(AO)
-
- MADD C33, C33, A3, B3
- MADD C43, C43, A4, B3
- LD A8, 7 * SIZE(AO)
-
- MADD C14, C14, A1, B4
- MADD C24, C24, A2, B4
- LD B1, 0 * SIZE(BO)
-
- MADD C34, C34, A3, B4
- MADD C44, C44, A4, B4
- LD B2, 1 * SIZE(BO)
-
- MADD C31, C31, B7, B3
- MADD C41, C41, B8, B3
- LD A1, 0 * SIZE(AO)
-
- MADD C12, C12, B5, B4
- LD A2, 1 * SIZE(AO)
-
- MADD C22, C22, B6, B4
- LD A3, 2 * SIZE(AO)
-
- LD A4, 3 * SIZE(AO)
- MADD C32, C32, B7, B4
- bgtz L, .L2810
- MADD C42, C42, B8, B4
-
- .align 4
- .L282:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L280
- LD ALPHA, 152($sp)
-
- MADD C13, C13, A5, B1
- MADD C23, C23, A6, B1
- MADD C33, C33, A7, B1
- MADD C43, C43, A8, B1
- MADD C14, C14, A5, B2
- MADD C24, C24, A6, B2
- MADD C34, C34, A7, B2
- MADD C44, C44, A8, B2
- daddiu AO, AO, 8 * SIZE
-
- MADD C11, C11, A1, B1
- MADD C21, C21, A2, B1
- MADD C31, C31, A3, B1
- MADD C41, C41, A4, B1
- MADD C12, C12, A1, B2
- MADD C22, C22, A2, B2
- MADD C32, C32, A3, B2
- MADD C42, C42, A4, B2
- daddiu BO, BO, 2 * SIZE
-
-
- .align 4
- .L280: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
-
- LD A1, 0 * SIZE(CO1)
- LD A2, 1 * SIZE(CO1)
- LD A3, 2 * SIZE(CO1)
- LD A4, 3 * SIZE(CO1)
- LD A5, 4 * SIZE(CO1)
- LD A6, 5 * SIZE(CO1)
- LD A7, 6 * SIZE(CO1)
- LD A8, 7 * SIZE(CO1)
-
- MADD A1, A1, C11, ALPHA
- LD B1, 0 * SIZE(CO2)
-
- MADD A2, A2, C21, ALPHA
- LD B2, 1 * SIZE(CO2)
-
- MADD A3, A3, C31, ALPHA
- LD B3, 2 * SIZE(CO2)
-
- MADD A4, A4, C41, ALPHA
- LD B4, 3 * SIZE(CO2)
-
- MADD A5, A5, C13, ALPHA
- LD B5, 4 * SIZE(CO2)
-
- MADD A6, A6, C23, ALPHA
- LD B6, 5 * SIZE(CO2)
-
- MADD A7, A7, C33, ALPHA
- LD B7, 6 * SIZE(CO2)
-
- MADD A8, A8, C43, ALPHA
- LD C11, 7 * SIZE(CO2)
-
- MADD B1, B1, C12, ALPHA
- ST A1, 0 * SIZE(CO1)
-
- MADD B2, B2, C22, ALPHA
- ST A2, 1 * SIZE(CO1)
-
- MADD B3, B3, C32, ALPHA
- ST A3, 2 * SIZE(CO1)
-
- MADD B4, B4, C42, ALPHA
- ST A4, 3 * SIZE(CO1)
-
- MADD B5, B5, C14, ALPHA
- ST A5, 4 * SIZE(CO1)
-
- MADD B6, B6, C24, ALPHA
- ST A6, 5 * SIZE(CO1)
-
- MADD B7, B7, C34, ALPHA
- ST A7, 6 * SIZE(CO1)
-
- MADD C11, C11, C44, ALPHA
- ST A8, 7 * SIZE(CO1)
-
- ST B1, 0 * SIZE(CO2)
- ST B2, 1 * SIZE(CO2)
- ST B3, 2 * SIZE(CO2)
- ST B4, 3 * SIZE(CO2)
- ST B5, 4 * SIZE(CO2)
- ST B6, 5 * SIZE(CO2)
- ST B7, 6 * SIZE(CO2)
- ST C11, 7 * SIZE(CO2)
-
- daddiu CO1, CO1, 8 * SIZE
- bgtz I, .L281
- daddiu CO2, CO2, 8 * SIZE
- #else
- daddiu I, I, -1
-
- MUL A1, C11, ALPHA
- MUL A2, C21, ALPHA
- MUL A3, C31, ALPHA
- MUL A4, C41, ALPHA
- MUL A5, C13, ALPHA
- MUL A6, C23, ALPHA
- MUL A7, C33, ALPHA
- MUL A8, C43, ALPHA
-
- MUL B1, C12, ALPHA
- ST A1, 0 * SIZE(CO1)
-
- MUL B2, C22, ALPHA
- ST A2, 1 * SIZE(CO1)
-
- MUL B3, C32, ALPHA
- ST A3, 2 * SIZE(CO1)
-
- MUL B4, C42, ALPHA
- ST A4, 3 * SIZE(CO1)
-
- MUL B5, C14, ALPHA
- ST A5, 4 * SIZE(CO1)
-
- MUL B6, C24, ALPHA
- ST A6, 5 * SIZE(CO1)
-
- MUL B7, C34, ALPHA
- ST A7, 6 * SIZE(CO1)
-
- MUL C11, C44, ALPHA
- ST A8, 7 * SIZE(CO1)
-
