libcuda.so.tramp.S 344 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585258625872588258925902591259225932594259525962597259825992600260126022603260426052606260726082609261026112612261326142615261626172618261926202621262226232624262526262627262826292630263126322633263426352636263726382639264026412642264326442645264626472648264926502651265226532654265526562657265826592660266126622663266426652666266726682669267026712672267326742675267626772678267926802681268226832684268526862687268826892690269126922693269426952696269726982699270027012702270327042705270627072708270927102711271227132714271527162717271827192720272127222723272427252726272727282729273027312732273327342735273627372738273927402741274227432744274527462747274827492750275127522753275427552756275727582759276027612762276327642765276627672768276927702771277227732774277527762777277827792780278127822783278427852786278727882789279027912792279327942795279627972798279928002801280228032804280528062807280828092810281128122813281428152816281728182819282028212822282328242825282628272828282928302831283228332834283528362837283828392840284128422843284428452846284728482849285028512852285328542855285628572858285928602861286228632864286528662867286828692870287128722873287428752876287728782879288028812882288328842885288628872888288928902891289228932894289528962897289828992900290129022903290429052906290729082909291029112912291329142915291629172918291929202921292229232924292529262927292829292930293129322933293429352936293729382939294029412942294329442945294629472948294929502951295229532954295529562957295829592960296129622963296429652966296729682969297029712972297329742975297629772978297929802981298229832984298529862987298829892990299129922993299429952996299729982999300030013002300330043005300630073008300930103011301230133014301530163017301830193020302130223023302430253026302730283029303030313032303330343035303630373038303930403041304230433044304530463047304830493050305130523053305430553056305730583059306030613062306330643065306630673068306930703071307230733074307530763077307830793080308130823083308430853086308730883089309030913092309330943095309630973098309931003101310231033104310531063107310831093110311131123113311431153116311731183119312031213122312331243125312631273128312931303131313231333134313531363137313831393140314131423143314431453146314731483149315031513152315331543155315631573158315931603161316231633164316531663167316831693170317131723173317431753176317731783179318031813182318331843185318631873188318931903191319231933194319531963197319831993200320132023203320432053206320732083209321032113212321332143215321632173218321932203221322232233224322532263227322832293230323132323233323432353236323732383239324032413242324332443245324632473248324932503251325232533254325532563257325832593260326132623263326432653266326732683269327032713272327332743275327632773278327932803281328232833284328532863287328832893290329132923293329432953296329732983299330033013302330333043305330633073308330933103311331233133314331533163317331833193320332133223323332433253326332733283329333033313332333333343335333633373338333933403341334233433344334533463347334833493350335133523353335433553356335733583359336033613362336333643365336633673368336933703371337233733374337533763377337833793380338133823383338433853386338733883389339033913392339333943395339633973398339934003401340234033404340534063407340834093410341134123413341434153416341734183419342034213422342334243425342634273428342934303431343234333434343534363437343834393440344134423443344434453446344734483449345034513452345334543455345634573458345934603461346234633464346534663467346834693470347134723473347434753476347734783479348034813482348334843485348634873488348934903491349234933494349534963497349834993500350135023503350435053506350735083509351035113512351335143515351635173518351935203521352235233524352535263527352835293530353135323533353435353536353735383539354035413542354335443545354635473548354935503551355235533554355535563557355835593560356135623563356435653566356735683569357035713572357335743575357635773578357935803581358235833584358535863587358835893590359135923593359435953596359735983599360036013602360336043605360636073608360936103611361236133614361536163617361836193620362136223623362436253626362736283629363036313632363336343635363636373638363936403641364236433644364536463647364836493650365136523653365436553656365736583659366036613662366336643665366636673668366936703671367236733674367536763677367836793680368136823683368436853686368736883689369036913692369336943695369636973698369937003701370237033704370537063707370837093710371137123713371437153716371737183719372037213722372337243725372637273728372937303731373237333734373537363737373837393740374137423743374437453746374737483749375037513752375337543755375637573758375937603761376237633764376537663767376837693770377137723773377437753776377737783779378037813782378337843785378637873788378937903791379237933794379537963797379837993800380138023803380438053806380738083809381038113812381338143815381638173818381938203821382238233824382538263827382838293830383138323833383438353836383738383839384038413842384338443845384638473848384938503851385238533854385538563857385838593860386138623863386438653866386738683869387038713872387338743875387638773878387938803881388238833884388538863887388838893890389138923893389438953896389738983899390039013902390339043905390639073908390939103911391239133914391539163917391839193920392139223923392439253926392739283929393039313932393339343935393639373938393939403941394239433944394539463947394839493950395139523953395439553956395739583959396039613962396339643965396639673968396939703971397239733974397539763977397839793980398139823983398439853986398739883989399039913992399339943995399639973998399940004001400240034004400540064007400840094010401140124013401440154016401740184019402040214022402340244025402640274028402940304031403240334034403540364037403840394040404140424043404440454046404740484049405040514052405340544055405640574058405940604061406240634064406540664067406840694070407140724073407440754076407740784079408040814082408340844085408640874088408940904091409240934094409540964097409840994100410141024103410441054106410741084109411041114112411341144115411641174118411941204121412241234124412541264127412841294130413141324133413441354136413741384139414041414142414341444145414641474148414941504151415241534154415541564157415841594160416141624163416441654166416741684169417041714172417341744175417641774178417941804181418241834184418541864187418841894190419141924193419441954196419741984199420042014202420342044205420642074208420942104211421242134214421542164217421842194220422142224223422442254226422742284229423042314232423342344235423642374238423942404241424242434244424542464247424842494250425142524253425442554256425742584259426042614262426342644265426642674268426942704271427242734274427542764277427842794280428142824283428442854286428742884289429042914292429342944295429642974298429943004301430243034304430543064307430843094310431143124313431443154316431743184319432043214322432343244325432643274328432943304331433243334334433543364337433843394340434143424343434443454346434743484349435043514352435343544355435643574358435943604361436243634364436543664367436843694370437143724373437443754376437743784379438043814382438343844385438643874388438943904391439243934394439543964397439843994400440144024403440444054406440744084409441044114412441344144415441644174418441944204421442244234424442544264427442844294430443144324433443444354436443744384439444044414442444344444445444644474448444944504451445244534454445544564457445844594460446144624463446444654466446744684469447044714472447344744475447644774478447944804481448244834484448544864487448844894490449144924493449444954496449744984499450045014502450345044505450645074508450945104511451245134514451545164517451845194520452145224523452445254526452745284529453045314532453345344535453645374538453945404541454245434544454545464547454845494550455145524553455445554556455745584559456045614562456345644565456645674568456945704571457245734574457545764577457845794580458145824583458445854586458745884589459045914592459345944595459645974598459946004601460246034604460546064607460846094610461146124613461446154616461746184619462046214622462346244625462646274628462946304631463246334634463546364637463846394640464146424643464446454646464746484649465046514652465346544655465646574658465946604661466246634664466546664667466846694670467146724673467446754676467746784679468046814682468346844685468646874688468946904691469246934694469546964697469846994700470147024703470447054706470747084709471047114712471347144715471647174718471947204721472247234724472547264727472847294730473147324733473447354736473747384739474047414742474347444745474647474748474947504751475247534754475547564757475847594760476147624763476447654766476747684769477047714772477347744775477647774778477947804781478247834784478547864787478847894790479147924793479447954796479747984799480048014802480348044805480648074808480948104811481248134814481548164817481848194820482148224823482448254826482748284829483048314832483348344835483648374838483948404841484248434844484548464847484848494850485148524853485448554856485748584859486048614862486348644865486648674868486948704871487248734874487548764877487848794880488148824883488448854886488748884889489048914892489348944895489648974898489949004901490249034904490549064907490849094910491149124913491449154916491749184919492049214922492349244925492649274928492949304931493249334934493549364937493849394940494149424943494449454946494749484949495049514952495349544955495649574958495949604961496249634964496549664967496849694970497149724973497449754976497749784979498049814982498349844985498649874988498949904991499249934994499549964997499849995000500150025003500450055006500750085009501050115012501350145015501650175018501950205021502250235024502550265027502850295030503150325033503450355036503750385039504050415042504350445045504650475048504950505051505250535054505550565057505850595060506150625063506450655066506750685069507050715072507350745075507650775078507950805081508250835084508550865087508850895090509150925093509450955096509750985099510051015102510351045105510651075108510951105111511251135114511551165117511851195120512151225123512451255126512751285129513051315132513351345135513651375138513951405141514251435144514551465147514851495150515151525153515451555156515751585159516051615162516351645165516651675168516951705171517251735174517551765177517851795180518151825183518451855186518751885189519051915192519351945195519651975198519952005201520252035204520552065207520852095210521152125213521452155216521752185219522052215222522352245225522652275228522952305231523252335234523552365237523852395240524152425243524452455246524752485249525052515252525352545255525652575258525952605261526252635264526552665267526852695270527152725273527452755276527752785279528052815282528352845285528652875288528952905291529252935294529552965297529852995300530153025303530453055306530753085309531053115312531353145315531653175318531953205321532253235324532553265327532853295330533153325333533453355336533753385339534053415342534353445345534653475348534953505351535253535354535553565357535853595360536153625363536453655366536753685369537053715372537353745375537653775378537953805381538253835384538553865387538853895390539153925393539453955396539753985399540054015402540354045405540654075408540954105411541254135414541554165417541854195420542154225423542454255426542754285429543054315432543354345435543654375438543954405441544254435444544554465447544854495450545154525453545454555456545754585459546054615462546354645465546654675468546954705471547254735474547554765477547854795480548154825483548454855486548754885489549054915492549354945495549654975498549955005501550255035504550555065507550855095510551155125513551455155516551755185519552055215522552355245525552655275528552955305531553255335534553555365537553855395540554155425543554455455546554755485549555055515552555355545555555655575558555955605561556255635564556555665567556855695570557155725573557455755576557755785579558055815582558355845585558655875588558955905591559255935594559555965597559855995600560156025603560456055606560756085609561056115612561356145615561656175618561956205621562256235624562556265627562856295630563156325633563456355636563756385639564056415642564356445645564656475648564956505651565256535654565556565657565856595660566156625663566456655666566756685669567056715672567356745675567656775678567956805681568256835684568556865687568856895690569156925693569456955696569756985699570057015702570357045705570657075708570957105711571257135714571557165717571857195720572157225723572457255726572757285729573057315732573357345735573657375738573957405741574257435744574557465747574857495750575157525753575457555756575757585759576057615762576357645765576657675768576957705771577257735774577557765777577857795780578157825783578457855786578757885789579057915792579357945795579657975798579958005801580258035804580558065807580858095810581158125813581458155816581758185819582058215822582358245825582658275828582958305831583258335834583558365837583858395840584158425843584458455846584758485849585058515852585358545855585658575858585958605861586258635864586558665867586858695870587158725873587458755876587758785879588058815882588358845885588658875888588958905891589258935894589558965897589858995900590159025903590459055906590759085909591059115912591359145915591659175918591959205921592259235924592559265927592859295930593159325933593459355936593759385939594059415942594359445945594659475948594959505951595259535954595559565957595859595960596159625963596459655966596759685969597059715972597359745975597659775978597959805981598259835984598559865987598859895990599159925993599459955996599759985999600060016002600360046005600660076008600960106011601260136014601560166017601860196020602160226023602460256026602760286029603060316032603360346035603660376038603960406041604260436044604560466047604860496050605160526053605460556056605760586059606060616062606360646065606660676068606960706071607260736074607560766077607860796080608160826083608460856086608760886089609060916092609360946095609660976098609961006101610261036104610561066107610861096110611161126113611461156116611761186119612061216122612361246125612661276128612961306131613261336134613561366137613861396140614161426143614461456146614761486149615061516152615361546155615661576158615961606161616261636164616561666167616861696170617161726173617461756176617761786179618061816182618361846185618661876188618961906191619261936194619561966197619861996200620162026203620462056206620762086209621062116212621362146215621662176218621962206221622262236224622562266227622862296230623162326233623462356236623762386239624062416242624362446245624662476248624962506251625262536254625562566257625862596260626162626263626462656266626762686269627062716272627362746275627662776278627962806281628262836284628562866287628862896290629162926293629462956296629762986299630063016302630363046305630663076308630963106311631263136314631563166317631863196320632163226323632463256326632763286329633063316332633363346335633663376338633963406341634263436344634563466347634863496350635163526353635463556356635763586359636063616362636363646365636663676368636963706371637263736374637563766377637863796380638163826383638463856386638763886389639063916392639363946395639663976398639964006401640264036404640564066407640864096410641164126413641464156416641764186419642064216422642364246425642664276428642964306431643264336434643564366437643864396440644164426443644464456446644764486449645064516452645364546455645664576458645964606461646264636464646564666467646864696470647164726473647464756476647764786479648064816482648364846485648664876488648964906491649264936494649564966497649864996500650165026503650465056506650765086509651065116512651365146515651665176518651965206521652265236524652565266527652865296530653165326533653465356536653765386539654065416542654365446545654665476548654965506551655265536554655565566557655865596560656165626563656465656566656765686569657065716572657365746575657665776578657965806581658265836584658565866587658865896590659165926593659465956596659765986599660066016602660366046605660666076608660966106611661266136614661566166617661866196620662166226623662466256626662766286629663066316632663366346635663666376638663966406641664266436644664566466647664866496650665166526653665466556656665766586659666066616662666366646665666666676668666966706671667266736674667566766677667866796680668166826683668466856686668766886689669066916692669366946695669666976698669967006701670267036704670567066707670867096710671167126713671467156716671767186719672067216722672367246725672667276728672967306731673267336734673567366737673867396740674167426743674467456746674767486749675067516752675367546755675667576758675967606761676267636764676567666767676867696770677167726773677467756776677767786779678067816782678367846785678667876788678967906791679267936794679567966797679867996800680168026803680468056806680768086809681068116812681368146815681668176818681968206821682268236824682568266827682868296830683168326833683468356836683768386839684068416842684368446845684668476848684968506851685268536854685568566857685868596860686168626863686468656866686768686869687068716872687368746875687668776878687968806881688268836884688568866887688868896890689168926893689468956896689768986899690069016902690369046905690669076908690969106911691269136914691569166917691869196920692169226923692469256926692769286929693069316932693369346935693669376938693969406941694269436944694569466947694869496950695169526953695469556956695769586959696069616962696369646965696669676968696969706971697269736974697569766977697869796980698169826983698469856986698769886989699069916992699369946995699669976998699970007001700270037004700570067007700870097010701170127013701470157016701770187019702070217022702370247025702670277028702970307031703270337034703570367037703870397040704170427043704470457046704770487049705070517052705370547055705670577058705970607061706270637064706570667067706870697070707170727073707470757076707770787079708070817082708370847085708670877088708970907091709270937094709570967097709870997100710171027103710471057106710771087109711071117112711371147115711671177118711971207121712271237124712571267127712871297130713171327133713471357136713771387139714071417142714371447145714671477148714971507151715271537154715571567157715871597160716171627163716471657166716771687169717071717172717371747175717671777178717971807181718271837184718571867187718871897190719171927193719471957196719771987199720072017202720372047205720672077208720972107211721272137214721572167217721872197220722172227223722472257226722772287229723072317232723372347235723672377238723972407241724272437244724572467247724872497250725172527253725472557256725772587259726072617262726372647265726672677268726972707271727272737274727572767277727872797280728172827283728472857286728772887289729072917292729372947295729672977298729973007301730273037304730573067307730873097310731173127313731473157316731773187319732073217322732373247325732673277328732973307331733273337334733573367337733873397340734173427343734473457346734773487349735073517352735373547355735673577358735973607361736273637364736573667367736873697370737173727373737473757376737773787379738073817382738373847385738673877388738973907391739273937394739573967397739873997400740174027403740474057406740774087409741074117412741374147415741674177418741974207421742274237424742574267427742874297430743174327433743474357436743774387439744074417442744374447445744674477448744974507451745274537454745574567457745874597460746174627463746474657466746774687469747074717472747374747475747674777478747974807481748274837484748574867487748874897490749174927493749474957496749774987499750075017502750375047505750675077508750975107511751275137514751575167517751875197520752175227523752475257526752775287529753075317532753375347535753675377538753975407541754275437544754575467547754875497550755175527553755475557556755775587559756075617562756375647565756675677568756975707571757275737574757575767577757875797580758175827583758475857586758775887589759075917592759375947595759675977598759976007601760276037604760576067607760876097610761176127613761476157616761776187619762076217622762376247625762676277628762976307631763276337634763576367637763876397640764176427643764476457646764776487649765076517652765376547655765676577658765976607661766276637664766576667667766876697670767176727673767476757676767776787679768076817682768376847685768676877688768976907691769276937694769576967697769876997700770177027703770477057706770777087709771077117712771377147715771677177718771977207721772277237724772577267727772877297730773177327733773477357736773777387739774077417742774377447745774677477748774977507751775277537754775577567757775877597760776177627763776477657766776777687769777077717772777377747775777677777778777977807781778277837784778577867787778877897790779177927793779477957796779777987799780078017802780378047805780678077808780978107811781278137814781578167817781878197820782178227823782478257826782778287829783078317832783378347835783678377838783978407841784278437844784578467847784878497850785178527853785478557856785778587859786078617862786378647865786678677868786978707871787278737874787578767877787878797880788178827883788478857886788778887889789078917892789378947895789678977898789979007901790279037904790579067907790879097910791179127913791479157916791779187919792079217922792379247925792679277928792979307931793279337934793579367937793879397940794179427943794479457946794779487949795079517952795379547955795679577958795979607961796279637964796579667967796879697970797179727973797479757976797779787979798079817982798379847985798679877988798979907991799279937994799579967997799879998000800180028003800480058006800780088009801080118012801380148015801680178018801980208021802280238024802580268027802880298030803180328033803480358036803780388039804080418042804380448045804680478048804980508051805280538054805580568057805880598060806180628063806480658066806780688069807080718072807380748075807680778078807980808081808280838084808580868087808880898090809180928093809480958096809780988099810081018102810381048105810681078108810981108111811281138114811581168117811881198120812181228123812481258126812781288129813081318132813381348135813681378138813981408141814281438144814581468147814881498150815181528153815481558156815781588159816081618162816381648165816681678168816981708171817281738174817581768177817881798180818181828183818481858186818781888189819081918192819381948195819681978198819982008201820282038204820582068207820882098210821182128213821482158216821782188219822082218222822382248225822682278228822982308231823282338234823582368237823882398240824182428243824482458246824782488249825082518252825382548255825682578258825982608261826282638264826582668267826882698270827182728273827482758276827782788279828082818282828382848285828682878288828982908291829282938294829582968297829882998300830183028303830483058306830783088309831083118312831383148315831683178318831983208321832283238324832583268327832883298330833183328333833483358336833783388339834083418342834383448345834683478348834983508351835283538354835583568357835883598360836183628363836483658366836783688369837083718372837383748375837683778378837983808381838283838384838583868387838883898390839183928393839483958396839783988399840084018402840384048405840684078408840984108411841284138414841584168417841884198420842184228423842484258426842784288429843084318432843384348435843684378438843984408441844284438444844584468447844884498450845184528453845484558456845784588459846084618462846384648465846684678468846984708471847284738474847584768477847884798480848184828483848484858486848784888489849084918492849384948495849684978498849985008501850285038504850585068507850885098510851185128513851485158516851785188519852085218522852385248525852685278528852985308531853285338534853585368537853885398540854185428543854485458546854785488549855085518552855385548555855685578558855985608561856285638564856585668567856885698570857185728573857485758576857785788579858085818582858385848585858685878588858985908591859285938594859585968597859885998600860186028603860486058606860786088609861086118612861386148615861686178618861986208621862286238624862586268627862886298630863186328633863486358636863786388639864086418642864386448645864686478648864986508651865286538654865586568657865886598660866186628663866486658666866786688669867086718672867386748675867686778678867986808681868286838684868586868687868886898690869186928693869486958696869786988699870087018702870387048705870687078708870987108711871287138714871587168717871887198720872187228723872487258726872787288729873087318732873387348735873687378738873987408741874287438744874587468747874887498750875187528753875487558756875787588759876087618762876387648765876687678768876987708771877287738774877587768777877887798780878187828783878487858786878787888789879087918792879387948795879687978798879988008801880288038804880588068807880888098810881188128813881488158816881788188819882088218822882388248825882688278828882988308831883288338834883588368837883888398840884188428843884488458846884788488849885088518852885388548855885688578858885988608861886288638864886588668867886888698870887188728873887488758876887788788879888088818882888388848885888688878888888988908891889288938894889588968897889888998900890189028903890489058906890789088909891089118912891389148915891689178918891989208921892289238924892589268927892889298930893189328933893489358936893789388939894089418942894389448945894689478948894989508951895289538954895589568957895889598960896189628963896489658966896789688969897089718972897389748975897689778978897989808981898289838984898589868987898889898990899189928993899489958996899789988999900090019002900390049005900690079008900990109011901290139014901590169017901890199020902190229023902490259026902790289029903090319032903390349035903690379038903990409041904290439044904590469047904890499050905190529053905490559056905790589059906090619062906390649065906690679068906990709071907290739074907590769077907890799080908190829083908490859086908790889089909090919092909390949095909690979098909991009101910291039104910591069107910891099110911191129113911491159116911791189119912091219122912391249125912691279128912991309131913291339134913591369137913891399140914191429143914491459146914791489149915091519152915391549155915691579158915991609161916291639164916591669167916891699170917191729173917491759176917791789179918091819182918391849185918691879188918991909191919291939194919591969197919891999200920192029203920492059206920792089209921092119212921392149215921692179218921992209221922292239224922592269227922892299230923192329233923492359236923792389239924092419242924392449245924692479248924992509251925292539254925592569257925892599260926192629263926492659266926792689269927092719272927392749275927692779278927992809281928292839284928592869287928892899290929192929293929492959296929792989299930093019302930393049305930693079308930993109311931293139314931593169317931893199320932193229323932493259326932793289329933093319332933393349335933693379338933993409341934293439344934593469347934893499350935193529353935493559356935793589359936093619362936393649365936693679368936993709371937293739374937593769377937893799380938193829383938493859386938793889389939093919392939393949395939693979398939994009401940294039404940594069407940894099410941194129413941494159416941794189419942094219422942394249425942694279428942994309431943294339434943594369437943894399440944194429443944494459446944794489449945094519452945394549455945694579458945994609461946294639464946594669467946894699470947194729473947494759476947794789479948094819482948394849485948694879488948994909491949294939494949594969497949894999500950195029503950495059506950795089509951095119512951395149515951695179518951995209521952295239524952595269527952895299530953195329533953495359536953795389539954095419542954395449545954695479548954995509551955295539554955595569557955895599560956195629563956495659566956795689569957095719572957395749575957695779578957995809581958295839584958595869587958895899590959195929593959495959596959795989599960096019602960396049605960696079608960996109611961296139614961596169617961896199620962196229623962496259626962796289629963096319632963396349635963696379638963996409641964296439644964596469647964896499650965196529653965496559656965796589659966096619662966396649665966696679668966996709671967296739674967596769677967896799680968196829683968496859686968796889689969096919692969396949695969696979698969997009701970297039704970597069707970897099710971197129713971497159716971797189719972097219722972397249725972697279728972997309731973297339734973597369737973897399740974197429743974497459746974797489749975097519752975397549755975697579758975997609761976297639764976597669767976897699770977197729773977497759776977797789779978097819782978397849785978697879788978997909791979297939794979597969797979897999800980198029803980498059806980798089809981098119812981398149815981698179818981998209821982298239824982598269827982898299830983198329833983498359836983798389839984098419842984398449845984698479848984998509851985298539854985598569857985898599860986198629863986498659866986798689869987098719872987398749875987698779878987998809881988298839884988598869887988898899890989198929893989498959896989798989899990099019902990399049905990699079908990999109911991299139914991599169917991899199920992199229923992499259926992799289929993099319932993399349935993699379938993999409941994299439944994599469947994899499950995199529953995499559956995799589959996099619962996399649965996699679968996999709971997299739974997599769977997899799980998199829983998499859986998799889989999099919992999399949995999699979998999910000100011000210003100041000510006100071000810009100101001110012100131001410015100161001710018100191002010021100221002310024100251002610027100281002910030100311003210033100341003510036100371003810039100401004110042100431004410045100461004710048100491005010051100521005310054100551005610057100581005910060100611006210063100641006510066100671006810069100701007110072100731007410075100761007710078100791008010081100821008310084100851008610087100881008910090100911009210093100941009510096100971009810099101001010110102101031010410105101061010710108101091011010111101121011310114101151011610117101181011910120101211012210123101241012510126101271012810129101301013110132101331013410135101361013710138101391014010141101421014310144101451014610147101481014910150101511015210153101541015510156101571015810159101601016110162101631016410165101661016710168101691017010171101721017310174101751017610177101781017910180101811018210183101841018510186101871018810189101901019110192101931019410195101961019710198101991020010201102021020310204102051020610207102081020910210102111021210213102141021510216102171021810219102201022110222102231022410225102261022710228102291023010231102321023310234102351023610237102381023910240102411024210243102441024510246102471024810249102501025110252102531025410255102561025710258102591026010261102621026310264102651026610267102681026910270102711027210273102741027510276102771027810279102801028110282102831028410285102861028710288102891029010291102921029310294102951029610297102981029910300103011030210303103041030510306103071030810309103101031110312103131031410315103161031710318103191032010321103221032310324103251032610327103281032910330103311033210333103341033510336103371033810339103401034110342103431034410345103461034710348103491035010351103521035310354103551035610357103581035910360103611036210363103641036510366103671036810369103701037110372103731037410375103761037710378103791038010381103821038310384103851038610387103881038910390103911039210393103941039510396103971039810399104001040110402104031040410405104061040710408104091041010411104121041310414104151041610417104181041910420104211042210423104241042510426104271042810429104301043110432104331043410435104361043710438104391044010441104421044310444104451044610447104481044910450104511045210453104541045510456104571045810459104601046110462104631046410465104661046710468104691047010471104721047310474104751047610477104781047910480104811048210483104841048510486104871048810489104901049110492104931049410495104961049710498104991050010501105021050310504105051050610507105081050910510105111051210513105141051510516105171051810519105201052110522105231052410525105261052710528105291053010531105321053310534105351053610537105381053910540105411054210543105441054510546105471054810549105501055110552105531055410555105561055710558105591056010561105621056310564105651056610567105681056910570105711057210573105741057510576105771057810579105801058110582105831058410585105861058710588105891059010591105921059310594105951059610597105981059910600106011060210603106041060510606106071060810609106101061110612106131061410615106161061710618106191062010621106221062310624106251062610627106281062910630106311063210633106341063510636106371063810639106401064110642106431064410645106461064710648106491065010651106521065310654106551065610657106581065910660106611066210663106641066510666106671066810669106701067110672106731067410675106761067710678106791068010681106821068310684106851068610687106881068910690106911069210693106941069510696106971069810699107001070110702107031070410705107061070710708107091071010711107121071310714107151071610717107181071910720107211072210723107241072510726107271072810729107301073110732107331073410735107361073710738107391074010741107421074310744107451074610747107481074910750107511075210753107541075510756107571075810759107601076110762107631076410765107661076710768107691077010771107721077310774107751077610777107781077910780107811078210783107841078510786107871078810789107901079110792107931079410795107961079710798107991080010801108021080310804108051080610807108081080910810108111081210813108141081510816108171081810819108201082110822108231082410825108261082710828108291083010831108321083310834108351083610837108381083910840108411084210843108441084510846108471084810849108501085110852108531085410855108561085710858108591086010861108621086310864108651086610867108681086910870108711087210873108741087510876108771087810879108801088110882108831088410885108861088710888108891089010891108921089310894108951089610897108981089910900109011090210903109041090510906109071090810909109101091110912109131091410915109161091710918109191092010921109221092310924109251092610927109281092910930109311093210933109341093510936109371093810939109401094110942109431094410945109461094710948109491095010951109521095310954109551095610957109581095910960109611096210963109641096510966109671096810969109701097110972109731097410975109761097710978109791098010981109821098310984109851098610987109881098910990109911099210993109941099510996109971099810999110001100111002110031100411005110061100711008110091101011011110121101311014110151101611017110181101911020110211102211023110241102511026110271102811029110301103111032110331103411035110361103711038110391104011041110421104311044110451104611047110481104911050110511105211053110541105511056110571105811059110601106111062110631106411065110661106711068110691107011071110721107311074110751107611077110781107911080110811108211083110841108511086110871108811089110901109111092110931109411095110961109711098110991110011101111021110311104111051110611107111081110911110111111111211113111141111511116111171111811119111201112111122111231112411125111261112711128111291113011131111321113311134111351113611137111381113911140111411114211143111441114511146111471114811149111501115111152111531115411155111561115711158111591116011161111621116311164111651116611167111681116911170111711117211173111741117511176111771117811179111801118111182111831118411185111861118711188111891119011191111921119311194111951119611197111981119911200112011120211203112041120511206112071120811209112101121111212112131121411215112161121711218112191122011221112221122311224112251122611227112281122911230112311123211233112341123511236112371123811239112401124111242112431124411245112461124711248112491125011251112521125311254112551125611257112581125911260112611126211263112641126511266112671126811269112701127111272112731127411275112761127711278112791128011281112821128311284112851128611287112881128911290112911129211293112941129511296112971129811299113001130111302113031130411305113061130711308113091131011311113121131311314113151131611317113181131911320113211132211323113241132511326113271132811329113301133111332113331133411335113361133711338113391134011341113421134311344113451134611347113481134911350113511135211353113541135511356113571135811359113601136111362113631136411365113661136711368113691137011371113721137311374113751137611377113781137911380113811138211383113841138511386113871138811389113901139111392113931139411395113961139711398113991140011401114021140311404114051140611407114081140911410114111141211413114141141511416114171141811419114201142111422114231142411425114261142711428114291143011431114321143311434114351143611437114381143911440114411144211443114441144511446114471144811449114501145111452114531145411455114561145711458114591146011461114621146311464114651146611467114681146911470114711147211473114741147511476114771147811479114801148111482114831148411485114861148711488114891149011491114921149311494114951149611497114981149911500115011150211503115041150511506115071150811509115101151111512115131151411515115161151711518115191152011521115221152311524115251152611527115281152911530115311153211533115341153511536115371153811539115401154111542115431154411545115461154711548115491155011551115521155311554115551155611557115581155911560115611156211563115641156511566115671156811569115701157111572115731157411575115761157711578115791158011581115821158311584115851158611587115881158911590115911159211593115941159511596115971159811599116001160111602116031160411605116061160711608116091161011611116121161311614116151161611617116181161911620116211162211623116241162511626116271162811629116301163111632116331163411635116361163711638116391164011641116421164311644116451164611647116481164911650116511165211653116541165511656116571165811659116601166111662116631166411665116661166711668116691167011671116721167311674116751167611677116781167911680116811168211683116841168511686116871168811689116901169111692116931169411695116961169711698116991170011701117021170311704117051170611707117081170911710117111171211713117141171511716117171171811719117201172111722117231172411725117261172711728117291173011731117321173311734117351173611737117381173911740117411174211743117441174511746117471174811749117501175111752117531175411755117561175711758117591176011761117621176311764117651176611767117681176911770117711177211773117741177511776117771177811779117801178111782117831178411785117861178711788117891179011791117921179311794117951179611797117981179911800118011180211803118041180511806118071180811809118101181111812118131181411815118161181711818118191182011821118221182311824118251182611827118281182911830118311183211833118341183511836118371183811839118401184111842118431184411845118461184711848118491185011851118521185311854118551185611857118581185911860118611186211863118641186511866118671186811869118701187111872118731187411875118761187711878118791188011881118821188311884118851188611887118881188911890118911189211893118941189511896118971189811899119001190111902119031190411905119061190711908119091191011911119121191311914119151191611917119181191911920119211192211923119241192511926119271192811929119301193111932119331193411935119361193711938119391194011941119421194311944119451194611947119481194911950119511195211953119541195511956119571195811959119601196111962119631196411965119661196711968119691197011971119721197311974119751197611977119781197911980119811198211983119841198511986119871198811989119901199111992119931199411995119961199711998119991200012001120021200312004120051200612007120081200912010120111201212013120141201512016120171201812019120201202112022120231202412025120261202712028120291203012031120321203312034120351203612037120381203912040120411204212043120441204512046120471204812049120501205112052120531205412055120561205712058120591206012061120621206312064120651206612067120681206912070120711207212073120741207512076120771207812079120801208112082120831208412085120861208712088120891209012091120921209312094120951209612097120981209912100121011210212103121041210512106121071210812109121101211112112121131211412115121161211712118121191212012121121221212312124121251212612127121281212912130121311213212133121341213512136121371213812139121401214112142121431214412145121461214712148121491215012151121521215312154121551215612157121581215912160121611216212163121641216512166121671216812169121701217112172121731217412175121761217712178121791218012181121821218312184121851218612187121881218912190121911219212193121941219512196121971219812199122001220112202122031220412205122061220712208122091221012211122121221312214122151221612217122181221912220122211222212223122241222512226122271222812229122301223112232122331223412235122361223712238122391224012241122421224312244122451224612247122481224912250122511225212253122541225512256122571225812259122601226112262122631226412265122661226712268122691227012271122721227312274122751227612277122781227912280122811228212283122841228512286122871228812289122901229112292122931229412295122961229712298122991230012301123021230312304123051230612307123081230912310123111231212313123141231512316123171231812319123201232112322123231232412325123261232712328123291233012331123321233312334123351233612337123381233912340123411234212343123441234512346123471234812349123501235112352123531235412355123561235712358123591236012361123621236312364123651236612367123681236912370123711237212373123741237512376123771237812379123801238112382123831238412385123861238712388123891239012391123921239312394123951239612397123981239912400124011240212403124041240512406124071240812409124101241112412124131241412415124161241712418124191242012421124221242312424124251242612427124281242912430124311243212433124341243512436124371243812439124401244112442124431244412445124461244712448124491245012451124521245312454124551245612457124581245912460124611246212463124641246512466124671246812469124701247112472124731247412475124761247712478124791248012481124821248312484124851248612487124881248912490124911249212493124941249512496124971249812499125001250112502125031250412505125061250712508125091251012511125121251312514125151251612517125181251912520125211252212523125241252512526125271252812529125301253112532125331253412535125361253712538125391254012541125421254312544125451254612547125481254912550125511255212553125541255512556125571255812559125601256112562125631256412565125661256712568125691257012571125721257312574125751257612577125781257912580125811258212583125841258512586125871258812589125901259112592125931259412595125961259712598125991260012601126021260312604126051260612607126081260912610126111261212613126141261512616126171261812619126201262112622126231262412625126261262712628126291263012631126321263312634126351263612637126381263912640126411264212643126441264512646126471264812649126501265112652126531265412655126561265712658126591266012661126621266312664126651266612667126681266912670126711267212673126741267512676126771267812679126801268112682126831268412685126861268712688126891269012691126921269312694126951269612697126981269912700127011270212703127041270512706127071270812709127101271112712127131271412715127161271712718127191272012721127221272312724127251272612727127281272912730127311273212733127341273512736127371273812739127401274112742127431274412745127461274712748127491275012751127521275312754127551275612757127581275912760127611276212763127641276512766127671276812769127701277112772127731277412775127761277712778127791278012781127821278312784127851278612787127881278912790127911279212793127941279512796127971279812799128001280112802128031280412805128061280712808128091281012811128121281312814128151281612817128181281912820128211282212823128241282512826128271282812829128301283112832128331283412835128361283712838128391284012841128421284312844128451284612847128481284912850128511285212853128541285512856128571285812859128601286112862128631286412865128661286712868128691287012871128721287312874128751287612877128781287912880128811288212883128841288512886128871288812889128901289112892128931289412895128961289712898128991290012901129021290312904129051290612907129081290912910129111291212913129141291512916129171291812919129201292112922129231292412925129261292712928129291293012931129321293312934129351293612937129381293912940129411294212943129441294512946129471294812949129501295112952129531295412955129561295712958129591296012961129621296312964129651296612967129681296912970129711297212973129741297512976129771297812979129801298112982129831298412985129861298712988129891299012991129921299312994129951299612997129981299913000130011300213003130041300513006130071300813009130101301113012130131301413015130161301713018130191302013021130221302313024130251302613027130281302913030130311303213033130341303513036130371303813039130401304113042130431304413045130461304713048130491305013051130521305313054130551305613057130581305913060130611306213063130641306513066130671306813069130701307113072130731307413075130761307713078130791308013081130821308313084130851308613087130881308913090130911309213093130941309513096130971309813099131001310113102131031310413105131061310713108131091311013111131121311313114131151311613117131181311913120131211312213123131241312513126131271312813129131301313113132131331313413135131361313713138131391314013141131421314313144131451314613147131481314913150131511315213153131541315513156131571315813159131601316113162131631316413165131661316713168131691317013171131721317313174131751317613177131781317913180131811318213183131841318513186131871318813189131901319113192131931319413195131961319713198131991320013201132021320313204132051320613207132081320913210132111321213213132141321513216132171321813219132201322113222132231322413225132261322713228132291323013231132321323313234132351323613237132381323913240132411324213243132441324513246132471324813249132501325113252132531325413255132561325713258132591326013261132621326313264132651326613267132681326913270132711327213273132741327513276132771327813279132801328113282132831328413285132861328713288132891329013291132921329313294132951329613297132981329913300133011330213303133041330513306133071330813309133101331113312133131331413315133161331713318133191332013321133221332313324133251332613327133281332913330133311333213333133341333513336133371333813339133401334113342133431334413345133461334713348133491335013351133521335313354133551335613357133581335913360133611336213363133641336513366133671336813369133701337113372133731337413375133761337713378133791338013381133821338313384133851338613387133881338913390133911339213393133941339513396133971339813399134001340113402134031340413405134061340713408134091341013411134121341313414134151341613417134181341913420134211342213423134241342513426134271342813429134301343113432134331343413435134361343713438134391344013441134421344313444134451344613447134481344913450134511345213453134541345513456134571345813459134601346113462134631346413465134661346713468134691347013471134721347313474134751347613477134781347913480134811348213483134841348513486134871348813489134901349113492134931349413495134961349713498134991350013501135021350313504135051350613507135081350913510135111351213513135141351513516135171351813519135201352113522135231352413525135261352713528135291353013531135321353313534135351353613537135381353913540135411354213543135441354513546135471354813549135501355113552135531355413555135561355713558135591356013561135621356313564135651356613567135681356913570135711357213573135741357513576135771357813579135801358113582135831358413585135861358713588135891359013591135921359313594135951359613597135981359913600136011360213603136041360513606136071360813609136101361113612136131361413615136161361713618136191362013621136221362313624136251362613627136281362913630136311363213633136341363513636136371363813639136401364113642136431364413645136461364713648136491365013651136521365313654136551365613657136581365913660136611366213663136641366513666136671366813669136701367113672136731367413675136761367713678136791368013681136821368313684136851368613687136881368913690136911369213693136941369513696136971369813699137001370113702137031370413705137061370713708137091371013711137121371313714137151371613717137181371913720137211372213723137241372513726137271372813729137301373113732137331373413735137361373713738137391374013741137421374313744137451374613747137481374913750137511375213753137541375513756137571375813759137601376113762137631376413765137661376713768137691377013771137721377313774137751377613777137781377913780137811378213783137841378513786137871378813789137901379113792137931379413795137961379713798137991380013801138021380313804138051380613807138081380913810138111381213813138141381513816138171381813819138201382113822138231382413825138261382713828138291383013831138321383313834138351383613837138381383913840138411384213843138441384513846138471384813849138501385113852138531385413855138561385713858138591386013861138621386313864138651386613867138681386913870138711387213873138741387513876138771387813879138801388113882138831388413885138861388713888138891389013891138921389313894138951389613897138981389913900139011390213903139041390513906139071390813909139101391113912139131391413915139161391713918139191392013921139221392313924139251392613927139281392913930139311393213933139341393513936139371393813939139401394113942139431394413945139461394713948139491395013951139521395313954139551395613957139581395913960139611396213963139641396513966139671396813969139701397113972139731397413975139761397713978139791398013981139821398313984139851398613987139881398913990139911399213993139941399513996139971399813999140001400114002140031400414005140061400714008140091401014011140121401314014140151401614017140181401914020140211402214023140241402514026140271402814029140301403114032140331403414035140361403714038140391404014041140421404314044140451404614047140481404914050140511405214053140541405514056140571405814059140601406114062140631406414065140661406714068140691407014071140721407314074140751407614077140781407914080140811408214083140841408514086140871408814089140901409114092140931409414095140961409714098140991410014101141021410314104141051410614107141081410914110141111411214113141141411514116141171411814119141201412114122141231412414125141261412714128141291413014131141321413314134141351413614137141381413914140141411414214143141441414514146141471414814149141501415114152141531415414155141561415714158141591416014161141621416314164141651416614167141681416914170141711417214173141741417514176141771417814179141801418114182141831418414185141861418714188141891419014191141921419314194141951419614197141981419914200142011420214203142041420514206142071420814209142101421114212142131421414215142161421714218142191422014221142221422314224142251422614227142281422914230142311423214233142341423514236142371423814239142401424114242142431424414245142461424714248142491425014251142521425314254142551425614257142581425914260142611426214263142641426514266142671426814269142701427114272142731427414275142761427714278142791428014281142821428314284142851428614287142881428914290142911429214293142941429514296142971429814299143001430114302143031430414305143061430714308143091431014311143121431314314143151431614317143181431914320143211432214323143241432514326143271432814329143301433114332143331433414335143361433714338143391434014341143421434314344143451434614347143481434914350143511435214353143541435514356143571435814359143601436114362143631436414365143661436714368143691437014371143721437314374143751437614377143781437914380143811438214383143841438514386143871438814389143901439114392143931439414395143961439714398143991440014401144021440314404144051440614407144081440914410144111441214413144141441514416144171441814419144201442114422144231442414425144261442714428144291443014431144321443314434144351443614437144381443914440144411444214443144441444514446144471444814449144501445114452144531445414455144561445714458144591446014461144621446314464144651446614467144681446914470
  1. /*
  2. * Copyright 2018-2020 Yury Gribov
  3. *
  4. * The MIT License (MIT)
  5. *
  6. * Use of this source code is governed by MIT license that can be
  7. * found in the LICENSE.txt file.
  8. */
  9. .data
  10. .globl _libcuda_so_tramp_table
  11. .hidden _libcuda_so_tramp_table
  12. .align 8
  13. _libcuda_so_tramp_table:
  14. .zero 3840
  15. .text
  16. .globl _libcuda_so_tramp_resolve
  17. .hidden _libcuda_so_tramp_resolve
  18. .globl _libcuda_so_save_regs_and_resolve
  19. .hidden _libcuda_so_save_regs_and_resolve
  20. .type _libcuda_so_save_regs_and_resolve, %function
  21. _libcuda_so_save_regs_and_resolve:
  22. .cfi_startproc
  23. #define PUSH_REG(reg) pushq %reg ; .cfi_adjust_cfa_offset 8; .cfi_rel_offset reg, 0
  24. #define POP_REG(reg) popq %reg ; .cfi_adjust_cfa_offset -8; .cfi_restore reg
  25. #define DEC_STACK(d) subq $d, %rsp; .cfi_adjust_cfa_offset d
  26. #define INC_STACK(d) addq $d, %rsp; .cfi_adjust_cfa_offset -d
  27. #define PUSH_XMM_REG(reg) DEC_STACK(16); movdqa %reg, (%rsp); .cfi_rel_offset reg, 0
  28. #define POP_XMM_REG(reg) movdqa (%rsp), %reg; .cfi_restore reg; INC_STACK(16)
  29. // Slow path which calls dlsym, taken only on first call.
  30. // All registers are stored to handle arbitrary calling conventions
  31. // (except x87 FPU registers which do not have to be preserved).
  32. // For Dwarf directives, read https://www.imperialviolet.org/2017/01/18/cfi.html.
  33. // FIXME: AVX (YMM, ZMM) registers are NOT saved to simplify code.
  34. PUSH_REG(rdi) // 16
  35. mov 0x10(%rsp), %rdi
  36. PUSH_REG(rax)
  37. PUSH_REG(rbx) // 16
  38. PUSH_REG(rcx)
  39. PUSH_REG(rdx) // 16
  40. PUSH_REG(rbp)
  41. PUSH_REG(rsi) // 16
  42. PUSH_REG(r8)
  43. PUSH_REG(r9) // 16
  44. PUSH_REG(r10)
  45. PUSH_REG(r11) // 16
  46. PUSH_REG(r12)
  47. PUSH_REG(r13) // 16
  48. PUSH_REG(r14)
  49. PUSH_REG(r15) // 16
  50. PUSH_XMM_REG(xmm0)
  51. PUSH_XMM_REG(xmm1)
  52. PUSH_XMM_REG(xmm2)
  53. PUSH_XMM_REG(xmm3)
  54. PUSH_XMM_REG(xmm4)
  55. PUSH_XMM_REG(xmm5)
  56. PUSH_XMM_REG(xmm6)
  57. PUSH_XMM_REG(xmm7)
  58. // Stack is just 8-byte aligned but callee will re-align to 16
  59. call _libcuda_so_tramp_resolve
  60. POP_XMM_REG(xmm7)
  61. POP_XMM_REG(xmm6)
  62. POP_XMM_REG(xmm5)
  63. POP_XMM_REG(xmm4)
  64. POP_XMM_REG(xmm3)
  65. POP_XMM_REG(xmm2)
  66. POP_XMM_REG(xmm1)
  67. POP_XMM_REG(xmm0) // 16
  68. POP_REG(r15)
  69. POP_REG(r14) // 16
  70. POP_REG(r13)
  71. POP_REG(r12) // 16
  72. POP_REG(r11)
  73. POP_REG(r10) // 16
  74. POP_REG(r9)
  75. POP_REG(r8) // 16
  76. POP_REG(rsi)
  77. POP_REG(rbp) // 16
  78. POP_REG(rdx)
  79. POP_REG(rcx) // 16
  80. POP_REG(rbx)
  81. POP_REG(rax) // 16
  82. POP_REG(rdi)
  83. ret
  84. .cfi_endproc
  85. /*
  86. * Copyright 2018-2019 Yury Gribov
  87. *
  88. * The MIT License (MIT)
  89. *
  90. * Use of this source code is governed by MIT license that can be
  91. * found in the LICENSE.txt file.
  92. */
  93. .globl cuArray3DCreate
  94. .p2align 4
  95. .type cuArray3DCreate, %function
  96. cuArray3DCreate:
  97. .cfi_startproc
  98. // Intel opt. manual says to
  99. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  100. // to hint static predictor.
  101. cmpq $0, _libcuda_so_tramp_table+0(%rip)
  102. je 2f
  103. 1:
  104. jmp *_libcuda_so_tramp_table+0(%rip)
  105. 2:
  106. pushq $0
  107. .cfi_adjust_cfa_offset 8
  108. call _libcuda_so_save_regs_and_resolve
  109. addq $8, %rsp
  110. .cfi_adjust_cfa_offset -8
  111. jmp 1b
  112. .cfi_endproc
  113. /*
  114. * Copyright 2018-2019 Yury Gribov
  115. *
  116. * The MIT License (MIT)
  117. *
  118. * Use of this source code is governed by MIT license that can be
  119. * found in the LICENSE.txt file.
  120. */
  121. .globl cuArray3DCreate_v2
  122. .p2align 4
  123. .type cuArray3DCreate_v2, %function
  124. cuArray3DCreate_v2:
  125. .cfi_startproc
  126. // Intel opt. manual says to
  127. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  128. // to hint static predictor.
  129. cmpq $0, _libcuda_so_tramp_table+8(%rip)
  130. je 2f
  131. 1:
  132. jmp *_libcuda_so_tramp_table+8(%rip)
  133. 2:
  134. pushq $1
  135. .cfi_adjust_cfa_offset 8
  136. call _libcuda_so_save_regs_and_resolve
  137. addq $8, %rsp
  138. .cfi_adjust_cfa_offset -8
  139. jmp 1b
  140. .cfi_endproc
  141. /*
  142. * Copyright 2018-2019 Yury Gribov
  143. *
  144. * The MIT License (MIT)
  145. *
  146. * Use of this source code is governed by MIT license that can be
  147. * found in the LICENSE.txt file.
  148. */
  149. .globl cuArray3DGetDescriptor
  150. .p2align 4
  151. .type cuArray3DGetDescriptor, %function
  152. cuArray3DGetDescriptor:
  153. .cfi_startproc
  154. // Intel opt. manual says to
  155. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  156. // to hint static predictor.
  157. cmpq $0, _libcuda_so_tramp_table+16(%rip)
  158. je 2f
  159. 1:
  160. jmp *_libcuda_so_tramp_table+16(%rip)
  161. 2:
  162. pushq $2
  163. .cfi_adjust_cfa_offset 8
  164. call _libcuda_so_save_regs_and_resolve
  165. addq $8, %rsp
  166. .cfi_adjust_cfa_offset -8
  167. jmp 1b
  168. .cfi_endproc
  169. /*
  170. * Copyright 2018-2019 Yury Gribov
  171. *
  172. * The MIT License (MIT)
  173. *
  174. * Use of this source code is governed by MIT license that can be
  175. * found in the LICENSE.txt file.
  176. */
  177. .globl cuArray3DGetDescriptor_v2
  178. .p2align 4
  179. .type cuArray3DGetDescriptor_v2, %function
  180. cuArray3DGetDescriptor_v2:
  181. .cfi_startproc
  182. // Intel opt. manual says to
  183. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  184. // to hint static predictor.
  185. cmpq $0, _libcuda_so_tramp_table+24(%rip)
  186. je 2f
  187. 1:
  188. jmp *_libcuda_so_tramp_table+24(%rip)
  189. 2:
  190. pushq $3
  191. .cfi_adjust_cfa_offset 8
  192. call _libcuda_so_save_regs_and_resolve
  193. addq $8, %rsp
  194. .cfi_adjust_cfa_offset -8
  195. jmp 1b
  196. .cfi_endproc
  197. /*
  198. * Copyright 2018-2019 Yury Gribov
  199. *
  200. * The MIT License (MIT)
  201. *
  202. * Use of this source code is governed by MIT license that can be
  203. * found in the LICENSE.txt file.
  204. */
  205. .globl cuArrayCreate
  206. .p2align 4
  207. .type cuArrayCreate, %function
  208. cuArrayCreate:
  209. .cfi_startproc
  210. // Intel opt. manual says to
  211. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  212. // to hint static predictor.
  213. cmpq $0, _libcuda_so_tramp_table+32(%rip)
  214. je 2f
  215. 1:
  216. jmp *_libcuda_so_tramp_table+32(%rip)
  217. 2:
  218. pushq $4
  219. .cfi_adjust_cfa_offset 8
  220. call _libcuda_so_save_regs_and_resolve
  221. addq $8, %rsp
  222. .cfi_adjust_cfa_offset -8
  223. jmp 1b
  224. .cfi_endproc
  225. /*
  226. * Copyright 2018-2019 Yury Gribov
  227. *
  228. * The MIT License (MIT)
  229. *
  230. * Use of this source code is governed by MIT license that can be
  231. * found in the LICENSE.txt file.
  232. */
  233. .globl cuArrayCreate_v2
  234. .p2align 4
  235. .type cuArrayCreate_v2, %function
  236. cuArrayCreate_v2:
  237. .cfi_startproc
  238. // Intel opt. manual says to
  239. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  240. // to hint static predictor.
  241. cmpq $0, _libcuda_so_tramp_table+40(%rip)
  242. je 2f
  243. 1:
  244. jmp *_libcuda_so_tramp_table+40(%rip)
  245. 2:
  246. pushq $5
  247. .cfi_adjust_cfa_offset 8
  248. call _libcuda_so_save_regs_and_resolve
  249. addq $8, %rsp
  250. .cfi_adjust_cfa_offset -8
  251. jmp 1b
  252. .cfi_endproc
  253. /*
  254. * Copyright 2018-2019 Yury Gribov
  255. *
  256. * The MIT License (MIT)
  257. *
  258. * Use of this source code is governed by MIT license that can be
  259. * found in the LICENSE.txt file.
  260. */
  261. .globl cuArrayDestroy
  262. .p2align 4
  263. .type cuArrayDestroy, %function
  264. cuArrayDestroy:
  265. .cfi_startproc
  266. // Intel opt. manual says to
  267. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  268. // to hint static predictor.
  269. cmpq $0, _libcuda_so_tramp_table+48(%rip)
  270. je 2f
  271. 1:
  272. jmp *_libcuda_so_tramp_table+48(%rip)
  273. 2:
  274. pushq $6
  275. .cfi_adjust_cfa_offset 8
  276. call _libcuda_so_save_regs_and_resolve
  277. addq $8, %rsp
  278. .cfi_adjust_cfa_offset -8
  279. jmp 1b
  280. .cfi_endproc
  281. /*
  282. * Copyright 2018-2019 Yury Gribov
  283. *
  284. * The MIT License (MIT)
  285. *
  286. * Use of this source code is governed by MIT license that can be
  287. * found in the LICENSE.txt file.
  288. */
  289. .globl cuArrayGetDescriptor
  290. .p2align 4
  291. .type cuArrayGetDescriptor, %function
  292. cuArrayGetDescriptor:
  293. .cfi_startproc
  294. // Intel opt. manual says to
  295. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  296. // to hint static predictor.
  297. cmpq $0, _libcuda_so_tramp_table+56(%rip)
  298. je 2f
  299. 1:
  300. jmp *_libcuda_so_tramp_table+56(%rip)
  301. 2:
  302. pushq $7
  303. .cfi_adjust_cfa_offset 8
  304. call _libcuda_so_save_regs_and_resolve
  305. addq $8, %rsp
  306. .cfi_adjust_cfa_offset -8
  307. jmp 1b
  308. .cfi_endproc
  309. /*
  310. * Copyright 2018-2019 Yury Gribov
  311. *
  312. * The MIT License (MIT)
  313. *
  314. * Use of this source code is governed by MIT license that can be
  315. * found in the LICENSE.txt file.
  316. */
  317. .globl cuArrayGetDescriptor_v2
  318. .p2align 4
  319. .type cuArrayGetDescriptor_v2, %function
  320. cuArrayGetDescriptor_v2:
  321. .cfi_startproc
  322. // Intel opt. manual says to
  323. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  324. // to hint static predictor.
  325. cmpq $0, _libcuda_so_tramp_table+64(%rip)
  326. je 2f
  327. 1:
  328. jmp *_libcuda_so_tramp_table+64(%rip)
  329. 2:
  330. pushq $8
  331. .cfi_adjust_cfa_offset 8
  332. call _libcuda_so_save_regs_and_resolve
  333. addq $8, %rsp
  334. .cfi_adjust_cfa_offset -8
  335. jmp 1b
  336. .cfi_endproc
  337. /*
  338. * Copyright 2018-2019 Yury Gribov
  339. *
  340. * The MIT License (MIT)
  341. *
  342. * Use of this source code is governed by MIT license that can be
  343. * found in the LICENSE.txt file.
  344. */
  345. .globl cuArrayGetSparseProperties
  346. .p2align 4
  347. .type cuArrayGetSparseProperties, %function
  348. cuArrayGetSparseProperties:
  349. .cfi_startproc
  350. // Intel opt. manual says to
  351. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  352. // to hint static predictor.
  353. cmpq $0, _libcuda_so_tramp_table+72(%rip)
  354. je 2f
  355. 1:
  356. jmp *_libcuda_so_tramp_table+72(%rip)
  357. 2:
  358. pushq $9
  359. .cfi_adjust_cfa_offset 8
  360. call _libcuda_so_save_regs_and_resolve
  361. addq $8, %rsp
  362. .cfi_adjust_cfa_offset -8
  363. jmp 1b
  364. .cfi_endproc
  365. /*
  366. * Copyright 2018-2019 Yury Gribov
  367. *
  368. * The MIT License (MIT)
  369. *
  370. * Use of this source code is governed by MIT license that can be
  371. * found in the LICENSE.txt file.
  372. */
  373. .globl cuCtxAttach
  374. .p2align 4
  375. .type cuCtxAttach, %function
  376. cuCtxAttach:
  377. .cfi_startproc
  378. // Intel opt. manual says to
  379. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  380. // to hint static predictor.
  381. cmpq $0, _libcuda_so_tramp_table+80(%rip)
  382. je 2f
  383. 1:
  384. jmp *_libcuda_so_tramp_table+80(%rip)
  385. 2:
  386. pushq $10
  387. .cfi_adjust_cfa_offset 8
  388. call _libcuda_so_save_regs_and_resolve
  389. addq $8, %rsp
  390. .cfi_adjust_cfa_offset -8
  391. jmp 1b
  392. .cfi_endproc
  393. /*
  394. * Copyright 2018-2019 Yury Gribov
  395. *
  396. * The MIT License (MIT)
  397. *
  398. * Use of this source code is governed by MIT license that can be
  399. * found in the LICENSE.txt file.
  400. */
  401. .globl cuCtxCreate
  402. .p2align 4
  403. .type cuCtxCreate, %function
  404. cuCtxCreate:
  405. .cfi_startproc
  406. // Intel opt. manual says to
  407. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  408. // to hint static predictor.
  409. cmpq $0, _libcuda_so_tramp_table+88(%rip)
  410. je 2f
  411. 1:
  412. jmp *_libcuda_so_tramp_table+88(%rip)
  413. 2:
  414. pushq $11
  415. .cfi_adjust_cfa_offset 8
  416. call _libcuda_so_save_regs_and_resolve
  417. addq $8, %rsp
  418. .cfi_adjust_cfa_offset -8
  419. jmp 1b
  420. .cfi_endproc
  421. /*
  422. * Copyright 2018-2019 Yury Gribov
  423. *
  424. * The MIT License (MIT)
  425. *
  426. * Use of this source code is governed by MIT license that can be
  427. * found in the LICENSE.txt file.
  428. */
  429. .globl cuCtxCreate_v2
  430. .p2align 4
  431. .type cuCtxCreate_v2, %function
  432. cuCtxCreate_v2:
  433. .cfi_startproc
  434. // Intel opt. manual says to
  435. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  436. // to hint static predictor.
  437. cmpq $0, _libcuda_so_tramp_table+96(%rip)
  438. je 2f
  439. 1:
  440. jmp *_libcuda_so_tramp_table+96(%rip)
  441. 2:
  442. pushq $12
  443. .cfi_adjust_cfa_offset 8
  444. call _libcuda_so_save_regs_and_resolve
  445. addq $8, %rsp
  446. .cfi_adjust_cfa_offset -8
  447. jmp 1b
  448. .cfi_endproc
  449. /*
  450. * Copyright 2018-2019 Yury Gribov
  451. *
  452. * The MIT License (MIT)
  453. *
  454. * Use of this source code is governed by MIT license that can be
  455. * found in the LICENSE.txt file.
  456. */
  457. .globl cuCtxDestroy
  458. .p2align 4
  459. .type cuCtxDestroy, %function
  460. cuCtxDestroy:
  461. .cfi_startproc
  462. // Intel opt. manual says to
  463. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  464. // to hint static predictor.
  465. cmpq $0, _libcuda_so_tramp_table+104(%rip)
  466. je 2f
  467. 1:
  468. jmp *_libcuda_so_tramp_table+104(%rip)
  469. 2:
  470. pushq $13
  471. .cfi_adjust_cfa_offset 8
  472. call _libcuda_so_save_regs_and_resolve
  473. addq $8, %rsp
  474. .cfi_adjust_cfa_offset -8
  475. jmp 1b
  476. .cfi_endproc
  477. /*
  478. * Copyright 2018-2019 Yury Gribov
  479. *
  480. * The MIT License (MIT)
  481. *
  482. * Use of this source code is governed by MIT license that can be
  483. * found in the LICENSE.txt file.
  484. */
  485. .globl cuCtxDestroy_v2
  486. .p2align 4
  487. .type cuCtxDestroy_v2, %function
  488. cuCtxDestroy_v2:
  489. .cfi_startproc
  490. // Intel opt. manual says to
  491. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  492. // to hint static predictor.
  493. cmpq $0, _libcuda_so_tramp_table+112(%rip)
  494. je 2f
  495. 1:
  496. jmp *_libcuda_so_tramp_table+112(%rip)
  497. 2:
  498. pushq $14
  499. .cfi_adjust_cfa_offset 8
  500. call _libcuda_so_save_regs_and_resolve
  501. addq $8, %rsp
  502. .cfi_adjust_cfa_offset -8
  503. jmp 1b
  504. .cfi_endproc
  505. /*
  506. * Copyright 2018-2019 Yury Gribov
  507. *
  508. * The MIT License (MIT)
  509. *
  510. * Use of this source code is governed by MIT license that can be
  511. * found in the LICENSE.txt file.
  512. */
  513. .globl cuCtxDetach
  514. .p2align 4
  515. .type cuCtxDetach, %function
  516. cuCtxDetach:
  517. .cfi_startproc
  518. // Intel opt. manual says to
  519. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  520. // to hint static predictor.
  521. cmpq $0, _libcuda_so_tramp_table+120(%rip)
  522. je 2f
  523. 1:
  524. jmp *_libcuda_so_tramp_table+120(%rip)
  525. 2:
  526. pushq $15
  527. .cfi_adjust_cfa_offset 8
  528. call _libcuda_so_save_regs_and_resolve
  529. addq $8, %rsp
  530. .cfi_adjust_cfa_offset -8
  531. jmp 1b
  532. .cfi_endproc
  533. /*
  534. * Copyright 2018-2019 Yury Gribov
  535. *
  536. * The MIT License (MIT)
  537. *
  538. * Use of this source code is governed by MIT license that can be
  539. * found in the LICENSE.txt file.
  540. */
  541. .globl cuCtxDisablePeerAccess
  542. .p2align 4
  543. .type cuCtxDisablePeerAccess, %function
  544. cuCtxDisablePeerAccess:
  545. .cfi_startproc
  546. // Intel opt. manual says to
  547. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  548. // to hint static predictor.
  549. cmpq $0, _libcuda_so_tramp_table+128(%rip)
  550. je 2f
  551. 1:
  552. jmp *_libcuda_so_tramp_table+128(%rip)
  553. 2:
  554. pushq $16
  555. .cfi_adjust_cfa_offset 8
  556. call _libcuda_so_save_regs_and_resolve
  557. addq $8, %rsp
  558. .cfi_adjust_cfa_offset -8
  559. jmp 1b
  560. .cfi_endproc
  561. /*
  562. * Copyright 2018-2019 Yury Gribov
  563. *
  564. * The MIT License (MIT)
  565. *
  566. * Use of this source code is governed by MIT license that can be
  567. * found in the LICENSE.txt file.
  568. */
  569. .globl cuCtxEnablePeerAccess
  570. .p2align 4
  571. .type cuCtxEnablePeerAccess, %function
  572. cuCtxEnablePeerAccess:
  573. .cfi_startproc
  574. // Intel opt. manual says to
  575. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  576. // to hint static predictor.
  577. cmpq $0, _libcuda_so_tramp_table+136(%rip)
  578. je 2f
  579. 1:
  580. jmp *_libcuda_so_tramp_table+136(%rip)
  581. 2:
  582. pushq $17
  583. .cfi_adjust_cfa_offset 8
  584. call _libcuda_so_save_regs_and_resolve
  585. addq $8, %rsp
  586. .cfi_adjust_cfa_offset -8
  587. jmp 1b
  588. .cfi_endproc
  589. /*
  590. * Copyright 2018-2019 Yury Gribov
  591. *
  592. * The MIT License (MIT)
  593. *
  594. * Use of this source code is governed by MIT license that can be
  595. * found in the LICENSE.txt file.
  596. */
  597. .globl cuCtxGetApiVersion
  598. .p2align 4
  599. .type cuCtxGetApiVersion, %function
  600. cuCtxGetApiVersion:
  601. .cfi_startproc
  602. // Intel opt. manual says to
  603. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  604. // to hint static predictor.
  605. cmpq $0, _libcuda_so_tramp_table+144(%rip)
  606. je 2f
  607. 1:
  608. jmp *_libcuda_so_tramp_table+144(%rip)
  609. 2:
  610. pushq $18
  611. .cfi_adjust_cfa_offset 8
  612. call _libcuda_so_save_regs_and_resolve
  613. addq $8, %rsp
  614. .cfi_adjust_cfa_offset -8
  615. jmp 1b
  616. .cfi_endproc
  617. /*
  618. * Copyright 2018-2019 Yury Gribov
  619. *
  620. * The MIT License (MIT)
  621. *
  622. * Use of this source code is governed by MIT license that can be
  623. * found in the LICENSE.txt file.
  624. */
  625. .globl cuCtxGetCacheConfig
  626. .p2align 4
  627. .type cuCtxGetCacheConfig, %function
  628. cuCtxGetCacheConfig:
  629. .cfi_startproc
  630. // Intel opt. manual says to
  631. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  632. // to hint static predictor.
  633. cmpq $0, _libcuda_so_tramp_table+152(%rip)
  634. je 2f
  635. 1:
  636. jmp *_libcuda_so_tramp_table+152(%rip)
  637. 2:
  638. pushq $19
  639. .cfi_adjust_cfa_offset 8
  640. call _libcuda_so_save_regs_and_resolve
  641. addq $8, %rsp
  642. .cfi_adjust_cfa_offset -8
  643. jmp 1b
  644. .cfi_endproc
  645. /*
  646. * Copyright 2018-2019 Yury Gribov
  647. *
  648. * The MIT License (MIT)
  649. *
  650. * Use of this source code is governed by MIT license that can be
  651. * found in the LICENSE.txt file.
  652. */
  653. .globl cuCtxGetCurrent
  654. .p2align 4
  655. .type cuCtxGetCurrent, %function
  656. cuCtxGetCurrent:
  657. .cfi_startproc
  658. // Intel opt. manual says to
  659. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  660. // to hint static predictor.
  661. cmpq $0, _libcuda_so_tramp_table+160(%rip)
  662. je 2f
  663. 1:
  664. jmp *_libcuda_so_tramp_table+160(%rip)
  665. 2:
  666. pushq $20
  667. .cfi_adjust_cfa_offset 8
  668. call _libcuda_so_save_regs_and_resolve
  669. addq $8, %rsp
  670. .cfi_adjust_cfa_offset -8
  671. jmp 1b
  672. .cfi_endproc
  673. /*
  674. * Copyright 2018-2019 Yury Gribov
  675. *
  676. * The MIT License (MIT)
  677. *
  678. * Use of this source code is governed by MIT license that can be
  679. * found in the LICENSE.txt file.
  680. */
  681. .globl cuCtxGetDevice
  682. .p2align 4
  683. .type cuCtxGetDevice, %function
  684. cuCtxGetDevice:
  685. .cfi_startproc
  686. // Intel opt. manual says to
  687. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  688. // to hint static predictor.
  689. cmpq $0, _libcuda_so_tramp_table+168(%rip)
  690. je 2f
  691. 1:
  692. jmp *_libcuda_so_tramp_table+168(%rip)
  693. 2:
  694. pushq $21
  695. .cfi_adjust_cfa_offset 8
  696. call _libcuda_so_save_regs_and_resolve
  697. addq $8, %rsp
  698. .cfi_adjust_cfa_offset -8
  699. jmp 1b
  700. .cfi_endproc
  701. /*
  702. * Copyright 2018-2019 Yury Gribov
  703. *
  704. * The MIT License (MIT)
  705. *
  706. * Use of this source code is governed by MIT license that can be
  707. * found in the LICENSE.txt file.
  708. */
  709. .globl cuCtxGetFlags
  710. .p2align 4
  711. .type cuCtxGetFlags, %function
  712. cuCtxGetFlags:
  713. .cfi_startproc
  714. // Intel opt. manual says to
  715. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  716. // to hint static predictor.
  717. cmpq $0, _libcuda_so_tramp_table+176(%rip)
  718. je 2f
  719. 1:
  720. jmp *_libcuda_so_tramp_table+176(%rip)
  721. 2:
  722. pushq $22
  723. .cfi_adjust_cfa_offset 8
  724. call _libcuda_so_save_regs_and_resolve
  725. addq $8, %rsp
  726. .cfi_adjust_cfa_offset -8
  727. jmp 1b
  728. .cfi_endproc
  729. /*
  730. * Copyright 2018-2019 Yury Gribov
  731. *
  732. * The MIT License (MIT)
  733. *
  734. * Use of this source code is governed by MIT license that can be
  735. * found in the LICENSE.txt file.
