|
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667266826692670267126722673267426752676267726782679268026812682268326842685268626872688268926902691269226932694269526962697269826992700270127022703270427052706270727082709271027112712271327142715271627172718271927202721272227232724272527262727272827292730273127322733273427352736273727382739274027412742274327442745274627472748274927502751275227532754275527562757275827592760276127622763276427652766276727682769277027712772277327742775277627772778277927802781278227832784278527862787278827892790279127922793279427952796279727982799280028012802280328042805280628072808280928102811281228132814281528162817281828192820282128222823282428252826282728282829283028312832283328342835283628372838283928402841284228432844284528462847284828492850285128522853285428552856285728582859286028612862286328642865286628672868286928702871287228732874287528762877287828792880288128822883288428852886288728882889289028912892289328942895289628972898289929002901290229032904290529062907290829092910291129122913291429152916291729182919292029212922292329242925292629272928292929302931293229332934293529362937293829392940294129422943294429452946294729482949295029512952295329542955295629572958295929602961296229632964296529662967296829692970297129722973297429752976297729782979298029812982298329842985298629872988298929902991299229932994299529962997299829993000300130023003300430053006300730083009301030113012301330143015301630173018301930203021302230233024302530263027302830293030303130323033303430353036303730383039304030413042304330443045304630473048304930503051305230533054305530563057305830593060306130623063306430653066306730683069307030713072307330743075307630773078307930803081308230833084308530863087308830893090309130923093309430953096309730983099310031013102310331043105310631073108310931103111311231133114311531163117311831193120312131223123312431253126312731283129313031313132313331343135313631373138313931403141314231433144314531463147314831493150315131523153315431553156315731583159316031613162316331643165316631673168316931703171317231733174317531763177317831793180318131823183318431853186318731883189319031913192319331943195319631973198319932003201320232033204320532063207320832093210321132123213321432153216321732183219322032213222322332243225322632273228322932303231323232333234323532363237323832393240324132423243324432453246324732483249325032513252325332543255325632573258325932603261326232633264326532663267326832693270327132723273327432753276327732783279328032813282328332843285328632873288328932903291329232933294329532963297329832993300330133023303330433053306330733083309331033113312331333143315331633173318331933203321332233233324332533263327332833293330333133323333333433353336333733383339334033413342334333443345334633473348334933503351335233533354335533563357335833593360336133623363336433653366336733683369337033713372337333743375337633773378337933803381338233833384338533863387338833893390339133923393339433953396339733983399340034013402340334043405340634073408340934103411341234133414341534163417341834193420342134223423342434253426342734283429343034313432343334343435343634373438343934403441344234433444344534463447344834493450345134523453345434553456345734583459346034613462346334643465346634673468346934703471347234733474347534763477347834793480348134823483348434853486348734883489349034913492349334943495349634973498349935003501350235033504350535063507350835093510351135123513351435153516351735183519352035213522352335243525352635273528352935303531353235333534353535363537353835393540354135423543354435453546354735483549355035513552355335543555355635573558355935603561356235633564356535663567356835693570357135723573357435753576357735783579358035813582358335843585358635873588358935903591359235933594359535963597359835993600360136023603360436053606360736083609361036113612361336143615361636173618361936203621362236233624362536263627362836293630363136323633363436353636363736383639364036413642364336443645364636473648364936503651365236533654365536563657365836593660366136623663366436653666366736683669367036713672367336743675367636773678367936803681368236833684368536863687368836893690369136923693369436953696369736983699370037013702370337043705370637073708370937103711371237133714371537163717371837193720372137223723372437253726372737283729373037313732373337343735373637373738373937403741374237433744374537463747374837493750375137523753375437553756375737583759376037613762376337643765376637673768376937703771377237733774377537763777377837793780378137823783378437853786378737883789379037913792379337943795379637973798379938003801380238033804380538063807380838093810381138123813381438153816381738183819382038213822382338243825382638273828382938303831383238333834383538363837383838393840384138423843384438453846384738483849385038513852385338543855385638573858385938603861386238633864386538663867386838693870387138723873387438753876387738783879388038813882388338843885388638873888388938903891389238933894389538963897389838993900390139023903390439053906390739083909391039113912391339143915391639173918391939203921392239233924392539263927392839293930393139323933393439353936393739383939394039413942394339443945394639473948394939503951395239533954395539563957395839593960396139623963396439653966396739683969397039713972397339743975397639773978397939803981398239833984398539863987398839893990399139923993399439953996399739983999400040014002400340044005400640074008400940104011401240134014401540164017401840194020402140224023402440254026 |
- ##define REALNAME gemm
- #define ASSEMBLER
- #include "common.h"
-
- #define FETCH ld
- #define STACKSIZE 192
- #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 $9