- ST B1, 0 * SIZE(CO2)
- ST B2, 1 * SIZE(CO2)
- ST B3, 2 * SIZE(CO2)
- ST B4, 3 * SIZE(CO2)
- ST B5, 4 * SIZE(CO2)
- ST B6, 5 * SIZE(CO2)
- ST B7, 6 * SIZE(CO2)
- ST C11, 7 * SIZE(CO2)
-
- #if ( defined(LEFT) && defined(TRANSA)) ||(!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
- #ifdef LEFT
- daddiu TEMP, TEMP, -8
- #else
- daddiu TEMP, TEMP, -2
- #endif
- dsll L, TEMP, 3 + BASE_SHIFT
- dsll TEMP, TEMP, 1 + BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 8
- #endif
- daddiu CO1, CO1, 8 * SIZE
- bgtz I, .L281
- daddiu CO2, CO2, 8 * SIZE
- #endif
-
-
- .align 4
- .L24:
- andi I, M, 4 # MR=4
- blez I, .L22
- NOP
-
- .align 4
- .L241:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, 2 + BASE_SHIFT
- dsll TEMP, KK, 1 + BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C21, C11
- MOV C22, C11
- LD A2, 1 * SIZE(AO)
-
- MOV C31, C11
- MOV C32, C11
- LD A3, 2 * SIZE(AO)
-
- MOV C41, C11
- MOV C42, C11
- LD A4, 3 * SIZE(AO)
-
- MOV C13, C11
- MOV C14, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C23, C11
- MOV C24, C11
- LD B2, 1 * SIZE(BO)
-
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- MOV C44, C11
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 4
- #else
- daddiu TEMP, KK, 2
- #endif
- dsra L, TEMP, 1
- blez L, .L242
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 1 # UnRoll K=4
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C21, C11
- MOV C22, C11
- LD A2, 1 * SIZE(AO)
-
- MOV C31, C11
- MOV C32, C11
- LD A3, 2 * SIZE(AO)
-
- MOV C41, C11
- MOV C42, C11
- LD A4, 3 * SIZE(AO)
-
- MOV C13, C11
- MOV C14, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C23, C11
- MOV C24, C11
- LD B2, 1 * SIZE(BO)
-
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- blez L, .L242
- MOV C44, C11
- #endif
-
- .align 4
- .L2410:
- daddiu L, L, -1
- MADD C11, C11, A1, B1
- LD A5, 4 * SIZE(AO)
-
- MADD C21, C21, A2, B1
- LD B3, 2 * SIZE(BO)
-
- MADD C31, C31, A3, B1
- LD B4, 3 * SIZE(BO)
-
- MADD C41, C41, A4, B1
- LD A6, 5 * SIZE(AO)
- daddiu BO, BO, 4 * SIZE
-
- MADD C12, C12, A1, B2
- LD A7, 6 * SIZE(AO)
-
- MADD C22, C22, A2, B2
- LD A8, 7 * SIZE(AO)
- daddiu AO, AO, 8 * SIZE
-
- MADD C32, C32, A3, B2
- MADD C42, C42, A4, B2
-
- MADD C11, C11, A5, B3
- LD A1, 0 * SIZE(AO)
-
- MADD C21, C21, A6, B3
- LD B1, 0 * SIZE(BO)
-
- MADD C31, C31, A7, B3
- LD B2, 1 * SIZE(BO)
-
- MADD C41, C41, A8, B3
- LD A2, 1 * SIZE(AO)
-
- MADD C12, C12, A5, B4
- LD A3, 2 * SIZE(AO)
-
- MADD C22, C22, A6, B4
- LD A4, 3 * SIZE(AO)
-
- MADD C32, C32, A7, B4
- bgtz L, .L2410
- MADD C42, C42, A8, B4
-
- .align 4
- .L242:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L240
- LD ALPHA, 152($sp)
-
- MADD C11, C11, A1, B1
- MADD C21, C21, A2, B1
- MADD C31, C31, A3, B1
- MADD C41, C41, A4, B1
- MADD C12, C12, A1, B2
- MADD C22, C22, A2, B2
- MADD C32, C32, A3, B2
- MADD C42, C42, A4, B2
- daddiu AO, AO, 4 * SIZE
- daddiu BO, BO, 2 * SIZE
-
-
- .align 4