  736. */
  737. .globl cuCtxGetLimit
  738. .p2align 4
  739. .type cuCtxGetLimit, %function
  740. cuCtxGetLimit:
  741. .cfi_startproc
  742. // Intel opt. manual says to
  743. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  744. // to hint static predictor.
  745. cmpq $0, _libcuda_so_tramp_table+184(%rip)
  746. je 2f
  747. 1:
  748. jmp *_libcuda_so_tramp_table+184(%rip)
  749. 2:
  750. pushq $23
  751. .cfi_adjust_cfa_offset 8
  752. call _libcuda_so_save_regs_and_resolve
  753. addq $8, %rsp
  754. .cfi_adjust_cfa_offset -8
  755. jmp 1b
  756. .cfi_endproc
  757. /*
  758. * Copyright 2018-2019 Yury Gribov
  759. *
  760. * The MIT License (MIT)
  761. *
  762. * Use of this source code is governed by MIT license that can be
  763. * found in the LICENSE.txt file.
  764. */
  765. .globl cuCtxGetSharedMemConfig
  766. .p2align 4
  767. .type cuCtxGetSharedMemConfig, %function
  768. cuCtxGetSharedMemConfig:
  769. .cfi_startproc
  770. // Intel opt. manual says to
  771. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  772. // to hint static predictor.
  773. cmpq $0, _libcuda_so_tramp_table+192(%rip)
  774. je 2f
  775. 1:
  776. jmp *_libcuda_so_tramp_table+192(%rip)
  777. 2:
  778. pushq $24
  779. .cfi_adjust_cfa_offset 8
  780. call _libcuda_so_save_regs_and_resolve
  781. addq $8, %rsp
  782. .cfi_adjust_cfa_offset -8
  783. jmp 1b
  784. .cfi_endproc
  785. /*
  786. * Copyright 2018-2019 Yury Gribov
  787. *
  788. * The MIT License (MIT)
  789. *
  790. * Use of this source code is governed by MIT license that can be
  791. * found in the LICENSE.txt file.
  792. */
  793. .globl cuCtxGetStreamPriorityRange
  794. .p2align 4
  795. .type cuCtxGetStreamPriorityRange, %function
  796. cuCtxGetStreamPriorityRange:
  797. .cfi_startproc
  798. // Intel opt. manual says to
  799. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  800. // to hint static predictor.
  801. cmpq $0, _libcuda_so_tramp_table+200(%rip)
  802. je 2f
  803. 1:
  804. jmp *_libcuda_so_tramp_table+200(%rip)
  805. 2:
  806. pushq $25
  807. .cfi_adjust_cfa_offset 8
  808. call _libcuda_so_save_regs_and_resolve
  809. addq $8, %rsp
  810. .cfi_adjust_cfa_offset -8
  811. jmp 1b
  812. .cfi_endproc
  813. /*
  814. * Copyright 2018-2019 Yury Gribov
  815. *
  816. * The MIT License (MIT)
  817. *
  818. * Use of this source code is governed by MIT license that can be
  819. * found in the LICENSE.txt file.
  820. */
  821. .globl cuCtxPopCurrent
  822. .p2align 4
  823. .type cuCtxPopCurrent, %function
  824. cuCtxPopCurrent:
  825. .cfi_startproc
  826. // Intel opt. manual says to
  827. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  828. // to hint static predictor.
  829. cmpq $0, _libcuda_so_tramp_table+208(%rip)
  830. je 2f
  831. 1:
  832. jmp *_libcuda_so_tramp_table+208(%rip)
  833. 2:
  834. pushq $26
  835. .cfi_adjust_cfa_offset 8
  836. call _libcuda_so_save_regs_and_resolve
  837. addq $8, %rsp
  838. .cfi_adjust_cfa_offset -8
  839. jmp 1b
  840. .cfi_endproc
  841. /*
  842. * Copyright 2018-2019 Yury Gribov
  843. *
  844. * The MIT License (MIT)
  845. *
  846. * Use of this source code is governed by MIT license that can be
  847. * found in the LICENSE.txt file.
  848. */
  849. .globl cuCtxPopCurrent_v2
  850. .p2align 4
  851. .type cuCtxPopCurrent_v2, %function
  852. cuCtxPopCurrent_v2:
  853. .cfi_startproc
  854. // Intel opt. manual says to
  855. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  856. // to hint static predictor.
  857. cmpq $0, _libcuda_so_tramp_table+216(%rip)
  858. je 2f
  859. 1:
  860. jmp *_libcuda_so_tramp_table+216(%rip)
  861. 2:
  862. pushq $27
  863. .cfi_adjust_cfa_offset 8
  864. call _libcuda_so_save_regs_and_resolve
  865. addq $8, %rsp
  866. .cfi_adjust_cfa_offset -8
  867. jmp 1b
  868. .cfi_endproc
  869. /*
  870. * Copyright 2018-2019 Yury Gribov
  871. *
  872. * The MIT License (MIT)
  873. *
  874. * Use of this source code is governed by MIT license that can be
  875. * found in the LICENSE.txt file.
  876. */
  877. .globl cuCtxPushCurrent
  878. .p2align 4
  879. .type cuCtxPushCurrent, %function
  880. cuCtxPushCurrent:
  881. .cfi_startproc
  882. // Intel opt. manual says to
  883. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  884. // to hint static predictor.
  885. cmpq $0, _libcuda_so_tramp_table+224(%rip)
  886. je 2f
  887. 1:
  888. jmp *_libcuda_so_tramp_table+224(%rip)
  889. 2:
  890. pushq $28
  891. .cfi_adjust_cfa_offset 8
  892. call _libcuda_so_save_regs_and_resolve
  893. addq $8, %rsp
  894. .cfi_adjust_cfa_offset -8
  895. jmp 1b
  896. .cfi_endproc
  897. /*
  898. * Copyright 2018-2019 Yury Gribov
  899. *
  900. * The MIT License (MIT)
  901. *
  902. * Use of this source code is governed by MIT license that can be
  903. * found in the LICENSE.txt file.
  904. */
  905. .globl cuCtxPushCurrent_v2
  906. .p2align 4
  907. .type cuCtxPushCurrent_v2, %function
  908. cuCtxPushCurrent_v2:
  909. .cfi_startproc
  910. // Intel opt. manual says to
  911. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  912. // to hint static predictor.
  913. cmpq $0, _libcuda_so_tramp_table+232(%rip)
  914. je 2f
  915. 1:
  916. jmp *_libcuda_so_tramp_table+232(%rip)
  917. 2:
  918. pushq $29
  919. .cfi_adjust_cfa_offset 8
  920. call _libcuda_so_save_regs_and_resolve
  921. addq $8, %rsp
  922. .cfi_adjust_cfa_offset -8
  923. jmp 1b
  924. .cfi_endproc
  925. /*
  926. * Copyright 2018-2019 Yury Gribov
  927. *
  928. * The MIT License (MIT)
  929. *
  930. * Use of this source code is governed by MIT license that can be
  931. * found in the LICENSE.txt file.
  932. */
  933. .globl cuCtxResetPersistingL2Cache
  934. .p2align 4
  935. .type cuCtxResetPersistingL2Cache, %function
  936. cuCtxResetPersistingL2Cache:
  937. .cfi_startproc
  938. // Intel opt. manual says to
  939. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  940. // to hint static predictor.
  941. cmpq $0, _libcuda_so_tramp_table+240(%rip)
  942. je 2f
  943. 1:
  944. jmp *_libcuda_so_tramp_table+240(%rip)
  945. 2:
  946. pushq $30
  947. .cfi_adjust_cfa_offset 8
  948. call _libcuda_so_save_regs_and_resolve
  949. addq $8, %rsp
  950. .cfi_adjust_cfa_offset -8
  951. jmp 1b
  952. .cfi_endproc
  953. /*
  954. * Copyright 2018-2019 Yury Gribov
  955. *
  956. * The MIT License (MIT)
  957. *
  958. * Use of this source code is governed by MIT license that can be
  959. * found in the LICENSE.txt file.
  960. */
  961. .globl cuCtxSetCacheConfig
  962. .p2align 4
  963. .type cuCtxSetCacheConfig, %function
  964. cuCtxSetCacheConfig:
  965. .cfi_startproc
  966. // Intel opt. manual says to
  967. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  968. // to hint static predictor.
  969. cmpq $0, _libcuda_so_tramp_table+248(%rip)
  970. je 2f
  971. 1:
  972. jmp *_libcuda_so_tramp_table+248(%rip)
  973. 2:
  974. pushq $31
  975. .cfi_adjust_cfa_offset 8
  976. call _libcuda_so_save_regs_and_resolve
  977. addq $8, %rsp
  978. .cfi_adjust_cfa_offset -8
  979. jmp 1b
  980. .cfi_endproc
  981. /*
  982. * Copyright 2018-2019 Yury Gribov
  983. *
  984. * The MIT License (MIT)
  985. *
  986. * Use of this source code is governed by MIT license that can be
  987. * found in the LICENSE.txt file.
  988. */
  989. .globl cuCtxSetCurrent
  990. .p2align 4
  991. .type cuCtxSetCurrent, %function
  992. cuCtxSetCurrent:
  993. .cfi_startproc
  994. // Intel opt. manual says to
  995. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  996. // to hint static predictor.
  997. cmpq $0, _libcuda_so_tramp_table+256(%rip)
  998. je 2f
  999. 1:
  1000. jmp *_libcuda_so_tramp_table+256(%rip)
  1001. 2:
  1002. pushq $32
  1003. .cfi_adjust_cfa_offset 8
  1004. call _libcuda_so_save_regs_and_resolve
  1005. addq $8, %rsp
  1006. .cfi_adjust_cfa_offset -8
  1007. jmp 1b
  1008. .cfi_endproc
  1009. /*
  1010. * Copyright 2018-2019 Yury Gribov
  1011. *
  1012. * The MIT License (MIT)
  1013. *
  1014. * Use of this source code is governed by MIT license that can be
  1015. * found in the LICENSE.txt file.
  1016. */
  1017. .globl cuCtxSetLimit
  1018. .p2align 4
  1019. .type cuCtxSetLimit, %function
  1020. cuCtxSetLimit:
  1021. .cfi_startproc
  1022. // Intel opt. manual says to
  1023. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1024. // to hint static predictor.
  1025. cmpq $0, _libcuda_so_tramp_table+264(%rip)
  1026. je 2f
  1027. 1:
  1028. jmp *_libcuda_so_tramp_table+264(%rip)
  1029. 2:
  1030. pushq $33
  1031. .cfi_adjust_cfa_offset 8
  1032. call _libcuda_so_save_regs_and_resolve
  1033. addq $8, %rsp
  1034. .cfi_adjust_cfa_offset -8
  1035. jmp 1b
  1036. .cfi_endproc
  1037. /*
  1038. * Copyright 2018-2019 Yury Gribov
  1039. *
  1040. * The MIT License (MIT)
  1041. *
  1042. * Use of this source code is governed by MIT license that can be
  1043. * found in the LICENSE.txt file.
  1044. */
  1045. .globl cuCtxSetSharedMemConfig
  1046. .p2align 4
  1047. .type cuCtxSetSharedMemConfig, %function
  1048. cuCtxSetSharedMemConfig:
  1049. .cfi_startproc
  1050. // Intel opt. manual says to
  1051. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1052. // to hint static predictor.
  1053. cmpq $0, _libcuda_so_tramp_table+272(%rip)
  1054. je 2f
  1055. 1:
  1056. jmp *_libcuda_so_tramp_table+272(%rip)
  1057. 2:
  1058. pushq $34
  1059. .cfi_adjust_cfa_offset 8
  1060. call _libcuda_so_save_regs_and_resolve
  1061. addq $8, %rsp
  1062. .cfi_adjust_cfa_offset -8
  1063. jmp 1b
  1064. .cfi_endproc
  1065. /*
  1066. * Copyright 2018-2019 Yury Gribov
  1067. *
  1068. * The MIT License (MIT)
  1069. *
  1070. * Use of this source code is governed by MIT license that can be
  1071. * found in the LICENSE.txt file.
  1072. */
  1073. .globl cuCtxSynchronize
  1074. .p2align 4
  1075. .type cuCtxSynchronize, %function
  1076. cuCtxSynchronize:
  1077. .cfi_startproc
  1078. // Intel opt. manual says to
  1079. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1080. // to hint static predictor.
  1081. cmpq $0, _libcuda_so_tramp_table+280(%rip)
  1082. je 2f
  1083. 1:
  1084. jmp *_libcuda_so_tramp_table+280(%rip)
  1085. 2:
  1086. pushq $35
  1087. .cfi_adjust_cfa_offset 8
  1088. call _libcuda_so_save_regs_and_resolve
  1089. addq $8, %rsp
  1090. .cfi_adjust_cfa_offset -8
  1091. jmp 1b
  1092. .cfi_endproc
  1093. /*
  1094. * Copyright 2018-2019 Yury Gribov
  1095. *
  1096. * The MIT License (MIT)
  1097. *
  1098. * Use of this source code is governed by MIT license that can be
  1099. * found in the LICENSE.txt file.
  1100. */
  1101. .globl cuDestroyExternalMemory
  1102. .p2align 4
  1103. .type cuDestroyExternalMemory, %function
  1104. cuDestroyExternalMemory:
  1105. .cfi_startproc
  1106. // Intel opt. manual says to
  1107. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1108. // to hint static predictor.
  1109. cmpq $0, _libcuda_so_tramp_table+288(%rip)
  1110. je 2f
  1111. 1:
  1112. jmp *_libcuda_so_tramp_table+288(%rip)
  1113. 2:
  1114. pushq $36
  1115. .cfi_adjust_cfa_offset 8
  1116. call _libcuda_so_save_regs_and_resolve
  1117. addq $8, %rsp
  1118. .cfi_adjust_cfa_offset -8
  1119. jmp 1b
  1120. .cfi_endproc
  1121. /*
  1122. * Copyright 2018-2019 Yury Gribov
  1123. *
  1124. * The MIT License (MIT)
  1125. *
  1126. * Use of this source code is governed by MIT license that can be
  1127. * found in the LICENSE.txt file.
  1128. */
  1129. .globl cuDestroyExternalSemaphore
  1130. .p2align 4
  1131. .type cuDestroyExternalSemaphore, %function
  1132. cuDestroyExternalSemaphore:
  1133. .cfi_startproc
  1134. // Intel opt. manual says to
  1135. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1136. // to hint static predictor.
  1137. cmpq $0, _libcuda_so_tramp_table+296(%rip)
  1138. je 2f
  1139. 1:
  1140. jmp *_libcuda_so_tramp_table+296(%rip)
  1141. 2:
  1142. pushq $37
  1143. .cfi_adjust_cfa_offset 8
  1144. call _libcuda_so_save_regs_and_resolve
  1145. addq $8, %rsp
  1146. .cfi_adjust_cfa_offset -8
  1147. jmp 1b
  1148. .cfi_endproc
  1149. /*
  1150. * Copyright 2018-2019 Yury Gribov
  1151. *
  1152. * The MIT License (MIT)
  1153. *
  1154. * Use of this source code is governed by MIT license that can be
  1155. * found in the LICENSE.txt file.
  1156. */
  1157. .globl cuDeviceCanAccessPeer
  1158. .p2align 4
  1159. .type cuDeviceCanAccessPeer, %function
  1160. cuDeviceCanAccessPeer:
  1161. .cfi_startproc
  1162. // Intel opt. manual says to
  1163. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1164. // to hint static predictor.
  1165. cmpq $0, _libcuda_so_tramp_table+304(%rip)
  1166. je 2f
  1167. 1:
  1168. jmp *_libcuda_so_tramp_table+304(%rip)
  1169. 2:
  1170. pushq $38
  1171. .cfi_adjust_cfa_offset 8
  1172. call _libcuda_so_save_regs_and_resolve
  1173. addq $8, %rsp
  1174. .cfi_adjust_cfa_offset -8
  1175. jmp 1b
  1176. .cfi_endproc
  1177. /*
  1178. * Copyright 2018-2019 Yury Gribov
  1179. *
  1180. * The MIT License (MIT)
  1181. *
  1182. * Use of this source code is governed by MIT license that can be
  1183. * found in the LICENSE.txt file.
  1184. */
  1185. .globl cuDeviceComputeCapability
  1186. .p2align 4
  1187. .type cuDeviceComputeCapability, %function
  1188. cuDeviceComputeCapability:
  1189. .cfi_startproc
  1190. // Intel opt. manual says to
  1191. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1192. // to hint static predictor.
  1193. cmpq $0, _libcuda_so_tramp_table+312(%rip)
  1194. je 2f
  1195. 1:
  1196. jmp *_libcuda_so_tramp_table+312(%rip)
  1197. 2:
  1198. pushq $39
  1199. .cfi_adjust_cfa_offset 8
  1200. call _libcuda_so_save_regs_and_resolve
  1201. addq $8, %rsp
  1202. .cfi_adjust_cfa_offset -8
  1203. jmp 1b
  1204. .cfi_endproc
  1205. /*
  1206. * Copyright 2018-2019 Yury Gribov
  1207. *
  1208. * The MIT License (MIT)
  1209. *
  1210. * Use of this source code is governed by MIT license that can be
  1211. * found in the LICENSE.txt file.
  1212. */
  1213. .globl cuDeviceGet
  1214. .p2align 4
  1215. .type cuDeviceGet, %function
  1216. cuDeviceGet:
  1217. .cfi_startproc
  1218. // Intel opt. manual says to
  1219. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1220. // to hint static predictor.
  1221. cmpq $0, _libcuda_so_tramp_table+320(%rip)
  1222. je 2f
  1223. 1:
  1224. jmp *_libcuda_so_tramp_table+320(%rip)
  1225. 2:
  1226. pushq $40
  1227. .cfi_adjust_cfa_offset 8
  1228. call _libcuda_so_save_regs_and_resolve
  1229. addq $8, %rsp
  1230. .cfi_adjust_cfa_offset -8
  1231. jmp 1b
  1232. .cfi_endproc
  1233. /*
  1234. * Copyright 2018-2019 Yury Gribov
  1235. *
  1236. * The MIT License (MIT)
  1237. *
  1238. * Use of this source code is governed by MIT license that can be
  1239. * found in the LICENSE.txt file.
  1240. */
  1241. .globl cuDeviceGetAttribute
  1242. .p2align 4
  1243. .type cuDeviceGetAttribute, %function
  1244. cuDeviceGetAttribute:
  1245. .cfi_startproc
  1246. // Intel opt. manual says to
  1247. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1248. // to hint static predictor.
  1249. cmpq $0, _libcuda_so_tramp_table+328(%rip)
  1250. je 2f
  1251. 1:
  1252. jmp *_libcuda_so_tramp_table+328(%rip)
  1253. 2:
  1254. pushq $41
  1255. .cfi_adjust_cfa_offset 8
  1256. call _libcuda_so_save_regs_and_resolve
  1257. addq $8, %rsp
  1258. .cfi_adjust_cfa_offset -8
  1259. jmp 1b
  1260. .cfi_endproc
  1261. /*
  1262. * Copyright 2018-2019 Yury Gribov
  1263. *
  1264. * The MIT License (MIT)
  1265. *
  1266. * Use of this source code is governed by MIT license that can be
  1267. * found in the LICENSE.txt file.
  1268. */
  1269. .globl cuDeviceGetByPCIBusId
  1270. .p2align 4
  1271. .type cuDeviceGetByPCIBusId, %function
  1272. cuDeviceGetByPCIBusId:
  1273. .cfi_startproc
  1274. // Intel opt. manual says to
  1275. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1276. // to hint static predictor.
  1277. cmpq $0, _libcuda_so_tramp_table+336(%rip)
  1278. je 2f
  1279. 1:
  1280. jmp *_libcuda_so_tramp_table+336(%rip)
  1281. 2:
  1282. pushq $42
  1283. .cfi_adjust_cfa_offset 8
  1284. call _libcuda_so_save_regs_and_resolve
  1285. addq $8, %rsp
  1286. .cfi_adjust_cfa_offset -8
  1287. jmp 1b
  1288. .cfi_endproc
  1289. /*
  1290. * Copyright 2018-2019 Yury Gribov
  1291. *
  1292. * The MIT License (MIT)
  1293. *
  1294. * Use of this source code is governed by MIT license that can be
  1295. * found in the LICENSE.txt file.
  1296. */
  1297. .globl cuDeviceGetCount
  1298. .p2align 4
  1299. .type cuDeviceGetCount, %function
  1300. cuDeviceGetCount:
  1301. .cfi_startproc
  1302. // Intel opt. manual says to
  1303. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1304. // to hint static predictor.
  1305. cmpq $0, _libcuda_so_tramp_table+344(%rip)
  1306. je 2f
  1307. 1:
  1308. jmp *_libcuda_so_tramp_table+344(%rip)
  1309. 2:
  1310. pushq $43
  1311. .cfi_adjust_cfa_offset 8
  1312. call _libcuda_so_save_regs_and_resolve
  1313. addq $8, %rsp
  1314. .cfi_adjust_cfa_offset -8
  1315. jmp 1b
  1316. .cfi_endproc
  1317. /*
  1318. * Copyright 2018-2019 Yury Gribov
  1319. *
  1320. * The MIT License (MIT)
  1321. *
  1322. * Use of this source code is governed by MIT license that can be
  1323. * found in the LICENSE.txt file.
  1324. */
  1325. .globl cuDeviceGetLuid
  1326. .p2align 4
  1327. .type cuDeviceGetLuid, %function
  1328. cuDeviceGetLuid:
  1329. .cfi_startproc
  1330. // Intel opt. manual says to
  1331. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1332. // to hint static predictor.
  1333. cmpq $0, _libcuda_so_tramp_table+352(%rip)
  1334. je 2f
  1335. 1:
  1336. jmp *_libcuda_so_tramp_table+352(%rip)
  1337. 2:
  1338. pushq $44
  1339. .cfi_adjust_cfa_offset 8
  1340. call _libcuda_so_save_regs_and_resolve
  1341. addq $8, %rsp
  1342. .cfi_adjust_cfa_offset -8
  1343. jmp 1b
  1344. .cfi_endproc
  1345. /*
  1346. * Copyright 2018-2019 Yury Gribov
  1347. *
  1348. * The MIT License (MIT)
  1349. *
  1350. * Use of this source code is governed by MIT license that can be
  1351. * found in the LICENSE.txt file.
  1352. */
  1353. .globl cuDeviceGetName
  1354. .p2align 4
  1355. .type cuDeviceGetName, %function
  1356. cuDeviceGetName:
  1357. .cfi_startproc
  1358. // Intel opt. manual says to
  1359. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1360. // to hint static predictor.
  1361. cmpq $0, _libcuda_so_tramp_table+360(%rip)
  1362. je 2f
  1363. 1:
  1364. jmp *_libcuda_so_tramp_table+360(%rip)
  1365. 2:
  1366. pushq $45
  1367. .cfi_adjust_cfa_offset 8
  1368. call _libcuda_so_save_regs_and_resolve
  1369. addq $8, %rsp
  1370. .cfi_adjust_cfa_offset -8
  1371. jmp 1b
  1372. .cfi_endproc
  1373. /*
  1374. * Copyright 2018-2019 Yury Gribov
  1375. *
  1376. * The MIT License (MIT)
  1377. *
  1378. * Use of this source code is governed by MIT license that can be
  1379. * found in the LICENSE.txt file.
  1380. */
  1381. .globl cuDeviceGetNvSciSyncAttributes
  1382. .p2align 4
  1383. .type cuDeviceGetNvSciSyncAttributes, %function
  1384. cuDeviceGetNvSciSyncAttributes:
  1385. .cfi_startproc
  1386. // Intel opt. manual says to
  1387. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1388. // to hint static predictor.
  1389. cmpq $0, _libcuda_so_tramp_table+368(%rip)
  1390. je 2f
  1391. 1:
  1392. jmp *_libcuda_so_tramp_table+368(%rip)
  1393. 2:
  1394. pushq $46
  1395. .cfi_adjust_cfa_offset 8
  1396. call _libcuda_so_save_regs_and_resolve
  1397. addq $8, %rsp
  1398. .cfi_adjust_cfa_offset -8
  1399. jmp 1b
  1400. .cfi_endproc
  1401. /*
  1402. * Copyright 2018-2019 Yury Gribov
  1403. *
  1404. * The MIT License (MIT)
  1405. *
  1406. * Use of this source code is governed by MIT license that can be
  1407. * found in the LICENSE.txt file.
  1408. */
  1409. .globl cuDeviceGetP2PAttribute
  1410. .p2align 4
  1411. .type cuDeviceGetP2PAttribute, %function
  1412. cuDeviceGetP2PAttribute:
  1413. .cfi_startproc
  1414. // Intel opt. manual says to
  1415. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1416. // to hint static predictor.
  1417. cmpq $0, _libcuda_so_tramp_table+376(%rip)
  1418. je 2f
  1419. 1:
  1420. jmp *_libcuda_so_tramp_table+376(%rip)
  1421. 2:
  1422. pushq $47
  1423. .cfi_adjust_cfa_offset 8
  1424. call _libcuda_so_save_regs_and_resolve
  1425. addq $8, %rsp
  1426. .cfi_adjust_cfa_offset -8
  1427. jmp 1b
  1428. .cfi_endproc
  1429. /*
  1430. * Copyright 2018-2019 Yury Gribov
  1431. *
  1432. * The MIT License (MIT)
  1433. *
  1434. * Use of this source code is governed by MIT license that can be
  1435. * found in the LICENSE.txt file.
  1436. */
  1437. .globl cuDeviceGetPCIBusId
  1438. .p2align 4
  1439. .type cuDeviceGetPCIBusId, %function
  1440. cuDeviceGetPCIBusId:
  1441. .cfi_startproc
  1442. // Intel opt. manual says to
  1443. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1444. // to hint static predictor.
  1445. cmpq $0, _libcuda_so_tramp_table+384(%rip)
  1446. je 2f
  1447. 1:
  1448. jmp *_libcuda_so_tramp_table+384(%rip)
  1449. 2:
  1450. pushq $48
  1451. .cfi_adjust_cfa_offset 8
  1452. call _libcuda_so_save_regs_and_resolve
  1453. addq $8, %rsp
  1454. .cfi_adjust_cfa_offset -8
  1455. jmp 1b
  1456. .cfi_endproc
  1457. /*
  1458. * Copyright 2018-2019 Yury Gribov
  1459. *
  1460. * The MIT License (MIT)
  1461. *
  1462. * Use of this source code is governed by MIT license that can be
  1463. * found in the LICENSE.txt file.
  1464. */
  1465. .globl cuDeviceGetProperties
  1466. .p2align 4
  1467. .type cuDeviceGetProperties, %function
  1468. cuDeviceGetProperties:
  1469. .cfi_startproc
  1470. // Intel opt. manual says to
  1471. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1472. // to hint static predictor.
  1473. cmpq $0, _libcuda_so_tramp_table+392(%rip)
  1474. je 2f
  1475. 1:
  1476. jmp *_libcuda_so_tramp_table+392(%rip)
  1477. 2:
  1478. pushq $49
  1479. .cfi_adjust_cfa_offset 8
  1480. call _libcuda_so_save_regs_and_resolve
  1481. addq $8, %rsp
  1482. .cfi_adjust_cfa_offset -8
  1483. jmp 1b
  1484. .cfi_endproc
  1485. /*
  1486. * Copyright 2018-2019 Yury Gribov
  1487. *
  1488. * The MIT License (MIT)
  1489. *
  1490. * Use of this source code is governed by MIT license that can be
  1491. * found in the LICENSE.txt file.
  1492. */
  1493. .globl cuDeviceGetTexture1DLinearMaxWidth
  1494. .p2align 4
  1495. .type cuDeviceGetTexture1DLinearMaxWidth, %function
  1496. cuDeviceGetTexture1DLinearMaxWidth:
  1497. .cfi_startproc
  1498. // Intel opt. manual says to
  1499. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1500. // to hint static predictor.
  1501. cmpq $0, _libcuda_so_tramp_table+400(%rip)
  1502. je 2f
  1503. 1:
  1504. jmp *_libcuda_so_tramp_table+400(%rip)
  1505. 2:
  1506. pushq $50
  1507. .cfi_adjust_cfa_offset 8
  1508. call _libcuda_so_save_regs_and_resolve
  1509. addq $8, %rsp
  1510. .cfi_adjust_cfa_offset -8
  1511. jmp 1b
  1512. .cfi_endproc
  1513. /*
  1514. * Copyright 2018-2019 Yury Gribov
  1515. *
  1516. * The MIT License (MIT)
  1517. *
  1518. * Use of this source code is governed by MIT license that can be
  1519. * found in the LICENSE.txt file.
  1520. */
  1521. .globl cuDeviceGetUuid
  1522. .p2align 4
  1523. .type cuDeviceGetUuid, %function
  1524. cuDeviceGetUuid:
  1525. .cfi_startproc
  1526. // Intel opt. manual says to
  1527. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1528. // to hint static predictor.
  1529. cmpq $0, _libcuda_so_tramp_table+408(%rip)
  1530. je 2f
  1531. 1:
  1532. jmp *_libcuda_so_tramp_table+408(%rip)
  1533. 2:
  1534. pushq $51
  1535. .cfi_adjust_cfa_offset 8
  1536. call _libcuda_so_save_regs_and_resolve
  1537. addq $8, %rsp
  1538. .cfi_adjust_cfa_offset -8
  1539. jmp 1b
  1540. .cfi_endproc
  1541. /*
  1542. * Copyright 2018-2019 Yury Gribov
  1543. *
  1544. * The MIT License (MIT)
  1545. *
  1546. * Use of this source code is governed by MIT license that can be
  1547. * found in the LICENSE.txt file.
  1548. */
  1549. .globl cuDevicePrimaryCtxGetState
  1550. .p2align 4
  1551. .type cuDevicePrimaryCtxGetState, %function
  1552. cuDevicePrimaryCtxGetState:
  1553. .cfi_startproc
  1554. // Intel opt. manual says to
  1555. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1556. // to hint static predictor.
  1557. cmpq $0, _libcuda_so_tramp_table+416(%rip)
  1558. je 2f
  1559. 1:
  1560. jmp *_libcuda_so_tramp_table+416(%rip)
  1561. 2:
  1562. pushq $52
  1563. .cfi_adjust_cfa_offset 8
  1564. call _libcuda_so_save_regs_and_resolve
  1565. addq $8, %rsp
  1566. .cfi_adjust_cfa_offset -8
  1567. jmp 1b
  1568. .cfi_endproc
  1569. /*
  1570. * Copyright 2018-2019 Yury Gribov
  1571. *
  1572. * The MIT License (MIT)
  1573. *
  1574. * Use of this source code is governed by MIT license that can be
  1575. * found in the LICENSE.txt file.
  1576. */
  1577. .globl cuDevicePrimaryCtxRelease
  1578. .p2align 4
  1579. .type cuDevicePrimaryCtxRelease, %function
  1580. cuDevicePrimaryCtxRelease:
  1581. .cfi_startproc
  1582. // Intel opt. manual says to
  1583. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1584. // to hint static predictor.
  1585. cmpq $0, _libcuda_so_tramp_table+424(%rip)
  1586. je 2f
  1587. 1:
  1588. jmp *_libcuda_so_tramp_table+424(%rip)
  1589. 2:
  1590. pushq $53
  1591. .cfi_adjust_cfa_offset 8
  1592. call _libcuda_so_save_regs_and_resolve
  1593. addq $8, %rsp
  1594. .cfi_adjust_cfa_offset -8
  1595. jmp 1b
  1596. .cfi_endproc
  1597. /*
  1598. * Copyright 2018-2019 Yury Gribov
  1599. *
  1600. * The MIT License (MIT)
  1601. *
  1602. * Use of this source code is governed by MIT license that can be
  1603. * found in the LICENSE.txt file.
  1604. */
  1605. .globl cuDevicePrimaryCtxRelease_v2
  1606. .p2align 4
  1607. .type cuDevicePrimaryCtxRelease_v2, %function
  1608. cuDevicePrimaryCtxRelease_v2:
  1609. .cfi_startproc
  1610. // Intel opt. manual says to
  1611. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1612. // to hint static predictor.
  1613. cmpq $0, _libcuda_so_tramp_table+432(%rip)
  1614. je 2f
  1615. 1:
  1616. jmp *_libcuda_so_tramp_table+432(%rip)
  1617. 2:
  1618. pushq $54
  1619. .cfi_adjust_cfa_offset 8
  1620. call _libcuda_so_save_regs_and_resolve
  1621. addq $8, %rsp
  1622. .cfi_adjust_cfa_offset -8
  1623. jmp 1b
  1624. .cfi_endproc
  1625. /*
  1626. * Copyright 2018-2019 Yury Gribov
  1627. *
  1628. * The MIT License (MIT)
  1629. *
  1630. * Use of this source code is governed by MIT license that can be
  1631. * found in the LICENSE.txt file.
  1632. */
  1633. .globl cuDevicePrimaryCtxReset
  1634. .p2align 4
  1635. .type cuDevicePrimaryCtxReset, %function
  1636. cuDevicePrimaryCtxReset:
  1637. .cfi_startproc
  1638. // Intel opt. manual says to
  1639. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1640. // to hint static predictor.
  1641. cmpq $0, _libcuda_so_tramp_table+440(%rip)
  1642. je 2f
  1643. 1:
  1644. jmp *_libcuda_so_tramp_table+440(%rip)
  1645. 2:
  1646. pushq $55
  1647. .cfi_adjust_cfa_offset 8
  1648. call _libcuda_so_save_regs_and_resolve
  1649. addq $8, %rsp
  1650. .cfi_adjust_cfa_offset -8
  1651. jmp 1b
  1652. .cfi_endproc
  1653. /*
  1654. * Copyright 2018-2019 Yury Gribov
  1655. *
  1656. * The MIT License (MIT)
  1657. *
  1658. * Use of this source code is governed by MIT license that can be
  1659. * found in the LICENSE.txt file.
  1660. */
  1661. .globl cuDevicePrimaryCtxReset_v2
  1662. .p2align 4
  1663. .type cuDevicePrimaryCtxReset_v2, %function
  1664. cuDevicePrimaryCtxReset_v2:
  1665. .cfi_startproc
  1666. // Intel opt. manual says to
  1667. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1668. // to hint static predictor.
  1669. cmpq $0, _libcuda_so_tramp_table+448(%rip)
  1670. je 2f
  1671. 1:
  1672. jmp *_libcuda_so_tramp_table+448(%rip)
  1673. 2:
  1674. pushq $56
  1675. .cfi_adjust_cfa_offset 8
  1676. call _libcuda_so_save_regs_and_resolve
  1677. addq $8, %rsp
  1678. .cfi_adjust_cfa_offset -8
  1679. jmp 1b
  1680. .cfi_endproc
  1681. /*
  1682. * Copyright 2018-2019 Yury Gribov
  1683. *
  1684. * The MIT License (MIT)
  1685. *
  1686. * Use of this source code is governed by MIT license that can be
  1687. * found in the LICENSE.txt file.
  1688. */
  1689. .globl cuDevicePrimaryCtxRetain
  1690. .p2align 4
  1691. .type cuDevicePrimaryCtxRetain, %function
  1692. cuDevicePrimaryCtxRetain:
  1693. .cfi_startproc
  1694. // Intel opt. manual says to
  1695. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1696. // to hint static predictor.
  1697. cmpq $0, _libcuda_so_tramp_table+456(%rip)
  1698. je 2f
  1699. 1:
  1700. jmp *_libcuda_so_tramp_table+456(%rip)
  1701. 2:
  1702. pushq $57
  1703. .cfi_adjust_cfa_offset 8
  1704. call _libcuda_so_save_regs_and_resolve
  1705. addq $8, %rsp
  1706. .cfi_adjust_cfa_offset -8
  1707. jmp 1b
  1708. .cfi_endproc
  1709. /*
  1710. * Copyright 2018-2019 Yury Gribov
  1711. *
  1712. * The MIT License (MIT)
  1713. *
  1714. * Use of this source code is governed by MIT license that can be
  1715. * found in the LICENSE.txt file.
  1716. */
  1717. .globl cuDevicePrimaryCtxSetFlags
  1718. .p2align 4
  1719. .type cuDevicePrimaryCtxSetFlags, %function
  1720. cuDevicePrimaryCtxSetFlags:
  1721. .cfi_startproc
  1722. // Intel opt. manual says to
  1723. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1724. // to hint static predictor.
  1725. cmpq $0, _libcuda_so_tramp_table+464(%rip)
  1726. je 2f
  1727. 1:
  1728. jmp *_libcuda_so_tramp_table+464(%rip)
  1729. 2:
  1730. pushq $58
  1731. .cfi_adjust_cfa_offset 8
  1732. call _libcuda_so_save_regs_and_resolve
  1733. addq $8, %rsp
  1734. .cfi_adjust_cfa_offset -8
  1735. jmp 1b
  1736. .cfi_endproc
  1737. /*
  1738. * Copyright 2018-2019 Yury Gribov
  1739. *
  1740. * The MIT License (MIT)
  1741. *
  1742. * Use of this source code is governed by MIT license that can be
  1743. * found in the LICENSE.txt file.
  1744. */
  1745. .globl cuDevicePrimaryCtxSetFlags_v2
  1746. .p2align 4
  1747. .type cuDevicePrimaryCtxSetFlags_v2, %function
  1748. cuDevicePrimaryCtxSetFlags_v2:
  1749. .cfi_startproc
  1750. // Intel opt. manual says to
  1751. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1752. // to hint static predictor.
  1753. cmpq $0, _libcuda_so_tramp_table+472(%rip)
  1754. je 2f
  1755. 1:
  1756. jmp *_libcuda_so_tramp_table+472(%rip)
  1757. 2:
  1758. pushq $59
  1759. .cfi_adjust_cfa_offset 8
  1760. call _libcuda_so_save_regs_and_resolve
  1761. addq $8, %rsp
  1762. .cfi_adjust_cfa_offset -8
  1763. jmp 1b
  1764. .cfi_endproc
  1765. /*
  1766. * Copyright 2018-2019 Yury Gribov
  1767. *
  1768. * The MIT License (MIT)
  1769. *
  1770. * Use of this source code is governed by MIT license that can be
  1771. * found in the LICENSE.txt file.
  1772. */
  1773. .globl cuDeviceTotalMem
  1774. .p2align 4
  1775. .type cuDeviceTotalMem, %function
  1776. cuDeviceTotalMem:
  1777. .cfi_startproc
  1778. // Intel opt. manual says to
  1779. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1780. // to hint static predictor.
  1781. cmpq $0, _libcuda_so_tramp_table+480(%rip)
  1782. je 2f
  1783. 1:
  1784. jmp *_libcuda_so_tramp_table+480(%rip)
  1785. 2:
  1786. pushq $60
  1787. .cfi_adjust_cfa_offset 8
  1788. call _libcuda_so_save_regs_and_resolve
  1789. addq $8, %rsp
  1790. .cfi_adjust_cfa_offset -8
  1791. jmp 1b
  1792. .cfi_endproc
  1793. /*
  1794. * Copyright 2018-2019 Yury Gribov
  1795. *
  1796. * The MIT License (MIT)
  1797. *
  1798. * Use of this source code is governed by MIT license that can be
  1799. * found in the LICENSE.txt file.
  1800. */
  1801. .globl cuDeviceTotalMem_v2
  1802. .p2align 4
  1803. .type cuDeviceTotalMem_v2, %function
  1804. cuDeviceTotalMem_v2:
  1805. .cfi_startproc
  1806. // Intel opt. manual says to
  1807. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1808. // to hint static predictor.
  1809. cmpq $0, _libcuda_so_tramp_table+488(%rip)
  1810. je 2f
  1811. 1:
  1812. jmp *_libcuda_so_tramp_table+488(%rip)
  1813. 2:
  1814. pushq $61
  1815. .cfi_adjust_cfa_offset 8
  1816. call _libcuda_so_save_regs_and_resolve
  1817. addq $8, %rsp
  1818. .cfi_adjust_cfa_offset -8
  1819. jmp 1b
  1820. .cfi_endproc
  1821. /*
  1822. * Copyright 2018-2019 Yury Gribov
  1823. *
  1824. * The MIT License (MIT)
  1825. *
  1826. * Use of this source code is governed by MIT license that can be
  1827. * found in the LICENSE.txt file.
  1828. */
  1829. .globl cuDriverGetVersion
  1830. .p2align 4
  1831. .type cuDriverGetVersion, %function
  1832. cuDriverGetVersion:
  1833. .cfi_startproc
  1834. // Intel opt. manual says to
  1835. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1836. // to hint static predictor.
  1837. cmpq $0, _libcuda_so_tramp_table+496(%rip)
  1838. je 2f
  1839. 1:
  1840. jmp *_libcuda_so_tramp_table+496(%rip)
  1841. 2:
  1842. pushq $62
  1843. .cfi_adjust_cfa_offset 8
  1844. call _libcuda_so_save_regs_and_resolve
  1845. addq $8, %rsp
  1846. .cfi_adjust_cfa_offset -8
  1847. jmp 1b
  1848. .cfi_endproc
  1849. /*
  1850. * Copyright 2018-2019 Yury Gribov
  1851. *
  1852. * The MIT License (MIT)
  1853. *
  1854. * Use of this source code is governed by MIT license that can be
  1855. * found in the LICENSE.txt file.
  1856. */
  1857. .globl cuEGLApiInit
  1858. .p2align 4
  1859. .type cuEGLApiInit, %function
  1860. cuEGLApiInit:
  1861. .cfi_startproc
  1862. // Intel opt. manual says to
  1863. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1864. // to hint static predictor.
  1865. cmpq $0, _libcuda_so_tramp_table+504(%rip)
  1866. je 2f
  1867. 1:
  1868. jmp *_libcuda_so_tramp_table+504(%rip)
  1869. 2:
  1870. pushq $63
  1871. .cfi_adjust_cfa_offset 8
  1872. call _libcuda_so_save_regs_and_resolve
  1873. addq $8, %rsp
  1874. .cfi_adjust_cfa_offset -8
  1875. jmp 1b
  1876. .cfi_endproc
  1877. /*
  1878. * Copyright 2018-2019 Yury Gribov
  1879. *
  1880. * The MIT License (MIT)
  1881. *
  1882. * Use of this source code is governed by MIT license that can be
  1883. * found in the LICENSE.txt file.
  1884. */
  1885. .globl cuEGLStreamConsumerAcquireFrame
  1886. .p2align 4
  1887. .type cuEGLStreamConsumerAcquireFrame, %function
  1888. cuEGLStreamConsumerAcquireFrame:
  1889. .cfi_startproc
  1890. // Intel opt. manual says to
  1891. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1892. // to hint static predictor.
  1893. cmpq $0, _libcuda_so_tramp_table+512(%rip)
  1894. je 2f
  1895. 1:
  1896. jmp *_libcuda_so_tramp_table+512(%rip)
  1897. 2:
  1898. pushq $64
  1899. .cfi_adjust_cfa_offset 8
  1900. call _libcuda_so_save_regs_and_resolve
  1901. addq $8, %rsp
  1902. .cfi_adjust_cfa_offset -8
  1903. jmp 1b
  1904. .cfi_endproc
  1905. /*
  1906. * Copyright 2018-2019 Yury Gribov
  1907. *
  1908. * The MIT License (MIT)
  1909. *
  1910. * Use of this source code is governed by MIT license that can be
  1911. * found in the LICENSE.txt file.
  1912. */
  1913. .globl cuEGLStreamConsumerConnect
  1914. .p2align 4
  1915. .type cuEGLStreamConsumerConnect, %function
  1916. cuEGLStreamConsumerConnect:
  1917. .cfi_startproc
  1918. // Intel opt. manual says to
  1919. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1920. // to hint static predictor.
  1921. cmpq $0, _libcuda_so_tramp_table+520(%rip)
  1922. je 2f
  1923. 1:
  1924. jmp *_libcuda_so_tramp_table+520(%rip)
  1925. 2:
  1926. pushq $65
  1927. .cfi_adjust_cfa_offset 8
  1928. call _libcuda_so_save_regs_and_resolve
  1929. addq $8, %rsp
  1930. .cfi_adjust_cfa_offset -8
  1931. jmp 1b
  1932. .cfi_endproc
  1933. /*
  1934. * Copyright 2018-2019 Yury Gribov
  1935. *
  1936. * The MIT License (MIT)
  1937. *
  1938. * Use of this source code is governed by MIT license that can be
  1939. * found in the LICENSE.txt file.
  1940. */
  1941. .globl cuEGLStreamConsumerConnectWithFlags
  1942. .p2align 4
  1943. .type cuEGLStreamConsumerConnectWithFlags, %function
  1944. cuEGLStreamConsumerConnectWithFlags:
  1945. .cfi_startproc
  1946. // Intel opt. manual says to
  1947. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1948. // to hint static predictor.
  1949. cmpq $0, _libcuda_so_tramp_table+528(%rip)
  1950. je 2f
  1951. 1:
  1952. jmp *_libcuda_so_tramp_table+528(%rip)
  1953. 2:
  1954. pushq $66
  1955. .cfi_adjust_cfa_offset 8
  1956. call _libcuda_so_save_regs_and_resolve
  1957. addq $8, %rsp
  1958. .cfi_adjust_cfa_offset -8
  1959. jmp 1b
  1960. .cfi_endproc
  1961. /*
  1962. * Copyright 2018-2019 Yury Gribov
  1963. *
  1964. * The MIT License (MIT)
  1965. *
  1966. * Use of this source code is governed by MIT license that can be
  1967. * found in the LICENSE.txt file.
  1968. */
  1969. .globl cuEGLStreamConsumerDisconnect
  1970. .p2align 4
  1971. .type cuEGLStreamConsumerDisconnect, %function
  1972. cuEGLStreamConsumerDisconnect:
  1973. .cfi_startproc
  1974. // Intel opt. manual says to
  1975. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  1976. // to hint static predictor.
  1977. cmpq $0, _libcuda_so_tramp_table+536(%rip)
  1978. je 2f
  1979. 1:
  1980. jmp *_libcuda_so_tramp_table+536(%rip)
  1981. 2:
  1982. pushq $67
  1983. .cfi_adjust_cfa_offset 8
  1984. call _libcuda_so_save_regs_and_resolve
  1985. addq $8, %rsp
  1986. .cfi_adjust_cfa_offset -8
  1987. jmp 1b
  1988. .cfi_endproc
  1989. /*
  1990. * Copyright 2018-2019 Yury Gribov
  1991. *
  1992. * The MIT License (MIT)
  1993. *
  1994. * Use of this source code is governed by MIT license that can be
  1995. * found in the LICENSE.txt file.
  1996. */
  1997. .globl cuEGLStreamConsumerReleaseFrame
  1998. .p2align 4
  1999. .type cuEGLStreamConsumerReleaseFrame, %function
  2000. cuEGLStreamConsumerReleaseFrame:
  2001. .cfi_startproc
  2002. // Intel opt. manual says to
  2003. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2004. // to hint static predictor.
  2005. cmpq $0, _libcuda_so_tramp_table+544(%rip)
  2006. je 2f
  2007. 1:
  2008. jmp *_libcuda_so_tramp_table+544(%rip)
  2009. 2:
  2010. pushq $68
  2011. .cfi_adjust_cfa_offset 8
  2012. call _libcuda_so_save_regs_and_resolve
  2013. addq $8, %rsp
  2014. .cfi_adjust_cfa_offset -8
  2015. jmp 1b
  2016. .cfi_endproc
  2017. /*
  2018. * Copyright 2018-2019 Yury Gribov
  2019. *
  2020. * The MIT License (MIT)
  2021. *
  2022. * Use of this source code is governed by MIT license that can be
  2023. * found in the LICENSE.txt file.
  2024. */
  2025. .globl cuEGLStreamProducerConnect
  2026. .p2align 4
  2027. .type cuEGLStreamProducerConnect, %function
  2028. cuEGLStreamProducerConnect:
  2029. .cfi_startproc
  2030. // Intel opt. manual says to
  2031. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2032. // to hint static predictor.
  2033. cmpq $0, _libcuda_so_tramp_table+552(%rip)
  2034. je 2f
  2035. 1:
  2036. jmp *_libcuda_so_tramp_table+552(%rip)
  2037. 2:
  2038. pushq $69
  2039. .cfi_adjust_cfa_offset 8
  2040. call _libcuda_so_save_regs_and_resolve
  2041. addq $8, %rsp
  2042. .cfi_adjust_cfa_offset -8
  2043. jmp 1b
  2044. .cfi_endproc
  2045. /*
  2046. * Copyright 2018-2019 Yury Gribov
  2047. *
  2048. * The MIT License (MIT)
  2049. *
  2050. * Use of this source code is governed by MIT license that can be
  2051. * found in the LICENSE.txt file.
  2052. */
  2053. .globl cuEGLStreamProducerDisconnect
  2054. .p2align 4
  2055. .type cuEGLStreamProducerDisconnect, %function
  2056. cuEGLStreamProducerDisconnect:
  2057. .cfi_startproc
  2058. // Intel opt. manual says to
  2059. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2060. // to hint static predictor.
  2061. cmpq $0, _libcuda_so_tramp_table+560(%rip)
  2062. je 2f
  2063. 1:
  2064. jmp *_libcuda_so_tramp_table+560(%rip)
  2065. 2:
  2066. pushq $70
  2067. .cfi_adjust_cfa_offset 8
  2068. call _libcuda_so_save_regs_and_resolve
  2069. addq $8, %rsp
  2070. .cfi_adjust_cfa_offset -8
  2071. jmp 1b
  2072. .cfi_endproc
  2073. /*
  2074. * Copyright 2018-2019 Yury Gribov
  2075. *
  2076. * The MIT License (MIT)
  2077. *
  2078. * Use of this source code is governed by MIT license that can be
  2079. * found in the LICENSE.txt file.
  2080. */
  2081. .globl cuEGLStreamProducerPresentFrame
  2082. .p2align 4
  2083. .type cuEGLStreamProducerPresentFrame, %function
  2084. cuEGLStreamProducerPresentFrame:
  2085. .cfi_startproc
  2086. // Intel opt. manual says to
  2087. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2088. // to hint static predictor.
  2089. cmpq $0, _libcuda_so_tramp_table+568(%rip)
  2090. je 2f
  2091. 1:
  2092. jmp *_libcuda_so_tramp_table+568(%rip)
  2093. 2:
  2094. pushq $71
  2095. .cfi_adjust_cfa_offset 8
  2096. call _libcuda_so_save_regs_and_resolve
  2097. addq $8, %rsp
  2098. .cfi_adjust_cfa_offset -8
  2099. jmp 1b
  2100. .cfi_endproc
  2101. /*
  2102. * Copyright 2018-2019 Yury Gribov
  2103. *
  2104. * The MIT License (MIT)
  2105. *
  2106. * Use of this source code is governed by MIT license that can be
  2107. * found in the LICENSE.txt file.
  2108. */
  2109. .globl cuEGLStreamProducerReturnFrame
  2110. .p2align 4
  2111. .type cuEGLStreamProducerReturnFrame, %function
  2112. cuEGLStreamProducerReturnFrame:
  2113. .cfi_startproc
  2114. // Intel opt. manual says to
  2115. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2116. // to hint static predictor.
  2117. cmpq $0, _libcuda_so_tramp_table+576(%rip)
  2118. je 2f
  2119. 1:
  2120. jmp *_libcuda_so_tramp_table+576(%rip)
  2121. 2:
  2122. pushq $72
  2123. .cfi_adjust_cfa_offset 8
  2124. call _libcuda_so_save_regs_and_resolve
  2125. addq $8, %rsp
  2126. .cfi_adjust_cfa_offset -8
  2127. jmp 1b
  2128. .cfi_endproc
  2129. /*
  2130. * Copyright 2018-2019 Yury Gribov
  2131. *
  2132. * The MIT License (MIT)
  2133. *
  2134. * Use of this source code is governed by MIT license that can be
  2135. * found in the LICENSE.txt file.
  2136. */
  2137. .globl cuEventCreate
  2138. .p2align 4
  2139. .type cuEventCreate, %function
  2140. cuEventCreate:
  2141. .cfi_startproc
  2142. // Intel opt. manual says to
  2143. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2144. // to hint static predictor.
  2145. cmpq $0, _libcuda_so_tramp_table+584(%rip)
  2146. je 2f
  2147. 1:
  2148. jmp *_libcuda_so_tramp_table+584(%rip)
  2149. 2:
  2150. pushq $73
  2151. .cfi_adjust_cfa_offset 8
  2152. call _libcuda_so_save_regs_and_resolve
  2153. addq $8, %rsp
  2154. .cfi_adjust_cfa_offset -8
  2155. jmp 1b
  2156. .cfi_endproc
  2157. /*
  2158. * Copyright 2018-2019 Yury Gribov
  2159. *
  2160. * The MIT License (MIT)
  2161. *
  2162. * Use of this source code is governed by MIT license that can be
  2163. * found in the LICENSE.txt file.
  2164. */
  2165. .globl cuEventDestroy
  2166. .p2align 4
  2167. .type cuEventDestroy, %function
  2168. cuEventDestroy:
  2169. .cfi_startproc
  2170. // Intel opt. manual says to
  2171. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2172. // to hint static predictor.
  2173. cmpq $0, _libcuda_so_tramp_table+592(%rip)
  2174. je 2f
  2175. 1:
  2176. jmp *_libcuda_so_tramp_table+592(%rip)
  2177. 2:
  2178. pushq $74
  2179. .cfi_adjust_cfa_offset 8
  2180. call _libcuda_so_save_regs_and_resolve
  2181. addq $8, %rsp
  2182. .cfi_adjust_cfa_offset -8
  2183. jmp 1b
  2184. .cfi_endproc
  2185. /*
  2186. * Copyright 2018-2019 Yury Gribov
  2187. *
  2188. * The MIT License (MIT)
  2189. *
  2190. * Use of this source code is governed by MIT license that can be
  2191. * found in the LICENSE.txt file.
  2192. */
  2193. .globl cuEventDestroy_v2
  2194. .p2align 4
  2195. .type cuEventDestroy_v2, %function
  2196. cuEventDestroy_v2:
  2197. .cfi_startproc
  2198. // Intel opt. manual says to
  2199. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2200. // to hint static predictor.
  2201. cmpq $0, _libcuda_so_tramp_table+600(%rip)
  2202. je 2f
  2203. 1:
  2204. jmp *_libcuda_so_tramp_table+600(%rip)
  2205. 2:
  2206. pushq $75
  2207. .cfi_adjust_cfa_offset 8
  2208. call _libcuda_so_save_regs_and_resolve
  2209. addq $8, %rsp
  2210. .cfi_adjust_cfa_offset -8
  2211. jmp 1b
  2212. .cfi_endproc
  2213. /*
  2214. * Copyright 2018-2019 Yury Gribov
  2215. *
  2216. * The MIT License (MIT)
  2217. *
  2218. * Use of this source code is governed by MIT license that can be
  2219. * found in the LICENSE.txt file.
  2220. */
  2221. .globl cuEventElapsedTime
  2222. .p2align 4
  2223. .type cuEventElapsedTime, %function
  2224. cuEventElapsedTime:
  2225. .cfi_startproc
  2226. // Intel opt. manual says to
  2227. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2228. // to hint static predictor.
  2229. cmpq $0, _libcuda_so_tramp_table+608(%rip)
  2230. je 2f
  2231. 1:
  2232. jmp *_libcuda_so_tramp_table+608(%rip)
  2233. 2:
  2234. pushq $76
  2235. .cfi_adjust_cfa_offset 8
  2236. call _libcuda_so_save_regs_and_resolve
  2237. addq $8, %rsp
  2238. .cfi_adjust_cfa_offset -8
  2239. jmp 1b
  2240. .cfi_endproc
  2241. /*
  2242. * Copyright 2018-2019 Yury Gribov
  2243. *
  2244. * The MIT License (MIT)
  2245. *
  2246. * Use of this source code is governed by MIT license that can be
  2247. * found in the LICENSE.txt file.
  2248. */
  2249. .globl cuEventQuery
  2250. .p2align 4
  2251. .type cuEventQuery, %function
  2252. cuEventQuery:
  2253. .cfi_startproc
  2254. // Intel opt. manual says to
  2255. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2256. // to hint static predictor.
  2257. cmpq $0, _libcuda_so_tramp_table+616(%rip)
  2258. je 2f
  2259. 1:
  2260. jmp *_libcuda_so_tramp_table+616(%rip)
  2261. 2:
  2262. pushq $77
  2263. .cfi_adjust_cfa_offset 8
  2264. call _libcuda_so_save_regs_and_resolve
  2265. addq $8, %rsp
  2266. .cfi_adjust_cfa_offset -8
  2267. jmp 1b
  2268. .cfi_endproc
  2269. /*
  2270. * Copyright 2018-2019 Yury Gribov
  2271. *
  2272. * The MIT License (MIT)
  2273. *
  2274. * Use of this source code is governed by MIT license that can be
  2275. * found in the LICENSE.txt file.
  2276. */
  2277. .globl cuEventRecord
  2278. .p2align 4
  2279. .type cuEventRecord, %function
  2280. cuEventRecord:
  2281. .cfi_startproc
  2282. // Intel opt. manual says to
  2283. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2284. // to hint static predictor.
  2285. cmpq $0, _libcuda_so_tramp_table+624(%rip)
  2286. je 2f
  2287. 1:
  2288. jmp *_libcuda_so_tramp_table+624(%rip)
  2289. 2:
  2290. pushq $78
  2291. .cfi_adjust_cfa_offset 8
  2292. call _libcuda_so_save_regs_and_resolve
  2293. addq $8, %rsp
  2294. .cfi_adjust_cfa_offset -8
  2295. jmp 1b
  2296. .cfi_endproc
  2297. /*
  2298. * Copyright 2018-2019 Yury Gribov
  2299. *
  2300. * The MIT License (MIT)
  2301. *
  2302. * Use of this source code is governed by MIT license that can be
  2303. * found in the LICENSE.txt file.
  2304. */
  2305. .globl cuEventRecordWithFlags
  2306. .p2align 4
  2307. .type cuEventRecordWithFlags, %function
  2308. cuEventRecordWithFlags:
  2309. .cfi_startproc
  2310. // Intel opt. manual says to
  2311. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2312. // to hint static predictor.
  2313. cmpq $0, _libcuda_so_tramp_table+632(%rip)
  2314. je 2f
  2315. 1:
  2316. jmp *_libcuda_so_tramp_table+632(%rip)
  2317. 2:
  2318. pushq $79
  2319. .cfi_adjust_cfa_offset 8
  2320. call _libcuda_so_save_regs_and_resolve
  2321. addq $8, %rsp
  2322. .cfi_adjust_cfa_offset -8
  2323. jmp 1b
  2324. .cfi_endproc
  2325. /*
  2326. * Copyright 2018-2019 Yury Gribov
  2327. *
  2328. * The MIT License (MIT)
  2329. *
  2330. * Use of this source code is governed by MIT license that can be
  2331. * found in the LICENSE.txt file.
  2332. */
  2333. .globl cuEventRecordWithFlags_ptsz
  2334. .p2align 4
  2335. .type cuEventRecordWithFlags_ptsz, %function
  2336. cuEventRecordWithFlags_ptsz:
  2337. .cfi_startproc
  2338. // Intel opt. manual says to
  2339. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2340. // to hint static predictor.
  2341. cmpq $0, _libcuda_so_tramp_table+640(%rip)
  2342. je 2f
  2343. 1:
  2344. jmp *_libcuda_so_tramp_table+640(%rip)
  2345. 2:
  2346. pushq $80
  2347. .cfi_adjust_cfa_offset 8
  2348. call _libcuda_so_save_regs_and_resolve
  2349. addq $8, %rsp
  2350. .cfi_adjust_cfa_offset -8
  2351. jmp 1b
  2352. .cfi_endproc
  2353. /*
  2354. * Copyright 2018-2019 Yury Gribov
  2355. *
  2356. * The MIT License (MIT)
  2357. *
  2358. * Use of this source code is governed by MIT license that can be
  2359. * found in the LICENSE.txt file.
  2360. */
  2361. .globl cuEventRecord_ptsz
  2362. .p2align 4
  2363. .type cuEventRecord_ptsz, %function
  2364. cuEventRecord_ptsz:
  2365. .cfi_startproc
  2366. // Intel opt. manual says to
  2367. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2368. // to hint static predictor.
  2369. cmpq $0, _libcuda_so_tramp_table+648(%rip)
  2370. je 2f
  2371. 1:
  2372. jmp *_libcuda_so_tramp_table+648(%rip)
  2373. 2:
  2374. pushq $81
  2375. .cfi_adjust_cfa_offset 8
  2376. call _libcuda_so_save_regs_and_resolve
  2377. addq $8, %rsp
  2378. .cfi_adjust_cfa_offset -8
  2379. jmp 1b
  2380. .cfi_endproc
  2381. /*
  2382. * Copyright 2018-2019 Yury Gribov
  2383. *
  2384. * The MIT License (MIT)
  2385. *
  2386. * Use of this source code is governed by MIT license that can be
  2387. * found in the LICENSE.txt file.
  2388. */
  2389. .globl cuEventSynchronize
  2390. .p2align 4
  2391. .type cuEventSynchronize, %function
  2392. cuEventSynchronize:
  2393. .cfi_startproc
  2394. // Intel opt. manual says to
  2395. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2396. // to hint static predictor.
  2397. cmpq $0, _libcuda_so_tramp_table+656(%rip)
  2398. je 2f
  2399. 1:
  2400. jmp *_libcuda_so_tramp_table+656(%rip)
  2401. 2:
  2402. pushq $82
  2403. .cfi_adjust_cfa_offset 8
  2404. call _libcuda_so_save_regs_and_resolve
  2405. addq $8, %rsp
  2406. .cfi_adjust_cfa_offset -8
  2407. jmp 1b
  2408. .cfi_endproc
  2409. /*
  2410. * Copyright 2018-2019 Yury Gribov
  2411. *
  2412. * The MIT License (MIT)
  2413. *
  2414. * Use of this source code is governed by MIT license that can be
  2415. * found in the LICENSE.txt file.
  2416. */
  2417. .globl cuExternalMemoryGetMappedBuffer
  2418. .p2align 4
  2419. .type cuExternalMemoryGetMappedBuffer, %function
  2420. cuExternalMemoryGetMappedBuffer:
  2421. .cfi_startproc
  2422. // Intel opt. manual says to
  2423. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2424. // to hint static predictor.
  2425. cmpq $0, _libcuda_so_tramp_table+664(%rip)
  2426. je 2f
  2427. 1:
  2428. jmp *_libcuda_so_tramp_table+664(%rip)
  2429. 2:
  2430. pushq $83
  2431. .cfi_adjust_cfa_offset 8
  2432. call _libcuda_so_save_regs_and_resolve
  2433. addq $8, %rsp
  2434. .cfi_adjust_cfa_offset -8
  2435. jmp 1b
  2436. .cfi_endproc
  2437. /*
  2438. * Copyright 2018-2019 Yury Gribov
  2439. *
  2440. * The MIT License (MIT)
  2441. *
  2442. * Use of this source code is governed by MIT license that can be
  2443. * found in the LICENSE.txt file.
  2444. */
  2445. .globl cuExternalMemoryGetMappedMipmappedArray
  2446. .p2align 4
  2447. .type cuExternalMemoryGetMappedMipmappedArray, %function
  2448. cuExternalMemoryGetMappedMipmappedArray:
  2449. .cfi_startproc
  2450. // Intel opt. manual says to
  2451. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2452. // to hint static predictor.
  2453. cmpq $0, _libcuda_so_tramp_table+672(%rip)
  2454. je 2f
  2455. 1:
  2456. jmp *_libcuda_so_tramp_table+672(%rip)
  2457. 2:
  2458. pushq $84
  2459. .cfi_adjust_cfa_offset 8
  2460. call _libcuda_so_save_regs_and_resolve
  2461. addq $8, %rsp
  2462. .cfi_adjust_cfa_offset -8
  2463. jmp 1b
  2464. .cfi_endproc
  2465. /*
  2466. * Copyright 2018-2019 Yury Gribov
  2467. *
  2468. * The MIT License (MIT)
  2469. *
  2470. * Use of this source code is governed by MIT license that can be
  2471. * found in the LICENSE.txt file.
  2472. */
  2473. .globl cuFuncGetAttribute
  2474. .p2align 4
  2475. .type cuFuncGetAttribute, %function
  2476. cuFuncGetAttribute:
  2477. .cfi_startproc
  2478. // Intel opt. manual says to
  2479. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2480. // to hint static predictor.
  2481. cmpq $0, _libcuda_so_tramp_table+680(%rip)
  2482. je 2f
  2483. 1:
  2484. jmp *_libcuda_so_tramp_table+680(%rip)
  2485. 2:
  2486. pushq $85
  2487. .cfi_adjust_cfa_offset 8
  2488. call _libcuda_so_save_regs_and_resolve
  2489. addq $8, %rsp
  2490. .cfi_adjust_cfa_offset -8
  2491. jmp 1b
  2492. .cfi_endproc
  2493. /*
  2494. * Copyright 2018-2019 Yury Gribov
  2495. *
  2496. * The MIT License (MIT)
  2497. *
  2498. * Use of this source code is governed by MIT license that can be
  2499. * found in the LICENSE.txt file.
  2500. */
  2501. .globl cuFuncGetModule
  2502. .p2align 4
  2503. .type cuFuncGetModule, %function
  2504. cuFuncGetModule:
  2505. .cfi_startproc
  2506. // Intel opt. manual says to
  2507. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2508. // to hint static predictor.
  2509. cmpq $0, _libcuda_so_tramp_table+688(%rip)
  2510. je 2f
  2511. 1:
  2512. jmp *_libcuda_so_tramp_table+688(%rip)
  2513. 2:
  2514. pushq $86
  2515. .cfi_adjust_cfa_offset 8
  2516. call _libcuda_so_save_regs_and_resolve
  2517. addq $8, %rsp
  2518. .cfi_adjust_cfa_offset -8
  2519. jmp 1b
  2520. .cfi_endproc
  2521. /*
  2522. * Copyright 2018-2019 Yury Gribov
  2523. *
  2524. * The MIT License (MIT)
  2525. *
  2526. * Use of this source code is governed by MIT license that can be
  2527. * found in the LICENSE.txt file.
  2528. */
  2529. .globl cuFuncSetAttribute
  2530. .p2align 4
  2531. .type cuFuncSetAttribute, %function
  2532. cuFuncSetAttribute:
  2533. .cfi_startproc
  2534. // Intel opt. manual says to
  2535. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2536. // to hint static predictor.
  2537. cmpq $0, _libcuda_so_tramp_table+696(%rip)
  2538. je 2f
  2539. 1:
  2540. jmp *_libcuda_so_tramp_table+696(%rip)
  2541. 2:
  2542. pushq $87
  2543. .cfi_adjust_cfa_offset 8
  2544. call _libcuda_so_save_regs_and_resolve
  2545. addq $8, %rsp
  2546. .cfi_adjust_cfa_offset -8
  2547. jmp 1b
  2548. .cfi_endproc
  2549. /*
  2550. * Copyright 2018-2019 Yury Gribov
  2551. *
  2552. * The MIT License (MIT)
  2553. *
  2554. * Use of this source code is governed by MIT license that can be
  2555. * found in the LICENSE.txt file.
  2556. */
  2557. .globl cuFuncSetBlockShape
  2558. .p2align 4
  2559. .type cuFuncSetBlockShape, %function
  2560. cuFuncSetBlockShape:
  2561. .cfi_startproc
  2562. // Intel opt. manual says to
  2563. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2564. // to hint static predictor.