- #define B $10
- #define C $11
- #define LDC $8
-
- #### Pointer A, B, C ####
- #define AO $12
- #define BO $13
-
- #define CO1 $14
- #define CO2 $15
-
- #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 $25
- #endif
-
-
- PROLOGUE
-
- LDARG LDC, 0($sp)
- 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, STACKSIZE+8($sp)
- #endif
-
- #ifndef __64BIT__
- sdc1 $f20,120($sp)
- sdc1 $f21,128($sp)
- sdc1 $f22,136($sp)
- sdc1 $f23,144($sp)
- #endif
-
- .align 4
- .L2:
- dsra J, N, 1 # NR=2
- ST $f15, 152($sp)
-
- #if defined(TRMMKERNEL) && !defined(LEFT)
- neg KK, OFFSET
- #endif
-
- dsll LDC, LDC, ZBASE_SHIFT# LDC*SIZE
- blez J, .L1
- ST $f16, 160($sp)
-
- .L24:
- #if defined(TRMMKERNEL) && defined(LEFT)
- move KK, OFFSET
- #endif
-
- dsra I, M, 2 # MR=8
- move AO, A # Reset A
-
- dsll PREA, K, 1 + ZBASE_SHIFT
- move CO1, C
-
- daddu CO2, C, LDC
- daddu PREA, AO, PREA
-
- blez I, .L22
- daddu C, CO2, LDC
-
- .align 4
- .L241:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, 2 + ZBASE_SHIFT
- dsll TEMP, KK, 1 + ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- dsll PREB, K, ZBASE_SHIFT
- MOV C21, C11
- MOV C22, C11
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MOV C31, C11
- MOV C32, C11
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C41, C11
- MOV C42, C11
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MOV C13, C11
- MOV C14, C11
-
- MOV C23, C11
- MOV C24, C11
-
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- MOV C44, C11
-
- PLU B3, B1, B1
- PLU B4, B2, B2
- daddu PREB, BO, PREB
-
- FETCH $0, 0 * SIZE(CO1)
- FETCH $0, 8 * SIZE(CO1)
- FETCH $0, 0 * SIZE(CO2)
- FETCH $0, 8 * SIZE(CO2)
- #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, 2
- blez L, .L242
- NOP
-
- #else
-
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=64
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- dsll PREB, K, ZBASE_SHIFT
- MOV C21, C11
- MOV C22, C11
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MOV C31, C11
- MOV C32, C11
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C41, C11
- MOV C42, C11
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MOV C13, C11
- MOV C14, C11
-
- FETCH $0, 0 * SIZE(CO1)
- MOV C23, C11
- MOV C24, C11
-
- FETCH $0, 0 * SIZE(CO2)
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- MOV C44, C11
- daddu PREB, BO, PREB
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- FETCH $0, 8 * SIZE(CO1)
- blez L, .L242
- FETCH $0, 8 * SIZE(CO2)
- #endif
-
- .L2410:
- daddiu L, L, -1
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F5, F4, 2) # A5 A6
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- gsLQC1(R12, F7, F6, 3) # A7 A8
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
-
- FETCH $0, 0 * SIZE(PREB)
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
-
- FETCH $0, 0 * SIZE(PREA)
- 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
- MADPS C24, C24, A2, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
- daddu PREB, PREB, 8 * SIZE
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- gsLQC1(R13, F9, F8, 2) # B1 B2
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
-
- gsLQC1(R12, F1, F0, 4) # A1 A2
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
-
- gsLQC1(R12, F3, F2, 5) # A3 A4
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
-
- FETCH $0, 8 * SIZE(PREA)
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
-
- 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
- MADPS C24, C24, A6, B8
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- gsLQC1(R13, F13, F12, 3) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F5, F4, 6) # A5 A6
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- gsLQC1(R12, F7, F6, 7) # A7 A8
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- daddiu BO, BO, 4 * 4 * SIZE # 4KR*4NR
-
- FETCH $0, 16 * SIZE(PREA)
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- daddiu AO, AO, 8 * 4 * SIZE # 4KR*8MR
-
- 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
- MADPS C24, C24, A2, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
-
- FETCH $0, 24 * SIZE(PREA)
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddu PREA, PREA, 32 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- MADPS C24, C24, A6, B8
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- MADPS C34, C34, A7, B8
- bgtz L, .L2410
- MADPS C44, C44, A8, B8
-
-
- .align 4
- .L242:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L247
- NOP
-
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F5, F4, 2) # A5 A6
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- gsLQC1(R12, F7, F6, 3) # A7 A8
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- daddiu BO, BO, 2 * 4 * SIZE # 4KR*4NR
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- daddiu AO, AO, 4 * 4 * SIZE
-
- 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
- MADPS C24, C24, A2, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
-
- 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
- MADPS C24, C24, A6, B8
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- .align 4
- .L247:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L240
- NOP
-
- 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
- daddiu BO, BO, 1 * 4 * SIZE # 4KR*4NR
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- daddiu AO, AO, 2 * 4 * SIZE
-
- 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
- MADPS C24, C24, A2, B4
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
-
- .align 4
- .L240: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C31
- CVTU A4, C41
-
- CVTU A5, C13
- CVTU A6, C23
-
- CVTU A7, C33
- CVTU A8, C43
-
- CVTU B1, C12
- CVTU B2, C22
-
- CVTU B3, C32
- CVTU B4, C42