- .L240: # Write Back
- #ifndef TRMMKERNEL
- LD A1, 0 * SIZE(CO1)
- LD A2, 1 * SIZE(CO1)
- LD A3, 2 * SIZE(CO1)
- LD A4, 3 * SIZE(CO1)
-
- MADD A1, A1, C11, ALPHA
- LD B1, 0 * SIZE(CO2)
-
- MADD A2, A2, C21, ALPHA
- LD B2, 1 * SIZE(CO2)
-
- MADD A3, A3, C31, ALPHA
- LD B3, 2 * SIZE(CO2)
-
- MADD A4, A4, C41, ALPHA
- LD B4, 3 * SIZE(CO2)
-
- MADD B1, B1, C12, ALPHA
- ST A1, 0 * SIZE(CO1)
-
- MADD B2, B2, C22, ALPHA
- ST A2, 1 * SIZE(CO1)
-
- MADD B3, B3, C32, ALPHA
- ST A3, 2 * SIZE(CO1)
-
- MADD B4, B4, C42, ALPHA
- ST A4, 3 * SIZE(CO1)
-
- ST B1, 0 * SIZE(CO2)
- ST B2, 1 * SIZE(CO2)
- ST B3, 2 * SIZE(CO2)
- ST B4, 3 * SIZE(CO2)
-
- daddiu CO1, CO1, 4 * SIZE
- daddiu CO2, CO2, 4 * SIZE
- #else
-
- MUL A1, C11, ALPHA
- MUL A2, C21, ALPHA
- MUL A3, C31, ALPHA
- MUL A4, C41, ALPHA
-
- MUL B1, C12, ALPHA
- ST A1, 0 * SIZE(CO1)
-
- MUL B2, C22, ALPHA
- ST A2, 1 * SIZE(CO1)
-
- MUL B3, C32, ALPHA
- ST A3, 2 * SIZE(CO1)
-
- MUL B4, C42, ALPHA
- ST A4, 3 * SIZE(CO1)
-
- ST B1, 0 * SIZE(CO2)
- ST B2, 1 * SIZE(CO2)
- ST B3, 2 * SIZE(CO2)
- ST B4, 3 * SIZE(CO2)
-
- daddiu CO1, CO1, 4 * SIZE
- daddiu CO2, CO2, 4 * SIZE
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
- #ifdef LEFT
- daddiu TEMP, TEMP, -4
- #else
- daddiu TEMP, TEMP, -2
- #endif
- dsll L, TEMP, 2 + BASE_SHIFT
- dsll TEMP, TEMP, 1 + BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 4
- #endif
- #endif
-
- .align 4
- .L22:
- andi I, M, 2
- blez I, .L21
- NOP
-
- .align 4
- .L221:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, 1 + BASE_SHIFT
- dsll TEMP, KK, 1 + BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C21, C11
- MOV C22, C11
- LD A2, 1 * SIZE(AO)
-
- MOV C31, C11
- MOV C32, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C41, C11
- MOV C42, C11
- LD B2, 1 * SIZE(BO)
-
- MOV C43, C11
- MOV C44, C11
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 2
- #else
- daddiu TEMP, KK, 2
- #endif
- dsra L, TEMP, 1
- blez L, .L222
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 1 # UnRoll K=4
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C21, C11
- MOV C22, C11
- LD A2, 1 * SIZE(AO)
-
- MOV C31, C11
- MOV C32, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C41, C11
- MOV C42, C11
- LD B2, 1 * SIZE(BO)
-
- MOV C43, C11
- blez L, .L222
- MOV C44, C11
- #endif
-
-
- .align 4
- .L2210:
- daddiu L, L, -1
- MADD C11, C11, A1, B1
- LD A3, 2 * SIZE(AO)
-
- MADD C21, C21, A2, B1
- LD B3, 2 * SIZE(BO)
-
- MADD C12, C12, A1, B2
- LD A4, 3 * SIZE(AO)
- daddiu AO, AO, 4 * SIZE
-
- MADD C22, C22, A2, B2
- LD B4, 3 * SIZE(BO)
- daddiu BO, BO, 4 * SIZE
-
- MADD C11, C11, A3, B3
- LD A1, 0 * SIZE(AO)
-
- MADD C21, C21, A4, B3
- LD B1, 0 * SIZE(BO)
-
- MADD C12, C12, A3, B4
- LD B2, 1 * SIZE(BO)
-
- MADD C22, C22, A4, B4
- bgtz L, .L2210
- LD A2, 1 * SIZE(AO)
-
-
- .align 4
- .L222:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L220
- LD ALPHA, 152($sp)
-
- MADD C11, C11, A1, B1
- MADD C21, C21, A2, B1