  2565. cmpq $0, _libcuda_so_tramp_table+704(%rip)
  2566. je 2f
  2567. 1:
  2568. jmp *_libcuda_so_tramp_table+704(%rip)
  2569. 2:
  2570. pushq $88
  2571. .cfi_adjust_cfa_offset 8
  2572. call _libcuda_so_save_regs_and_resolve
  2573. addq $8, %rsp
  2574. .cfi_adjust_cfa_offset -8
  2575. jmp 1b
  2576. .cfi_endproc
  2577. /*
  2578. * Copyright 2018-2019 Yury Gribov
  2579. *
  2580. * The MIT License (MIT)
  2581. *
  2582. * Use of this source code is governed by MIT license that can be
  2583. * found in the LICENSE.txt file.
  2584. */
  2585. .globl cuFuncSetCacheConfig
  2586. .p2align 4
  2587. .type cuFuncSetCacheConfig, %function
  2588. cuFuncSetCacheConfig:
  2589. .cfi_startproc
  2590. // Intel opt. manual says to
  2591. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2592. // to hint static predictor.
  2593. cmpq $0, _libcuda_so_tramp_table+712(%rip)
  2594. je 2f
  2595. 1:
  2596. jmp *_libcuda_so_tramp_table+712(%rip)
  2597. 2:
  2598. pushq $89
  2599. .cfi_adjust_cfa_offset 8
  2600. call _libcuda_so_save_regs_and_resolve
  2601. addq $8, %rsp
  2602. .cfi_adjust_cfa_offset -8
  2603. jmp 1b
  2604. .cfi_endproc
  2605. /*
  2606. * Copyright 2018-2019 Yury Gribov
  2607. *
  2608. * The MIT License (MIT)
  2609. *
  2610. * Use of this source code is governed by MIT license that can be
  2611. * found in the LICENSE.txt file.
  2612. */
  2613. .globl cuFuncSetSharedMemConfig
  2614. .p2align 4
  2615. .type cuFuncSetSharedMemConfig, %function
  2616. cuFuncSetSharedMemConfig:
  2617. .cfi_startproc
  2618. // Intel opt. manual says to
  2619. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2620. // to hint static predictor.
  2621. cmpq $0, _libcuda_so_tramp_table+720(%rip)
  2622. je 2f
  2623. 1:
  2624. jmp *_libcuda_so_tramp_table+720(%rip)
  2625. 2:
  2626. pushq $90
  2627. .cfi_adjust_cfa_offset 8
  2628. call _libcuda_so_save_regs_and_resolve
  2629. addq $8, %rsp
  2630. .cfi_adjust_cfa_offset -8
  2631. jmp 1b
  2632. .cfi_endproc
  2633. /*
  2634. * Copyright 2018-2019 Yury Gribov
  2635. *
  2636. * The MIT License (MIT)
  2637. *
  2638. * Use of this source code is governed by MIT license that can be
  2639. * found in the LICENSE.txt file.
  2640. */
  2641. .globl cuFuncSetSharedSize
  2642. .p2align 4
  2643. .type cuFuncSetSharedSize, %function
  2644. cuFuncSetSharedSize:
  2645. .cfi_startproc
  2646. // Intel opt. manual says to
  2647. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2648. // to hint static predictor.
  2649. cmpq $0, _libcuda_so_tramp_table+728(%rip)
  2650. je 2f
  2651. 1:
  2652. jmp *_libcuda_so_tramp_table+728(%rip)
  2653. 2:
  2654. pushq $91
  2655. .cfi_adjust_cfa_offset 8
  2656. call _libcuda_so_save_regs_and_resolve
  2657. addq $8, %rsp
  2658. .cfi_adjust_cfa_offset -8
  2659. jmp 1b
  2660. .cfi_endproc
  2661. /*
  2662. * Copyright 2018-2019 Yury Gribov
  2663. *
  2664. * The MIT License (MIT)
  2665. *
  2666. * Use of this source code is governed by MIT license that can be
  2667. * found in the LICENSE.txt file.
  2668. */
  2669. .globl cuGLCtxCreate
  2670. .p2align 4
  2671. .type cuGLCtxCreate, %function
  2672. cuGLCtxCreate:
  2673. .cfi_startproc
  2674. // Intel opt. manual says to
  2675. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2676. // to hint static predictor.
  2677. cmpq $0, _libcuda_so_tramp_table+736(%rip)
  2678. je 2f
  2679. 1:
  2680. jmp *_libcuda_so_tramp_table+736(%rip)
  2681. 2:
  2682. pushq $92
  2683. .cfi_adjust_cfa_offset 8
  2684. call _libcuda_so_save_regs_and_resolve
  2685. addq $8, %rsp
  2686. .cfi_adjust_cfa_offset -8
  2687. jmp 1b
  2688. .cfi_endproc
  2689. /*
  2690. * Copyright 2018-2019 Yury Gribov
  2691. *
  2692. * The MIT License (MIT)
  2693. *
  2694. * Use of this source code is governed by MIT license that can be
  2695. * found in the LICENSE.txt file.
  2696. */
  2697. .globl cuGLCtxCreate_v2
  2698. .p2align 4
  2699. .type cuGLCtxCreate_v2, %function
  2700. cuGLCtxCreate_v2:
  2701. .cfi_startproc
  2702. // Intel opt. manual says to
  2703. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2704. // to hint static predictor.
  2705. cmpq $0, _libcuda_so_tramp_table+744(%rip)
  2706. je 2f
  2707. 1:
  2708. jmp *_libcuda_so_tramp_table+744(%rip)
  2709. 2:
  2710. pushq $93
  2711. .cfi_adjust_cfa_offset 8
  2712. call _libcuda_so_save_regs_and_resolve
  2713. addq $8, %rsp
  2714. .cfi_adjust_cfa_offset -8
  2715. jmp 1b
  2716. .cfi_endproc
  2717. /*
  2718. * Copyright 2018-2019 Yury Gribov
  2719. *
  2720. * The MIT License (MIT)
  2721. *
  2722. * Use of this source code is governed by MIT license that can be
  2723. * found in the LICENSE.txt file.
  2724. */
  2725. .globl cuGLGetDevices
  2726. .p2align 4
  2727. .type cuGLGetDevices, %function
  2728. cuGLGetDevices:
  2729. .cfi_startproc
  2730. // Intel opt. manual says to
  2731. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2732. // to hint static predictor.
  2733. cmpq $0, _libcuda_so_tramp_table+752(%rip)
  2734. je 2f
  2735. 1:
  2736. jmp *_libcuda_so_tramp_table+752(%rip)
  2737. 2:
  2738. pushq $94
  2739. .cfi_adjust_cfa_offset 8
  2740. call _libcuda_so_save_regs_and_resolve
  2741. addq $8, %rsp
  2742. .cfi_adjust_cfa_offset -8
  2743. jmp 1b
  2744. .cfi_endproc
  2745. /*
  2746. * Copyright 2018-2019 Yury Gribov
  2747. *
  2748. * The MIT License (MIT)
  2749. *
  2750. * Use of this source code is governed by MIT license that can be
  2751. * found in the LICENSE.txt file.
  2752. */
  2753. .globl cuGLGetDevices_v2
  2754. .p2align 4
  2755. .type cuGLGetDevices_v2, %function
  2756. cuGLGetDevices_v2:
  2757. .cfi_startproc
  2758. // Intel opt. manual says to
  2759. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2760. // to hint static predictor.
  2761. cmpq $0, _libcuda_so_tramp_table+760(%rip)
  2762. je 2f
  2763. 1:
  2764. jmp *_libcuda_so_tramp_table+760(%rip)
  2765. 2:
  2766. pushq $95
  2767. .cfi_adjust_cfa_offset 8
  2768. call _libcuda_so_save_regs_and_resolve
  2769. addq $8, %rsp
  2770. .cfi_adjust_cfa_offset -8
  2771. jmp 1b
  2772. .cfi_endproc
  2773. /*
  2774. * Copyright 2018-2019 Yury Gribov
  2775. *
  2776. * The MIT License (MIT)
  2777. *
  2778. * Use of this source code is governed by MIT license that can be
  2779. * found in the LICENSE.txt file.
  2780. */
  2781. .globl cuGLInit
  2782. .p2align 4
  2783. .type cuGLInit, %function
  2784. cuGLInit:
  2785. .cfi_startproc
  2786. // Intel opt. manual says to
  2787. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2788. // to hint static predictor.
  2789. cmpq $0, _libcuda_so_tramp_table+768(%rip)
  2790. je 2f
  2791. 1:
  2792. jmp *_libcuda_so_tramp_table+768(%rip)
  2793. 2:
  2794. pushq $96
  2795. .cfi_adjust_cfa_offset 8
  2796. call _libcuda_so_save_regs_and_resolve
  2797. addq $8, %rsp
  2798. .cfi_adjust_cfa_offset -8
  2799. jmp 1b
  2800. .cfi_endproc
  2801. /*
  2802. * Copyright 2018-2019 Yury Gribov
  2803. *
  2804. * The MIT License (MIT)
  2805. *
  2806. * Use of this source code is governed by MIT license that can be
  2807. * found in the LICENSE.txt file.
  2808. */
  2809. .globl cuGLMapBufferObject
  2810. .p2align 4
  2811. .type cuGLMapBufferObject, %function
  2812. cuGLMapBufferObject:
  2813. .cfi_startproc
  2814. // Intel opt. manual says to
  2815. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2816. // to hint static predictor.
  2817. cmpq $0, _libcuda_so_tramp_table+776(%rip)
  2818. je 2f
  2819. 1:
  2820. jmp *_libcuda_so_tramp_table+776(%rip)
  2821. 2:
  2822. pushq $97
  2823. .cfi_adjust_cfa_offset 8
  2824. call _libcuda_so_save_regs_and_resolve
  2825. addq $8, %rsp
  2826. .cfi_adjust_cfa_offset -8
  2827. jmp 1b
  2828. .cfi_endproc
  2829. /*
  2830. * Copyright 2018-2019 Yury Gribov
  2831. *
  2832. * The MIT License (MIT)
  2833. *
  2834. * Use of this source code is governed by MIT license that can be
  2835. * found in the LICENSE.txt file.
  2836. */
  2837. .globl cuGLMapBufferObjectAsync
  2838. .p2align 4
  2839. .type cuGLMapBufferObjectAsync, %function
  2840. cuGLMapBufferObjectAsync:
  2841. .cfi_startproc
  2842. // Intel opt. manual says to
  2843. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2844. // to hint static predictor.
  2845. cmpq $0, _libcuda_so_tramp_table+784(%rip)
  2846. je 2f
  2847. 1:
  2848. jmp *_libcuda_so_tramp_table+784(%rip)
  2849. 2:
  2850. pushq $98
  2851. .cfi_adjust_cfa_offset 8
  2852. call _libcuda_so_save_regs_and_resolve
  2853. addq $8, %rsp
  2854. .cfi_adjust_cfa_offset -8
  2855. jmp 1b
  2856. .cfi_endproc
  2857. /*
  2858. * Copyright 2018-2019 Yury Gribov
  2859. *
  2860. * The MIT License (MIT)
  2861. *
  2862. * Use of this source code is governed by MIT license that can be
  2863. * found in the LICENSE.txt file.
  2864. */
  2865. .globl cuGLMapBufferObjectAsync_v2
  2866. .p2align 4
  2867. .type cuGLMapBufferObjectAsync_v2, %function
  2868. cuGLMapBufferObjectAsync_v2:
  2869. .cfi_startproc
  2870. // Intel opt. manual says to
  2871. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2872. // to hint static predictor.
  2873. cmpq $0, _libcuda_so_tramp_table+792(%rip)
  2874. je 2f
  2875. 1:
  2876. jmp *_libcuda_so_tramp_table+792(%rip)
  2877. 2:
  2878. pushq $99
  2879. .cfi_adjust_cfa_offset 8
  2880. call _libcuda_so_save_regs_and_resolve
  2881. addq $8, %rsp
  2882. .cfi_adjust_cfa_offset -8
  2883. jmp 1b
  2884. .cfi_endproc
  2885. /*
  2886. * Copyright 2018-2019 Yury Gribov
  2887. *
  2888. * The MIT License (MIT)
  2889. *
  2890. * Use of this source code is governed by MIT license that can be
  2891. * found in the LICENSE.txt file.
  2892. */
  2893. .globl cuGLMapBufferObjectAsync_v2_ptsz
  2894. .p2align 4
  2895. .type cuGLMapBufferObjectAsync_v2_ptsz, %function
  2896. cuGLMapBufferObjectAsync_v2_ptsz:
  2897. .cfi_startproc
  2898. // Intel opt. manual says to
  2899. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2900. // to hint static predictor.
  2901. cmpq $0, _libcuda_so_tramp_table+800(%rip)
  2902. je 2f
  2903. 1:
  2904. jmp *_libcuda_so_tramp_table+800(%rip)
  2905. 2:
  2906. pushq $100
  2907. .cfi_adjust_cfa_offset 8
  2908. call _libcuda_so_save_regs_and_resolve
  2909. addq $8, %rsp
  2910. .cfi_adjust_cfa_offset -8
  2911. jmp 1b
  2912. .cfi_endproc
  2913. /*
  2914. * Copyright 2018-2019 Yury Gribov
  2915. *
  2916. * The MIT License (MIT)
  2917. *
  2918. * Use of this source code is governed by MIT license that can be
  2919. * found in the LICENSE.txt file.
  2920. */
  2921. .globl cuGLMapBufferObject_v2
  2922. .p2align 4
  2923. .type cuGLMapBufferObject_v2, %function
  2924. cuGLMapBufferObject_v2:
  2925. .cfi_startproc
  2926. // Intel opt. manual says to
  2927. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2928. // to hint static predictor.
  2929. cmpq $0, _libcuda_so_tramp_table+808(%rip)
  2930. je 2f
  2931. 1:
  2932. jmp *_libcuda_so_tramp_table+808(%rip)
  2933. 2:
  2934. pushq $101
  2935. .cfi_adjust_cfa_offset 8
  2936. call _libcuda_so_save_regs_and_resolve
  2937. addq $8, %rsp
  2938. .cfi_adjust_cfa_offset -8
  2939. jmp 1b
  2940. .cfi_endproc
  2941. /*
  2942. * Copyright 2018-2019 Yury Gribov
  2943. *
  2944. * The MIT License (MIT)
  2945. *
  2946. * Use of this source code is governed by MIT license that can be
  2947. * found in the LICENSE.txt file.
  2948. */
  2949. .globl cuGLMapBufferObject_v2_ptds
  2950. .p2align 4
  2951. .type cuGLMapBufferObject_v2_ptds, %function
  2952. cuGLMapBufferObject_v2_ptds:
  2953. .cfi_startproc
  2954. // Intel opt. manual says to
  2955. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2956. // to hint static predictor.
  2957. cmpq $0, _libcuda_so_tramp_table+816(%rip)
  2958. je 2f
  2959. 1:
  2960. jmp *_libcuda_so_tramp_table+816(%rip)
  2961. 2:
  2962. pushq $102
  2963. .cfi_adjust_cfa_offset 8
  2964. call _libcuda_so_save_regs_and_resolve
  2965. addq $8, %rsp
  2966. .cfi_adjust_cfa_offset -8
  2967. jmp 1b
  2968. .cfi_endproc
  2969. /*
  2970. * Copyright 2018-2019 Yury Gribov
  2971. *
  2972. * The MIT License (MIT)
  2973. *
  2974. * Use of this source code is governed by MIT license that can be
  2975. * found in the LICENSE.txt file.
  2976. */
  2977. .globl cuGLRegisterBufferObject
  2978. .p2align 4
  2979. .type cuGLRegisterBufferObject, %function
  2980. cuGLRegisterBufferObject:
  2981. .cfi_startproc
  2982. // Intel opt. manual says to
  2983. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  2984. // to hint static predictor.
  2985. cmpq $0, _libcuda_so_tramp_table+824(%rip)
  2986. je 2f
  2987. 1:
  2988. jmp *_libcuda_so_tramp_table+824(%rip)
  2989. 2:
  2990. pushq $103
  2991. .cfi_adjust_cfa_offset 8
  2992. call _libcuda_so_save_regs_and_resolve
  2993. addq $8, %rsp
  2994. .cfi_adjust_cfa_offset -8
  2995. jmp 1b
  2996. .cfi_endproc
  2997. /*
  2998. * Copyright 2018-2019 Yury Gribov
  2999. *
  3000. * The MIT License (MIT)
  3001. *
  3002. * Use of this source code is governed by MIT license that can be
  3003. * found in the LICENSE.txt file.
  3004. */
  3005. .globl cuGLSetBufferObjectMapFlags
  3006. .p2align 4
  3007. .type cuGLSetBufferObjectMapFlags, %function
  3008. cuGLSetBufferObjectMapFlags:
  3009. .cfi_startproc
  3010. // Intel opt. manual says to
  3011. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3012. // to hint static predictor.
  3013. cmpq $0, _libcuda_so_tramp_table+832(%rip)
  3014. je 2f
  3015. 1:
  3016. jmp *_libcuda_so_tramp_table+832(%rip)
  3017. 2:
  3018. pushq $104
  3019. .cfi_adjust_cfa_offset 8
  3020. call _libcuda_so_save_regs_and_resolve
  3021. addq $8, %rsp
  3022. .cfi_adjust_cfa_offset -8
  3023. jmp 1b
  3024. .cfi_endproc
  3025. /*
  3026. * Copyright 2018-2019 Yury Gribov
  3027. *
  3028. * The MIT License (MIT)
  3029. *
  3030. * Use of this source code is governed by MIT license that can be
  3031. * found in the LICENSE.txt file.
  3032. */
  3033. .globl cuGLUnmapBufferObject
  3034. .p2align 4
  3035. .type cuGLUnmapBufferObject, %function
  3036. cuGLUnmapBufferObject:
  3037. .cfi_startproc
  3038. // Intel opt. manual says to
  3039. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3040. // to hint static predictor.
  3041. cmpq $0, _libcuda_so_tramp_table+840(%rip)
  3042. je 2f
  3043. 1:
  3044. jmp *_libcuda_so_tramp_table+840(%rip)
  3045. 2:
  3046. pushq $105
  3047. .cfi_adjust_cfa_offset 8
  3048. call _libcuda_so_save_regs_and_resolve
  3049. addq $8, %rsp
  3050. .cfi_adjust_cfa_offset -8
  3051. jmp 1b
  3052. .cfi_endproc
  3053. /*
  3054. * Copyright 2018-2019 Yury Gribov
  3055. *
  3056. * The MIT License (MIT)
  3057. *
  3058. * Use of this source code is governed by MIT license that can be
  3059. * found in the LICENSE.txt file.
  3060. */
  3061. .globl cuGLUnmapBufferObjectAsync
  3062. .p2align 4
  3063. .type cuGLUnmapBufferObjectAsync, %function
  3064. cuGLUnmapBufferObjectAsync:
  3065. .cfi_startproc
  3066. // Intel opt. manual says to
  3067. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3068. // to hint static predictor.
  3069. cmpq $0, _libcuda_so_tramp_table+848(%rip)
  3070. je 2f
  3071. 1:
  3072. jmp *_libcuda_so_tramp_table+848(%rip)
  3073. 2:
  3074. pushq $106
  3075. .cfi_adjust_cfa_offset 8
  3076. call _libcuda_so_save_regs_and_resolve
  3077. addq $8, %rsp
  3078. .cfi_adjust_cfa_offset -8
  3079. jmp 1b
  3080. .cfi_endproc
  3081. /*
  3082. * Copyright 2018-2019 Yury Gribov
  3083. *
  3084. * The MIT License (MIT)
  3085. *
  3086. * Use of this source code is governed by MIT license that can be
  3087. * found in the LICENSE.txt file.
  3088. */
  3089. .globl cuGLUnregisterBufferObject
  3090. .p2align 4
  3091. .type cuGLUnregisterBufferObject, %function
  3092. cuGLUnregisterBufferObject:
  3093. .cfi_startproc
  3094. // Intel opt. manual says to
  3095. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3096. // to hint static predictor.
  3097. cmpq $0, _libcuda_so_tramp_table+856(%rip)
  3098. je 2f
  3099. 1:
  3100. jmp *_libcuda_so_tramp_table+856(%rip)
  3101. 2:
  3102. pushq $107
  3103. .cfi_adjust_cfa_offset 8
  3104. call _libcuda_so_save_regs_and_resolve
  3105. addq $8, %rsp
  3106. .cfi_adjust_cfa_offset -8
  3107. jmp 1b
  3108. .cfi_endproc
  3109. /*
  3110. * Copyright 2018-2019 Yury Gribov
  3111. *
  3112. * The MIT License (MIT)
  3113. *
  3114. * Use of this source code is governed by MIT license that can be
  3115. * found in the LICENSE.txt file.
  3116. */
  3117. .globl cuGetErrorName
  3118. .p2align 4
  3119. .type cuGetErrorName, %function
  3120. cuGetErrorName:
  3121. .cfi_startproc
  3122. // Intel opt. manual says to
  3123. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3124. // to hint static predictor.
  3125. cmpq $0, _libcuda_so_tramp_table+864(%rip)
  3126. je 2f
  3127. 1:
  3128. jmp *_libcuda_so_tramp_table+864(%rip)
  3129. 2:
  3130. pushq $108
  3131. .cfi_adjust_cfa_offset 8
  3132. call _libcuda_so_save_regs_and_resolve
  3133. addq $8, %rsp
  3134. .cfi_adjust_cfa_offset -8
  3135. jmp 1b
  3136. .cfi_endproc
  3137. /*
  3138. * Copyright 2018-2019 Yury Gribov
  3139. *
  3140. * The MIT License (MIT)
  3141. *
  3142. * Use of this source code is governed by MIT license that can be
  3143. * found in the LICENSE.txt file.
  3144. */
  3145. .globl cuGetErrorString
  3146. .p2align 4
  3147. .type cuGetErrorString, %function
  3148. cuGetErrorString:
  3149. .cfi_startproc
  3150. // Intel opt. manual says to
  3151. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3152. // to hint static predictor.
  3153. cmpq $0, _libcuda_so_tramp_table+872(%rip)
  3154. je 2f
  3155. 1:
  3156. jmp *_libcuda_so_tramp_table+872(%rip)
  3157. 2:
  3158. pushq $109
  3159. .cfi_adjust_cfa_offset 8
  3160. call _libcuda_so_save_regs_and_resolve
  3161. addq $8, %rsp
  3162. .cfi_adjust_cfa_offset -8
  3163. jmp 1b
  3164. .cfi_endproc
  3165. /*
  3166. * Copyright 2018-2019 Yury Gribov
  3167. *
  3168. * The MIT License (MIT)
  3169. *
  3170. * Use of this source code is governed by MIT license that can be
  3171. * found in the LICENSE.txt file.
  3172. */
  3173. .globl cuGetExportTable
  3174. .p2align 4
  3175. .type cuGetExportTable, %function
  3176. cuGetExportTable:
  3177. .cfi_startproc
  3178. // Intel opt. manual says to
  3179. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3180. // to hint static predictor.
  3181. cmpq $0, _libcuda_so_tramp_table+880(%rip)
  3182. je 2f
  3183. 1:
  3184. jmp *_libcuda_so_tramp_table+880(%rip)
  3185. 2:
  3186. pushq $110
  3187. .cfi_adjust_cfa_offset 8
  3188. call _libcuda_so_save_regs_and_resolve
  3189. addq $8, %rsp
  3190. .cfi_adjust_cfa_offset -8
  3191. jmp 1b
  3192. .cfi_endproc
  3193. /*
  3194. * Copyright 2018-2019 Yury Gribov
  3195. *
  3196. * The MIT License (MIT)
  3197. *
  3198. * Use of this source code is governed by MIT license that can be
  3199. * found in the LICENSE.txt file.
  3200. */
  3201. .globl cuGraphAddChildGraphNode
  3202. .p2align 4
  3203. .type cuGraphAddChildGraphNode, %function
  3204. cuGraphAddChildGraphNode:
  3205. .cfi_startproc
  3206. // Intel opt. manual says to
  3207. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3208. // to hint static predictor.
  3209. cmpq $0, _libcuda_so_tramp_table+888(%rip)
  3210. je 2f
  3211. 1:
  3212. jmp *_libcuda_so_tramp_table+888(%rip)
  3213. 2:
  3214. pushq $111
  3215. .cfi_adjust_cfa_offset 8
  3216. call _libcuda_so_save_regs_and_resolve
  3217. addq $8, %rsp
  3218. .cfi_adjust_cfa_offset -8
  3219. jmp 1b
  3220. .cfi_endproc
  3221. /*
  3222. * Copyright 2018-2019 Yury Gribov
  3223. *
  3224. * The MIT License (MIT)
  3225. *
  3226. * Use of this source code is governed by MIT license that can be
  3227. * found in the LICENSE.txt file.
  3228. */
  3229. .globl cuGraphAddDependencies
  3230. .p2align 4
  3231. .type cuGraphAddDependencies, %function
  3232. cuGraphAddDependencies:
  3233. .cfi_startproc
  3234. // Intel opt. manual says to
  3235. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3236. // to hint static predictor.
  3237. cmpq $0, _libcuda_so_tramp_table+896(%rip)
  3238. je 2f
  3239. 1:
  3240. jmp *_libcuda_so_tramp_table+896(%rip)
  3241. 2:
  3242. pushq $112
  3243. .cfi_adjust_cfa_offset 8
  3244. call _libcuda_so_save_regs_and_resolve
  3245. addq $8, %rsp
  3246. .cfi_adjust_cfa_offset -8
  3247. jmp 1b
  3248. .cfi_endproc
  3249. /*
  3250. * Copyright 2018-2019 Yury Gribov
  3251. *
  3252. * The MIT License (MIT)
  3253. *
  3254. * Use of this source code is governed by MIT license that can be
  3255. * found in the LICENSE.txt file.
  3256. */
  3257. .globl cuGraphAddEmptyNode
  3258. .p2align 4
  3259. .type cuGraphAddEmptyNode, %function
  3260. cuGraphAddEmptyNode:
  3261. .cfi_startproc
  3262. // Intel opt. manual says to
  3263. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3264. // to hint static predictor.
  3265. cmpq $0, _libcuda_so_tramp_table+904(%rip)
  3266. je 2f
  3267. 1:
  3268. jmp *_libcuda_so_tramp_table+904(%rip)
  3269. 2:
  3270. pushq $113
  3271. .cfi_adjust_cfa_offset 8
  3272. call _libcuda_so_save_regs_and_resolve
  3273. addq $8, %rsp
  3274. .cfi_adjust_cfa_offset -8
  3275. jmp 1b
  3276. .cfi_endproc
  3277. /*
  3278. * Copyright 2018-2019 Yury Gribov
  3279. *
  3280. * The MIT License (MIT)
  3281. *
  3282. * Use of this source code is governed by MIT license that can be
  3283. * found in the LICENSE.txt file.
  3284. */
  3285. .globl cuGraphAddEventRecordNode
  3286. .p2align 4
  3287. .type cuGraphAddEventRecordNode, %function
  3288. cuGraphAddEventRecordNode:
  3289. .cfi_startproc
  3290. // Intel opt. manual says to
  3291. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3292. // to hint static predictor.
  3293. cmpq $0, _libcuda_so_tramp_table+912(%rip)
  3294. je 2f
  3295. 1:
  3296. jmp *_libcuda_so_tramp_table+912(%rip)
  3297. 2:
  3298. pushq $114
  3299. .cfi_adjust_cfa_offset 8
  3300. call _libcuda_so_save_regs_and_resolve
  3301. addq $8, %rsp
  3302. .cfi_adjust_cfa_offset -8
  3303. jmp 1b
  3304. .cfi_endproc
  3305. /*
  3306. * Copyright 2018-2019 Yury Gribov
  3307. *
  3308. * The MIT License (MIT)
  3309. *
  3310. * Use of this source code is governed by MIT license that can be
  3311. * found in the LICENSE.txt file.
  3312. */
  3313. .globl cuGraphAddEventWaitNode
  3314. .p2align 4
  3315. .type cuGraphAddEventWaitNode, %function
  3316. cuGraphAddEventWaitNode:
  3317. .cfi_startproc
  3318. // Intel opt. manual says to
  3319. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3320. // to hint static predictor.
  3321. cmpq $0, _libcuda_so_tramp_table+920(%rip)
  3322. je 2f
  3323. 1:
  3324. jmp *_libcuda_so_tramp_table+920(%rip)
  3325. 2:
  3326. pushq $115
  3327. .cfi_adjust_cfa_offset 8
  3328. call _libcuda_so_save_regs_and_resolve
  3329. addq $8, %rsp
  3330. .cfi_adjust_cfa_offset -8
  3331. jmp 1b
  3332. .cfi_endproc
  3333. /*
  3334. * Copyright 2018-2019 Yury Gribov
  3335. *
  3336. * The MIT License (MIT)
  3337. *
  3338. * Use of this source code is governed by MIT license that can be
  3339. * found in the LICENSE.txt file.
  3340. */
  3341. .globl cuGraphAddHostNode
  3342. .p2align 4
  3343. .type cuGraphAddHostNode, %function
  3344. cuGraphAddHostNode:
  3345. .cfi_startproc
  3346. // Intel opt. manual says to
  3347. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3348. // to hint static predictor.
  3349. cmpq $0, _libcuda_so_tramp_table+928(%rip)
  3350. je 2f
  3351. 1:
  3352. jmp *_libcuda_so_tramp_table+928(%rip)
  3353. 2:
  3354. pushq $116
  3355. .cfi_adjust_cfa_offset 8
  3356. call _libcuda_so_save_regs_and_resolve
  3357. addq $8, %rsp
  3358. .cfi_adjust_cfa_offset -8
  3359. jmp 1b
  3360. .cfi_endproc
  3361. /*
  3362. * Copyright 2018-2019 Yury Gribov
  3363. *
  3364. * The MIT License (MIT)
  3365. *
  3366. * Use of this source code is governed by MIT license that can be
  3367. * found in the LICENSE.txt file.
  3368. */
  3369. .globl cuGraphAddKernelNode
  3370. .p2align 4
  3371. .type cuGraphAddKernelNode, %function
  3372. cuGraphAddKernelNode:
  3373. .cfi_startproc
  3374. // Intel opt. manual says to
  3375. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3376. // to hint static predictor.
  3377. cmpq $0, _libcuda_so_tramp_table+936(%rip)
  3378. je 2f
  3379. 1:
  3380. jmp *_libcuda_so_tramp_table+936(%rip)
  3381. 2:
  3382. pushq $117
  3383. .cfi_adjust_cfa_offset 8
  3384. call _libcuda_so_save_regs_and_resolve
  3385. addq $8, %rsp
  3386. .cfi_adjust_cfa_offset -8
  3387. jmp 1b
  3388. .cfi_endproc
  3389. /*
  3390. * Copyright 2018-2019 Yury Gribov
  3391. *
  3392. * The MIT License (MIT)
  3393. *
  3394. * Use of this source code is governed by MIT license that can be
  3395. * found in the LICENSE.txt file.
  3396. */
  3397. .globl cuGraphAddMemcpyNode
  3398. .p2align 4
  3399. .type cuGraphAddMemcpyNode, %function
  3400. cuGraphAddMemcpyNode:
  3401. .cfi_startproc
  3402. // Intel opt. manual says to
  3403. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3404. // to hint static predictor.
  3405. cmpq $0, _libcuda_so_tramp_table+944(%rip)
  3406. je 2f
  3407. 1:
  3408. jmp *_libcuda_so_tramp_table+944(%rip)
  3409. 2:
  3410. pushq $118
  3411. .cfi_adjust_cfa_offset 8
  3412. call _libcuda_so_save_regs_and_resolve
  3413. addq $8, %rsp
  3414. .cfi_adjust_cfa_offset -8
  3415. jmp 1b
  3416. .cfi_endproc
  3417. /*
  3418. * Copyright 2018-2019 Yury Gribov
  3419. *
  3420. * The MIT License (MIT)
  3421. *
  3422. * Use of this source code is governed by MIT license that can be
  3423. * found in the LICENSE.txt file.
  3424. */
  3425. .globl cuGraphAddMemsetNode
  3426. .p2align 4
  3427. .type cuGraphAddMemsetNode, %function
  3428. cuGraphAddMemsetNode:
  3429. .cfi_startproc
  3430. // Intel opt. manual says to
  3431. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3432. // to hint static predictor.
  3433. cmpq $0, _libcuda_so_tramp_table+952(%rip)
  3434. je 2f
  3435. 1:
  3436. jmp *_libcuda_so_tramp_table+952(%rip)
  3437. 2:
  3438. pushq $119
  3439. .cfi_adjust_cfa_offset 8
  3440. call _libcuda_so_save_regs_and_resolve
  3441. addq $8, %rsp
  3442. .cfi_adjust_cfa_offset -8
  3443. jmp 1b
  3444. .cfi_endproc
  3445. /*
  3446. * Copyright 2018-2019 Yury Gribov
  3447. *
  3448. * The MIT License (MIT)
  3449. *
  3450. * Use of this source code is governed by MIT license that can be
  3451. * found in the LICENSE.txt file.
  3452. */
  3453. .globl cuGraphChildGraphNodeGetGraph
  3454. .p2align 4
  3455. .type cuGraphChildGraphNodeGetGraph, %function
  3456. cuGraphChildGraphNodeGetGraph:
  3457. .cfi_startproc
  3458. // Intel opt. manual says to
  3459. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3460. // to hint static predictor.
  3461. cmpq $0, _libcuda_so_tramp_table+960(%rip)
  3462. je 2f
  3463. 1:
  3464. jmp *_libcuda_so_tramp_table+960(%rip)
  3465. 2:
  3466. pushq $120
  3467. .cfi_adjust_cfa_offset 8
  3468. call _libcuda_so_save_regs_and_resolve
  3469. addq $8, %rsp
  3470. .cfi_adjust_cfa_offset -8
  3471. jmp 1b
  3472. .cfi_endproc
  3473. /*
  3474. * Copyright 2018-2019 Yury Gribov
  3475. *
  3476. * The MIT License (MIT)
  3477. *
  3478. * Use of this source code is governed by MIT license that can be
  3479. * found in the LICENSE.txt file.
  3480. */
  3481. .globl cuGraphClone
  3482. .p2align 4
  3483. .type cuGraphClone, %function
  3484. cuGraphClone:
  3485. .cfi_startproc
  3486. // Intel opt. manual says to
  3487. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3488. // to hint static predictor.
  3489. cmpq $0, _libcuda_so_tramp_table+968(%rip)
  3490. je 2f
  3491. 1:
  3492. jmp *_libcuda_so_tramp_table+968(%rip)
  3493. 2:
  3494. pushq $121
  3495. .cfi_adjust_cfa_offset 8
  3496. call _libcuda_so_save_regs_and_resolve
  3497. addq $8, %rsp
  3498. .cfi_adjust_cfa_offset -8
  3499. jmp 1b
  3500. .cfi_endproc
  3501. /*
  3502. * Copyright 2018-2019 Yury Gribov
  3503. *
  3504. * The MIT License (MIT)
  3505. *
  3506. * Use of this source code is governed by MIT license that can be
  3507. * found in the LICENSE.txt file.
  3508. */
  3509. .globl cuGraphCreate
  3510. .p2align 4
  3511. .type cuGraphCreate, %function
  3512. cuGraphCreate:
  3513. .cfi_startproc
  3514. // Intel opt. manual says to
  3515. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3516. // to hint static predictor.
  3517. cmpq $0, _libcuda_so_tramp_table+976(%rip)
  3518. je 2f
  3519. 1:
  3520. jmp *_libcuda_so_tramp_table+976(%rip)
  3521. 2:
  3522. pushq $122
  3523. .cfi_adjust_cfa_offset 8
  3524. call _libcuda_so_save_regs_and_resolve
  3525. addq $8, %rsp
  3526. .cfi_adjust_cfa_offset -8
  3527. jmp 1b
  3528. .cfi_endproc
  3529. /*
  3530. * Copyright 2018-2019 Yury Gribov
  3531. *
  3532. * The MIT License (MIT)
  3533. *
  3534. * Use of this source code is governed by MIT license that can be
  3535. * found in the LICENSE.txt file.
  3536. */
  3537. .globl cuGraphDestroy
  3538. .p2align 4
  3539. .type cuGraphDestroy, %function
  3540. cuGraphDestroy:
  3541. .cfi_startproc
  3542. // Intel opt. manual says to
  3543. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3544. // to hint static predictor.
  3545. cmpq $0, _libcuda_so_tramp_table+984(%rip)
  3546. je 2f
  3547. 1:
  3548. jmp *_libcuda_so_tramp_table+984(%rip)
  3549. 2:
  3550. pushq $123
  3551. .cfi_adjust_cfa_offset 8
  3552. call _libcuda_so_save_regs_and_resolve
  3553. addq $8, %rsp
  3554. .cfi_adjust_cfa_offset -8
  3555. jmp 1b
  3556. .cfi_endproc
  3557. /*
  3558. * Copyright 2018-2019 Yury Gribov
  3559. *
  3560. * The MIT License (MIT)
  3561. *
  3562. * Use of this source code is governed by MIT license that can be
  3563. * found in the LICENSE.txt file.
  3564. */
  3565. .globl cuGraphDestroyNode
  3566. .p2align 4
  3567. .type cuGraphDestroyNode, %function
  3568. cuGraphDestroyNode:
  3569. .cfi_startproc
  3570. // Intel opt. manual says to
  3571. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3572. // to hint static predictor.
  3573. cmpq $0, _libcuda_so_tramp_table+992(%rip)
  3574. je 2f
  3575. 1:
  3576. jmp *_libcuda_so_tramp_table+992(%rip)
  3577. 2:
  3578. pushq $124
  3579. .cfi_adjust_cfa_offset 8
  3580. call _libcuda_so_save_regs_and_resolve
  3581. addq $8, %rsp
  3582. .cfi_adjust_cfa_offset -8
  3583. jmp 1b
  3584. .cfi_endproc
  3585. /*
  3586. * Copyright 2018-2019 Yury Gribov
  3587. *
  3588. * The MIT License (MIT)
  3589. *
  3590. * Use of this source code is governed by MIT license that can be
  3591. * found in the LICENSE.txt file.
  3592. */
  3593. .globl cuGraphEventRecordNodeGetEvent
  3594. .p2align 4
  3595. .type cuGraphEventRecordNodeGetEvent, %function
  3596. cuGraphEventRecordNodeGetEvent:
  3597. .cfi_startproc
  3598. // Intel opt. manual says to
  3599. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3600. // to hint static predictor.
  3601. cmpq $0, _libcuda_so_tramp_table+1000(%rip)
  3602. je 2f
  3603. 1:
  3604. jmp *_libcuda_so_tramp_table+1000(%rip)
  3605. 2:
  3606. pushq $125
  3607. .cfi_adjust_cfa_offset 8
  3608. call _libcuda_so_save_regs_and_resolve
  3609. addq $8, %rsp
  3610. .cfi_adjust_cfa_offset -8
  3611. jmp 1b
  3612. .cfi_endproc
  3613. /*
  3614. * Copyright 2018-2019 Yury Gribov
  3615. *
  3616. * The MIT License (MIT)
  3617. *
  3618. * Use of this source code is governed by MIT license that can be
  3619. * found in the LICENSE.txt file.
  3620. */
  3621. .globl cuGraphEventRecordNodeSetEvent
  3622. .p2align 4
  3623. .type cuGraphEventRecordNodeSetEvent, %function
  3624. cuGraphEventRecordNodeSetEvent:
  3625. .cfi_startproc
  3626. // Intel opt. manual says to
  3627. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3628. // to hint static predictor.
  3629. cmpq $0, _libcuda_so_tramp_table+1008(%rip)
  3630. je 2f
  3631. 1:
  3632. jmp *_libcuda_so_tramp_table+1008(%rip)
  3633. 2:
  3634. pushq $126
  3635. .cfi_adjust_cfa_offset 8
  3636. call _libcuda_so_save_regs_and_resolve
  3637. addq $8, %rsp
  3638. .cfi_adjust_cfa_offset -8
  3639. jmp 1b
  3640. .cfi_endproc
  3641. /*
  3642. * Copyright 2018-2019 Yury Gribov
  3643. *
  3644. * The MIT License (MIT)
  3645. *
  3646. * Use of this source code is governed by MIT license that can be
  3647. * found in the LICENSE.txt file.
  3648. */
  3649. .globl cuGraphEventWaitNodeGetEvent
  3650. .p2align 4
  3651. .type cuGraphEventWaitNodeGetEvent, %function
  3652. cuGraphEventWaitNodeGetEvent:
  3653. .cfi_startproc
  3654. // Intel opt. manual says to
  3655. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3656. // to hint static predictor.
  3657. cmpq $0, _libcuda_so_tramp_table+1016(%rip)
  3658. je 2f
  3659. 1:
  3660. jmp *_libcuda_so_tramp_table+1016(%rip)
  3661. 2:
  3662. pushq $127
  3663. .cfi_adjust_cfa_offset 8
  3664. call _libcuda_so_save_regs_and_resolve
  3665. addq $8, %rsp
  3666. .cfi_adjust_cfa_offset -8
  3667. jmp 1b
  3668. .cfi_endproc
  3669. /*
  3670. * Copyright 2018-2019 Yury Gribov
  3671. *
  3672. * The MIT License (MIT)
  3673. *
  3674. * Use of this source code is governed by MIT license that can be
  3675. * found in the LICENSE.txt file.
  3676. */
  3677. .globl cuGraphEventWaitNodeSetEvent
  3678. .p2align 4
  3679. .type cuGraphEventWaitNodeSetEvent, %function
  3680. cuGraphEventWaitNodeSetEvent:
  3681. .cfi_startproc
  3682. // Intel opt. manual says to
  3683. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3684. // to hint static predictor.
  3685. cmpq $0, _libcuda_so_tramp_table+1024(%rip)
  3686. je 2f
  3687. 1:
  3688. jmp *_libcuda_so_tramp_table+1024(%rip)
  3689. 2:
  3690. pushq $128
  3691. .cfi_adjust_cfa_offset 8
  3692. call _libcuda_so_save_regs_and_resolve
  3693. addq $8, %rsp
  3694. .cfi_adjust_cfa_offset -8
  3695. jmp 1b
  3696. .cfi_endproc
  3697. /*
  3698. * Copyright 2018-2019 Yury Gribov
  3699. *
  3700. * The MIT License (MIT)
  3701. *
  3702. * Use of this source code is governed by MIT license that can be
  3703. * found in the LICENSE.txt file.
  3704. */
  3705. .globl cuGraphExecChildGraphNodeSetParams
  3706. .p2align 4
  3707. .type cuGraphExecChildGraphNodeSetParams, %function
  3708. cuGraphExecChildGraphNodeSetParams:
  3709. .cfi_startproc
  3710. // Intel opt. manual says to
  3711. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3712. // to hint static predictor.
  3713. cmpq $0, _libcuda_so_tramp_table+1032(%rip)
  3714. je 2f
  3715. 1:
  3716. jmp *_libcuda_so_tramp_table+1032(%rip)
  3717. 2:
  3718. pushq $129
  3719. .cfi_adjust_cfa_offset 8
  3720. call _libcuda_so_save_regs_and_resolve
  3721. addq $8, %rsp
  3722. .cfi_adjust_cfa_offset -8
  3723. jmp 1b
  3724. .cfi_endproc
  3725. /*
  3726. * Copyright 2018-2019 Yury Gribov
  3727. *
  3728. * The MIT License (MIT)
  3729. *
  3730. * Use of this source code is governed by MIT license that can be
  3731. * found in the LICENSE.txt file.
  3732. */
  3733. .globl cuGraphExecDestroy
  3734. .p2align 4
  3735. .type cuGraphExecDestroy, %function
  3736. cuGraphExecDestroy:
  3737. .cfi_startproc
  3738. // Intel opt. manual says to
  3739. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3740. // to hint static predictor.
  3741. cmpq $0, _libcuda_so_tramp_table+1040(%rip)
  3742. je 2f
  3743. 1:
  3744. jmp *_libcuda_so_tramp_table+1040(%rip)
  3745. 2:
  3746. pushq $130
  3747. .cfi_adjust_cfa_offset 8
  3748. call _libcuda_so_save_regs_and_resolve
  3749. addq $8, %rsp
  3750. .cfi_adjust_cfa_offset -8
  3751. jmp 1b
  3752. .cfi_endproc
  3753. /*
  3754. * Copyright 2018-2019 Yury Gribov
  3755. *
  3756. * The MIT License (MIT)
  3757. *
  3758. * Use of this source code is governed by MIT license that can be
  3759. * found in the LICENSE.txt file.
  3760. */
  3761. .globl cuGraphExecEventRecordNodeSetEvent
  3762. .p2align 4
  3763. .type cuGraphExecEventRecordNodeSetEvent, %function
  3764. cuGraphExecEventRecordNodeSetEvent:
  3765. .cfi_startproc
  3766. // Intel opt. manual says to
  3767. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3768. // to hint static predictor.
  3769. cmpq $0, _libcuda_so_tramp_table+1048(%rip)
  3770. je 2f
  3771. 1:
  3772. jmp *_libcuda_so_tramp_table+1048(%rip)
  3773. 2:
  3774. pushq $131
  3775. .cfi_adjust_cfa_offset 8
  3776. call _libcuda_so_save_regs_and_resolve
  3777. addq $8, %rsp
  3778. .cfi_adjust_cfa_offset -8
  3779. jmp 1b
  3780. .cfi_endproc
  3781. /*
  3782. * Copyright 2018-2019 Yury Gribov
  3783. *
  3784. * The MIT License (MIT)
  3785. *
  3786. * Use of this source code is governed by MIT license that can be
  3787. * found in the LICENSE.txt file.
  3788. */
  3789. .globl cuGraphExecEventWaitNodeSetEvent
  3790. .p2align 4
  3791. .type cuGraphExecEventWaitNodeSetEvent, %function
  3792. cuGraphExecEventWaitNodeSetEvent:
  3793. .cfi_startproc
  3794. // Intel opt. manual says to
  3795. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3796. // to hint static predictor.
  3797. cmpq $0, _libcuda_so_tramp_table+1056(%rip)
  3798. je 2f
  3799. 1:
  3800. jmp *_libcuda_so_tramp_table+1056(%rip)
  3801. 2:
  3802. pushq $132
  3803. .cfi_adjust_cfa_offset 8
  3804. call _libcuda_so_save_regs_and_resolve
  3805. addq $8, %rsp
  3806. .cfi_adjust_cfa_offset -8
  3807. jmp 1b
  3808. .cfi_endproc
  3809. /*
  3810. * Copyright 2018-2019 Yury Gribov
  3811. *
  3812. * The MIT License (MIT)
  3813. *
  3814. * Use of this source code is governed by MIT license that can be
  3815. * found in the LICENSE.txt file.
  3816. */
  3817. .globl cuGraphExecHostNodeSetParams
  3818. .p2align 4
  3819. .type cuGraphExecHostNodeSetParams, %function
  3820. cuGraphExecHostNodeSetParams:
  3821. .cfi_startproc
  3822. // Intel opt. manual says to
  3823. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3824. // to hint static predictor.
  3825. cmpq $0, _libcuda_so_tramp_table+1064(%rip)
  3826. je 2f
  3827. 1:
  3828. jmp *_libcuda_so_tramp_table+1064(%rip)
  3829. 2:
  3830. pushq $133
  3831. .cfi_adjust_cfa_offset 8
  3832. call _libcuda_so_save_regs_and_resolve
  3833. addq $8, %rsp
  3834. .cfi_adjust_cfa_offset -8
  3835. jmp 1b
  3836. .cfi_endproc
  3837. /*
  3838. * Copyright 2018-2019 Yury Gribov
  3839. *
  3840. * The MIT License (MIT)
  3841. *
  3842. * Use of this source code is governed by MIT license that can be
  3843. * found in the LICENSE.txt file.
  3844. */
  3845. .globl cuGraphExecKernelNodeSetParams
  3846. .p2align 4
  3847. .type cuGraphExecKernelNodeSetParams, %function
  3848. cuGraphExecKernelNodeSetParams:
  3849. .cfi_startproc
  3850. // Intel opt. manual says to
  3851. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3852. // to hint static predictor.
  3853. cmpq $0, _libcuda_so_tramp_table+1072(%rip)
  3854. je 2f
  3855. 1:
  3856. jmp *_libcuda_so_tramp_table+1072(%rip)
  3857. 2:
  3858. pushq $134
  3859. .cfi_adjust_cfa_offset 8
  3860. call _libcuda_so_save_regs_and_resolve
  3861. addq $8, %rsp
  3862. .cfi_adjust_cfa_offset -8
  3863. jmp 1b
  3864. .cfi_endproc
  3865. /*
  3866. * Copyright 2018-2019 Yury Gribov
  3867. *
  3868. * The MIT License (MIT)
  3869. *
  3870. * Use of this source code is governed by MIT license that can be
  3871. * found in the LICENSE.txt file.
  3872. */
  3873. .globl cuGraphExecMemcpyNodeSetParams
  3874. .p2align 4
  3875. .type cuGraphExecMemcpyNodeSetParams, %function
  3876. cuGraphExecMemcpyNodeSetParams:
  3877. .cfi_startproc
  3878. // Intel opt. manual says to
  3879. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3880. // to hint static predictor.
  3881. cmpq $0, _libcuda_so_tramp_table+1080(%rip)
  3882. je 2f
  3883. 1:
  3884. jmp *_libcuda_so_tramp_table+1080(%rip)
  3885. 2:
  3886. pushq $135
  3887. .cfi_adjust_cfa_offset 8
  3888. call _libcuda_so_save_regs_and_resolve
  3889. addq $8, %rsp
  3890. .cfi_adjust_cfa_offset -8
  3891. jmp 1b
  3892. .cfi_endproc
  3893. /*
  3894. * Copyright 2018-2019 Yury Gribov
  3895. *
  3896. * The MIT License (MIT)
  3897. *
  3898. * Use of this source code is governed by MIT license that can be
  3899. * found in the LICENSE.txt file.
  3900. */
  3901. .globl cuGraphExecMemsetNodeSetParams
  3902. .p2align 4
  3903. .type cuGraphExecMemsetNodeSetParams, %function
  3904. cuGraphExecMemsetNodeSetParams:
  3905. .cfi_startproc
  3906. // Intel opt. manual says to
  3907. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3908. // to hint static predictor.
  3909. cmpq $0, _libcuda_so_tramp_table+1088(%rip)
  3910. je 2f
  3911. 1:
  3912. jmp *_libcuda_so_tramp_table+1088(%rip)
  3913. 2:
  3914. pushq $136
  3915. .cfi_adjust_cfa_offset 8
  3916. call _libcuda_so_save_regs_and_resolve
  3917. addq $8, %rsp
  3918. .cfi_adjust_cfa_offset -8
  3919. jmp 1b
  3920. .cfi_endproc
  3921. /*
  3922. * Copyright 2018-2019 Yury Gribov
  3923. *
  3924. * The MIT License (MIT)
  3925. *
  3926. * Use of this source code is governed by MIT license that can be
  3927. * found in the LICENSE.txt file.
  3928. */
  3929. .globl cuGraphExecUpdate
  3930. .p2align 4
  3931. .type cuGraphExecUpdate, %function
  3932. cuGraphExecUpdate:
  3933. .cfi_startproc
  3934. // Intel opt. manual says to
  3935. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3936. // to hint static predictor.
  3937. cmpq $0, _libcuda_so_tramp_table+1096(%rip)
  3938. je 2f
  3939. 1:
  3940. jmp *_libcuda_so_tramp_table+1096(%rip)
  3941. 2:
  3942. pushq $137
  3943. .cfi_adjust_cfa_offset 8
  3944. call _libcuda_so_save_regs_and_resolve
  3945. addq $8, %rsp
  3946. .cfi_adjust_cfa_offset -8
  3947. jmp 1b
  3948. .cfi_endproc
  3949. /*
  3950. * Copyright 2018-2019 Yury Gribov
  3951. *
  3952. * The MIT License (MIT)
  3953. *
  3954. * Use of this source code is governed by MIT license that can be
  3955. * found in the LICENSE.txt file.
  3956. */
  3957. .globl cuGraphGetEdges
  3958. .p2align 4
  3959. .type cuGraphGetEdges, %function
  3960. cuGraphGetEdges:
  3961. .cfi_startproc
  3962. // Intel opt. manual says to
  3963. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3964. // to hint static predictor.
  3965. cmpq $0, _libcuda_so_tramp_table+1104(%rip)
  3966. je 2f
  3967. 1:
  3968. jmp *_libcuda_so_tramp_table+1104(%rip)
  3969. 2:
  3970. pushq $138
  3971. .cfi_adjust_cfa_offset 8
  3972. call _libcuda_so_save_regs_and_resolve
  3973. addq $8, %rsp
  3974. .cfi_adjust_cfa_offset -8
  3975. jmp 1b
  3976. .cfi_endproc
  3977. /*
  3978. * Copyright 2018-2019 Yury Gribov
  3979. *
  3980. * The MIT License (MIT)
  3981. *
  3982. * Use of this source code is governed by MIT license that can be
  3983. * found in the LICENSE.txt file.
  3984. */
  3985. .globl cuGraphGetNodes
  3986. .p2align 4
  3987. .type cuGraphGetNodes, %function
  3988. cuGraphGetNodes:
  3989. .cfi_startproc
  3990. // Intel opt. manual says to
  3991. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  3992. // to hint static predictor.
  3993. cmpq $0, _libcuda_so_tramp_table+1112(%rip)
  3994. je 2f
  3995. 1:
  3996. jmp *_libcuda_so_tramp_table+1112(%rip)
  3997. 2:
  3998. pushq $139
  3999. .cfi_adjust_cfa_offset 8
  4000. call _libcuda_so_save_regs_and_resolve
  4001. addq $8, %rsp
  4002. .cfi_adjust_cfa_offset -8
  4003. jmp 1b
  4004. .cfi_endproc
  4005. /*
  4006. * Copyright 2018-2019 Yury Gribov
  4007. *
  4008. * The MIT License (MIT)
  4009. *
  4010. * Use of this source code is governed by MIT license that can be
  4011. * found in the LICENSE.txt file.
  4012. */
  4013. .globl cuGraphGetRootNodes
  4014. .p2align 4
  4015. .type cuGraphGetRootNodes, %function
  4016. cuGraphGetRootNodes:
  4017. .cfi_startproc
  4018. // Intel opt. manual says to
  4019. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4020. // to hint static predictor.
  4021. cmpq $0, _libcuda_so_tramp_table+1120(%rip)
  4022. je 2f
  4023. 1:
  4024. jmp *_libcuda_so_tramp_table+1120(%rip)
  4025. 2:
  4026. pushq $140
  4027. .cfi_adjust_cfa_offset 8
  4028. call _libcuda_so_save_regs_and_resolve
  4029. addq $8, %rsp
  4030. .cfi_adjust_cfa_offset -8
  4031. jmp 1b
  4032. .cfi_endproc
  4033. /*
  4034. * Copyright 2018-2019 Yury Gribov
  4035. *
  4036. * The MIT License (MIT)
  4037. *
  4038. * Use of this source code is governed by MIT license that can be
  4039. * found in the LICENSE.txt file.
  4040. */
  4041. .globl cuGraphHostNodeGetParams
  4042. .p2align 4
  4043. .type cuGraphHostNodeGetParams, %function
  4044. cuGraphHostNodeGetParams:
  4045. .cfi_startproc
  4046. // Intel opt. manual says to
  4047. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4048. // to hint static predictor.
  4049. cmpq $0, _libcuda_so_tramp_table+1128(%rip)
  4050. je 2f
  4051. 1:
  4052. jmp *_libcuda_so_tramp_table+1128(%rip)
  4053. 2:
  4054. pushq $141
  4055. .cfi_adjust_cfa_offset 8
  4056. call _libcuda_so_save_regs_and_resolve
  4057. addq $8, %rsp
  4058. .cfi_adjust_cfa_offset -8
  4059. jmp 1b
  4060. .cfi_endproc
  4061. /*
  4062. * Copyright 2018-2019 Yury Gribov
  4063. *
  4064. * The MIT License (MIT)
  4065. *
  4066. * Use of this source code is governed by MIT license that can be
  4067. * found in the LICENSE.txt file.
  4068. */
  4069. .globl cuGraphHostNodeSetParams
  4070. .p2align 4
  4071. .type cuGraphHostNodeSetParams, %function
  4072. cuGraphHostNodeSetParams:
  4073. .cfi_startproc
  4074. // Intel opt. manual says to
  4075. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4076. // to hint static predictor.
  4077. cmpq $0, _libcuda_so_tramp_table+1136(%rip)
  4078. je 2f
  4079. 1:
  4080. jmp *_libcuda_so_tramp_table+1136(%rip)
  4081. 2:
  4082. pushq $142
  4083. .cfi_adjust_cfa_offset 8
  4084. call _libcuda_so_save_regs_and_resolve
  4085. addq $8, %rsp
  4086. .cfi_adjust_cfa_offset -8
  4087. jmp 1b
  4088. .cfi_endproc
  4089. /*
  4090. * Copyright 2018-2019 Yury Gribov
  4091. *
  4092. * The MIT License (MIT)
  4093. *
  4094. * Use of this source code is governed by MIT license that can be
  4095. * found in the LICENSE.txt file.
  4096. */
  4097. .globl cuGraphInstantiate
  4098. .p2align 4
  4099. .type cuGraphInstantiate, %function
  4100. cuGraphInstantiate:
  4101. .cfi_startproc
  4102. // Intel opt. manual says to
  4103. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4104. // to hint static predictor.
  4105. cmpq $0, _libcuda_so_tramp_table+1144(%rip)
  4106. je 2f
  4107. 1:
  4108. jmp *_libcuda_so_tramp_table+1144(%rip)
  4109. 2:
  4110. pushq $143
  4111. .cfi_adjust_cfa_offset 8
  4112. call _libcuda_so_save_regs_and_resolve
  4113. addq $8, %rsp
  4114. .cfi_adjust_cfa_offset -8
  4115. jmp 1b
  4116. .cfi_endproc
  4117. /*
  4118. * Copyright 2018-2019 Yury Gribov
  4119. *
  4120. * The MIT License (MIT)
  4121. *
  4122. * Use of this source code is governed by MIT license that can be
  4123. * found in the LICENSE.txt file.
  4124. */
  4125. .globl cuGraphInstantiate_v2
  4126. .p2align 4
  4127. .type cuGraphInstantiate_v2, %function
  4128. cuGraphInstantiate_v2:
  4129. .cfi_startproc
  4130. // Intel opt. manual says to
  4131. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4132. // to hint static predictor.
  4133. cmpq $0, _libcuda_so_tramp_table+1152(%rip)
  4134. je 2f
  4135. 1:
  4136. jmp *_libcuda_so_tramp_table+1152(%rip)
  4137. 2:
  4138. pushq $144
  4139. .cfi_adjust_cfa_offset 8
  4140. call _libcuda_so_save_regs_and_resolve
  4141. addq $8, %rsp
  4142. .cfi_adjust_cfa_offset -8
  4143. jmp 1b
  4144. .cfi_endproc
  4145. /*
  4146. * Copyright 2018-2019 Yury Gribov
  4147. *
  4148. * The MIT License (MIT)
  4149. *
  4150. * Use of this source code is governed by MIT license that can be
  4151. * found in the LICENSE.txt file.
  4152. */
  4153. .globl cuGraphKernelNodeCopyAttributes
  4154. .p2align 4
  4155. .type cuGraphKernelNodeCopyAttributes, %function
  4156. cuGraphKernelNodeCopyAttributes:
  4157. .cfi_startproc
  4158. // Intel opt. manual says to
  4159. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4160. // to hint static predictor.
  4161. cmpq $0, _libcuda_so_tramp_table+1160(%rip)
  4162. je 2f
  4163. 1:
  4164. jmp *_libcuda_so_tramp_table+1160(%rip)
  4165. 2:
  4166. pushq $145
  4167. .cfi_adjust_cfa_offset 8
  4168. call _libcuda_so_save_regs_and_resolve
  4169. addq $8, %rsp
  4170. .cfi_adjust_cfa_offset -8
  4171. jmp 1b
  4172. .cfi_endproc
  4173. /*
  4174. * Copyright 2018-2019 Yury Gribov
  4175. *
  4176. * The MIT License (MIT)
  4177. *
  4178. * Use of this source code is governed by MIT license that can be
  4179. * found in the LICENSE.txt file.
  4180. */
  4181. .globl cuGraphKernelNodeGetAttribute
  4182. .p2align 4
  4183. .type cuGraphKernelNodeGetAttribute, %function
  4184. cuGraphKernelNodeGetAttribute:
  4185. .cfi_startproc
  4186. // Intel opt. manual says to
  4187. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4188. // to hint static predictor.
  4189. cmpq $0, _libcuda_so_tramp_table+1168(%rip)
  4190. je 2f
  4191. 1:
  4192. jmp *_libcuda_so_tramp_table+1168(%rip)
  4193. 2:
  4194. pushq $146
  4195. .cfi_adjust_cfa_offset 8
  4196. call _libcuda_so_save_regs_and_resolve
  4197. addq $8, %rsp
  4198. .cfi_adjust_cfa_offset -8
  4199. jmp 1b
  4200. .cfi_endproc
  4201. /*
  4202. * Copyright 2018-2019 Yury Gribov
  4203. *
  4204. * The MIT License (MIT)
  4205. *
  4206. * Use of this source code is governed by MIT license that can be
  4207. * found in the LICENSE.txt file.
  4208. */
  4209. .globl cuGraphKernelNodeGetParams
  4210. .p2align 4
  4211. .type cuGraphKernelNodeGetParams, %function
  4212. cuGraphKernelNodeGetParams:
  4213. .cfi_startproc
  4214. // Intel opt. manual says to
  4215. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4216. // to hint static predictor.
  4217. cmpq $0, _libcuda_so_tramp_table+1176(%rip)
  4218. je 2f
  4219. 1:
  4220. jmp *_libcuda_so_tramp_table+1176(%rip)
  4221. 2:
  4222. pushq $147
  4223. .cfi_adjust_cfa_offset 8
  4224. call _libcuda_so_save_regs_and_resolve
  4225. addq $8, %rsp
  4226. .cfi_adjust_cfa_offset -8
  4227. jmp 1b
  4228. .cfi_endproc
  4229. /*
  4230. * Copyright 2018-2019 Yury Gribov
  4231. *
  4232. * The MIT License (MIT)
  4233. *
  4234. * Use of this source code is governed by MIT license that can be
  4235. * found in the LICENSE.txt file.
  4236. */
  4237. .globl cuGraphKernelNodeSetAttribute
  4238. .p2align 4
  4239. .type cuGraphKernelNodeSetAttribute, %function
  4240. cuGraphKernelNodeSetAttribute:
  4241. .cfi_startproc
  4242. // Intel opt. manual says to
  4243. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4244. // to hint static predictor.
  4245. cmpq $0, _libcuda_so_tramp_table+1184(%rip)
  4246. je 2f
  4247. 1:
  4248. jmp *_libcuda_so_tramp_table+1184(%rip)
  4249. 2:
  4250. pushq $148
  4251. .cfi_adjust_cfa_offset 8
  4252. call _libcuda_so_save_regs_and_resolve
  4253. addq $8, %rsp
  4254. .cfi_adjust_cfa_offset -8
  4255. jmp 1b
  4256. .cfi_endproc
  4257. /*
  4258. * Copyright 2018-2019 Yury Gribov
  4259. *
  4260. * The MIT License (MIT)
  4261. *
  4262. * Use of this source code is governed by MIT license that can be
  4263. * found in the LICENSE.txt file.
  4264. */
  4265. .globl cuGraphKernelNodeSetParams
  4266. .p2align 4
  4267. .type cuGraphKernelNodeSetParams, %function
  4268. cuGraphKernelNodeSetParams:
  4269. .cfi_startproc
  4270. // Intel opt. manual says to
  4271. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4272. // to hint static predictor.
  4273. cmpq $0, _libcuda_so_tramp_table+1192(%rip)
  4274. je 2f
  4275. 1:
  4276. jmp *_libcuda_so_tramp_table+1192(%rip)
  4277. 2:
  4278. pushq $149
  4279. .cfi_adjust_cfa_offset 8
  4280. call _libcuda_so_save_regs_and_resolve
  4281. addq $8, %rsp
  4282. .cfi_adjust_cfa_offset -8
  4283. jmp 1b
  4284. .cfi_endproc
  4285. /*
  4286. * Copyright 2018-2019 Yury Gribov
  4287. *
  4288. * The MIT License (MIT)
  4289. *
  4290. * Use of this source code is governed by MIT license that can be
  4291. * found in the LICENSE.txt file.
  4292. */
  4293. .globl cuGraphLaunch
  4294. .p2align 4
  4295. .type cuGraphLaunch, %function
  4296. cuGraphLaunch:
  4297. .cfi_startproc
  4298. // Intel opt. manual says to
  4299. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4300. // to hint static predictor.
  4301. cmpq $0, _libcuda_so_tramp_table+1200(%rip)
  4302. je 2f
  4303. 1:
  4304. jmp *_libcuda_so_tramp_table+1200(%rip)
  4305. 2:
  4306. pushq $150
  4307. .cfi_adjust_cfa_offset 8
  4308. call _libcuda_so_save_regs_and_resolve
  4309. addq $8, %rsp
  4310. .cfi_adjust_cfa_offset -8
  4311. jmp 1b
  4312. .cfi_endproc
  4313. /*
  4314. * Copyright 2018-2019 Yury Gribov
  4315. *
  4316. * The MIT License (MIT)
  4317. *
  4318. * Use of this source code is governed by MIT license that can be
  4319. * found in the LICENSE.txt file.
  4320. */
  4321. .globl cuGraphLaunch_ptsz
  4322. .p2align 4
  4323. .type cuGraphLaunch_ptsz, %function
  4324. cuGraphLaunch_ptsz:
  4325. .cfi_startproc
  4326. // Intel opt. manual says to
  4327. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4328. // to hint static predictor.
  4329. cmpq $0, _libcuda_so_tramp_table+1208(%rip)
  4330. je 2f
  4331. 1:
  4332. jmp *_libcuda_so_tramp_table+1208(%rip)
  4333. 2:
  4334. pushq $151
  4335. .cfi_adjust_cfa_offset 8
  4336. call _libcuda_so_save_regs_and_resolve
  4337. addq $8, %rsp
  4338. .cfi_adjust_cfa_offset -8
  4339. jmp 1b
  4340. .cfi_endproc
  4341. /*
  4342. * Copyright 2018-2019 Yury Gribov
  4343. *
  4344. * The MIT License (MIT)
  4345. *
  4346. * Use of this source code is governed by MIT license that can be
  4347. * found in the LICENSE.txt file.
  4348. */
  4349. .globl cuGraphMemcpyNodeGetParams
  4350. .p2align 4
  4351. .type cuGraphMemcpyNodeGetParams, %function
  4352. cuGraphMemcpyNodeGetParams:
  4353. .cfi_startproc
  4354. // Intel opt. manual says to
  4355. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4356. // to hint static predictor.
  4357. cmpq $0, _libcuda_so_tramp_table+1216(%rip)
  4358. je 2f
  4359. 1:
  4360. jmp *_libcuda_so_tramp_table+1216(%rip)
  4361. 2:
  4362. pushq $152
  4363. .cfi_adjust_cfa_offset 8
  4364. call _libcuda_so_save_regs_and_resolve
  4365. addq $8, %rsp
  4366. .cfi_adjust_cfa_offset -8
  4367. jmp 1b
  4368. .cfi_endproc
  4369. /*
  4370. * Copyright 2018-2019 Yury Gribov
  4371. *
  4372. * The MIT License (MIT)
  4373. *
  4374. * Use of this source code is governed by MIT license that can be
  4375. * found in the LICENSE.txt file.
  4376. */
  4377. .globl cuGraphMemcpyNodeSetParams
  4378. .p2align 4
  4379. .type cuGraphMemcpyNodeSetParams, %function
  4380. cuGraphMemcpyNodeSetParams:
  4381. .cfi_startproc
  4382. // Intel opt. manual says to
  4383. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4384. // to hint static predictor.
