You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

trmmkernel_16x8.c 59 kB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248224922502251225222532254225522562257225822592260226122622263226422652266226722682269227022712272227322742275227622772278227922802281228222832284228522862287228822892290229122922293229422952296229722982299230023012302230323042305230623072308230923102311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380238123822383238423852386238723882389239023912392239323942395239623972398239924002401240224032404240524062407240824092410241124122413241424152416241724182419242024212422242324242425242624272428242924302431243224332434243524362437243824392440244124422443244424452446244724482449245024512452245324542455245624572458245924602461246224632464246524662467246824692470247124722473247424752476247724782479248024812482248324842485248624872488248924902491249224932494249524962497249824992500250125022503250425052506250725082509251025112512251325142515251625172518251925202521252225232524252525262527252825292530253125322533253425352536253725382539254025412542254325442545254625472548254925502551255225532554255525562557255825592560256125622563256425652566256725682569257025712572257325742575257625772578257925802581258225832584258525862587258825892590259125922593259425952596259725982599260026012602260326042605260626072608260926102611261226132614261526162617261826192620262126222623262426252626262726282629263026312632263326342635263626372638263926402641264226432644264526462647264826492650265126522653265426552656265726582659266026612662266326642665266626672668266926702671267226732674267526762677267826792680268126822683268426852686268726882689269026912692269326942695269626972698269927002701270227032704270527062707270827092710271127122713271427152716271727182719272027212722272327242725272627272728272927302731273227332734273527362737273827392740274127422743274427452746274727482749275027512752275327542755275627572758275927602761276227632764276527662767276827692770277127722773277427752776277727782779278027812782278327842785278627872788278927902791279227932794279527962797279827992800280128022803280428052806280728082809281028112812281328142815281628172818281928202821282228232824282528262827282828292830283128322833283428352836283728382839284028412842284328442845284628472848284928502851285228532854285528562857285828592860286128622863286428652866286728682869287028712872287328742875287628772878287928802881288228832884288528862887288828892890289128922893289428952896289728982899290029012902290329042905290629072908290929102911291229132914291529162917291829192920292129222923292429252926292729282929293029312932293329342935293629372938293929402941294229432944294529462947294829492950295129522953295429552956295729582959296029612962296329642965296629672968296929702971297229732974297529762977297829792980298129822983298429852986298729882989299029912992299329942995299629972998299930003001300230033004300530063007300830093010301130123013301430153016301730183019302030213022302330243025302630273028302930303031303230333034303530363037303830393040304130423043304430453046304730483049305030513052305330543055305630573058305930603061306230633064306530663067306830693070307130723073307430753076307730783079308030813082308330843085308630873088308930903091309230933094309530963097309830993100310131023103310431053106310731083109311031113112311331143115311631173118311931203121312231233124312531263127312831293130313131323133313431353136313731383139314031413142314331443145314631473148314931503151315231533154315531563157315831593160316131623163316431653166316731683169317031713172317331743175317631773178317931803181318231833184318531863187318831893190319131923193319431953196319731983199320032013202320332043205320632073208320932103211321232133214321532163217321832193220322132223223322432253226322732283229323032313232323332343235323632373238323932403241324232433244324532463247324832493250325132523253325432553256325732583259326032613262326332643265326632673268326932703271327232733274327532763277327832793280328132823283328432853286328732883289329032913292329332943295329632973298329933003301330233033304330533063307330833093310331133123313331433153316331733183319332033213322332333243325332633273328332933303331333233333334333533363337333833393340334133423343334433453346334733483349335033513352335333543355335633573358335933603361336233633364336533663367336833693370337133723373337433753376337733783379338033813382338333843385338633873388338933903391339233933394339533963397339833993400340134023403340434053406340734083409341034113412341334143415341634173418341934203421342234233424342534263427342834293430343134323433343434353436343734383439344034413442344334443445344634473448344934503451345234533454345534563457345834593460346134623463346434653466346734683469347034713472347334743475347634773478347934803481348234833484348534863487348834893490349134923493349434953496349734983499350035013502350335043505350635073508350935103511351235133514351535163517351835193520352135223523352435253526352735283529353035313532353335343535353635373538353935403541354235433544354535463547354835493550355135523553355435553556355735583559356035613562356335643565356635673568356935703571357235733574357535763577357835793580358135823583358435853586358735883589359035913592359335943595359635973598359936003601360236033604360536063607360836093610361136123613361436153616361736183619362036213622362336243625362636273628362936303631363236333634363536363637363836393640364136423643364436453646364736483649365036513652365336543655365636573658365936603661366236633664366536663667366836693670367136723673367436753676
  1. #include "common.h"
  2. int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FLOAT* C,BLASLONG ldc ,BLASLONG offset)
  3. {
  4. BLASLONG i,j,k;
  5. FLOAT *C0,*C1,*C2,*C3,*C4,*C5,*C6,*C7,*ptrba,*ptrbb;
  6. FLOAT res0_0;
  7. FLOAT res0_1;
  8. FLOAT res0_2;
  9. FLOAT res0_3;
  10. FLOAT res0_4;
  11. FLOAT res0_5;
  12. FLOAT res0_6;
  13. FLOAT res0_7;
  14. FLOAT res0_8;
  15. FLOAT res0_9;
  16. FLOAT res0_10;
  17. FLOAT res0_11;
  18. FLOAT res0_12;
  19. FLOAT res0_13;
  20. FLOAT res0_14;
  21. FLOAT res0_15;
  22. FLOAT res1_0;
  23. FLOAT res1_1;
  24. FLOAT res1_2;
  25. FLOAT res1_3;
  26. FLOAT res1_4;
  27. FLOAT res1_5;
  28. FLOAT res1_6;
  29. FLOAT res1_7;
  30. FLOAT res1_8;
  31. FLOAT res1_9;
  32. FLOAT res1_10;
  33. FLOAT res1_11;
  34. FLOAT res1_12;
  35. FLOAT res1_13;
  36. FLOAT res1_14;
  37. FLOAT res1_15;
  38. FLOAT res2_0;
  39. FLOAT res2_1;
  40. FLOAT res2_2;
  41. FLOAT res2_3;
  42. FLOAT res2_4;
  43. FLOAT res2_5;
  44. FLOAT res2_6;
  45. FLOAT res2_7;
  46. FLOAT res2_8;
  47. FLOAT res2_9;
  48. FLOAT res2_10;
  49. FLOAT res2_11;
  50. FLOAT res2_12;
  51. FLOAT res2_13;
  52. FLOAT res2_14;
  53. FLOAT res2_15;
  54. FLOAT res3_0;
  55. FLOAT res3_1;
  56. FLOAT res3_2;
  57. FLOAT res3_3;
  58. FLOAT res3_4;
  59. FLOAT res3_5;
  60. FLOAT res3_6;
  61. FLOAT res3_7;
  62. FLOAT res3_8;
  63. FLOAT res3_9;
  64. FLOAT res3_10;
  65. FLOAT res3_11;
  66. FLOAT res3_12;
  67. FLOAT res3_13;
  68. FLOAT res3_14;
  69. FLOAT res3_15;
  70. FLOAT res4_0;
  71. FLOAT res4_1;
  72. FLOAT res4_2;
  73. FLOAT res4_3;
  74. FLOAT res4_4;
  75. FLOAT res4_5;
  76. FLOAT res4_6;
  77. FLOAT res4_7;
  78. FLOAT res4_8;
  79. FLOAT res4_9;
  80. FLOAT res4_10;
  81. FLOAT res4_11;
  82. FLOAT res4_12;
  83. FLOAT res4_13;
  84. FLOAT res4_14;
  85. FLOAT res4_15;
  86. FLOAT res5_0;
  87. FLOAT res5_1;
  88. FLOAT res5_2;
  89. FLOAT res5_3;
  90. FLOAT res5_4;
  91. FLOAT res5_5;
  92. FLOAT res5_6;