-
- CVTU B5, C14
- CVTU B6, C24
-
- CVTU B7, C34
- CVTU B8, C44
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- # LD A1, 0 * SIZE(A) # load alpha_r
- SUB C31, C31, A3
- LD A1, 152($sp) # load alpha_r
- SUB C41, C41, A4
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- ADD C13, A5, C13 # ad'+'cb
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
- SUB C12, C12, B1
- SUB C22, C22, B2
- SUB C32, C32, B3
- SUB C42, C42, B4
- ADD C14, B5, C14
- ADD C24, B6, C24
- ADD C34, B7, C34
- ADD C44, B8, C44
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- LD C13, 0 * SIZE(CO2)
- LD C23, 2 * SIZE(CO2)
- LD C33, 4 * SIZE(CO2)
- LD C43, 6 * SIZE(CO2)
- LD C11, 1 * SIZE(CO2)
- LD C21, 3 * SIZE(CO2)
- LD C31, 5 * SIZE(CO2)
- LD C41, 7 * SIZE(CO2)
-
- MADD C13, C13, C12, A1
- MADD C23, C23, C22, A1
-
- MADD C33, C33, C32, A1
- ST B1, 0 * SIZE(CO1)
-
- MADD C43, C43, C42, A1
- ST B3, 2 * SIZE(CO1)
-
- MADD C11, C11, C14, A1
- ST B5, 4 * SIZE(CO1)
-
- MADD C21, C21, C24, A1
- ST B7, 6 * SIZE(CO1)
-
- MADD C31, C31, C34, A1
- ST B2, 1 * SIZE(CO1)
-
- MADD C41, C41, C44, A1
- ST B4, 3 * SIZE(CO1)
-
- NMSUB C13, C13, C14, A2
- ST B6, 5 * SIZE(CO1)
-
- NMSUB C23, C23, C24, A2
- ST B8, 7 * SIZE(CO1)
-
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- SUB C13, A5, C13 # ad'+'cb
- SUB C23, A6, C23
- SUB C33, A7, C33
- SUB C43, A8, C43
- ADD C12, B1, C12
- ADD C22, B2, C22
- ADD C32, B3, C32
- ADD C42, B4, C42
- SUB C14, B5, C14
- SUB C24, B6, C24
- SUB C34, B7, C34
- SUB C44, B8, C44
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- LD C13, 0 * SIZE(CO2)
- LD C23, 2 * SIZE(CO2)
- LD C33, 4 * SIZE(CO2)
- LD C43, 6 * SIZE(CO2)
- LD C11, 1 * SIZE(CO2)
- LD C21, 3 * SIZE(CO2)
- LD C31, 5 * SIZE(CO2)
- LD C41, 7 * SIZE(CO2)
-
- MADD C13, C13, C12, A1
- MADD C23, C23, C22, A1
-
- MADD C33, C33, C32, A1
- ST B1, 0 * SIZE(CO1)
-
- MADD C43, C43, C42, A1
- ST B3, 2 * SIZE(CO1)
-
- MADD C11, C11, C14, A1
- ST B5, 4 * SIZE(CO1)
-
- MADD C21, C21, C24, A1
- ST B7, 6 * SIZE(CO1)
-
- MADD C31, C31, C34, A1
- ST B2, 1 * SIZE(CO1)
-
- MADD C41, C41, C44, A1
- ST B4, 3 * SIZE(CO1)
-
- NMSUB C13, C13, C14, A2
- ST B6, 5 * SIZE(CO1)
-
- NMSUB C23, C23, C24, A2
- ST B8, 7 * SIZE(CO1)
-
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
-
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- SUB C13, C13, A5 # ad'+'cb
- SUB C23, C23, A6
- SUB C33, C33, A7
- SUB C43, C43, A8
- ADD C12, B1, C12
- ADD C22, B2, C22
- ADD C32, B3, C32
- ADD C42, B4, C42
- SUB C14, C14, B5
- SUB C24, C24, B6
- SUB C34, C34, B7
- SUB C44, C44, B8
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- LD C13, 0 * SIZE(CO2)
- LD C23, 2 * SIZE(CO2)
- LD C33, 4 * SIZE(CO2)
- LD C43, 6 * SIZE(CO2)
- LD C11, 1 * SIZE(CO2)
- LD C21, 3 * SIZE(CO2)
- LD C31, 5 * SIZE(CO2)
- LD C41, 7 * SIZE(CO2)
-
- MADD C13, C13, C12, A1
- MADD C23, C23, C22, A1
-
- MADD C33, C33, C32, A1
- ST B1, 0 * SIZE(CO1)
-
- MADD C43, C43, C42, A1
- ST B3, 2 * SIZE(CO1)
-
- MADD C11, C11, C14, A1
- ST B5, 4 * SIZE(CO1)
-
- MADD C21, C21, C24, A1
- ST B7, 6 * SIZE(CO1)
-
- MADD C31, C31, C34, A1
- ST B2, 1 * SIZE(CO1)
-
- MADD C41, C41, C44, A1
- ST B4, 3 * SIZE(CO1)
-
- NMSUB C13, C13, C14, A2
- ST B6, 5 * SIZE(CO1)
-
- NMSUB C23, C23, C24, A2
- ST B8, 7 * SIZE(CO1)
-
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
-
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- SUB C31, C31, A3
- LD A1, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- SUB C41, C41, A4
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
-
- ADD C13, A5, C13 # ad'+'cb
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
- SUB C12, C12, B1
- SUB C22, C22, B2
- SUB C32, C32, B3
- SUB C42, C42, B4
- ADD C14, B5, C14
- ADD C24, B6, C24
- ADD C34, B7, C34
- ADD C44, B8, C44
- NEG C13, C13
- NEG C23, C23
- NEG C33, C33
- NEG C43, C43
- NEG C14, C14
- NEG C24, C24
- NEG C34, C34
- NEG C44, C44
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- LD C13, 0 * SIZE(CO2)
- LD C43, 6 * SIZE(CO2)
- LD C23, 2 * SIZE(CO2)
- LD C33, 4 * SIZE(CO2)
- LD C11, 1 * SIZE(CO2)
- LD C21, 3 * SIZE(CO2)
- LD C31, 5 * SIZE(CO2)
- LD C41, 7 * SIZE(CO2)
-
- MADD C13, C13, C12, A1
- ST B1, 0 * SIZE(CO1)
-
- MADD C23, C23, C22, A1
- ST B3, 2 * SIZE(CO1)
-
- MADD C33, C33, C32, A1
- ST B5, 4 * SIZE(CO1)
-
- MADD C43, C43, C42, A1
- ST B7, 6 * SIZE(CO1)
-
- MADD C11, C11, C14, A1
- ST B2, 1 * SIZE(CO1)
-
- MADD C21, C21, C24, A1
- ST B4, 3 * SIZE(CO1)
-
- MADD C31, C31, C34, A1
- ST B6, 5 * SIZE(CO1)
-
- MADD C41, C41, C44, A1
- ST B8, 7 * SIZE(CO1)
-
- NMSUB C13, C13, C14, A2
- NMSUB C23, C23, C24, A2
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
-
- #endif
-
- #else
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C31
- CVTU A4, C41
-
- CVTU A5, C13
- CVTU A6, C23
-
- CVTU A7, C33
- CVTU A8, C43
-
- CVTU B1, C12
- CVTU B2, C22
-
- CVTU B3, C32
- CVTU B4, C42
-
- CVTU B5, C14
- CVTU B6, C24
-
- CVTU B7, C34
- CVTU B8, C44
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- SUB C31, C31, A3
- LD A1, 152($sp) # load alpha_r
- SUB C41, C41, A4
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- ADD C13, A5, C13 # ad'+'cb
- ADD C23, A6, C23
- # LD A2, 0 * SIZE(A) # load alpha_i
- ADD C33, A7, C33
- ADD C43, A8, C43
- SUB C12, C12, B1
- SUB C22, C22, B2
- SUB C32, C32, B3
- SUB C42, C42, B4
- ADD C14, B5, C14
- ADD C24, B6, C24
- ADD C34, B7, C34
- ADD C44, B8, C44
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- MUL C13, C12, A1
- MUL C23, C22, A1
-
- ST B3, 2 * SIZE(CO1)
- MUL C33, C32, A1
- MUL C43, C42, A1
-
- ST B5, 4 * SIZE(CO1)
- MUL C11, C14, A1
- MUL C21, C24, A1
-
- ST B7, 6 * SIZE(CO1)
- MUL C31, C34, A1
- MUL C41, C44, A1
-
- ST B2, 1 * SIZE(CO1)