  4385. cmpq $0, _libcuda_so_tramp_table+1224(%rip)
  4386. je 2f
  4387. 1:
  4388. jmp *_libcuda_so_tramp_table+1224(%rip)
  4389. 2:
  4390. pushq $153
  4391. .cfi_adjust_cfa_offset 8
  4392. call _libcuda_so_save_regs_and_resolve
  4393. addq $8, %rsp
  4394. .cfi_adjust_cfa_offset -8
  4395. jmp 1b
  4396. .cfi_endproc
  4397. /*
  4398. * Copyright 2018-2019 Yury Gribov
  4399. *
  4400. * The MIT License (MIT)
  4401. *
  4402. * Use of this source code is governed by MIT license that can be
  4403. * found in the LICENSE.txt file.
  4404. */
  4405. .globl cuGraphMemsetNodeGetParams
  4406. .p2align 4
  4407. .type cuGraphMemsetNodeGetParams, %function
  4408. cuGraphMemsetNodeGetParams:
  4409. .cfi_startproc
  4410. // Intel opt. manual says to
  4411. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4412. // to hint static predictor.
  4413. cmpq $0, _libcuda_so_tramp_table+1232(%rip)
  4414. je 2f
  4415. 1:
  4416. jmp *_libcuda_so_tramp_table+1232(%rip)
  4417. 2:
  4418. pushq $154
  4419. .cfi_adjust_cfa_offset 8
  4420. call _libcuda_so_save_regs_and_resolve
  4421. addq $8, %rsp
  4422. .cfi_adjust_cfa_offset -8
  4423. jmp 1b
  4424. .cfi_endproc
  4425. /*
  4426. * Copyright 2018-2019 Yury Gribov
  4427. *
  4428. * The MIT License (MIT)
  4429. *
  4430. * Use of this source code is governed by MIT license that can be
  4431. * found in the LICENSE.txt file.
  4432. */
  4433. .globl cuGraphMemsetNodeSetParams
  4434. .p2align 4
  4435. .type cuGraphMemsetNodeSetParams, %function
  4436. cuGraphMemsetNodeSetParams:
  4437. .cfi_startproc
  4438. // Intel opt. manual says to
  4439. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4440. // to hint static predictor.
  4441. cmpq $0, _libcuda_so_tramp_table+1240(%rip)
  4442. je 2f
  4443. 1:
  4444. jmp *_libcuda_so_tramp_table+1240(%rip)
  4445. 2:
  4446. pushq $155
  4447. .cfi_adjust_cfa_offset 8
  4448. call _libcuda_so_save_regs_and_resolve
  4449. addq $8, %rsp
  4450. .cfi_adjust_cfa_offset -8
  4451. jmp 1b
  4452. .cfi_endproc
  4453. /*
  4454. * Copyright 2018-2019 Yury Gribov
  4455. *
  4456. * The MIT License (MIT)
  4457. *
  4458. * Use of this source code is governed by MIT license that can be
  4459. * found in the LICENSE.txt file.
  4460. */
  4461. .globl cuGraphNodeFindInClone
  4462. .p2align 4
  4463. .type cuGraphNodeFindInClone, %function
  4464. cuGraphNodeFindInClone:
  4465. .cfi_startproc
  4466. // Intel opt. manual says to
  4467. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4468. // to hint static predictor.
  4469. cmpq $0, _libcuda_so_tramp_table+1248(%rip)
  4470. je 2f
  4471. 1:
  4472. jmp *_libcuda_so_tramp_table+1248(%rip)
  4473. 2:
  4474. pushq $156
  4475. .cfi_adjust_cfa_offset 8
  4476. call _libcuda_so_save_regs_and_resolve
  4477. addq $8, %rsp
  4478. .cfi_adjust_cfa_offset -8
  4479. jmp 1b
  4480. .cfi_endproc
  4481. /*
  4482. * Copyright 2018-2019 Yury Gribov
  4483. *
  4484. * The MIT License (MIT)
  4485. *
  4486. * Use of this source code is governed by MIT license that can be
  4487. * found in the LICENSE.txt file.
  4488. */
  4489. .globl cuGraphNodeGetDependencies
  4490. .p2align 4
  4491. .type cuGraphNodeGetDependencies, %function
  4492. cuGraphNodeGetDependencies:
  4493. .cfi_startproc
  4494. // Intel opt. manual says to
  4495. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4496. // to hint static predictor.
  4497. cmpq $0, _libcuda_so_tramp_table+1256(%rip)
  4498. je 2f
  4499. 1:
  4500. jmp *_libcuda_so_tramp_table+1256(%rip)
  4501. 2:
  4502. pushq $157
  4503. .cfi_adjust_cfa_offset 8
  4504. call _libcuda_so_save_regs_and_resolve
  4505. addq $8, %rsp
  4506. .cfi_adjust_cfa_offset -8
  4507. jmp 1b
  4508. .cfi_endproc
  4509. /*
  4510. * Copyright 2018-2019 Yury Gribov
  4511. *
  4512. * The MIT License (MIT)
  4513. *
  4514. * Use of this source code is governed by MIT license that can be
  4515. * found in the LICENSE.txt file.
  4516. */
  4517. .globl cuGraphNodeGetDependentNodes
  4518. .p2align 4
  4519. .type cuGraphNodeGetDependentNodes, %function
  4520. cuGraphNodeGetDependentNodes:
  4521. .cfi_startproc
  4522. // Intel opt. manual says to
  4523. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4524. // to hint static predictor.
  4525. cmpq $0, _libcuda_so_tramp_table+1264(%rip)
  4526. je 2f
  4527. 1:
  4528. jmp *_libcuda_so_tramp_table+1264(%rip)
  4529. 2:
  4530. pushq $158
  4531. .cfi_adjust_cfa_offset 8
  4532. call _libcuda_so_save_regs_and_resolve
  4533. addq $8, %rsp
  4534. .cfi_adjust_cfa_offset -8
  4535. jmp 1b
  4536. .cfi_endproc
  4537. /*
  4538. * Copyright 2018-2019 Yury Gribov
  4539. *
  4540. * The MIT License (MIT)
  4541. *
  4542. * Use of this source code is governed by MIT license that can be
  4543. * found in the LICENSE.txt file.
  4544. */
  4545. .globl cuGraphNodeGetType
  4546. .p2align 4
  4547. .type cuGraphNodeGetType, %function
  4548. cuGraphNodeGetType:
  4549. .cfi_startproc
  4550. // Intel opt. manual says to
  4551. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4552. // to hint static predictor.
  4553. cmpq $0, _libcuda_so_tramp_table+1272(%rip)
  4554. je 2f
  4555. 1:
  4556. jmp *_libcuda_so_tramp_table+1272(%rip)
  4557. 2:
  4558. pushq $159
  4559. .cfi_adjust_cfa_offset 8
  4560. call _libcuda_so_save_regs_and_resolve
  4561. addq $8, %rsp
  4562. .cfi_adjust_cfa_offset -8
  4563. jmp 1b
  4564. .cfi_endproc
  4565. /*
  4566. * Copyright 2018-2019 Yury Gribov
  4567. *
  4568. * The MIT License (MIT)
  4569. *
  4570. * Use of this source code is governed by MIT license that can be
  4571. * found in the LICENSE.txt file.
  4572. */
  4573. .globl cuGraphRemoveDependencies
  4574. .p2align 4
  4575. .type cuGraphRemoveDependencies, %function
  4576. cuGraphRemoveDependencies:
  4577. .cfi_startproc
  4578. // Intel opt. manual says to
  4579. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4580. // to hint static predictor.
  4581. cmpq $0, _libcuda_so_tramp_table+1280(%rip)
  4582. je 2f
  4583. 1:
  4584. jmp *_libcuda_so_tramp_table+1280(%rip)
  4585. 2:
  4586. pushq $160
  4587. .cfi_adjust_cfa_offset 8
  4588. call _libcuda_so_save_regs_and_resolve
  4589. addq $8, %rsp
  4590. .cfi_adjust_cfa_offset -8
  4591. jmp 1b
  4592. .cfi_endproc
  4593. /*
  4594. * Copyright 2018-2019 Yury Gribov
  4595. *
  4596. * The MIT License (MIT)
  4597. *
  4598. * Use of this source code is governed by MIT license that can be
  4599. * found in the LICENSE.txt file.
  4600. */
  4601. .globl cuGraphUpload
  4602. .p2align 4
  4603. .type cuGraphUpload, %function
  4604. cuGraphUpload:
  4605. .cfi_startproc
  4606. // Intel opt. manual says to
  4607. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4608. // to hint static predictor.
  4609. cmpq $0, _libcuda_so_tramp_table+1288(%rip)
  4610. je 2f
  4611. 1:
  4612. jmp *_libcuda_so_tramp_table+1288(%rip)
  4613. 2:
  4614. pushq $161
  4615. .cfi_adjust_cfa_offset 8
  4616. call _libcuda_so_save_regs_and_resolve
  4617. addq $8, %rsp
  4618. .cfi_adjust_cfa_offset -8
  4619. jmp 1b
  4620. .cfi_endproc
  4621. /*
  4622. * Copyright 2018-2019 Yury Gribov
  4623. *
  4624. * The MIT License (MIT)
  4625. *
  4626. * Use of this source code is governed by MIT license that can be
  4627. * found in the LICENSE.txt file.
  4628. */
  4629. .globl cuGraphUpload_ptsz
  4630. .p2align 4
  4631. .type cuGraphUpload_ptsz, %function
  4632. cuGraphUpload_ptsz:
  4633. .cfi_startproc
  4634. // Intel opt. manual says to
  4635. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4636. // to hint static predictor.
  4637. cmpq $0, _libcuda_so_tramp_table+1296(%rip)
  4638. je 2f
  4639. 1:
  4640. jmp *_libcuda_so_tramp_table+1296(%rip)
  4641. 2:
  4642. pushq $162
  4643. .cfi_adjust_cfa_offset 8
  4644. call _libcuda_so_save_regs_and_resolve
  4645. addq $8, %rsp
  4646. .cfi_adjust_cfa_offset -8
  4647. jmp 1b
  4648. .cfi_endproc
  4649. /*
  4650. * Copyright 2018-2019 Yury Gribov
  4651. *
  4652. * The MIT License (MIT)
  4653. *
  4654. * Use of this source code is governed by MIT license that can be
  4655. * found in the LICENSE.txt file.
  4656. */
  4657. .globl cuGraphicsEGLRegisterImage
  4658. .p2align 4
  4659. .type cuGraphicsEGLRegisterImage, %function
  4660. cuGraphicsEGLRegisterImage:
  4661. .cfi_startproc
  4662. // Intel opt. manual says to
  4663. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4664. // to hint static predictor.
  4665. cmpq $0, _libcuda_so_tramp_table+1304(%rip)
  4666. je 2f
  4667. 1:
  4668. jmp *_libcuda_so_tramp_table+1304(%rip)
  4669. 2:
  4670. pushq $163
  4671. .cfi_adjust_cfa_offset 8
  4672. call _libcuda_so_save_regs_and_resolve
  4673. addq $8, %rsp
  4674. .cfi_adjust_cfa_offset -8
  4675. jmp 1b
  4676. .cfi_endproc
  4677. /*
  4678. * Copyright 2018-2019 Yury Gribov
  4679. *
  4680. * The MIT License (MIT)
  4681. *
  4682. * Use of this source code is governed by MIT license that can be
  4683. * found in the LICENSE.txt file.
  4684. */
  4685. .globl cuGraphicsGLRegisterBuffer
  4686. .p2align 4
  4687. .type cuGraphicsGLRegisterBuffer, %function
  4688. cuGraphicsGLRegisterBuffer:
  4689. .cfi_startproc
  4690. // Intel opt. manual says to
  4691. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4692. // to hint static predictor.
  4693. cmpq $0, _libcuda_so_tramp_table+1312(%rip)
  4694. je 2f
  4695. 1:
  4696. jmp *_libcuda_so_tramp_table+1312(%rip)
  4697. 2:
  4698. pushq $164
  4699. .cfi_adjust_cfa_offset 8
  4700. call _libcuda_so_save_regs_and_resolve
  4701. addq $8, %rsp
  4702. .cfi_adjust_cfa_offset -8
  4703. jmp 1b
  4704. .cfi_endproc
  4705. /*
  4706. * Copyright 2018-2019 Yury Gribov
  4707. *
  4708. * The MIT License (MIT)
  4709. *
  4710. * Use of this source code is governed by MIT license that can be
  4711. * found in the LICENSE.txt file.
  4712. */
  4713. .globl cuGraphicsGLRegisterImage
  4714. .p2align 4
  4715. .type cuGraphicsGLRegisterImage, %function
  4716. cuGraphicsGLRegisterImage:
  4717. .cfi_startproc
  4718. // Intel opt. manual says to
  4719. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4720. // to hint static predictor.
  4721. cmpq $0, _libcuda_so_tramp_table+1320(%rip)
  4722. je 2f
  4723. 1:
  4724. jmp *_libcuda_so_tramp_table+1320(%rip)
  4725. 2:
  4726. pushq $165
  4727. .cfi_adjust_cfa_offset 8
  4728. call _libcuda_so_save_regs_and_resolve
  4729. addq $8, %rsp
  4730. .cfi_adjust_cfa_offset -8
  4731. jmp 1b
  4732. .cfi_endproc
  4733. /*
  4734. * Copyright 2018-2019 Yury Gribov
  4735. *
  4736. * The MIT License (MIT)
  4737. *
  4738. * Use of this source code is governed by MIT license that can be
  4739. * found in the LICENSE.txt file.
  4740. */
  4741. .globl cuGraphicsMapResources
  4742. .p2align 4
  4743. .type cuGraphicsMapResources, %function
  4744. cuGraphicsMapResources:
  4745. .cfi_startproc
  4746. // Intel opt. manual says to
  4747. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4748. // to hint static predictor.
  4749. cmpq $0, _libcuda_so_tramp_table+1328(%rip)
  4750. je 2f
  4751. 1:
  4752. jmp *_libcuda_so_tramp_table+1328(%rip)
  4753. 2:
  4754. pushq $166
  4755. .cfi_adjust_cfa_offset 8
  4756. call _libcuda_so_save_regs_and_resolve
  4757. addq $8, %rsp
  4758. .cfi_adjust_cfa_offset -8
  4759. jmp 1b
  4760. .cfi_endproc
  4761. /*
  4762. * Copyright 2018-2019 Yury Gribov
  4763. *
  4764. * The MIT License (MIT)
  4765. *
  4766. * Use of this source code is governed by MIT license that can be
  4767. * found in the LICENSE.txt file.
  4768. */
  4769. .globl cuGraphicsMapResources_ptsz
  4770. .p2align 4
  4771. .type cuGraphicsMapResources_ptsz, %function
  4772. cuGraphicsMapResources_ptsz:
  4773. .cfi_startproc
  4774. // Intel opt. manual says to
  4775. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4776. // to hint static predictor.
  4777. cmpq $0, _libcuda_so_tramp_table+1336(%rip)
  4778. je 2f
  4779. 1:
  4780. jmp *_libcuda_so_tramp_table+1336(%rip)
  4781. 2:
  4782. pushq $167
  4783. .cfi_adjust_cfa_offset 8
  4784. call _libcuda_so_save_regs_and_resolve
  4785. addq $8, %rsp
  4786. .cfi_adjust_cfa_offset -8
  4787. jmp 1b
  4788. .cfi_endproc
  4789. /*
  4790. * Copyright 2018-2019 Yury Gribov
  4791. *
  4792. * The MIT License (MIT)
  4793. *
  4794. * Use of this source code is governed by MIT license that can be
  4795. * found in the LICENSE.txt file.
  4796. */
  4797. .globl cuGraphicsResourceGetMappedEglFrame
  4798. .p2align 4
  4799. .type cuGraphicsResourceGetMappedEglFrame, %function
  4800. cuGraphicsResourceGetMappedEglFrame:
  4801. .cfi_startproc
  4802. // Intel opt. manual says to
  4803. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4804. // to hint static predictor.
  4805. cmpq $0, _libcuda_so_tramp_table+1344(%rip)
  4806. je 2f
  4807. 1:
  4808. jmp *_libcuda_so_tramp_table+1344(%rip)
  4809. 2:
  4810. pushq $168
  4811. .cfi_adjust_cfa_offset 8
  4812. call _libcuda_so_save_regs_and_resolve
  4813. addq $8, %rsp
  4814. .cfi_adjust_cfa_offset -8
  4815. jmp 1b
  4816. .cfi_endproc
  4817. /*
  4818. * Copyright 2018-2019 Yury Gribov
  4819. *
  4820. * The MIT License (MIT)
  4821. *
  4822. * Use of this source code is governed by MIT license that can be
  4823. * found in the LICENSE.txt file.
  4824. */
  4825. .globl cuGraphicsResourceGetMappedMipmappedArray
  4826. .p2align 4
  4827. .type cuGraphicsResourceGetMappedMipmappedArray, %function
  4828. cuGraphicsResourceGetMappedMipmappedArray:
  4829. .cfi_startproc
  4830. // Intel opt. manual says to
  4831. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4832. // to hint static predictor.
  4833. cmpq $0, _libcuda_so_tramp_table+1352(%rip)
  4834. je 2f
  4835. 1:
  4836. jmp *_libcuda_so_tramp_table+1352(%rip)
  4837. 2:
  4838. pushq $169
  4839. .cfi_adjust_cfa_offset 8
  4840. call _libcuda_so_save_regs_and_resolve
  4841. addq $8, %rsp
  4842. .cfi_adjust_cfa_offset -8
  4843. jmp 1b
  4844. .cfi_endproc
  4845. /*
  4846. * Copyright 2018-2019 Yury Gribov
  4847. *
  4848. * The MIT License (MIT)
  4849. *
  4850. * Use of this source code is governed by MIT license that can be
  4851. * found in the LICENSE.txt file.
  4852. */
  4853. .globl cuGraphicsResourceGetMappedPointer
  4854. .p2align 4
  4855. .type cuGraphicsResourceGetMappedPointer, %function
  4856. cuGraphicsResourceGetMappedPointer:
  4857. .cfi_startproc
  4858. // Intel opt. manual says to
  4859. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4860. // to hint static predictor.
  4861. cmpq $0, _libcuda_so_tramp_table+1360(%rip)
  4862. je 2f
  4863. 1:
  4864. jmp *_libcuda_so_tramp_table+1360(%rip)
  4865. 2:
  4866. pushq $170
  4867. .cfi_adjust_cfa_offset 8
  4868. call _libcuda_so_save_regs_and_resolve
  4869. addq $8, %rsp
  4870. .cfi_adjust_cfa_offset -8
  4871. jmp 1b
  4872. .cfi_endproc
  4873. /*
  4874. * Copyright 2018-2019 Yury Gribov
  4875. *
  4876. * The MIT License (MIT)
  4877. *
  4878. * Use of this source code is governed by MIT license that can be
  4879. * found in the LICENSE.txt file.
  4880. */
  4881. .globl cuGraphicsResourceGetMappedPointer_v2
  4882. .p2align 4
  4883. .type cuGraphicsResourceGetMappedPointer_v2, %function
  4884. cuGraphicsResourceGetMappedPointer_v2:
  4885. .cfi_startproc
  4886. // Intel opt. manual says to
  4887. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4888. // to hint static predictor.
  4889. cmpq $0, _libcuda_so_tramp_table+1368(%rip)
  4890. je 2f
  4891. 1:
  4892. jmp *_libcuda_so_tramp_table+1368(%rip)
  4893. 2:
  4894. pushq $171
  4895. .cfi_adjust_cfa_offset 8
  4896. call _libcuda_so_save_regs_and_resolve
  4897. addq $8, %rsp
  4898. .cfi_adjust_cfa_offset -8
  4899. jmp 1b
  4900. .cfi_endproc
  4901. /*
  4902. * Copyright 2018-2019 Yury Gribov
  4903. *
  4904. * The MIT License (MIT)
  4905. *
  4906. * Use of this source code is governed by MIT license that can be
  4907. * found in the LICENSE.txt file.
  4908. */
  4909. .globl cuGraphicsResourceSetMapFlags
  4910. .p2align 4
  4911. .type cuGraphicsResourceSetMapFlags, %function
  4912. cuGraphicsResourceSetMapFlags:
  4913. .cfi_startproc
  4914. // Intel opt. manual says to
  4915. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4916. // to hint static predictor.
  4917. cmpq $0, _libcuda_so_tramp_table+1376(%rip)
  4918. je 2f
  4919. 1:
  4920. jmp *_libcuda_so_tramp_table+1376(%rip)
  4921. 2:
  4922. pushq $172
  4923. .cfi_adjust_cfa_offset 8
  4924. call _libcuda_so_save_regs_and_resolve
  4925. addq $8, %rsp
  4926. .cfi_adjust_cfa_offset -8
  4927. jmp 1b
  4928. .cfi_endproc
  4929. /*
  4930. * Copyright 2018-2019 Yury Gribov
  4931. *
  4932. * The MIT License (MIT)
  4933. *
  4934. * Use of this source code is governed by MIT license that can be
  4935. * found in the LICENSE.txt file.
  4936. */
  4937. .globl cuGraphicsResourceSetMapFlags_v2
  4938. .p2align 4
  4939. .type cuGraphicsResourceSetMapFlags_v2, %function
  4940. cuGraphicsResourceSetMapFlags_v2:
  4941. .cfi_startproc
  4942. // Intel opt. manual says to
  4943. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4944. // to hint static predictor.
  4945. cmpq $0, _libcuda_so_tramp_table+1384(%rip)
  4946. je 2f
  4947. 1:
  4948. jmp *_libcuda_so_tramp_table+1384(%rip)
  4949. 2:
  4950. pushq $173
  4951. .cfi_adjust_cfa_offset 8
  4952. call _libcuda_so_save_regs_and_resolve
  4953. addq $8, %rsp
  4954. .cfi_adjust_cfa_offset -8
  4955. jmp 1b
  4956. .cfi_endproc
  4957. /*
  4958. * Copyright 2018-2019 Yury Gribov
  4959. *
  4960. * The MIT License (MIT)
  4961. *
  4962. * Use of this source code is governed by MIT license that can be
  4963. * found in the LICENSE.txt file.
  4964. */
  4965. .globl cuGraphicsSubResourceGetMappedArray
  4966. .p2align 4
  4967. .type cuGraphicsSubResourceGetMappedArray, %function
  4968. cuGraphicsSubResourceGetMappedArray:
  4969. .cfi_startproc
  4970. // Intel opt. manual says to
  4971. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  4972. // to hint static predictor.
  4973. cmpq $0, _libcuda_so_tramp_table+1392(%rip)
  4974. je 2f
  4975. 1:
  4976. jmp *_libcuda_so_tramp_table+1392(%rip)
  4977. 2:
  4978. pushq $174
  4979. .cfi_adjust_cfa_offset 8
  4980. call _libcuda_so_save_regs_and_resolve
  4981. addq $8, %rsp
  4982. .cfi_adjust_cfa_offset -8
  4983. jmp 1b
  4984. .cfi_endproc
  4985. /*
  4986. * Copyright 2018-2019 Yury Gribov
  4987. *
  4988. * The MIT License (MIT)
  4989. *
  4990. * Use of this source code is governed by MIT license that can be
  4991. * found in the LICENSE.txt file.
  4992. */
  4993. .globl cuGraphicsUnmapResources
  4994. .p2align 4
  4995. .type cuGraphicsUnmapResources, %function
  4996. cuGraphicsUnmapResources:
  4997. .cfi_startproc
  4998. // Intel opt. manual says to
  4999. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5000. // to hint static predictor.
  5001. cmpq $0, _libcuda_so_tramp_table+1400(%rip)
  5002. je 2f
  5003. 1:
  5004. jmp *_libcuda_so_tramp_table+1400(%rip)
  5005. 2:
  5006. pushq $175
  5007. .cfi_adjust_cfa_offset 8
  5008. call _libcuda_so_save_regs_and_resolve
  5009. addq $8, %rsp
  5010. .cfi_adjust_cfa_offset -8
  5011. jmp 1b
  5012. .cfi_endproc
  5013. /*
  5014. * Copyright 2018-2019 Yury Gribov
  5015. *
  5016. * The MIT License (MIT)
  5017. *
  5018. * Use of this source code is governed by MIT license that can be
  5019. * found in the LICENSE.txt file.
  5020. */
  5021. .globl cuGraphicsUnmapResources_ptsz
  5022. .p2align 4
  5023. .type cuGraphicsUnmapResources_ptsz, %function
  5024. cuGraphicsUnmapResources_ptsz:
  5025. .cfi_startproc
  5026. // Intel opt. manual says to
  5027. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5028. // to hint static predictor.
  5029. cmpq $0, _libcuda_so_tramp_table+1408(%rip)
  5030. je 2f
  5031. 1:
  5032. jmp *_libcuda_so_tramp_table+1408(%rip)
  5033. 2:
  5034. pushq $176
  5035. .cfi_adjust_cfa_offset 8
  5036. call _libcuda_so_save_regs_and_resolve
  5037. addq $8, %rsp
  5038. .cfi_adjust_cfa_offset -8
  5039. jmp 1b
  5040. .cfi_endproc
  5041. /*
  5042. * Copyright 2018-2019 Yury Gribov
  5043. *
  5044. * The MIT License (MIT)
  5045. *
  5046. * Use of this source code is governed by MIT license that can be
  5047. * found in the LICENSE.txt file.
  5048. */
  5049. .globl cuGraphicsUnregisterResource
  5050. .p2align 4
  5051. .type cuGraphicsUnregisterResource, %function
  5052. cuGraphicsUnregisterResource:
  5053. .cfi_startproc
  5054. // Intel opt. manual says to
  5055. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5056. // to hint static predictor.
  5057. cmpq $0, _libcuda_so_tramp_table+1416(%rip)
  5058. je 2f
  5059. 1:
  5060. jmp *_libcuda_so_tramp_table+1416(%rip)
  5061. 2:
  5062. pushq $177
  5063. .cfi_adjust_cfa_offset 8
  5064. call _libcuda_so_save_regs_and_resolve
  5065. addq $8, %rsp
  5066. .cfi_adjust_cfa_offset -8
  5067. jmp 1b
  5068. .cfi_endproc
  5069. /*
  5070. * Copyright 2018-2019 Yury Gribov
  5071. *
  5072. * The MIT License (MIT)
  5073. *
  5074. * Use of this source code is governed by MIT license that can be
  5075. * found in the LICENSE.txt file.
  5076. */
  5077. .globl cuGraphicsVDPAURegisterOutputSurface
  5078. .p2align 4
  5079. .type cuGraphicsVDPAURegisterOutputSurface, %function
  5080. cuGraphicsVDPAURegisterOutputSurface:
  5081. .cfi_startproc
  5082. // Intel opt. manual says to
  5083. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5084. // to hint static predictor.
  5085. cmpq $0, _libcuda_so_tramp_table+1424(%rip)
  5086. je 2f
  5087. 1:
  5088. jmp *_libcuda_so_tramp_table+1424(%rip)
  5089. 2:
  5090. pushq $178
  5091. .cfi_adjust_cfa_offset 8
  5092. call _libcuda_so_save_regs_and_resolve
  5093. addq $8, %rsp
  5094. .cfi_adjust_cfa_offset -8
  5095. jmp 1b
  5096. .cfi_endproc
  5097. /*
  5098. * Copyright 2018-2019 Yury Gribov
  5099. *
  5100. * The MIT License (MIT)
  5101. *
  5102. * Use of this source code is governed by MIT license that can be
  5103. * found in the LICENSE.txt file.
  5104. */
  5105. .globl cuGraphicsVDPAURegisterVideoSurface
  5106. .p2align 4
  5107. .type cuGraphicsVDPAURegisterVideoSurface, %function
  5108. cuGraphicsVDPAURegisterVideoSurface:
  5109. .cfi_startproc
  5110. // Intel opt. manual says to
  5111. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5112. // to hint static predictor.
  5113. cmpq $0, _libcuda_so_tramp_table+1432(%rip)
  5114. je 2f
  5115. 1:
  5116. jmp *_libcuda_so_tramp_table+1432(%rip)
  5117. 2:
  5118. pushq $179
  5119. .cfi_adjust_cfa_offset 8
  5120. call _libcuda_so_save_regs_and_resolve
  5121. addq $8, %rsp
  5122. .cfi_adjust_cfa_offset -8
  5123. jmp 1b
  5124. .cfi_endproc
  5125. /*
  5126. * Copyright 2018-2019 Yury Gribov
  5127. *
  5128. * The MIT License (MIT)
  5129. *
  5130. * Use of this source code is governed by MIT license that can be
  5131. * found in the LICENSE.txt file.
  5132. */
  5133. .globl cuImportExternalMemory
  5134. .p2align 4
  5135. .type cuImportExternalMemory, %function
  5136. cuImportExternalMemory:
  5137. .cfi_startproc
  5138. // Intel opt. manual says to
  5139. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5140. // to hint static predictor.
  5141. cmpq $0, _libcuda_so_tramp_table+1440(%rip)
  5142. je 2f
  5143. 1:
  5144. jmp *_libcuda_so_tramp_table+1440(%rip)
  5145. 2:
  5146. pushq $180
  5147. .cfi_adjust_cfa_offset 8
  5148. call _libcuda_so_save_regs_and_resolve
  5149. addq $8, %rsp
  5150. .cfi_adjust_cfa_offset -8
  5151. jmp 1b
  5152. .cfi_endproc
  5153. /*
  5154. * Copyright 2018-2019 Yury Gribov
  5155. *
  5156. * The MIT License (MIT)
  5157. *
  5158. * Use of this source code is governed by MIT license that can be
  5159. * found in the LICENSE.txt file.
  5160. */
  5161. .globl cuImportExternalSemaphore
  5162. .p2align 4
  5163. .type cuImportExternalSemaphore, %function
  5164. cuImportExternalSemaphore:
  5165. .cfi_startproc
  5166. // Intel opt. manual says to
  5167. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5168. // to hint static predictor.
  5169. cmpq $0, _libcuda_so_tramp_table+1448(%rip)
  5170. je 2f
  5171. 1:
  5172. jmp *_libcuda_so_tramp_table+1448(%rip)
  5173. 2:
  5174. pushq $181
  5175. .cfi_adjust_cfa_offset 8
  5176. call _libcuda_so_save_regs_and_resolve
  5177. addq $8, %rsp
  5178. .cfi_adjust_cfa_offset -8
  5179. jmp 1b
  5180. .cfi_endproc
  5181. /*
  5182. * Copyright 2018-2019 Yury Gribov
  5183. *
  5184. * The MIT License (MIT)
  5185. *
  5186. * Use of this source code is governed by MIT license that can be
  5187. * found in the LICENSE.txt file.
  5188. */
  5189. .globl cuInit
  5190. .p2align 4
  5191. .type cuInit, %function
  5192. cuInit:
  5193. .cfi_startproc
  5194. // Intel opt. manual says to
  5195. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5196. // to hint static predictor.
  5197. cmpq $0, _libcuda_so_tramp_table+1456(%rip)
  5198. je 2f
  5199. 1:
  5200. jmp *_libcuda_so_tramp_table+1456(%rip)
  5201. 2:
  5202. pushq $182
  5203. .cfi_adjust_cfa_offset 8
  5204. call _libcuda_so_save_regs_and_resolve
  5205. addq $8, %rsp
  5206. .cfi_adjust_cfa_offset -8
  5207. jmp 1b
  5208. .cfi_endproc
  5209. /*
  5210. * Copyright 2018-2019 Yury Gribov
  5211. *
  5212. * The MIT License (MIT)
  5213. *
  5214. * Use of this source code is governed by MIT license that can be
  5215. * found in the LICENSE.txt file.
  5216. */
  5217. .globl cuIpcCloseMemHandle
  5218. .p2align 4
  5219. .type cuIpcCloseMemHandle, %function
  5220. cuIpcCloseMemHandle:
  5221. .cfi_startproc
  5222. // Intel opt. manual says to
  5223. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5224. // to hint static predictor.
  5225. cmpq $0, _libcuda_so_tramp_table+1464(%rip)
  5226. je 2f
  5227. 1:
  5228. jmp *_libcuda_so_tramp_table+1464(%rip)
  5229. 2:
  5230. pushq $183
  5231. .cfi_adjust_cfa_offset 8
  5232. call _libcuda_so_save_regs_and_resolve
  5233. addq $8, %rsp
  5234. .cfi_adjust_cfa_offset -8
  5235. jmp 1b
  5236. .cfi_endproc
  5237. /*
  5238. * Copyright 2018-2019 Yury Gribov
  5239. *
  5240. * The MIT License (MIT)
  5241. *
  5242. * Use of this source code is governed by MIT license that can be
  5243. * found in the LICENSE.txt file.
  5244. */
  5245. .globl cuIpcGetEventHandle
  5246. .p2align 4
  5247. .type cuIpcGetEventHandle, %function
  5248. cuIpcGetEventHandle:
  5249. .cfi_startproc
  5250. // Intel opt. manual says to
  5251. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5252. // to hint static predictor.
  5253. cmpq $0, _libcuda_so_tramp_table+1472(%rip)
  5254. je 2f
  5255. 1:
  5256. jmp *_libcuda_so_tramp_table+1472(%rip)
  5257. 2:
  5258. pushq $184
  5259. .cfi_adjust_cfa_offset 8
  5260. call _libcuda_so_save_regs_and_resolve
  5261. addq $8, %rsp
  5262. .cfi_adjust_cfa_offset -8
  5263. jmp 1b
  5264. .cfi_endproc
  5265. /*
  5266. * Copyright 2018-2019 Yury Gribov
  5267. *
  5268. * The MIT License (MIT)
  5269. *
  5270. * Use of this source code is governed by MIT license that can be
  5271. * found in the LICENSE.txt file.
  5272. */
  5273. .globl cuIpcGetMemHandle
  5274. .p2align 4
  5275. .type cuIpcGetMemHandle, %function
  5276. cuIpcGetMemHandle:
  5277. .cfi_startproc
  5278. // Intel opt. manual says to
  5279. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5280. // to hint static predictor.
  5281. cmpq $0, _libcuda_so_tramp_table+1480(%rip)
  5282. je 2f
  5283. 1:
  5284. jmp *_libcuda_so_tramp_table+1480(%rip)
  5285. 2:
  5286. pushq $185
  5287. .cfi_adjust_cfa_offset 8
  5288. call _libcuda_so_save_regs_and_resolve
  5289. addq $8, %rsp
  5290. .cfi_adjust_cfa_offset -8
  5291. jmp 1b
  5292. .cfi_endproc
  5293. /*
  5294. * Copyright 2018-2019 Yury Gribov
  5295. *
  5296. * The MIT License (MIT)
  5297. *
  5298. * Use of this source code is governed by MIT license that can be
  5299. * found in the LICENSE.txt file.
  5300. */
  5301. .globl cuIpcOpenEventHandle
  5302. .p2align 4
  5303. .type cuIpcOpenEventHandle, %function
  5304. cuIpcOpenEventHandle:
  5305. .cfi_startproc
  5306. // Intel opt. manual says to
  5307. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5308. // to hint static predictor.
  5309. cmpq $0, _libcuda_so_tramp_table+1488(%rip)
  5310. je 2f
  5311. 1:
  5312. jmp *_libcuda_so_tramp_table+1488(%rip)
  5313. 2:
  5314. pushq $186
  5315. .cfi_adjust_cfa_offset 8
  5316. call _libcuda_so_save_regs_and_resolve
  5317. addq $8, %rsp
  5318. .cfi_adjust_cfa_offset -8
  5319. jmp 1b
  5320. .cfi_endproc
  5321. /*
  5322. * Copyright 2018-2019 Yury Gribov
  5323. *
  5324. * The MIT License (MIT)
  5325. *
  5326. * Use of this source code is governed by MIT license that can be
  5327. * found in the LICENSE.txt file.
  5328. */
  5329. .globl cuIpcOpenMemHandle
  5330. .p2align 4
  5331. .type cuIpcOpenMemHandle, %function
  5332. cuIpcOpenMemHandle:
  5333. .cfi_startproc
  5334. // Intel opt. manual says to
  5335. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5336. // to hint static predictor.
  5337. cmpq $0, _libcuda_so_tramp_table+1496(%rip)
  5338. je 2f
  5339. 1:
  5340. jmp *_libcuda_so_tramp_table+1496(%rip)
  5341. 2:
  5342. pushq $187
  5343. .cfi_adjust_cfa_offset 8
  5344. call _libcuda_so_save_regs_and_resolve
  5345. addq $8, %rsp
  5346. .cfi_adjust_cfa_offset -8
  5347. jmp 1b
  5348. .cfi_endproc
  5349. /*
  5350. * Copyright 2018-2019 Yury Gribov
  5351. *
  5352. * The MIT License (MIT)
  5353. *
  5354. * Use of this source code is governed by MIT license that can be
  5355. * found in the LICENSE.txt file.
  5356. */
  5357. .globl cuIpcOpenMemHandle_v2
  5358. .p2align 4
  5359. .type cuIpcOpenMemHandle_v2, %function
  5360. cuIpcOpenMemHandle_v2:
  5361. .cfi_startproc
  5362. // Intel opt. manual says to
  5363. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5364. // to hint static predictor.
  5365. cmpq $0, _libcuda_so_tramp_table+1504(%rip)
  5366. je 2f
  5367. 1:
  5368. jmp *_libcuda_so_tramp_table+1504(%rip)
  5369. 2:
  5370. pushq $188
  5371. .cfi_adjust_cfa_offset 8
  5372. call _libcuda_so_save_regs_and_resolve
  5373. addq $8, %rsp
  5374. .cfi_adjust_cfa_offset -8
  5375. jmp 1b
  5376. .cfi_endproc
  5377. /*
  5378. * Copyright 2018-2019 Yury Gribov
  5379. *
  5380. * The MIT License (MIT)
  5381. *
  5382. * Use of this source code is governed by MIT license that can be
  5383. * found in the LICENSE.txt file.
  5384. */
  5385. .globl cuLaunch
  5386. .p2align 4
  5387. .type cuLaunch, %function
  5388. cuLaunch:
  5389. .cfi_startproc
  5390. // Intel opt. manual says to
  5391. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5392. // to hint static predictor.
  5393. cmpq $0, _libcuda_so_tramp_table+1512(%rip)
  5394. je 2f
  5395. 1:
  5396. jmp *_libcuda_so_tramp_table+1512(%rip)
  5397. 2:
  5398. pushq $189
  5399. .cfi_adjust_cfa_offset 8
  5400. call _libcuda_so_save_regs_and_resolve
  5401. addq $8, %rsp
  5402. .cfi_adjust_cfa_offset -8
  5403. jmp 1b
  5404. .cfi_endproc
  5405. /*
  5406. * Copyright 2018-2019 Yury Gribov
  5407. *
  5408. * The MIT License (MIT)
  5409. *
  5410. * Use of this source code is governed by MIT license that can be
  5411. * found in the LICENSE.txt file.
  5412. */
  5413. .globl cuLaunchCooperativeKernel
  5414. .p2align 4
  5415. .type cuLaunchCooperativeKernel, %function
  5416. cuLaunchCooperativeKernel:
  5417. .cfi_startproc
  5418. // Intel opt. manual says to
  5419. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5420. // to hint static predictor.
  5421. cmpq $0, _libcuda_so_tramp_table+1520(%rip)
  5422. je 2f
  5423. 1:
  5424. jmp *_libcuda_so_tramp_table+1520(%rip)
  5425. 2:
  5426. pushq $190
  5427. .cfi_adjust_cfa_offset 8
  5428. call _libcuda_so_save_regs_and_resolve
  5429. addq $8, %rsp
  5430. .cfi_adjust_cfa_offset -8
  5431. jmp 1b
  5432. .cfi_endproc
  5433. /*
  5434. * Copyright 2018-2019 Yury Gribov
  5435. *
  5436. * The MIT License (MIT)
  5437. *
  5438. * Use of this source code is governed by MIT license that can be
  5439. * found in the LICENSE.txt file.
  5440. */
  5441. .globl cuLaunchCooperativeKernelMultiDevice
  5442. .p2align 4
  5443. .type cuLaunchCooperativeKernelMultiDevice, %function
  5444. cuLaunchCooperativeKernelMultiDevice:
  5445. .cfi_startproc
  5446. // Intel opt. manual says to
  5447. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5448. // to hint static predictor.
  5449. cmpq $0, _libcuda_so_tramp_table+1528(%rip)
  5450. je 2f
  5451. 1:
  5452. jmp *_libcuda_so_tramp_table+1528(%rip)
  5453. 2:
  5454. pushq $191
  5455. .cfi_adjust_cfa_offset 8
  5456. call _libcuda_so_save_regs_and_resolve
  5457. addq $8, %rsp
  5458. .cfi_adjust_cfa_offset -8
  5459. jmp 1b
  5460. .cfi_endproc
  5461. /*
  5462. * Copyright 2018-2019 Yury Gribov
  5463. *
  5464. * The MIT License (MIT)
  5465. *
  5466. * Use of this source code is governed by MIT license that can be
  5467. * found in the LICENSE.txt file.
  5468. */
  5469. .globl cuLaunchCooperativeKernel_ptsz
  5470. .p2align 4
  5471. .type cuLaunchCooperativeKernel_ptsz, %function
  5472. cuLaunchCooperativeKernel_ptsz:
  5473. .cfi_startproc
  5474. // Intel opt. manual says to
  5475. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5476. // to hint static predictor.
  5477. cmpq $0, _libcuda_so_tramp_table+1536(%rip)
  5478. je 2f
  5479. 1:
  5480. jmp *_libcuda_so_tramp_table+1536(%rip)
  5481. 2:
  5482. pushq $192
  5483. .cfi_adjust_cfa_offset 8
  5484. call _libcuda_so_save_regs_and_resolve
  5485. addq $8, %rsp
  5486. .cfi_adjust_cfa_offset -8
  5487. jmp 1b
  5488. .cfi_endproc
  5489. /*
  5490. * Copyright 2018-2019 Yury Gribov
  5491. *
  5492. * The MIT License (MIT)
  5493. *
  5494. * Use of this source code is governed by MIT license that can be
  5495. * found in the LICENSE.txt file.
  5496. */
  5497. .globl cuLaunchGrid
  5498. .p2align 4
  5499. .type cuLaunchGrid, %function
  5500. cuLaunchGrid:
  5501. .cfi_startproc
  5502. // Intel opt. manual says to
  5503. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5504. // to hint static predictor.
  5505. cmpq $0, _libcuda_so_tramp_table+1544(%rip)
  5506. je 2f
  5507. 1:
  5508. jmp *_libcuda_so_tramp_table+1544(%rip)
  5509. 2:
  5510. pushq $193
  5511. .cfi_adjust_cfa_offset 8
  5512. call _libcuda_so_save_regs_and_resolve
  5513. addq $8, %rsp
  5514. .cfi_adjust_cfa_offset -8
  5515. jmp 1b
  5516. .cfi_endproc
  5517. /*
  5518. * Copyright 2018-2019 Yury Gribov
  5519. *
  5520. * The MIT License (MIT)
  5521. *
  5522. * Use of this source code is governed by MIT license that can be
  5523. * found in the LICENSE.txt file.
  5524. */
  5525. .globl cuLaunchGridAsync
  5526. .p2align 4
  5527. .type cuLaunchGridAsync, %function
  5528. cuLaunchGridAsync:
  5529. .cfi_startproc
  5530. // Intel opt. manual says to
  5531. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5532. // to hint static predictor.
  5533. cmpq $0, _libcuda_so_tramp_table+1552(%rip)
  5534. je 2f
  5535. 1:
  5536. jmp *_libcuda_so_tramp_table+1552(%rip)
  5537. 2:
  5538. pushq $194
  5539. .cfi_adjust_cfa_offset 8
  5540. call _libcuda_so_save_regs_and_resolve
  5541. addq $8, %rsp
  5542. .cfi_adjust_cfa_offset -8
  5543. jmp 1b
  5544. .cfi_endproc
  5545. /*
  5546. * Copyright 2018-2019 Yury Gribov
  5547. *
  5548. * The MIT License (MIT)
  5549. *
  5550. * Use of this source code is governed by MIT license that can be
  5551. * found in the LICENSE.txt file.
  5552. */
  5553. .globl cuLaunchHostFunc
  5554. .p2align 4
  5555. .type cuLaunchHostFunc, %function
  5556. cuLaunchHostFunc:
  5557. .cfi_startproc
  5558. // Intel opt. manual says to
  5559. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5560. // to hint static predictor.
  5561. cmpq $0, _libcuda_so_tramp_table+1560(%rip)
  5562. je 2f
  5563. 1:
  5564. jmp *_libcuda_so_tramp_table+1560(%rip)
  5565. 2:
  5566. pushq $195
  5567. .cfi_adjust_cfa_offset 8
  5568. call _libcuda_so_save_regs_and_resolve
  5569. addq $8, %rsp
  5570. .cfi_adjust_cfa_offset -8
  5571. jmp 1b
  5572. .cfi_endproc
  5573. /*
  5574. * Copyright 2018-2019 Yury Gribov
  5575. *
  5576. * The MIT License (MIT)
  5577. *
  5578. * Use of this source code is governed by MIT license that can be
  5579. * found in the LICENSE.txt file.
  5580. */
  5581. .globl cuLaunchHostFunc_ptsz
  5582. .p2align 4
  5583. .type cuLaunchHostFunc_ptsz, %function
  5584. cuLaunchHostFunc_ptsz:
  5585. .cfi_startproc
  5586. // Intel opt. manual says to
  5587. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5588. // to hint static predictor.
  5589. cmpq $0, _libcuda_so_tramp_table+1568(%rip)
  5590. je 2f
  5591. 1:
  5592. jmp *_libcuda_so_tramp_table+1568(%rip)
  5593. 2:
  5594. pushq $196
  5595. .cfi_adjust_cfa_offset 8
  5596. call _libcuda_so_save_regs_and_resolve
  5597. addq $8, %rsp
  5598. .cfi_adjust_cfa_offset -8
  5599. jmp 1b
  5600. .cfi_endproc
  5601. /*
  5602. * Copyright 2018-2019 Yury Gribov
  5603. *
  5604. * The MIT License (MIT)
  5605. *
  5606. * Use of this source code is governed by MIT license that can be
  5607. * found in the LICENSE.txt file.
  5608. */
  5609. .globl cuLaunchKernel
  5610. .p2align 4
  5611. .type cuLaunchKernel, %function
  5612. cuLaunchKernel:
  5613. .cfi_startproc
  5614. // Intel opt. manual says to
  5615. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5616. // to hint static predictor.
  5617. cmpq $0, _libcuda_so_tramp_table+1576(%rip)
  5618. je 2f
  5619. 1:
  5620. jmp *_libcuda_so_tramp_table+1576(%rip)
  5621. 2:
  5622. pushq $197
  5623. .cfi_adjust_cfa_offset 8
  5624. call _libcuda_so_save_regs_and_resolve
  5625. addq $8, %rsp
  5626. .cfi_adjust_cfa_offset -8
  5627. jmp 1b
  5628. .cfi_endproc
  5629. /*
  5630. * Copyright 2018-2019 Yury Gribov
  5631. *
  5632. * The MIT License (MIT)
  5633. *
  5634. * Use of this source code is governed by MIT license that can be
  5635. * found in the LICENSE.txt file.
  5636. */
  5637. .globl cuLaunchKernel_ptsz
  5638. .p2align 4
  5639. .type cuLaunchKernel_ptsz, %function
  5640. cuLaunchKernel_ptsz:
  5641. .cfi_startproc
  5642. // Intel opt. manual says to
  5643. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5644. // to hint static predictor.
  5645. cmpq $0, _libcuda_so_tramp_table+1584(%rip)
  5646. je 2f
  5647. 1:
  5648. jmp *_libcuda_so_tramp_table+1584(%rip)
  5649. 2:
  5650. pushq $198
  5651. .cfi_adjust_cfa_offset 8
  5652. call _libcuda_so_save_regs_and_resolve
  5653. addq $8, %rsp
  5654. .cfi_adjust_cfa_offset -8
  5655. jmp 1b
  5656. .cfi_endproc
  5657. /*
  5658. * Copyright 2018-2019 Yury Gribov
  5659. *
  5660. * The MIT License (MIT)
  5661. *
  5662. * Use of this source code is governed by MIT license that can be
  5663. * found in the LICENSE.txt file.
  5664. */
  5665. .globl cuLinkAddData
  5666. .p2align 4
  5667. .type cuLinkAddData, %function
  5668. cuLinkAddData:
  5669. .cfi_startproc
  5670. // Intel opt. manual says to
  5671. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5672. // to hint static predictor.
  5673. cmpq $0, _libcuda_so_tramp_table+1592(%rip)
  5674. je 2f
  5675. 1:
  5676. jmp *_libcuda_so_tramp_table+1592(%rip)
  5677. 2:
  5678. pushq $199
  5679. .cfi_adjust_cfa_offset 8
  5680. call _libcuda_so_save_regs_and_resolve
  5681. addq $8, %rsp
  5682. .cfi_adjust_cfa_offset -8
  5683. jmp 1b
  5684. .cfi_endproc
  5685. /*
  5686. * Copyright 2018-2019 Yury Gribov
  5687. *
  5688. * The MIT License (MIT)
  5689. *
  5690. * Use of this source code is governed by MIT license that can be
  5691. * found in the LICENSE.txt file.
  5692. */
  5693. .globl cuLinkAddData_v2
  5694. .p2align 4
  5695. .type cuLinkAddData_v2, %function
  5696. cuLinkAddData_v2:
  5697. .cfi_startproc
  5698. // Intel opt. manual says to
  5699. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5700. // to hint static predictor.
  5701. cmpq $0, _libcuda_so_tramp_table+1600(%rip)
  5702. je 2f
  5703. 1:
  5704. jmp *_libcuda_so_tramp_table+1600(%rip)
  5705. 2:
  5706. pushq $200
  5707. .cfi_adjust_cfa_offset 8
  5708. call _libcuda_so_save_regs_and_resolve
  5709. addq $8, %rsp
  5710. .cfi_adjust_cfa_offset -8
  5711. jmp 1b
  5712. .cfi_endproc
  5713. /*
  5714. * Copyright 2018-2019 Yury Gribov
  5715. *
  5716. * The MIT License (MIT)
  5717. *
  5718. * Use of this source code is governed by MIT license that can be
  5719. * found in the LICENSE.txt file.
  5720. */
  5721. .globl cuLinkAddFile
  5722. .p2align 4
  5723. .type cuLinkAddFile, %function
  5724. cuLinkAddFile:
  5725. .cfi_startproc
  5726. // Intel opt. manual says to
  5727. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5728. // to hint static predictor.
  5729. cmpq $0, _libcuda_so_tramp_table+1608(%rip)
  5730. je 2f
  5731. 1:
  5732. jmp *_libcuda_so_tramp_table+1608(%rip)
  5733. 2:
  5734. pushq $201
  5735. .cfi_adjust_cfa_offset 8
  5736. call _libcuda_so_save_regs_and_resolve
  5737. addq $8, %rsp
  5738. .cfi_adjust_cfa_offset -8
  5739. jmp 1b
  5740. .cfi_endproc
  5741. /*
  5742. * Copyright 2018-2019 Yury Gribov
  5743. *
  5744. * The MIT License (MIT)
  5745. *
  5746. * Use of this source code is governed by MIT license that can be
  5747. * found in the LICENSE.txt file.
  5748. */
  5749. .globl cuLinkAddFile_v2
  5750. .p2align 4
  5751. .type cuLinkAddFile_v2, %function
  5752. cuLinkAddFile_v2:
  5753. .cfi_startproc
  5754. // Intel opt. manual says to
  5755. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5756. // to hint static predictor.
  5757. cmpq $0, _libcuda_so_tramp_table+1616(%rip)
  5758. je 2f
  5759. 1:
  5760. jmp *_libcuda_so_tramp_table+1616(%rip)
  5761. 2:
  5762. pushq $202
  5763. .cfi_adjust_cfa_offset 8
  5764. call _libcuda_so_save_regs_and_resolve
  5765. addq $8, %rsp
  5766. .cfi_adjust_cfa_offset -8
  5767. jmp 1b
  5768. .cfi_endproc
  5769. /*
  5770. * Copyright 2018-2019 Yury Gribov
  5771. *
  5772. * The MIT License (MIT)
  5773. *
  5774. * Use of this source code is governed by MIT license that can be
  5775. * found in the LICENSE.txt file.
  5776. */
  5777. .globl cuLinkComplete
  5778. .p2align 4
  5779. .type cuLinkComplete, %function
  5780. cuLinkComplete:
  5781. .cfi_startproc
  5782. // Intel opt. manual says to
  5783. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5784. // to hint static predictor.
  5785. cmpq $0, _libcuda_so_tramp_table+1624(%rip)
  5786. je 2f
  5787. 1:
  5788. jmp *_libcuda_so_tramp_table+1624(%rip)
  5789. 2:
  5790. pushq $203
  5791. .cfi_adjust_cfa_offset 8
  5792. call _libcuda_so_save_regs_and_resolve
  5793. addq $8, %rsp
  5794. .cfi_adjust_cfa_offset -8
  5795. jmp 1b
  5796. .cfi_endproc
  5797. /*
  5798. * Copyright 2018-2019 Yury Gribov
  5799. *
  5800. * The MIT License (MIT)
  5801. *
  5802. * Use of this source code is governed by MIT license that can be
  5803. * found in the LICENSE.txt file.
  5804. */
  5805. .globl cuLinkCreate
  5806. .p2align 4
  5807. .type cuLinkCreate, %function
  5808. cuLinkCreate:
  5809. .cfi_startproc
  5810. // Intel opt. manual says to
  5811. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5812. // to hint static predictor.
  5813. cmpq $0, _libcuda_so_tramp_table+1632(%rip)
  5814. je 2f
  5815. 1:
  5816. jmp *_libcuda_so_tramp_table+1632(%rip)
  5817. 2:
  5818. pushq $204
  5819. .cfi_adjust_cfa_offset 8
  5820. call _libcuda_so_save_regs_and_resolve
  5821. addq $8, %rsp
  5822. .cfi_adjust_cfa_offset -8
  5823. jmp 1b
  5824. .cfi_endproc
  5825. /*
  5826. * Copyright 2018-2019 Yury Gribov
  5827. *
  5828. * The MIT License (MIT)
  5829. *
  5830. * Use of this source code is governed by MIT license that can be
  5831. * found in the LICENSE.txt file.
  5832. */
  5833. .globl cuLinkCreate_v2
  5834. .p2align 4
  5835. .type cuLinkCreate_v2, %function
  5836. cuLinkCreate_v2:
  5837. .cfi_startproc
  5838. // Intel opt. manual says to
  5839. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5840. // to hint static predictor.
  5841. cmpq $0, _libcuda_so_tramp_table+1640(%rip)
  5842. je 2f
  5843. 1:
  5844. jmp *_libcuda_so_tramp_table+1640(%rip)
  5845. 2:
  5846. pushq $205
  5847. .cfi_adjust_cfa_offset 8
  5848. call _libcuda_so_save_regs_and_resolve
  5849. addq $8, %rsp
  5850. .cfi_adjust_cfa_offset -8
  5851. jmp 1b
  5852. .cfi_endproc
  5853. /*
  5854. * Copyright 2018-2019 Yury Gribov
  5855. *
  5856. * The MIT License (MIT)
  5857. *
  5858. * Use of this source code is governed by MIT license that can be
  5859. * found in the LICENSE.txt file.
  5860. */
  5861. .globl cuLinkDestroy
  5862. .p2align 4
  5863. .type cuLinkDestroy, %function
  5864. cuLinkDestroy:
  5865. .cfi_startproc
  5866. // Intel opt. manual says to
  5867. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5868. // to hint static predictor.
  5869. cmpq $0, _libcuda_so_tramp_table+1648(%rip)
  5870. je 2f
  5871. 1:
  5872. jmp *_libcuda_so_tramp_table+1648(%rip)
  5873. 2:
  5874. pushq $206
  5875. .cfi_adjust_cfa_offset 8
  5876. call _libcuda_so_save_regs_and_resolve
  5877. addq $8, %rsp
  5878. .cfi_adjust_cfa_offset -8
  5879. jmp 1b
  5880. .cfi_endproc
  5881. /*
  5882. * Copyright 2018-2019 Yury Gribov
  5883. *
  5884. * The MIT License (MIT)
  5885. *
  5886. * Use of this source code is governed by MIT license that can be
  5887. * found in the LICENSE.txt file.
  5888. */
  5889. .globl cuMemAddressFree
  5890. .p2align 4
  5891. .type cuMemAddressFree, %function
  5892. cuMemAddressFree:
  5893. .cfi_startproc
  5894. // Intel opt. manual says to
  5895. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5896. // to hint static predictor.
  5897. cmpq $0, _libcuda_so_tramp_table+1656(%rip)
  5898. je 2f
  5899. 1:
  5900. jmp *_libcuda_so_tramp_table+1656(%rip)
  5901. 2:
  5902. pushq $207
  5903. .cfi_adjust_cfa_offset 8
  5904. call _libcuda_so_save_regs_and_resolve
  5905. addq $8, %rsp
  5906. .cfi_adjust_cfa_offset -8
  5907. jmp 1b
  5908. .cfi_endproc
  5909. /*
  5910. * Copyright 2018-2019 Yury Gribov
  5911. *
  5912. * The MIT License (MIT)
  5913. *
  5914. * Use of this source code is governed by MIT license that can be
  5915. * found in the LICENSE.txt file.
  5916. */
  5917. .globl cuMemAddressReserve
  5918. .p2align 4
  5919. .type cuMemAddressReserve, %function
  5920. cuMemAddressReserve:
  5921. .cfi_startproc
  5922. // Intel opt. manual says to
  5923. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5924. // to hint static predictor.
  5925. cmpq $0, _libcuda_so_tramp_table+1664(%rip)
  5926. je 2f
  5927. 1:
  5928. jmp *_libcuda_so_tramp_table+1664(%rip)
  5929. 2:
  5930. pushq $208
  5931. .cfi_adjust_cfa_offset 8
  5932. call _libcuda_so_save_regs_and_resolve
  5933. addq $8, %rsp
  5934. .cfi_adjust_cfa_offset -8
  5935. jmp 1b
  5936. .cfi_endproc
  5937. /*
  5938. * Copyright 2018-2019 Yury Gribov
  5939. *
  5940. * The MIT License (MIT)
  5941. *
  5942. * Use of this source code is governed by MIT license that can be
  5943. * found in the LICENSE.txt file.
  5944. */
  5945. .globl cuMemAdvise
  5946. .p2align 4
  5947. .type cuMemAdvise, %function
  5948. cuMemAdvise:
  5949. .cfi_startproc
  5950. // Intel opt. manual says to
  5951. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5952. // to hint static predictor.
  5953. cmpq $0, _libcuda_so_tramp_table+1672(%rip)
  5954. je 2f
  5955. 1:
  5956. jmp *_libcuda_so_tramp_table+1672(%rip)
  5957. 2:
  5958. pushq $209
  5959. .cfi_adjust_cfa_offset 8
  5960. call _libcuda_so_save_regs_and_resolve
  5961. addq $8, %rsp
  5962. .cfi_adjust_cfa_offset -8
  5963. jmp 1b
  5964. .cfi_endproc
  5965. /*
  5966. * Copyright 2018-2019 Yury Gribov
  5967. *
  5968. * The MIT License (MIT)
  5969. *
  5970. * Use of this source code is governed by MIT license that can be
  5971. * found in the LICENSE.txt file.
  5972. */
  5973. .globl cuMemAlloc
  5974. .p2align 4
  5975. .type cuMemAlloc, %function
  5976. cuMemAlloc:
  5977. .cfi_startproc
  5978. // Intel opt. manual says to
  5979. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  5980. // to hint static predictor.
  5981. cmpq $0, _libcuda_so_tramp_table+1680(%rip)
  5982. je 2f
  5983. 1:
  5984. jmp *_libcuda_so_tramp_table+1680(%rip)
  5985. 2:
  5986. pushq $210
  5987. .cfi_adjust_cfa_offset 8
  5988. call _libcuda_so_save_regs_and_resolve
  5989. addq $8, %rsp
  5990. .cfi_adjust_cfa_offset -8
  5991. jmp 1b
  5992. .cfi_endproc
  5993. /*
  5994. * Copyright 2018-2019 Yury Gribov
  5995. *
  5996. * The MIT License (MIT)
  5997. *
  5998. * Use of this source code is governed by MIT license that can be
  5999. * found in the LICENSE.txt file.
  6000. */
  6001. .globl cuMemAllocHost
  6002. .p2align 4
  6003. .type cuMemAllocHost, %function
  6004. cuMemAllocHost:
  6005. .cfi_startproc
  6006. // Intel opt. manual says to
  6007. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6008. // to hint static predictor.
  6009. cmpq $0, _libcuda_so_tramp_table+1688(%rip)
  6010. je 2f
  6011. 1:
  6012. jmp *_libcuda_so_tramp_table+1688(%rip)
  6013. 2:
  6014. pushq $211
  6015. .cfi_adjust_cfa_offset 8
  6016. call _libcuda_so_save_regs_and_resolve
  6017. addq $8, %rsp
  6018. .cfi_adjust_cfa_offset -8
  6019. jmp 1b
  6020. .cfi_endproc
  6021. /*
  6022. * Copyright 2018-2019 Yury Gribov
  6023. *
  6024. * The MIT License (MIT)
  6025. *
  6026. * Use of this source code is governed by MIT license that can be
  6027. * found in the LICENSE.txt file.
  6028. */
  6029. .globl cuMemAllocHost_v2
  6030. .p2align 4
  6031. .type cuMemAllocHost_v2, %function
  6032. cuMemAllocHost_v2:
  6033. .cfi_startproc
  6034. // Intel opt. manual says to
  6035. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6036. // to hint static predictor.
  6037. cmpq $0, _libcuda_so_tramp_table+1696(%rip)
  6038. je 2f
  6039. 1:
  6040. jmp *_libcuda_so_tramp_table+1696(%rip)
  6041. 2:
  6042. pushq $212
  6043. .cfi_adjust_cfa_offset 8
  6044. call _libcuda_so_save_regs_and_resolve
  6045. addq $8, %rsp
  6046. .cfi_adjust_cfa_offset -8
  6047. jmp 1b
  6048. .cfi_endproc
  6049. /*
  6050. * Copyright 2018-2019 Yury Gribov
  6051. *
  6052. * The MIT License (MIT)
  6053. *
  6054. * Use of this source code is governed by MIT license that can be
  6055. * found in the LICENSE.txt file.
  6056. */
  6057. .globl cuMemAllocManaged
  6058. .p2align 4
  6059. .type cuMemAllocManaged, %function
  6060. cuMemAllocManaged:
  6061. .cfi_startproc
  6062. // Intel opt. manual says to
  6063. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6064. // to hint static predictor.
  6065. cmpq $0, _libcuda_so_tramp_table+1704(%rip)
  6066. je 2f
  6067. 1:
  6068. jmp *_libcuda_so_tramp_table+1704(%rip)
  6069. 2:
  6070. pushq $213
  6071. .cfi_adjust_cfa_offset 8
  6072. call _libcuda_so_save_regs_and_resolve
  6073. addq $8, %rsp
  6074. .cfi_adjust_cfa_offset -8
  6075. jmp 1b
  6076. .cfi_endproc
  6077. /*
  6078. * Copyright 2018-2019 Yury Gribov
  6079. *
  6080. * The MIT License (MIT)
  6081. *
  6082. * Use of this source code is governed by MIT license that can be
  6083. * found in the LICENSE.txt file.
  6084. */
  6085. .globl cuMemAllocPitch
  6086. .p2align 4
  6087. .type cuMemAllocPitch, %function
  6088. cuMemAllocPitch:
  6089. .cfi_startproc
  6090. // Intel opt. manual says to
  6091. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6092. // to hint static predictor.
  6093. cmpq $0, _libcuda_so_tramp_table+1712(%rip)
  6094. je 2f
  6095. 1:
  6096. jmp *_libcuda_so_tramp_table+1712(%rip)
  6097. 2:
  6098. pushq $214
  6099. .cfi_adjust_cfa_offset 8
  6100. call _libcuda_so_save_regs_and_resolve
  6101. addq $8, %rsp
  6102. .cfi_adjust_cfa_offset -8
  6103. jmp 1b
  6104. .cfi_endproc
  6105. /*
  6106. * Copyright 2018-2019 Yury Gribov
  6107. *
  6108. * The MIT License (MIT)
  6109. *
  6110. * Use of this source code is governed by MIT license that can be
  6111. * found in the LICENSE.txt file.
  6112. */
  6113. .globl cuMemAllocPitch_v2
  6114. .p2align 4
  6115. .type cuMemAllocPitch_v2, %function
  6116. cuMemAllocPitch_v2:
  6117. .cfi_startproc
  6118. // Intel opt. manual says to
  6119. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6120. // to hint static predictor.
  6121. cmpq $0, _libcuda_so_tramp_table+1720(%rip)
  6122. je 2f
  6123. 1:
  6124. jmp *_libcuda_so_tramp_table+1720(%rip)
  6125. 2:
  6126. pushq $215
  6127. .cfi_adjust_cfa_offset 8
  6128. call _libcuda_so_save_regs_and_resolve
  6129. addq $8, %rsp
  6130. .cfi_adjust_cfa_offset -8
  6131. jmp 1b
  6132. .cfi_endproc
  6133. /*
  6134. * Copyright 2018-2019 Yury Gribov
  6135. *
  6136. * The MIT License (MIT)
  6137. *
  6138. * Use of this source code is governed by MIT license that can be
  6139. * found in the LICENSE.txt file.
  6140. */
  6141. .globl cuMemAlloc_v2
  6142. .p2align 4
  6143. .type cuMemAlloc_v2, %function
  6144. cuMemAlloc_v2:
  6145. .cfi_startproc
  6146. // Intel opt. manual says to
  6147. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6148. // to hint static predictor.
  6149. cmpq $0, _libcuda_so_tramp_table+1728(%rip)
  6150. je 2f
  6151. 1:
  6152. jmp *_libcuda_so_tramp_table+1728(%rip)
  6153. 2:
  6154. pushq $216
  6155. .cfi_adjust_cfa_offset 8
  6156. call _libcuda_so_save_regs_and_resolve
  6157. addq $8, %rsp
  6158. .cfi_adjust_cfa_offset -8
  6159. jmp 1b
  6160. .cfi_endproc
  6161. /*
  6162. * Copyright 2018-2019 Yury Gribov
  6163. *
  6164. * The MIT License (MIT)
  6165. *
  6166. * Use of this source code is governed by MIT license that can be
  6167. * found in the LICENSE.txt file.
  6168. */
  6169. .globl cuMemCreate
  6170. .p2align 4
  6171. .type cuMemCreate, %function
  6172. cuMemCreate:
  6173. .cfi_startproc
  6174. // Intel opt. manual says to
  6175. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6176. // to hint static predictor.
  6177. cmpq $0, _libcuda_so_tramp_table+1736(%rip)
  6178. je 2f
  6179. 1:
  6180. jmp *_libcuda_so_tramp_table+1736(%rip)
  6181. 2:
  6182. pushq $217
  6183. .cfi_adjust_cfa_offset 8
  6184. call _libcuda_so_save_regs_and_resolve
  6185. addq $8, %rsp
  6186. .cfi_adjust_cfa_offset -8
  6187. jmp 1b
  6188. .cfi_endproc
  6189. /*
  6190. * Copyright 2018-2019 Yury Gribov
  6191. *
  6192. * The MIT License (MIT)
  6193. *
  6194. * Use of this source code is governed by MIT license that can be
  6195. * found in the LICENSE.txt file.
  6196. */
  6197. .globl cuMemExportToShareableHandle
  6198. .p2align 4
  6199. .type cuMemExportToShareableHandle, %function
  6200. cuMemExportToShareableHandle:
  6201. .cfi_startproc
  6202. // Intel opt. manual says to
  6203. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6204. // to hint static predictor.
  6205. cmpq $0, _libcuda_so_tramp_table+1744(%rip)
  6206. je 2f
  6207. 1:
  6208. jmp *_libcuda_so_tramp_table+1744(%rip)
  6209. 2:
  6210. pushq $218
  6211. .cfi_adjust_cfa_offset 8
  6212. call _libcuda_so_save_regs_and_resolve
  6213. addq $8, %rsp
  6214. .cfi_adjust_cfa_offset -8
  6215. jmp 1b
  6216. .cfi_endproc
  6217. /*
  6218. * Copyright 2018-2019 Yury Gribov
  6219. *
  6220. * The MIT License (MIT)
  6221. *
  6222. * Use of this source code is governed by MIT license that can be
  6223. * found in the LICENSE.txt file.