  93. FLOAT res5_7;
  94. FLOAT res5_8;
  95. FLOAT res5_9;
  96. FLOAT res5_10;
  97. FLOAT res5_11;
  98. FLOAT res5_12;
  99. FLOAT res5_13;
  100. FLOAT res5_14;
  101. FLOAT res5_15;
  102. FLOAT res6_0;
  103. FLOAT res6_1;
  104. FLOAT res6_2;
  105. FLOAT res6_3;
  106. FLOAT res6_4;
  107. FLOAT res6_5;
  108. FLOAT res6_6;
  109. FLOAT res6_7;
  110. FLOAT res6_8;
  111. FLOAT res6_9;
  112. FLOAT res6_10;
  113. FLOAT res6_11;
  114. FLOAT res6_12;
  115. FLOAT res6_13;
  116. FLOAT res6_14;
  117. FLOAT res6_15;
  118. FLOAT res7_0;
  119. FLOAT res7_1;
  120. FLOAT res7_2;
  121. FLOAT res7_3;
  122. FLOAT res7_4;
  123. FLOAT res7_5;
  124. FLOAT res7_6;
  125. FLOAT res7_7;
  126. FLOAT res7_8;
  127. FLOAT res7_9;
  128. FLOAT res7_10;
  129. FLOAT res7_11;
  130. FLOAT res7_12;
  131. FLOAT res7_13;
  132. FLOAT res7_14;
  133. FLOAT res7_15;
  134. FLOAT a0;
  135. FLOAT a1;
  136. FLOAT b0;
  137. FLOAT b1;
  138. FLOAT b2;
  139. FLOAT b3;
  140. FLOAT b4;
  141. FLOAT b5;
  142. FLOAT b6;
  143. FLOAT b7;
  144. BLASLONG off, temp;
  145. #if !defined(LEFT)
  146. off = -offset;
  147. #else
  148. off = 0;
  149. #endif
  150. for (j=0; j<bn/8; j+=1)
  151. {
  152. C0 = C;
  153. C1 = C0+ldc;
  154. C2 = C0+2*ldc;
  155. C3 = C0+3*ldc;
  156. C4 = C0+4*ldc;
  157. C5 = C0+5*ldc;
  158. C6 = C0+6*ldc;
  159. C7 = C0+7*ldc;
  160. #if defined(TRMMKERNEL) && defined(LEFT)
  161. off = offset;
  162. #endif
  163. ptrba = ba;
  164. for (i=0; i<bm/16; i+=1)
  165. {
  166. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  167. ptrbb = bb;
  168. #else
  169. ptrba += off*16;
  170. ptrbb = bb + off*8;
  171. #endif
  172. res0_0 = 0;
  173. res0_1 = 0;
  174. res0_2 = 0;
  175. res0_3 = 0;
  176. res0_4 = 0;
  177. res0_5 = 0;
  178. res0_6 = 0;
  179. res0_7 = 0;
  180. res0_8 = 0;
  181. res0_9 = 0;
  182. res0_10 = 0;
  183. res0_11 = 0;
  184. res0_12 = 0;
  185. res0_13 = 0;
  186. res0_14 = 0;
  187. res0_15 = 0;
  188. res1_0 = 0;
  189. res1_1 = 0;
  190. res1_2 = 0;
  191. res1_3 = 0;
  192. res1_4 = 0;
  193. res1_5 = 0;
  194. res1_6 = 0;
  195. res1_7 = 0;
  196. res1_8 = 0;
  197. res1_9 = 0;
  198. res1_10 = 0;
  199. res1_11 = 0;
  200. res1_12 = 0;
  201. res1_13 = 0;
  202. res1_14 = 0;
  203. res1_15 = 0;
  204. res2_0 = 0;
  205. res2_1 = 0;
  206. res2_2 = 0;
  207. res2_3 = 0;
  208. res2_4 = 0;
  209. res2_5 = 0;
  210. res2_6 = 0;
  211. res2_7 = 0;
  212. res2_8 = 0;
  213. res2_9 = 0;
  214. res2_10 = 0;
  215. res2_11 = 0;
  216. res2_12 = 0;
  217. res2_13 = 0;
  218. res2_14 = 0;
  219. res2_15 = 0;
  220. res3_0 = 0;
  221. res3_1 = 0;
  222. res3_2 = 0;
  223. res3_3 = 0;
  224. res3_4 = 0;
  225. res3_5 = 0;
  226. res3_6 = 0;
  227. res3_7 = 0;
  228. res3_8 = 0;
  229. res3_9 = 0;
  230. res3_10 = 0;
  231. res3_11 = 0;
  232. res3_12 = 0;
  233. res3_13 = 0;
  234. res3_14 = 0;
  235. res3_15 = 0;
  236. res4_0 = 0;
  237. res4_1 = 0;
  238. res4_2 = 0;
  239. res4_3 = 0;
  240. res4_4 = 0;
  241. res4_5 = 0;
  242. res4_6 = 0;
  243. res4_7 = 0;
  244. res4_8 = 0;
  245. res4_9 = 0;
  246. res4_10 = 0;
  247. res4_11 = 0;
  248. res4_12 = 0;
  249. res4_13 = 0;
  250. res4_14 = 0;
  251. res4_15 = 0;
  252. res5_0 = 0;
  253. res5_1 = 0;
  254. res5_2 = 0;
  255. res5_3 = 0;
  256. res5_4 = 0;
  257. res5_5 = 0;
  258. res5_6 = 0;
  259. res5_7 = 0;
  260. res5_8 = 0;
  261. res5_9 = 0;
  262. res5_10 = 0;
  263. res5_11 = 0;
  264. res5_12 = 0;
  265. res5_13 = 0;
  266. res5_14 = 0;
  267. res5_15 = 0;
  268. res6_0 = 0;
  269. res6_1 = 0;
  270. res6_2 = 0;
  271. res6_3 = 0;
  272. res6_4 = 0;
  273. res6_5 = 0;
  274. res6_6 = 0;
  275. res6_7 = 0;
  276. res6_8 = 0;
  277. res6_9 = 0;
  278. res6_10 = 0;
  279. res6_11 = 0;
  280. res6_12 = 0;
  281. res6_13 = 0;
  282. res6_14 = 0;
  283. res6_15 = 0;
  284. res7_0 = 0;
  285. res7_1 = 0;
  286. res7_2 = 0;
  287. res7_3 = 0;
  288. res7_4 = 0;
  289. res7_5 = 0;
  290. res7_6 = 0;
  291. res7_7 = 0;
  292. res7_8 = 0;
  293. res7_9 = 0;
  294. res7_10 = 0;
  295. res7_11 = 0;
  296. res7_12 = 0;
  297. res7_13 = 0;
  298. res7_14 = 0;
  299. res7_15 = 0;
  300. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  301. temp = bk-off;
  302. #elif defined(LEFT)
  303. temp = off+16; // number of values in A
  304. #else
  305. temp = off+8; // number of values in B
  306. #endif
  307. for (k=0; k<temp; k++)
  308. {
  309. b0 = ptrbb[0];
  310. b1 = ptrbb[1];
  311. b2 = ptrbb[2];
  312. b3 = ptrbb[3];
  313. b4 = ptrbb[4];
  314. b5 = ptrbb[5];
  315. b6 = ptrbb[6];
  316. b7 = ptrbb[7];
  317. a0 = ptrba[0];
  318. res0_0 += a0*b0;
  319. res1_0 += a0*b1;
  320. res2_0 += a0*b2;
  321. res3_0 += a0*b3;
  322. res4_0 += a0*b4;
  323. res5_0 += a0*b5;
  324. res6_0 += a0*b6;
  325. res7_0 += a0*b7;
  326. a1 = ptrba[1];
  327. res0_1 += a1*b0;
  328. res1_1 += a1*b1;
  329. res2_1 += a1*b2;
  330. res3_1 += a1*b3;
  331. res4_1 += a1*b4;
  332. res5_1 += a1*b5;
  333. res6_1 += a1*b6;
  334. res7_1 += a1*b7;
  335. a0 = ptrba[2];
  336. res0_2 += a0*b0;
  337. res1_2 += a0*b1;
  338. res2_2 += a0*b2;
  339. res3_2 += a0*b3;
  340. res4_2 += a0*b4;
  341. res5_2 += a0*b5;
  342. res6_2 += a0*b6;
  343. res7_2 += a0*b7;
  344. a1 = ptrba[3];
  345. res0_3 += a1*b0;
  346. res1_3 += a1*b1;
  347. res2_3 += a1*b2;
  348. res3_3 += a1*b3;
  349. res4_3 += a1*b4;
  350. res5_3 += a1*b5;
  351. res6_3 += a1*b6;
  352. res7_3 += a1*b7;
  353. a0 = ptrba[4];
  354. res0_4 += a0*b0;
  355. res1_4 += a0*b1;
  356. res2_4 += a0*b2;
  357. res3_4 += a0*b3;
  358. res4_4 += a0*b4;
  359. res5_4 += a0*b5;
  360. res6_4 += a0*b6;
  361. res7_4 += a0*b7;
  362. a1 = ptrba[5];
  363. res0_5 += a1*b0;
  364. res1_5 += a1*b1;
  365. res2_5 += a1*b2;
  366. res3_5 += a1*b3;
  367. res4_5 += a1*b4;
  368. res5_5 += a1*b5;
  369. res6_5 += a1*b6;
  370. res7_5 += a1*b7;
  371. a0 = ptrba[6];
  372. res0_6 += a0*b0;
  373. res1_6 += a0*b1;
  374. res2_6 += a0*b2;
  375. res3_6 += a0*b3;
  376. res4_6 += a0*b4;
  377. res5_6 += a0*b5;
  378. res6_6 += a0*b6;
  379. res7_6 += a0*b7;
  380. a1 = ptrba[7];
  381. res0_7 += a1*b0;
  382. res1_7 += a1*b1;
  383. res2_7 += a1*b2;
  384. res3_7 += a1*b3;
  385. res4_7 += a1*b4;
  386. res5_7 += a1*b5;
  387. res6_7 += a1*b6;
  388. res7_7 += a1*b7;
  389. a0 = ptrba[8];
  390. res0_8 += a0*b0;
  391. res1_8 += a0*b1;
  392. res2_8 += a0*b2;
  393. res3_8 += a0*b3;
  394. res4_8 += a0*b4;
  395. res5_8 += a0*b5;
  396. res6_8 += a0*b6;
  397. res7_8 += a0*b7;
  398. a1 = ptrba[9];
  399. res0_9 += a1*b0;
  400. res1_9 += a1*b1;
  401. res2_9 += a1*b2;
  402. res3_9 += a1*b3;
  403. res4_9 += a1*b4;
  404. res5_9 += a1*b5;
  405. res6_9 += a1*b6;
  406. res7_9 += a1*b7;
  407. a0 = ptrba[10];
  408. res0_10 += a0*b0;
  409. res1_10 += a0*b1;
  410. res2_10 += a0*b2;
  411. res3_10 += a0*b3;
  412. res4_10 += a0*b4;
  413. res5_10 += a0*b5;
  414. res6_10 += a0*b6;
  415. res7_10 += a0*b7;
  416. a1 = ptrba[11];
  417. res0_11 += a1*b0;
  418. res1_11 += a1*b1;
  419. res2_11 += a1*b2;
  420. res3_11 += a1*b3;
  421. res4_11 += a1*b4;
  422. res5_11 += a1*b5;
  423. res6_11 += a1*b6;
  424. res7_11 += a1*b7;
  425. a0 = ptrba[12];
  426. res0_12 += a0*b0;
  427. res1_12 += a0*b1;
  428. res2_12 += a0*b2;
  429. res3_12 += a0*b3;
  430. res4_12 += a0*b4;
  431. res5_12 += a0*b5;
  432. res6_12 += a0*b6;
  433. res7_12 += a0*b7;
  434. a1 = ptrba[13];
  435. res0_13 += a1*b0;
  436. res1_13 += a1*b1;
  437. res2_13 += a1*b2;
  438. res3_13 += a1*b3;
  439. res4_13 += a1*b4;
  440. res5_13 += a1*b5;
  441. res6_13 += a1*b6;
  442. res7_13 += a1*b7;