- MADD C12, C12, A1, B2
- MADD C22, C22, A2, B2
- daddiu AO, AO, 2 * SIZE
- daddiu BO, BO, 2 * SIZE
-
-
- .align 4
- .L220: # Write Back
- #ifndef TRMMKERNEL
- LD A1, 0 * SIZE(CO1)
- LD A2, 1 * SIZE(CO1)
-
- MADD A1, A1, C11, ALPHA
- LD B1, 0 * SIZE(CO2)
-
- MADD A2, A2, C21, ALPHA
- LD B2, 1 * SIZE(CO2)
-
- MADD B1, B1, C12, ALPHA
- ST A1, 0 * SIZE(CO1)
-
- MADD B2, B2, C22, ALPHA
- ST A2, 1 * SIZE(CO1)
-
- ST B1, 0 * SIZE(CO2)
- ST B2, 1 * SIZE(CO2)
-
- daddiu CO1, CO1, 2 * SIZE
- daddiu CO2, CO2, 2 * SIZE
- #else
-
- MUL A1, C11, ALPHA
- MUL A2, C21, ALPHA
- MUL B1, C12, ALPHA
- MUL B2, C22, ALPHA
-
- ST A1, 0 * SIZE(CO1)
- ST A2, 1 * SIZE(CO1)
- ST B1, 0 * SIZE(CO2)
- ST B2, 1 * SIZE(CO2)
-
- daddiu CO1, CO1, 2 * SIZE
- daddiu CO2, CO2, 2 * SIZE
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
- #ifdef LEFT
- daddiu TEMP, TEMP, -2
- #else
- daddiu TEMP, TEMP, -2
- #endif
- dsll L, TEMP, 1 + BASE_SHIFT
- dsll TEMP, TEMP, 1 + BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddu KK, KK, 2
- #endif
- #endif
-
- .align 4
- .L21:
- andi I, M, 1
- blez I, .L20
- NOP
-
- .align 4
- .L211:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B # Reset B
- #else
- dsll L, KK, BASE_SHIFT
- dsll TEMP, KK, 1 + BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C21, C11
- MOV C22, C11
-
- MOV C31, C11
- MOV C32, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C41, C11
- MOV C42, C11
- LD B2, 1 * SIZE(BO)
-
- MOV C43, C11
- MOV C44, C11
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 1
- #else
- daddiu TEMP, KK, 2
- #endif
- dsra L, TEMP, 1
- blez L, .L212
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 1 # UnRoll K=4
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C21, C11
- MOV C22, C11
-
- MOV C31, C11
- MOV C32, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C41, C11
- MOV C42, C11
- LD B2, 1 * SIZE(BO)
-
- MOV C43, C11
- blez L, .L212
- MOV C44, C11
- #endif
-
- .align 4
- .L2110:
- daddiu L, L, -1
- MADD C11, C11, A1, B1
- LD A2, 1 * SIZE(AO)
-
- MADD C12, C12, A1, B2
- LD B3, 2 * SIZE(BO)
-
- LD B4, 3 * SIZE(BO)
- daddiu AO, AO, 2 * SIZE
- daddiu BO, BO, 4 * SIZE
-
- MADD C11, C11, A2, B3
- LD A1, 0 * SIZE(AO)
-
- MADD C12, C12, A2, B4
- LD B1, 0 * SIZE(BO)
-
- bgtz L, .L2110
- LD B2, 1 * SIZE(BO)
-
-
- .align 4
- .L212:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L210
- LD ALPHA, 152($sp)
-
- MADD C11, C11, A1, B1
- MADD C12, C12, A1, B2
- daddiu AO, AO, 1 * SIZE
- daddiu BO, BO, 2 * SIZE
-
-
- .align 4
- .L210: # Write Back
- #ifndef TRMMKERNEL
- LD A1, 0 * SIZE(CO1)
-
- MADD A1, A1, C11, ALPHA
- LD B1, 0 * SIZE(CO2)
-
- MADD B1, B1, C12, ALPHA
- ST A1, 0 * SIZE(CO1)
-
- ST B1, 0 * SIZE(CO2)
-
- daddiu CO1, CO1, 1 * SIZE
- daddiu CO2, CO2, 1 * SIZE
- #else
-
- MUL A1, C11, ALPHA
- MUL B1, C12, ALPHA
-
- ST A1, 0 * SIZE(CO1)
- ST B1, 0 * SIZE(CO2)
-
- daddiu CO1, CO1, 1 * SIZE
- daddiu CO2, CO2, 1 * SIZE
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