  6224. */
  6225. .globl cuMemFree
  6226. .p2align 4
  6227. .type cuMemFree, %function
  6228. cuMemFree:
  6229. .cfi_startproc
  6230. // Intel opt. manual says to
  6231. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6232. // to hint static predictor.
  6233. cmpq $0, _libcuda_so_tramp_table+1752(%rip)
  6234. je 2f
  6235. 1:
  6236. jmp *_libcuda_so_tramp_table+1752(%rip)
  6237. 2:
  6238. pushq $219
  6239. .cfi_adjust_cfa_offset 8
  6240. call _libcuda_so_save_regs_and_resolve
  6241. addq $8, %rsp
  6242. .cfi_adjust_cfa_offset -8
  6243. jmp 1b
  6244. .cfi_endproc
  6245. /*
  6246. * Copyright 2018-2019 Yury Gribov
  6247. *
  6248. * The MIT License (MIT)
  6249. *
  6250. * Use of this source code is governed by MIT license that can be
  6251. * found in the LICENSE.txt file.
  6252. */
  6253. .globl cuMemFreeHost
  6254. .p2align 4
  6255. .type cuMemFreeHost, %function
  6256. cuMemFreeHost:
  6257. .cfi_startproc
  6258. // Intel opt. manual says to
  6259. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6260. // to hint static predictor.
  6261. cmpq $0, _libcuda_so_tramp_table+1760(%rip)
  6262. je 2f
  6263. 1:
  6264. jmp *_libcuda_so_tramp_table+1760(%rip)
  6265. 2:
  6266. pushq $220
  6267. .cfi_adjust_cfa_offset 8
  6268. call _libcuda_so_save_regs_and_resolve
  6269. addq $8, %rsp
  6270. .cfi_adjust_cfa_offset -8
  6271. jmp 1b
  6272. .cfi_endproc
  6273. /*
  6274. * Copyright 2018-2019 Yury Gribov
  6275. *
  6276. * The MIT License (MIT)
  6277. *
  6278. * Use of this source code is governed by MIT license that can be
  6279. * found in the LICENSE.txt file.
  6280. */
  6281. .globl cuMemFree_v2
  6282. .p2align 4
  6283. .type cuMemFree_v2, %function
  6284. cuMemFree_v2:
  6285. .cfi_startproc
  6286. // Intel opt. manual says to
  6287. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6288. // to hint static predictor.
  6289. cmpq $0, _libcuda_so_tramp_table+1768(%rip)
  6290. je 2f
  6291. 1:
  6292. jmp *_libcuda_so_tramp_table+1768(%rip)
  6293. 2:
  6294. pushq $221
  6295. .cfi_adjust_cfa_offset 8
  6296. call _libcuda_so_save_regs_and_resolve
  6297. addq $8, %rsp
  6298. .cfi_adjust_cfa_offset -8
  6299. jmp 1b
  6300. .cfi_endproc
  6301. /*
  6302. * Copyright 2018-2019 Yury Gribov
  6303. *
  6304. * The MIT License (MIT)
  6305. *
  6306. * Use of this source code is governed by MIT license that can be
  6307. * found in the LICENSE.txt file.
  6308. */
  6309. .globl cuMemGetAccess
  6310. .p2align 4
  6311. .type cuMemGetAccess, %function
  6312. cuMemGetAccess:
  6313. .cfi_startproc
  6314. // Intel opt. manual says to
  6315. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6316. // to hint static predictor.
  6317. cmpq $0, _libcuda_so_tramp_table+1776(%rip)
  6318. je 2f
  6319. 1:
  6320. jmp *_libcuda_so_tramp_table+1776(%rip)
  6321. 2:
  6322. pushq $222
  6323. .cfi_adjust_cfa_offset 8
  6324. call _libcuda_so_save_regs_and_resolve
  6325. addq $8, %rsp
  6326. .cfi_adjust_cfa_offset -8
  6327. jmp 1b
  6328. .cfi_endproc
  6329. /*
  6330. * Copyright 2018-2019 Yury Gribov
  6331. *
  6332. * The MIT License (MIT)
  6333. *
  6334. * Use of this source code is governed by MIT license that can be
  6335. * found in the LICENSE.txt file.
  6336. */
  6337. .globl cuMemGetAddressRange
  6338. .p2align 4
  6339. .type cuMemGetAddressRange, %function
  6340. cuMemGetAddressRange:
  6341. .cfi_startproc
  6342. // Intel opt. manual says to
  6343. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6344. // to hint static predictor.
  6345. cmpq $0, _libcuda_so_tramp_table+1784(%rip)
  6346. je 2f
  6347. 1:
  6348. jmp *_libcuda_so_tramp_table+1784(%rip)
  6349. 2:
  6350. pushq $223
  6351. .cfi_adjust_cfa_offset 8
  6352. call _libcuda_so_save_regs_and_resolve
  6353. addq $8, %rsp
  6354. .cfi_adjust_cfa_offset -8
  6355. jmp 1b
  6356. .cfi_endproc
  6357. /*
  6358. * Copyright 2018-2019 Yury Gribov
  6359. *
  6360. * The MIT License (MIT)
  6361. *
  6362. * Use of this source code is governed by MIT license that can be
  6363. * found in the LICENSE.txt file.
  6364. */
  6365. .globl cuMemGetAddressRange_v2
  6366. .p2align 4
  6367. .type cuMemGetAddressRange_v2, %function
  6368. cuMemGetAddressRange_v2:
  6369. .cfi_startproc
  6370. // Intel opt. manual says to
  6371. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6372. // to hint static predictor.
  6373. cmpq $0, _libcuda_so_tramp_table+1792(%rip)
  6374. je 2f
  6375. 1:
  6376. jmp *_libcuda_so_tramp_table+1792(%rip)
  6377. 2:
  6378. pushq $224
  6379. .cfi_adjust_cfa_offset 8
  6380. call _libcuda_so_save_regs_and_resolve
  6381. addq $8, %rsp
  6382. .cfi_adjust_cfa_offset -8
  6383. jmp 1b
  6384. .cfi_endproc
  6385. /*
  6386. * Copyright 2018-2019 Yury Gribov
  6387. *
  6388. * The MIT License (MIT)
  6389. *
  6390. * Use of this source code is governed by MIT license that can be
  6391. * found in the LICENSE.txt file.
  6392. */
  6393. .globl cuMemGetAllocationGranularity
  6394. .p2align 4
  6395. .type cuMemGetAllocationGranularity, %function
  6396. cuMemGetAllocationGranularity:
  6397. .cfi_startproc
  6398. // Intel opt. manual says to
  6399. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6400. // to hint static predictor.
  6401. cmpq $0, _libcuda_so_tramp_table+1800(%rip)
  6402. je 2f
  6403. 1:
  6404. jmp *_libcuda_so_tramp_table+1800(%rip)
  6405. 2:
  6406. pushq $225
  6407. .cfi_adjust_cfa_offset 8
  6408. call _libcuda_so_save_regs_and_resolve
  6409. addq $8, %rsp
  6410. .cfi_adjust_cfa_offset -8
  6411. jmp 1b
  6412. .cfi_endproc
  6413. /*
  6414. * Copyright 2018-2019 Yury Gribov
  6415. *
  6416. * The MIT License (MIT)
  6417. *
  6418. * Use of this source code is governed by MIT license that can be
  6419. * found in the LICENSE.txt file.
  6420. */
  6421. .globl cuMemGetAllocationPropertiesFromHandle
  6422. .p2align 4
  6423. .type cuMemGetAllocationPropertiesFromHandle, %function
  6424. cuMemGetAllocationPropertiesFromHandle:
  6425. .cfi_startproc
  6426. // Intel opt. manual says to
  6427. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6428. // to hint static predictor.
  6429. cmpq $0, _libcuda_so_tramp_table+1808(%rip)
  6430. je 2f
  6431. 1:
  6432. jmp *_libcuda_so_tramp_table+1808(%rip)
  6433. 2:
  6434. pushq $226
  6435. .cfi_adjust_cfa_offset 8
  6436. call _libcuda_so_save_regs_and_resolve
  6437. addq $8, %rsp
  6438. .cfi_adjust_cfa_offset -8
  6439. jmp 1b
  6440. .cfi_endproc
  6441. /*
  6442. * Copyright 2018-2019 Yury Gribov
  6443. *
  6444. * The MIT License (MIT)
  6445. *
  6446. * Use of this source code is governed by MIT license that can be
  6447. * found in the LICENSE.txt file.
  6448. */
  6449. .globl cuMemGetAttribute
  6450. .p2align 4
  6451. .type cuMemGetAttribute, %function
  6452. cuMemGetAttribute:
  6453. .cfi_startproc
  6454. // Intel opt. manual says to
  6455. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6456. // to hint static predictor.
  6457. cmpq $0, _libcuda_so_tramp_table+1816(%rip)
  6458. je 2f
  6459. 1:
  6460. jmp *_libcuda_so_tramp_table+1816(%rip)
  6461. 2:
  6462. pushq $227
  6463. .cfi_adjust_cfa_offset 8
  6464. call _libcuda_so_save_regs_and_resolve
  6465. addq $8, %rsp
  6466. .cfi_adjust_cfa_offset -8
  6467. jmp 1b
  6468. .cfi_endproc
  6469. /*
  6470. * Copyright 2018-2019 Yury Gribov
  6471. *
  6472. * The MIT License (MIT)
  6473. *
  6474. * Use of this source code is governed by MIT license that can be
  6475. * found in the LICENSE.txt file.
  6476. */
  6477. .globl cuMemGetAttribute_v2
  6478. .p2align 4
  6479. .type cuMemGetAttribute_v2, %function
  6480. cuMemGetAttribute_v2:
  6481. .cfi_startproc
  6482. // Intel opt. manual says to
  6483. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6484. // to hint static predictor.
  6485. cmpq $0, _libcuda_so_tramp_table+1824(%rip)
  6486. je 2f
  6487. 1:
  6488. jmp *_libcuda_so_tramp_table+1824(%rip)
  6489. 2:
  6490. pushq $228
  6491. .cfi_adjust_cfa_offset 8
  6492. call _libcuda_so_save_regs_and_resolve
  6493. addq $8, %rsp
  6494. .cfi_adjust_cfa_offset -8
  6495. jmp 1b
  6496. .cfi_endproc
  6497. /*
  6498. * Copyright 2018-2019 Yury Gribov
  6499. *
  6500. * The MIT License (MIT)
  6501. *
  6502. * Use of this source code is governed by MIT license that can be
  6503. * found in the LICENSE.txt file.
  6504. */
  6505. .globl cuMemGetInfo
  6506. .p2align 4
  6507. .type cuMemGetInfo, %function
  6508. cuMemGetInfo:
  6509. .cfi_startproc
  6510. // Intel opt. manual says to
  6511. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6512. // to hint static predictor.
  6513. cmpq $0, _libcuda_so_tramp_table+1832(%rip)
  6514. je 2f
  6515. 1:
  6516. jmp *_libcuda_so_tramp_table+1832(%rip)
  6517. 2:
  6518. pushq $229
  6519. .cfi_adjust_cfa_offset 8
  6520. call _libcuda_so_save_regs_and_resolve
  6521. addq $8, %rsp
  6522. .cfi_adjust_cfa_offset -8
  6523. jmp 1b
  6524. .cfi_endproc
  6525. /*
  6526. * Copyright 2018-2019 Yury Gribov
  6527. *
  6528. * The MIT License (MIT)
  6529. *
  6530. * Use of this source code is governed by MIT license that can be
  6531. * found in the LICENSE.txt file.
  6532. */
  6533. .globl cuMemGetInfo_v2
  6534. .p2align 4
  6535. .type cuMemGetInfo_v2, %function
  6536. cuMemGetInfo_v2:
  6537. .cfi_startproc
  6538. // Intel opt. manual says to
  6539. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6540. // to hint static predictor.
  6541. cmpq $0, _libcuda_so_tramp_table+1840(%rip)
  6542. je 2f
  6543. 1:
  6544. jmp *_libcuda_so_tramp_table+1840(%rip)
  6545. 2:
  6546. pushq $230
  6547. .cfi_adjust_cfa_offset 8
  6548. call _libcuda_so_save_regs_and_resolve
  6549. addq $8, %rsp
  6550. .cfi_adjust_cfa_offset -8
  6551. jmp 1b
  6552. .cfi_endproc
  6553. /*
  6554. * Copyright 2018-2019 Yury Gribov
  6555. *
  6556. * The MIT License (MIT)
  6557. *
  6558. * Use of this source code is governed by MIT license that can be
  6559. * found in the LICENSE.txt file.
  6560. */
  6561. .globl cuMemHostAlloc
  6562. .p2align 4
  6563. .type cuMemHostAlloc, %function
  6564. cuMemHostAlloc:
  6565. .cfi_startproc
  6566. // Intel opt. manual says to
  6567. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6568. // to hint static predictor.
  6569. cmpq $0, _libcuda_so_tramp_table+1848(%rip)
  6570. je 2f
  6571. 1:
  6572. jmp *_libcuda_so_tramp_table+1848(%rip)
  6573. 2:
  6574. pushq $231
  6575. .cfi_adjust_cfa_offset 8
  6576. call _libcuda_so_save_regs_and_resolve
  6577. addq $8, %rsp
  6578. .cfi_adjust_cfa_offset -8
  6579. jmp 1b
  6580. .cfi_endproc
  6581. /*
  6582. * Copyright 2018-2019 Yury Gribov
  6583. *
  6584. * The MIT License (MIT)
  6585. *
  6586. * Use of this source code is governed by MIT license that can be
  6587. * found in the LICENSE.txt file.
  6588. */
  6589. .globl cuMemHostGetDevicePointer
  6590. .p2align 4
  6591. .type cuMemHostGetDevicePointer, %function
  6592. cuMemHostGetDevicePointer:
  6593. .cfi_startproc
  6594. // Intel opt. manual says to
  6595. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6596. // to hint static predictor.
  6597. cmpq $0, _libcuda_so_tramp_table+1856(%rip)
  6598. je 2f
  6599. 1:
  6600. jmp *_libcuda_so_tramp_table+1856(%rip)
  6601. 2:
  6602. pushq $232
  6603. .cfi_adjust_cfa_offset 8
  6604. call _libcuda_so_save_regs_and_resolve
  6605. addq $8, %rsp
  6606. .cfi_adjust_cfa_offset -8
  6607. jmp 1b
  6608. .cfi_endproc
  6609. /*
  6610. * Copyright 2018-2019 Yury Gribov
  6611. *
  6612. * The MIT License (MIT)
  6613. *
  6614. * Use of this source code is governed by MIT license that can be
  6615. * found in the LICENSE.txt file.
  6616. */
  6617. .globl cuMemHostGetDevicePointer_v2
  6618. .p2align 4
  6619. .type cuMemHostGetDevicePointer_v2, %function
  6620. cuMemHostGetDevicePointer_v2:
  6621. .cfi_startproc
  6622. // Intel opt. manual says to
  6623. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6624. // to hint static predictor.
  6625. cmpq $0, _libcuda_so_tramp_table+1864(%rip)
  6626. je 2f
  6627. 1:
  6628. jmp *_libcuda_so_tramp_table+1864(%rip)
  6629. 2:
  6630. pushq $233
  6631. .cfi_adjust_cfa_offset 8
  6632. call _libcuda_so_save_regs_and_resolve
  6633. addq $8, %rsp
  6634. .cfi_adjust_cfa_offset -8
  6635. jmp 1b
  6636. .cfi_endproc
  6637. /*
  6638. * Copyright 2018-2019 Yury Gribov
  6639. *
  6640. * The MIT License (MIT)
  6641. *
  6642. * Use of this source code is governed by MIT license that can be
  6643. * found in the LICENSE.txt file.
  6644. */
  6645. .globl cuMemHostGetFlags
  6646. .p2align 4
  6647. .type cuMemHostGetFlags, %function
  6648. cuMemHostGetFlags:
  6649. .cfi_startproc
  6650. // Intel opt. manual says to
  6651. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6652. // to hint static predictor.
  6653. cmpq $0, _libcuda_so_tramp_table+1872(%rip)
  6654. je 2f
  6655. 1:
  6656. jmp *_libcuda_so_tramp_table+1872(%rip)
  6657. 2:
  6658. pushq $234
  6659. .cfi_adjust_cfa_offset 8
  6660. call _libcuda_so_save_regs_and_resolve
  6661. addq $8, %rsp
  6662. .cfi_adjust_cfa_offset -8
  6663. jmp 1b
  6664. .cfi_endproc
  6665. /*
  6666. * Copyright 2018-2019 Yury Gribov
  6667. *
  6668. * The MIT License (MIT)
  6669. *
  6670. * Use of this source code is governed by MIT license that can be
  6671. * found in the LICENSE.txt file.
  6672. */
  6673. .globl cuMemHostRegister
  6674. .p2align 4
  6675. .type cuMemHostRegister, %function
  6676. cuMemHostRegister:
  6677. .cfi_startproc
  6678. // Intel opt. manual says to
  6679. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6680. // to hint static predictor.
  6681. cmpq $0, _libcuda_so_tramp_table+1880(%rip)
  6682. je 2f
  6683. 1:
  6684. jmp *_libcuda_so_tramp_table+1880(%rip)
  6685. 2:
  6686. pushq $235
  6687. .cfi_adjust_cfa_offset 8
  6688. call _libcuda_so_save_regs_and_resolve
  6689. addq $8, %rsp
  6690. .cfi_adjust_cfa_offset -8
  6691. jmp 1b
  6692. .cfi_endproc
  6693. /*
  6694. * Copyright 2018-2019 Yury Gribov
  6695. *
  6696. * The MIT License (MIT)
  6697. *
  6698. * Use of this source code is governed by MIT license that can be
  6699. * found in the LICENSE.txt file.
  6700. */
  6701. .globl cuMemHostRegister_v2
  6702. .p2align 4
  6703. .type cuMemHostRegister_v2, %function
  6704. cuMemHostRegister_v2:
  6705. .cfi_startproc
  6706. // Intel opt. manual says to
  6707. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6708. // to hint static predictor.
  6709. cmpq $0, _libcuda_so_tramp_table+1888(%rip)
  6710. je 2f
  6711. 1:
  6712. jmp *_libcuda_so_tramp_table+1888(%rip)
  6713. 2:
  6714. pushq $236
  6715. .cfi_adjust_cfa_offset 8
  6716. call _libcuda_so_save_regs_and_resolve
  6717. addq $8, %rsp
  6718. .cfi_adjust_cfa_offset -8
  6719. jmp 1b
  6720. .cfi_endproc
  6721. /*
  6722. * Copyright 2018-2019 Yury Gribov
  6723. *
  6724. * The MIT License (MIT)
  6725. *
  6726. * Use of this source code is governed by MIT license that can be
  6727. * found in the LICENSE.txt file.
  6728. */
  6729. .globl cuMemHostUnregister
  6730. .p2align 4
  6731. .type cuMemHostUnregister, %function
  6732. cuMemHostUnregister:
  6733. .cfi_startproc
  6734. // Intel opt. manual says to
  6735. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6736. // to hint static predictor.
  6737. cmpq $0, _libcuda_so_tramp_table+1896(%rip)
  6738. je 2f
  6739. 1:
  6740. jmp *_libcuda_so_tramp_table+1896(%rip)
  6741. 2:
  6742. pushq $237
  6743. .cfi_adjust_cfa_offset 8
  6744. call _libcuda_so_save_regs_and_resolve
  6745. addq $8, %rsp
  6746. .cfi_adjust_cfa_offset -8
  6747. jmp 1b
  6748. .cfi_endproc
  6749. /*
  6750. * Copyright 2018-2019 Yury Gribov
  6751. *
  6752. * The MIT License (MIT)
  6753. *
  6754. * Use of this source code is governed by MIT license that can be
  6755. * found in the LICENSE.txt file.
  6756. */
  6757. .globl cuMemImportFromShareableHandle
  6758. .p2align 4
  6759. .type cuMemImportFromShareableHandle, %function
  6760. cuMemImportFromShareableHandle:
  6761. .cfi_startproc
  6762. // Intel opt. manual says to
  6763. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6764. // to hint static predictor.
  6765. cmpq $0, _libcuda_so_tramp_table+1904(%rip)
  6766. je 2f
  6767. 1:
  6768. jmp *_libcuda_so_tramp_table+1904(%rip)
  6769. 2:
  6770. pushq $238
  6771. .cfi_adjust_cfa_offset 8
  6772. call _libcuda_so_save_regs_and_resolve
  6773. addq $8, %rsp
  6774. .cfi_adjust_cfa_offset -8
  6775. jmp 1b
  6776. .cfi_endproc
  6777. /*
  6778. * Copyright 2018-2019 Yury Gribov
  6779. *
  6780. * The MIT License (MIT)
  6781. *
  6782. * Use of this source code is governed by MIT license that can be
  6783. * found in the LICENSE.txt file.
  6784. */
  6785. .globl cuMemMap
  6786. .p2align 4
  6787. .type cuMemMap, %function
  6788. cuMemMap:
  6789. .cfi_startproc
  6790. // Intel opt. manual says to
  6791. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6792. // to hint static predictor.
  6793. cmpq $0, _libcuda_so_tramp_table+1912(%rip)
  6794. je 2f
  6795. 1:
  6796. jmp *_libcuda_so_tramp_table+1912(%rip)
  6797. 2:
  6798. pushq $239
  6799. .cfi_adjust_cfa_offset 8
  6800. call _libcuda_so_save_regs_and_resolve
  6801. addq $8, %rsp
  6802. .cfi_adjust_cfa_offset -8
  6803. jmp 1b
  6804. .cfi_endproc
  6805. /*
  6806. * Copyright 2018-2019 Yury Gribov
  6807. *
  6808. * The MIT License (MIT)
  6809. *
  6810. * Use of this source code is governed by MIT license that can be
  6811. * found in the LICENSE.txt file.
  6812. */
  6813. .globl cuMemMapArrayAsync
  6814. .p2align 4
  6815. .type cuMemMapArrayAsync, %function
  6816. cuMemMapArrayAsync:
  6817. .cfi_startproc
  6818. // Intel opt. manual says to
  6819. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6820. // to hint static predictor.
  6821. cmpq $0, _libcuda_so_tramp_table+1920(%rip)
  6822. je 2f
  6823. 1:
  6824. jmp *_libcuda_so_tramp_table+1920(%rip)
  6825. 2:
  6826. pushq $240
  6827. .cfi_adjust_cfa_offset 8
  6828. call _libcuda_so_save_regs_and_resolve
  6829. addq $8, %rsp
  6830. .cfi_adjust_cfa_offset -8
  6831. jmp 1b
  6832. .cfi_endproc
  6833. /*
  6834. * Copyright 2018-2019 Yury Gribov
  6835. *
  6836. * The MIT License (MIT)
  6837. *
  6838. * Use of this source code is governed by MIT license that can be
  6839. * found in the LICENSE.txt file.
  6840. */
  6841. .globl cuMemMapArrayAsync_ptsz
  6842. .p2align 4
  6843. .type cuMemMapArrayAsync_ptsz, %function
  6844. cuMemMapArrayAsync_ptsz:
  6845. .cfi_startproc
  6846. // Intel opt. manual says to
  6847. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6848. // to hint static predictor.
  6849. cmpq $0, _libcuda_so_tramp_table+1928(%rip)
  6850. je 2f
  6851. 1:
  6852. jmp *_libcuda_so_tramp_table+1928(%rip)
  6853. 2:
  6854. pushq $241
  6855. .cfi_adjust_cfa_offset 8
  6856. call _libcuda_so_save_regs_and_resolve
  6857. addq $8, %rsp
  6858. .cfi_adjust_cfa_offset -8
  6859. jmp 1b
  6860. .cfi_endproc
  6861. /*
  6862. * Copyright 2018-2019 Yury Gribov
  6863. *
  6864. * The MIT License (MIT)
  6865. *
  6866. * Use of this source code is governed by MIT license that can be
  6867. * found in the LICENSE.txt file.
  6868. */
  6869. .globl cuMemPrefetchAsync
  6870. .p2align 4
  6871. .type cuMemPrefetchAsync, %function
  6872. cuMemPrefetchAsync:
  6873. .cfi_startproc
  6874. // Intel opt. manual says to
  6875. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6876. // to hint static predictor.
  6877. cmpq $0, _libcuda_so_tramp_table+1936(%rip)
  6878. je 2f
  6879. 1:
  6880. jmp *_libcuda_so_tramp_table+1936(%rip)
  6881. 2:
  6882. pushq $242
  6883. .cfi_adjust_cfa_offset 8
  6884. call _libcuda_so_save_regs_and_resolve
  6885. addq $8, %rsp
  6886. .cfi_adjust_cfa_offset -8
  6887. jmp 1b
  6888. .cfi_endproc
  6889. /*
  6890. * Copyright 2018-2019 Yury Gribov
  6891. *
  6892. * The MIT License (MIT)
  6893. *
  6894. * Use of this source code is governed by MIT license that can be
  6895. * found in the LICENSE.txt file.
  6896. */
  6897. .globl cuMemPrefetchAsync_ptsz
  6898. .p2align 4
  6899. .type cuMemPrefetchAsync_ptsz, %function
  6900. cuMemPrefetchAsync_ptsz:
  6901. .cfi_startproc
  6902. // Intel opt. manual says to
  6903. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6904. // to hint static predictor.
  6905. cmpq $0, _libcuda_so_tramp_table+1944(%rip)
  6906. je 2f
  6907. 1:
  6908. jmp *_libcuda_so_tramp_table+1944(%rip)
  6909. 2:
  6910. pushq $243
  6911. .cfi_adjust_cfa_offset 8
  6912. call _libcuda_so_save_regs_and_resolve
  6913. addq $8, %rsp
  6914. .cfi_adjust_cfa_offset -8
  6915. jmp 1b
  6916. .cfi_endproc
  6917. /*
  6918. * Copyright 2018-2019 Yury Gribov
  6919. *
  6920. * The MIT License (MIT)
  6921. *
  6922. * Use of this source code is governed by MIT license that can be
  6923. * found in the LICENSE.txt file.
  6924. */
  6925. .globl cuMemRangeGetAttribute
  6926. .p2align 4
  6927. .type cuMemRangeGetAttribute, %function
  6928. cuMemRangeGetAttribute:
  6929. .cfi_startproc
  6930. // Intel opt. manual says to
  6931. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6932. // to hint static predictor.
  6933. cmpq $0, _libcuda_so_tramp_table+1952(%rip)
  6934. je 2f
  6935. 1:
  6936. jmp *_libcuda_so_tramp_table+1952(%rip)
  6937. 2:
  6938. pushq $244
  6939. .cfi_adjust_cfa_offset 8
  6940. call _libcuda_so_save_regs_and_resolve
  6941. addq $8, %rsp
  6942. .cfi_adjust_cfa_offset -8
  6943. jmp 1b
  6944. .cfi_endproc
  6945. /*
  6946. * Copyright 2018-2019 Yury Gribov
  6947. *
  6948. * The MIT License (MIT)
  6949. *
  6950. * Use of this source code is governed by MIT license that can be
  6951. * found in the LICENSE.txt file.
  6952. */
  6953. .globl cuMemRangeGetAttributes
  6954. .p2align 4
  6955. .type cuMemRangeGetAttributes, %function
  6956. cuMemRangeGetAttributes:
  6957. .cfi_startproc
  6958. // Intel opt. manual says to
  6959. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6960. // to hint static predictor.
  6961. cmpq $0, _libcuda_so_tramp_table+1960(%rip)
  6962. je 2f
  6963. 1:
  6964. jmp *_libcuda_so_tramp_table+1960(%rip)
  6965. 2:
  6966. pushq $245
  6967. .cfi_adjust_cfa_offset 8
  6968. call _libcuda_so_save_regs_and_resolve
  6969. addq $8, %rsp
  6970. .cfi_adjust_cfa_offset -8
  6971. jmp 1b
  6972. .cfi_endproc
  6973. /*
  6974. * Copyright 2018-2019 Yury Gribov
  6975. *
  6976. * The MIT License (MIT)
  6977. *
  6978. * Use of this source code is governed by MIT license that can be
  6979. * found in the LICENSE.txt file.
  6980. */
  6981. .globl cuMemRelease
  6982. .p2align 4
  6983. .type cuMemRelease, %function
  6984. cuMemRelease:
  6985. .cfi_startproc
  6986. // Intel opt. manual says to
  6987. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  6988. // to hint static predictor.
  6989. cmpq $0, _libcuda_so_tramp_table+1968(%rip)
  6990. je 2f
  6991. 1:
  6992. jmp *_libcuda_so_tramp_table+1968(%rip)
  6993. 2:
  6994. pushq $246
  6995. .cfi_adjust_cfa_offset 8
  6996. call _libcuda_so_save_regs_and_resolve
  6997. addq $8, %rsp
  6998. .cfi_adjust_cfa_offset -8
  6999. jmp 1b
  7000. .cfi_endproc
  7001. /*
  7002. * Copyright 2018-2019 Yury Gribov
  7003. *
  7004. * The MIT License (MIT)
  7005. *
  7006. * Use of this source code is governed by MIT license that can be
  7007. * found in the LICENSE.txt file.
  7008. */
  7009. .globl cuMemRetainAllocationHandle
  7010. .p2align 4
  7011. .type cuMemRetainAllocationHandle, %function
  7012. cuMemRetainAllocationHandle:
  7013. .cfi_startproc
  7014. // Intel opt. manual says to
  7015. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7016. // to hint static predictor.
  7017. cmpq $0, _libcuda_so_tramp_table+1976(%rip)
  7018. je 2f
  7019. 1:
  7020. jmp *_libcuda_so_tramp_table+1976(%rip)
  7021. 2:
  7022. pushq $247
  7023. .cfi_adjust_cfa_offset 8
  7024. call _libcuda_so_save_regs_and_resolve
  7025. addq $8, %rsp
  7026. .cfi_adjust_cfa_offset -8
  7027. jmp 1b
  7028. .cfi_endproc
  7029. /*
  7030. * Copyright 2018-2019 Yury Gribov
  7031. *
  7032. * The MIT License (MIT)
  7033. *
  7034. * Use of this source code is governed by MIT license that can be
  7035. * found in the LICENSE.txt file.
  7036. */
  7037. .globl cuMemSetAccess
  7038. .p2align 4
  7039. .type cuMemSetAccess, %function
  7040. cuMemSetAccess:
  7041. .cfi_startproc
  7042. // Intel opt. manual says to
  7043. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7044. // to hint static predictor.
  7045. cmpq $0, _libcuda_so_tramp_table+1984(%rip)
  7046. je 2f
  7047. 1:
  7048. jmp *_libcuda_so_tramp_table+1984(%rip)
  7049. 2:
  7050. pushq $248
  7051. .cfi_adjust_cfa_offset 8
  7052. call _libcuda_so_save_regs_and_resolve
  7053. addq $8, %rsp
  7054. .cfi_adjust_cfa_offset -8
  7055. jmp 1b
  7056. .cfi_endproc
  7057. /*
  7058. * Copyright 2018-2019 Yury Gribov
  7059. *
  7060. * The MIT License (MIT)
  7061. *
  7062. * Use of this source code is governed by MIT license that can be
  7063. * found in the LICENSE.txt file.
  7064. */
  7065. .globl cuMemUnmap
  7066. .p2align 4
  7067. .type cuMemUnmap, %function
  7068. cuMemUnmap:
  7069. .cfi_startproc
  7070. // Intel opt. manual says to
  7071. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7072. // to hint static predictor.
  7073. cmpq $0, _libcuda_so_tramp_table+1992(%rip)
  7074. je 2f
  7075. 1:
  7076. jmp *_libcuda_so_tramp_table+1992(%rip)
  7077. 2:
  7078. pushq $249
  7079. .cfi_adjust_cfa_offset 8
  7080. call _libcuda_so_save_regs_and_resolve
  7081. addq $8, %rsp
  7082. .cfi_adjust_cfa_offset -8
  7083. jmp 1b
  7084. .cfi_endproc
  7085. /*
  7086. * Copyright 2018-2019 Yury Gribov
  7087. *
  7088. * The MIT License (MIT)
  7089. *
  7090. * Use of this source code is governed by MIT license that can be
  7091. * found in the LICENSE.txt file.
  7092. */
  7093. .globl cuMemcpy
  7094. .p2align 4
  7095. .type cuMemcpy, %function
  7096. cuMemcpy:
  7097. .cfi_startproc
  7098. // Intel opt. manual says to
  7099. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7100. // to hint static predictor.
  7101. cmpq $0, _libcuda_so_tramp_table+2000(%rip)
  7102. je 2f
  7103. 1:
  7104. jmp *_libcuda_so_tramp_table+2000(%rip)
  7105. 2:
  7106. pushq $250
  7107. .cfi_adjust_cfa_offset 8
  7108. call _libcuda_so_save_regs_and_resolve
  7109. addq $8, %rsp
  7110. .cfi_adjust_cfa_offset -8
  7111. jmp 1b
  7112. .cfi_endproc
  7113. /*
  7114. * Copyright 2018-2019 Yury Gribov
  7115. *
  7116. * The MIT License (MIT)
  7117. *
  7118. * Use of this source code is governed by MIT license that can be
  7119. * found in the LICENSE.txt file.
  7120. */
  7121. .globl cuMemcpy2D
  7122. .p2align 4
  7123. .type cuMemcpy2D, %function
  7124. cuMemcpy2D:
  7125. .cfi_startproc
  7126. // Intel opt. manual says to
  7127. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7128. // to hint static predictor.
  7129. cmpq $0, _libcuda_so_tramp_table+2008(%rip)
  7130. je 2f
  7131. 1:
  7132. jmp *_libcuda_so_tramp_table+2008(%rip)
  7133. 2:
  7134. pushq $251
  7135. .cfi_adjust_cfa_offset 8
  7136. call _libcuda_so_save_regs_and_resolve
  7137. addq $8, %rsp
  7138. .cfi_adjust_cfa_offset -8
  7139. jmp 1b
  7140. .cfi_endproc
  7141. /*
  7142. * Copyright 2018-2019 Yury Gribov
  7143. *
  7144. * The MIT License (MIT)
  7145. *
  7146. * Use of this source code is governed by MIT license that can be
  7147. * found in the LICENSE.txt file.
  7148. */
  7149. .globl cuMemcpy2DAsync
  7150. .p2align 4
  7151. .type cuMemcpy2DAsync, %function
  7152. cuMemcpy2DAsync:
  7153. .cfi_startproc
  7154. // Intel opt. manual says to
  7155. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7156. // to hint static predictor.
  7157. cmpq $0, _libcuda_so_tramp_table+2016(%rip)
  7158. je 2f
  7159. 1:
  7160. jmp *_libcuda_so_tramp_table+2016(%rip)
  7161. 2:
  7162. pushq $252
  7163. .cfi_adjust_cfa_offset 8
  7164. call _libcuda_so_save_regs_and_resolve
  7165. addq $8, %rsp
  7166. .cfi_adjust_cfa_offset -8
  7167. jmp 1b
  7168. .cfi_endproc
  7169. /*
  7170. * Copyright 2018-2019 Yury Gribov
  7171. *
  7172. * The MIT License (MIT)
  7173. *
  7174. * Use of this source code is governed by MIT license that can be
  7175. * found in the LICENSE.txt file.
  7176. */
  7177. .globl cuMemcpy2DAsync_v2
  7178. .p2align 4
  7179. .type cuMemcpy2DAsync_v2, %function
  7180. cuMemcpy2DAsync_v2:
  7181. .cfi_startproc
  7182. // Intel opt. manual says to
  7183. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7184. // to hint static predictor.
  7185. cmpq $0, _libcuda_so_tramp_table+2024(%rip)
  7186. je 2f
  7187. 1:
  7188. jmp *_libcuda_so_tramp_table+2024(%rip)
  7189. 2:
  7190. pushq $253
  7191. .cfi_adjust_cfa_offset 8
  7192. call _libcuda_so_save_regs_and_resolve
  7193. addq $8, %rsp
  7194. .cfi_adjust_cfa_offset -8
  7195. jmp 1b
  7196. .cfi_endproc
  7197. /*
  7198. * Copyright 2018-2019 Yury Gribov
  7199. *
  7200. * The MIT License (MIT)
  7201. *
  7202. * Use of this source code is governed by MIT license that can be
  7203. * found in the LICENSE.txt file.
  7204. */
  7205. .globl cuMemcpy2DAsync_v2_ptsz
  7206. .p2align 4
  7207. .type cuMemcpy2DAsync_v2_ptsz, %function
  7208. cuMemcpy2DAsync_v2_ptsz:
  7209. .cfi_startproc
  7210. // Intel opt. manual says to
  7211. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7212. // to hint static predictor.
  7213. cmpq $0, _libcuda_so_tramp_table+2032(%rip)
  7214. je 2f
  7215. 1:
  7216. jmp *_libcuda_so_tramp_table+2032(%rip)
  7217. 2:
  7218. pushq $254
  7219. .cfi_adjust_cfa_offset 8
  7220. call _libcuda_so_save_regs_and_resolve
  7221. addq $8, %rsp
  7222. .cfi_adjust_cfa_offset -8
  7223. jmp 1b
  7224. .cfi_endproc
  7225. /*
  7226. * Copyright 2018-2019 Yury Gribov
  7227. *
  7228. * The MIT License (MIT)
  7229. *
  7230. * Use of this source code is governed by MIT license that can be
  7231. * found in the LICENSE.txt file.
  7232. */
  7233. .globl cuMemcpy2DUnaligned
  7234. .p2align 4
  7235. .type cuMemcpy2DUnaligned, %function
  7236. cuMemcpy2DUnaligned:
  7237. .cfi_startproc
  7238. // Intel opt. manual says to
  7239. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7240. // to hint static predictor.
  7241. cmpq $0, _libcuda_so_tramp_table+2040(%rip)
  7242. je 2f
  7243. 1:
  7244. jmp *_libcuda_so_tramp_table+2040(%rip)
  7245. 2:
  7246. pushq $255
  7247. .cfi_adjust_cfa_offset 8
  7248. call _libcuda_so_save_regs_and_resolve
  7249. addq $8, %rsp
  7250. .cfi_adjust_cfa_offset -8
  7251. jmp 1b
  7252. .cfi_endproc
  7253. /*
  7254. * Copyright 2018-2019 Yury Gribov
  7255. *
  7256. * The MIT License (MIT)
  7257. *
  7258. * Use of this source code is governed by MIT license that can be
  7259. * found in the LICENSE.txt file.
  7260. */
  7261. .globl cuMemcpy2DUnaligned_v2
  7262. .p2align 4
  7263. .type cuMemcpy2DUnaligned_v2, %function
  7264. cuMemcpy2DUnaligned_v2:
  7265. .cfi_startproc
  7266. // Intel opt. manual says to
  7267. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7268. // to hint static predictor.
  7269. cmpq $0, _libcuda_so_tramp_table+2048(%rip)
  7270. je 2f
  7271. 1:
  7272. jmp *_libcuda_so_tramp_table+2048(%rip)
  7273. 2:
  7274. pushq $256
  7275. .cfi_adjust_cfa_offset 8
  7276. call _libcuda_so_save_regs_and_resolve
  7277. addq $8, %rsp
  7278. .cfi_adjust_cfa_offset -8
  7279. jmp 1b
  7280. .cfi_endproc
  7281. /*
  7282. * Copyright 2018-2019 Yury Gribov
  7283. *
  7284. * The MIT License (MIT)
  7285. *
  7286. * Use of this source code is governed by MIT license that can be
  7287. * found in the LICENSE.txt file.
  7288. */
  7289. .globl cuMemcpy2DUnaligned_v2_ptds
  7290. .p2align 4
  7291. .type cuMemcpy2DUnaligned_v2_ptds, %function
  7292. cuMemcpy2DUnaligned_v2_ptds:
  7293. .cfi_startproc
  7294. // Intel opt. manual says to
  7295. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7296. // to hint static predictor.
  7297. cmpq $0, _libcuda_so_tramp_table+2056(%rip)
  7298. je 2f
  7299. 1:
  7300. jmp *_libcuda_so_tramp_table+2056(%rip)
  7301. 2:
  7302. pushq $257
  7303. .cfi_adjust_cfa_offset 8
  7304. call _libcuda_so_save_regs_and_resolve
  7305. addq $8, %rsp
  7306. .cfi_adjust_cfa_offset -8
  7307. jmp 1b
  7308. .cfi_endproc
  7309. /*
  7310. * Copyright 2018-2019 Yury Gribov
  7311. *
  7312. * The MIT License (MIT)
  7313. *
  7314. * Use of this source code is governed by MIT license that can be
  7315. * found in the LICENSE.txt file.
  7316. */
  7317. .globl cuMemcpy2D_v2
  7318. .p2align 4
  7319. .type cuMemcpy2D_v2, %function
  7320. cuMemcpy2D_v2:
  7321. .cfi_startproc
  7322. // Intel opt. manual says to
  7323. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7324. // to hint static predictor.
  7325. cmpq $0, _libcuda_so_tramp_table+2064(%rip)
  7326. je 2f
  7327. 1:
  7328. jmp *_libcuda_so_tramp_table+2064(%rip)
  7329. 2:
  7330. pushq $258
  7331. .cfi_adjust_cfa_offset 8
  7332. call _libcuda_so_save_regs_and_resolve
  7333. addq $8, %rsp
  7334. .cfi_adjust_cfa_offset -8
  7335. jmp 1b
  7336. .cfi_endproc
  7337. /*
  7338. * Copyright 2018-2019 Yury Gribov
  7339. *
  7340. * The MIT License (MIT)
  7341. *
  7342. * Use of this source code is governed by MIT license that can be
  7343. * found in the LICENSE.txt file.
  7344. */
  7345. .globl cuMemcpy2D_v2_ptds
  7346. .p2align 4
  7347. .type cuMemcpy2D_v2_ptds, %function
  7348. cuMemcpy2D_v2_ptds:
  7349. .cfi_startproc
  7350. // Intel opt. manual says to
  7351. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7352. // to hint static predictor.
  7353. cmpq $0, _libcuda_so_tramp_table+2072(%rip)
  7354. je 2f
  7355. 1:
  7356. jmp *_libcuda_so_tramp_table+2072(%rip)
  7357. 2:
  7358. pushq $259
  7359. .cfi_adjust_cfa_offset 8
  7360. call _libcuda_so_save_regs_and_resolve
  7361. addq $8, %rsp
  7362. .cfi_adjust_cfa_offset -8
  7363. jmp 1b
  7364. .cfi_endproc
  7365. /*
  7366. * Copyright 2018-2019 Yury Gribov
  7367. *
  7368. * The MIT License (MIT)
  7369. *
  7370. * Use of this source code is governed by MIT license that can be
  7371. * found in the LICENSE.txt file.
  7372. */
  7373. .globl cuMemcpy3D
  7374. .p2align 4
  7375. .type cuMemcpy3D, %function
  7376. cuMemcpy3D:
  7377. .cfi_startproc
  7378. // Intel opt. manual says to
  7379. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7380. // to hint static predictor.
  7381. cmpq $0, _libcuda_so_tramp_table+2080(%rip)
  7382. je 2f
  7383. 1:
  7384. jmp *_libcuda_so_tramp_table+2080(%rip)
  7385. 2:
  7386. pushq $260
  7387. .cfi_adjust_cfa_offset 8
  7388. call _libcuda_so_save_regs_and_resolve
  7389. addq $8, %rsp
  7390. .cfi_adjust_cfa_offset -8
  7391. jmp 1b
  7392. .cfi_endproc
  7393. /*
  7394. * Copyright 2018-2019 Yury Gribov
  7395. *
  7396. * The MIT License (MIT)
  7397. *
  7398. * Use of this source code is governed by MIT license that can be
  7399. * found in the LICENSE.txt file.
  7400. */
  7401. .globl cuMemcpy3DAsync
  7402. .p2align 4
  7403. .type cuMemcpy3DAsync, %function
  7404. cuMemcpy3DAsync:
  7405. .cfi_startproc
  7406. // Intel opt. manual says to
  7407. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7408. // to hint static predictor.
  7409. cmpq $0, _libcuda_so_tramp_table+2088(%rip)
  7410. je 2f
  7411. 1:
  7412. jmp *_libcuda_so_tramp_table+2088(%rip)
  7413. 2:
  7414. pushq $261
  7415. .cfi_adjust_cfa_offset 8
  7416. call _libcuda_so_save_regs_and_resolve
  7417. addq $8, %rsp
  7418. .cfi_adjust_cfa_offset -8
  7419. jmp 1b
  7420. .cfi_endproc
  7421. /*
  7422. * Copyright 2018-2019 Yury Gribov
  7423. *
  7424. * The MIT License (MIT)
  7425. *
  7426. * Use of this source code is governed by MIT license that can be
  7427. * found in the LICENSE.txt file.
  7428. */
  7429. .globl cuMemcpy3DAsync_v2
  7430. .p2align 4
  7431. .type cuMemcpy3DAsync_v2, %function
  7432. cuMemcpy3DAsync_v2:
  7433. .cfi_startproc
  7434. // Intel opt. manual says to
  7435. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7436. // to hint static predictor.
  7437. cmpq $0, _libcuda_so_tramp_table+2096(%rip)
  7438. je 2f
  7439. 1:
  7440. jmp *_libcuda_so_tramp_table+2096(%rip)
  7441. 2:
  7442. pushq $262
  7443. .cfi_adjust_cfa_offset 8
  7444. call _libcuda_so_save_regs_and_resolve
  7445. addq $8, %rsp
  7446. .cfi_adjust_cfa_offset -8
  7447. jmp 1b
  7448. .cfi_endproc
  7449. /*
  7450. * Copyright 2018-2019 Yury Gribov
  7451. *
  7452. * The MIT License (MIT)
  7453. *
  7454. * Use of this source code is governed by MIT license that can be
  7455. * found in the LICENSE.txt file.
  7456. */
  7457. .globl cuMemcpy3DAsync_v2_ptsz
  7458. .p2align 4
  7459. .type cuMemcpy3DAsync_v2_ptsz, %function
  7460. cuMemcpy3DAsync_v2_ptsz:
  7461. .cfi_startproc
  7462. // Intel opt. manual says to
  7463. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7464. // to hint static predictor.
  7465. cmpq $0, _libcuda_so_tramp_table+2104(%rip)
  7466. je 2f
  7467. 1:
  7468. jmp *_libcuda_so_tramp_table+2104(%rip)
  7469. 2:
  7470. pushq $263
  7471. .cfi_adjust_cfa_offset 8
  7472. call _libcuda_so_save_regs_and_resolve
  7473. addq $8, %rsp
  7474. .cfi_adjust_cfa_offset -8
  7475. jmp 1b
  7476. .cfi_endproc
  7477. /*
  7478. * Copyright 2018-2019 Yury Gribov
  7479. *
  7480. * The MIT License (MIT)
  7481. *
  7482. * Use of this source code is governed by MIT license that can be
  7483. * found in the LICENSE.txt file.
  7484. */
  7485. .globl cuMemcpy3DPeer
  7486. .p2align 4
  7487. .type cuMemcpy3DPeer, %function
  7488. cuMemcpy3DPeer:
  7489. .cfi_startproc
  7490. // Intel opt. manual says to
  7491. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7492. // to hint static predictor.
  7493. cmpq $0, _libcuda_so_tramp_table+2112(%rip)
  7494. je 2f
  7495. 1:
  7496. jmp *_libcuda_so_tramp_table+2112(%rip)
  7497. 2:
  7498. pushq $264
  7499. .cfi_adjust_cfa_offset 8
  7500. call _libcuda_so_save_regs_and_resolve
  7501. addq $8, %rsp
  7502. .cfi_adjust_cfa_offset -8
  7503. jmp 1b
  7504. .cfi_endproc
  7505. /*
  7506. * Copyright 2018-2019 Yury Gribov
  7507. *
  7508. * The MIT License (MIT)
  7509. *
  7510. * Use of this source code is governed by MIT license that can be
  7511. * found in the LICENSE.txt file.
  7512. */
  7513. .globl cuMemcpy3DPeerAsync
  7514. .p2align 4
  7515. .type cuMemcpy3DPeerAsync, %function
  7516. cuMemcpy3DPeerAsync:
  7517. .cfi_startproc
  7518. // Intel opt. manual says to
  7519. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7520. // to hint static predictor.
  7521. cmpq $0, _libcuda_so_tramp_table+2120(%rip)
  7522. je 2f
  7523. 1:
  7524. jmp *_libcuda_so_tramp_table+2120(%rip)
  7525. 2:
  7526. pushq $265
  7527. .cfi_adjust_cfa_offset 8
  7528. call _libcuda_so_save_regs_and_resolve
  7529. addq $8, %rsp
  7530. .cfi_adjust_cfa_offset -8
  7531. jmp 1b
  7532. .cfi_endproc
  7533. /*
  7534. * Copyright 2018-2019 Yury Gribov
  7535. *
  7536. * The MIT License (MIT)
  7537. *
  7538. * Use of this source code is governed by MIT license that can be
  7539. * found in the LICENSE.txt file.
  7540. */
  7541. .globl cuMemcpy3DPeerAsync_ptsz
  7542. .p2align 4
  7543. .type cuMemcpy3DPeerAsync_ptsz, %function
  7544. cuMemcpy3DPeerAsync_ptsz:
  7545. .cfi_startproc
  7546. // Intel opt. manual says to
  7547. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7548. // to hint static predictor.
  7549. cmpq $0, _libcuda_so_tramp_table+2128(%rip)
  7550. je 2f
  7551. 1:
  7552. jmp *_libcuda_so_tramp_table+2128(%rip)
  7553. 2:
  7554. pushq $266
  7555. .cfi_adjust_cfa_offset 8
  7556. call _libcuda_so_save_regs_and_resolve
  7557. addq $8, %rsp
  7558. .cfi_adjust_cfa_offset -8
  7559. jmp 1b
  7560. .cfi_endproc
  7561. /*
  7562. * Copyright 2018-2019 Yury Gribov
  7563. *
  7564. * The MIT License (MIT)
  7565. *
  7566. * Use of this source code is governed by MIT license that can be
  7567. * found in the LICENSE.txt file.
  7568. */
  7569. .globl cuMemcpy3DPeer_ptds
  7570. .p2align 4
  7571. .type cuMemcpy3DPeer_ptds, %function
  7572. cuMemcpy3DPeer_ptds:
  7573. .cfi_startproc
  7574. // Intel opt. manual says to
  7575. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7576. // to hint static predictor.
  7577. cmpq $0, _libcuda_so_tramp_table+2136(%rip)
  7578. je 2f
  7579. 1:
  7580. jmp *_libcuda_so_tramp_table+2136(%rip)
  7581. 2:
  7582. pushq $267
  7583. .cfi_adjust_cfa_offset 8
  7584. call _libcuda_so_save_regs_and_resolve
  7585. addq $8, %rsp
  7586. .cfi_adjust_cfa_offset -8
  7587. jmp 1b
  7588. .cfi_endproc
  7589. /*
  7590. * Copyright 2018-2019 Yury Gribov
  7591. *
  7592. * The MIT License (MIT)
  7593. *
  7594. * Use of this source code is governed by MIT license that can be
  7595. * found in the LICENSE.txt file.
  7596. */
  7597. .globl cuMemcpy3D_v2
  7598. .p2align 4
  7599. .type cuMemcpy3D_v2, %function
  7600. cuMemcpy3D_v2:
  7601. .cfi_startproc
  7602. // Intel opt. manual says to
  7603. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7604. // to hint static predictor.
  7605. cmpq $0, _libcuda_so_tramp_table+2144(%rip)
  7606. je 2f
  7607. 1:
  7608. jmp *_libcuda_so_tramp_table+2144(%rip)
  7609. 2:
  7610. pushq $268
  7611. .cfi_adjust_cfa_offset 8
  7612. call _libcuda_so_save_regs_and_resolve
  7613. addq $8, %rsp
  7614. .cfi_adjust_cfa_offset -8
  7615. jmp 1b
  7616. .cfi_endproc
  7617. /*
  7618. * Copyright 2018-2019 Yury Gribov
  7619. *
  7620. * The MIT License (MIT)
  7621. *
  7622. * Use of this source code is governed by MIT license that can be
  7623. * found in the LICENSE.txt file.
  7624. */
  7625. .globl cuMemcpy3D_v2_ptds
  7626. .p2align 4
  7627. .type cuMemcpy3D_v2_ptds, %function
  7628. cuMemcpy3D_v2_ptds:
  7629. .cfi_startproc
  7630. // Intel opt. manual says to
  7631. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7632. // to hint static predictor.
  7633. cmpq $0, _libcuda_so_tramp_table+2152(%rip)
  7634. je 2f
  7635. 1:
  7636. jmp *_libcuda_so_tramp_table+2152(%rip)
  7637. 2:
  7638. pushq $269
  7639. .cfi_adjust_cfa_offset 8
  7640. call _libcuda_so_save_regs_and_resolve
  7641. addq $8, %rsp
  7642. .cfi_adjust_cfa_offset -8
  7643. jmp 1b
  7644. .cfi_endproc
  7645. /*
  7646. * Copyright 2018-2019 Yury Gribov
  7647. *
  7648. * The MIT License (MIT)
  7649. *
  7650. * Use of this source code is governed by MIT license that can be
  7651. * found in the LICENSE.txt file.
  7652. */
  7653. .globl cuMemcpyAsync
  7654. .p2align 4
  7655. .type cuMemcpyAsync, %function
  7656. cuMemcpyAsync:
  7657. .cfi_startproc
  7658. // Intel opt. manual says to
  7659. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7660. // to hint static predictor.
  7661. cmpq $0, _libcuda_so_tramp_table+2160(%rip)
  7662. je 2f
  7663. 1:
  7664. jmp *_libcuda_so_tramp_table+2160(%rip)
  7665. 2:
  7666. pushq $270
  7667. .cfi_adjust_cfa_offset 8
  7668. call _libcuda_so_save_regs_and_resolve
  7669. addq $8, %rsp
  7670. .cfi_adjust_cfa_offset -8
  7671. jmp 1b
  7672. .cfi_endproc
  7673. /*
  7674. * Copyright 2018-2019 Yury Gribov
  7675. *
  7676. * The MIT License (MIT)
  7677. *
  7678. * Use of this source code is governed by MIT license that can be
  7679. * found in the LICENSE.txt file.
  7680. */
  7681. .globl cuMemcpyAsync_ptsz
  7682. .p2align 4
  7683. .type cuMemcpyAsync_ptsz, %function
  7684. cuMemcpyAsync_ptsz:
  7685. .cfi_startproc
  7686. // Intel opt. manual says to
  7687. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7688. // to hint static predictor.
  7689. cmpq $0, _libcuda_so_tramp_table+2168(%rip)
  7690. je 2f
  7691. 1:
  7692. jmp *_libcuda_so_tramp_table+2168(%rip)
  7693. 2:
  7694. pushq $271
  7695. .cfi_adjust_cfa_offset 8
  7696. call _libcuda_so_save_regs_and_resolve
  7697. addq $8, %rsp
  7698. .cfi_adjust_cfa_offset -8
  7699. jmp 1b
  7700. .cfi_endproc
  7701. /*
  7702. * Copyright 2018-2019 Yury Gribov
  7703. *
  7704. * The MIT License (MIT)
  7705. *
  7706. * Use of this source code is governed by MIT license that can be
  7707. * found in the LICENSE.txt file.
  7708. */
  7709. .globl cuMemcpyAtoA
  7710. .p2align 4
  7711. .type cuMemcpyAtoA, %function
  7712. cuMemcpyAtoA:
  7713. .cfi_startproc
  7714. // Intel opt. manual says to
  7715. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7716. // to hint static predictor.
  7717. cmpq $0, _libcuda_so_tramp_table+2176(%rip)
  7718. je 2f
  7719. 1:
  7720. jmp *_libcuda_so_tramp_table+2176(%rip)
  7721. 2:
  7722. pushq $272
  7723. .cfi_adjust_cfa_offset 8
  7724. call _libcuda_so_save_regs_and_resolve
  7725. addq $8, %rsp
  7726. .cfi_adjust_cfa_offset -8
  7727. jmp 1b
  7728. .cfi_endproc
  7729. /*
  7730. * Copyright 2018-2019 Yury Gribov
  7731. *
  7732. * The MIT License (MIT)
  7733. *
  7734. * Use of this source code is governed by MIT license that can be
  7735. * found in the LICENSE.txt file.
  7736. */
  7737. .globl cuMemcpyAtoA_v2
  7738. .p2align 4
  7739. .type cuMemcpyAtoA_v2, %function
  7740. cuMemcpyAtoA_v2:
  7741. .cfi_startproc
  7742. // Intel opt. manual says to
  7743. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7744. // to hint static predictor.
  7745. cmpq $0, _libcuda_so_tramp_table+2184(%rip)
  7746. je 2f
  7747. 1:
  7748. jmp *_libcuda_so_tramp_table+2184(%rip)
  7749. 2:
  7750. pushq $273
  7751. .cfi_adjust_cfa_offset 8
  7752. call _libcuda_so_save_regs_and_resolve
  7753. addq $8, %rsp
  7754. .cfi_adjust_cfa_offset -8
  7755. jmp 1b
  7756. .cfi_endproc
  7757. /*
  7758. * Copyright 2018-2019 Yury Gribov
  7759. *
  7760. * The MIT License (MIT)
  7761. *
  7762. * Use of this source code is governed by MIT license that can be
  7763. * found in the LICENSE.txt file.
  7764. */
  7765. .globl cuMemcpyAtoA_v2_ptds
  7766. .p2align 4
  7767. .type cuMemcpyAtoA_v2_ptds, %function
  7768. cuMemcpyAtoA_v2_ptds:
  7769. .cfi_startproc
  7770. // Intel opt. manual says to
  7771. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7772. // to hint static predictor.
  7773. cmpq $0, _libcuda_so_tramp_table+2192(%rip)
  7774. je 2f
  7775. 1:
  7776. jmp *_libcuda_so_tramp_table+2192(%rip)
  7777. 2:
  7778. pushq $274
  7779. .cfi_adjust_cfa_offset 8
  7780. call _libcuda_so_save_regs_and_resolve
  7781. addq $8, %rsp
  7782. .cfi_adjust_cfa_offset -8
  7783. jmp 1b
  7784. .cfi_endproc
  7785. /*
  7786. * Copyright 2018-2019 Yury Gribov
  7787. *
  7788. * The MIT License (MIT)
  7789. *
  7790. * Use of this source code is governed by MIT license that can be
  7791. * found in the LICENSE.txt file.
  7792. */
  7793. .globl cuMemcpyAtoD
  7794. .p2align 4
  7795. .type cuMemcpyAtoD, %function
  7796. cuMemcpyAtoD:
  7797. .cfi_startproc
  7798. // Intel opt. manual says to
  7799. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7800. // to hint static predictor.
  7801. cmpq $0, _libcuda_so_tramp_table+2200(%rip)
  7802. je 2f
  7803. 1:
  7804. jmp *_libcuda_so_tramp_table+2200(%rip)
  7805. 2:
  7806. pushq $275
  7807. .cfi_adjust_cfa_offset 8
  7808. call _libcuda_so_save_regs_and_resolve
  7809. addq $8, %rsp
  7810. .cfi_adjust_cfa_offset -8
  7811. jmp 1b
  7812. .cfi_endproc
  7813. /*
  7814. * Copyright 2018-2019 Yury Gribov
  7815. *
  7816. * The MIT License (MIT)
  7817. *
  7818. * Use of this source code is governed by MIT license that can be
  7819. * found in the LICENSE.txt file.
  7820. */
  7821. .globl cuMemcpyAtoD_v2
  7822. .p2align 4
  7823. .type cuMemcpyAtoD_v2, %function
  7824. cuMemcpyAtoD_v2:
  7825. .cfi_startproc
  7826. // Intel opt. manual says to
  7827. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7828. // to hint static predictor.
  7829. cmpq $0, _libcuda_so_tramp_table+2208(%rip)
  7830. je 2f
  7831. 1:
  7832. jmp *_libcuda_so_tramp_table+2208(%rip)
  7833. 2:
  7834. pushq $276
  7835. .cfi_adjust_cfa_offset 8
  7836. call _libcuda_so_save_regs_and_resolve
  7837. addq $8, %rsp
  7838. .cfi_adjust_cfa_offset -8
  7839. jmp 1b
  7840. .cfi_endproc
  7841. /*
  7842. * Copyright 2018-2019 Yury Gribov
  7843. *
  7844. * The MIT License (MIT)
  7845. *
  7846. * Use of this source code is governed by MIT license that can be
  7847. * found in the LICENSE.txt file.
  7848. */
  7849. .globl cuMemcpyAtoD_v2_ptds
  7850. .p2align 4
  7851. .type cuMemcpyAtoD_v2_ptds, %function
  7852. cuMemcpyAtoD_v2_ptds:
  7853. .cfi_startproc
  7854. // Intel opt. manual says to
  7855. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7856. // to hint static predictor.
  7857. cmpq $0, _libcuda_so_tramp_table+2216(%rip)
  7858. je 2f
  7859. 1:
  7860. jmp *_libcuda_so_tramp_table+2216(%rip)
  7861. 2:
  7862. pushq $277
  7863. .cfi_adjust_cfa_offset 8
  7864. call _libcuda_so_save_regs_and_resolve
  7865. addq $8, %rsp
  7866. .cfi_adjust_cfa_offset -8
  7867. jmp 1b
  7868. .cfi_endproc
  7869. /*
  7870. * Copyright 2018-2019 Yury Gribov
  7871. *
  7872. * The MIT License (MIT)
  7873. *
  7874. * Use of this source code is governed by MIT license that can be
  7875. * found in the LICENSE.txt file.
  7876. */
  7877. .globl cuMemcpyAtoH
  7878. .p2align 4
  7879. .type cuMemcpyAtoH, %function
  7880. cuMemcpyAtoH:
  7881. .cfi_startproc
  7882. // Intel opt. manual says to
  7883. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7884. // to hint static predictor.
  7885. cmpq $0, _libcuda_so_tramp_table+2224(%rip)
  7886. je 2f
  7887. 1:
  7888. jmp *_libcuda_so_tramp_table+2224(%rip)
  7889. 2:
  7890. pushq $278
  7891. .cfi_adjust_cfa_offset 8
  7892. call _libcuda_so_save_regs_and_resolve
  7893. addq $8, %rsp
  7894. .cfi_adjust_cfa_offset -8
  7895. jmp 1b
  7896. .cfi_endproc
  7897. /*
  7898. * Copyright 2018-2019 Yury Gribov
  7899. *
  7900. * The MIT License (MIT)
  7901. *
  7902. * Use of this source code is governed by MIT license that can be
  7903. * found in the LICENSE.txt file.
  7904. */
  7905. .globl cuMemcpyAtoHAsync
  7906. .p2align 4
  7907. .type cuMemcpyAtoHAsync, %function
  7908. cuMemcpyAtoHAsync:
  7909. .cfi_startproc
  7910. // Intel opt. manual says to
  7911. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7912. // to hint static predictor.
  7913. cmpq $0, _libcuda_so_tramp_table+2232(%rip)
  7914. je 2f
  7915. 1:
  7916. jmp *_libcuda_so_tramp_table+2232(%rip)
  7917. 2:
  7918. pushq $279
  7919. .cfi_adjust_cfa_offset 8
  7920. call _libcuda_so_save_regs_and_resolve
  7921. addq $8, %rsp
  7922. .cfi_adjust_cfa_offset -8
  7923. jmp 1b
  7924. .cfi_endproc
  7925. /*
  7926. * Copyright 2018-2019 Yury Gribov
  7927. *
  7928. * The MIT License (MIT)
  7929. *
  7930. * Use of this source code is governed by MIT license that can be
  7931. * found in the LICENSE.txt file.
  7932. */
  7933. .globl cuMemcpyAtoHAsync_v2
  7934. .p2align 4
  7935. .type cuMemcpyAtoHAsync_v2, %function
  7936. cuMemcpyAtoHAsync_v2:
  7937. .cfi_startproc
  7938. // Intel opt. manual says to
  7939. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7940. // to hint static predictor.
  7941. cmpq $0, _libcuda_so_tramp_table+2240(%rip)
  7942. je 2f
  7943. 1:
  7944. jmp *_libcuda_so_tramp_table+2240(%rip)
  7945. 2:
  7946. pushq $280
  7947. .cfi_adjust_cfa_offset 8
  7948. call _libcuda_so_save_regs_and_resolve
  7949. addq $8, %rsp
  7950. .cfi_adjust_cfa_offset -8
  7951. jmp 1b
  7952. .cfi_endproc
  7953. /*
  7954. * Copyright 2018-2019 Yury Gribov
  7955. *
  7956. * The MIT License (MIT)
  7957. *
  7958. * Use of this source code is governed by MIT license that can be
  7959. * found in the LICENSE.txt file.
  7960. */
  7961. .globl cuMemcpyAtoHAsync_v2_ptsz
  7962. .p2align 4
  7963. .type cuMemcpyAtoHAsync_v2_ptsz, %function
  7964. cuMemcpyAtoHAsync_v2_ptsz:
  7965. .cfi_startproc
  7966. // Intel opt. manual says to
  7967. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7968. // to hint static predictor.
  7969. cmpq $0, _libcuda_so_tramp_table+2248(%rip)
  7970. je 2f
  7971. 1:
  7972. jmp *_libcuda_so_tramp_table+2248(%rip)
  7973. 2:
  7974. pushq $281
  7975. .cfi_adjust_cfa_offset 8
  7976. call _libcuda_so_save_regs_and_resolve
  7977. addq $8, %rsp
  7978. .cfi_adjust_cfa_offset -8
  7979. jmp 1b
  7980. .cfi_endproc
  7981. /*
  7982. * Copyright 2018-2019 Yury Gribov
  7983. *
  7984. * The MIT License (MIT)
  7985. *
  7986. * Use of this source code is governed by MIT license that can be
  7987. * found in the LICENSE.txt file.
  7988. */
  7989. .globl cuMemcpyAtoH_v2
  7990. .p2align 4
  7991. .type cuMemcpyAtoH_v2, %function
  7992. cuMemcpyAtoH_v2:
  7993. .cfi_startproc
  7994. // Intel opt. manual says to
  7995. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  7996. // to hint static predictor.
  7997. cmpq $0, _libcuda_so_tramp_table+2256(%rip)
  7998. je 2f
  7999. 1:
  8000. jmp *_libcuda_so_tramp_table+2256(%rip)
  8001. 2:
  8002. pushq $282
  8003. .cfi_adjust_cfa_offset 8
  8004. call _libcuda_so_save_regs_and_resolve
  8005. addq $8, %rsp
  8006. .cfi_adjust_cfa_offset -8
  8007. jmp 1b
  8008. .cfi_endproc
  8009. /*
  8010. * Copyright 2018-2019 Yury Gribov
  8011. *
  8012. * The MIT License (MIT)
  8013. *
  8014. * Use of this source code is governed by MIT license that can be
  8015. * found in the LICENSE.txt file.
  8016. */
  8017. .globl cuMemcpyAtoH_v2_ptds
  8018. .p2align 4
  8019. .type cuMemcpyAtoH_v2_ptds, %function
  8020. cuMemcpyAtoH_v2_ptds:
  8021. .cfi_startproc
  8022. // Intel opt. manual says to
  8023. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8024. // to hint static predictor.
  8025. cmpq $0, _libcuda_so_tramp_table+2264(%rip)
  8026. je 2f
  8027. 1:
  8028. jmp *_libcuda_so_tramp_table+2264(%rip)
  8029. 2:
  8030. pushq $283
  8031. .cfi_adjust_cfa_offset 8
  8032. call _libcuda_so_save_regs_and_resolve
  8033. addq $8, %rsp
  8034. .cfi_adjust_cfa_offset -8
  8035. jmp 1b
  8036. .cfi_endproc
  8037. /*
  8038. * Copyright 2018-2019 Yury Gribov
  8039. *
  8040. * The MIT License (MIT)
  8041. *
  8042. * Use of this source code is governed by MIT license that can be
  8043. * found in the LICENSE.txt file.