  443. a0 = ptrba[14];
  444. res0_14 += a0*b0;
  445. res1_14 += a0*b1;
  446. res2_14 += a0*b2;
  447. res3_14 += a0*b3;
  448. res4_14 += a0*b4;
  449. res5_14 += a0*b5;
  450. res6_14 += a0*b6;
  451. res7_14 += a0*b7;
  452. a1 = ptrba[15];
  453. res0_15 += a1*b0;
  454. res1_15 += a1*b1;
  455. res2_15 += a1*b2;
  456. res3_15 += a1*b3;
  457. res4_15 += a1*b4;
  458. res5_15 += a1*b5;
  459. res6_15 += a1*b6;
  460. res7_15 += a1*b7;
  461. ptrba = ptrba+16;
  462. ptrbb = ptrbb+8;
  463. }
  464. res0_0 *= alpha;
  465. res0_1 *= alpha;
  466. res0_2 *= alpha;
  467. res0_3 *= alpha;
  468. res0_4 *= alpha;
  469. res0_5 *= alpha;
  470. res0_6 *= alpha;
  471. res0_7 *= alpha;
  472. res0_8 *= alpha;
  473. res0_9 *= alpha;
  474. res0_10 *= alpha;
  475. res0_11 *= alpha;
  476. res0_12 *= alpha;
  477. res0_13 *= alpha;
  478. res0_14 *= alpha;
  479. res0_15 *= alpha;
  480. res1_0 *= alpha;
  481. res1_1 *= alpha;
  482. res1_2 *= alpha;
  483. res1_3 *= alpha;
  484. res1_4 *= alpha;
  485. res1_5 *= alpha;
  486. res1_6 *= alpha;
  487. res1_7 *= alpha;
  488. res1_8 *= alpha;
  489. res1_9 *= alpha;
  490. res1_10 *= alpha;
  491. res1_11 *= alpha;
  492. res1_12 *= alpha;
  493. res1_13 *= alpha;
  494. res1_14 *= alpha;
  495. res1_15 *= alpha;
  496. res2_0 *= alpha;
  497. res2_1 *= alpha;
  498. res2_2 *= alpha;
  499. res2_3 *= alpha;
  500. res2_4 *= alpha;
  501. res2_5 *= alpha;
  502. res2_6 *= alpha;
  503. res2_7 *= alpha;
  504. res2_8 *= alpha;
  505. res2_9 *= alpha;
  506. res2_10 *= alpha;
  507. res2_11 *= alpha;
  508. res2_12 *= alpha;
  509. res2_13 *= alpha;
  510. res2_14 *= alpha;
  511. res2_15 *= alpha;
  512. res3_0 *= alpha;
  513. res3_1 *= alpha;
  514. res3_2 *= alpha;
  515. res3_3 *= alpha;
  516. res3_4 *= alpha;
  517. res3_5 *= alpha;
  518. res3_6 *= alpha;
  519. res3_7 *= alpha;
  520. res3_8 *= alpha;
  521. res3_9 *= alpha;
  522. res3_10 *= alpha;
  523. res3_11 *= alpha;
  524. res3_12 *= alpha;
  525. res3_13 *= alpha;
  526. res3_14 *= alpha;
  527. res3_15 *= alpha;
  528. res4_0 *= alpha;
  529. res4_1 *= alpha;
  530. res4_2 *= alpha;
  531. res4_3 *= alpha;
  532. res4_4 *= alpha;
  533. res4_5 *= alpha;
  534. res4_6 *= alpha;
  535. res4_7 *= alpha;
  536. res4_8 *= alpha;
  537. res4_9 *= alpha;
  538. res4_10 *= alpha;
  539. res4_11 *= alpha;
  540. res4_12 *= alpha;
  541. res4_13 *= alpha;
  542. res4_14 *= alpha;
  543. res4_15 *= alpha;
  544. res5_0 *= alpha;
  545. res5_1 *= alpha;
  546. res5_2 *= alpha;
  547. res5_3 *= alpha;
  548. res5_4 *= alpha;
  549. res5_5 *= alpha;
  550. res5_6 *= alpha;
  551. res5_7 *= alpha;
  552. res5_8 *= alpha;
  553. res5_9 *= alpha;
  554. res5_10 *= alpha;
  555. res5_11 *= alpha;
  556. res5_12 *= alpha;
  557. res5_13 *= alpha;
  558. res5_14 *= alpha;
  559. res5_15 *= alpha;
  560. res6_0 *= alpha;
  561. res6_1 *= alpha;
  562. res6_2 *= alpha;
  563. res6_3 *= alpha;
  564. res6_4 *= alpha;
  565. res6_5 *= alpha;
  566. res6_6 *= alpha;
  567. res6_7 *= alpha;
  568. res6_8 *= alpha;
  569. res6_9 *= alpha;
  570. res6_10 *= alpha;
  571. res6_11 *= alpha;
  572. res6_12 *= alpha;
  573. res6_13 *= alpha;
  574. res6_14 *= alpha;
  575. res6_15 *= alpha;
  576. res7_0 *= alpha;
  577. res7_1 *= alpha;
  578. res7_2 *= alpha;
  579. res7_3 *= alpha;
  580. res7_4 *= alpha;
  581. res7_5 *= alpha;
  582. res7_6 *= alpha;
  583. res7_7 *= alpha;
  584. res7_8 *= alpha;
  585. res7_9 *= alpha;
  586. res7_10 *= alpha;
  587. res7_11 *= alpha;
  588. res7_12 *= alpha;
  589. res7_13 *= alpha;
  590. res7_14 *= alpha;
  591. res7_15 *= alpha;
  592. C0[0] = res0_0;
  593. C0[1] = res0_1;
  594. C0[2] = res0_2;
  595. C0[3] = res0_3;
  596. C0[4] = res0_4;
  597. C0[5] = res0_5;
  598. C0[6] = res0_6;
  599. C0[7] = res0_7;
  600. C0[8] = res0_8;
  601. C0[9] = res0_9;
  602. C0[10] = res0_10;
  603. C0[11] = res0_11;
  604. C0[12] = res0_12;
  605. C0[13] = res0_13;
  606. C0[14] = res0_14;
  607. C0[15] = res0_15;
  608. C1[0] = res1_0;
  609. C1[1] = res1_1;
  610. C1[2] = res1_2;
  611. C1[3] = res1_3;
  612. C1[4] = res1_4;
  613. C1[5] = res1_5;
  614. C1[6] = res1_6;
  615. C1[7] = res1_7;
  616. C1[8] = res1_8;
  617. C1[9] = res1_9;
  618. C1[10] = res1_10;
  619. C1[11] = res1_11;
  620. C1[12] = res1_12;
  621. C1[13] = res1_13;
  622. C1[14] = res1_14;
  623. C1[15] = res1_15;
  624. C2[0] = res2_0;
  625. C2[1] = res2_1;
  626. C2[2] = res2_2;
  627. C2[3] = res2_3;
  628. C2[4] = res2_4;
  629. C2[5] = res2_5;
  630. C2[6] = res2_6;
  631. C2[7] = res2_7;
  632. C2[8] = res2_8;
  633. C2[9] = res2_9;
  634. C2[10] = res2_10;
  635. C2[11] = res2_11;
  636. C2[12] = res2_12;
  637. C2[13] = res2_13;
  638. C2[14] = res2_14;
  639. C2[15] = res2_15;
  640. C3[0] = res3_0;
  641. C3[1] = res3_1;
  642. C3[2] = res3_2;
  643. C3[3] = res3_3;
  644. C3[4] = res3_4;
  645. C3[5] = res3_5;
  646. C3[6] = res3_6;
  647. C3[7] = res3_7;
  648. C3[8] = res3_8;
  649. C3[9] = res3_9;
  650. C3[10] = res3_10;
  651. C3[11] = res3_11;
  652. C3[12] = res3_12;
  653. C3[13] = res3_13;
  654. C3[14] = res3_14;
  655. C3[15] = res3_15;
  656. C4[0] = res4_0;
  657. C4[1] = res4_1;
  658. C4[2] = res4_2;
  659. C4[3] = res4_3;
  660. C4[4] = res4_4;
  661. C4[5] = res4_5;
  662. C4[6] = res4_6;
  663. C4[7] = res4_7;
  664. C4[8] = res4_8;
  665. C4[9] = res4_9;
  666. C4[10] = res4_10;
  667. C4[11] = res4_11;
  668. C4[12] = res4_12;
  669. C4[13] = res4_13;
  670. C4[14] = res4_14;
  671. C4[15] = res4_15;
  672. C5[0] = res5_0;
  673. C5[1] = res5_1;
  674. C5[2] = res5_2;
  675. C5[3] = res5_3;
  676. C5[4] = res5_4;
  677. C5[5] = res5_5;
  678. C5[6] = res5_6;
  679. C5[7] = res5_7;
  680. C5[8] = res5_8;
  681. C5[9] = res5_9;
  682. C5[10] = res5_10;
  683. C5[11] = res5_11;
  684. C5[12] = res5_12;
  685. C5[13] = res5_13;
  686. C5[14] = res5_14;
  687. C5[15] = res5_15;
  688. C6[0] = res6_0;
  689. C6[1] = res6_1;
  690. C6[2] = res6_2;
  691. C6[3] = res6_3;
  692. C6[4] = res6_4;
  693. C6[5] = res6_5;
  694. C6[6] = res6_6;
  695. C6[7] = res6_7;
  696. C6[8] = res6_8;
  697. C6[9] = res6_9;
  698. C6[10] = res6_10;
  699. C6[11] = res6_11;
  700. C6[12] = res6_12;
  701. C6[13] = res6_13;
  702. C6[14] = res6_14;
  703. C6[15] = res6_15;
  704. C7[0] = res7_0;
  705. C7[1] = res7_1;
  706. C7[2] = res7_2;
  707. C7[3] = res7_3;
  708. C7[4] = res7_4;
  709. C7[5] = res7_5;
  710. C7[6] = res7_6;
  711. C7[7] = res7_7;
  712. C7[8] = res7_8;
  713. C7[9] = res7_9;
  714. C7[10] = res7_10;
  715. C7[11] = res7_11;
  716. C7[12] = res7_12;
  717. C7[13] = res7_13;
  718. C7[14] = res7_14;
  719. C7[15] = res7_15;
  720. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  721. temp = bk - off;
  722. #ifdef LEFT
  723. temp -= 16; // number of values in A
  724. #else
  725. temp -= 8; // number of values in B
  726. #endif
  727. ptrba += temp*16;
  728. ptrbb += temp*8;
  729. #endif
  730. #ifdef LEFT
  731. off += 16; // number of values in A
  732. #endif
  733. C0 = C0+16;
  734. C1 = C1+16;
  735. C2 = C2+16;
  736. C3 = C3+16;
  737. C4 = C4+16;
  738. C5 = C5+16;
  739. C6 = C6+16;
  740. C7 = C7+16;
  741. }
  742. if ( bm & 8)
  743. {
  744. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  745. ptrbb = bb;
  746. #else
  747. ptrba += off*8;
  748. ptrbb = bb + off*8;
  749. #endif
  750. res0_0 = 0;
  751. res0_1 = 0;
  752. res0_2 = 0;
  753. res0_3 = 0;
  754. res0_4 = 0;
  755. res0_5 = 0;
  756. res0_6 = 0;
  757. res0_7 = 0;
  758. res1_0 = 0;
  759. res1_1 = 0;
  760. res1_2 = 0;
  761. res1_3 = 0;
  762. res1_4 = 0;