- #ifdef LEFT
- daddiu TEMP, TEMP, -1
- #else
- daddiu TEMP, TEMP, -2
- #endif
- dsll L, TEMP, BASE_SHIFT
- dsll TEMP, TEMP, 1 + BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 1
- #endif
- #endif
-
-
- .align 4
- .L20:
- #if defined(TRMMKERNEL) && !defined(LEFT)
- daddiu KK, KK, 2
- #endif
- move B, BO
-
-
-
- .align 4
- .L1:
- andi J, N, 1
- blez J, .L999
- NOP
-
- .L18:
- dsra I, M, 3 # MR=8
- move AO, A # Reset A
-
- #if defined(TRMMKERNEL) && defined(LEFT)
- move KK, OFFSET
- #endif
- blez I, .L14
- NOP
-
-
- .align 4
- .L181:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B # Reset B
- #else
- dsll L, KK, 3 + BASE_SHIFT
- dsll TEMP, KK, BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- LD A1, 0 * SIZE(AO)
-
- MOV C12, C11
- LD A2, 1 * SIZE(AO)
-
- MOV C21, C11
- LD A3, 2 * SIZE(AO)
-
- MOV C22, C11
- LD A4, 3 * SIZE(AO)
-
- MOV C31, C11
- LD A5, 4 * SIZE(AO)
-
- MOV C32, C11
- LD A6, 5 * SIZE(AO)
-
- MOV C41, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C42, C11
- LD A7, 6 * SIZE(AO)
-
- MOV C13, C11
- LD A8, 7 * SIZE(AO)
-
- MOV C14, C11
-
- MOV C23, C11
- MOV C24, C11
-
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- MOV C44, C11
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 8
- #else
- daddiu TEMP, KK, 1
- #endif
- dsra L, TEMP, 1
- blez L, .L182
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 1 # UnRoll K=4
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- LD A1, 0 * SIZE(AO)
-
- MOV C12, C11
- LD A2, 1 * SIZE(AO)
-
- MOV C21, C11
- LD A3, 2 * SIZE(AO)
-
- MOV C22, C11
- LD A4, 3 * SIZE(AO)
-
- MOV C31, C11
- LD A5, 4 * SIZE(AO)
-
- MOV C32, C11
- LD A6, 5 * SIZE(AO)
-
- MOV C41, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C42, C11
- LD A7, 6 * SIZE(AO)
-
- MOV C13, C11
- LD A8, 7 * SIZE(AO)
-
- MOV C14, C11
-
- MOV C23, C11
- MOV C24, C11
-
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- blez L, .L182
- MOV C44, C11
- #endif
-
-
- .align 4
- .L1810:
- daddiu L, L, -1
- MADD C11, C11, A1, B1
- LD B5, 8 * SIZE(AO)
-
- MADD C21, C21, A2, B1
- LD B6, 9 * SIZE(AO)
-
- MADD C31, C31, A3, B1
- LD B7, 10 * SIZE(AO)
-
- MADD C41, C41, A4, B1
- LD B8, 11 * SIZE(AO)
-
- MADD C13, C13, A5, B1
- LD B2, 1 * SIZE(BO)
- daddiu BO, BO, 2 * SIZE
-
- MADD C23, C23, A6, B1
- LD A1, 12 * SIZE(AO)
-
- MADD C33, C33, A7, B1
- LD A2, 13 * SIZE(AO)
-
- MADD C43, C43, A8, B1
- LD A3, 14 * SIZE(AO)
-
- LD A4, 15 * SIZE(AO)
- daddiu AO, AO, 16 * SIZE
-
- MADD C11, C11, B5, B2
- LD A5, 4 * SIZE(AO)
-
- MADD C21, C21, B6, B2
- LD A6, 5 * SIZE(AO)
-
- MADD C13, C13, A1, B2
- LD A7, 6 * SIZE(AO)
-
- MADD C23, C23, A2, B2
- LD A8, 7 * SIZE(AO)
-
- MADD C33, C33, A3, B2
- LD B1, 0 * SIZE(BO)
-
- MADD C43, C43, A4, B2
- LD A1, 0 * SIZE(AO)
-
- MADD C31, C31, B7, B2
- LD A2, 1 * SIZE(AO)
-
- MADD C41, C41, B8, B2
- LD A3, 2 * SIZE(AO)
-
- bgtz L, .L1810
- LD A4, 3 * SIZE(AO)
-
- .align 4