- NMSUB C13, C13, C14, A2
- NMSUB C23, C23, C24, A2
-
- ST B4, 3 * SIZE(CO1)
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- ST B6, 5 * SIZE(CO1)
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- ST B8, 7 * SIZE(CO1)
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- SUB C13, A5, C13 # ad'+'cb
- SUB C23, A6, C23
- SUB C33, A7, C33
- SUB C43, A8, C43
- ADD C12, B1, C12
- ADD C22, B2, C22
- ADD C32, B3, C32
- ADD C42, B4, C42
- SUB C14, B5, C14
- SUB C24, B6, C24
- SUB C34, B7, C34
- SUB C44, B8, C44
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- MUL C13, C12, A1
- MUL C23, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- MUL C33, C32, A1
- MUL C43, C42, A1
-
- ST B3, 2 * SIZE(CO1)
- MUL C11, C14, A1
- MUL C21, C24, A1
-
- ST B5, 4 * SIZE(CO1)
- MUL C31, C34, A1
- MUL C41, C44, A1
-
- ST B7, 6 * SIZE(CO1)
- NMSUB C13, C13, C14, A2
- NMSUB C23, C23, C24, A2
-
- ST B2, 1 * SIZE(CO1)
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- ST B4, 3 * SIZE(CO1)
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- ST B6, 5 * SIZE(CO1)
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST B8, 7 * SIZE(CO1)
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
-
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- SUB C13, C13, A5 # ad'+'cb
- SUB C23, C23, A6
- SUB C33, C33, A7
- SUB C43, C43, A8
- ADD C12, B1, C12
- ADD C22, B2, C22
- ADD C32, B3, C32
- ADD C42, B4, C42
- SUB C14, C14, B5
- SUB C24, C24, B6
-
- SUB C34, C34, B7
- SUB C44, C44, B8
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- MUL C13, C12, A1
- MUL C23, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- MUL C33, C32, A1
- MUL C43, C42, A1
-
- ST B3, 2 * SIZE(CO1)
- MUL C11, C14, A1
- MUL C21, C24, A1
-
- ST B5, 4 * SIZE(CO1)
- MUL C31, C34, A1
- MUL C41, C44, A1
-
- ST B7, 6 * SIZE(CO1)
- NMSUB C13, C13, C14, A2
- NMSUB C23, C23, C24, A2
-
- ST B2, 1 * SIZE(CO1)
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- ST B4, 3 * SIZE(CO1)
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- ST B6, 5 * SIZE(CO1)
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST B8, 7 * SIZE(CO1)
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
-
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- SUB C31, C31, A3
- LD A1, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- SUB C41, C41, A4
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
-
- ADD C13, A5, C13 # ad'+'cb
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
- SUB C12, C12, B1
- SUB C22, C22, B2
- SUB C32, C32, B3
- SUB C42, C42, B4
- ADD C14, B5, C14
- ADD C24, B6, C24
- ADD C34, B7, C34
- ADD C44, B8, C44
-
- NEG C13, C13
- NEG C23, C23
- NEG C33, C33
- NEG C43, C43
- NEG C14, C14
- NEG C24, C24
- NEG C34, C34
- NEG C44, C44
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- MUL C13, C12, A1
- MUL C23, C22, A1
-
- ST B3, 2 * SIZE(CO1)
- MUL C33, C32, A1
- MUL C43, C42, A1
-
- ST B5, 4 * SIZE(CO1)
- MUL C11, C14, A1
- MUL C21, C24, A1
-
- ST B7, 6 * SIZE(CO1)
- MUL C31, C34, A1
- MUL C41, C44, A1
-
- ST B2, 1 * SIZE(CO1)
- NMSUB C13, C13, C14, A2
- NMSUB C23, C23, C24, A2
-
- ST B4, 3 * SIZE(CO1)
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- ST B6, 5 * SIZE(CO1)
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- ST B8, 7 * SIZE(CO1)
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
- #endif
-
-
- #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 + ZBASE_SHIFT
- dsll TEMP, TEMP, 1 + ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 4
- #endif
-
- #endif
- daddiu CO1, CO1, 8 * SIZE
- bgtz I, .L241
- daddiu CO2, CO2, 8 * SIZE
-
- .align 4
- .L22:
- andi I, M, 2 # MR=4
- blez I, .L21
- NOP
-
- .align 4
- .L221:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll TEMP, KK, 1 + ZBASE_SHIFT # NR=2
-
- daddu AO, AO, TEMP
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- MOV C21, C11
- MOV C22, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
- MOV C14, C11
-
- MOV C23, C11
- FETCH $0, 0 * SIZE(CO1)
-
- FETCH $0, 8 * SIZE(CO1)
- MOV C24, C11
-
- FETCH $0, 0 * SIZE(CO2)
- FETCH $0, 8 * SIZE(CO2)
-
- PLU B3, B1, B1
- PLU B4, B2, B2
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 2 # MR=2
- #else
- daddiu TEMP, KK, 2 # NR=2
- #endif
- dsra L, TEMP, 2
- blez L, .L222
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=64
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- MOV C21, C11
- MOV C22, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
- MOV C14, C11
-
- MOV C23, C11
- FETCH $0, 0 * SIZE(CO1)
-
- FETCH $0, 8 * SIZE(CO1)
- MOV C24, C11
-
- FETCH $0, 0 * SIZE(CO2)
- FETCH $0, 8 * SIZE(CO2)
-
- PLU B3, B1, B1
- blez L, .L222
- PLU B4, B2, B2
- #endif
-
- .L2210:
- daddiu L, L, -1
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C14, C14, A1, B4
- MADPS C24, C24, A2, B4
-
- gsLQC1(R12, F5, F4, 2) # A5 A6
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- gsLQC1(R13, F9, F8, 2) # B1 B2
- MADPS C11, C11, A3, B5
- MADPS C21, C21, A4, B5
-
- MADPS C12, C12, A3, B6
- MADPS C22, C22, A4, B6
-
- MADPS C13, C13, A3, B7
- MADPS C23, C23, A4, B7
-
- MADPS C14, C14, A3, B8
- MADPS C24, C24, A4, B8
-
- gsLQC1(R12, F7, F6, 3) # A7 A8
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- gsLQC1(R13, F13, F12, 3) # B3 B4
- MADPS C11, C11, A5, B1
- MADPS C21, C21, A6, B1
-
- MADPS C12, C12, A5, B2
- MADPS C22, C22, A6, B2
- daddiu BO, BO, 4 * 4 * SIZE # 4KR*4NR
-
- daddiu AO, AO, 4 * 4 * SIZE # 4KR*8MR
- MADPS C13, C13, A5, B3
- MADPS C23, C23, A6, B3
-
- MADPS C14, C14, A5, B4
- MADPS C24, C24, A6, B4
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C11, C11, A7, B5
- MADPS C21, C21, A8, B5
-
- MADPS C12, C12, A7, B6
- MADPS C22, C22, A8, B6