  763. res1_5 = 0;
  764. res1_6 = 0;
  765. res1_7 = 0;
  766. res2_0 = 0;
  767. res2_1 = 0;
  768. res2_2 = 0;
  769. res2_3 = 0;
  770. res2_4 = 0;
  771. res2_5 = 0;
  772. res2_6 = 0;
  773. res2_7 = 0;
  774. res3_0 = 0;
  775. res3_1 = 0;
  776. res3_2 = 0;
  777. res3_3 = 0;
  778. res3_4 = 0;
  779. res3_5 = 0;
  780. res3_6 = 0;
  781. res3_7 = 0;
  782. res4_0 = 0;
  783. res4_1 = 0;
  784. res4_2 = 0;
  785. res4_3 = 0;
  786. res4_4 = 0;
  787. res4_5 = 0;
  788. res4_6 = 0;
  789. res4_7 = 0;
  790. res5_0 = 0;
  791. res5_1 = 0;
  792. res5_2 = 0;
  793. res5_3 = 0;
  794. res5_4 = 0;
  795. res5_5 = 0;
  796. res5_6 = 0;
  797. res5_7 = 0;
  798. res6_0 = 0;
  799. res6_1 = 0;
  800. res6_2 = 0;
  801. res6_3 = 0;
  802. res6_4 = 0;
  803. res6_5 = 0;
  804. res6_6 = 0;
  805. res6_7 = 0;
  806. res7_0 = 0;
  807. res7_1 = 0;
  808. res7_2 = 0;
  809. res7_3 = 0;
  810. res7_4 = 0;
  811. res7_5 = 0;
  812. res7_6 = 0;
  813. res7_7 = 0;
  814. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  815. temp = bk-off;
  816. #elif defined(LEFT)
  817. temp = off+8; // number of values in A
  818. #else
  819. temp = off+8; // number of values in B
  820. #endif
  821. for (k=0; k<temp; k++)
  822. {
  823. b0 = ptrbb[0];
  824. b1 = ptrbb[1];
  825. b2 = ptrbb[2];
  826. b3 = ptrbb[3];
  827. b4 = ptrbb[4];
  828. b5 = ptrbb[5];
  829. b6 = ptrbb[6];
  830. b7 = ptrbb[7];
  831. a0 = ptrba[0];
  832. res0_0 += a0*b0;
  833. res1_0 += a0*b1;
  834. res2_0 += a0*b2;
  835. res3_0 += a0*b3;
  836. res4_0 += a0*b4;
  837. res5_0 += a0*b5;
  838. res6_0 += a0*b6;
  839. res7_0 += a0*b7;
  840. a1 = ptrba[1];
  841. res0_1 += a1*b0;
  842. res1_1 += a1*b1;
  843. res2_1 += a1*b2;
  844. res3_1 += a1*b3;
  845. res4_1 += a1*b4;
  846. res5_1 += a1*b5;
  847. res6_1 += a1*b6;
  848. res7_1 += a1*b7;
  849. a0 = ptrba[2];
  850. res0_2 += a0*b0;
  851. res1_2 += a0*b1;
  852. res2_2 += a0*b2;
  853. res3_2 += a0*b3;
  854. res4_2 += a0*b4;
  855. res5_2 += a0*b5;
  856. res6_2 += a0*b6;
  857. res7_2 += a0*b7;
  858. a1 = ptrba[3];
  859. res0_3 += a1*b0;
  860. res1_3 += a1*b1;
  861. res2_3 += a1*b2;
  862. res3_3 += a1*b3;
  863. res4_3 += a1*b4;
  864. res5_3 += a1*b5;
  865. res6_3 += a1*b6;
  866. res7_3 += a1*b7;
  867. a0 = ptrba[4];
  868. res0_4 += a0*b0;
  869. res1_4 += a0*b1;
  870. res2_4 += a0*b2;
  871. res3_4 += a0*b3;
  872. res4_4 += a0*b4;
  873. res5_4 += a0*b5;
  874. res6_4 += a0*b6;
  875. res7_4 += a0*b7;
  876. a1 = ptrba[5];
  877. res0_5 += a1*b0;
  878. res1_5 += a1*b1;
  879. res2_5 += a1*b2;
  880. res3_5 += a1*b3;
  881. res4_5 += a1*b4;
  882. res5_5 += a1*b5;
  883. res6_5 += a1*b6;
  884. res7_5 += a1*b7;
  885. a0 = ptrba[6];
  886. res0_6 += a0*b0;
  887. res1_6 += a0*b1;
  888. res2_6 += a0*b2;
  889. res3_6 += a0*b3;
  890. res4_6 += a0*b4;
  891. res5_6 += a0*b5;
  892. res6_6 += a0*b6;
  893. res7_6 += a0*b7;
  894. a1 = ptrba[7];
  895. res0_7 += a1*b0;
  896. res1_7 += a1*b1;
  897. res2_7 += a1*b2;
  898. res3_7 += a1*b3;
  899. res4_7 += a1*b4;
  900. res5_7 += a1*b5;
  901. res6_7 += a1*b6;
  902. res7_7 += a1*b7;
  903. ptrba = ptrba+8;
  904. ptrbb = ptrbb+8;
  905. }
  906. res0_0 *= alpha;
  907. res0_1 *= alpha;
  908. res0_2 *= alpha;
  909. res0_3 *= alpha;
  910. res0_4 *= alpha;
  911. res0_5 *= alpha;
  912. res0_6 *= alpha;
  913. res0_7 *= alpha;
  914. res1_0 *= alpha;
  915. res1_1 *= alpha;
  916. res1_2 *= alpha;
  917. res1_3 *= alpha;
  918. res1_4 *= alpha;
  919. res1_5 *= alpha;
  920. res1_6 *= alpha;
  921. res1_7 *= alpha;
  922. res2_0 *= alpha;
  923. res2_1 *= alpha;
  924. res2_2 *= alpha;
  925. res2_3 *= alpha;
  926. res2_4 *= alpha;
  927. res2_5 *= alpha;
  928. res2_6 *= alpha;
  929. res2_7 *= alpha;
  930. res3_0 *= alpha;
  931. res3_1 *= alpha;
  932. res3_2 *= alpha;
  933. res3_3 *= alpha;
  934. res3_4 *= alpha;
  935. res3_5 *= alpha;
  936. res3_6 *= alpha;
  937. res3_7 *= alpha;
  938. res4_0 *= alpha;
  939. res4_1 *= alpha;
  940. res4_2 *= alpha;
  941. res4_3 *= alpha;
  942. res4_4 *= alpha;
  943. res4_5 *= alpha;
  944. res4_6 *= alpha;
  945. res4_7 *= alpha;
  946. res5_0 *= alpha;
  947. res5_1 *= alpha;
  948. res5_2 *= alpha;
  949. res5_3 *= alpha;
  950. res5_4 *= alpha;
  951. res5_5 *= alpha;
  952. res5_6 *= alpha;
  953. res5_7 *= alpha;
  954. res6_0 *= alpha;
  955. res6_1 *= alpha;
  956. res6_2 *= alpha;
  957. res6_3 *= alpha;
  958. res6_4 *= alpha;
  959. res6_5 *= alpha;
  960. res6_6 *= alpha;
  961. res6_7 *= alpha;
  962. res7_0 *= alpha;
  963. res7_1 *= alpha;
  964. res7_2 *= alpha;
  965. res7_3 *= alpha;
  966. res7_4 *= alpha;
  967. res7_5 *= alpha;
  968. res7_6 *= alpha;
  969. res7_7 *= alpha;
  970. C0[0] = res0_0;
  971. C0[1] = res0_1;
  972. C0[2] = res0_2;
  973. C0[3] = res0_3;
  974. C0[4] = res0_4;
  975. C0[5] = res0_5;
  976. C0[6] = res0_6;
  977. C0[7] = res0_7;
  978. C1[0] = res1_0;
  979. C1[1] = res1_1;
  980. C1[2] = res1_2;
  981. C1[3] = res1_3;
  982. C1[4] = res1_4;
  983. C1[5] = res1_5;
  984. C1[6] = res1_6;
  985. C1[7] = res1_7;
  986. C2[0] = res2_0;
  987. C2[1] = res2_1;
  988. C2[2] = res2_2;
  989. C2[3] = res2_3;
  990. C2[4] = res2_4;
  991. C2[5] = res2_5;
  992. C2[6] = res2_6;
  993. C2[7] = res2_7;
  994. C3[0] = res3_0;
  995. C3[1] = res3_1;
  996. C3[2] = res3_2;
  997. C3[3] = res3_3;
  998. C3[4] = res3_4;
  999. C3[5] = res3_5;
  1000. C3[6] = res3_6;
  1001. C3[7] = res3_7;
  1002. C4[0] = res4_0;
  1003. C4[1] = res4_1;
  1004. C4[2] = res4_2;
  1005. C4[3] = res4_3;
  1006. C4[4] = res4_4;
  1007. C4[5] = res4_5;
  1008. C4[6] = res4_6;
  1009. C4[7] = res4_7;
  1010. C5[0] = res5_0;
  1011. C5[1] = res5_1;
  1012. C5[2] = res5_2;
  1013. C5[3] = res5_3;
  1014. C5[4] = res5_4;
  1015. C5[5] = res5_5;
  1016. C5[6] = res5_6;
  1017. C5[7] = res5_7;
  1018. C6[0] = res6_0;
  1019. C6[1] = res6_1;
  1020. C6[2] = res6_2;
  1021. C6[3] = res6_3;
  1022. C6[4] = res6_4;
  1023. C6[5] = res6_5;
  1024. C6[6] = res6_6;
  1025. C6[7] = res6_7;
  1026. C7[0] = res7_0;
  1027. C7[1] = res7_1;
  1028. C7[2] = res7_2;
  1029. C7[3] = res7_3;
  1030. C7[4] = res7_4;
  1031. C7[5] = res7_5;
  1032. C7[6] = res7_6;
  1033. C7[7] = res7_7;
  1034. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  1035. temp = bk - off;
  1036. #ifdef LEFT
  1037. temp -= 8; // number of values in A
  1038. #else
  1039. temp -= 8; // number of values in B
  1040. #endif
  1041. ptrba += temp*8;
  1042. ptrbb += temp*8;
  1043. #endif
  1044. #ifdef LEFT
  1045. off += 8; // number of values in A
  1046. #endif
  1047. C0 = C0+8;
  1048. C1 = C1+8;
  1049. C2 = C2+8;
  1050. C3 = C3+8;
  1051. C4 = C4+8;
  1052. C5 = C5+8;
  1053. C6 = C6+8;
  1054. C7 = C7+8;
  1055. }
  1056. if ( bm & 4 )
  1057. {
  1058. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  1059. ptrbb = bb;
  1060. #else
  1061. ptrba += off*4;
  1062. ptrbb = bb + off*8;
  1063. #endif
  1064. res0_0 = 0;
  1065. res1_0 = 0;
  1066. res2_0 = 0;
  1067. res3_0 = 0;
  1068. res4_0 = 0;
  1069. res5_0 = 0;
  1070. res6_0 = 0;
  1071. res7_0 = 0;
  1072. res0_1 = 0;
  1073. res1_1 = 0;
  1074. res2_1 = 0;
  1075. res3_1 = 0;
  1076. res4_1 = 0;
  1077. res5_1 = 0;
  1078. res6_1 = 0;
  1079. res7_1 = 0;
  1080. res0_2 = 0;
  1081. res1_2 = 0;
  1082. res2_2 = 0;
  1083. res3_2 = 0;
  1084. res4_2 = 0;
  1085. res5_2 = 0;
  1086. res6_2 = 0;
  1087. res7_2 = 0;
  1088. res0_3 = 0;
  1089. res1_3 = 0;
  1090. res2_3 = 0;
  1091. res3_3 = 0;
  1092. res4_3 = 0;
  1093. res5_3 = 0;
  1094. res6_3 = 0;
  1095. res7_3 = 0;
  1096. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  1097. temp = bk-off;