  8044. */
  8045. .globl cuMemcpyDtoA
  8046. .p2align 4
  8047. .type cuMemcpyDtoA, %function
  8048. cuMemcpyDtoA:
  8049. .cfi_startproc
  8050. // Intel opt. manual says to
  8051. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8052. // to hint static predictor.
  8053. cmpq $0, _libcuda_so_tramp_table+2272(%rip)
  8054. je 2f
  8055. 1:
  8056. jmp *_libcuda_so_tramp_table+2272(%rip)
  8057. 2:
  8058. pushq $284
  8059. .cfi_adjust_cfa_offset 8
  8060. call _libcuda_so_save_regs_and_resolve
  8061. addq $8, %rsp
  8062. .cfi_adjust_cfa_offset -8
  8063. jmp 1b
  8064. .cfi_endproc
  8065. /*
  8066. * Copyright 2018-2019 Yury Gribov
  8067. *
  8068. * The MIT License (MIT)
  8069. *
  8070. * Use of this source code is governed by MIT license that can be
  8071. * found in the LICENSE.txt file.
  8072. */
  8073. .globl cuMemcpyDtoA_v2
  8074. .p2align 4
  8075. .type cuMemcpyDtoA_v2, %function
  8076. cuMemcpyDtoA_v2:
  8077. .cfi_startproc
  8078. // Intel opt. manual says to
  8079. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8080. // to hint static predictor.
  8081. cmpq $0, _libcuda_so_tramp_table+2280(%rip)
  8082. je 2f
  8083. 1:
  8084. jmp *_libcuda_so_tramp_table+2280(%rip)
  8085. 2:
  8086. pushq $285
  8087. .cfi_adjust_cfa_offset 8
  8088. call _libcuda_so_save_regs_and_resolve
  8089. addq $8, %rsp
  8090. .cfi_adjust_cfa_offset -8
  8091. jmp 1b
  8092. .cfi_endproc
  8093. /*
  8094. * Copyright 2018-2019 Yury Gribov
  8095. *
  8096. * The MIT License (MIT)
  8097. *
  8098. * Use of this source code is governed by MIT license that can be
  8099. * found in the LICENSE.txt file.
  8100. */
  8101. .globl cuMemcpyDtoA_v2_ptds
  8102. .p2align 4
  8103. .type cuMemcpyDtoA_v2_ptds, %function
  8104. cuMemcpyDtoA_v2_ptds:
  8105. .cfi_startproc
  8106. // Intel opt. manual says to
  8107. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8108. // to hint static predictor.
  8109. cmpq $0, _libcuda_so_tramp_table+2288(%rip)
  8110. je 2f
  8111. 1:
  8112. jmp *_libcuda_so_tramp_table+2288(%rip)
  8113. 2:
  8114. pushq $286
  8115. .cfi_adjust_cfa_offset 8
  8116. call _libcuda_so_save_regs_and_resolve
  8117. addq $8, %rsp
  8118. .cfi_adjust_cfa_offset -8
  8119. jmp 1b
  8120. .cfi_endproc
  8121. /*
  8122. * Copyright 2018-2019 Yury Gribov
  8123. *
  8124. * The MIT License (MIT)
  8125. *
  8126. * Use of this source code is governed by MIT license that can be
  8127. * found in the LICENSE.txt file.
  8128. */
  8129. .globl cuMemcpyDtoD
  8130. .p2align 4
  8131. .type cuMemcpyDtoD, %function
  8132. cuMemcpyDtoD:
  8133. .cfi_startproc
  8134. // Intel opt. manual says to
  8135. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8136. // to hint static predictor.
  8137. cmpq $0, _libcuda_so_tramp_table+2296(%rip)
  8138. je 2f
  8139. 1:
  8140. jmp *_libcuda_so_tramp_table+2296(%rip)
  8141. 2:
  8142. pushq $287
  8143. .cfi_adjust_cfa_offset 8
  8144. call _libcuda_so_save_regs_and_resolve
  8145. addq $8, %rsp
  8146. .cfi_adjust_cfa_offset -8
  8147. jmp 1b
  8148. .cfi_endproc
  8149. /*
  8150. * Copyright 2018-2019 Yury Gribov
  8151. *
  8152. * The MIT License (MIT)
  8153. *
  8154. * Use of this source code is governed by MIT license that can be
  8155. * found in the LICENSE.txt file.
  8156. */
  8157. .globl cuMemcpyDtoDAsync
  8158. .p2align 4
  8159. .type cuMemcpyDtoDAsync, %function
  8160. cuMemcpyDtoDAsync:
  8161. .cfi_startproc
  8162. // Intel opt. manual says to
  8163. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8164. // to hint static predictor.
  8165. cmpq $0, _libcuda_so_tramp_table+2304(%rip)
  8166. je 2f
  8167. 1:
  8168. jmp *_libcuda_so_tramp_table+2304(%rip)
  8169. 2:
  8170. pushq $288
  8171. .cfi_adjust_cfa_offset 8
  8172. call _libcuda_so_save_regs_and_resolve
  8173. addq $8, %rsp
  8174. .cfi_adjust_cfa_offset -8
  8175. jmp 1b
  8176. .cfi_endproc
  8177. /*
  8178. * Copyright 2018-2019 Yury Gribov
  8179. *
  8180. * The MIT License (MIT)
  8181. *
  8182. * Use of this source code is governed by MIT license that can be
  8183. * found in the LICENSE.txt file.
  8184. */
  8185. .globl cuMemcpyDtoDAsync_v2
  8186. .p2align 4
  8187. .type cuMemcpyDtoDAsync_v2, %function
  8188. cuMemcpyDtoDAsync_v2:
  8189. .cfi_startproc
  8190. // Intel opt. manual says to
  8191. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8192. // to hint static predictor.
  8193. cmpq $0, _libcuda_so_tramp_table+2312(%rip)
  8194. je 2f
  8195. 1:
  8196. jmp *_libcuda_so_tramp_table+2312(%rip)
  8197. 2:
  8198. pushq $289
  8199. .cfi_adjust_cfa_offset 8
  8200. call _libcuda_so_save_regs_and_resolve
  8201. addq $8, %rsp
  8202. .cfi_adjust_cfa_offset -8
  8203. jmp 1b
  8204. .cfi_endproc
  8205. /*
  8206. * Copyright 2018-2019 Yury Gribov
  8207. *
  8208. * The MIT License (MIT)
  8209. *
  8210. * Use of this source code is governed by MIT license that can be
  8211. * found in the LICENSE.txt file.
  8212. */
  8213. .globl cuMemcpyDtoDAsync_v2_ptsz
  8214. .p2align 4
  8215. .type cuMemcpyDtoDAsync_v2_ptsz, %function
  8216. cuMemcpyDtoDAsync_v2_ptsz:
  8217. .cfi_startproc
  8218. // Intel opt. manual says to
  8219. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8220. // to hint static predictor.
  8221. cmpq $0, _libcuda_so_tramp_table+2320(%rip)
  8222. je 2f
  8223. 1:
  8224. jmp *_libcuda_so_tramp_table+2320(%rip)
  8225. 2:
  8226. pushq $290
  8227. .cfi_adjust_cfa_offset 8
  8228. call _libcuda_so_save_regs_and_resolve
  8229. addq $8, %rsp
  8230. .cfi_adjust_cfa_offset -8
  8231. jmp 1b
  8232. .cfi_endproc
  8233. /*
  8234. * Copyright 2018-2019 Yury Gribov
  8235. *
  8236. * The MIT License (MIT)
  8237. *
  8238. * Use of this source code is governed by MIT license that can be
  8239. * found in the LICENSE.txt file.
  8240. */
  8241. .globl cuMemcpyDtoD_v2
  8242. .p2align 4
  8243. .type cuMemcpyDtoD_v2, %function
  8244. cuMemcpyDtoD_v2:
  8245. .cfi_startproc
  8246. // Intel opt. manual says to
  8247. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8248. // to hint static predictor.
  8249. cmpq $0, _libcuda_so_tramp_table+2328(%rip)
  8250. je 2f
  8251. 1:
  8252. jmp *_libcuda_so_tramp_table+2328(%rip)
  8253. 2:
  8254. pushq $291
  8255. .cfi_adjust_cfa_offset 8
  8256. call _libcuda_so_save_regs_and_resolve
  8257. addq $8, %rsp
  8258. .cfi_adjust_cfa_offset -8
  8259. jmp 1b
  8260. .cfi_endproc
  8261. /*
  8262. * Copyright 2018-2019 Yury Gribov
  8263. *
  8264. * The MIT License (MIT)
  8265. *
  8266. * Use of this source code is governed by MIT license that can be
  8267. * found in the LICENSE.txt file.
  8268. */
  8269. .globl cuMemcpyDtoD_v2_ptds
  8270. .p2align 4
  8271. .type cuMemcpyDtoD_v2_ptds, %function
  8272. cuMemcpyDtoD_v2_ptds:
  8273. .cfi_startproc
  8274. // Intel opt. manual says to
  8275. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8276. // to hint static predictor.
  8277. cmpq $0, _libcuda_so_tramp_table+2336(%rip)
  8278. je 2f
  8279. 1:
  8280. jmp *_libcuda_so_tramp_table+2336(%rip)
  8281. 2:
  8282. pushq $292
  8283. .cfi_adjust_cfa_offset 8
  8284. call _libcuda_so_save_regs_and_resolve
  8285. addq $8, %rsp
  8286. .cfi_adjust_cfa_offset -8
  8287. jmp 1b
  8288. .cfi_endproc
  8289. /*
  8290. * Copyright 2018-2019 Yury Gribov
  8291. *
  8292. * The MIT License (MIT)
  8293. *
  8294. * Use of this source code is governed by MIT license that can be
  8295. * found in the LICENSE.txt file.
  8296. */
  8297. .globl cuMemcpyDtoH
  8298. .p2align 4
  8299. .type cuMemcpyDtoH, %function
  8300. cuMemcpyDtoH:
  8301. .cfi_startproc
  8302. // Intel opt. manual says to
  8303. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8304. // to hint static predictor.
  8305. cmpq $0, _libcuda_so_tramp_table+2344(%rip)
  8306. je 2f
  8307. 1:
  8308. jmp *_libcuda_so_tramp_table+2344(%rip)
  8309. 2:
  8310. pushq $293
  8311. .cfi_adjust_cfa_offset 8
  8312. call _libcuda_so_save_regs_and_resolve
  8313. addq $8, %rsp
  8314. .cfi_adjust_cfa_offset -8
  8315. jmp 1b
  8316. .cfi_endproc
  8317. /*
  8318. * Copyright 2018-2019 Yury Gribov
  8319. *
  8320. * The MIT License (MIT)
  8321. *
  8322. * Use of this source code is governed by MIT license that can be
  8323. * found in the LICENSE.txt file.
  8324. */
  8325. .globl cuMemcpyDtoHAsync
  8326. .p2align 4
  8327. .type cuMemcpyDtoHAsync, %function
  8328. cuMemcpyDtoHAsync:
  8329. .cfi_startproc
  8330. // Intel opt. manual says to
  8331. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8332. // to hint static predictor.
  8333. cmpq $0, _libcuda_so_tramp_table+2352(%rip)
  8334. je 2f
  8335. 1:
  8336. jmp *_libcuda_so_tramp_table+2352(%rip)
  8337. 2:
  8338. pushq $294
  8339. .cfi_adjust_cfa_offset 8
  8340. call _libcuda_so_save_regs_and_resolve
  8341. addq $8, %rsp
  8342. .cfi_adjust_cfa_offset -8
  8343. jmp 1b
  8344. .cfi_endproc
  8345. /*
  8346. * Copyright 2018-2019 Yury Gribov
  8347. *
  8348. * The MIT License (MIT)
  8349. *
  8350. * Use of this source code is governed by MIT license that can be
  8351. * found in the LICENSE.txt file.
  8352. */
  8353. .globl cuMemcpyDtoHAsync_v2
  8354. .p2align 4
  8355. .type cuMemcpyDtoHAsync_v2, %function
  8356. cuMemcpyDtoHAsync_v2:
  8357. .cfi_startproc
  8358. // Intel opt. manual says to
  8359. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8360. // to hint static predictor.
  8361. cmpq $0, _libcuda_so_tramp_table+2360(%rip)
  8362. je 2f
  8363. 1:
  8364. jmp *_libcuda_so_tramp_table+2360(%rip)
  8365. 2:
  8366. pushq $295
  8367. .cfi_adjust_cfa_offset 8
  8368. call _libcuda_so_save_regs_and_resolve
  8369. addq $8, %rsp
  8370. .cfi_adjust_cfa_offset -8
  8371. jmp 1b
  8372. .cfi_endproc
  8373. /*
  8374. * Copyright 2018-2019 Yury Gribov
  8375. *
  8376. * The MIT License (MIT)
  8377. *
  8378. * Use of this source code is governed by MIT license that can be
  8379. * found in the LICENSE.txt file.
  8380. */
  8381. .globl cuMemcpyDtoHAsync_v2_ptsz
  8382. .p2align 4
  8383. .type cuMemcpyDtoHAsync_v2_ptsz, %function
  8384. cuMemcpyDtoHAsync_v2_ptsz:
  8385. .cfi_startproc
  8386. // Intel opt. manual says to
  8387. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8388. // to hint static predictor.
  8389. cmpq $0, _libcuda_so_tramp_table+2368(%rip)
  8390. je 2f
  8391. 1:
  8392. jmp *_libcuda_so_tramp_table+2368(%rip)
  8393. 2:
  8394. pushq $296
  8395. .cfi_adjust_cfa_offset 8
  8396. call _libcuda_so_save_regs_and_resolve
  8397. addq $8, %rsp
  8398. .cfi_adjust_cfa_offset -8
  8399. jmp 1b
  8400. .cfi_endproc
  8401. /*
  8402. * Copyright 2018-2019 Yury Gribov
  8403. *
  8404. * The MIT License (MIT)
  8405. *
  8406. * Use of this source code is governed by MIT license that can be
  8407. * found in the LICENSE.txt file.
  8408. */
  8409. .globl cuMemcpyDtoH_v2
  8410. .p2align 4
  8411. .type cuMemcpyDtoH_v2, %function
  8412. cuMemcpyDtoH_v2:
  8413. .cfi_startproc
  8414. // Intel opt. manual says to
  8415. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8416. // to hint static predictor.
  8417. cmpq $0, _libcuda_so_tramp_table+2376(%rip)
  8418. je 2f
  8419. 1:
  8420. jmp *_libcuda_so_tramp_table+2376(%rip)
  8421. 2:
  8422. pushq $297
  8423. .cfi_adjust_cfa_offset 8
  8424. call _libcuda_so_save_regs_and_resolve
  8425. addq $8, %rsp
  8426. .cfi_adjust_cfa_offset -8
  8427. jmp 1b
  8428. .cfi_endproc
  8429. /*
  8430. * Copyright 2018-2019 Yury Gribov
  8431. *
  8432. * The MIT License (MIT)
  8433. *
  8434. * Use of this source code is governed by MIT license that can be
  8435. * found in the LICENSE.txt file.
  8436. */
  8437. .globl cuMemcpyDtoH_v2_ptds
  8438. .p2align 4
  8439. .type cuMemcpyDtoH_v2_ptds, %function
  8440. cuMemcpyDtoH_v2_ptds:
  8441. .cfi_startproc
  8442. // Intel opt. manual says to
  8443. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8444. // to hint static predictor.
  8445. cmpq $0, _libcuda_so_tramp_table+2384(%rip)
  8446. je 2f
  8447. 1:
  8448. jmp *_libcuda_so_tramp_table+2384(%rip)
  8449. 2:
  8450. pushq $298
  8451. .cfi_adjust_cfa_offset 8
  8452. call _libcuda_so_save_regs_and_resolve
  8453. addq $8, %rsp
  8454. .cfi_adjust_cfa_offset -8
  8455. jmp 1b
  8456. .cfi_endproc
  8457. /*
  8458. * Copyright 2018-2019 Yury Gribov
  8459. *
  8460. * The MIT License (MIT)
  8461. *
  8462. * Use of this source code is governed by MIT license that can be
  8463. * found in the LICENSE.txt file.
  8464. */
  8465. .globl cuMemcpyHtoA
  8466. .p2align 4
  8467. .type cuMemcpyHtoA, %function
  8468. cuMemcpyHtoA:
  8469. .cfi_startproc
  8470. // Intel opt. manual says to
  8471. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8472. // to hint static predictor.
  8473. cmpq $0, _libcuda_so_tramp_table+2392(%rip)
  8474. je 2f
  8475. 1:
  8476. jmp *_libcuda_so_tramp_table+2392(%rip)
  8477. 2:
  8478. pushq $299
  8479. .cfi_adjust_cfa_offset 8
  8480. call _libcuda_so_save_regs_and_resolve
  8481. addq $8, %rsp
  8482. .cfi_adjust_cfa_offset -8
  8483. jmp 1b
  8484. .cfi_endproc
  8485. /*
  8486. * Copyright 2018-2019 Yury Gribov
  8487. *
  8488. * The MIT License (MIT)
  8489. *
  8490. * Use of this source code is governed by MIT license that can be
  8491. * found in the LICENSE.txt file.
  8492. */
  8493. .globl cuMemcpyHtoAAsync
  8494. .p2align 4
  8495. .type cuMemcpyHtoAAsync, %function
  8496. cuMemcpyHtoAAsync:
  8497. .cfi_startproc
  8498. // Intel opt. manual says to
  8499. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8500. // to hint static predictor.
  8501. cmpq $0, _libcuda_so_tramp_table+2400(%rip)
  8502. je 2f
  8503. 1:
  8504. jmp *_libcuda_so_tramp_table+2400(%rip)
  8505. 2:
  8506. pushq $300
  8507. .cfi_adjust_cfa_offset 8
  8508. call _libcuda_so_save_regs_and_resolve
  8509. addq $8, %rsp
  8510. .cfi_adjust_cfa_offset -8
  8511. jmp 1b
  8512. .cfi_endproc
  8513. /*
  8514. * Copyright 2018-2019 Yury Gribov
  8515. *
  8516. * The MIT License (MIT)
  8517. *
  8518. * Use of this source code is governed by MIT license that can be
  8519. * found in the LICENSE.txt file.
  8520. */
  8521. .globl cuMemcpyHtoAAsync_v2
  8522. .p2align 4
  8523. .type cuMemcpyHtoAAsync_v2, %function
  8524. cuMemcpyHtoAAsync_v2:
  8525. .cfi_startproc
  8526. // Intel opt. manual says to
  8527. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8528. // to hint static predictor.
  8529. cmpq $0, _libcuda_so_tramp_table+2408(%rip)
  8530. je 2f
  8531. 1:
  8532. jmp *_libcuda_so_tramp_table+2408(%rip)
  8533. 2:
  8534. pushq $301
  8535. .cfi_adjust_cfa_offset 8
  8536. call _libcuda_so_save_regs_and_resolve
  8537. addq $8, %rsp
  8538. .cfi_adjust_cfa_offset -8
  8539. jmp 1b
  8540. .cfi_endproc
  8541. /*
  8542. * Copyright 2018-2019 Yury Gribov
  8543. *
  8544. * The MIT License (MIT)
  8545. *
  8546. * Use of this source code is governed by MIT license that can be
  8547. * found in the LICENSE.txt file.
  8548. */
  8549. .globl cuMemcpyHtoAAsync_v2_ptsz
  8550. .p2align 4
  8551. .type cuMemcpyHtoAAsync_v2_ptsz, %function
  8552. cuMemcpyHtoAAsync_v2_ptsz:
  8553. .cfi_startproc
  8554. // Intel opt. manual says to
  8555. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8556. // to hint static predictor.
  8557. cmpq $0, _libcuda_so_tramp_table+2416(%rip)
  8558. je 2f
  8559. 1:
  8560. jmp *_libcuda_so_tramp_table+2416(%rip)
  8561. 2:
  8562. pushq $302
  8563. .cfi_adjust_cfa_offset 8
  8564. call _libcuda_so_save_regs_and_resolve
  8565. addq $8, %rsp
  8566. .cfi_adjust_cfa_offset -8
  8567. jmp 1b
  8568. .cfi_endproc
  8569. /*
  8570. * Copyright 2018-2019 Yury Gribov
  8571. *
  8572. * The MIT License (MIT)
  8573. *
  8574. * Use of this source code is governed by MIT license that can be
  8575. * found in the LICENSE.txt file.
  8576. */
  8577. .globl cuMemcpyHtoA_v2
  8578. .p2align 4
  8579. .type cuMemcpyHtoA_v2, %function
  8580. cuMemcpyHtoA_v2:
  8581. .cfi_startproc
  8582. // Intel opt. manual says to
  8583. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8584. // to hint static predictor.
  8585. cmpq $0, _libcuda_so_tramp_table+2424(%rip)
  8586. je 2f
  8587. 1:
  8588. jmp *_libcuda_so_tramp_table+2424(%rip)
  8589. 2:
  8590. pushq $303
  8591. .cfi_adjust_cfa_offset 8
  8592. call _libcuda_so_save_regs_and_resolve
  8593. addq $8, %rsp
  8594. .cfi_adjust_cfa_offset -8
  8595. jmp 1b
  8596. .cfi_endproc
  8597. /*
  8598. * Copyright 2018-2019 Yury Gribov
  8599. *
  8600. * The MIT License (MIT)
  8601. *
  8602. * Use of this source code is governed by MIT license that can be
  8603. * found in the LICENSE.txt file.
  8604. */
  8605. .globl cuMemcpyHtoA_v2_ptds
  8606. .p2align 4
  8607. .type cuMemcpyHtoA_v2_ptds, %function
  8608. cuMemcpyHtoA_v2_ptds:
  8609. .cfi_startproc
  8610. // Intel opt. manual says to
  8611. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8612. // to hint static predictor.
  8613. cmpq $0, _libcuda_so_tramp_table+2432(%rip)
  8614. je 2f
  8615. 1:
  8616. jmp *_libcuda_so_tramp_table+2432(%rip)
  8617. 2:
  8618. pushq $304
  8619. .cfi_adjust_cfa_offset 8
  8620. call _libcuda_so_save_regs_and_resolve
  8621. addq $8, %rsp
  8622. .cfi_adjust_cfa_offset -8
  8623. jmp 1b
  8624. .cfi_endproc
  8625. /*
  8626. * Copyright 2018-2019 Yury Gribov
  8627. *
  8628. * The MIT License (MIT)
  8629. *
  8630. * Use of this source code is governed by MIT license that can be
  8631. * found in the LICENSE.txt file.
  8632. */
  8633. .globl cuMemcpyHtoD
  8634. .p2align 4
  8635. .type cuMemcpyHtoD, %function
  8636. cuMemcpyHtoD:
  8637. .cfi_startproc
  8638. // Intel opt. manual says to
  8639. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8640. // to hint static predictor.
  8641. cmpq $0, _libcuda_so_tramp_table+2440(%rip)
  8642. je 2f
  8643. 1:
  8644. jmp *_libcuda_so_tramp_table+2440(%rip)
  8645. 2:
  8646. pushq $305
  8647. .cfi_adjust_cfa_offset 8
  8648. call _libcuda_so_save_regs_and_resolve
  8649. addq $8, %rsp
  8650. .cfi_adjust_cfa_offset -8
  8651. jmp 1b
  8652. .cfi_endproc
  8653. /*
  8654. * Copyright 2018-2019 Yury Gribov
  8655. *
  8656. * The MIT License (MIT)
  8657. *
  8658. * Use of this source code is governed by MIT license that can be
  8659. * found in the LICENSE.txt file.
  8660. */
  8661. .globl cuMemcpyHtoDAsync
  8662. .p2align 4
  8663. .type cuMemcpyHtoDAsync, %function
  8664. cuMemcpyHtoDAsync:
  8665. .cfi_startproc
  8666. // Intel opt. manual says to
  8667. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8668. // to hint static predictor.
  8669. cmpq $0, _libcuda_so_tramp_table+2448(%rip)
  8670. je 2f
  8671. 1:
  8672. jmp *_libcuda_so_tramp_table+2448(%rip)
  8673. 2:
  8674. pushq $306
  8675. .cfi_adjust_cfa_offset 8
  8676. call _libcuda_so_save_regs_and_resolve
  8677. addq $8, %rsp
  8678. .cfi_adjust_cfa_offset -8
  8679. jmp 1b
  8680. .cfi_endproc
  8681. /*
  8682. * Copyright 2018-2019 Yury Gribov
  8683. *
  8684. * The MIT License (MIT)
  8685. *
  8686. * Use of this source code is governed by MIT license that can be
  8687. * found in the LICENSE.txt file.
  8688. */
  8689. .globl cuMemcpyHtoDAsync_v2
  8690. .p2align 4
  8691. .type cuMemcpyHtoDAsync_v2, %function
  8692. cuMemcpyHtoDAsync_v2:
  8693. .cfi_startproc
  8694. // Intel opt. manual says to
  8695. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8696. // to hint static predictor.
  8697. cmpq $0, _libcuda_so_tramp_table+2456(%rip)
  8698. je 2f
  8699. 1:
  8700. jmp *_libcuda_so_tramp_table+2456(%rip)
  8701. 2:
  8702. pushq $307
  8703. .cfi_adjust_cfa_offset 8
  8704. call _libcuda_so_save_regs_and_resolve
  8705. addq $8, %rsp
  8706. .cfi_adjust_cfa_offset -8
  8707. jmp 1b
  8708. .cfi_endproc
  8709. /*
  8710. * Copyright 2018-2019 Yury Gribov
  8711. *
  8712. * The MIT License (MIT)
  8713. *
  8714. * Use of this source code is governed by MIT license that can be
  8715. * found in the LICENSE.txt file.
  8716. */
  8717. .globl cuMemcpyHtoDAsync_v2_ptsz
  8718. .p2align 4
  8719. .type cuMemcpyHtoDAsync_v2_ptsz, %function
  8720. cuMemcpyHtoDAsync_v2_ptsz:
  8721. .cfi_startproc
  8722. // Intel opt. manual says to
  8723. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8724. // to hint static predictor.
  8725. cmpq $0, _libcuda_so_tramp_table+2464(%rip)
  8726. je 2f
  8727. 1:
  8728. jmp *_libcuda_so_tramp_table+2464(%rip)
  8729. 2:
  8730. pushq $308
  8731. .cfi_adjust_cfa_offset 8
  8732. call _libcuda_so_save_regs_and_resolve
  8733. addq $8, %rsp
  8734. .cfi_adjust_cfa_offset -8
  8735. jmp 1b
  8736. .cfi_endproc
  8737. /*
  8738. * Copyright 2018-2019 Yury Gribov
  8739. *
  8740. * The MIT License (MIT)
  8741. *
  8742. * Use of this source code is governed by MIT license that can be
  8743. * found in the LICENSE.txt file.
  8744. */
  8745. .globl cuMemcpyHtoD_v2
  8746. .p2align 4
  8747. .type cuMemcpyHtoD_v2, %function
  8748. cuMemcpyHtoD_v2:
  8749. .cfi_startproc
  8750. // Intel opt. manual says to
  8751. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8752. // to hint static predictor.
  8753. cmpq $0, _libcuda_so_tramp_table+2472(%rip)
  8754. je 2f
  8755. 1:
  8756. jmp *_libcuda_so_tramp_table+2472(%rip)
  8757. 2:
  8758. pushq $309
  8759. .cfi_adjust_cfa_offset 8
  8760. call _libcuda_so_save_regs_and_resolve
  8761. addq $8, %rsp
  8762. .cfi_adjust_cfa_offset -8
  8763. jmp 1b
  8764. .cfi_endproc
  8765. /*
  8766. * Copyright 2018-2019 Yury Gribov
  8767. *
  8768. * The MIT License (MIT)
  8769. *
  8770. * Use of this source code is governed by MIT license that can be
  8771. * found in the LICENSE.txt file.
  8772. */
  8773. .globl cuMemcpyHtoD_v2_ptds
  8774. .p2align 4
  8775. .type cuMemcpyHtoD_v2_ptds, %function
  8776. cuMemcpyHtoD_v2_ptds:
  8777. .cfi_startproc
  8778. // Intel opt. manual says to
  8779. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8780. // to hint static predictor.
  8781. cmpq $0, _libcuda_so_tramp_table+2480(%rip)
  8782. je 2f
  8783. 1:
  8784. jmp *_libcuda_so_tramp_table+2480(%rip)
  8785. 2:
  8786. pushq $310
  8787. .cfi_adjust_cfa_offset 8
  8788. call _libcuda_so_save_regs_and_resolve
  8789. addq $8, %rsp
  8790. .cfi_adjust_cfa_offset -8
  8791. jmp 1b
  8792. .cfi_endproc
  8793. /*
  8794. * Copyright 2018-2019 Yury Gribov
  8795. *
  8796. * The MIT License (MIT)
  8797. *
  8798. * Use of this source code is governed by MIT license that can be
  8799. * found in the LICENSE.txt file.
  8800. */
  8801. .globl cuMemcpyPeer
  8802. .p2align 4
  8803. .type cuMemcpyPeer, %function
  8804. cuMemcpyPeer:
  8805. .cfi_startproc
  8806. // Intel opt. manual says to
  8807. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8808. // to hint static predictor.
  8809. cmpq $0, _libcuda_so_tramp_table+2488(%rip)
  8810. je 2f
  8811. 1:
  8812. jmp *_libcuda_so_tramp_table+2488(%rip)
  8813. 2:
  8814. pushq $311
  8815. .cfi_adjust_cfa_offset 8
  8816. call _libcuda_so_save_regs_and_resolve
  8817. addq $8, %rsp
  8818. .cfi_adjust_cfa_offset -8
  8819. jmp 1b
  8820. .cfi_endproc
  8821. /*
  8822. * Copyright 2018-2019 Yury Gribov
  8823. *
  8824. * The MIT License (MIT)
  8825. *
  8826. * Use of this source code is governed by MIT license that can be
  8827. * found in the LICENSE.txt file.
  8828. */
  8829. .globl cuMemcpyPeerAsync
  8830. .p2align 4
  8831. .type cuMemcpyPeerAsync, %function
  8832. cuMemcpyPeerAsync:
  8833. .cfi_startproc
  8834. // Intel opt. manual says to
  8835. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8836. // to hint static predictor.
  8837. cmpq $0, _libcuda_so_tramp_table+2496(%rip)
  8838. je 2f
  8839. 1:
  8840. jmp *_libcuda_so_tramp_table+2496(%rip)
  8841. 2:
  8842. pushq $312
  8843. .cfi_adjust_cfa_offset 8
  8844. call _libcuda_so_save_regs_and_resolve
  8845. addq $8, %rsp
  8846. .cfi_adjust_cfa_offset -8
  8847. jmp 1b
  8848. .cfi_endproc
  8849. /*
  8850. * Copyright 2018-2019 Yury Gribov
  8851. *
  8852. * The MIT License (MIT)
  8853. *
  8854. * Use of this source code is governed by MIT license that can be
  8855. * found in the LICENSE.txt file.
  8856. */
  8857. .globl cuMemcpyPeerAsync_ptsz
  8858. .p2align 4
  8859. .type cuMemcpyPeerAsync_ptsz, %function
  8860. cuMemcpyPeerAsync_ptsz:
  8861. .cfi_startproc
  8862. // Intel opt. manual says to
  8863. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8864. // to hint static predictor.
  8865. cmpq $0, _libcuda_so_tramp_table+2504(%rip)
  8866. je 2f
  8867. 1:
  8868. jmp *_libcuda_so_tramp_table+2504(%rip)
  8869. 2:
  8870. pushq $313
  8871. .cfi_adjust_cfa_offset 8
  8872. call _libcuda_so_save_regs_and_resolve
  8873. addq $8, %rsp
  8874. .cfi_adjust_cfa_offset -8
  8875. jmp 1b
  8876. .cfi_endproc
  8877. /*
  8878. * Copyright 2018-2019 Yury Gribov
  8879. *
  8880. * The MIT License (MIT)
  8881. *
  8882. * Use of this source code is governed by MIT license that can be
  8883. * found in the LICENSE.txt file.
  8884. */
  8885. .globl cuMemcpyPeer_ptds
  8886. .p2align 4
  8887. .type cuMemcpyPeer_ptds, %function
  8888. cuMemcpyPeer_ptds:
  8889. .cfi_startproc
  8890. // Intel opt. manual says to
  8891. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8892. // to hint static predictor.
  8893. cmpq $0, _libcuda_so_tramp_table+2512(%rip)
  8894. je 2f
  8895. 1:
  8896. jmp *_libcuda_so_tramp_table+2512(%rip)
  8897. 2:
  8898. pushq $314
  8899. .cfi_adjust_cfa_offset 8
  8900. call _libcuda_so_save_regs_and_resolve
  8901. addq $8, %rsp
  8902. .cfi_adjust_cfa_offset -8
  8903. jmp 1b
  8904. .cfi_endproc
  8905. /*
  8906. * Copyright 2018-2019 Yury Gribov
  8907. *
  8908. * The MIT License (MIT)
  8909. *
  8910. * Use of this source code is governed by MIT license that can be
  8911. * found in the LICENSE.txt file.
  8912. */
  8913. .globl cuMemcpy_ptds
  8914. .p2align 4
  8915. .type cuMemcpy_ptds, %function
  8916. cuMemcpy_ptds:
  8917. .cfi_startproc
  8918. // Intel opt. manual says to
  8919. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8920. // to hint static predictor.
  8921. cmpq $0, _libcuda_so_tramp_table+2520(%rip)
  8922. je 2f
  8923. 1:
  8924. jmp *_libcuda_so_tramp_table+2520(%rip)
  8925. 2:
  8926. pushq $315
  8927. .cfi_adjust_cfa_offset 8
  8928. call _libcuda_so_save_regs_and_resolve
  8929. addq $8, %rsp
  8930. .cfi_adjust_cfa_offset -8
  8931. jmp 1b
  8932. .cfi_endproc
  8933. /*
  8934. * Copyright 2018-2019 Yury Gribov
  8935. *
  8936. * The MIT License (MIT)
  8937. *
  8938. * Use of this source code is governed by MIT license that can be
  8939. * found in the LICENSE.txt file.
  8940. */
  8941. .globl cuMemsetD16
  8942. .p2align 4
  8943. .type cuMemsetD16, %function
  8944. cuMemsetD16:
  8945. .cfi_startproc
  8946. // Intel opt. manual says to
  8947. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8948. // to hint static predictor.
  8949. cmpq $0, _libcuda_so_tramp_table+2528(%rip)
  8950. je 2f
  8951. 1:
  8952. jmp *_libcuda_so_tramp_table+2528(%rip)
  8953. 2:
  8954. pushq $316
  8955. .cfi_adjust_cfa_offset 8
  8956. call _libcuda_so_save_regs_and_resolve
  8957. addq $8, %rsp
  8958. .cfi_adjust_cfa_offset -8
  8959. jmp 1b
  8960. .cfi_endproc
  8961. /*
  8962. * Copyright 2018-2019 Yury Gribov
  8963. *
  8964. * The MIT License (MIT)
  8965. *
  8966. * Use of this source code is governed by MIT license that can be
  8967. * found in the LICENSE.txt file.
  8968. */
  8969. .globl cuMemsetD16Async
  8970. .p2align 4
  8971. .type cuMemsetD16Async, %function
  8972. cuMemsetD16Async:
  8973. .cfi_startproc
  8974. // Intel opt. manual says to
  8975. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  8976. // to hint static predictor.
  8977. cmpq $0, _libcuda_so_tramp_table+2536(%rip)
  8978. je 2f
  8979. 1:
  8980. jmp *_libcuda_so_tramp_table+2536(%rip)
  8981. 2:
  8982. pushq $317
  8983. .cfi_adjust_cfa_offset 8
  8984. call _libcuda_so_save_regs_and_resolve
  8985. addq $8, %rsp
  8986. .cfi_adjust_cfa_offset -8
  8987. jmp 1b
  8988. .cfi_endproc
  8989. /*
  8990. * Copyright 2018-2019 Yury Gribov
  8991. *
  8992. * The MIT License (MIT)
  8993. *
  8994. * Use of this source code is governed by MIT license that can be
  8995. * found in the LICENSE.txt file.
  8996. */
  8997. .globl cuMemsetD16Async_ptsz
  8998. .p2align 4
  8999. .type cuMemsetD16Async_ptsz, %function
  9000. cuMemsetD16Async_ptsz:
  9001. .cfi_startproc
  9002. // Intel opt. manual says to
  9003. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9004. // to hint static predictor.
  9005. cmpq $0, _libcuda_so_tramp_table+2544(%rip)
  9006. je 2f
  9007. 1:
  9008. jmp *_libcuda_so_tramp_table+2544(%rip)
  9009. 2:
  9010. pushq $318
  9011. .cfi_adjust_cfa_offset 8
  9012. call _libcuda_so_save_regs_and_resolve
  9013. addq $8, %rsp
  9014. .cfi_adjust_cfa_offset -8
  9015. jmp 1b
  9016. .cfi_endproc
  9017. /*
  9018. * Copyright 2018-2019 Yury Gribov
  9019. *
  9020. * The MIT License (MIT)
  9021. *
  9022. * Use of this source code is governed by MIT license that can be
  9023. * found in the LICENSE.txt file.
  9024. */
  9025. .globl cuMemsetD16_v2
  9026. .p2align 4
  9027. .type cuMemsetD16_v2, %function
  9028. cuMemsetD16_v2:
  9029. .cfi_startproc
  9030. // Intel opt. manual says to
  9031. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9032. // to hint static predictor.
  9033. cmpq $0, _libcuda_so_tramp_table+2552(%rip)
  9034. je 2f
  9035. 1:
  9036. jmp *_libcuda_so_tramp_table+2552(%rip)
  9037. 2:
  9038. pushq $319
  9039. .cfi_adjust_cfa_offset 8
  9040. call _libcuda_so_save_regs_and_resolve
  9041. addq $8, %rsp
  9042. .cfi_adjust_cfa_offset -8
  9043. jmp 1b
  9044. .cfi_endproc
  9045. /*
  9046. * Copyright 2018-2019 Yury Gribov
  9047. *
  9048. * The MIT License (MIT)
  9049. *
  9050. * Use of this source code is governed by MIT license that can be
  9051. * found in the LICENSE.txt file.
  9052. */
  9053. .globl cuMemsetD16_v2_ptds
  9054. .p2align 4
  9055. .type cuMemsetD16_v2_ptds, %function
  9056. cuMemsetD16_v2_ptds:
  9057. .cfi_startproc
  9058. // Intel opt. manual says to
  9059. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9060. // to hint static predictor.
  9061. cmpq $0, _libcuda_so_tramp_table+2560(%rip)
  9062. je 2f
  9063. 1:
  9064. jmp *_libcuda_so_tramp_table+2560(%rip)
  9065. 2:
  9066. pushq $320
  9067. .cfi_adjust_cfa_offset 8
  9068. call _libcuda_so_save_regs_and_resolve
  9069. addq $8, %rsp
  9070. .cfi_adjust_cfa_offset -8
  9071. jmp 1b
  9072. .cfi_endproc
  9073. /*
  9074. * Copyright 2018-2019 Yury Gribov
  9075. *
  9076. * The MIT License (MIT)
  9077. *
  9078. * Use of this source code is governed by MIT license that can be
  9079. * found in the LICENSE.txt file.
  9080. */
  9081. .globl cuMemsetD2D16
  9082. .p2align 4
  9083. .type cuMemsetD2D16, %function
  9084. cuMemsetD2D16:
  9085. .cfi_startproc
  9086. // Intel opt. manual says to
  9087. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9088. // to hint static predictor.
  9089. cmpq $0, _libcuda_so_tramp_table+2568(%rip)
  9090. je 2f
  9091. 1:
  9092. jmp *_libcuda_so_tramp_table+2568(%rip)
  9093. 2:
  9094. pushq $321
  9095. .cfi_adjust_cfa_offset 8
  9096. call _libcuda_so_save_regs_and_resolve
  9097. addq $8, %rsp
  9098. .cfi_adjust_cfa_offset -8
  9099. jmp 1b
  9100. .cfi_endproc
  9101. /*
  9102. * Copyright 2018-2019 Yury Gribov
  9103. *
  9104. * The MIT License (MIT)
  9105. *
  9106. * Use of this source code is governed by MIT license that can be
  9107. * found in the LICENSE.txt file.
  9108. */
  9109. .globl cuMemsetD2D16Async
  9110. .p2align 4
  9111. .type cuMemsetD2D16Async, %function
  9112. cuMemsetD2D16Async:
  9113. .cfi_startproc
  9114. // Intel opt. manual says to
  9115. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9116. // to hint static predictor.
  9117. cmpq $0, _libcuda_so_tramp_table+2576(%rip)
  9118. je 2f
  9119. 1:
  9120. jmp *_libcuda_so_tramp_table+2576(%rip)
  9121. 2:
  9122. pushq $322
  9123. .cfi_adjust_cfa_offset 8
  9124. call _libcuda_so_save_regs_and_resolve
  9125. addq $8, %rsp
  9126. .cfi_adjust_cfa_offset -8
  9127. jmp 1b
  9128. .cfi_endproc
  9129. /*
  9130. * Copyright 2018-2019 Yury Gribov
  9131. *
  9132. * The MIT License (MIT)
  9133. *
  9134. * Use of this source code is governed by MIT license that can be
  9135. * found in the LICENSE.txt file.
  9136. */
  9137. .globl cuMemsetD2D16Async_ptsz
  9138. .p2align 4
  9139. .type cuMemsetD2D16Async_ptsz, %function
  9140. cuMemsetD2D16Async_ptsz:
  9141. .cfi_startproc
  9142. // Intel opt. manual says to
  9143. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9144. // to hint static predictor.
  9145. cmpq $0, _libcuda_so_tramp_table+2584(%rip)
  9146. je 2f
  9147. 1:
  9148. jmp *_libcuda_so_tramp_table+2584(%rip)
  9149. 2:
  9150. pushq $323
  9151. .cfi_adjust_cfa_offset 8
  9152. call _libcuda_so_save_regs_and_resolve
  9153. addq $8, %rsp
  9154. .cfi_adjust_cfa_offset -8
  9155. jmp 1b
  9156. .cfi_endproc
  9157. /*
  9158. * Copyright 2018-2019 Yury Gribov
  9159. *
  9160. * The MIT License (MIT)
  9161. *
  9162. * Use of this source code is governed by MIT license that can be
  9163. * found in the LICENSE.txt file.
  9164. */
  9165. .globl cuMemsetD2D16_v2
  9166. .p2align 4
  9167. .type cuMemsetD2D16_v2, %function
  9168. cuMemsetD2D16_v2:
  9169. .cfi_startproc
  9170. // Intel opt. manual says to
  9171. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9172. // to hint static predictor.
  9173. cmpq $0, _libcuda_so_tramp_table+2592(%rip)
  9174. je 2f
  9175. 1:
  9176. jmp *_libcuda_so_tramp_table+2592(%rip)
  9177. 2:
  9178. pushq $324
  9179. .cfi_adjust_cfa_offset 8
  9180. call _libcuda_so_save_regs_and_resolve
  9181. addq $8, %rsp
  9182. .cfi_adjust_cfa_offset -8
  9183. jmp 1b
  9184. .cfi_endproc
  9185. /*
  9186. * Copyright 2018-2019 Yury Gribov
  9187. *
  9188. * The MIT License (MIT)
  9189. *
  9190. * Use of this source code is governed by MIT license that can be
  9191. * found in the LICENSE.txt file.
  9192. */
  9193. .globl cuMemsetD2D16_v2_ptds
  9194. .p2align 4
  9195. .type cuMemsetD2D16_v2_ptds, %function
  9196. cuMemsetD2D16_v2_ptds:
  9197. .cfi_startproc
  9198. // Intel opt. manual says to
  9199. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9200. // to hint static predictor.
  9201. cmpq $0, _libcuda_so_tramp_table+2600(%rip)
  9202. je 2f
  9203. 1:
  9204. jmp *_libcuda_so_tramp_table+2600(%rip)
  9205. 2:
  9206. pushq $325
  9207. .cfi_adjust_cfa_offset 8
  9208. call _libcuda_so_save_regs_and_resolve
  9209. addq $8, %rsp
  9210. .cfi_adjust_cfa_offset -8
  9211. jmp 1b
  9212. .cfi_endproc
  9213. /*
  9214. * Copyright 2018-2019 Yury Gribov
  9215. *
  9216. * The MIT License (MIT)
  9217. *
  9218. * Use of this source code is governed by MIT license that can be
  9219. * found in the LICENSE.txt file.
  9220. */
  9221. .globl cuMemsetD2D32
  9222. .p2align 4
  9223. .type cuMemsetD2D32, %function
  9224. cuMemsetD2D32:
  9225. .cfi_startproc
  9226. // Intel opt. manual says to
  9227. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9228. // to hint static predictor.
  9229. cmpq $0, _libcuda_so_tramp_table+2608(%rip)
  9230. je 2f
  9231. 1:
  9232. jmp *_libcuda_so_tramp_table+2608(%rip)
  9233. 2:
  9234. pushq $326
  9235. .cfi_adjust_cfa_offset 8
  9236. call _libcuda_so_save_regs_and_resolve
  9237. addq $8, %rsp
  9238. .cfi_adjust_cfa_offset -8
  9239. jmp 1b
  9240. .cfi_endproc
  9241. /*
  9242. * Copyright 2018-2019 Yury Gribov
  9243. *
  9244. * The MIT License (MIT)
  9245. *
  9246. * Use of this source code is governed by MIT license that can be
  9247. * found in the LICENSE.txt file.
  9248. */
  9249. .globl cuMemsetD2D32Async
  9250. .p2align 4
  9251. .type cuMemsetD2D32Async, %function
  9252. cuMemsetD2D32Async:
  9253. .cfi_startproc
  9254. // Intel opt. manual says to
  9255. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9256. // to hint static predictor.
  9257. cmpq $0, _libcuda_so_tramp_table+2616(%rip)
  9258. je 2f
  9259. 1:
  9260. jmp *_libcuda_so_tramp_table+2616(%rip)
  9261. 2:
  9262. pushq $327
  9263. .cfi_adjust_cfa_offset 8
  9264. call _libcuda_so_save_regs_and_resolve
  9265. addq $8, %rsp
  9266. .cfi_adjust_cfa_offset -8
  9267. jmp 1b
  9268. .cfi_endproc
  9269. /*
  9270. * Copyright 2018-2019 Yury Gribov
  9271. *
  9272. * The MIT License (MIT)
  9273. *
  9274. * Use of this source code is governed by MIT license that can be
  9275. * found in the LICENSE.txt file.
  9276. */
  9277. .globl cuMemsetD2D32Async_ptsz
  9278. .p2align 4
  9279. .type cuMemsetD2D32Async_ptsz, %function
  9280. cuMemsetD2D32Async_ptsz:
  9281. .cfi_startproc
  9282. // Intel opt. manual says to
  9283. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9284. // to hint static predictor.
  9285. cmpq $0, _libcuda_so_tramp_table+2624(%rip)
  9286. je 2f
  9287. 1:
  9288. jmp *_libcuda_so_tramp_table+2624(%rip)
  9289. 2:
  9290. pushq $328
  9291. .cfi_adjust_cfa_offset 8
  9292. call _libcuda_so_save_regs_and_resolve
  9293. addq $8, %rsp
  9294. .cfi_adjust_cfa_offset -8
  9295. jmp 1b
  9296. .cfi_endproc
  9297. /*
  9298. * Copyright 2018-2019 Yury Gribov
  9299. *
  9300. * The MIT License (MIT)
  9301. *
  9302. * Use of this source code is governed by MIT license that can be
  9303. * found in the LICENSE.txt file.
  9304. */
  9305. .globl cuMemsetD2D32_v2
  9306. .p2align 4
  9307. .type cuMemsetD2D32_v2, %function
  9308. cuMemsetD2D32_v2:
  9309. .cfi_startproc
  9310. // Intel opt. manual says to
  9311. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9312. // to hint static predictor.
  9313. cmpq $0, _libcuda_so_tramp_table+2632(%rip)
  9314. je 2f
  9315. 1:
  9316. jmp *_libcuda_so_tramp_table+2632(%rip)
  9317. 2:
  9318. pushq $329
  9319. .cfi_adjust_cfa_offset 8
  9320. call _libcuda_so_save_regs_and_resolve
  9321. addq $8, %rsp
  9322. .cfi_adjust_cfa_offset -8
  9323. jmp 1b
  9324. .cfi_endproc
  9325. /*
  9326. * Copyright 2018-2019 Yury Gribov
  9327. *
  9328. * The MIT License (MIT)
  9329. *
  9330. * Use of this source code is governed by MIT license that can be
  9331. * found in the LICENSE.txt file.
  9332. */
  9333. .globl cuMemsetD2D32_v2_ptds
  9334. .p2align 4
  9335. .type cuMemsetD2D32_v2_ptds, %function
  9336. cuMemsetD2D32_v2_ptds:
  9337. .cfi_startproc
  9338. // Intel opt. manual says to
  9339. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9340. // to hint static predictor.
  9341. cmpq $0, _libcuda_so_tramp_table+2640(%rip)
  9342. je 2f
  9343. 1:
  9344. jmp *_libcuda_so_tramp_table+2640(%rip)
  9345. 2:
  9346. pushq $330
  9347. .cfi_adjust_cfa_offset 8
  9348. call _libcuda_so_save_regs_and_resolve
  9349. addq $8, %rsp
  9350. .cfi_adjust_cfa_offset -8
  9351. jmp 1b
  9352. .cfi_endproc
  9353. /*
  9354. * Copyright 2018-2019 Yury Gribov
  9355. *
  9356. * The MIT License (MIT)
  9357. *
  9358. * Use of this source code is governed by MIT license that can be
  9359. * found in the LICENSE.txt file.
  9360. */
  9361. .globl cuMemsetD2D8
  9362. .p2align 4
  9363. .type cuMemsetD2D8, %function
  9364. cuMemsetD2D8:
  9365. .cfi_startproc
  9366. // Intel opt. manual says to
  9367. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9368. // to hint static predictor.
  9369. cmpq $0, _libcuda_so_tramp_table+2648(%rip)
  9370. je 2f
  9371. 1:
  9372. jmp *_libcuda_so_tramp_table+2648(%rip)
  9373. 2:
  9374. pushq $331
  9375. .cfi_adjust_cfa_offset 8
  9376. call _libcuda_so_save_regs_and_resolve
  9377. addq $8, %rsp
  9378. .cfi_adjust_cfa_offset -8
  9379. jmp 1b
  9380. .cfi_endproc
  9381. /*
  9382. * Copyright 2018-2019 Yury Gribov
  9383. *
  9384. * The MIT License (MIT)
  9385. *
  9386. * Use of this source code is governed by MIT license that can be
  9387. * found in the LICENSE.txt file.
  9388. */
  9389. .globl cuMemsetD2D8Async
  9390. .p2align 4
  9391. .type cuMemsetD2D8Async, %function
  9392. cuMemsetD2D8Async:
  9393. .cfi_startproc
  9394. // Intel opt. manual says to
  9395. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9396. // to hint static predictor.
  9397. cmpq $0, _libcuda_so_tramp_table+2656(%rip)
  9398. je 2f
  9399. 1:
  9400. jmp *_libcuda_so_tramp_table+2656(%rip)
  9401. 2:
  9402. pushq $332
  9403. .cfi_adjust_cfa_offset 8
  9404. call _libcuda_so_save_regs_and_resolve
  9405. addq $8, %rsp
  9406. .cfi_adjust_cfa_offset -8
  9407. jmp 1b
  9408. .cfi_endproc
  9409. /*
  9410. * Copyright 2018-2019 Yury Gribov
  9411. *
  9412. * The MIT License (MIT)
  9413. *
  9414. * Use of this source code is governed by MIT license that can be
  9415. * found in the LICENSE.txt file.
  9416. */
  9417. .globl cuMemsetD2D8Async_ptsz
  9418. .p2align 4
  9419. .type cuMemsetD2D8Async_ptsz, %function
  9420. cuMemsetD2D8Async_ptsz:
  9421. .cfi_startproc
  9422. // Intel opt. manual says to
  9423. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9424. // to hint static predictor.
  9425. cmpq $0, _libcuda_so_tramp_table+2664(%rip)
  9426. je 2f
  9427. 1:
  9428. jmp *_libcuda_so_tramp_table+2664(%rip)
  9429. 2:
  9430. pushq $333
  9431. .cfi_adjust_cfa_offset 8
  9432. call _libcuda_so_save_regs_and_resolve
  9433. addq $8, %rsp
  9434. .cfi_adjust_cfa_offset -8
  9435. jmp 1b
  9436. .cfi_endproc
  9437. /*
  9438. * Copyright 2018-2019 Yury Gribov
  9439. *
  9440. * The MIT License (MIT)
  9441. *
  9442. * Use of this source code is governed by MIT license that can be
  9443. * found in the LICENSE.txt file.
  9444. */
  9445. .globl cuMemsetD2D8_v2
  9446. .p2align 4
  9447. .type cuMemsetD2D8_v2, %function
  9448. cuMemsetD2D8_v2:
  9449. .cfi_startproc
  9450. // Intel opt. manual says to
  9451. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9452. // to hint static predictor.
  9453. cmpq $0, _libcuda_so_tramp_table+2672(%rip)
  9454. je 2f
  9455. 1:
  9456. jmp *_libcuda_so_tramp_table+2672(%rip)
  9457. 2:
  9458. pushq $334
  9459. .cfi_adjust_cfa_offset 8
  9460. call _libcuda_so_save_regs_and_resolve
  9461. addq $8, %rsp
  9462. .cfi_adjust_cfa_offset -8
  9463. jmp 1b
  9464. .cfi_endproc
  9465. /*
  9466. * Copyright 2018-2019 Yury Gribov
  9467. *
  9468. * The MIT License (MIT)
  9469. *
  9470. * Use of this source code is governed by MIT license that can be
  9471. * found in the LICENSE.txt file.
  9472. */
  9473. .globl cuMemsetD2D8_v2_ptds
  9474. .p2align 4
  9475. .type cuMemsetD2D8_v2_ptds, %function
  9476. cuMemsetD2D8_v2_ptds:
  9477. .cfi_startproc
  9478. // Intel opt. manual says to
  9479. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9480. // to hint static predictor.
  9481. cmpq $0, _libcuda_so_tramp_table+2680(%rip)
  9482. je 2f
  9483. 1:
  9484. jmp *_libcuda_so_tramp_table+2680(%rip)
  9485. 2:
  9486. pushq $335
  9487. .cfi_adjust_cfa_offset 8
  9488. call _libcuda_so_save_regs_and_resolve
  9489. addq $8, %rsp
  9490. .cfi_adjust_cfa_offset -8
  9491. jmp 1b
  9492. .cfi_endproc
  9493. /*
  9494. * Copyright 2018-2019 Yury Gribov
  9495. *
  9496. * The MIT License (MIT)
  9497. *
  9498. * Use of this source code is governed by MIT license that can be
  9499. * found in the LICENSE.txt file.
  9500. */
  9501. .globl cuMemsetD32
  9502. .p2align 4
  9503. .type cuMemsetD32, %function
  9504. cuMemsetD32:
  9505. .cfi_startproc
  9506. // Intel opt. manual says to
  9507. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9508. // to hint static predictor.
  9509. cmpq $0, _libcuda_so_tramp_table+2688(%rip)
  9510. je 2f
  9511. 1:
  9512. jmp *_libcuda_so_tramp_table+2688(%rip)
  9513. 2:
  9514. pushq $336
  9515. .cfi_adjust_cfa_offset 8
  9516. call _libcuda_so_save_regs_and_resolve
  9517. addq $8, %rsp
  9518. .cfi_adjust_cfa_offset -8
  9519. jmp 1b
  9520. .cfi_endproc
  9521. /*
  9522. * Copyright 2018-2019 Yury Gribov
  9523. *
  9524. * The MIT License (MIT)
  9525. *
  9526. * Use of this source code is governed by MIT license that can be
  9527. * found in the LICENSE.txt file.
  9528. */
  9529. .globl cuMemsetD32Async
  9530. .p2align 4
  9531. .type cuMemsetD32Async, %function
  9532. cuMemsetD32Async:
  9533. .cfi_startproc
  9534. // Intel opt. manual says to
  9535. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9536. // to hint static predictor.
  9537. cmpq $0, _libcuda_so_tramp_table+2696(%rip)
  9538. je 2f
  9539. 1:
  9540. jmp *_libcuda_so_tramp_table+2696(%rip)
  9541. 2:
  9542. pushq $337
  9543. .cfi_adjust_cfa_offset 8
  9544. call _libcuda_so_save_regs_and_resolve
  9545. addq $8, %rsp
  9546. .cfi_adjust_cfa_offset -8
  9547. jmp 1b
  9548. .cfi_endproc
  9549. /*
  9550. * Copyright 2018-2019 Yury Gribov
  9551. *
  9552. * The MIT License (MIT)
  9553. *
  9554. * Use of this source code is governed by MIT license that can be
  9555. * found in the LICENSE.txt file.
  9556. */
  9557. .globl cuMemsetD32Async_ptsz
  9558. .p2align 4
  9559. .type cuMemsetD32Async_ptsz, %function
  9560. cuMemsetD32Async_ptsz:
  9561. .cfi_startproc
  9562. // Intel opt. manual says to
  9563. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9564. // to hint static predictor.
  9565. cmpq $0, _libcuda_so_tramp_table+2704(%rip)
  9566. je 2f
  9567. 1:
  9568. jmp *_libcuda_so_tramp_table+2704(%rip)
  9569. 2:
  9570. pushq $338
  9571. .cfi_adjust_cfa_offset 8
  9572. call _libcuda_so_save_regs_and_resolve
  9573. addq $8, %rsp
  9574. .cfi_adjust_cfa_offset -8
  9575. jmp 1b
  9576. .cfi_endproc
  9577. /*
  9578. * Copyright 2018-2019 Yury Gribov
  9579. *
  9580. * The MIT License (MIT)
  9581. *
  9582. * Use of this source code is governed by MIT license that can be
  9583. * found in the LICENSE.txt file.
  9584. */
  9585. .globl cuMemsetD32_v2
  9586. .p2align 4
  9587. .type cuMemsetD32_v2, %function
  9588. cuMemsetD32_v2:
  9589. .cfi_startproc
  9590. // Intel opt. manual says to
  9591. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9592. // to hint static predictor.
  9593. cmpq $0, _libcuda_so_tramp_table+2712(%rip)
  9594. je 2f
  9595. 1:
  9596. jmp *_libcuda_so_tramp_table+2712(%rip)
  9597. 2:
  9598. pushq $339
  9599. .cfi_adjust_cfa_offset 8
  9600. call _libcuda_so_save_regs_and_resolve
  9601. addq $8, %rsp
  9602. .cfi_adjust_cfa_offset -8
  9603. jmp 1b
  9604. .cfi_endproc
  9605. /*
  9606. * Copyright 2018-2019 Yury Gribov
  9607. *
  9608. * The MIT License (MIT)
  9609. *
  9610. * Use of this source code is governed by MIT license that can be
  9611. * found in the LICENSE.txt file.
  9612. */
  9613. .globl cuMemsetD32_v2_ptds
  9614. .p2align 4
  9615. .type cuMemsetD32_v2_ptds, %function
  9616. cuMemsetD32_v2_ptds:
  9617. .cfi_startproc
  9618. // Intel opt. manual says to
  9619. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9620. // to hint static predictor.
  9621. cmpq $0, _libcuda_so_tramp_table+2720(%rip)
  9622. je 2f
  9623. 1:
  9624. jmp *_libcuda_so_tramp_table+2720(%rip)
  9625. 2:
  9626. pushq $340
  9627. .cfi_adjust_cfa_offset 8
  9628. call _libcuda_so_save_regs_and_resolve
  9629. addq $8, %rsp
  9630. .cfi_adjust_cfa_offset -8
  9631. jmp 1b
  9632. .cfi_endproc
  9633. /*
  9634. * Copyright 2018-2019 Yury Gribov
  9635. *
  9636. * The MIT License (MIT)
  9637. *
  9638. * Use of this source code is governed by MIT license that can be
  9639. * found in the LICENSE.txt file.
  9640. */
  9641. .globl cuMemsetD8
  9642. .p2align 4
  9643. .type cuMemsetD8, %function
  9644. cuMemsetD8:
  9645. .cfi_startproc
  9646. // Intel opt. manual says to
  9647. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9648. // to hint static predictor.
  9649. cmpq $0, _libcuda_so_tramp_table+2728(%rip)
  9650. je 2f
  9651. 1:
  9652. jmp *_libcuda_so_tramp_table+2728(%rip)
  9653. 2:
  9654. pushq $341
  9655. .cfi_adjust_cfa_offset 8
  9656. call _libcuda_so_save_regs_and_resolve
  9657. addq $8, %rsp
  9658. .cfi_adjust_cfa_offset -8
  9659. jmp 1b
  9660. .cfi_endproc
  9661. /*
  9662. * Copyright 2018-2019 Yury Gribov
  9663. *
  9664. * The MIT License (MIT)
  9665. *
  9666. * Use of this source code is governed by MIT license that can be
  9667. * found in the LICENSE.txt file.
  9668. */
  9669. .globl cuMemsetD8Async
  9670. .p2align 4
  9671. .type cuMemsetD8Async, %function
  9672. cuMemsetD8Async:
  9673. .cfi_startproc
  9674. // Intel opt. manual says to
  9675. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9676. // to hint static predictor.
  9677. cmpq $0, _libcuda_so_tramp_table+2736(%rip)
  9678. je 2f
  9679. 1:
  9680. jmp *_libcuda_so_tramp_table+2736(%rip)
  9681. 2:
  9682. pushq $342
  9683. .cfi_adjust_cfa_offset 8
  9684. call _libcuda_so_save_regs_and_resolve
  9685. addq $8, %rsp
  9686. .cfi_adjust_cfa_offset -8
  9687. jmp 1b
  9688. .cfi_endproc
  9689. /*
  9690. * Copyright 2018-2019 Yury Gribov
  9691. *
  9692. * The MIT License (MIT)
  9693. *
  9694. * Use of this source code is governed by MIT license that can be
  9695. * found in the LICENSE.txt file.
  9696. */
  9697. .globl cuMemsetD8Async_ptsz
  9698. .p2align 4
  9699. .type cuMemsetD8Async_ptsz, %function
  9700. cuMemsetD8Async_ptsz:
  9701. .cfi_startproc
  9702. // Intel opt. manual says to
  9703. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9704. // to hint static predictor.
  9705. cmpq $0, _libcuda_so_tramp_table+2744(%rip)
  9706. je 2f
  9707. 1:
  9708. jmp *_libcuda_so_tramp_table+2744(%rip)
  9709. 2:
  9710. pushq $343
  9711. .cfi_adjust_cfa_offset 8
  9712. call _libcuda_so_save_regs_and_resolve
  9713. addq $8, %rsp
  9714. .cfi_adjust_cfa_offset -8
  9715. jmp 1b
  9716. .cfi_endproc
  9717. /*
  9718. * Copyright 2018-2019 Yury Gribov
  9719. *
  9720. * The MIT License (MIT)
  9721. *
  9722. * Use of this source code is governed by MIT license that can be
  9723. * found in the LICENSE.txt file.
  9724. */
  9725. .globl cuMemsetD8_v2
  9726. .p2align 4
  9727. .type cuMemsetD8_v2, %function
  9728. cuMemsetD8_v2:
  9729. .cfi_startproc
  9730. // Intel opt. manual says to
  9731. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9732. // to hint static predictor.
  9733. cmpq $0, _libcuda_so_tramp_table+2752(%rip)
  9734. je 2f
  9735. 1:
  9736. jmp *_libcuda_so_tramp_table+2752(%rip)
  9737. 2:
  9738. pushq $344
  9739. .cfi_adjust_cfa_offset 8
  9740. call _libcuda_so_save_regs_and_resolve
  9741. addq $8, %rsp
  9742. .cfi_adjust_cfa_offset -8
  9743. jmp 1b
  9744. .cfi_endproc
  9745. /*
  9746. * Copyright 2018-2019 Yury Gribov
  9747. *
  9748. * The MIT License (MIT)
  9749. *
  9750. * Use of this source code is governed by MIT license that can be
  9751. * found in the LICENSE.txt file.
  9752. */
  9753. .globl cuMemsetD8_v2_ptds
  9754. .p2align 4
  9755. .type cuMemsetD8_v2_ptds, %function
  9756. cuMemsetD8_v2_ptds:
  9757. .cfi_startproc
  9758. // Intel opt. manual says to
  9759. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9760. // to hint static predictor.
  9761. cmpq $0, _libcuda_so_tramp_table+2760(%rip)
  9762. je 2f
  9763. 1:
  9764. jmp *_libcuda_so_tramp_table+2760(%rip)
  9765. 2:
  9766. pushq $345
  9767. .cfi_adjust_cfa_offset 8
  9768. call _libcuda_so_save_regs_and_resolve
  9769. addq $8, %rsp
  9770. .cfi_adjust_cfa_offset -8
  9771. jmp 1b
  9772. .cfi_endproc
  9773. /*
  9774. * Copyright 2018-2019 Yury Gribov
  9775. *
  9776. * The MIT License (MIT)
  9777. *
  9778. * Use of this source code is governed by MIT license that can be
  9779. * found in the LICENSE.txt file.
  9780. */
  9781. .globl cuMipmappedArrayCreate
  9782. .p2align 4
  9783. .type cuMipmappedArrayCreate, %function
  9784. cuMipmappedArrayCreate:
  9785. .cfi_startproc
  9786. // Intel opt. manual says to
  9787. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9788. // to hint static predictor.
  9789. cmpq $0, _libcuda_so_tramp_table+2768(%rip)
  9790. je 2f
  9791. 1:
  9792. jmp *_libcuda_so_tramp_table+2768(%rip)
  9793. 2:
  9794. pushq $346
  9795. .cfi_adjust_cfa_offset 8
  9796. call _libcuda_so_save_regs_and_resolve
  9797. addq $8, %rsp
  9798. .cfi_adjust_cfa_offset -8
  9799. jmp 1b
  9800. .cfi_endproc
  9801. /*
  9802. * Copyright 2018-2019 Yury Gribov
  9803. *
  9804. * The MIT License (MIT)
  9805. *
  9806. * Use of this source code is governed by MIT license that can be
  9807. * found in the LICENSE.txt file.
  9808. */
  9809. .globl cuMipmappedArrayDestroy
  9810. .p2align 4
  9811. .type cuMipmappedArrayDestroy, %function
  9812. cuMipmappedArrayDestroy:
  9813. .cfi_startproc
  9814. // Intel opt. manual says to
  9815. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9816. // to hint static predictor.
  9817. cmpq $0, _libcuda_so_tramp_table+2776(%rip)
  9818. je 2f
  9819. 1:
  9820. jmp *_libcuda_so_tramp_table+2776(%rip)
  9821. 2:
  9822. pushq $347
  9823. .cfi_adjust_cfa_offset 8
  9824. call _libcuda_so_save_regs_and_resolve
  9825. addq $8, %rsp
  9826. .cfi_adjust_cfa_offset -8
  9827. jmp 1b
  9828. .cfi_endproc
  9829. /*
  9830. * Copyright 2018-2019 Yury Gribov
  9831. *
  9832. * The MIT License (MIT)
  9833. *
  9834. * Use of this source code is governed by MIT license that can be
  9835. * found in the LICENSE.txt file.
  9836. */
  9837. .globl cuMipmappedArrayGetLevel
  9838. .p2align 4
  9839. .type cuMipmappedArrayGetLevel, %function
  9840. cuMipmappedArrayGetLevel:
  9841. .cfi_startproc
  9842. // Intel opt. manual says to
  9843. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9844. // to hint static predictor.