- .L182:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L180
- LD ALPHA, 152($sp)
-
- MADD C13, C13, A5, B1
- MADD C23, C23, A6, B1
- MADD C33, C33, A7, B1
- MADD C43, C43, A8, B1
- daddiu AO, AO, 8 * SIZE
-
- MADD C11, C11, A1, B1
- MADD C21, C21, A2, B1
- MADD C31, C31, A3, B1
- MADD C41, C41, A4, B1
- daddiu BO, BO, 1 * SIZE
-
-
- .align 4
- .L180: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
-
- LD A1, 0 * SIZE(C)
- LD A2, 1 * SIZE(C)
- LD A3, 2 * SIZE(C)
- LD A4, 3 * SIZE(C)
- LD A5, 4 * SIZE(C)
- LD A6, 5 * SIZE(C)
- LD A7, 6 * SIZE(C)
- LD A8, 7 * SIZE(C)
-
- MADD A1, A1, C11, ALPHA
- MADD A2, A2, C21, ALPHA
- MADD A3, A3, C31, ALPHA
- MADD A4, A4, C41, ALPHA
- MADD A5, A5, C13, ALPHA
- MADD A6, A6, C23, ALPHA
- MADD A7, A7, C33, ALPHA
- MADD A8, A8, C43, ALPHA
-
- ST A1, 0 * SIZE(C)
- ST A2, 1 * SIZE(C)
- ST A3, 2 * SIZE(C)
- ST A4, 3 * SIZE(C)
- ST A5, 4 * SIZE(C)
- ST A6, 5 * SIZE(C)
- ST A7, 6 * SIZE(C)
- ST A8, 7 * SIZE(C)
-
- daddiu C, C, 8 * SIZE
- bgtz I, .L181
- NOP
- #else
- daddiu I, I, -1
-
- MUL A1, C11, ALPHA
- MUL A2, C21, ALPHA
- MUL A3, C31, ALPHA
- MUL A4, C41, ALPHA
- MUL A5, C13, ALPHA
- MUL A6, C23, ALPHA
- MUL A7, C33, ALPHA
- MUL A8, C43, ALPHA
-
- ST A1, 0 * SIZE(C)
- ST A2, 1 * SIZE(C)
- ST A3, 2 * SIZE(C)
- ST A4, 3 * SIZE(C)
- ST A5, 4 * SIZE(C)
- ST A6, 5 * SIZE(C)
- ST A7, 6 * SIZE(C)
- ST A8, 7 * SIZE(C)
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
-
- #ifdef LEFT
- daddiu TEMP, TEMP, -8
- #else
- daddiu TEMP, TEMP, -1
- #endif
-
- dsll L, TEMP, 3 + BASE_SHIFT
- dsll TEMP, TEMP, BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 8
- #endif
-
- daddiu C, C, 8 * SIZE
- bgtz I, .L181
- NOP
- #endif
-
- .align 4
- .L14:
- andi I, M, 4 # MR=4
- blez I, .L12
- NOP
-
- .align 4
- .L141:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, 2 + BASE_SHIFT
- dsll TEMP, KK, BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C21, C11
- MOV C22, C11
- LD A2, 1 * SIZE(AO)
-
- MOV C31, C11
- MOV C32, C11
- LD A3, 2 * SIZE(AO)
-
- MOV C41, C11
- MOV C42, C11
- LD A4, 3 * SIZE(AO)
-
- MOV C13, C11
- MOV C14, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C23, C11
- MOV C24, C11
-
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- MOV C44, C11
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 4
- #else
- daddiu TEMP, KK, 1
- #endif
- dsra L, TEMP, 1
- blez L, .L142
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 1 # UnRoll K=4
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C21, C11
- MOV C22, C11
- LD A2, 1 * SIZE(AO)
-
- MOV C31, C11
- MOV C32, C11
- LD A3, 2 * SIZE(AO)
-
- MOV C41, C11
- MOV C42, C11
- LD A4, 3 * SIZE(AO)
-
- MOV C13, C11
- MOV C14, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C23, C11
- MOV C24, C11
-
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- blez L, .L142
- MOV C44, C11
- #endif
-
- .align 4
- .L1410:
- daddiu L, L, -1
- MADD C11, C11, A1, B1
- LD A5, 4 * SIZE(AO)
-
- MADD C21, C21, A2, B1