-
- MADPS C13, C13, A7, B7
- MADPS C23, C23, A8, B7
-
- MADPS C14, C14, A7, B8
- MADPS C24, C24, A8, B8
-
- PLU B3, B1, B1
- bgtz L, .L2210
- PLU B4, B2, B2
-
-
- .align 4
- .L222:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L227
- NOP
-
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- 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
- daddiu BO, BO, 2 * 4 * SIZE
-
- daddiu AO, AO, 2 * 4 * SIZE
- MADPS C11, C11, A3, B5
- MADPS C21, C21, A4, B5
- gsLQC1(R13, F9, F8, 0) # A1 A2
-
- MADPS C12, C12, A3, B6
- MADPS C22, C22, A4, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- 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
-
-
- .align 4
- .L227:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L220
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- daddiu BO, BO, 4 * SIZE
- daddiu AO, AO, 4 * SIZE
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C14, C14, A1, B4
- MADPS C24, C24, A2, B4
-
- .align 4
- .L220: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C13
- CVTU A4, C23
-
- CVTU A5, C12
- CVTU A6, C22
-
- CVTU A7, C14
- CVTU A8, C24
-
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- SUB C22, C22, A6
- ADD C14, A7, C14
- ADD C24, A8, C24
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B7, 2 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
- LD B8, 3 * SIZE(CO2)
-
- MADD B5, B5, C12, A1
- MADD B7, B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MADD B6, B6, C14, A1
- MADD B8, B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, A3, C13 # ad'+'cb
- SUB C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C12, A5, C12
- ADD C22, A6, C22
- SUB C14, A7, C14
- SUB C24, A8, C24
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B7, 2 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
- LD B8, 3 * SIZE(CO2)
-
- MADD B5, B5, C12, A1
- MADD B7, B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MADD B6, B6, C14, A1
- MADD B8, B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
-
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, C13, A3 # ad'+'cb
- SUB C23, C23, A4
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- ADD C12, A5, C12
- ADD C22, A6, C22
- SUB C14, C14, A7
- SUB C24, C24, A8
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B7, 2 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
- LD B8, 3 * SIZE(CO2)
-
- MADD B5, B5, C12, A1
- MADD B7, B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MADD B6, B6, C14, A1
- MADD B8, B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
-
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- LD A1, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- SUB C22, C22, A6
- ADD C14, A7, C14
- ADD C24, A8, C24
- NEG C13, C13
- NEG C23, C23
- NEG C14, C14
- NEG C24, C24
-
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B7, 2 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
- LD B8, 3 * SIZE(CO2)
-
- MADD B5, B5, C12, A1
- MADD B7, B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MADD B6, B6, C14, A1
- MADD B8, B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
- #endif
-
- #else
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C13
- CVTU A4, C23
-
- CVTU A5, C12
- CVTU A6, C22
-
- CVTU A7, C14
- CVTU A8, C24
-
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- SUB C22, C22, A6
- ADD C14, A7, C14
- ADD C24, A8, C24
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
-
- MUL B5, C12, A1
- MUL B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MUL B6, C14, A1
- MUL B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, A3, C13 # ad'+'cb
- SUB C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C12, A5, C12
- ADD C22, A6, C22
- SUB C14, A7, C14
- SUB C24, A8, C24
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- MUL B5, C12, A1
- MUL B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MUL B6, C14, A1
- MUL B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
-
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, C13, A3 # ad'+'cb
- SUB C23, C23, A4
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- ADD C12, A5, C12
- ADD C22, A6, C22
- SUB C14, C14, A7
- SUB C24, C24, A8
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- MUL B5, C12, A1
- MUL B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MUL B6, C14, A1
- MUL B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
-
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- LD A1, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- SUB C22, C22, A6
- ADD C14, A7, C14
- ADD C24, A8, C24
- NEG C13, C13
- NEG C23, C23
- NEG C14, C14
- NEG C24, C24
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- MUL B5, C12, A1
- MUL B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MUL B6, C14, A1
- MUL B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
- #endif
-
- #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 TEMP, TEMP, 1 + ZBASE_SHIFT
-
- daddu AO, AO, TEMP
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 2
- #endif
-
- #endif
- daddiu CO1, CO1, 4 * SIZE
- daddiu CO2, CO2, 4 * SIZE
-
-
- .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
- #else
- dsll L, KK, ZBASE_SHIFT # MR=1
- dsll TEMP, KK, 1 + ZBASE_SHIFT # NR=2
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
- MOV C14, C11
-
- FETCH $0, 0 * SIZE(CO1)
- FETCH $0, 0 * SIZE(CO2)
-
- PLU B3, B1, B1
- PLU B4, B2, B2
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 1 # MR=1
- #else
- daddiu TEMP, KK, 2 # NR=2
- #endif
- dsra L, TEMP, 2
- blez L, .L212
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=64
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
- MOV C14, C11
-
- FETCH $0, 0 * SIZE(CO1)
- FETCH $0, 0 * SIZE(CO2)
-
- PLU B3, B1, B1