  9845. cmpq $0, _libcuda_so_tramp_table+2784(%rip)
  9846. je 2f
  9847. 1:
  9848. jmp *_libcuda_so_tramp_table+2784(%rip)
  9849. 2:
  9850. pushq $348
  9851. .cfi_adjust_cfa_offset 8
  9852. call _libcuda_so_save_regs_and_resolve
  9853. addq $8, %rsp
  9854. .cfi_adjust_cfa_offset -8
  9855. jmp 1b
  9856. .cfi_endproc
  9857. /*
  9858. * Copyright 2018-2019 Yury Gribov
  9859. *
  9860. * The MIT License (MIT)
  9861. *
  9862. * Use of this source code is governed by MIT license that can be
  9863. * found in the LICENSE.txt file.
  9864. */
  9865. .globl cuMipmappedArrayGetSparseProperties
  9866. .p2align 4
  9867. .type cuMipmappedArrayGetSparseProperties, %function
  9868. cuMipmappedArrayGetSparseProperties:
  9869. .cfi_startproc
  9870. // Intel opt. manual says to
  9871. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9872. // to hint static predictor.
  9873. cmpq $0, _libcuda_so_tramp_table+2792(%rip)
  9874. je 2f
  9875. 1:
  9876. jmp *_libcuda_so_tramp_table+2792(%rip)
  9877. 2:
  9878. pushq $349
  9879. .cfi_adjust_cfa_offset 8
  9880. call _libcuda_so_save_regs_and_resolve
  9881. addq $8, %rsp
  9882. .cfi_adjust_cfa_offset -8
  9883. jmp 1b
  9884. .cfi_endproc
  9885. /*
  9886. * Copyright 2018-2019 Yury Gribov
  9887. *
  9888. * The MIT License (MIT)
  9889. *
  9890. * Use of this source code is governed by MIT license that can be
  9891. * found in the LICENSE.txt file.
  9892. */
  9893. .globl cuModuleGetFunction
  9894. .p2align 4
  9895. .type cuModuleGetFunction, %function
  9896. cuModuleGetFunction:
  9897. .cfi_startproc
  9898. // Intel opt. manual says to
  9899. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9900. // to hint static predictor.
  9901. cmpq $0, _libcuda_so_tramp_table+2800(%rip)
  9902. je 2f
  9903. 1:
  9904. jmp *_libcuda_so_tramp_table+2800(%rip)
  9905. 2:
  9906. pushq $350
  9907. .cfi_adjust_cfa_offset 8
  9908. call _libcuda_so_save_regs_and_resolve
  9909. addq $8, %rsp
  9910. .cfi_adjust_cfa_offset -8
  9911. jmp 1b
  9912. .cfi_endproc
  9913. /*
  9914. * Copyright 2018-2019 Yury Gribov
  9915. *
  9916. * The MIT License (MIT)
  9917. *
  9918. * Use of this source code is governed by MIT license that can be
  9919. * found in the LICENSE.txt file.
  9920. */
  9921. .globl cuModuleGetGlobal
  9922. .p2align 4
  9923. .type cuModuleGetGlobal, %function
  9924. cuModuleGetGlobal:
  9925. .cfi_startproc
  9926. // Intel opt. manual says to
  9927. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9928. // to hint static predictor.
  9929. cmpq $0, _libcuda_so_tramp_table+2808(%rip)
  9930. je 2f
  9931. 1:
  9932. jmp *_libcuda_so_tramp_table+2808(%rip)
  9933. 2:
  9934. pushq $351
  9935. .cfi_adjust_cfa_offset 8
  9936. call _libcuda_so_save_regs_and_resolve
  9937. addq $8, %rsp
  9938. .cfi_adjust_cfa_offset -8
  9939. jmp 1b
  9940. .cfi_endproc
  9941. /*
  9942. * Copyright 2018-2019 Yury Gribov
  9943. *
  9944. * The MIT License (MIT)
  9945. *
  9946. * Use of this source code is governed by MIT license that can be
  9947. * found in the LICENSE.txt file.
  9948. */
  9949. .globl cuModuleGetGlobal_v2
  9950. .p2align 4
  9951. .type cuModuleGetGlobal_v2, %function
  9952. cuModuleGetGlobal_v2:
  9953. .cfi_startproc
  9954. // Intel opt. manual says to
  9955. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9956. // to hint static predictor.
  9957. cmpq $0, _libcuda_so_tramp_table+2816(%rip)
  9958. je 2f
  9959. 1:
  9960. jmp *_libcuda_so_tramp_table+2816(%rip)
  9961. 2:
  9962. pushq $352
  9963. .cfi_adjust_cfa_offset 8
  9964. call _libcuda_so_save_regs_and_resolve
  9965. addq $8, %rsp
  9966. .cfi_adjust_cfa_offset -8
  9967. jmp 1b
  9968. .cfi_endproc
  9969. /*
  9970. * Copyright 2018-2019 Yury Gribov
  9971. *
  9972. * The MIT License (MIT)
  9973. *
  9974. * Use of this source code is governed by MIT license that can be
  9975. * found in the LICENSE.txt file.
  9976. */
  9977. .globl cuModuleGetSurfRef
  9978. .p2align 4
  9979. .type cuModuleGetSurfRef, %function
  9980. cuModuleGetSurfRef:
  9981. .cfi_startproc
  9982. // Intel opt. manual says to
  9983. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  9984. // to hint static predictor.
  9985. cmpq $0, _libcuda_so_tramp_table+2824(%rip)
  9986. je 2f
  9987. 1:
  9988. jmp *_libcuda_so_tramp_table+2824(%rip)
  9989. 2:
  9990. pushq $353
  9991. .cfi_adjust_cfa_offset 8
  9992. call _libcuda_so_save_regs_and_resolve
  9993. addq $8, %rsp
  9994. .cfi_adjust_cfa_offset -8
  9995. jmp 1b
  9996. .cfi_endproc
  9997. /*
  9998. * Copyright 2018-2019 Yury Gribov
  9999. *
  10000. * The MIT License (MIT)
  10001. *
  10002. * Use of this source code is governed by MIT license that can be
  10003. * found in the LICENSE.txt file.
  10004. */
  10005. .globl cuModuleGetTexRef
  10006. .p2align 4
  10007. .type cuModuleGetTexRef, %function
  10008. cuModuleGetTexRef:
  10009. .cfi_startproc
  10010. // Intel opt. manual says to
  10011. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10012. // to hint static predictor.
  10013. cmpq $0, _libcuda_so_tramp_table+2832(%rip)
  10014. je 2f
  10015. 1:
  10016. jmp *_libcuda_so_tramp_table+2832(%rip)
  10017. 2:
  10018. pushq $354
  10019. .cfi_adjust_cfa_offset 8
  10020. call _libcuda_so_save_regs_and_resolve
  10021. addq $8, %rsp
  10022. .cfi_adjust_cfa_offset -8
  10023. jmp 1b
  10024. .cfi_endproc
  10025. /*
  10026. * Copyright 2018-2019 Yury Gribov
  10027. *
  10028. * The MIT License (MIT)
  10029. *
  10030. * Use of this source code is governed by MIT license that can be
  10031. * found in the LICENSE.txt file.
  10032. */
  10033. .globl cuModuleLoad
  10034. .p2align 4
  10035. .type cuModuleLoad, %function
  10036. cuModuleLoad:
  10037. .cfi_startproc
  10038. // Intel opt. manual says to
  10039. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10040. // to hint static predictor.
  10041. cmpq $0, _libcuda_so_tramp_table+2840(%rip)
  10042. je 2f
  10043. 1:
  10044. jmp *_libcuda_so_tramp_table+2840(%rip)
  10045. 2:
  10046. pushq $355
  10047. .cfi_adjust_cfa_offset 8
  10048. call _libcuda_so_save_regs_and_resolve
  10049. addq $8, %rsp
  10050. .cfi_adjust_cfa_offset -8
  10051. jmp 1b
  10052. .cfi_endproc
  10053. /*
  10054. * Copyright 2018-2019 Yury Gribov
  10055. *
  10056. * The MIT License (MIT)
  10057. *
  10058. * Use of this source code is governed by MIT license that can be
  10059. * found in the LICENSE.txt file.
  10060. */
  10061. .globl cuModuleLoadData
  10062. .p2align 4
  10063. .type cuModuleLoadData, %function
  10064. cuModuleLoadData:
  10065. .cfi_startproc
  10066. // Intel opt. manual says to
  10067. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10068. // to hint static predictor.
  10069. cmpq $0, _libcuda_so_tramp_table+2848(%rip)
  10070. je 2f
  10071. 1:
  10072. jmp *_libcuda_so_tramp_table+2848(%rip)
  10073. 2:
  10074. pushq $356
  10075. .cfi_adjust_cfa_offset 8
  10076. call _libcuda_so_save_regs_and_resolve
  10077. addq $8, %rsp
  10078. .cfi_adjust_cfa_offset -8
  10079. jmp 1b
  10080. .cfi_endproc
  10081. /*
  10082. * Copyright 2018-2019 Yury Gribov
  10083. *
  10084. * The MIT License (MIT)
  10085. *
  10086. * Use of this source code is governed by MIT license that can be
  10087. * found in the LICENSE.txt file.
  10088. */
  10089. .globl cuModuleLoadDataEx
  10090. .p2align 4
  10091. .type cuModuleLoadDataEx, %function
  10092. cuModuleLoadDataEx:
  10093. .cfi_startproc
  10094. // Intel opt. manual says to
  10095. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10096. // to hint static predictor.
  10097. cmpq $0, _libcuda_so_tramp_table+2856(%rip)
  10098. je 2f
  10099. 1:
  10100. jmp *_libcuda_so_tramp_table+2856(%rip)
  10101. 2:
  10102. pushq $357
  10103. .cfi_adjust_cfa_offset 8
  10104. call _libcuda_so_save_regs_and_resolve
  10105. addq $8, %rsp
  10106. .cfi_adjust_cfa_offset -8
  10107. jmp 1b
  10108. .cfi_endproc
  10109. /*
  10110. * Copyright 2018-2019 Yury Gribov
  10111. *
  10112. * The MIT License (MIT)
  10113. *
  10114. * Use of this source code is governed by MIT license that can be
  10115. * found in the LICENSE.txt file.
  10116. */
  10117. .globl cuModuleLoadFatBinary
  10118. .p2align 4
  10119. .type cuModuleLoadFatBinary, %function
  10120. cuModuleLoadFatBinary:
  10121. .cfi_startproc
  10122. // Intel opt. manual says to
  10123. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10124. // to hint static predictor.
  10125. cmpq $0, _libcuda_so_tramp_table+2864(%rip)
  10126. je 2f
  10127. 1:
  10128. jmp *_libcuda_so_tramp_table+2864(%rip)
  10129. 2:
  10130. pushq $358
  10131. .cfi_adjust_cfa_offset 8
  10132. call _libcuda_so_save_regs_and_resolve
  10133. addq $8, %rsp
  10134. .cfi_adjust_cfa_offset -8
  10135. jmp 1b
  10136. .cfi_endproc
  10137. /*
  10138. * Copyright 2018-2019 Yury Gribov
  10139. *
  10140. * The MIT License (MIT)
  10141. *
  10142. * Use of this source code is governed by MIT license that can be
  10143. * found in the LICENSE.txt file.
  10144. */
  10145. .globl cuModuleUnload
  10146. .p2align 4
  10147. .type cuModuleUnload, %function
  10148. cuModuleUnload:
  10149. .cfi_startproc
  10150. // Intel opt. manual says to
  10151. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10152. // to hint static predictor.
  10153. cmpq $0, _libcuda_so_tramp_table+2872(%rip)
  10154. je 2f
  10155. 1:
  10156. jmp *_libcuda_so_tramp_table+2872(%rip)
  10157. 2:
  10158. pushq $359
  10159. .cfi_adjust_cfa_offset 8
  10160. call _libcuda_so_save_regs_and_resolve
  10161. addq $8, %rsp
  10162. .cfi_adjust_cfa_offset -8
  10163. jmp 1b
  10164. .cfi_endproc
  10165. /*
  10166. * Copyright 2018-2019 Yury Gribov
  10167. *
  10168. * The MIT License (MIT)
  10169. *
  10170. * Use of this source code is governed by MIT license that can be
  10171. * found in the LICENSE.txt file.
  10172. */
  10173. .globl cuOccupancyAvailableDynamicSMemPerBlock
  10174. .p2align 4
  10175. .type cuOccupancyAvailableDynamicSMemPerBlock, %function
  10176. cuOccupancyAvailableDynamicSMemPerBlock:
  10177. .cfi_startproc
  10178. // Intel opt. manual says to
  10179. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10180. // to hint static predictor.
  10181. cmpq $0, _libcuda_so_tramp_table+2880(%rip)
  10182. je 2f
  10183. 1:
  10184. jmp *_libcuda_so_tramp_table+2880(%rip)
  10185. 2:
  10186. pushq $360
  10187. .cfi_adjust_cfa_offset 8
  10188. call _libcuda_so_save_regs_and_resolve
  10189. addq $8, %rsp
  10190. .cfi_adjust_cfa_offset -8
  10191. jmp 1b
  10192. .cfi_endproc
  10193. /*
  10194. * Copyright 2018-2019 Yury Gribov
  10195. *
  10196. * The MIT License (MIT)
  10197. *
  10198. * Use of this source code is governed by MIT license that can be
  10199. * found in the LICENSE.txt file.
  10200. */
  10201. .globl cuOccupancyMaxActiveBlocksPerMultiprocessor
  10202. .p2align 4
  10203. .type cuOccupancyMaxActiveBlocksPerMultiprocessor, %function
  10204. cuOccupancyMaxActiveBlocksPerMultiprocessor:
  10205. .cfi_startproc
  10206. // Intel opt. manual says to
  10207. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10208. // to hint static predictor.
  10209. cmpq $0, _libcuda_so_tramp_table+2888(%rip)
  10210. je 2f
  10211. 1:
  10212. jmp *_libcuda_so_tramp_table+2888(%rip)
  10213. 2:
  10214. pushq $361
  10215. .cfi_adjust_cfa_offset 8
  10216. call _libcuda_so_save_regs_and_resolve
  10217. addq $8, %rsp
  10218. .cfi_adjust_cfa_offset -8
  10219. jmp 1b
  10220. .cfi_endproc
  10221. /*
  10222. * Copyright 2018-2019 Yury Gribov
  10223. *
  10224. * The MIT License (MIT)
  10225. *
  10226. * Use of this source code is governed by MIT license that can be
  10227. * found in the LICENSE.txt file.
  10228. */
  10229. .globl cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags
  10230. .p2align 4
  10231. .type cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags, %function
  10232. cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags:
  10233. .cfi_startproc
  10234. // Intel opt. manual says to
  10235. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10236. // to hint static predictor.
  10237. cmpq $0, _libcuda_so_tramp_table+2896(%rip)
  10238. je 2f
  10239. 1:
  10240. jmp *_libcuda_so_tramp_table+2896(%rip)
  10241. 2:
  10242. pushq $362
  10243. .cfi_adjust_cfa_offset 8
  10244. call _libcuda_so_save_regs_and_resolve
  10245. addq $8, %rsp
  10246. .cfi_adjust_cfa_offset -8
  10247. jmp 1b
  10248. .cfi_endproc
  10249. /*
  10250. * Copyright 2018-2019 Yury Gribov
  10251. *
  10252. * The MIT License (MIT)
  10253. *
  10254. * Use of this source code is governed by MIT license that can be
  10255. * found in the LICENSE.txt file.
  10256. */
  10257. .globl cuOccupancyMaxPotentialBlockSize
  10258. .p2align 4
  10259. .type cuOccupancyMaxPotentialBlockSize, %function
  10260. cuOccupancyMaxPotentialBlockSize:
  10261. .cfi_startproc
  10262. // Intel opt. manual says to
  10263. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10264. // to hint static predictor.
  10265. cmpq $0, _libcuda_so_tramp_table+2904(%rip)
  10266. je 2f
  10267. 1:
  10268. jmp *_libcuda_so_tramp_table+2904(%rip)
  10269. 2:
  10270. pushq $363
  10271. .cfi_adjust_cfa_offset 8
  10272. call _libcuda_so_save_regs_and_resolve
  10273. addq $8, %rsp
  10274. .cfi_adjust_cfa_offset -8
  10275. jmp 1b
  10276. .cfi_endproc
  10277. /*
  10278. * Copyright 2018-2019 Yury Gribov
  10279. *
  10280. * The MIT License (MIT)
  10281. *
  10282. * Use of this source code is governed by MIT license that can be
  10283. * found in the LICENSE.txt file.
  10284. */
  10285. .globl cuOccupancyMaxPotentialBlockSizeWithFlags
  10286. .p2align 4
  10287. .type cuOccupancyMaxPotentialBlockSizeWithFlags, %function
  10288. cuOccupancyMaxPotentialBlockSizeWithFlags:
  10289. .cfi_startproc
  10290. // Intel opt. manual says to
  10291. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10292. // to hint static predictor.
  10293. cmpq $0, _libcuda_so_tramp_table+2912(%rip)
  10294. je 2f
  10295. 1:
  10296. jmp *_libcuda_so_tramp_table+2912(%rip)
  10297. 2:
  10298. pushq $364
  10299. .cfi_adjust_cfa_offset 8
  10300. call _libcuda_so_save_regs_and_resolve
  10301. addq $8, %rsp
  10302. .cfi_adjust_cfa_offset -8
  10303. jmp 1b
  10304. .cfi_endproc
  10305. /*
  10306. * Copyright 2018-2019 Yury Gribov
  10307. *
  10308. * The MIT License (MIT)
  10309. *
  10310. * Use of this source code is governed by MIT license that can be
  10311. * found in the LICENSE.txt file.
  10312. */
  10313. .globl cuParamSetSize
  10314. .p2align 4
  10315. .type cuParamSetSize, %function
  10316. cuParamSetSize:
  10317. .cfi_startproc
  10318. // Intel opt. manual says to
  10319. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10320. // to hint static predictor.
  10321. cmpq $0, _libcuda_so_tramp_table+2920(%rip)
  10322. je 2f
  10323. 1:
  10324. jmp *_libcuda_so_tramp_table+2920(%rip)
  10325. 2:
  10326. pushq $365
  10327. .cfi_adjust_cfa_offset 8
  10328. call _libcuda_so_save_regs_and_resolve
  10329. addq $8, %rsp
  10330. .cfi_adjust_cfa_offset -8
  10331. jmp 1b
  10332. .cfi_endproc
  10333. /*
  10334. * Copyright 2018-2019 Yury Gribov
  10335. *
  10336. * The MIT License (MIT)
  10337. *
  10338. * Use of this source code is governed by MIT license that can be
  10339. * found in the LICENSE.txt file.
  10340. */
  10341. .globl cuParamSetTexRef
  10342. .p2align 4
  10343. .type cuParamSetTexRef, %function
  10344. cuParamSetTexRef:
  10345. .cfi_startproc
  10346. // Intel opt. manual says to
  10347. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10348. // to hint static predictor.
  10349. cmpq $0, _libcuda_so_tramp_table+2928(%rip)
  10350. je 2f
  10351. 1:
  10352. jmp *_libcuda_so_tramp_table+2928(%rip)
  10353. 2:
  10354. pushq $366
  10355. .cfi_adjust_cfa_offset 8
  10356. call _libcuda_so_save_regs_and_resolve
  10357. addq $8, %rsp
  10358. .cfi_adjust_cfa_offset -8
  10359. jmp 1b
  10360. .cfi_endproc
  10361. /*
  10362. * Copyright 2018-2019 Yury Gribov
  10363. *
  10364. * The MIT License (MIT)
  10365. *
  10366. * Use of this source code is governed by MIT license that can be
  10367. * found in the LICENSE.txt file.
  10368. */
  10369. .globl cuParamSetf
  10370. .p2align 4
  10371. .type cuParamSetf, %function
  10372. cuParamSetf:
  10373. .cfi_startproc
  10374. // Intel opt. manual says to
  10375. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10376. // to hint static predictor.
  10377. cmpq $0, _libcuda_so_tramp_table+2936(%rip)
  10378. je 2f
  10379. 1:
  10380. jmp *_libcuda_so_tramp_table+2936(%rip)
  10381. 2:
  10382. pushq $367
  10383. .cfi_adjust_cfa_offset 8
  10384. call _libcuda_so_save_regs_and_resolve
  10385. addq $8, %rsp
  10386. .cfi_adjust_cfa_offset -8
  10387. jmp 1b
  10388. .cfi_endproc
  10389. /*
  10390. * Copyright 2018-2019 Yury Gribov
  10391. *
  10392. * The MIT License (MIT)
  10393. *
  10394. * Use of this source code is governed by MIT license that can be
  10395. * found in the LICENSE.txt file.
  10396. */
  10397. .globl cuParamSeti
  10398. .p2align 4
  10399. .type cuParamSeti, %function
  10400. cuParamSeti:
  10401. .cfi_startproc
  10402. // Intel opt. manual says to
  10403. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10404. // to hint static predictor.
  10405. cmpq $0, _libcuda_so_tramp_table+2944(%rip)
  10406. je 2f
  10407. 1:
  10408. jmp *_libcuda_so_tramp_table+2944(%rip)
  10409. 2:
  10410. pushq $368
  10411. .cfi_adjust_cfa_offset 8
  10412. call _libcuda_so_save_regs_and_resolve
  10413. addq $8, %rsp
  10414. .cfi_adjust_cfa_offset -8
  10415. jmp 1b
  10416. .cfi_endproc
  10417. /*
  10418. * Copyright 2018-2019 Yury Gribov
  10419. *
  10420. * The MIT License (MIT)
  10421. *
  10422. * Use of this source code is governed by MIT license that can be
  10423. * found in the LICENSE.txt file.
  10424. */
  10425. .globl cuParamSetv
  10426. .p2align 4
  10427. .type cuParamSetv, %function
  10428. cuParamSetv:
  10429. .cfi_startproc
  10430. // Intel opt. manual says to
  10431. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10432. // to hint static predictor.
  10433. cmpq $0, _libcuda_so_tramp_table+2952(%rip)
  10434. je 2f
  10435. 1:
  10436. jmp *_libcuda_so_tramp_table+2952(%rip)
  10437. 2:
  10438. pushq $369
  10439. .cfi_adjust_cfa_offset 8
  10440. call _libcuda_so_save_regs_and_resolve
  10441. addq $8, %rsp
  10442. .cfi_adjust_cfa_offset -8
  10443. jmp 1b
  10444. .cfi_endproc
  10445. /*
  10446. * Copyright 2018-2019 Yury Gribov
  10447. *
  10448. * The MIT License (MIT)
  10449. *
  10450. * Use of this source code is governed by MIT license that can be
  10451. * found in the LICENSE.txt file.
  10452. */
  10453. .globl cuPointerGetAttribute
  10454. .p2align 4
  10455. .type cuPointerGetAttribute, %function
  10456. cuPointerGetAttribute:
  10457. .cfi_startproc
  10458. // Intel opt. manual says to
  10459. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10460. // to hint static predictor.
  10461. cmpq $0, _libcuda_so_tramp_table+2960(%rip)
  10462. je 2f
  10463. 1:
  10464. jmp *_libcuda_so_tramp_table+2960(%rip)
  10465. 2:
  10466. pushq $370
  10467. .cfi_adjust_cfa_offset 8
  10468. call _libcuda_so_save_regs_and_resolve
  10469. addq $8, %rsp
  10470. .cfi_adjust_cfa_offset -8
  10471. jmp 1b
  10472. .cfi_endproc
  10473. /*
  10474. * Copyright 2018-2019 Yury Gribov
  10475. *
  10476. * The MIT License (MIT)
  10477. *
  10478. * Use of this source code is governed by MIT license that can be
  10479. * found in the LICENSE.txt file.
  10480. */
  10481. .globl cuPointerGetAttributes
  10482. .p2align 4
  10483. .type cuPointerGetAttributes, %function
  10484. cuPointerGetAttributes:
  10485. .cfi_startproc
  10486. // Intel opt. manual says to
  10487. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10488. // to hint static predictor.
  10489. cmpq $0, _libcuda_so_tramp_table+2968(%rip)
  10490. je 2f
  10491. 1:
  10492. jmp *_libcuda_so_tramp_table+2968(%rip)
  10493. 2:
  10494. pushq $371
  10495. .cfi_adjust_cfa_offset 8
  10496. call _libcuda_so_save_regs_and_resolve
  10497. addq $8, %rsp
  10498. .cfi_adjust_cfa_offset -8
  10499. jmp 1b
  10500. .cfi_endproc
  10501. /*
  10502. * Copyright 2018-2019 Yury Gribov
  10503. *
  10504. * The MIT License (MIT)
  10505. *
  10506. * Use of this source code is governed by MIT license that can be
  10507. * found in the LICENSE.txt file.
  10508. */
  10509. .globl cuPointerSetAttribute
  10510. .p2align 4
  10511. .type cuPointerSetAttribute, %function
  10512. cuPointerSetAttribute:
  10513. .cfi_startproc
  10514. // Intel opt. manual says to
  10515. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10516. // to hint static predictor.
  10517. cmpq $0, _libcuda_so_tramp_table+2976(%rip)
  10518. je 2f
  10519. 1:
  10520. jmp *_libcuda_so_tramp_table+2976(%rip)
  10521. 2:
  10522. pushq $372
  10523. .cfi_adjust_cfa_offset 8
  10524. call _libcuda_so_save_regs_and_resolve
  10525. addq $8, %rsp
  10526. .cfi_adjust_cfa_offset -8
  10527. jmp 1b
  10528. .cfi_endproc
  10529. /*
  10530. * Copyright 2018-2019 Yury Gribov
  10531. *
  10532. * The MIT License (MIT)
  10533. *
  10534. * Use of this source code is governed by MIT license that can be
  10535. * found in the LICENSE.txt file.
  10536. */
  10537. .globl cuProfilerInitialize
  10538. .p2align 4
  10539. .type cuProfilerInitialize, %function
  10540. cuProfilerInitialize:
  10541. .cfi_startproc
  10542. // Intel opt. manual says to
  10543. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10544. // to hint static predictor.
  10545. cmpq $0, _libcuda_so_tramp_table+2984(%rip)
  10546. je 2f
  10547. 1:
  10548. jmp *_libcuda_so_tramp_table+2984(%rip)
  10549. 2:
  10550. pushq $373
  10551. .cfi_adjust_cfa_offset 8
  10552. call _libcuda_so_save_regs_and_resolve
  10553. addq $8, %rsp
  10554. .cfi_adjust_cfa_offset -8
  10555. jmp 1b
  10556. .cfi_endproc
  10557. /*
  10558. * Copyright 2018-2019 Yury Gribov
  10559. *
  10560. * The MIT License (MIT)
  10561. *
  10562. * Use of this source code is governed by MIT license that can be
  10563. * found in the LICENSE.txt file.
  10564. */
  10565. .globl cuProfilerStart
  10566. .p2align 4
  10567. .type cuProfilerStart, %function
  10568. cuProfilerStart:
  10569. .cfi_startproc
  10570. // Intel opt. manual says to
  10571. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10572. // to hint static predictor.
  10573. cmpq $0, _libcuda_so_tramp_table+2992(%rip)
  10574. je 2f
  10575. 1:
  10576. jmp *_libcuda_so_tramp_table+2992(%rip)
  10577. 2:
  10578. pushq $374
  10579. .cfi_adjust_cfa_offset 8
  10580. call _libcuda_so_save_regs_and_resolve
  10581. addq $8, %rsp
  10582. .cfi_adjust_cfa_offset -8
  10583. jmp 1b
  10584. .cfi_endproc
  10585. /*
  10586. * Copyright 2018-2019 Yury Gribov
  10587. *
  10588. * The MIT License (MIT)
  10589. *
  10590. * Use of this source code is governed by MIT license that can be
  10591. * found in the LICENSE.txt file.
  10592. */
  10593. .globl cuProfilerStop
  10594. .p2align 4
  10595. .type cuProfilerStop, %function
  10596. cuProfilerStop:
  10597. .cfi_startproc
  10598. // Intel opt. manual says to
  10599. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10600. // to hint static predictor.
  10601. cmpq $0, _libcuda_so_tramp_table+3000(%rip)
  10602. je 2f
  10603. 1:
  10604. jmp *_libcuda_so_tramp_table+3000(%rip)
  10605. 2:
  10606. pushq $375
  10607. .cfi_adjust_cfa_offset 8
  10608. call _libcuda_so_save_regs_and_resolve
  10609. addq $8, %rsp
  10610. .cfi_adjust_cfa_offset -8
  10611. jmp 1b
  10612. .cfi_endproc
  10613. /*
  10614. * Copyright 2018-2019 Yury Gribov
  10615. *
  10616. * The MIT License (MIT)
  10617. *
  10618. * Use of this source code is governed by MIT license that can be
  10619. * found in the LICENSE.txt file.
  10620. */
  10621. .globl cuSignalExternalSemaphoresAsync
  10622. .p2align 4
  10623. .type cuSignalExternalSemaphoresAsync, %function
  10624. cuSignalExternalSemaphoresAsync:
  10625. .cfi_startproc
  10626. // Intel opt. manual says to
  10627. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10628. // to hint static predictor.
  10629. cmpq $0, _libcuda_so_tramp_table+3008(%rip)
  10630. je 2f
  10631. 1:
  10632. jmp *_libcuda_so_tramp_table+3008(%rip)
  10633. 2:
  10634. pushq $376
  10635. .cfi_adjust_cfa_offset 8
  10636. call _libcuda_so_save_regs_and_resolve
  10637. addq $8, %rsp
  10638. .cfi_adjust_cfa_offset -8
  10639. jmp 1b
  10640. .cfi_endproc
  10641. /*
  10642. * Copyright 2018-2019 Yury Gribov
  10643. *
  10644. * The MIT License (MIT)
  10645. *
  10646. * Use of this source code is governed by MIT license that can be
  10647. * found in the LICENSE.txt file.
  10648. */
  10649. .globl cuSignalExternalSemaphoresAsync_ptsz
  10650. .p2align 4
  10651. .type cuSignalExternalSemaphoresAsync_ptsz, %function
  10652. cuSignalExternalSemaphoresAsync_ptsz:
  10653. .cfi_startproc
  10654. // Intel opt. manual says to
  10655. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10656. // to hint static predictor.
  10657. cmpq $0, _libcuda_so_tramp_table+3016(%rip)
  10658. je 2f
  10659. 1:
  10660. jmp *_libcuda_so_tramp_table+3016(%rip)
  10661. 2:
  10662. pushq $377
  10663. .cfi_adjust_cfa_offset 8
  10664. call _libcuda_so_save_regs_and_resolve
  10665. addq $8, %rsp
  10666. .cfi_adjust_cfa_offset -8
  10667. jmp 1b
  10668. .cfi_endproc
  10669. /*
  10670. * Copyright 2018-2019 Yury Gribov
  10671. *
  10672. * The MIT License (MIT)
  10673. *
  10674. * Use of this source code is governed by MIT license that can be
  10675. * found in the LICENSE.txt file.
  10676. */
  10677. .globl cuStreamAddCallback
  10678. .p2align 4
  10679. .type cuStreamAddCallback, %function
  10680. cuStreamAddCallback:
  10681. .cfi_startproc
  10682. // Intel opt. manual says to
  10683. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10684. // to hint static predictor.
  10685. cmpq $0, _libcuda_so_tramp_table+3024(%rip)
  10686. je 2f
  10687. 1:
  10688. jmp *_libcuda_so_tramp_table+3024(%rip)
  10689. 2:
  10690. pushq $378
  10691. .cfi_adjust_cfa_offset 8
  10692. call _libcuda_so_save_regs_and_resolve
  10693. addq $8, %rsp
  10694. .cfi_adjust_cfa_offset -8
  10695. jmp 1b
  10696. .cfi_endproc
  10697. /*
  10698. * Copyright 2018-2019 Yury Gribov
  10699. *
  10700. * The MIT License (MIT)
  10701. *
  10702. * Use of this source code is governed by MIT license that can be
  10703. * found in the LICENSE.txt file.
  10704. */
  10705. .globl cuStreamAddCallback_ptsz
  10706. .p2align 4
  10707. .type cuStreamAddCallback_ptsz, %function
  10708. cuStreamAddCallback_ptsz:
  10709. .cfi_startproc
  10710. // Intel opt. manual says to
  10711. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10712. // to hint static predictor.
  10713. cmpq $0, _libcuda_so_tramp_table+3032(%rip)
  10714. je 2f
  10715. 1:
  10716. jmp *_libcuda_so_tramp_table+3032(%rip)
  10717. 2:
  10718. pushq $379
  10719. .cfi_adjust_cfa_offset 8
  10720. call _libcuda_so_save_regs_and_resolve
  10721. addq $8, %rsp
  10722. .cfi_adjust_cfa_offset -8
  10723. jmp 1b
  10724. .cfi_endproc
  10725. /*
  10726. * Copyright 2018-2019 Yury Gribov
  10727. *
  10728. * The MIT License (MIT)
  10729. *
  10730. * Use of this source code is governed by MIT license that can be
  10731. * found in the LICENSE.txt file.
  10732. */
  10733. .globl cuStreamAttachMemAsync
  10734. .p2align 4
  10735. .type cuStreamAttachMemAsync, %function
  10736. cuStreamAttachMemAsync:
  10737. .cfi_startproc
  10738. // Intel opt. manual says to
  10739. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10740. // to hint static predictor.
  10741. cmpq $0, _libcuda_so_tramp_table+3040(%rip)
  10742. je 2f
  10743. 1:
  10744. jmp *_libcuda_so_tramp_table+3040(%rip)
  10745. 2:
  10746. pushq $380
  10747. .cfi_adjust_cfa_offset 8
  10748. call _libcuda_so_save_regs_and_resolve
  10749. addq $8, %rsp
  10750. .cfi_adjust_cfa_offset -8
  10751. jmp 1b
  10752. .cfi_endproc
  10753. /*
  10754. * Copyright 2018-2019 Yury Gribov
  10755. *
  10756. * The MIT License (MIT)
  10757. *
  10758. * Use of this source code is governed by MIT license that can be
  10759. * found in the LICENSE.txt file.
  10760. */
  10761. .globl cuStreamAttachMemAsync_ptsz
  10762. .p2align 4
  10763. .type cuStreamAttachMemAsync_ptsz, %function
  10764. cuStreamAttachMemAsync_ptsz:
  10765. .cfi_startproc
  10766. // Intel opt. manual says to
  10767. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10768. // to hint static predictor.
  10769. cmpq $0, _libcuda_so_tramp_table+3048(%rip)
  10770. je 2f
  10771. 1:
  10772. jmp *_libcuda_so_tramp_table+3048(%rip)
  10773. 2:
  10774. pushq $381
  10775. .cfi_adjust_cfa_offset 8
  10776. call _libcuda_so_save_regs_and_resolve
  10777. addq $8, %rsp
  10778. .cfi_adjust_cfa_offset -8
  10779. jmp 1b
  10780. .cfi_endproc
  10781. /*
  10782. * Copyright 2018-2019 Yury Gribov
  10783. *
  10784. * The MIT License (MIT)
  10785. *
  10786. * Use of this source code is governed by MIT license that can be
  10787. * found in the LICENSE.txt file.
  10788. */
  10789. .globl cuStreamBatchMemOp
  10790. .p2align 4
  10791. .type cuStreamBatchMemOp, %function
  10792. cuStreamBatchMemOp:
  10793. .cfi_startproc
  10794. // Intel opt. manual says to
  10795. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10796. // to hint static predictor.
  10797. cmpq $0, _libcuda_so_tramp_table+3056(%rip)
  10798. je 2f
  10799. 1:
  10800. jmp *_libcuda_so_tramp_table+3056(%rip)
  10801. 2:
  10802. pushq $382
  10803. .cfi_adjust_cfa_offset 8
  10804. call _libcuda_so_save_regs_and_resolve
  10805. addq $8, %rsp
  10806. .cfi_adjust_cfa_offset -8
  10807. jmp 1b
  10808. .cfi_endproc
  10809. /*
  10810. * Copyright 2018-2019 Yury Gribov
  10811. *
  10812. * The MIT License (MIT)
  10813. *
  10814. * Use of this source code is governed by MIT license that can be
  10815. * found in the LICENSE.txt file.
  10816. */
  10817. .globl cuStreamBatchMemOp_ptsz
  10818. .p2align 4
  10819. .type cuStreamBatchMemOp_ptsz, %function
  10820. cuStreamBatchMemOp_ptsz:
  10821. .cfi_startproc
  10822. // Intel opt. manual says to
  10823. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10824. // to hint static predictor.
  10825. cmpq $0, _libcuda_so_tramp_table+3064(%rip)
  10826. je 2f
  10827. 1:
  10828. jmp *_libcuda_so_tramp_table+3064(%rip)
  10829. 2:
  10830. pushq $383
  10831. .cfi_adjust_cfa_offset 8
  10832. call _libcuda_so_save_regs_and_resolve
  10833. addq $8, %rsp
  10834. .cfi_adjust_cfa_offset -8
  10835. jmp 1b
  10836. .cfi_endproc
  10837. /*
  10838. * Copyright 2018-2019 Yury Gribov
  10839. *
  10840. * The MIT License (MIT)
  10841. *
  10842. * Use of this source code is governed by MIT license that can be
  10843. * found in the LICENSE.txt file.
  10844. */
  10845. .globl cuStreamBeginCapture
  10846. .p2align 4
  10847. .type cuStreamBeginCapture, %function
  10848. cuStreamBeginCapture:
  10849. .cfi_startproc
  10850. // Intel opt. manual says to
  10851. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10852. // to hint static predictor.
  10853. cmpq $0, _libcuda_so_tramp_table+3072(%rip)
  10854. je 2f
  10855. 1:
  10856. jmp *_libcuda_so_tramp_table+3072(%rip)
  10857. 2:
  10858. pushq $384
  10859. .cfi_adjust_cfa_offset 8
  10860. call _libcuda_so_save_regs_and_resolve
  10861. addq $8, %rsp
  10862. .cfi_adjust_cfa_offset -8
  10863. jmp 1b
  10864. .cfi_endproc
  10865. /*
  10866. * Copyright 2018-2019 Yury Gribov
  10867. *
  10868. * The MIT License (MIT)
  10869. *
  10870. * Use of this source code is governed by MIT license that can be
  10871. * found in the LICENSE.txt file.
  10872. */
  10873. .globl cuStreamBeginCapture_ptsz
  10874. .p2align 4
  10875. .type cuStreamBeginCapture_ptsz, %function
  10876. cuStreamBeginCapture_ptsz:
  10877. .cfi_startproc
  10878. // Intel opt. manual says to
  10879. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10880. // to hint static predictor.
  10881. cmpq $0, _libcuda_so_tramp_table+3080(%rip)
  10882. je 2f
  10883. 1:
  10884. jmp *_libcuda_so_tramp_table+3080(%rip)
  10885. 2:
  10886. pushq $385
  10887. .cfi_adjust_cfa_offset 8
  10888. call _libcuda_so_save_regs_and_resolve
  10889. addq $8, %rsp
  10890. .cfi_adjust_cfa_offset -8
  10891. jmp 1b
  10892. .cfi_endproc
  10893. /*
  10894. * Copyright 2018-2019 Yury Gribov
  10895. *
  10896. * The MIT License (MIT)
  10897. *
  10898. * Use of this source code is governed by MIT license that can be
  10899. * found in the LICENSE.txt file.
  10900. */
  10901. .globl cuStreamBeginCapture_v2
  10902. .p2align 4
  10903. .type cuStreamBeginCapture_v2, %function
  10904. cuStreamBeginCapture_v2:
  10905. .cfi_startproc
  10906. // Intel opt. manual says to
  10907. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10908. // to hint static predictor.
  10909. cmpq $0, _libcuda_so_tramp_table+3088(%rip)
  10910. je 2f
  10911. 1:
  10912. jmp *_libcuda_so_tramp_table+3088(%rip)
  10913. 2:
  10914. pushq $386
  10915. .cfi_adjust_cfa_offset 8
  10916. call _libcuda_so_save_regs_and_resolve
  10917. addq $8, %rsp
  10918. .cfi_adjust_cfa_offset -8
  10919. jmp 1b
  10920. .cfi_endproc
  10921. /*
  10922. * Copyright 2018-2019 Yury Gribov
  10923. *
  10924. * The MIT License (MIT)
  10925. *
  10926. * Use of this source code is governed by MIT license that can be
  10927. * found in the LICENSE.txt file.
  10928. */
  10929. .globl cuStreamBeginCapture_v2_ptsz
  10930. .p2align 4
  10931. .type cuStreamBeginCapture_v2_ptsz, %function
  10932. cuStreamBeginCapture_v2_ptsz:
  10933. .cfi_startproc
  10934. // Intel opt. manual says to
  10935. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10936. // to hint static predictor.
  10937. cmpq $0, _libcuda_so_tramp_table+3096(%rip)
  10938. je 2f
  10939. 1:
  10940. jmp *_libcuda_so_tramp_table+3096(%rip)
  10941. 2:
  10942. pushq $387
  10943. .cfi_adjust_cfa_offset 8
  10944. call _libcuda_so_save_regs_and_resolve
  10945. addq $8, %rsp
  10946. .cfi_adjust_cfa_offset -8
  10947. jmp 1b
  10948. .cfi_endproc
  10949. /*
  10950. * Copyright 2018-2019 Yury Gribov
  10951. *
  10952. * The MIT License (MIT)
  10953. *
  10954. * Use of this source code is governed by MIT license that can be
  10955. * found in the LICENSE.txt file.
  10956. */
  10957. .globl cuStreamCopyAttributes
  10958. .p2align 4
  10959. .type cuStreamCopyAttributes, %function
  10960. cuStreamCopyAttributes:
  10961. .cfi_startproc
  10962. // Intel opt. manual says to
  10963. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10964. // to hint static predictor.
  10965. cmpq $0, _libcuda_so_tramp_table+3104(%rip)
  10966. je 2f
  10967. 1:
  10968. jmp *_libcuda_so_tramp_table+3104(%rip)
  10969. 2:
  10970. pushq $388
  10971. .cfi_adjust_cfa_offset 8
  10972. call _libcuda_so_save_regs_and_resolve
  10973. addq $8, %rsp
  10974. .cfi_adjust_cfa_offset -8
  10975. jmp 1b
  10976. .cfi_endproc
  10977. /*
  10978. * Copyright 2018-2019 Yury Gribov
  10979. *
  10980. * The MIT License (MIT)
  10981. *
  10982. * Use of this source code is governed by MIT license that can be
  10983. * found in the LICENSE.txt file.
  10984. */
  10985. .globl cuStreamCopyAttributes_ptsz
  10986. .p2align 4
  10987. .type cuStreamCopyAttributes_ptsz, %function
  10988. cuStreamCopyAttributes_ptsz:
  10989. .cfi_startproc
  10990. // Intel opt. manual says to
  10991. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  10992. // to hint static predictor.
  10993. cmpq $0, _libcuda_so_tramp_table+3112(%rip)
  10994. je 2f
  10995. 1:
  10996. jmp *_libcuda_so_tramp_table+3112(%rip)
  10997. 2:
  10998. pushq $389
  10999. .cfi_adjust_cfa_offset 8
  11000. call _libcuda_so_save_regs_and_resolve
  11001. addq $8, %rsp
  11002. .cfi_adjust_cfa_offset -8
  11003. jmp 1b
  11004. .cfi_endproc
  11005. /*
  11006. * Copyright 2018-2019 Yury Gribov
  11007. *
  11008. * The MIT License (MIT)
  11009. *
  11010. * Use of this source code is governed by MIT license that can be
  11011. * found in the LICENSE.txt file.
  11012. */
  11013. .globl cuStreamCreate
  11014. .p2align 4
  11015. .type cuStreamCreate, %function
  11016. cuStreamCreate:
  11017. .cfi_startproc
  11018. // Intel opt. manual says to
  11019. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11020. // to hint static predictor.
  11021. cmpq $0, _libcuda_so_tramp_table+3120(%rip)
  11022. je 2f
  11023. 1:
  11024. jmp *_libcuda_so_tramp_table+3120(%rip)
  11025. 2:
  11026. pushq $390
  11027. .cfi_adjust_cfa_offset 8
  11028. call _libcuda_so_save_regs_and_resolve
  11029. addq $8, %rsp
  11030. .cfi_adjust_cfa_offset -8
  11031. jmp 1b
  11032. .cfi_endproc
  11033. /*
  11034. * Copyright 2018-2019 Yury Gribov
  11035. *
  11036. * The MIT License (MIT)
  11037. *
  11038. * Use of this source code is governed by MIT license that can be
  11039. * found in the LICENSE.txt file.
  11040. */
  11041. .globl cuStreamCreateWithPriority
  11042. .p2align 4
  11043. .type cuStreamCreateWithPriority, %function
  11044. cuStreamCreateWithPriority:
  11045. .cfi_startproc
  11046. // Intel opt. manual says to
  11047. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11048. // to hint static predictor.
  11049. cmpq $0, _libcuda_so_tramp_table+3128(%rip)
  11050. je 2f
  11051. 1:
  11052. jmp *_libcuda_so_tramp_table+3128(%rip)
  11053. 2:
  11054. pushq $391
  11055. .cfi_adjust_cfa_offset 8
  11056. call _libcuda_so_save_regs_and_resolve
  11057. addq $8, %rsp
  11058. .cfi_adjust_cfa_offset -8
  11059. jmp 1b
  11060. .cfi_endproc
  11061. /*
  11062. * Copyright 2018-2019 Yury Gribov
  11063. *
  11064. * The MIT License (MIT)
  11065. *
  11066. * Use of this source code is governed by MIT license that can be
  11067. * found in the LICENSE.txt file.
  11068. */
  11069. .globl cuStreamDestroy
  11070. .p2align 4
  11071. .type cuStreamDestroy, %function
  11072. cuStreamDestroy:
  11073. .cfi_startproc
  11074. // Intel opt. manual says to
  11075. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11076. // to hint static predictor.
  11077. cmpq $0, _libcuda_so_tramp_table+3136(%rip)
  11078. je 2f
  11079. 1:
  11080. jmp *_libcuda_so_tramp_table+3136(%rip)
  11081. 2:
  11082. pushq $392
  11083. .cfi_adjust_cfa_offset 8
  11084. call _libcuda_so_save_regs_and_resolve
  11085. addq $8, %rsp
  11086. .cfi_adjust_cfa_offset -8
  11087. jmp 1b
  11088. .cfi_endproc
  11089. /*
  11090. * Copyright 2018-2019 Yury Gribov
  11091. *
  11092. * The MIT License (MIT)
  11093. *
  11094. * Use of this source code is governed by MIT license that can be
  11095. * found in the LICENSE.txt file.
  11096. */
  11097. .globl cuStreamDestroy_v2
  11098. .p2align 4
  11099. .type cuStreamDestroy_v2, %function
  11100. cuStreamDestroy_v2:
  11101. .cfi_startproc
  11102. // Intel opt. manual says to
  11103. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11104. // to hint static predictor.
  11105. cmpq $0, _libcuda_so_tramp_table+3144(%rip)
  11106. je 2f
  11107. 1:
  11108. jmp *_libcuda_so_tramp_table+3144(%rip)
  11109. 2:
  11110. pushq $393
  11111. .cfi_adjust_cfa_offset 8
  11112. call _libcuda_so_save_regs_and_resolve
  11113. addq $8, %rsp
  11114. .cfi_adjust_cfa_offset -8
  11115. jmp 1b
  11116. .cfi_endproc
  11117. /*
  11118. * Copyright 2018-2019 Yury Gribov
  11119. *
  11120. * The MIT License (MIT)
  11121. *
  11122. * Use of this source code is governed by MIT license that can be
  11123. * found in the LICENSE.txt file.
  11124. */
  11125. .globl cuStreamEndCapture
  11126. .p2align 4
  11127. .type cuStreamEndCapture, %function
  11128. cuStreamEndCapture:
  11129. .cfi_startproc
  11130. // Intel opt. manual says to
  11131. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11132. // to hint static predictor.
  11133. cmpq $0, _libcuda_so_tramp_table+3152(%rip)
  11134. je 2f
  11135. 1:
  11136. jmp *_libcuda_so_tramp_table+3152(%rip)
  11137. 2:
  11138. pushq $394
  11139. .cfi_adjust_cfa_offset 8
  11140. call _libcuda_so_save_regs_and_resolve
  11141. addq $8, %rsp
  11142. .cfi_adjust_cfa_offset -8
  11143. jmp 1b
  11144. .cfi_endproc
  11145. /*
  11146. * Copyright 2018-2019 Yury Gribov
  11147. *
  11148. * The MIT License (MIT)
  11149. *
  11150. * Use of this source code is governed by MIT license that can be
  11151. * found in the LICENSE.txt file.
  11152. */
  11153. .globl cuStreamEndCapture_ptsz
  11154. .p2align 4
  11155. .type cuStreamEndCapture_ptsz, %function
  11156. cuStreamEndCapture_ptsz:
  11157. .cfi_startproc
  11158. // Intel opt. manual says to
  11159. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11160. // to hint static predictor.
  11161. cmpq $0, _libcuda_so_tramp_table+3160(%rip)
  11162. je 2f
  11163. 1:
  11164. jmp *_libcuda_so_tramp_table+3160(%rip)
  11165. 2:
  11166. pushq $395
  11167. .cfi_adjust_cfa_offset 8
  11168. call _libcuda_so_save_regs_and_resolve
  11169. addq $8, %rsp
  11170. .cfi_adjust_cfa_offset -8
  11171. jmp 1b
  11172. .cfi_endproc
  11173. /*
  11174. * Copyright 2018-2019 Yury Gribov
  11175. *
  11176. * The MIT License (MIT)
  11177. *
  11178. * Use of this source code is governed by MIT license that can be
  11179. * found in the LICENSE.txt file.
  11180. */
  11181. .globl cuStreamGetAttribute
  11182. .p2align 4
  11183. .type cuStreamGetAttribute, %function
  11184. cuStreamGetAttribute:
  11185. .cfi_startproc
  11186. // Intel opt. manual says to
  11187. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11188. // to hint static predictor.
  11189. cmpq $0, _libcuda_so_tramp_table+3168(%rip)
  11190. je 2f
  11191. 1:
  11192. jmp *_libcuda_so_tramp_table+3168(%rip)
  11193. 2:
  11194. pushq $396
  11195. .cfi_adjust_cfa_offset 8
  11196. call _libcuda_so_save_regs_and_resolve
  11197. addq $8, %rsp
  11198. .cfi_adjust_cfa_offset -8
  11199. jmp 1b
  11200. .cfi_endproc
  11201. /*
  11202. * Copyright 2018-2019 Yury Gribov
  11203. *
  11204. * The MIT License (MIT)
  11205. *
  11206. * Use of this source code is governed by MIT license that can be
  11207. * found in the LICENSE.txt file.
  11208. */
  11209. .globl cuStreamGetAttribute_ptsz
  11210. .p2align 4
  11211. .type cuStreamGetAttribute_ptsz, %function
  11212. cuStreamGetAttribute_ptsz:
  11213. .cfi_startproc
  11214. // Intel opt. manual says to
  11215. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11216. // to hint static predictor.
  11217. cmpq $0, _libcuda_so_tramp_table+3176(%rip)
  11218. je 2f
  11219. 1:
  11220. jmp *_libcuda_so_tramp_table+3176(%rip)
  11221. 2:
  11222. pushq $397
  11223. .cfi_adjust_cfa_offset 8
  11224. call _libcuda_so_save_regs_and_resolve
  11225. addq $8, %rsp
  11226. .cfi_adjust_cfa_offset -8
  11227. jmp 1b
  11228. .cfi_endproc
  11229. /*
  11230. * Copyright 2018-2019 Yury Gribov
  11231. *
  11232. * The MIT License (MIT)
  11233. *
  11234. * Use of this source code is governed by MIT license that can be
  11235. * found in the LICENSE.txt file.
  11236. */
  11237. .globl cuStreamGetCaptureInfo
  11238. .p2align 4
  11239. .type cuStreamGetCaptureInfo, %function
  11240. cuStreamGetCaptureInfo:
  11241. .cfi_startproc
  11242. // Intel opt. manual says to
  11243. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11244. // to hint static predictor.
  11245. cmpq $0, _libcuda_so_tramp_table+3184(%rip)
  11246. je 2f
  11247. 1:
  11248. jmp *_libcuda_so_tramp_table+3184(%rip)
  11249. 2:
  11250. pushq $398
  11251. .cfi_adjust_cfa_offset 8
  11252. call _libcuda_so_save_regs_and_resolve
  11253. addq $8, %rsp
  11254. .cfi_adjust_cfa_offset -8
  11255. jmp 1b
  11256. .cfi_endproc
  11257. /*
  11258. * Copyright 2018-2019 Yury Gribov
  11259. *
  11260. * The MIT License (MIT)
  11261. *
  11262. * Use of this source code is governed by MIT license that can be
  11263. * found in the LICENSE.txt file.
  11264. */
  11265. .globl cuStreamGetCaptureInfo_ptsz
  11266. .p2align 4
  11267. .type cuStreamGetCaptureInfo_ptsz, %function
  11268. cuStreamGetCaptureInfo_ptsz:
  11269. .cfi_startproc
  11270. // Intel opt. manual says to
  11271. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11272. // to hint static predictor.
  11273. cmpq $0, _libcuda_so_tramp_table+3192(%rip)
  11274. je 2f
  11275. 1:
  11276. jmp *_libcuda_so_tramp_table+3192(%rip)
  11277. 2:
  11278. pushq $399
  11279. .cfi_adjust_cfa_offset 8
  11280. call _libcuda_so_save_regs_and_resolve
  11281. addq $8, %rsp
  11282. .cfi_adjust_cfa_offset -8
  11283. jmp 1b
  11284. .cfi_endproc
  11285. /*
  11286. * Copyright 2018-2019 Yury Gribov
  11287. *
  11288. * The MIT License (MIT)
  11289. *
  11290. * Use of this source code is governed by MIT license that can be
  11291. * found in the LICENSE.txt file.
  11292. */
  11293. .globl cuStreamGetCtx
  11294. .p2align 4
  11295. .type cuStreamGetCtx, %function
  11296. cuStreamGetCtx:
  11297. .cfi_startproc
  11298. // Intel opt. manual says to
  11299. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11300. // to hint static predictor.
  11301. cmpq $0, _libcuda_so_tramp_table+3200(%rip)
  11302. je 2f
  11303. 1:
  11304. jmp *_libcuda_so_tramp_table+3200(%rip)
  11305. 2:
  11306. pushq $400
  11307. .cfi_adjust_cfa_offset 8
  11308. call _libcuda_so_save_regs_and_resolve
  11309. addq $8, %rsp
  11310. .cfi_adjust_cfa_offset -8
  11311. jmp 1b
  11312. .cfi_endproc
  11313. /*
  11314. * Copyright 2018-2019 Yury Gribov
  11315. *
  11316. * The MIT License (MIT)
  11317. *
  11318. * Use of this source code is governed by MIT license that can be
  11319. * found in the LICENSE.txt file.
  11320. */
  11321. .globl cuStreamGetCtx_ptsz
  11322. .p2align 4
  11323. .type cuStreamGetCtx_ptsz, %function
  11324. cuStreamGetCtx_ptsz:
  11325. .cfi_startproc
  11326. // Intel opt. manual says to
  11327. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11328. // to hint static predictor.
  11329. cmpq $0, _libcuda_so_tramp_table+3208(%rip)
  11330. je 2f
  11331. 1:
  11332. jmp *_libcuda_so_tramp_table+3208(%rip)
  11333. 2:
  11334. pushq $401
  11335. .cfi_adjust_cfa_offset 8
  11336. call _libcuda_so_save_regs_and_resolve
  11337. addq $8, %rsp
  11338. .cfi_adjust_cfa_offset -8
  11339. jmp 1b
  11340. .cfi_endproc
  11341. /*
  11342. * Copyright 2018-2019 Yury Gribov
  11343. *
  11344. * The MIT License (MIT)
  11345. *
  11346. * Use of this source code is governed by MIT license that can be
  11347. * found in the LICENSE.txt file.
  11348. */
  11349. .globl cuStreamGetFlags
  11350. .p2align 4
  11351. .type cuStreamGetFlags, %function
  11352. cuStreamGetFlags:
  11353. .cfi_startproc
  11354. // Intel opt. manual says to
  11355. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11356. // to hint static predictor.
  11357. cmpq $0, _libcuda_so_tramp_table+3216(%rip)
  11358. je 2f
  11359. 1:
  11360. jmp *_libcuda_so_tramp_table+3216(%rip)
  11361. 2:
  11362. pushq $402
  11363. .cfi_adjust_cfa_offset 8
  11364. call _libcuda_so_save_regs_and_resolve
  11365. addq $8, %rsp
  11366. .cfi_adjust_cfa_offset -8
  11367. jmp 1b
  11368. .cfi_endproc
  11369. /*
  11370. * Copyright 2018-2019 Yury Gribov
  11371. *
  11372. * The MIT License (MIT)
  11373. *
  11374. * Use of this source code is governed by MIT license that can be
  11375. * found in the LICENSE.txt file.
  11376. */
  11377. .globl cuStreamGetFlags_ptsz
  11378. .p2align 4
  11379. .type cuStreamGetFlags_ptsz, %function
  11380. cuStreamGetFlags_ptsz:
  11381. .cfi_startproc
  11382. // Intel opt. manual says to
  11383. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11384. // to hint static predictor.
  11385. cmpq $0, _libcuda_so_tramp_table+3224(%rip)
  11386. je 2f
  11387. 1:
  11388. jmp *_libcuda_so_tramp_table+3224(%rip)
  11389. 2:
  11390. pushq $403
  11391. .cfi_adjust_cfa_offset 8
  11392. call _libcuda_so_save_regs_and_resolve
  11393. addq $8, %rsp
  11394. .cfi_adjust_cfa_offset -8
  11395. jmp 1b
  11396. .cfi_endproc
  11397. /*
  11398. * Copyright 2018-2019 Yury Gribov
  11399. *
  11400. * The MIT License (MIT)
  11401. *
  11402. * Use of this source code is governed by MIT license that can be
  11403. * found in the LICENSE.txt file.
  11404. */
  11405. .globl cuStreamGetPriority
  11406. .p2align 4
  11407. .type cuStreamGetPriority, %function
  11408. cuStreamGetPriority:
  11409. .cfi_startproc
  11410. // Intel opt. manual says to
  11411. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11412. // to hint static predictor.
  11413. cmpq $0, _libcuda_so_tramp_table+3232(%rip)
  11414. je 2f
  11415. 1:
  11416. jmp *_libcuda_so_tramp_table+3232(%rip)
  11417. 2:
  11418. pushq $404
  11419. .cfi_adjust_cfa_offset 8
  11420. call _libcuda_so_save_regs_and_resolve
  11421. addq $8, %rsp
  11422. .cfi_adjust_cfa_offset -8
  11423. jmp 1b
  11424. .cfi_endproc
  11425. /*
  11426. * Copyright 2018-2019 Yury Gribov
  11427. *
  11428. * The MIT License (MIT)
  11429. *
  11430. * Use of this source code is governed by MIT license that can be
  11431. * found in the LICENSE.txt file.
  11432. */
  11433. .globl cuStreamGetPriority_ptsz
  11434. .p2align 4
  11435. .type cuStreamGetPriority_ptsz, %function
  11436. cuStreamGetPriority_ptsz:
  11437. .cfi_startproc
  11438. // Intel opt. manual says to
  11439. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11440. // to hint static predictor.
  11441. cmpq $0, _libcuda_so_tramp_table+3240(%rip)
  11442. je 2f
  11443. 1:
  11444. jmp *_libcuda_so_tramp_table+3240(%rip)
  11445. 2:
  11446. pushq $405
  11447. .cfi_adjust_cfa_offset 8
  11448. call _libcuda_so_save_regs_and_resolve
  11449. addq $8, %rsp
  11450. .cfi_adjust_cfa_offset -8
  11451. jmp 1b
  11452. .cfi_endproc
  11453. /*
  11454. * Copyright 2018-2019 Yury Gribov
  11455. *
  11456. * The MIT License (MIT)
  11457. *
  11458. * Use of this source code is governed by MIT license that can be
  11459. * found in the LICENSE.txt file.
  11460. */
  11461. .globl cuStreamIsCapturing
  11462. .p2align 4
  11463. .type cuStreamIsCapturing, %function
  11464. cuStreamIsCapturing:
  11465. .cfi_startproc
  11466. // Intel opt. manual says to
  11467. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11468. // to hint static predictor.
  11469. cmpq $0, _libcuda_so_tramp_table+3248(%rip)
  11470. je 2f
  11471. 1:
  11472. jmp *_libcuda_so_tramp_table+3248(%rip)
  11473. 2:
  11474. pushq $406
  11475. .cfi_adjust_cfa_offset 8
  11476. call _libcuda_so_save_regs_and_resolve
  11477. addq $8, %rsp
  11478. .cfi_adjust_cfa_offset -8
  11479. jmp 1b
  11480. .cfi_endproc
  11481. /*
  11482. * Copyright 2018-2019 Yury Gribov
  11483. *
  11484. * The MIT License (MIT)
  11485. *
  11486. * Use of this source code is governed by MIT license that can be
  11487. * found in the LICENSE.txt file.
  11488. */
  11489. .globl cuStreamIsCapturing_ptsz
  11490. .p2align 4
  11491. .type cuStreamIsCapturing_ptsz, %function
  11492. cuStreamIsCapturing_ptsz:
  11493. .cfi_startproc
  11494. // Intel opt. manual says to
  11495. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11496. // to hint static predictor.
  11497. cmpq $0, _libcuda_so_tramp_table+3256(%rip)
  11498. je 2f
  11499. 1:
  11500. jmp *_libcuda_so_tramp_table+3256(%rip)
  11501. 2:
  11502. pushq $407
  11503. .cfi_adjust_cfa_offset 8
  11504. call _libcuda_so_save_regs_and_resolve
  11505. addq $8, %rsp
  11506. .cfi_adjust_cfa_offset -8
  11507. jmp 1b
  11508. .cfi_endproc
  11509. /*
  11510. * Copyright 2018-2019 Yury Gribov
  11511. *
  11512. * The MIT License (MIT)
  11513. *
  11514. * Use of this source code is governed by MIT license that can be
  11515. * found in the LICENSE.txt file.
  11516. */
  11517. .globl cuStreamQuery
  11518. .p2align 4
  11519. .type cuStreamQuery, %function
  11520. cuStreamQuery:
  11521. .cfi_startproc
  11522. // Intel opt. manual says to
  11523. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11524. // to hint static predictor.
  11525. cmpq $0, _libcuda_so_tramp_table+3264(%rip)
  11526. je 2f
  11527. 1:
  11528. jmp *_libcuda_so_tramp_table+3264(%rip)
  11529. 2:
  11530. pushq $408
  11531. .cfi_adjust_cfa_offset 8
  11532. call _libcuda_so_save_regs_and_resolve
  11533. addq $8, %rsp
  11534. .cfi_adjust_cfa_offset -8
  11535. jmp 1b
  11536. .cfi_endproc
  11537. /*
  11538. * Copyright 2018-2019 Yury Gribov
  11539. *
  11540. * The MIT License (MIT)
  11541. *
  11542. * Use of this source code is governed by MIT license that can be
  11543. * found in the LICENSE.txt file.
  11544. */
  11545. .globl cuStreamQuery_ptsz
  11546. .p2align 4
  11547. .type cuStreamQuery_ptsz, %function
  11548. cuStreamQuery_ptsz:
  11549. .cfi_startproc
  11550. // Intel opt. manual says to
  11551. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11552. // to hint static predictor.
  11553. cmpq $0, _libcuda_so_tramp_table+3272(%rip)
  11554. je 2f
  11555. 1:
  11556. jmp *_libcuda_so_tramp_table+3272(%rip)
  11557. 2:
  11558. pushq $409
  11559. .cfi_adjust_cfa_offset 8
  11560. call _libcuda_so_save_regs_and_resolve
  11561. addq $8, %rsp
  11562. .cfi_adjust_cfa_offset -8
  11563. jmp 1b
  11564. .cfi_endproc
  11565. /*
  11566. * Copyright 2018-2019 Yury Gribov
  11567. *
  11568. * The MIT License (MIT)
  11569. *
  11570. * Use of this source code is governed by MIT license that can be
  11571. * found in the LICENSE.txt file.
  11572. */
  11573. .globl cuStreamSetAttribute
  11574. .p2align 4
  11575. .type cuStreamSetAttribute, %function
  11576. cuStreamSetAttribute:
  11577. .cfi_startproc
  11578. // Intel opt. manual says to
  11579. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11580. // to hint static predictor.
  11581. cmpq $0, _libcuda_so_tramp_table+3280(%rip)
  11582. je 2f
  11583. 1:
  11584. jmp *_libcuda_so_tramp_table+3280(%rip)
  11585. 2:
  11586. pushq $410
  11587. .cfi_adjust_cfa_offset 8
  11588. call _libcuda_so_save_regs_and_resolve
  11589. addq $8, %rsp
  11590. .cfi_adjust_cfa_offset -8
  11591. jmp 1b
  11592. .cfi_endproc
  11593. /*
  11594. * Copyright 2018-2019 Yury Gribov
  11595. *
  11596. * The MIT License (MIT)
  11597. *
  11598. * Use of this source code is governed by MIT license that can be
  11599. * found in the LICENSE.txt file.
  11600. */
  11601. .globl cuStreamSetAttribute_ptsz
  11602. .p2align 4
  11603. .type cuStreamSetAttribute_ptsz, %function
  11604. cuStreamSetAttribute_ptsz:
  11605. .cfi_startproc
  11606. // Intel opt. manual says to
  11607. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11608. // to hint static predictor.
  11609. cmpq $0, _libcuda_so_tramp_table+3288(%rip)
  11610. je 2f
  11611. 1:
  11612. jmp *_libcuda_so_tramp_table+3288(%rip)
  11613. 2:
  11614. pushq $411
  11615. .cfi_adjust_cfa_offset 8
  11616. call _libcuda_so_save_regs_and_resolve
  11617. addq $8, %rsp
  11618. .cfi_adjust_cfa_offset -8
  11619. jmp 1b
  11620. .cfi_endproc
  11621. /*
  11622. * Copyright 2018-2019 Yury Gribov
  11623. *
  11624. * The MIT License (MIT)
  11625. *
  11626. * Use of this source code is governed by MIT license that can be
  11627. * found in the LICENSE.txt file.
  11628. */
  11629. .globl cuStreamSynchronize
  11630. .p2align 4
  11631. .type cuStreamSynchronize, %function
  11632. cuStreamSynchronize:
  11633. .cfi_startproc
  11634. // Intel opt. manual says to
  11635. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11636. // to hint static predictor.
  11637. cmpq $0, _libcuda_so_tramp_table+3296(%rip)
  11638. je 2f
  11639. 1:
  11640. jmp *_libcuda_so_tramp_table+3296(%rip)
  11641. 2:
  11642. pushq $412
  11643. .cfi_adjust_cfa_offset 8
  11644. call _libcuda_so_save_regs_and_resolve
  11645. addq $8, %rsp
  11646. .cfi_adjust_cfa_offset -8
  11647. jmp 1b
  11648. .cfi_endproc
  11649. /*
  11650. * Copyright 2018-2019 Yury Gribov
  11651. *
  11652. * The MIT License (MIT)
  11653. *
  11654. * Use of this source code is governed by MIT license that can be
  11655. * found in the LICENSE.txt file.
  11656. */
  11657. .globl cuStreamSynchronize_ptsz
  11658. .p2align 4
  11659. .type cuStreamSynchronize_ptsz, %function
  11660. cuStreamSynchronize_ptsz:
  11661. .cfi_startproc
  11662. // Intel opt. manual says to
  11663. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11664. // to hint static predictor.
  11665. cmpq $0, _libcuda_so_tramp_table+3304(%rip)
  11666. je 2f
  11667. 1:
  11668. jmp *_libcuda_so_tramp_table+3304(%rip)
  11669. 2:
  11670. pushq $413
  11671. .cfi_adjust_cfa_offset 8
  11672. call _libcuda_so_save_regs_and_resolve
  11673. addq $8, %rsp
  11674. .cfi_adjust_cfa_offset -8
  11675. jmp 1b
  11676. .cfi_endproc
  11677. /*
  11678. * Copyright 2018-2019 Yury Gribov
  11679. *
  11680. * The MIT License (MIT)
  11681. *
  11682. * Use of this source code is governed by MIT license that can be
  11683. * found in the LICENSE.txt file.