  1098. #elif defined(LEFT)
  1099. temp = off+4; // number of values in A
  1100. #else
  1101. temp = off+8; // number of values in B
  1102. #endif
  1103. for (k=0; k<temp; k++)
  1104. {
  1105. b0 = ptrbb[0];
  1106. b1 = ptrbb[1];
  1107. b2 = ptrbb[2];
  1108. b3 = ptrbb[3];
  1109. b4 = ptrbb[4];
  1110. b5 = ptrbb[5];
  1111. b6 = ptrbb[6];
  1112. b7 = ptrbb[7];
  1113. a0 = ptrba[0];
  1114. res0_0 += a0*b0;
  1115. res1_0 += a0*b1;
  1116. res2_0 += a0*b2;
  1117. res3_0 += a0*b3;
  1118. res4_0 += a0*b4;
  1119. res5_0 += a0*b5;
  1120. res6_0 += a0*b6;
  1121. res7_0 += a0*b7;
  1122. a1 = ptrba[1];
  1123. res0_1 += a1*b0;
  1124. res1_1 += a1*b1;
  1125. res2_1 += a1*b2;
  1126. res3_1 += a1*b3;
  1127. res4_1 += a1*b4;
  1128. res5_1 += a1*b5;
  1129. res6_1 += a1*b6;
  1130. res7_1 += a1*b7;
  1131. a0 = ptrba[2];
  1132. res0_2 += a0*b0;
  1133. res1_2 += a0*b1;
  1134. res2_2 += a0*b2;
  1135. res3_2 += a0*b3;
  1136. res4_2 += a0*b4;
  1137. res5_2 += a0*b5;
  1138. res6_2 += a0*b6;
  1139. res7_2 += a0*b7;
  1140. a1 = ptrba[3];
  1141. res0_3 += a1*b0;
  1142. res1_3 += a1*b1;
  1143. res2_3 += a1*b2;
  1144. res3_3 += a1*b3;
  1145. res4_3 += a1*b4;
  1146. res5_3 += a1*b5;
  1147. res6_3 += a1*b6;
  1148. res7_3 += a1*b7;
  1149. ptrba = ptrba+4;
  1150. ptrbb = ptrbb+8;
  1151. }
  1152. res0_0 *= alpha;
  1153. res0_1 *= alpha;
  1154. res0_2 *= alpha;
  1155. res0_3 *= alpha;
  1156. res1_0 *= alpha;
  1157. res1_1 *= alpha;
  1158. res1_2 *= alpha;
  1159. res1_3 *= alpha;
  1160. res2_0 *= alpha;
  1161. res2_1 *= alpha;
  1162. res2_2 *= alpha;
  1163. res2_3 *= alpha;
  1164. res3_0 *= alpha;
  1165. res3_1 *= alpha;
  1166. res3_2 *= alpha;
  1167. res3_3 *= alpha;
  1168. res4_0 *= alpha;
  1169. res4_1 *= alpha;
  1170. res4_2 *= alpha;
  1171. res4_3 *= alpha;
  1172. res5_0 *= alpha;
  1173. res5_1 *= alpha;
  1174. res5_2 *= alpha;
  1175. res5_3 *= alpha;
  1176. res6_0 *= alpha;
  1177. res6_1 *= alpha;
  1178. res6_2 *= alpha;
  1179. res6_3 *= alpha;
  1180. res7_0 *= alpha;
  1181. res7_1 *= alpha;
  1182. res7_2 *= alpha;
  1183. res7_3 *= alpha;
  1184. C0[0] = res0_0;
  1185. C0[1] = res0_1;
  1186. C0[2] = res0_2;
  1187. C0[3] = res0_3;
  1188. C1[0] = res1_0;
  1189. C1[1] = res1_1;
  1190. C1[2] = res1_2;
  1191. C1[3] = res1_3;
  1192. C2[0] = res2_0;
  1193. C2[1] = res2_1;
  1194. C2[2] = res2_2;
  1195. C2[3] = res2_3;
  1196. C3[0] = res3_0;
  1197. C3[1] = res3_1;
  1198. C3[2] = res3_2;
  1199. C3[3] = res3_3;
  1200. C4[0] = res4_0;
  1201. C4[1] = res4_1;
  1202. C4[2] = res4_2;
  1203. C4[3] = res4_3;
  1204. C5[0] = res5_0;
  1205. C5[1] = res5_1;
  1206. C5[2] = res5_2;
  1207. C5[3] = res5_3;
  1208. C6[0] = res6_0;
  1209. C6[1] = res6_1;
  1210. C6[2] = res6_2;
  1211. C6[3] = res6_3;
  1212. C7[0] = res7_0;
  1213. C7[1] = res7_1;
  1214. C7[2] = res7_2;
  1215. C7[3] = res7_3;
  1216. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  1217. temp = bk - off;
  1218. #ifdef LEFT
  1219. temp -= 4; // number of values in A
  1220. #else
  1221. temp -= 8; // number of values in B
  1222. #endif
  1223. ptrba += temp*4;
  1224. ptrbb += temp*8;
  1225. #endif
  1226. #ifdef LEFT
  1227. off += 4; // number of values in A
  1228. #endif
  1229. C0 = C0+4;
  1230. C1 = C1+4;
  1231. C2 = C2+4;
  1232. C3 = C3+4;
  1233. C4 = C4+4;
  1234. C5 = C5+4;
  1235. C6 = C6+4;
  1236. C7 = C7+4;
  1237. }
  1238. if ( bm & 2 )
  1239. {
  1240. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  1241. ptrbb = bb;
  1242. #else
  1243. ptrba += off*2;
  1244. ptrbb = bb + off*8;
  1245. #endif
  1246. res0_0 = 0;
  1247. res0_1 = 0;
  1248. res1_0 = 0;
  1249. res1_1 = 0;
  1250. res2_0 = 0;
  1251. res2_1 = 0;
  1252. res3_0 = 0;
  1253. res3_1 = 0;
  1254. res4_0 = 0;
  1255. res4_1 = 0;
  1256. res5_0 = 0;
  1257. res5_1 = 0;
  1258. res6_0 = 0;
  1259. res6_1 = 0;
  1260. res7_0 = 0;
  1261. res7_1 = 0;
  1262. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  1263. temp = bk-off;
  1264. #elif defined(LEFT)
  1265. temp = off+2; // number of values in A
  1266. #else
  1267. temp = off+8; // number of values in B
  1268. #endif
  1269. for (k=0; k<temp; k++)
  1270. {
  1271. b0 = ptrbb[0];
  1272. b1 = ptrbb[1];
  1273. b2 = ptrbb[2];
  1274. b3 = ptrbb[3];
  1275. b4 = ptrbb[4];
  1276. b5 = ptrbb[5];
  1277. b6 = ptrbb[6];
  1278. b7 = ptrbb[7];
  1279. a0 = ptrba[0];
  1280. res0_0 += a0*b0;
  1281. res1_0 += a0*b1;
  1282. res2_0 += a0*b2;
  1283. res3_0 += a0*b3;
  1284. res4_0 += a0*b4;
  1285. res5_0 += a0*b5;
  1286. res6_0 += a0*b6;
  1287. res7_0 += a0*b7;
  1288. a1 = ptrba[1];
  1289. res0_1 += a1*b0;
  1290. res1_1 += a1*b1;
  1291. res2_1 += a1*b2;
  1292. res3_1 += a1*b3;
  1293. res4_1 += a1*b4;
  1294. res5_1 += a1*b5;
  1295. res6_1 += a1*b6;
  1296. res7_1 += a1*b7;
  1297. ptrba = ptrba+2;
  1298. ptrbb = ptrbb+8;
  1299. }
  1300. res0_0 *= alpha;
  1301. res0_1 *= alpha;
  1302. res1_0 *= alpha;
  1303. res1_1 *= alpha;
  1304. res2_0 *= alpha;
  1305. res2_1 *= alpha;
  1306. res3_0 *= alpha;
  1307. res3_1 *= alpha;
  1308. res4_0 *= alpha;
  1309. res4_1 *= alpha;
  1310. res5_0 *= alpha;
  1311. res5_1 *= alpha;
  1312. res6_0 *= alpha;
  1313. res6_1 *= alpha;
  1314. res7_0 *= alpha;
  1315. res7_1 *= alpha;
  1316. C0[0] = res0_0;
  1317. C0[1] = res0_1;
  1318. C1[0] = res1_0;
  1319. C1[1] = res1_1;
  1320. C2[0] = res2_0;
  1321. C2[1] = res2_1;
  1322. C3[0] = res3_0;
  1323. C3[1] = res3_1;
  1324. C4[0] = res4_0;
  1325. C4[1] = res4_1;
  1326. C5[0] = res5_0;
  1327. C5[1] = res5_1;
  1328. C6[0] = res6_0;
  1329. C6[1] = res6_1;
  1330. C7[0] = res7_0;
  1331. C7[1] = res7_1;
  1332. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  1333. temp = bk - off;
  1334. #ifdef LEFT
  1335. temp -= 2; // number of values in A
  1336. #else
  1337. temp -= 8; // number of values in B
  1338. #endif
  1339. ptrba += temp*2;
  1340. ptrbb += temp*8;
  1341. #endif
  1342. #ifdef LEFT
  1343. off += 2; // number of values in A
  1344. #endif
  1345. C0 = C0+2;
  1346. C1 = C1+2;
  1347. C2 = C2+2;
  1348. C3 = C3+2;
  1349. C4 = C4+2;
  1350. C5 = C5+2;
  1351. C6 = C6+2;
  1352. C7 = C7+2;
  1353. }
  1354. if ( bm & 1 )
  1355. {
  1356. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  1357. ptrbb = bb;
  1358. #else
  1359. ptrba += off*1;
  1360. ptrbb = bb + off*8;
  1361. #endif
  1362. res0_0 = 0;
  1363. res1_0 = 0;
  1364. res2_0 = 0;
  1365. res3_0 = 0;
  1366. res4_0 = 0;
  1367. res5_0 = 0;
  1368. res6_0 = 0;
  1369. res7_0 = 0;
  1370. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  1371. temp = bk-off;
  1372. #elif defined(LEFT)
  1373. temp = off+1; // number of values in A
  1374. #else
  1375. temp = off+8; // number of values in B
  1376. #endif
  1377. for (k=0; k<temp; k++)
  1378. {
  1379. b0 = ptrbb[0];
  1380. b1 = ptrbb[1];
  1381. b2 = ptrbb[2];
  1382. b3 = ptrbb[3];
  1383. b4 = ptrbb[4];
  1384. b5 = ptrbb[5];
  1385. b6 = ptrbb[6];
  1386. b7 = ptrbb[7];
  1387. a0 = ptrba[0];
  1388. res0_0 += a0*b0;
  1389. res1_0 += a0*b1;
  1390. res2_0 += a0*b2;
  1391. res3_0 += a0*b3;
  1392. res4_0 += a0*b4;
  1393. res5_0 += a0*b5;
  1394. res6_0 += a0*b6;
  1395. res7_0 += a0*b7;
  1396. ptrba = ptrba+1;
  1397. ptrbb = ptrbb+8;
  1398. }
  1399. res0_0 *= alpha;
  1400. res1_0 *= alpha;
  1401. res2_0 *= alpha;
  1402. res3_0 *= alpha;
  1403. res4_0 *= alpha;
  1404. res5_0 *= alpha;
  1405. res6_0 *= alpha;
  1406. res7_0 *= alpha;
  1407. C0[0] = res0_0;
  1408. C1[0] = res1_0;
  1409. C2[0] = res2_0;