- blez L, .L212
- PLU B4, B2, B2
- #endif
-
- .L2110:
- daddiu L, L, -1
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C12, C12, A1, B2
-
- MADPS C13, C13, A1, B3
- MADPS C14, C14, A1, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- gsLQC1(R13, F9, F8, 2) # B1 B2
- MADPS C11, C11, A2, B5
- MADPS C12, C12, A2, B6
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C13, C13, A2, B7
- MADPS C14, C14, A2, B8
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- gsLQC1(R13, F13, F12, 3) # B3 B4
- MADPS C11, C11, A3, B1
- MADPS C12, C12, A3, B2
- daddiu BO, BO, 4 * 4 * SIZE # 4KR*4NR
-
- daddiu AO, AO, 2 * 4 * SIZE # 4KR*8MR
- MADPS C13, C13, A3, B3
- MADPS C14, C14, A3, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C11, C11, A4, B5
- MADPS C12, C12, A4, B6
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C13, C13, A4, B7
- MADPS C14, C14, A4, B8
-
- PLU B3, B1, B1
- bgtz L, .L2110
- PLU B4, B2, B2
-
-
- .align 4
- .L212:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L217
- NOP
-
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C12, C12, A1, B2
-
- MADPS C13, C13, A1, B3
- MADPS C14, C14, A1, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
- daddiu BO, BO, 2 * 4 * SIZE
-
- MADPS C11, C11, A2, B5
- MADPS C12, C12, A2, B6
- daddiu AO, AO, 4 * SIZE
-
- MADPS C13, C13, A2, B7
- MADPS C14, C14, A2, B8
-
- gsLQC1(R12, F1, F0, 0) # A5 A6
- gsLQC1(R13, F9, F8, 0) # B1 B2
- PLU B3, B1, B1
- PLU B4, B2, B2
-
-
- .align 4
- .L217:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L210
- NOP
-
- MADPS C11, C11, A1, B1
- daddiu BO, BO, 4 * SIZE
- MADPS C12, C12, A1, B2
- daddiu AO, AO, 2 * SIZE
-
- MADPS C13, C13, A1, B3
- MADPS C14, C14, A1, B4
-
- .align 4
- .L210: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A3, C13
- CVTU A5, C12
- CVTU A7, C14
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- ADD C14, A7, C14
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
-
- MADD B5, B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MADD B6, B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, A3, C13 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C12, A5, C12
- SUB C14, A7, C14
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
-
- MADD B5, B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MADD B6, B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
-
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, C13, A3 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- ADD C12, A5, C12
- SUB C14, C14, A7
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
-
- MADD B5, B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MADD B6, B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- LD A4, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- ADD C14, A7, C14
- NEG C13, C13
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- NEG C14, C14
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
-
- MADD B5, B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MADD B6, B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- #endif
-
- #else
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A3, C13
- CVTU A5, C12
- CVTU A7, C14
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- ADD C14, A7, C14
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- MUL B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MUL B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, A3, C13 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C12, A5, C12
- SUB C14, A7, C14
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- MUL B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MUL B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
-
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, C13, A3 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- ADD C12, A5, C12
- SUB C14, C14, A7
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- MUL B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MUL B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- LD A4, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- ADD C14, A7, C14
- NEG C13, C13
- NEG C14, C14
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- MUL B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MUL B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- #endif
-
-
- #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, ZBASE_SHIFT
- dsll TEMP, TEMP, 1 + ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 1
- #endif
-
- #endif
- daddiu CO1, CO1, 2 * SIZE
- daddiu CO2, CO2, 2 * SIZE
-
-
- .align 4
- .L20:
- daddiu J, J, -1
- move B, BO
-
- #if defined(TRMMKERNEL) && !defined(LEFT)
- daddiu KK, KK, 2
- #endif
-
- bgtz J, .L24
- NOP
-
-
- .align 4
- .L1:
- andi J, N, 1
- blez J, .L999
- NOP
-
- .L14:
- dsra I, M, 2 # MR=8
- move AO, A # Reset A
-
- #if defined(TRMMKERNEL) && defined(LEFT)
- move KK, OFFSET
- #endif
-
- move CO1, C
- blez I, .L12
- daddu C, CO1, LDC
-
- .align 4
- .L141:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, 2 + ZBASE_SHIFT
- dsll TEMP, KK, ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C21, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C31, C11
- MOV C41, C11
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MOV C13, C11
- MOV C23, C11
-
- FETCH $0, 0 * SIZE(CO1)
- MOV C33, C11
- MOV C43, C11
-
- FETCH $0, 8 * SIZE(CO1)
- PLU B3, B1, B1
- PLU B4, B2, B2
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 4 # define Mr=4
- #else
- daddiu TEMP, KK, 1 # define NR=1
- #endif
- dsra L, TEMP, 2
- blez L, .L142
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=64
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C21, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C31, C11