  11684. */
  11685. .globl cuStreamWaitEvent
  11686. .p2align 4
  11687. .type cuStreamWaitEvent, %function
  11688. cuStreamWaitEvent:
  11689. .cfi_startproc
  11690. // Intel opt. manual says to
  11691. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11692. // to hint static predictor.
  11693. cmpq $0, _libcuda_so_tramp_table+3312(%rip)
  11694. je 2f
  11695. 1:
  11696. jmp *_libcuda_so_tramp_table+3312(%rip)
  11697. 2:
  11698. pushq $414
  11699. .cfi_adjust_cfa_offset 8
  11700. call _libcuda_so_save_regs_and_resolve
  11701. addq $8, %rsp
  11702. .cfi_adjust_cfa_offset -8
  11703. jmp 1b
  11704. .cfi_endproc
  11705. /*
  11706. * Copyright 2018-2019 Yury Gribov
  11707. *
  11708. * The MIT License (MIT)
  11709. *
  11710. * Use of this source code is governed by MIT license that can be
  11711. * found in the LICENSE.txt file.
  11712. */
  11713. .globl cuStreamWaitEvent_ptsz
  11714. .p2align 4
  11715. .type cuStreamWaitEvent_ptsz, %function
  11716. cuStreamWaitEvent_ptsz:
  11717. .cfi_startproc
  11718. // Intel opt. manual says to
  11719. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11720. // to hint static predictor.
  11721. cmpq $0, _libcuda_so_tramp_table+3320(%rip)
  11722. je 2f
  11723. 1:
  11724. jmp *_libcuda_so_tramp_table+3320(%rip)
  11725. 2:
  11726. pushq $415
  11727. .cfi_adjust_cfa_offset 8
  11728. call _libcuda_so_save_regs_and_resolve
  11729. addq $8, %rsp
  11730. .cfi_adjust_cfa_offset -8
  11731. jmp 1b
  11732. .cfi_endproc
  11733. /*
  11734. * Copyright 2018-2019 Yury Gribov
  11735. *
  11736. * The MIT License (MIT)
  11737. *
  11738. * Use of this source code is governed by MIT license that can be
  11739. * found in the LICENSE.txt file.
  11740. */
  11741. .globl cuStreamWaitValue32
  11742. .p2align 4
  11743. .type cuStreamWaitValue32, %function
  11744. cuStreamWaitValue32:
  11745. .cfi_startproc
  11746. // Intel opt. manual says to
  11747. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11748. // to hint static predictor.
  11749. cmpq $0, _libcuda_so_tramp_table+3328(%rip)
  11750. je 2f
  11751. 1:
  11752. jmp *_libcuda_so_tramp_table+3328(%rip)
  11753. 2:
  11754. pushq $416
  11755. .cfi_adjust_cfa_offset 8
  11756. call _libcuda_so_save_regs_and_resolve
  11757. addq $8, %rsp
  11758. .cfi_adjust_cfa_offset -8
  11759. jmp 1b
  11760. .cfi_endproc
  11761. /*
  11762. * Copyright 2018-2019 Yury Gribov
  11763. *
  11764. * The MIT License (MIT)
  11765. *
  11766. * Use of this source code is governed by MIT license that can be
  11767. * found in the LICENSE.txt file.
  11768. */
  11769. .globl cuStreamWaitValue32_ptsz
  11770. .p2align 4
  11771. .type cuStreamWaitValue32_ptsz, %function
  11772. cuStreamWaitValue32_ptsz:
  11773. .cfi_startproc
  11774. // Intel opt. manual says to
  11775. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11776. // to hint static predictor.
  11777. cmpq $0, _libcuda_so_tramp_table+3336(%rip)
  11778. je 2f
  11779. 1:
  11780. jmp *_libcuda_so_tramp_table+3336(%rip)
  11781. 2:
  11782. pushq $417
  11783. .cfi_adjust_cfa_offset 8
  11784. call _libcuda_so_save_regs_and_resolve
  11785. addq $8, %rsp
  11786. .cfi_adjust_cfa_offset -8
  11787. jmp 1b
  11788. .cfi_endproc
  11789. /*
  11790. * Copyright 2018-2019 Yury Gribov
  11791. *
  11792. * The MIT License (MIT)
  11793. *
  11794. * Use of this source code is governed by MIT license that can be
  11795. * found in the LICENSE.txt file.
  11796. */
  11797. .globl cuStreamWaitValue64
  11798. .p2align 4
  11799. .type cuStreamWaitValue64, %function
  11800. cuStreamWaitValue64:
  11801. .cfi_startproc
  11802. // Intel opt. manual says to
  11803. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11804. // to hint static predictor.
  11805. cmpq $0, _libcuda_so_tramp_table+3344(%rip)
  11806. je 2f
  11807. 1:
  11808. jmp *_libcuda_so_tramp_table+3344(%rip)
  11809. 2:
  11810. pushq $418
  11811. .cfi_adjust_cfa_offset 8
  11812. call _libcuda_so_save_regs_and_resolve
  11813. addq $8, %rsp
  11814. .cfi_adjust_cfa_offset -8
  11815. jmp 1b
  11816. .cfi_endproc
  11817. /*
  11818. * Copyright 2018-2019 Yury Gribov
  11819. *
  11820. * The MIT License (MIT)
  11821. *
  11822. * Use of this source code is governed by MIT license that can be
  11823. * found in the LICENSE.txt file.
  11824. */
  11825. .globl cuStreamWaitValue64_ptsz
  11826. .p2align 4
  11827. .type cuStreamWaitValue64_ptsz, %function
  11828. cuStreamWaitValue64_ptsz:
  11829. .cfi_startproc
  11830. // Intel opt. manual says to
  11831. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11832. // to hint static predictor.
  11833. cmpq $0, _libcuda_so_tramp_table+3352(%rip)
  11834. je 2f
  11835. 1:
  11836. jmp *_libcuda_so_tramp_table+3352(%rip)
  11837. 2:
  11838. pushq $419
  11839. .cfi_adjust_cfa_offset 8
  11840. call _libcuda_so_save_regs_and_resolve
  11841. addq $8, %rsp
  11842. .cfi_adjust_cfa_offset -8
  11843. jmp 1b
  11844. .cfi_endproc
  11845. /*
  11846. * Copyright 2018-2019 Yury Gribov
  11847. *
  11848. * The MIT License (MIT)
  11849. *
  11850. * Use of this source code is governed by MIT license that can be
  11851. * found in the LICENSE.txt file.
  11852. */
  11853. .globl cuStreamWriteValue32
  11854. .p2align 4
  11855. .type cuStreamWriteValue32, %function
  11856. cuStreamWriteValue32:
  11857. .cfi_startproc
  11858. // Intel opt. manual says to
  11859. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11860. // to hint static predictor.
  11861. cmpq $0, _libcuda_so_tramp_table+3360(%rip)
  11862. je 2f
  11863. 1:
  11864. jmp *_libcuda_so_tramp_table+3360(%rip)
  11865. 2:
  11866. pushq $420
  11867. .cfi_adjust_cfa_offset 8
  11868. call _libcuda_so_save_regs_and_resolve
  11869. addq $8, %rsp
  11870. .cfi_adjust_cfa_offset -8
  11871. jmp 1b
  11872. .cfi_endproc
  11873. /*
  11874. * Copyright 2018-2019 Yury Gribov
  11875. *
  11876. * The MIT License (MIT)
  11877. *
  11878. * Use of this source code is governed by MIT license that can be
  11879. * found in the LICENSE.txt file.
  11880. */
  11881. .globl cuStreamWriteValue32_ptsz
  11882. .p2align 4
  11883. .type cuStreamWriteValue32_ptsz, %function
  11884. cuStreamWriteValue32_ptsz:
  11885. .cfi_startproc
  11886. // Intel opt. manual says to
  11887. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11888. // to hint static predictor.
  11889. cmpq $0, _libcuda_so_tramp_table+3368(%rip)
  11890. je 2f
  11891. 1:
  11892. jmp *_libcuda_so_tramp_table+3368(%rip)
  11893. 2:
  11894. pushq $421
  11895. .cfi_adjust_cfa_offset 8
  11896. call _libcuda_so_save_regs_and_resolve
  11897. addq $8, %rsp
  11898. .cfi_adjust_cfa_offset -8
  11899. jmp 1b
  11900. .cfi_endproc
  11901. /*
  11902. * Copyright 2018-2019 Yury Gribov
  11903. *
  11904. * The MIT License (MIT)
  11905. *
  11906. * Use of this source code is governed by MIT license that can be
  11907. * found in the LICENSE.txt file.
  11908. */
  11909. .globl cuStreamWriteValue64
  11910. .p2align 4
  11911. .type cuStreamWriteValue64, %function
  11912. cuStreamWriteValue64:
  11913. .cfi_startproc
  11914. // Intel opt. manual says to
  11915. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11916. // to hint static predictor.
  11917. cmpq $0, _libcuda_so_tramp_table+3376(%rip)
  11918. je 2f
  11919. 1:
  11920. jmp *_libcuda_so_tramp_table+3376(%rip)
  11921. 2:
  11922. pushq $422
  11923. .cfi_adjust_cfa_offset 8
  11924. call _libcuda_so_save_regs_and_resolve
  11925. addq $8, %rsp
  11926. .cfi_adjust_cfa_offset -8
  11927. jmp 1b
  11928. .cfi_endproc
  11929. /*
  11930. * Copyright 2018-2019 Yury Gribov
  11931. *
  11932. * The MIT License (MIT)
  11933. *
  11934. * Use of this source code is governed by MIT license that can be
  11935. * found in the LICENSE.txt file.
  11936. */
  11937. .globl cuStreamWriteValue64_ptsz
  11938. .p2align 4
  11939. .type cuStreamWriteValue64_ptsz, %function
  11940. cuStreamWriteValue64_ptsz:
  11941. .cfi_startproc
  11942. // Intel opt. manual says to
  11943. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11944. // to hint static predictor.
  11945. cmpq $0, _libcuda_so_tramp_table+3384(%rip)
  11946. je 2f
  11947. 1:
  11948. jmp *_libcuda_so_tramp_table+3384(%rip)
  11949. 2:
  11950. pushq $423
  11951. .cfi_adjust_cfa_offset 8
  11952. call _libcuda_so_save_regs_and_resolve
  11953. addq $8, %rsp
  11954. .cfi_adjust_cfa_offset -8
  11955. jmp 1b
  11956. .cfi_endproc
  11957. /*
  11958. * Copyright 2018-2019 Yury Gribov
  11959. *
  11960. * The MIT License (MIT)
  11961. *
  11962. * Use of this source code is governed by MIT license that can be
  11963. * found in the LICENSE.txt file.
  11964. */
  11965. .globl cuSurfObjectCreate
  11966. .p2align 4
  11967. .type cuSurfObjectCreate, %function
  11968. cuSurfObjectCreate:
  11969. .cfi_startproc
  11970. // Intel opt. manual says to
  11971. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  11972. // to hint static predictor.
  11973. cmpq $0, _libcuda_so_tramp_table+3392(%rip)
  11974. je 2f
  11975. 1:
  11976. jmp *_libcuda_so_tramp_table+3392(%rip)
  11977. 2:
  11978. pushq $424
  11979. .cfi_adjust_cfa_offset 8
  11980. call _libcuda_so_save_regs_and_resolve
  11981. addq $8, %rsp
  11982. .cfi_adjust_cfa_offset -8
  11983. jmp 1b
  11984. .cfi_endproc
  11985. /*
  11986. * Copyright 2018-2019 Yury Gribov
  11987. *
  11988. * The MIT License (MIT)
  11989. *
  11990. * Use of this source code is governed by MIT license that can be
  11991. * found in the LICENSE.txt file.
  11992. */
  11993. .globl cuSurfObjectDestroy
  11994. .p2align 4
  11995. .type cuSurfObjectDestroy, %function
  11996. cuSurfObjectDestroy:
  11997. .cfi_startproc
  11998. // Intel opt. manual says to
  11999. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12000. // to hint static predictor.
  12001. cmpq $0, _libcuda_so_tramp_table+3400(%rip)
  12002. je 2f
  12003. 1:
  12004. jmp *_libcuda_so_tramp_table+3400(%rip)
  12005. 2:
  12006. pushq $425
  12007. .cfi_adjust_cfa_offset 8
  12008. call _libcuda_so_save_regs_and_resolve
  12009. addq $8, %rsp
  12010. .cfi_adjust_cfa_offset -8
  12011. jmp 1b
  12012. .cfi_endproc
  12013. /*
  12014. * Copyright 2018-2019 Yury Gribov
  12015. *
  12016. * The MIT License (MIT)
  12017. *
  12018. * Use of this source code is governed by MIT license that can be
  12019. * found in the LICENSE.txt file.
  12020. */
  12021. .globl cuSurfObjectGetResourceDesc
  12022. .p2align 4
  12023. .type cuSurfObjectGetResourceDesc, %function
  12024. cuSurfObjectGetResourceDesc:
  12025. .cfi_startproc
  12026. // Intel opt. manual says to
  12027. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12028. // to hint static predictor.
  12029. cmpq $0, _libcuda_so_tramp_table+3408(%rip)
  12030. je 2f
  12031. 1:
  12032. jmp *_libcuda_so_tramp_table+3408(%rip)
  12033. 2:
  12034. pushq $426
  12035. .cfi_adjust_cfa_offset 8
  12036. call _libcuda_so_save_regs_and_resolve
  12037. addq $8, %rsp
  12038. .cfi_adjust_cfa_offset -8
  12039. jmp 1b
  12040. .cfi_endproc
  12041. /*
  12042. * Copyright 2018-2019 Yury Gribov
  12043. *
  12044. * The MIT License (MIT)
  12045. *
  12046. * Use of this source code is governed by MIT license that can be
  12047. * found in the LICENSE.txt file.
  12048. */
  12049. .globl cuSurfRefGetArray
  12050. .p2align 4
  12051. .type cuSurfRefGetArray, %function
  12052. cuSurfRefGetArray:
  12053. .cfi_startproc
  12054. // Intel opt. manual says to
  12055. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12056. // to hint static predictor.
  12057. cmpq $0, _libcuda_so_tramp_table+3416(%rip)
  12058. je 2f
  12059. 1:
  12060. jmp *_libcuda_so_tramp_table+3416(%rip)
  12061. 2:
  12062. pushq $427
  12063. .cfi_adjust_cfa_offset 8
  12064. call _libcuda_so_save_regs_and_resolve
  12065. addq $8, %rsp
  12066. .cfi_adjust_cfa_offset -8
  12067. jmp 1b
  12068. .cfi_endproc
  12069. /*
  12070. * Copyright 2018-2019 Yury Gribov
  12071. *
  12072. * The MIT License (MIT)
  12073. *
  12074. * Use of this source code is governed by MIT license that can be
  12075. * found in the LICENSE.txt file.
  12076. */
  12077. .globl cuSurfRefSetArray
  12078. .p2align 4
  12079. .type cuSurfRefSetArray, %function
  12080. cuSurfRefSetArray:
  12081. .cfi_startproc
  12082. // Intel opt. manual says to
  12083. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12084. // to hint static predictor.
  12085. cmpq $0, _libcuda_so_tramp_table+3424(%rip)
  12086. je 2f
  12087. 1:
  12088. jmp *_libcuda_so_tramp_table+3424(%rip)
  12089. 2:
  12090. pushq $428
  12091. .cfi_adjust_cfa_offset 8
  12092. call _libcuda_so_save_regs_and_resolve
  12093. addq $8, %rsp
  12094. .cfi_adjust_cfa_offset -8
  12095. jmp 1b
  12096. .cfi_endproc
  12097. /*
  12098. * Copyright 2018-2019 Yury Gribov
  12099. *
  12100. * The MIT License (MIT)
  12101. *
  12102. * Use of this source code is governed by MIT license that can be
  12103. * found in the LICENSE.txt file.
  12104. */
  12105. .globl cuTexObjectCreate
  12106. .p2align 4
  12107. .type cuTexObjectCreate, %function
  12108. cuTexObjectCreate:
  12109. .cfi_startproc
  12110. // Intel opt. manual says to
  12111. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12112. // to hint static predictor.
  12113. cmpq $0, _libcuda_so_tramp_table+3432(%rip)
  12114. je 2f
  12115. 1:
  12116. jmp *_libcuda_so_tramp_table+3432(%rip)
  12117. 2:
  12118. pushq $429
  12119. .cfi_adjust_cfa_offset 8
  12120. call _libcuda_so_save_regs_and_resolve
  12121. addq $8, %rsp
  12122. .cfi_adjust_cfa_offset -8
  12123. jmp 1b
  12124. .cfi_endproc
  12125. /*
  12126. * Copyright 2018-2019 Yury Gribov
  12127. *
  12128. * The MIT License (MIT)
  12129. *
  12130. * Use of this source code is governed by MIT license that can be
  12131. * found in the LICENSE.txt file.
  12132. */
  12133. .globl cuTexObjectDestroy
  12134. .p2align 4
  12135. .type cuTexObjectDestroy, %function
  12136. cuTexObjectDestroy:
  12137. .cfi_startproc
  12138. // Intel opt. manual says to
  12139. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12140. // to hint static predictor.
  12141. cmpq $0, _libcuda_so_tramp_table+3440(%rip)
  12142. je 2f
  12143. 1:
  12144. jmp *_libcuda_so_tramp_table+3440(%rip)
  12145. 2:
  12146. pushq $430
  12147. .cfi_adjust_cfa_offset 8
  12148. call _libcuda_so_save_regs_and_resolve
  12149. addq $8, %rsp
  12150. .cfi_adjust_cfa_offset -8
  12151. jmp 1b
  12152. .cfi_endproc
  12153. /*
  12154. * Copyright 2018-2019 Yury Gribov
  12155. *
  12156. * The MIT License (MIT)
  12157. *
  12158. * Use of this source code is governed by MIT license that can be
  12159. * found in the LICENSE.txt file.
  12160. */
  12161. .globl cuTexObjectGetResourceDesc
  12162. .p2align 4
  12163. .type cuTexObjectGetResourceDesc, %function
  12164. cuTexObjectGetResourceDesc:
  12165. .cfi_startproc
  12166. // Intel opt. manual says to
  12167. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12168. // to hint static predictor.
  12169. cmpq $0, _libcuda_so_tramp_table+3448(%rip)
  12170. je 2f
  12171. 1:
  12172. jmp *_libcuda_so_tramp_table+3448(%rip)
  12173. 2:
  12174. pushq $431
  12175. .cfi_adjust_cfa_offset 8
  12176. call _libcuda_so_save_regs_and_resolve
  12177. addq $8, %rsp
  12178. .cfi_adjust_cfa_offset -8
  12179. jmp 1b
  12180. .cfi_endproc
  12181. /*
  12182. * Copyright 2018-2019 Yury Gribov
  12183. *
  12184. * The MIT License (MIT)
  12185. *
  12186. * Use of this source code is governed by MIT license that can be
  12187. * found in the LICENSE.txt file.
  12188. */
  12189. .globl cuTexObjectGetResourceViewDesc
  12190. .p2align 4
  12191. .type cuTexObjectGetResourceViewDesc, %function
  12192. cuTexObjectGetResourceViewDesc:
  12193. .cfi_startproc
  12194. // Intel opt. manual says to
  12195. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12196. // to hint static predictor.
  12197. cmpq $0, _libcuda_so_tramp_table+3456(%rip)
  12198. je 2f
  12199. 1:
  12200. jmp *_libcuda_so_tramp_table+3456(%rip)
  12201. 2:
  12202. pushq $432
  12203. .cfi_adjust_cfa_offset 8
  12204. call _libcuda_so_save_regs_and_resolve
  12205. addq $8, %rsp
  12206. .cfi_adjust_cfa_offset -8
  12207. jmp 1b
  12208. .cfi_endproc
  12209. /*
  12210. * Copyright 2018-2019 Yury Gribov
  12211. *
  12212. * The MIT License (MIT)
  12213. *
  12214. * Use of this source code is governed by MIT license that can be
  12215. * found in the LICENSE.txt file.
  12216. */
  12217. .globl cuTexObjectGetTextureDesc
  12218. .p2align 4
  12219. .type cuTexObjectGetTextureDesc, %function
  12220. cuTexObjectGetTextureDesc:
  12221. .cfi_startproc
  12222. // Intel opt. manual says to
  12223. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12224. // to hint static predictor.
  12225. cmpq $0, _libcuda_so_tramp_table+3464(%rip)
  12226. je 2f
  12227. 1:
  12228. jmp *_libcuda_so_tramp_table+3464(%rip)
  12229. 2:
  12230. pushq $433
  12231. .cfi_adjust_cfa_offset 8
  12232. call _libcuda_so_save_regs_and_resolve
  12233. addq $8, %rsp
  12234. .cfi_adjust_cfa_offset -8
  12235. jmp 1b
  12236. .cfi_endproc
  12237. /*
  12238. * Copyright 2018-2019 Yury Gribov
  12239. *
  12240. * The MIT License (MIT)
  12241. *
  12242. * Use of this source code is governed by MIT license that can be
  12243. * found in the LICENSE.txt file.
  12244. */
  12245. .globl cuTexRefCreate
  12246. .p2align 4
  12247. .type cuTexRefCreate, %function
  12248. cuTexRefCreate:
  12249. .cfi_startproc
  12250. // Intel opt. manual says to
  12251. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12252. // to hint static predictor.
  12253. cmpq $0, _libcuda_so_tramp_table+3472(%rip)
  12254. je 2f
  12255. 1:
  12256. jmp *_libcuda_so_tramp_table+3472(%rip)
  12257. 2:
  12258. pushq $434
  12259. .cfi_adjust_cfa_offset 8
  12260. call _libcuda_so_save_regs_and_resolve
  12261. addq $8, %rsp
  12262. .cfi_adjust_cfa_offset -8
  12263. jmp 1b
  12264. .cfi_endproc
  12265. /*
  12266. * Copyright 2018-2019 Yury Gribov
  12267. *
  12268. * The MIT License (MIT)
  12269. *
  12270. * Use of this source code is governed by MIT license that can be
  12271. * found in the LICENSE.txt file.
  12272. */
  12273. .globl cuTexRefDestroy
  12274. .p2align 4
  12275. .type cuTexRefDestroy, %function
  12276. cuTexRefDestroy:
  12277. .cfi_startproc
  12278. // Intel opt. manual says to
  12279. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12280. // to hint static predictor.
  12281. cmpq $0, _libcuda_so_tramp_table+3480(%rip)
  12282. je 2f
  12283. 1:
  12284. jmp *_libcuda_so_tramp_table+3480(%rip)
  12285. 2:
  12286. pushq $435
  12287. .cfi_adjust_cfa_offset 8
  12288. call _libcuda_so_save_regs_and_resolve
  12289. addq $8, %rsp
  12290. .cfi_adjust_cfa_offset -8
  12291. jmp 1b
  12292. .cfi_endproc
  12293. /*
  12294. * Copyright 2018-2019 Yury Gribov
  12295. *
  12296. * The MIT License (MIT)
  12297. *
  12298. * Use of this source code is governed by MIT license that can be
  12299. * found in the LICENSE.txt file.
  12300. */
  12301. .globl cuTexRefGetAddress
  12302. .p2align 4
  12303. .type cuTexRefGetAddress, %function
  12304. cuTexRefGetAddress:
  12305. .cfi_startproc
  12306. // Intel opt. manual says to
  12307. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12308. // to hint static predictor.
  12309. cmpq $0, _libcuda_so_tramp_table+3488(%rip)
  12310. je 2f
  12311. 1:
  12312. jmp *_libcuda_so_tramp_table+3488(%rip)
  12313. 2:
  12314. pushq $436
  12315. .cfi_adjust_cfa_offset 8
  12316. call _libcuda_so_save_regs_and_resolve
  12317. addq $8, %rsp
  12318. .cfi_adjust_cfa_offset -8
  12319. jmp 1b
  12320. .cfi_endproc
  12321. /*
  12322. * Copyright 2018-2019 Yury Gribov
  12323. *
  12324. * The MIT License (MIT)
  12325. *
  12326. * Use of this source code is governed by MIT license that can be
  12327. * found in the LICENSE.txt file.
  12328. */
  12329. .globl cuTexRefGetAddressMode
  12330. .p2align 4
  12331. .type cuTexRefGetAddressMode, %function
  12332. cuTexRefGetAddressMode:
  12333. .cfi_startproc
  12334. // Intel opt. manual says to
  12335. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12336. // to hint static predictor.
  12337. cmpq $0, _libcuda_so_tramp_table+3496(%rip)
  12338. je 2f
  12339. 1:
  12340. jmp *_libcuda_so_tramp_table+3496(%rip)
  12341. 2:
  12342. pushq $437
  12343. .cfi_adjust_cfa_offset 8
  12344. call _libcuda_so_save_regs_and_resolve
  12345. addq $8, %rsp
  12346. .cfi_adjust_cfa_offset -8
  12347. jmp 1b
  12348. .cfi_endproc
  12349. /*
  12350. * Copyright 2018-2019 Yury Gribov
  12351. *
  12352. * The MIT License (MIT)
  12353. *
  12354. * Use of this source code is governed by MIT license that can be
  12355. * found in the LICENSE.txt file.
  12356. */
  12357. .globl cuTexRefGetAddress_v2
  12358. .p2align 4
  12359. .type cuTexRefGetAddress_v2, %function
  12360. cuTexRefGetAddress_v2:
  12361. .cfi_startproc
  12362. // Intel opt. manual says to
  12363. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12364. // to hint static predictor.
  12365. cmpq $0, _libcuda_so_tramp_table+3504(%rip)
  12366. je 2f
  12367. 1:
  12368. jmp *_libcuda_so_tramp_table+3504(%rip)
  12369. 2:
  12370. pushq $438
  12371. .cfi_adjust_cfa_offset 8
  12372. call _libcuda_so_save_regs_and_resolve
  12373. addq $8, %rsp
  12374. .cfi_adjust_cfa_offset -8
  12375. jmp 1b
  12376. .cfi_endproc
  12377. /*
  12378. * Copyright 2018-2019 Yury Gribov
  12379. *
  12380. * The MIT License (MIT)
  12381. *
  12382. * Use of this source code is governed by MIT license that can be
  12383. * found in the LICENSE.txt file.
  12384. */
  12385. .globl cuTexRefGetArray
  12386. .p2align 4
  12387. .type cuTexRefGetArray, %function
  12388. cuTexRefGetArray:
  12389. .cfi_startproc
  12390. // Intel opt. manual says to
  12391. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12392. // to hint static predictor.
  12393. cmpq $0, _libcuda_so_tramp_table+3512(%rip)
  12394. je 2f
  12395. 1:
  12396. jmp *_libcuda_so_tramp_table+3512(%rip)
  12397. 2:
  12398. pushq $439
  12399. .cfi_adjust_cfa_offset 8
  12400. call _libcuda_so_save_regs_and_resolve
  12401. addq $8, %rsp
  12402. .cfi_adjust_cfa_offset -8
  12403. jmp 1b
  12404. .cfi_endproc
  12405. /*
  12406. * Copyright 2018-2019 Yury Gribov
  12407. *
  12408. * The MIT License (MIT)
  12409. *
  12410. * Use of this source code is governed by MIT license that can be
  12411. * found in the LICENSE.txt file.
  12412. */
  12413. .globl cuTexRefGetBorderColor
  12414. .p2align 4
  12415. .type cuTexRefGetBorderColor, %function
  12416. cuTexRefGetBorderColor:
  12417. .cfi_startproc
  12418. // Intel opt. manual says to
  12419. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12420. // to hint static predictor.
  12421. cmpq $0, _libcuda_so_tramp_table+3520(%rip)
  12422. je 2f
  12423. 1:
  12424. jmp *_libcuda_so_tramp_table+3520(%rip)
  12425. 2:
  12426. pushq $440
  12427. .cfi_adjust_cfa_offset 8
  12428. call _libcuda_so_save_regs_and_resolve
  12429. addq $8, %rsp
  12430. .cfi_adjust_cfa_offset -8
  12431. jmp 1b
  12432. .cfi_endproc
  12433. /*
  12434. * Copyright 2018-2019 Yury Gribov
  12435. *
  12436. * The MIT License (MIT)
  12437. *
  12438. * Use of this source code is governed by MIT license that can be
  12439. * found in the LICENSE.txt file.
  12440. */
  12441. .globl cuTexRefGetFilterMode
  12442. .p2align 4
  12443. .type cuTexRefGetFilterMode, %function
  12444. cuTexRefGetFilterMode:
  12445. .cfi_startproc
  12446. // Intel opt. manual says to
  12447. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12448. // to hint static predictor.
  12449. cmpq $0, _libcuda_so_tramp_table+3528(%rip)
  12450. je 2f
  12451. 1:
  12452. jmp *_libcuda_so_tramp_table+3528(%rip)
  12453. 2:
  12454. pushq $441
  12455. .cfi_adjust_cfa_offset 8
  12456. call _libcuda_so_save_regs_and_resolve
  12457. addq $8, %rsp
  12458. .cfi_adjust_cfa_offset -8
  12459. jmp 1b
  12460. .cfi_endproc
  12461. /*
  12462. * Copyright 2018-2019 Yury Gribov
  12463. *
  12464. * The MIT License (MIT)
  12465. *
  12466. * Use of this source code is governed by MIT license that can be
  12467. * found in the LICENSE.txt file.
  12468. */
  12469. .globl cuTexRefGetFlags
  12470. .p2align 4
  12471. .type cuTexRefGetFlags, %function
  12472. cuTexRefGetFlags:
  12473. .cfi_startproc
  12474. // Intel opt. manual says to
  12475. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12476. // to hint static predictor.
  12477. cmpq $0, _libcuda_so_tramp_table+3536(%rip)
  12478. je 2f
  12479. 1:
  12480. jmp *_libcuda_so_tramp_table+3536(%rip)
  12481. 2:
  12482. pushq $442
  12483. .cfi_adjust_cfa_offset 8
  12484. call _libcuda_so_save_regs_and_resolve
  12485. addq $8, %rsp
  12486. .cfi_adjust_cfa_offset -8
  12487. jmp 1b
  12488. .cfi_endproc
  12489. /*
  12490. * Copyright 2018-2019 Yury Gribov
  12491. *
  12492. * The MIT License (MIT)
  12493. *
  12494. * Use of this source code is governed by MIT license that can be
  12495. * found in the LICENSE.txt file.
  12496. */
  12497. .globl cuTexRefGetFormat
  12498. .p2align 4
  12499. .type cuTexRefGetFormat, %function
  12500. cuTexRefGetFormat:
  12501. .cfi_startproc
  12502. // Intel opt. manual says to
  12503. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12504. // to hint static predictor.
  12505. cmpq $0, _libcuda_so_tramp_table+3544(%rip)
  12506. je 2f
  12507. 1:
  12508. jmp *_libcuda_so_tramp_table+3544(%rip)
  12509. 2:
  12510. pushq $443
  12511. .cfi_adjust_cfa_offset 8
  12512. call _libcuda_so_save_regs_and_resolve
  12513. addq $8, %rsp
  12514. .cfi_adjust_cfa_offset -8
  12515. jmp 1b
  12516. .cfi_endproc
  12517. /*
  12518. * Copyright 2018-2019 Yury Gribov
  12519. *
  12520. * The MIT License (MIT)
  12521. *
  12522. * Use of this source code is governed by MIT license that can be
  12523. * found in the LICENSE.txt file.
  12524. */
  12525. .globl cuTexRefGetMaxAnisotropy
  12526. .p2align 4
  12527. .type cuTexRefGetMaxAnisotropy, %function
  12528. cuTexRefGetMaxAnisotropy:
  12529. .cfi_startproc
  12530. // Intel opt. manual says to
  12531. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12532. // to hint static predictor.
  12533. cmpq $0, _libcuda_so_tramp_table+3552(%rip)
  12534. je 2f
  12535. 1:
  12536. jmp *_libcuda_so_tramp_table+3552(%rip)
  12537. 2:
  12538. pushq $444
  12539. .cfi_adjust_cfa_offset 8
  12540. call _libcuda_so_save_regs_and_resolve
  12541. addq $8, %rsp
  12542. .cfi_adjust_cfa_offset -8
  12543. jmp 1b
  12544. .cfi_endproc
  12545. /*
  12546. * Copyright 2018-2019 Yury Gribov
  12547. *
  12548. * The MIT License (MIT)
  12549. *
  12550. * Use of this source code is governed by MIT license that can be
  12551. * found in the LICENSE.txt file.
  12552. */
  12553. .globl cuTexRefGetMipmapFilterMode
  12554. .p2align 4
  12555. .type cuTexRefGetMipmapFilterMode, %function
  12556. cuTexRefGetMipmapFilterMode:
  12557. .cfi_startproc
  12558. // Intel opt. manual says to
  12559. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12560. // to hint static predictor.
  12561. cmpq $0, _libcuda_so_tramp_table+3560(%rip)
  12562. je 2f
  12563. 1:
  12564. jmp *_libcuda_so_tramp_table+3560(%rip)
  12565. 2:
  12566. pushq $445
  12567. .cfi_adjust_cfa_offset 8
  12568. call _libcuda_so_save_regs_and_resolve
  12569. addq $8, %rsp
  12570. .cfi_adjust_cfa_offset -8
  12571. jmp 1b
  12572. .cfi_endproc
  12573. /*
  12574. * Copyright 2018-2019 Yury Gribov
  12575. *
  12576. * The MIT License (MIT)
  12577. *
  12578. * Use of this source code is governed by MIT license that can be
  12579. * found in the LICENSE.txt file.
  12580. */
  12581. .globl cuTexRefGetMipmapLevelBias
  12582. .p2align 4
  12583. .type cuTexRefGetMipmapLevelBias, %function
  12584. cuTexRefGetMipmapLevelBias:
  12585. .cfi_startproc
  12586. // Intel opt. manual says to
  12587. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12588. // to hint static predictor.
  12589. cmpq $0, _libcuda_so_tramp_table+3568(%rip)
  12590. je 2f
  12591. 1:
  12592. jmp *_libcuda_so_tramp_table+3568(%rip)
  12593. 2:
  12594. pushq $446
  12595. .cfi_adjust_cfa_offset 8
  12596. call _libcuda_so_save_regs_and_resolve
  12597. addq $8, %rsp
  12598. .cfi_adjust_cfa_offset -8
  12599. jmp 1b
  12600. .cfi_endproc
  12601. /*
  12602. * Copyright 2018-2019 Yury Gribov
  12603. *
  12604. * The MIT License (MIT)
  12605. *
  12606. * Use of this source code is governed by MIT license that can be
  12607. * found in the LICENSE.txt file.
  12608. */
  12609. .globl cuTexRefGetMipmapLevelClamp
  12610. .p2align 4
  12611. .type cuTexRefGetMipmapLevelClamp, %function
  12612. cuTexRefGetMipmapLevelClamp:
  12613. .cfi_startproc
  12614. // Intel opt. manual says to
  12615. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12616. // to hint static predictor.
  12617. cmpq $0, _libcuda_so_tramp_table+3576(%rip)
  12618. je 2f
  12619. 1:
  12620. jmp *_libcuda_so_tramp_table+3576(%rip)
  12621. 2:
  12622. pushq $447
  12623. .cfi_adjust_cfa_offset 8
  12624. call _libcuda_so_save_regs_and_resolve
  12625. addq $8, %rsp
  12626. .cfi_adjust_cfa_offset -8
  12627. jmp 1b
  12628. .cfi_endproc
  12629. /*
  12630. * Copyright 2018-2019 Yury Gribov
  12631. *
  12632. * The MIT License (MIT)
  12633. *
  12634. * Use of this source code is governed by MIT license that can be
  12635. * found in the LICENSE.txt file.
  12636. */
  12637. .globl cuTexRefGetMipmappedArray
  12638. .p2align 4
  12639. .type cuTexRefGetMipmappedArray, %function
  12640. cuTexRefGetMipmappedArray:
  12641. .cfi_startproc
  12642. // Intel opt. manual says to
  12643. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12644. // to hint static predictor.
  12645. cmpq $0, _libcuda_so_tramp_table+3584(%rip)
  12646. je 2f
  12647. 1:
  12648. jmp *_libcuda_so_tramp_table+3584(%rip)
  12649. 2:
  12650. pushq $448
  12651. .cfi_adjust_cfa_offset 8
  12652. call _libcuda_so_save_regs_and_resolve
  12653. addq $8, %rsp
  12654. .cfi_adjust_cfa_offset -8
  12655. jmp 1b
  12656. .cfi_endproc
  12657. /*
  12658. * Copyright 2018-2019 Yury Gribov
  12659. *
  12660. * The MIT License (MIT)
  12661. *
  12662. * Use of this source code is governed by MIT license that can be
  12663. * found in the LICENSE.txt file.
  12664. */
  12665. .globl cuTexRefSetAddress
  12666. .p2align 4
  12667. .type cuTexRefSetAddress, %function
  12668. cuTexRefSetAddress:
  12669. .cfi_startproc
  12670. // Intel opt. manual says to
  12671. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12672. // to hint static predictor.
  12673. cmpq $0, _libcuda_so_tramp_table+3592(%rip)
  12674. je 2f
  12675. 1:
  12676. jmp *_libcuda_so_tramp_table+3592(%rip)
  12677. 2:
  12678. pushq $449
  12679. .cfi_adjust_cfa_offset 8
  12680. call _libcuda_so_save_regs_and_resolve
  12681. addq $8, %rsp
  12682. .cfi_adjust_cfa_offset -8
  12683. jmp 1b
  12684. .cfi_endproc
  12685. /*
  12686. * Copyright 2018-2019 Yury Gribov
  12687. *
  12688. * The MIT License (MIT)
  12689. *
  12690. * Use of this source code is governed by MIT license that can be
  12691. * found in the LICENSE.txt file.
  12692. */
  12693. .globl cuTexRefSetAddress2D
  12694. .p2align 4
  12695. .type cuTexRefSetAddress2D, %function
  12696. cuTexRefSetAddress2D:
  12697. .cfi_startproc
  12698. // Intel opt. manual says to
  12699. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12700. // to hint static predictor.
  12701. cmpq $0, _libcuda_so_tramp_table+3600(%rip)
  12702. je 2f
  12703. 1:
  12704. jmp *_libcuda_so_tramp_table+3600(%rip)
  12705. 2:
  12706. pushq $450
  12707. .cfi_adjust_cfa_offset 8
  12708. call _libcuda_so_save_regs_and_resolve
  12709. addq $8, %rsp
  12710. .cfi_adjust_cfa_offset -8
  12711. jmp 1b
  12712. .cfi_endproc
  12713. /*
  12714. * Copyright 2018-2019 Yury Gribov
  12715. *
  12716. * The MIT License (MIT)
  12717. *
  12718. * Use of this source code is governed by MIT license that can be
  12719. * found in the LICENSE.txt file.
  12720. */
  12721. .globl cuTexRefSetAddress2D_v2
  12722. .p2align 4
  12723. .type cuTexRefSetAddress2D_v2, %function
  12724. cuTexRefSetAddress2D_v2:
  12725. .cfi_startproc
  12726. // Intel opt. manual says to
  12727. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12728. // to hint static predictor.
  12729. cmpq $0, _libcuda_so_tramp_table+3608(%rip)
  12730. je 2f
  12731. 1:
  12732. jmp *_libcuda_so_tramp_table+3608(%rip)
  12733. 2:
  12734. pushq $451
  12735. .cfi_adjust_cfa_offset 8
  12736. call _libcuda_so_save_regs_and_resolve
  12737. addq $8, %rsp
  12738. .cfi_adjust_cfa_offset -8
  12739. jmp 1b
  12740. .cfi_endproc
  12741. /*
  12742. * Copyright 2018-2019 Yury Gribov
  12743. *
  12744. * The MIT License (MIT)
  12745. *
  12746. * Use of this source code is governed by MIT license that can be
  12747. * found in the LICENSE.txt file.
  12748. */
  12749. .globl cuTexRefSetAddress2D_v3
  12750. .p2align 4
  12751. .type cuTexRefSetAddress2D_v3, %function
  12752. cuTexRefSetAddress2D_v3:
  12753. .cfi_startproc
  12754. // Intel opt. manual says to
  12755. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12756. // to hint static predictor.
  12757. cmpq $0, _libcuda_so_tramp_table+3616(%rip)
  12758. je 2f
  12759. 1:
  12760. jmp *_libcuda_so_tramp_table+3616(%rip)
  12761. 2:
  12762. pushq $452
  12763. .cfi_adjust_cfa_offset 8
  12764. call _libcuda_so_save_regs_and_resolve
  12765. addq $8, %rsp
  12766. .cfi_adjust_cfa_offset -8
  12767. jmp 1b
  12768. .cfi_endproc
  12769. /*
  12770. * Copyright 2018-2019 Yury Gribov
  12771. *
  12772. * The MIT License (MIT)
  12773. *
  12774. * Use of this source code is governed by MIT license that can be
  12775. * found in the LICENSE.txt file.
  12776. */
  12777. .globl cuTexRefSetAddressMode
  12778. .p2align 4
  12779. .type cuTexRefSetAddressMode, %function
  12780. cuTexRefSetAddressMode:
  12781. .cfi_startproc
  12782. // Intel opt. manual says to
  12783. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12784. // to hint static predictor.
  12785. cmpq $0, _libcuda_so_tramp_table+3624(%rip)
  12786. je 2f
  12787. 1:
  12788. jmp *_libcuda_so_tramp_table+3624(%rip)
  12789. 2:
  12790. pushq $453
  12791. .cfi_adjust_cfa_offset 8
  12792. call _libcuda_so_save_regs_and_resolve
  12793. addq $8, %rsp
  12794. .cfi_adjust_cfa_offset -8
  12795. jmp 1b
  12796. .cfi_endproc
  12797. /*
  12798. * Copyright 2018-2019 Yury Gribov
  12799. *
  12800. * The MIT License (MIT)
  12801. *
  12802. * Use of this source code is governed by MIT license that can be
  12803. * found in the LICENSE.txt file.
  12804. */
  12805. .globl cuTexRefSetAddress_v2
  12806. .p2align 4
  12807. .type cuTexRefSetAddress_v2, %function
  12808. cuTexRefSetAddress_v2:
  12809. .cfi_startproc
  12810. // Intel opt. manual says to
  12811. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12812. // to hint static predictor.
  12813. cmpq $0, _libcuda_so_tramp_table+3632(%rip)
  12814. je 2f
  12815. 1:
  12816. jmp *_libcuda_so_tramp_table+3632(%rip)
  12817. 2:
  12818. pushq $454
  12819. .cfi_adjust_cfa_offset 8
  12820. call _libcuda_so_save_regs_and_resolve
  12821. addq $8, %rsp
  12822. .cfi_adjust_cfa_offset -8
  12823. jmp 1b
  12824. .cfi_endproc
  12825. /*
  12826. * Copyright 2018-2019 Yury Gribov
  12827. *
  12828. * The MIT License (MIT)
  12829. *
  12830. * Use of this source code is governed by MIT license that can be
  12831. * found in the LICENSE.txt file.
  12832. */
  12833. .globl cuTexRefSetArray
  12834. .p2align 4
  12835. .type cuTexRefSetArray, %function
  12836. cuTexRefSetArray:
  12837. .cfi_startproc
  12838. // Intel opt. manual says to
  12839. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12840. // to hint static predictor.
  12841. cmpq $0, _libcuda_so_tramp_table+3640(%rip)
  12842. je 2f
  12843. 1:
  12844. jmp *_libcuda_so_tramp_table+3640(%rip)
  12845. 2:
  12846. pushq $455
  12847. .cfi_adjust_cfa_offset 8
  12848. call _libcuda_so_save_regs_and_resolve
  12849. addq $8, %rsp
  12850. .cfi_adjust_cfa_offset -8
  12851. jmp 1b
  12852. .cfi_endproc
  12853. /*
  12854. * Copyright 2018-2019 Yury Gribov
  12855. *
  12856. * The MIT License (MIT)
  12857. *
  12858. * Use of this source code is governed by MIT license that can be
  12859. * found in the LICENSE.txt file.
  12860. */
  12861. .globl cuTexRefSetBorderColor
  12862. .p2align 4
  12863. .type cuTexRefSetBorderColor, %function
  12864. cuTexRefSetBorderColor:
  12865. .cfi_startproc
  12866. // Intel opt. manual says to
  12867. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12868. // to hint static predictor.
  12869. cmpq $0, _libcuda_so_tramp_table+3648(%rip)
  12870. je 2f
  12871. 1:
  12872. jmp *_libcuda_so_tramp_table+3648(%rip)
  12873. 2:
  12874. pushq $456
  12875. .cfi_adjust_cfa_offset 8
  12876. call _libcuda_so_save_regs_and_resolve
  12877. addq $8, %rsp
  12878. .cfi_adjust_cfa_offset -8
  12879. jmp 1b
  12880. .cfi_endproc
  12881. /*
  12882. * Copyright 2018-2019 Yury Gribov
  12883. *
  12884. * The MIT License (MIT)
  12885. *
  12886. * Use of this source code is governed by MIT license that can be
  12887. * found in the LICENSE.txt file.
  12888. */
  12889. .globl cuTexRefSetFilterMode
  12890. .p2align 4
  12891. .type cuTexRefSetFilterMode, %function
  12892. cuTexRefSetFilterMode:
  12893. .cfi_startproc
  12894. // Intel opt. manual says to
  12895. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12896. // to hint static predictor.
  12897. cmpq $0, _libcuda_so_tramp_table+3656(%rip)
  12898. je 2f
  12899. 1:
  12900. jmp *_libcuda_so_tramp_table+3656(%rip)
  12901. 2:
  12902. pushq $457
  12903. .cfi_adjust_cfa_offset 8
  12904. call _libcuda_so_save_regs_and_resolve
  12905. addq $8, %rsp
  12906. .cfi_adjust_cfa_offset -8
  12907. jmp 1b
  12908. .cfi_endproc
  12909. /*
  12910. * Copyright 2018-2019 Yury Gribov
  12911. *
  12912. * The MIT License (MIT)
  12913. *
  12914. * Use of this source code is governed by MIT license that can be
  12915. * found in the LICENSE.txt file.
  12916. */
  12917. .globl cuTexRefSetFlags
  12918. .p2align 4
  12919. .type cuTexRefSetFlags, %function
  12920. cuTexRefSetFlags:
  12921. .cfi_startproc
  12922. // Intel opt. manual says to
  12923. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12924. // to hint static predictor.
  12925. cmpq $0, _libcuda_so_tramp_table+3664(%rip)
  12926. je 2f
  12927. 1:
  12928. jmp *_libcuda_so_tramp_table+3664(%rip)
  12929. 2:
  12930. pushq $458
  12931. .cfi_adjust_cfa_offset 8
  12932. call _libcuda_so_save_regs_and_resolve
  12933. addq $8, %rsp
  12934. .cfi_adjust_cfa_offset -8
  12935. jmp 1b
  12936. .cfi_endproc
  12937. /*
  12938. * Copyright 2018-2019 Yury Gribov
  12939. *
  12940. * The MIT License (MIT)
  12941. *
  12942. * Use of this source code is governed by MIT license that can be
  12943. * found in the LICENSE.txt file.
  12944. */
  12945. .globl cuTexRefSetFormat
  12946. .p2align 4
  12947. .type cuTexRefSetFormat, %function
  12948. cuTexRefSetFormat:
  12949. .cfi_startproc
  12950. // Intel opt. manual says to
  12951. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12952. // to hint static predictor.
  12953. cmpq $0, _libcuda_so_tramp_table+3672(%rip)
  12954. je 2f
  12955. 1:
  12956. jmp *_libcuda_so_tramp_table+3672(%rip)
  12957. 2:
  12958. pushq $459
  12959. .cfi_adjust_cfa_offset 8
  12960. call _libcuda_so_save_regs_and_resolve
  12961. addq $8, %rsp
  12962. .cfi_adjust_cfa_offset -8
  12963. jmp 1b
  12964. .cfi_endproc
  12965. /*
  12966. * Copyright 2018-2019 Yury Gribov
  12967. *
  12968. * The MIT License (MIT)
  12969. *
  12970. * Use of this source code is governed by MIT license that can be
  12971. * found in the LICENSE.txt file.
  12972. */
  12973. .globl cuTexRefSetMaxAnisotropy
  12974. .p2align 4
  12975. .type cuTexRefSetMaxAnisotropy, %function
  12976. cuTexRefSetMaxAnisotropy:
  12977. .cfi_startproc
  12978. // Intel opt. manual says to
  12979. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  12980. // to hint static predictor.
  12981. cmpq $0, _libcuda_so_tramp_table+3680(%rip)
  12982. je 2f
  12983. 1:
  12984. jmp *_libcuda_so_tramp_table+3680(%rip)
  12985. 2:
  12986. pushq $460
  12987. .cfi_adjust_cfa_offset 8
  12988. call _libcuda_so_save_regs_and_resolve
  12989. addq $8, %rsp
  12990. .cfi_adjust_cfa_offset -8
  12991. jmp 1b
  12992. .cfi_endproc
  12993. /*
  12994. * Copyright 2018-2019 Yury Gribov
  12995. *
  12996. * The MIT License (MIT)
  12997. *
  12998. * Use of this source code is governed by MIT license that can be
  12999. * found in the LICENSE.txt file.
  13000. */
  13001. .globl cuTexRefSetMipmapFilterMode
  13002. .p2align 4
  13003. .type cuTexRefSetMipmapFilterMode, %function
  13004. cuTexRefSetMipmapFilterMode:
  13005. .cfi_startproc
  13006. // Intel opt. manual says to
  13007. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13008. // to hint static predictor.
  13009. cmpq $0, _libcuda_so_tramp_table+3688(%rip)
  13010. je 2f
  13011. 1:
  13012. jmp *_libcuda_so_tramp_table+3688(%rip)
  13013. 2:
  13014. pushq $461
  13015. .cfi_adjust_cfa_offset 8
  13016. call _libcuda_so_save_regs_and_resolve
  13017. addq $8, %rsp
  13018. .cfi_adjust_cfa_offset -8
  13019. jmp 1b
  13020. .cfi_endproc
  13021. /*
  13022. * Copyright 2018-2019 Yury Gribov
  13023. *
  13024. * The MIT License (MIT)
  13025. *
  13026. * Use of this source code is governed by MIT license that can be
  13027. * found in the LICENSE.txt file.
  13028. */
  13029. .globl cuTexRefSetMipmapLevelBias
  13030. .p2align 4
  13031. .type cuTexRefSetMipmapLevelBias, %function
  13032. cuTexRefSetMipmapLevelBias:
  13033. .cfi_startproc
  13034. // Intel opt. manual says to
  13035. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13036. // to hint static predictor.
  13037. cmpq $0, _libcuda_so_tramp_table+3696(%rip)
  13038. je 2f
  13039. 1:
  13040. jmp *_libcuda_so_tramp_table+3696(%rip)
  13041. 2:
  13042. pushq $462
  13043. .cfi_adjust_cfa_offset 8
  13044. call _libcuda_so_save_regs_and_resolve
  13045. addq $8, %rsp
  13046. .cfi_adjust_cfa_offset -8
  13047. jmp 1b
  13048. .cfi_endproc
  13049. /*
  13050. * Copyright 2018-2019 Yury Gribov
  13051. *
  13052. * The MIT License (MIT)
  13053. *
  13054. * Use of this source code is governed by MIT license that can be
  13055. * found in the LICENSE.txt file.
  13056. */
  13057. .globl cuTexRefSetMipmapLevelClamp
  13058. .p2align 4
  13059. .type cuTexRefSetMipmapLevelClamp, %function
  13060. cuTexRefSetMipmapLevelClamp:
  13061. .cfi_startproc
  13062. // Intel opt. manual says to
  13063. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13064. // to hint static predictor.
  13065. cmpq $0, _libcuda_so_tramp_table+3704(%rip)
  13066. je 2f
  13067. 1:
  13068. jmp *_libcuda_so_tramp_table+3704(%rip)
  13069. 2:
  13070. pushq $463
  13071. .cfi_adjust_cfa_offset 8
  13072. call _libcuda_so_save_regs_and_resolve
  13073. addq $8, %rsp
  13074. .cfi_adjust_cfa_offset -8
  13075. jmp 1b
  13076. .cfi_endproc
  13077. /*
  13078. * Copyright 2018-2019 Yury Gribov
  13079. *
  13080. * The MIT License (MIT)
  13081. *
  13082. * Use of this source code is governed by MIT license that can be
  13083. * found in the LICENSE.txt file.
  13084. */
  13085. .globl cuTexRefSetMipmappedArray
  13086. .p2align 4
  13087. .type cuTexRefSetMipmappedArray, %function
  13088. cuTexRefSetMipmappedArray:
  13089. .cfi_startproc
  13090. // Intel opt. manual says to
  13091. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13092. // to hint static predictor.
  13093. cmpq $0, _libcuda_so_tramp_table+3712(%rip)
  13094. je 2f
  13095. 1:
  13096. jmp *_libcuda_so_tramp_table+3712(%rip)
  13097. 2:
  13098. pushq $464
  13099. .cfi_adjust_cfa_offset 8
  13100. call _libcuda_so_save_regs_and_resolve
  13101. addq $8, %rsp
  13102. .cfi_adjust_cfa_offset -8
  13103. jmp 1b
  13104. .cfi_endproc
  13105. /*
  13106. * Copyright 2018-2019 Yury Gribov
  13107. *
  13108. * The MIT License (MIT)
  13109. *
  13110. * Use of this source code is governed by MIT license that can be
  13111. * found in the LICENSE.txt file.
  13112. */
  13113. .globl cuThreadExchangeStreamCaptureMode
  13114. .p2align 4
  13115. .type cuThreadExchangeStreamCaptureMode, %function
  13116. cuThreadExchangeStreamCaptureMode:
  13117. .cfi_startproc
  13118. // Intel opt. manual says to
  13119. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13120. // to hint static predictor.
  13121. cmpq $0, _libcuda_so_tramp_table+3720(%rip)
  13122. je 2f
  13123. 1:
  13124. jmp *_libcuda_so_tramp_table+3720(%rip)
  13125. 2:
  13126. pushq $465
  13127. .cfi_adjust_cfa_offset 8
  13128. call _libcuda_so_save_regs_and_resolve
  13129. addq $8, %rsp
  13130. .cfi_adjust_cfa_offset -8
  13131. jmp 1b
  13132. .cfi_endproc
  13133. /*
  13134. * Copyright 2018-2019 Yury Gribov
  13135. *
  13136. * The MIT License (MIT)
  13137. *
  13138. * Use of this source code is governed by MIT license that can be
  13139. * found in the LICENSE.txt file.
  13140. */
  13141. .globl cuVDPAUCtxCreate
  13142. .p2align 4
  13143. .type cuVDPAUCtxCreate, %function
  13144. cuVDPAUCtxCreate:
  13145. .cfi_startproc
  13146. // Intel opt. manual says to
  13147. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13148. // to hint static predictor.
  13149. cmpq $0, _libcuda_so_tramp_table+3728(%rip)
  13150. je 2f
  13151. 1:
  13152. jmp *_libcuda_so_tramp_table+3728(%rip)
  13153. 2:
  13154. pushq $466
  13155. .cfi_adjust_cfa_offset 8
  13156. call _libcuda_so_save_regs_and_resolve
  13157. addq $8, %rsp
  13158. .cfi_adjust_cfa_offset -8
  13159. jmp 1b
  13160. .cfi_endproc
  13161. /*
  13162. * Copyright 2018-2019 Yury Gribov
  13163. *
  13164. * The MIT License (MIT)
  13165. *
  13166. * Use of this source code is governed by MIT license that can be
  13167. * found in the LICENSE.txt file.
  13168. */
  13169. .globl cuVDPAUCtxCreate_v2
  13170. .p2align 4
  13171. .type cuVDPAUCtxCreate_v2, %function
  13172. cuVDPAUCtxCreate_v2:
  13173. .cfi_startproc
  13174. // Intel opt. manual says to
  13175. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13176. // to hint static predictor.
  13177. cmpq $0, _libcuda_so_tramp_table+3736(%rip)
  13178. je 2f
  13179. 1:
  13180. jmp *_libcuda_so_tramp_table+3736(%rip)
  13181. 2:
  13182. pushq $467
  13183. .cfi_adjust_cfa_offset 8
  13184. call _libcuda_so_save_regs_and_resolve
  13185. addq $8, %rsp
  13186. .cfi_adjust_cfa_offset -8
  13187. jmp 1b
  13188. .cfi_endproc
  13189. /*
  13190. * Copyright 2018-2019 Yury Gribov
  13191. *
  13192. * The MIT License (MIT)
  13193. *
  13194. * Use of this source code is governed by MIT license that can be
  13195. * found in the LICENSE.txt file.
  13196. */
  13197. .globl cuVDPAUGetDevice
  13198. .p2align 4
  13199. .type cuVDPAUGetDevice, %function
  13200. cuVDPAUGetDevice:
  13201. .cfi_startproc
  13202. // Intel opt. manual says to
  13203. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13204. // to hint static predictor.
  13205. cmpq $0, _libcuda_so_tramp_table+3744(%rip)
  13206. je 2f
  13207. 1:
  13208. jmp *_libcuda_so_tramp_table+3744(%rip)
  13209. 2:
  13210. pushq $468
  13211. .cfi_adjust_cfa_offset 8
  13212. call _libcuda_so_save_regs_and_resolve
  13213. addq $8, %rsp
  13214. .cfi_adjust_cfa_offset -8
  13215. jmp 1b
  13216. .cfi_endproc
  13217. /*
  13218. * Copyright 2018-2019 Yury Gribov
  13219. *
  13220. * The MIT License (MIT)
  13221. *
  13222. * Use of this source code is governed by MIT license that can be
  13223. * found in the LICENSE.txt file.
  13224. */
  13225. .globl cuWaitExternalSemaphoresAsync
  13226. .p2align 4
  13227. .type cuWaitExternalSemaphoresAsync, %function
  13228. cuWaitExternalSemaphoresAsync:
  13229. .cfi_startproc
  13230. // Intel opt. manual says to
  13231. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13232. // to hint static predictor.
  13233. cmpq $0, _libcuda_so_tramp_table+3752(%rip)
  13234. je 2f
  13235. 1:
  13236. jmp *_libcuda_so_tramp_table+3752(%rip)
  13237. 2:
  13238. pushq $469
  13239. .cfi_adjust_cfa_offset 8
  13240. call _libcuda_so_save_regs_and_resolve
  13241. addq $8, %rsp
  13242. .cfi_adjust_cfa_offset -8
  13243. jmp 1b
  13244. .cfi_endproc
  13245. /*
  13246. * Copyright 2018-2019 Yury Gribov
  13247. *
  13248. * The MIT License (MIT)
  13249. *
  13250. * Use of this source code is governed by MIT license that can be
  13251. * found in the LICENSE.txt file.
  13252. */
  13253. .globl cuWaitExternalSemaphoresAsync_ptsz
  13254. .p2align 4
  13255. .type cuWaitExternalSemaphoresAsync_ptsz, %function
  13256. cuWaitExternalSemaphoresAsync_ptsz:
  13257. .cfi_startproc
  13258. // Intel opt. manual says to
  13259. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13260. // to hint static predictor.
  13261. cmpq $0, _libcuda_so_tramp_table+3760(%rip)
  13262. je 2f
  13263. 1:
  13264. jmp *_libcuda_so_tramp_table+3760(%rip)
  13265. 2:
  13266. pushq $470
  13267. .cfi_adjust_cfa_offset 8
  13268. call _libcuda_so_save_regs_and_resolve
  13269. addq $8, %rsp
  13270. .cfi_adjust_cfa_offset -8
  13271. jmp 1b
  13272. .cfi_endproc
  13273. /*
  13274. * Copyright 2018-2019 Yury Gribov
  13275. *
  13276. * The MIT License (MIT)
  13277. *
  13278. * Use of this source code is governed by MIT license that can be
  13279. * found in the LICENSE.txt file.
  13280. */
  13281. .globl cudbgApiAttach
  13282. .p2align 4
  13283. .type cudbgApiAttach, %function
  13284. cudbgApiAttach:
  13285. .cfi_startproc
  13286. // Intel opt. manual says to
  13287. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13288. // to hint static predictor.
  13289. cmpq $0, _libcuda_so_tramp_table+3768(%rip)
  13290. je 2f
  13291. 1:
  13292. jmp *_libcuda_so_tramp_table+3768(%rip)
  13293. 2:
  13294. pushq $471
  13295. .cfi_adjust_cfa_offset 8
  13296. call _libcuda_so_save_regs_and_resolve
  13297. addq $8, %rsp
  13298. .cfi_adjust_cfa_offset -8
  13299. jmp 1b
  13300. .cfi_endproc
  13301. /*
  13302. * Copyright 2018-2019 Yury Gribov
  13303. *
  13304. * The MIT License (MIT)
  13305. *
  13306. * Use of this source code is governed by MIT license that can be
  13307. * found in the LICENSE.txt file.
  13308. */
  13309. .globl cudbgApiDetach
  13310. .p2align 4
  13311. .type cudbgApiDetach, %function
  13312. cudbgApiDetach:
  13313. .cfi_startproc
  13314. // Intel opt. manual says to
  13315. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13316. // to hint static predictor.
  13317. cmpq $0, _libcuda_so_tramp_table+3776(%rip)
  13318. je 2f
  13319. 1:
  13320. jmp *_libcuda_so_tramp_table+3776(%rip)
  13321. 2:
  13322. pushq $472
  13323. .cfi_adjust_cfa_offset 8
  13324. call _libcuda_so_save_regs_and_resolve
  13325. addq $8, %rsp
  13326. .cfi_adjust_cfa_offset -8
  13327. jmp 1b
  13328. .cfi_endproc
  13329. /*
  13330. * Copyright 2018-2019 Yury Gribov
  13331. *
  13332. * The MIT License (MIT)
  13333. *
  13334. * Use of this source code is governed by MIT license that can be
  13335. * found in the LICENSE.txt file.
  13336. */
  13337. .globl cudbgApiInit
  13338. .p2align 4
  13339. .type cudbgApiInit, %function
  13340. cudbgApiInit:
  13341. .cfi_startproc
  13342. // Intel opt. manual says to
  13343. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13344. // to hint static predictor.
  13345. cmpq $0, _libcuda_so_tramp_table+3784(%rip)
  13346. je 2f
  13347. 1:
  13348. jmp *_libcuda_so_tramp_table+3784(%rip)
  13349. 2:
  13350. pushq $473
  13351. .cfi_adjust_cfa_offset 8
  13352. call _libcuda_so_save_regs_and_resolve
  13353. addq $8, %rsp
  13354. .cfi_adjust_cfa_offset -8
  13355. jmp 1b
  13356. .cfi_endproc
  13357. /*
  13358. * Copyright 2018-2019 Yury Gribov
  13359. *
  13360. * The MIT License (MIT)
  13361. *
  13362. * Use of this source code is governed by MIT license that can be
  13363. * found in the LICENSE.txt file.
  13364. */
  13365. .globl cudbgGetAPI
  13366. .p2align 4
  13367. .type cudbgGetAPI, %function
  13368. cudbgGetAPI:
  13369. .cfi_startproc
  13370. // Intel opt. manual says to
  13371. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13372. // to hint static predictor.
  13373. cmpq $0, _libcuda_so_tramp_table+3792(%rip)
  13374. je 2f
  13375. 1:
  13376. jmp *_libcuda_so_tramp_table+3792(%rip)
  13377. 2:
  13378. pushq $474
  13379. .cfi_adjust_cfa_offset 8
  13380. call _libcuda_so_save_regs_and_resolve
  13381. addq $8, %rsp
  13382. .cfi_adjust_cfa_offset -8
  13383. jmp 1b
  13384. .cfi_endproc
  13385. /*
  13386. * Copyright 2018-2019 Yury Gribov
  13387. *
  13388. * The MIT License (MIT)
  13389. *
  13390. * Use of this source code is governed by MIT license that can be
  13391. * found in the LICENSE.txt file.
  13392. */
  13393. .globl cudbgGetAPIVersion
  13394. .p2align 4
  13395. .type cudbgGetAPIVersion, %function
  13396. cudbgGetAPIVersion:
  13397. .cfi_startproc
  13398. // Intel opt. manual says to
  13399. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13400. // to hint static predictor.
  13401. cmpq $0, _libcuda_so_tramp_table+3800(%rip)
  13402. je 2f
  13403. 1:
  13404. jmp *_libcuda_so_tramp_table+3800(%rip)
  13405. 2:
  13406. pushq $475
  13407. .cfi_adjust_cfa_offset 8
  13408. call _libcuda_so_save_regs_and_resolve
  13409. addq $8, %rsp
  13410. .cfi_adjust_cfa_offset -8
  13411. jmp 1b
  13412. .cfi_endproc
  13413. /*
  13414. * Copyright 2018-2019 Yury Gribov
  13415. *
  13416. * The MIT License (MIT)
  13417. *
  13418. * Use of this source code is governed by MIT license that can be
  13419. * found in the LICENSE.txt file.
  13420. */
  13421. .globl cudbgMain
  13422. .p2align 4
  13423. .type cudbgMain, %function
  13424. cudbgMain:
  13425. .cfi_startproc
  13426. // Intel opt. manual says to
  13427. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13428. // to hint static predictor.
  13429. cmpq $0, _libcuda_so_tramp_table+3808(%rip)
  13430. je 2f
  13431. 1:
  13432. jmp *_libcuda_so_tramp_table+3808(%rip)
  13433. 2:
  13434. pushq $476
  13435. .cfi_adjust_cfa_offset 8
  13436. call _libcuda_so_save_regs_and_resolve
  13437. addq $8, %rsp
  13438. .cfi_adjust_cfa_offset -8
  13439. jmp 1b
  13440. .cfi_endproc
  13441. /*
  13442. * Copyright 2018-2019 Yury Gribov
  13443. *
  13444. * The MIT License (MIT)
  13445. *
  13446. * Use of this source code is governed by MIT license that can be
  13447. * found in the LICENSE.txt file.
  13448. */
  13449. .globl cudbgReportDriverApiError
  13450. .p2align 4
  13451. .type cudbgReportDriverApiError, %function
  13452. cudbgReportDriverApiError:
  13453. .cfi_startproc
  13454. // Intel opt. manual says to
  13455. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13456. // to hint static predictor.
  13457. cmpq $0, _libcuda_so_tramp_table+3816(%rip)
  13458. je 2f
  13459. 1:
  13460. jmp *_libcuda_so_tramp_table+3816(%rip)
  13461. 2:
  13462. pushq $477
  13463. .cfi_adjust_cfa_offset 8
  13464. call _libcuda_so_save_regs_and_resolve
  13465. addq $8, %rsp
  13466. .cfi_adjust_cfa_offset -8
  13467. jmp 1b
  13468. .cfi_endproc
  13469. /*
  13470. * Copyright 2018-2019 Yury Gribov
  13471. *
  13472. * The MIT License (MIT)
  13473. *
  13474. * Use of this source code is governed by MIT license that can be
  13475. * found in the LICENSE.txt file.
  13476. */
  13477. .globl cudbgReportDriverInternalError
  13478. .p2align 4
  13479. .type cudbgReportDriverInternalError, %function
  13480. cudbgReportDriverInternalError:
  13481. .cfi_startproc
  13482. // Intel opt. manual says to
  13483. // "make the fall-through code following a conditional branch be the likely target for a branch with a forward target"
  13484. // to hint static predictor.
  13485. cmpq $0, _libcuda_so_tramp_table+3824(%rip)
  13486. je 2f
  13487. 1:
  13488. jmp *_libcuda_so_tramp_table+3824(%rip)
  13489. 2:
  13490. pushq $478
  13491. .cfi_adjust_cfa_offset 8
  13492. call _libcuda_so_save_regs_and_resolve
  13493. addq $8, %rsp
  13494. .cfi_adjust_cfa_offset -8
  13495. jmp 1b
  13496. .cfi_endproc