  1410. C3[0] = res3_0;
  1411. C4[0] = res4_0;
  1412. C5[0] = res5_0;
  1413. C6[0] = res6_0;
  1414. C7[0] = res7_0;
  1415. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  1416. temp = bk - off;
  1417. #ifdef LEFT
  1418. temp -= 1; // number of values in A
  1419. #else
  1420. temp -= 8; // number of values in B
  1421. #endif
  1422. ptrba += temp*1;
  1423. ptrbb += temp*8;
  1424. #endif
  1425. #ifdef LEFT
  1426. off += 1; // number of values in A
  1427. #endif
  1428. C0 = C0+1;
  1429. C1 = C1+1;
  1430. C2 = C2+1;
  1431. C3 = C3+1;
  1432. C4 = C4+1;
  1433. C5 = C5+1;
  1434. C6 = C6+1;
  1435. C7 = C7+1;
  1436. }
  1437. #if defined(TRMMKERNEL) && !defined(LEFT)
  1438. off += 8;
  1439. #endif
  1440. k = (bk<<3);
  1441. bb = bb+k;
  1442. i = (ldc<<3);
  1443. C = C+i;
  1444. }
  1445. if( bn & 4 )
  1446. {
  1447. C0 = C;
  1448. C1 = C0+ldc;
  1449. C2 = C0+2*ldc;
  1450. C3 = C0+3*ldc;
  1451. #if defined(TRMMKERNEL) && defined(LEFT)
  1452. off = offset;
  1453. #endif
  1454. ptrba = ba;
  1455. for (i=0; i<bm/16; i+=1)
  1456. {
  1457. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  1458. ptrbb = bb;
  1459. #else
  1460. ptrba += off*16;
  1461. ptrbb = bb + off*4;
  1462. #endif
  1463. res0_0 = 0;
  1464. res0_1 = 0;
  1465. res0_2 = 0;
  1466. res0_3 = 0;
  1467. res0_4 = 0;
  1468. res0_5 = 0;
  1469. res0_6 = 0;
  1470. res0_7 = 0;
  1471. res0_8 = 0;
  1472. res0_9 = 0;
  1473. res0_10 = 0;
  1474. res0_11 = 0;
  1475. res0_12 = 0;
  1476. res0_13 = 0;
  1477. res0_14 = 0;
  1478. res0_15 = 0;
  1479. res1_0 = 0;
  1480. res1_1 = 0;
  1481. res1_2 = 0;
  1482. res1_3 = 0;
  1483. res1_4 = 0;
  1484. res1_5 = 0;
  1485. res1_6 = 0;
  1486. res1_7 = 0;
  1487. res1_8 = 0;
  1488. res1_9 = 0;
  1489. res1_10 = 0;
  1490. res1_11 = 0;
  1491. res1_12 = 0;
  1492. res1_13 = 0;
  1493. res1_14 = 0;
  1494. res1_15 = 0;
  1495. res2_0 = 0;
  1496. res2_1 = 0;
  1497. res2_2 = 0;
  1498. res2_3 = 0;
  1499. res2_4 = 0;
  1500. res2_5 = 0;
  1501. res2_6 = 0;
  1502. res2_7 = 0;
  1503. res2_8 = 0;
  1504. res2_9 = 0;
  1505. res2_10 = 0;
  1506. res2_11 = 0;
  1507. res2_12 = 0;
  1508. res2_13 = 0;
  1509. res2_14 = 0;
  1510. res2_15 = 0;
  1511. res3_0 = 0;
  1512. res3_1 = 0;
  1513. res3_2 = 0;
  1514. res3_3 = 0;
  1515. res3_4 = 0;
  1516. res3_5 = 0;
  1517. res3_6 = 0;
  1518. res3_7 = 0;
  1519. res3_8 = 0;
  1520. res3_9 = 0;
  1521. res3_10 = 0;
  1522. res3_11 = 0;
  1523. res3_12 = 0;
  1524. res3_13 = 0;
  1525. res3_14 = 0;
  1526. res3_15 = 0;
  1527. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  1528. temp = bk-off;
  1529. #elif defined(LEFT)
  1530. temp = off+16; // number of values in A
  1531. #else
  1532. temp = off+4; // number of values in B
  1533. #endif
  1534. for (k=0; k<temp; k++)
  1535. {
  1536. b0 = ptrbb[0];
  1537. b1 = ptrbb[1];
  1538. b2 = ptrbb[2];
  1539. b3 = ptrbb[3];
  1540. a0 = ptrba[0];
  1541. res0_0 += a0*b0;
  1542. res1_0 += a0*b1;
  1543. res2_0 += a0*b2;
  1544. res3_0 += a0*b3;
  1545. a1 = ptrba[1];
  1546. res0_1 += a1*b0;
  1547. res1_1 += a1*b1;
  1548. res2_1 += a1*b2;
  1549. res3_1 += a1*b3;
  1550. a0 = ptrba[2];
  1551. res0_2 += a0*b0;
  1552. res1_2 += a0*b1;
  1553. res2_2 += a0*b2;
  1554. res3_2 += a0*b3;
  1555. a1 = ptrba[3];
  1556. res0_3 += a1*b0;
  1557. res1_3 += a1*b1;
  1558. res2_3 += a1*b2;
  1559. res3_3 += a1*b3;
  1560. a0 = ptrba[4];
  1561. res0_4 += a0*b0;
  1562. res1_4 += a0*b1;
  1563. res2_4 += a0*b2;
  1564. res3_4 += a0*b3;
  1565. a1 = ptrba[5];
  1566. res0_5 += a1*b0;
  1567. res1_5 += a1*b1;
  1568. res2_5 += a1*b2;
  1569. res3_5 += a1*b3;
  1570. a0 = ptrba[6];
  1571. res0_6 += a0*b0;
  1572. res1_6 += a0*b1;
  1573. res2_6 += a0*b2;
  1574. res3_6 += a0*b3;
  1575. a1 = ptrba[7];
  1576. res0_7 += a1*b0;
  1577. res1_7 += a1*b1;
  1578. res2_7 += a1*b2;
  1579. res3_7 += a1*b3;
  1580. a0 = ptrba[8];
  1581. res0_8 += a0*b0;
  1582. res1_8 += a0*b1;
  1583. res2_8 += a0*b2;
  1584. res3_8 += a0*b3;
  1585. a1 = ptrba[9];
  1586. res0_9 += a1*b0;
  1587. res1_9 += a1*b1;
  1588. res2_9 += a1*b2;
  1589. res3_9 += a1*b3;
  1590. a0 = ptrba[10];
  1591. res0_10 += a0*b0;
  1592. res1_10 += a0*b1;
  1593. res2_10 += a0*b2;
  1594. res3_10 += a0*b3;
  1595. a1 = ptrba[11];
  1596. res0_11 += a1*b0;
  1597. res1_11 += a1*b1;
  1598. res2_11 += a1*b2;
  1599. res3_11 += a1*b3;
  1600. a0 = ptrba[12];
  1601. res0_12 += a0*b0;
  1602. res1_12 += a0*b1;
  1603. res2_12 += a0*b2;
  1604. res3_12 += a0*b3;
  1605. a1 = ptrba[13];
  1606. res0_13 += a1*b0;
  1607. res1_13 += a1*b1;
  1608. res2_13 += a1*b2;
  1609. res3_13 += a1*b3;
  1610. a0 = ptrba[14];
  1611. res0_14 += a0*b0;
  1612. res1_14 += a0*b1;
  1613. res2_14 += a0*b2;
  1614. res3_14 += a0*b3;
  1615. a1 = ptrba[15];
  1616. res0_15 += a1*b0;
  1617. res1_15 += a1*b1;
  1618. res2_15 += a1*b2;
  1619. res3_15 += a1*b3;
  1620. ptrba = ptrba+16;
  1621. ptrbb = ptrbb+4;
  1622. }
  1623. res0_0 *= alpha;
  1624. res0_1 *= alpha;
  1625. res0_2 *= alpha;
  1626. res0_3 *= alpha;
  1627. res0_4 *= alpha;
  1628. res0_5 *= alpha;
  1629. res0_6 *= alpha;
  1630. res0_7 *= alpha;
  1631. res0_8 *= alpha;
  1632. res0_9 *= alpha;
  1633. res0_10 *= alpha;
  1634. res0_11 *= alpha;
  1635. res0_12 *= alpha;
  1636. res0_13 *= alpha;
  1637. res0_14 *= alpha;
  1638. res0_15 *= alpha;
  1639. res1_0 *= alpha;
  1640. res1_1 *= alpha;
  1641. res1_2 *= alpha;
  1642. res1_3 *= alpha;
  1643. res1_4 *= alpha;
  1644. res1_5 *= alpha;
  1645. res1_6 *= alpha;
  1646. res1_7 *= alpha;
  1647. res1_8 *= alpha;
  1648. res1_9 *= alpha;
  1649. res1_10 *= alpha;
  1650. res1_11 *= alpha;
  1651. res1_12 *= alpha;
  1652. res1_13 *= alpha;
  1653. res1_14 *= alpha;
  1654. res1_15 *= alpha;
  1655. res2_0 *= alpha;
  1656. res2_1 *= alpha;
  1657. res2_2 *= alpha;
  1658. res2_3 *= alpha;
  1659. res2_4 *= alpha;
  1660. res2_5 *= alpha;
  1661. res2_6 *= alpha;
  1662. res2_7 *= alpha;
  1663. res2_8 *= alpha;
  1664. res2_9 *= alpha;
  1665. res2_10 *= alpha;
  1666. res2_11 *= alpha;
  1667. res2_12 *= alpha;
  1668. res2_13 *= alpha;
  1669. res2_14 *= alpha;
  1670. res2_15 *= alpha;
  1671. res3_0 *= alpha;
  1672. res3_1 *= alpha;
  1673. res3_2 *= alpha;
  1674. res3_3 *= alpha;
  1675. res3_4 *= alpha;
  1676. res3_5 *= alpha;
  1677. res3_6 *= alpha;
  1678. res3_7 *= alpha;
  1679. res3_8 *= alpha;
  1680. res3_9 *= alpha;
  1681. res3_10 *= alpha;
  1682. res3_11 *= alpha;
  1683. res3_12 *= alpha;
  1684. res3_13 *= alpha;
  1685. res3_14 *= alpha;
  1686. res3_15 *= alpha;
  1687. C0[0] = res0_0;
  1688. C0[1] = res0_1;
  1689. C0[2] = res0_2;
  1690. C0[3] = res0_3;
  1691. C0[4] = res0_4;
  1692. C0[5] = res0_5;
  1693. C0[6] = res0_6;
  1694. C0[7] = res0_7;
  1695. C0[8] = res0_8;
  1696. C0[9] = res0_9;
  1697. C0[10] = res0_10;
  1698. C0[11] = res0_11;
  1699. C0[12] = res0_12;
  1700. C0[13] = res0_13;
  1701. C0[14] = res0_14;
  1702. C0[15] = res0_15;
  1703. C1[0] = res1_0;
  1704. C1[1] = res1_1;
  1705. C1[2] = res1_2;
  1706. C1[3] = res1_3;
  1707. C1[4] = res1_4;
  1708. C1[5] = res1_5;
  1709. C1[6] = res1_6;
  1710. C1[7] = res1_7;
  1711. C1[8] = res1_8;
  1712. C1[9] = res1_9;
  1713. C1[10] = res1_10;
  1714. C1[11] = res1_11;
  1715. C1[12] = res1_12;
  1716. C1[13] = res1_13;
  1717. C1[14] = res1_14;
  1718. C1[15] = res1_15;
  1719. C2[0] = res2_0;
  1720. C2[1] = res2_1;
  1721. C2[2] = res2_2;
  1722. C2[3] = res2_3;
  1723. C2[4] = res2_4;
  1724. C2[5] = res2_5;
  1725. C2[6] = res2_6;
  1726. C2[7] = res2_7;
  1727. C2[8] = res2_8;