- MOV C41, C11
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MOV C13, C11
- MOV C23, C11
-
- FETCH $0, 0 * SIZE(CO1)
- MOV C33, C11
- MOV C43, C11
-
- FETCH $0, 8 * SIZE(CO1)
- PLU B3, B1, B1
- blez L, .L142
- PLU B4, B2, B2
- #endif
-
- .L1410:
- daddiu L, L, -1
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- gsLQC1(R12, F7, F6, 3) # A7 A8
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- gsLQC1(R12, F1, F0, 4) # A1 A2
- MADPS C11, C11, A5, B2
- MADPS C21, C21, A6, B2
-
- gsLQC1(R12, F3, F2, 5) # A3 A4
- MADPS C31, C31, A7, B2
- MADPS C41, C41, A8, B2
- daddiu BO, BO, 2 * 4 * SIZE # 4KR*4NR
-
- MADPS C13, C13, A5, B4
- MADPS C23, C23, A6, B4
-
- MADPS C33, C33, A7, B4
- MADPS C43, C43, A8, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- MADPS C11, C11, A1, B5
- MADPS C21, C21, A2, B5
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- gsLQC1(R12, F7, F6, 7) # A7 A8
- MADPS C31, C31, A3, B5
- MADPS C41, C41, A4, B5
-
- daddiu AO, AO, 8 * 4 * SIZE # 4KR*8MR
- MADPS C13, C13, A1, B7
- MADPS C23, C23, A2, B7
-
- MADPS C33, C33, A3, B7
- MADPS C43, C43, A4, B7
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C11, C11, A5, B6
- MADPS C21, C21, A6, B6
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C31, C31, A7, B6
- MADPS C41, C41, A8, B6
-
- MADPS C13, C13, A5, B8
- MADPS C23, C23, A6, B8
-
- MADPS C33, C33, A7, B8
- MADPS C43, C43, A8, B8
-
- PLU B3, B1, B1
- bgtz L, .L1410
- PLU B4, B2, B2
-
-
- .align 4
- .L142:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L147
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- gsLQC1(R12, F7, F6, 3) # A7 A8
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- daddiu AO, AO, 4 * 4 * SIZE # 4KR*8MR
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
- gsLQC1(R13, F13, F8, 1) # B3 B4
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C11, C11, A5, B2
- MADPS C21, C21, A6, B2
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C31, C31, A7, B2
- MADPS C41, C41, A8, B2
- daddiu BO, BO, 4 * SIZE # 4KR*4NR
-
- MADPS C13, C13, A5, B4
- MADPS C23, C23, A6, B4
-
- MADPS C33, C33, A7, B4
- MADPS C43, C43, A8, B4
- PLU B3, B1, B1
-
-
- .align 4
- .L147:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L140
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- daddiu BO, BO, 2 * SIZE
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- daddiu AO, AO, 2 * 4 * SIZE
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
-
- .align 4
- .L140: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C31
- CVTU A4, C41
-
- CVTU A5, C13
- CVTU A6, C23
-
- CVTU A7, C33
- CVTU A8, C43
-
- CVTU B1, C12
- CVTU B2, C22
-
- CVTU B3, C32
- CVTU B4, C42
-
- CVTU B5, C14
- CVTU B6, C24
-
- CVTU B7, C34
- CVTU B8, C44
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- # LD A1, 0 * SIZE(A) # load alpha_r
- SUB C31, C31, A3
- LD A1, 152($sp) # load alpha_r
- SUB C41, C41, A4
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- ADD C13, A5, C13 # ad'+'cb
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- SUB C13, A5, C13 # ad'+'cb
- SUB C23, A6, C23
- SUB C33, A7, C33
- SUB C43, A8, C43
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- SUB C13, C13, A5 # ad'+'cb
- SUB C23, C23, A6
- SUB C33, C33, A7
- SUB C43, C43, A8
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # AC'+'BD
- SUB C21, C21, A2
- SUB C31, C31, A3
- LD A1, 152($sp) # LOAD ALPHA_R
- # LD A1, 0 * SIZE(A) # LOAD ALPHA_R
- SUB C41, C41, A4
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # LOAD ALPHA_I
-
- ADD C13, A5, C13 # AD'+'CB
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
- NEG C13, C13 # AD'+'CB
- NEG C23, C23
- NEG C33, C33
- NEG C43, C43
-
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = ALPHA_R
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = ALPHA_I
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #else
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C31
- CVTU A4, C41
-
- CVTU A5, C13
- CVTU A6, C23
-
- CVTU A7, C33
- CVTU A8, C43
-
- CVTU B1, C12
- CVTU B2, C22
-
- CVTU B3, C32
- CVTU B4, C42
-
- CVTU B5, C14
- CVTU B6, C24
-
- CVTU B7, C34
- CVTU B8, C44
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- # LD A1, 0 * SIZE(A) # load alpha_r
- SUB C31, C31, A3
- LD A1, 152($sp) # load alpha_r
- SUB C41, C41, A4
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- ADD C13, A5, C13 # ad'+'cb
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- SUB C13, A5, C13 # ad'+'cb
- SUB C23, A6, C23
- SUB C33, A7, C33
- SUB C43, A8, C43
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- SUB C13, C13, A5 # ad'+'cb
- SUB C23, C23, A6
- SUB C33, C33, A7
- SUB C43, C43, A8
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # AC'+'BD
- SUB C21, C21, A2
- SUB C31, C31, A3
- LD A1, 152($sp) # LOAD ALPHA_R
- # LD A1, 0 * SIZE(A) # LOAD ALPHA_R
- SUB C41, C41, A4
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # LOAD ALPHA_I
-
- ADD C13, A5, C13 # AD'+'CB
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
- NEG C13, C13 # AD'+'CB
- NEG C23, C23
- NEG C33, C33
- NEG C43, C43
-
- MUL B1, C11, A1 # A1 = ALPHA_R
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = ALPHA_I
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
-
- #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 + ZBASE_SHIFT
- dsll TEMP, TEMP, ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 4
- #endif
-
- #endif
- bgtz I, .L141
- daddiu CO1, CO1, 8 * SIZE
-
- .align 4
- .L12:
- andi I, M, 2 # MR=4
- blez I, .L11
- NOP
-
- .align 4
- .L121:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, 1 + ZBASE_SHIFT
- dsll TEMP, KK, ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C21, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