- LD B3, 1 * SIZE(BO)
-
- MADD C31, C31, A3, B1
- LD A6, 5 * SIZE(AO)
- daddiu BO, BO, 2 * SIZE
-
- MADD C41, C41, A4, B1
- LD A7, 6 * SIZE(AO)
-
- LD A8, 7 * SIZE(AO)
- daddiu AO, AO, 8 * SIZE
-
-
- MADD C11, C11, A5, B3
- LD A1, 0 * SIZE(AO)
-
- MADD C21, C21, A6, B3
- LD B1, 0 * SIZE(BO)
-
- MADD C31, C31, A7, B3
- LD A2, 1 * SIZE(AO)
-
- MADD C41, C41, A8, B3
- LD A3, 2 * SIZE(AO)
-
- bgtz L, .L1410
- LD A4, 3 * SIZE(AO)
-
- .align 4
- .L142:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L140
- LD ALPHA, 152($sp)
-
- MADD C11, C11, A1, B1
- MADD C21, C21, A2, B1
- MADD C31, C31, A3, B1
- MADD C41, C41, A4, B1
- daddiu AO, AO, 4 * SIZE
- daddiu BO, BO, 1 * SIZE
-
-
- .align 4
- .L140: # Write Back
- #ifndef TRMMKERNEL
- LD A1, 0 * SIZE(C)
- LD A2, 1 * SIZE(C)
- LD A3, 2 * SIZE(C)
- LD A4, 3 * SIZE(C)
-
- MADD A1, A1, C11, ALPHA
- MADD A2, A2, C21, ALPHA
- MADD A3, A3, C31, ALPHA
- MADD A4, A4, C41, ALPHA
-
- ST A1, 0 * SIZE(C)
- ST A2, 1 * SIZE(C)
- ST A3, 2 * SIZE(C)
- ST A4, 3 * SIZE(C)
- daddiu C, C, 4 * SIZE
- #else
- MUL A1, C11, ALPHA
- MUL A2, C21, ALPHA
- MUL A3, C31, ALPHA
- MUL A4, C41, ALPHA
-
- ST A1, 0 * SIZE(C)
- ST A2, 1 * SIZE(C)
- ST A3, 2 * SIZE(C)
- ST A4, 3 * SIZE(C)
- daddiu C, C, 4 * SIZE
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
- #ifdef LEFT
- daddiu TEMP, TEMP, -4
- #else
- daddiu TEMP, TEMP, -1
- #endif
-
- dsll L, TEMP, 2 + BASE_SHIFT
- dsll TEMP, TEMP, BASE_SHIFT
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 4
- #endif
- #endif
-
- .align 4
- .L12:
- andi I, M, 2
- blez I, .L11
- NOP
-
- .align 4
- .L121:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) ||\
- (!defined(LEFT) && !defined(TRANSA))
- move BO, B # Reset B
- #else
- dsll L, KK, 1 + BASE_SHIFT
- dsll TEMP, KK, BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C21, C11
- MOV C22, C11
- LD A2, 1 * SIZE(AO)
-
- MOV C31, C11
- MOV C32, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C41, C11
- MOV C42, C11
-
- MOV C43, C11
- MOV C44, C11
- #if (defined(LEFT) && !defined(TRANSA)) ||\
- (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 2
- #else
- daddiu TEMP, KK, 1
- #endif
- dsra L, TEMP, 1
- blez L, .L122
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 1 # UnRoll K=4
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C21, C11
- MOV C22, C11
- LD A2, 1 * SIZE(AO)
-
- MOV C31, C11
- MOV C32, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C41, C11
- MOV C42, C11
-
- MOV C43, C11
- blez L, .L122
- MOV C44, C11
- #endif
-
- .align 4
- .L1210:
- daddiu L, L, -1
- MADD C11, C11, A1, B1
- LD B3, 1 * SIZE(BO)
-
- MADD C21, C21, A2, B1
- daddiu BO, BO, 2 * SIZE
-
- LD A3, 2 * SIZE(AO)
- LD A4, 3 * SIZE(AO)
- daddiu AO, AO, 4 * SIZE
-
- MADD C11, C11, A3, B3
- LD B1, 0 * SIZE(BO)
-
- MADD C21, C21, A4, B3
- LD A1, 0 * SIZE(AO)
- bgtz L, .L1210
- LD A2, 1 * SIZE(AO)
-
-
- .align 4
- .L122:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L120