  1728. C2[9] = res2_9;
  1729. C2[10] = res2_10;
  1730. C2[11] = res2_11;
  1731. C2[12] = res2_12;
  1732. C2[13] = res2_13;
  1733. C2[14] = res2_14;
  1734. C2[15] = res2_15;
  1735. C3[0] = res3_0;
  1736. C3[1] = res3_1;
  1737. C3[2] = res3_2;
  1738. C3[3] = res3_3;
  1739. C3[4] = res3_4;
  1740. C3[5] = res3_5;
  1741. C3[6] = res3_6;
  1742. C3[7] = res3_7;
  1743. C3[8] = res3_8;
  1744. C3[9] = res3_9;
  1745. C3[10] = res3_10;
  1746. C3[11] = res3_11;
  1747. C3[12] = res3_12;
  1748. C3[13] = res3_13;
  1749. C3[14] = res3_14;
  1750. C3[15] = res3_15;
  1751. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  1752. temp = bk - off;
  1753. #ifdef LEFT
  1754. temp -= 16; // number of values in A
  1755. #else
  1756. temp -= 4; // number of values in B
  1757. #endif
  1758. ptrba += temp*16;
  1759. ptrbb += temp*4;
  1760. #endif
  1761. #ifdef LEFT
  1762. off += 16; // number of values in A
  1763. #endif
  1764. C0 = C0+16;
  1765. C1 = C1+16;
  1766. C2 = C2+16;
  1767. C3 = C3+16;
  1768. }
  1769. if ( bm & 8)
  1770. {
  1771. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  1772. ptrbb = bb;
  1773. #else
  1774. ptrba += off*8;
  1775. ptrbb = bb + off*4;
  1776. #endif
  1777. res0_0 = 0;
  1778. res0_1 = 0;
  1779. res0_2 = 0;
  1780. res0_3 = 0;
  1781. res0_4 = 0;
  1782. res0_5 = 0;
  1783. res0_6 = 0;
  1784. res0_7 = 0;
  1785. res1_0 = 0;
  1786. res1_1 = 0;
  1787. res1_2 = 0;
  1788. res1_3 = 0;
  1789. res1_4 = 0;
  1790. res1_5 = 0;
  1791. res1_6 = 0;
  1792. res1_7 = 0;
  1793. res2_0 = 0;
  1794. res2_1 = 0;
  1795. res2_2 = 0;
  1796. res2_3 = 0;
  1797. res2_4 = 0;
  1798. res2_5 = 0;
  1799. res2_6 = 0;
  1800. res2_7 = 0;
  1801. res3_0 = 0;
  1802. res3_1 = 0;
  1803. res3_2 = 0;
  1804. res3_3 = 0;
  1805. res3_4 = 0;
  1806. res3_5 = 0;
  1807. res3_6 = 0;
  1808. res3_7 = 0;
  1809. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  1810. temp = bk-off;
  1811. #elif defined(LEFT)
  1812. temp = off+8; // number of values in A
  1813. #else
  1814. temp = off+4; // number of values in B
  1815. #endif
  1816. for (k=0; k<temp; k++)
  1817. {
  1818. b0 = ptrbb[0];
  1819. b1 = ptrbb[1];
  1820. b2 = ptrbb[2];
  1821. b3 = ptrbb[3];
  1822. a0 = ptrba[0];
  1823. res0_0 += a0*b0;
  1824. res1_0 += a0*b1;
  1825. res2_0 += a0*b2;
  1826. res3_0 += a0*b3;
  1827. a1 = ptrba[1];
  1828. res0_1 += a1*b0;
  1829. res1_1 += a1*b1;
  1830. res2_1 += a1*b2;
  1831. res3_1 += a1*b3;
  1832. a0 = ptrba[2];
  1833. res0_2 += a0*b0;
  1834. res1_2 += a0*b1;
  1835. res2_2 += a0*b2;
  1836. res3_2 += a0*b3;
  1837. a1 = ptrba[3];
  1838. res0_3 += a1*b0;
  1839. res1_3 += a1*b1;
  1840. res2_3 += a1*b2;
  1841. res3_3 += a1*b3;
  1842. a0 = ptrba[4];
  1843. res0_4 += a0*b0;
  1844. res1_4 += a0*b1;
  1845. res2_4 += a0*b2;
  1846. res3_4 += a0*b3;
  1847. a1 = ptrba[5];
  1848. res0_5 += a1*b0;
  1849. res1_5 += a1*b1;
  1850. res2_5 += a1*b2;
  1851. res3_5 += a1*b3;
  1852. a0 = ptrba[6];
  1853. res0_6 += a0*b0;
  1854. res1_6 += a0*b1;
  1855. res2_6 += a0*b2;
  1856. res3_6 += a0*b3;
  1857. a1 = ptrba[7];
  1858. res0_7 += a1*b0;
  1859. res1_7 += a1*b1;
  1860. res2_7 += a1*b2;
  1861. res3_7 += a1*b3;
  1862. ptrba = ptrba+8;
  1863. ptrbb = ptrbb+4;
  1864. }
  1865. res0_0 *= alpha;
  1866. res0_1 *= alpha;
  1867. res0_2 *= alpha;
  1868. res0_3 *= alpha;
  1869. res0_4 *= alpha;
  1870. res0_5 *= alpha;
  1871. res0_6 *= alpha;
  1872. res0_7 *= alpha;
  1873. res1_0 *= alpha;
  1874. res1_1 *= alpha;
  1875. res1_2 *= alpha;
  1876. res1_3 *= alpha;
  1877. res1_4 *= alpha;
  1878. res1_5 *= alpha;
  1879. res1_6 *= alpha;
  1880. res1_7 *= alpha;
  1881. res2_0 *= alpha;
  1882. res2_1 *= alpha;
  1883. res2_2 *= alpha;
  1884. res2_3 *= alpha;
  1885. res2_4 *= alpha;
  1886. res2_5 *= alpha;
  1887. res2_6 *= alpha;
  1888. res2_7 *= alpha;
  1889. res3_0 *= alpha;
  1890. res3_1 *= alpha;
  1891. res3_2 *= alpha;
  1892. res3_3 *= alpha;
  1893. res3_4 *= alpha;
  1894. res3_5 *= alpha;
  1895. res3_6 *= alpha;
  1896. res3_7 *= alpha;
  1897. C0[0] = res0_0;
  1898. C0[1] = res0_1;
  1899. C0[2] = res0_2;
  1900. C0[3] = res0_3;
  1901. C0[4] = res0_4;
  1902. C0[5] = res0_5;
  1903. C0[6] = res0_6;
  1904. C0[7] = res0_7;
  1905. C1[0] = res1_0;
  1906. C1[1] = res1_1;
  1907. C1[2] = res1_2;
  1908. C1[3] = res1_3;
  1909. C1[4] = res1_4;
  1910. C1[5] = res1_5;
  1911. C1[6] = res1_6;
  1912. C1[7] = res1_7;
  1913. C2[0] = res2_0;
  1914. C2[1] = res2_1;
  1915. C2[2] = res2_2;
  1916. C2[3] = res2_3;
  1917. C2[4] = res2_4;
  1918. C2[5] = res2_5;
  1919. C2[6] = res2_6;
  1920. C2[7] = res2_7;
  1921. C3[0] = res3_0;
  1922. C3[1] = res3_1;
  1923. C3[2] = res3_2;
  1924. C3[3] = res3_3;
  1925. C3[4] = res3_4;
  1926. C3[5] = res3_5;
  1927. C3[6] = res3_6;
  1928. C3[7] = res3_7;
  1929. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  1930. temp = bk - off;
  1931. #ifdef LEFT
  1932. temp -= 8; // number of values in A
  1933. #else
  1934. temp -= 4; // number of values in B
  1935. #endif
  1936. ptrba += temp*8;
  1937. ptrbb += temp*4;
  1938. #endif
  1939. #ifdef LEFT
  1940. off += 8; // number of values in A
  1941. #endif
  1942. C0 = C0+8;
  1943. C1 = C1+8;
  1944. C2 = C2+8;
  1945. C3 = C3+8;
  1946. }
  1947. if ( bm & 4 )
  1948. {
  1949. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  1950. ptrbb = bb;
  1951. #else
  1952. ptrba += off*4;
  1953. ptrbb = bb + off*4;
  1954. #endif
  1955. res0_0 = 0;
  1956. res0_1 = 0;
  1957. res0_2 = 0;
  1958. res0_3 = 0;
  1959. res1_0 = 0;
  1960. res1_1 = 0;
  1961. res1_2 = 0;
  1962. res1_3 = 0;
  1963. res2_0 = 0;
  1964. res2_1 = 0;
  1965. res2_2 = 0;
  1966. res2_3 = 0;
  1967. res3_0 = 0;
  1968. res3_1 = 0;
  1969. res3_2 = 0;
  1970. res3_3 = 0;
  1971. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  1972. temp = bk-off;
  1973. #elif defined(LEFT)
  1974. temp = off+4; // number of values in A
  1975. #else
  1976. temp = off+4; // number of values in B
  1977. #endif
  1978. for (k=0; k<temp; k++)
  1979. {
  1980. b0 = ptrbb[0];
  1981. b1 = ptrbb[1];
  1982. b2 = ptrbb[2];
  1983. b3 = ptrbb[3];
  1984. a0 = ptrba[0];
  1985. res0_0 += a0*b0;
  1986. res1_0 += a0*b1;
  1987. res2_0 += a0*b2;
  1988. res3_0 += a0*b3;
  1989. a1 = ptrba[1];
  1990. res0_1 += a1*b0;
  1991. res1_1 += a1*b1;
  1992. res2_1 += a1*b2;
  1993. res3_1 += a1*b3;
  1994. a0 = ptrba[2];
  1995. res0_2 += a0*b0;
  1996. res1_2 += a0*b1;
  1997. res2_2 += a0*b2;
  1998. res3_2 += a0*b3;
  1999. a1 = ptrba[3];
  2000. res0_3 += a1*b0;
  2001. res1_3 += a1*b1;
  2002. res2_3 += a1*b2;
  2003. res3_3 += a1*b3;
  2004. ptrba = ptrba+4;
  2005. ptrbb = ptrbb+4;
  2006. }
  2007. res0_0 *= alpha;
  2008. res0_1 *= alpha;
  2009. res0_2 *= alpha;
  2010. res0_3 *= alpha;
  2011. res1_0 *= alpha;
  2012. res1_1 *= alpha;
  2013. res1_2 *= alpha;
  2014. res1_3 *= alpha;
  2015. res2_0 *= alpha;
  2016. res2_1 *= alpha;
  2017. res2_2 *= alpha;
  2018. res2_3 *= alpha;
  2019. res3_0 *= alpha;
  2020. res3_1 *= alpha;
  2021. res3_2 *= alpha;
  2022. res3_3 *= alpha;
  2023. C0[0] = res0_0;
  2024. C0[1] = res0_1;
  2025. C0[2] = res0_2;
  2026. C0[3] = res0_3;
  2027. C1[0] = res1_0;
  2028. C1[1] = res1_1;
  2029. C1[2] = res1_2;
  2030. C1[3] = res1_3;
  2031. C2[0] = res2_0;
  2032. C2[1] = res2_1;
  2033. C2[2] = res2_2;
  2034. C2[3] = res2_3;
  2035. C3[0] = res3_0;
  2036. C3[1] = res3_1;
  2037. C3[2] = res3_2;
  2038. C3[3] = res3_3;
  2039. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2040. temp = bk - off;
  2041. #ifdef LEFT