- MOV C23, C11
-
- FETCH $0, 0 * SIZE(CO1)
- FETCH $0, 8 * SIZE(CO1)
-
- PLU B3, B1, B1
- 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, 1
- #endif
- dsra L, TEMP, 2
- blez L, .L122
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=64
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C21, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
- MOV C23, C11
-
- FETCH $0, 0 * SIZE(CO1)
- FETCH $0, 8 * SIZE(CO1)
-
- PLU B3, B1, B1
- blez L, .L122
- PLU B4, B2, B2
- #endif
-
- .L1210:
- daddiu L, L, -1
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- gsLQC1(R12, F5, F4, 2) # A5 A6
- PLU B7, B5, B5
- PLU B8, B6, B6
- daddiu BO, BO, 2 * 4 * SIZE # 4KR*4NR
-
- MADPS C11, C11, A3, B2
- MADPS C21, C21, A4, B2
-
- gsLQC1(R12, F7, F6, 3) # A7 A8
- MADPS C13, C13, A3, B4
- MADPS C23, C23, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- daddiu AO, AO, 4 * 4 * SIZE # 4KR*8MR
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C11, C11, A7, B6
- MADPS C21, C21, A8, B6
-
- MADPS C13, C13, A7, B8
- MADPS C23, C23, A8, B8
-
- PLU B3, B1, B1
- bgtz L, .L1210
- PLU B4, B2, B2
-
-
- .align 4
- .L122:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L127
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- PLU B7, B5, B5
- daddiu BO, BO, 1 * 4 * SIZE
-
- daddiu AO, AO, 2 * 4 * SIZE
- MADPS C11, C11, A3, B2
- MADPS C21, C21, A4, B2
-
- MADPS C13, C13, A3, B4
- MADPS C23, C23, A4, B4
-
- gsLQC1(R13, F9, F8, 0)
- gsLQC1(R12, F1, F0, 0)
- PLU B3, B1, B1
-
- .align 4
- .L127:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L120
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- daddiu BO, BO, 2 * SIZE
- daddiu AO, AO, 4 * SIZE
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- .align 4
- .L120: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C13
- CVTU A4, C23
-
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, A3, C13 # ad'+'cb
- SUB C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, C13, A3 # ad'+'cb
- SUB C23, C23, A4
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- LD A1, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
- NEG C13, C13 # ad'+'cb
- NEG C23, C23
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #else
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C13
- CVTU A4, C23
-
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, A3, C13 # ad'+'cb
- SUB C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, C13, A3 # ad'+'cb
- SUB C23, C23, A4
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- LD A1, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
- NEG C13, C13 # ad'+'cb
- NEG C23, C23
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
- #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 + ZBASE_SHIFT
- dsll TEMP, TEMP, ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 2
- #endif
-
- #endif
- daddiu CO1, CO1, 4 * SIZE
- daddiu CO2, CO2, 4 * SIZE
-
-
- .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 TEMP, KK, ZBASE_SHIFT
-
- daddu AO, AO, TEMP
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
-
- FETCH $0, 0 * SIZE(CO1)
-
- PLU B3, B1, B1
- PLU B4, B2, B2
- #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, 2
- blez L, .L112
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=64
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
-
- FETCH $0, 0 * SIZE(CO1)
-
- PLU B3, B1, B1
- blez L, .L112
- PLU B4, B2, B2
- #endif
-
- .L1110:
- daddiu L, L, -1
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 2 * 4 * SIZE # 4KR*4NR
-
- PLU B7, B5, B5
- PLU B8, B6, B6
- daddiu AO, AO, 2 * 4 * SIZE # 4KR*8MR
-
- MADPS C11, C11, A2, B2
- MADPS C13, C13, A2, B4
-
- MADPS C11, C11, A3, B5
- MADPS C13, C13, A3, B7
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C11, C11, A4, B6
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C13, C13, A4, B8
-
- PLU B3, B1, B1
- bgtz L, .L1110
- PLU B4, B2, B2
-
-
- .align 4
- .L112:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L117
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 4 * SIZE
- daddiu AO, AO, 4 * SIZE
-
- MADPS C11, C11, A2, B2
- MADPS C13, C13, A2, B4
-
- gsLQC1(R13, F9, F8, 0)
- gsLQC1(R12, F1, F0, 0)
- PLU B3, B1, B1
-
-
- .align 4
- .L117:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L110
- NOP
-
- daddiu BO, BO, 2 * SIZE
- daddiu AO, AO, 2 * SIZE
-
- MADPS C11, C11, A1, B1
- MADPS C13, C13, A1, B3
-
-
- .align 4
- .L110: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A3, C13
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, A3, C13 # ad'+'cb
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, C13, A3 # ad'+'cb
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- NEG C13, C13
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp)
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #else
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A3, C13
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, A3, C13 # ad'+'cb
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, C13, A3 # ad'+'cb
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- NEG C13, C13
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp)
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || \
- (!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
- #ifdef LEFT
- daddiu TEMP, TEMP, -1
- #else
- daddiu TEMP, TEMP, -1
- #endif
-
- dsll TEMP, TEMP, ZBASE_SHIFT
-
- daddu AO, AO, TEMP
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 1
- #endif
-
- #endif
- daddiu CO1, CO1, 2 * SIZE
- daddiu CO2, CO2, 2 * SIZE
-
-
- .align 4
- .L10:
- move B, BO
- #if defined(TRMMKERNEL) && !defined(LEFT)
- daddiu KK, KK, 1
- #endif
-
- .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
|