- LD ALPHA, 152($sp)
-
- MADD C11, C11, A1, B1
- MADD C21, C21, A2, B1
- daddiu AO, AO, 2 * SIZE
- daddiu BO, BO, 1 * SIZE
-
-
- .align 4
- .L120: # Write Back
- #ifndef TRMMKERNEL
- LD A1, 0 * SIZE(C)
- LD A2, 1 * SIZE(C)
-
- MADD A1, A1, C11, ALPHA
- MADD A2, A2, C21, ALPHA
-
- ST A1, 0 * SIZE(C)
- ST A2, 1 * SIZE(C)
-
- daddiu C, C, 2 * SIZE
- #else
- MUL A1, C11, ALPHA
- MUL A2, C21, ALPHA
-
- ST A1, 0 * SIZE(C)
- ST A2, 1 * SIZE(C)
-
- daddiu C, C, 2 * SIZE
- #if ( defined(LEFT) && defined(TRANSA))||\
- (!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
- #ifdef LEFT
- daddiu TEMP, TEMP, -2
- #else
- daddiu TEMP, TEMP, -1
- #endif
- dsll L, TEMP, 1 + BASE_SHIFT
- dsll TEMP, TEMP, BASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 2
- #endif
- #endif
-
- .align 4
- .L11:
- andi I, M, 1
- blez I, .L10
- NOP
-
- .align 4
- .L111:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA))||\
- (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, BASE_SHIFT
- daddu AO, AO, L
- daddu BO, B, L
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C21, C11
- MOV C22, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C31, C11
- MOV C32, C11
- #if (defined(LEFT) && !defined(TRANSA))||\
- (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 1
- #else
- daddiu TEMP, KK, 1
- #endif
- dsra L, TEMP, 1
- blez L, .L112
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 1 # UnRoll K=4
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
-
- MOV C21, C11
- MOV C22, C11
- LD B1, 0 * SIZE(BO)
-
- MOV C31, C11
- blez L, .L112
- MOV C32, C11
- #endif
-
-
- .align 4
- .L1110:
- daddiu L, L, -1
- MADD C11, C11, A1, B1
-
- LD A2, 1 * SIZE(AO)
- LD B2, 1 * SIZE(BO)
-
- daddiu AO, AO, 2 * SIZE
- daddiu BO, BO, 2 * SIZE
-
- MADD C11, C11, A2, B2
- LD A1, 0 * SIZE(AO)
- LD B1, 0 * SIZE(BO)
-
- bgtz L, .L1110
- NOP
-
-
- .align 4
- .L112:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L110
- LD ALPHA, 152($sp)
-
- MADD C11, C11, A1, B1
- daddiu AO, AO, 1 * SIZE
- daddiu BO, BO, 1 * SIZE
-
-
- .align 4
- .L110: # Write Back
- #ifndef TRMMKERNEL
- LD A1, 0 * SIZE(C)
-
- MADD A1, A1, C11, ALPHA
-
- ST A1, 0 * SIZE(C)
-
- daddiu C, C, 1 * SIZE
- #else
- MUL A1, C11, ALPHA
-
- ST A1, 0 * SIZE(C)
-
- daddiu C, C, 1 * SIZE
-
- #endif
-
- .align 4
- .L10:
- move B, BO
- NOP
-
- .L999:
- ld $16, 0($sp)
- ld $17, 8($sp)
- ld $18, 16($sp)
- ld $19, 24($sp)
- ld $20, 32($sp)
- ld $21, 40($sp)
- ld $22, 48($sp)
-
- ldc1 $f24, 56($sp)
- ldc1 $f25, 64($sp)
- ldc1 $f26, 72($sp)
- ldc1 $f27, 80($sp)
- ldc1 $f28, 88($sp)
-
- #if defined(TRMMKERNEL)
- ld $23, 96($sp)
- ld $24, 104($sp)
- ld $25, 112($sp)
- #endif
-
- #ifndef __64BIT__
- ldc1 $f20,120($sp)
- ldc1 $f21,128($sp)
- ldc1 $f22,136($sp)
- ldc1 $f23,144($sp)
- #endif
-
- daddiu $sp,$sp,STACKSIZE
- j $31
- nop
-
- EPILOGUE
- # .set macro
- # .set reorder
- # .end gemm
- # .size gemm, .-gemm
- # .ident "GCC: (Debian 4.4.6-6) 4.4.6"
|