  2042. temp -= 4; // number of values in A
  2043. #else
  2044. temp -= 4; // number of values in B
  2045. #endif
  2046. ptrba += temp*4;
  2047. ptrbb += temp*4;
  2048. #endif
  2049. #ifdef LEFT
  2050. off += 4; // number of values in A
  2051. #endif
  2052. C0 = C0+4;
  2053. C1 = C1+4;
  2054. C2 = C2+4;
  2055. C3 = C3+4;
  2056. }
  2057. if ( bm & 2 )
  2058. {
  2059. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2060. ptrbb = bb;
  2061. #else
  2062. ptrba += off*2;
  2063. ptrbb = bb + off*4;
  2064. #endif
  2065. res0_0 = 0;
  2066. res0_1 = 0;
  2067. res1_0 = 0;
  2068. res1_1 = 0;
  2069. res2_0 = 0;
  2070. res2_1 = 0;
  2071. res3_0 = 0;
  2072. res3_1 = 0;
  2073. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2074. temp = bk-off;
  2075. #elif defined(LEFT)
  2076. temp = off+2; // number of values in A
  2077. #else
  2078. temp = off+4; // number of values in B
  2079. #endif
  2080. for (k=0; k<temp; k++)
  2081. {
  2082. b0 = ptrbb[0];
  2083. b1 = ptrbb[1];
  2084. b2 = ptrbb[2];
  2085. b3 = ptrbb[3];
  2086. a0 = ptrba[0];
  2087. res0_0 += a0*b0;
  2088. res1_0 += a0*b1;
  2089. res2_0 += a0*b2;
  2090. res3_0 += a0*b3;
  2091. a1 = ptrba[1];
  2092. res0_1 += a1*b0;
  2093. res1_1 += a1*b1;
  2094. res2_1 += a1*b2;
  2095. res3_1 += a1*b3;
  2096. ptrba = ptrba+2;
  2097. ptrbb = ptrbb+4;
  2098. }
  2099. res0_0 *= alpha;
  2100. res0_1 *= alpha;
  2101. res1_0 *= alpha;
  2102. res1_1 *= alpha;
  2103. res2_0 *= alpha;
  2104. res2_1 *= alpha;
  2105. res3_0 *= alpha;
  2106. res3_1 *= alpha;
  2107. C0[0] = res0_0;
  2108. C0[1] = res0_1;
  2109. C1[0] = res1_0;
  2110. C1[1] = res1_1;
  2111. C2[0] = res2_0;
  2112. C2[1] = res2_1;
  2113. C3[0] = res3_0;
  2114. C3[1] = res3_1;
  2115. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2116. temp = bk - off;
  2117. #ifdef LEFT
  2118. temp -= 2; // number of values in A
  2119. #else
  2120. temp -= 4; // number of values in B
  2121. #endif
  2122. ptrba += temp*2;
  2123. ptrbb += temp*4;
  2124. #endif
  2125. #ifdef LEFT
  2126. off += 2; // number of values in A
  2127. #endif
  2128. C0 = C0+2;
  2129. C1 = C1+2;
  2130. C2 = C2+2;
  2131. C3 = C3+2;
  2132. }
  2133. if ( bm & 1 )
  2134. {
  2135. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2136. ptrbb = bb;
  2137. #else
  2138. ptrba += off*1;
  2139. ptrbb = bb + off*4;
  2140. #endif
  2141. res0_0 = 0;
  2142. res1_0 = 0;
  2143. res2_0 = 0;
  2144. res3_0 = 0;
  2145. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2146. temp = bk-off;
  2147. #elif defined(LEFT)
  2148. temp = off+1; // number of values in A
  2149. #else
  2150. temp = off+4; // number of values in B
  2151. #endif
  2152. for (k=0; k<temp; k++)
  2153. {
  2154. b0 = ptrbb[0];
  2155. b1 = ptrbb[1];
  2156. b2 = ptrbb[2];
  2157. b3 = ptrbb[3];
  2158. a0 = ptrba[0];
  2159. res0_0 += a0*b0;
  2160. res1_0 += a0*b1;
  2161. res2_0 += a0*b2;
  2162. res3_0 += a0*b3;
  2163. ptrba = ptrba+1;
  2164. ptrbb = ptrbb+4;
  2165. }
  2166. res0_0 *= alpha;
  2167. res1_0 *= alpha;
  2168. res2_0 *= alpha;
  2169. res3_0 *= alpha;
  2170. C0[0] = res0_0;
  2171. C1[0] = res1_0;
  2172. C2[0] = res2_0;
  2173. C3[0] = res3_0;
  2174. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2175. temp = bk - off;
  2176. #ifdef LEFT
  2177. temp -= 1; // number of values in A
  2178. #else
  2179. temp -= 4; // number of values in B
  2180. #endif
  2181. ptrba += temp*1;
  2182. ptrbb += temp*4;
  2183. #endif
  2184. #ifdef LEFT
  2185. off += 1; // number of values in A
  2186. #endif
  2187. C0 = C0+1;
  2188. C1 = C1+1;
  2189. C2 = C2+1;
  2190. C3 = C3+1;
  2191. }
  2192. #if defined(TRMMKERNEL) && !defined(LEFT)
  2193. off += 4;
  2194. #endif
  2195. k = (bk<<2);
  2196. bb = bb+k;
  2197. i = (ldc<<2);
  2198. C = C+i;
  2199. }
  2200. if(bn&2)
  2201. {
  2202. C0 = C;
  2203. C1 = C0+ldc;
  2204. #if defined(TRMMKERNEL) && defined(LEFT)
  2205. off = offset;
  2206. #endif
  2207. ptrba = ba;
  2208. for (i=0; i<bm/16; i+=1)
  2209. {
  2210. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2211. ptrbb = bb;
  2212. #else
  2213. ptrba += off*16;
  2214. ptrbb = bb + off*2;
  2215. #endif
  2216. res0_0 = 0;
  2217. res0_1 = 0;
  2218. res0_2 = 0;
  2219. res0_3 = 0;
  2220. res0_4 = 0;
  2221. res0_5 = 0;
  2222. res0_6 = 0;
  2223. res0_7 = 0;
  2224. res0_8 = 0;
  2225. res0_9 = 0;
  2226. res0_10 = 0;
  2227. res0_11 = 0;
  2228. res0_12 = 0;
  2229. res0_13 = 0;
  2230. res0_14 = 0;
  2231. res0_15 = 0;
  2232. res1_0 = 0;
  2233. res1_1 = 0;
  2234. res1_2 = 0;
  2235. res1_3 = 0;
  2236. res1_4 = 0;
  2237. res1_5 = 0;
  2238. res1_6 = 0;
  2239. res1_7 = 0;
  2240. res1_8 = 0;
  2241. res1_9 = 0;
  2242. res1_10 = 0;
  2243. res1_11 = 0;
  2244. res1_12 = 0;
  2245. res1_13 = 0;
  2246. res1_14 = 0;
  2247. res1_15 = 0;
  2248. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2249. temp = bk-off;
  2250. #elif defined(LEFT)
  2251. temp = off+16; // number of values in A
  2252. #else
  2253. temp = off+2; // number of values in B
  2254. #endif
  2255. for (k=0; k<temp; k++)
  2256. {
  2257. b0 = ptrbb[0];
  2258. b1 = ptrbb[1];
  2259. a0 = ptrba[0];
  2260. res0_0 += a0*b0;
  2261. res1_0 += a0*b1;
  2262. a1 = ptrba[1];
  2263. res0_1 += a1*b0;
  2264. res1_1 += a1*b1;
  2265. a0 = ptrba[2];
  2266. res0_2 += a0*b0;
  2267. res1_2 += a0*b1;
  2268. a1 = ptrba[3];
  2269. res0_3 += a1*b0;
  2270. res1_3 += a1*b1;
  2271. a0 = ptrba[4];
  2272. res0_4 += a0*b0;
  2273. res1_4 += a0*b1;
  2274. a1 = ptrba[5];
  2275. res0_5 += a1*b0;
  2276. res1_5 += a1*b1;
  2277. a0 = ptrba[6];
  2278. res0_6 += a0*b0;
  2279. res1_6 += a0*b1;
  2280. a1 = ptrba[7];
  2281. res0_7 += a1*b0;
  2282. res1_7 += a1*b1;
  2283. a0 = ptrba[8];
  2284. res0_8 += a0*b0;
  2285. res1_8 += a0*b1;
  2286. a1 = ptrba[9];
  2287. res0_9 += a1*b0;
  2288. res1_9 += a1*b1;
  2289. a0 = ptrba[10];
  2290. res0_10 += a0*b0;
  2291. res1_10 += a0*b1;
  2292. a1 = ptrba[11];
  2293. res0_11 += a1*b0;
  2294. res1_11 += a1*b1;
  2295. a0 = ptrba[12];
  2296. res0_12 += a0*b0;
  2297. res1_12 += a0*b1;
  2298. a1 = ptrba[13];
  2299. res0_13 += a1*b0;
  2300. res1_13 += a1*b1;
  2301. a0 = ptrba[14];
  2302. res0_14 += a0*b0;
  2303. res1_14 += a0*b1;
  2304. a1 = ptrba[15];
  2305. res0_15 += a1*b0;
  2306. res1_15 += a1*b1;
  2307. ptrba = ptrba+16;
  2308. ptrbb = ptrbb+2;
  2309. }
  2310. res0_0 *= alpha;
  2311. res0_1 *= alpha;
  2312. res0_2 *= alpha;
  2313. res0_3 *= alpha;
  2314. res0_4 *= alpha;
  2315. res0_5 *= alpha;
  2316. res0_6 *= alpha;
  2317. res0_7 *= alpha;
  2318. res0_8 *= alpha;
  2319. res0_9 *= alpha;
  2320. res0_10 *= alpha;
  2321. res0_11 *= alpha;
  2322. res0_12 *= alpha;
  2323. res0_13 *= alpha;
  2324. res0_14 *= alpha;
  2325. res0_15 *= alpha;
  2326. res1_0 *= alpha;
  2327. res1_1 *= alpha;
  2328. res1_2 *= alpha;
  2329. res1_3 *= alpha;
  2330. res1_4 *= alpha;
  2331. res1_5 *= alpha;
  2332. res1_6 *= alpha;
  2333. res1_7 *= alpha;
  2334. res1_8 *= alpha;
  2335. res1_9 *= alpha;
  2336. res1_10 *= alpha;
  2337. res1_11 *= alpha;
  2338. res1_12 *= alpha;
  2339. res1_13 *= alpha;
  2340. res1_14 *= alpha;
  2341. res1_15 *= alpha;
  2342. C0[0] = res0_0;
  2343. C0[1] = res0_1;
  2344. C0[2] = res0_2;
  2345. C0[3] = res0_3;
  2346. C0[4] = res0_4;
  2347. C0[5] = res0_5;
  2348. C0[6] = res0_6;
  2349. C0[7] = res0_7;
  2350. C0[8] = res0_8;
  2351. C0[9] = res0_9;
  2352. C0[10] = res0_10;
  2353. C0[11] = res0_11;
  2354. C0[12] = res0_12;
  2355. C0[13] = res0_13;
  2356. C0[14] = res0_14;
  2357. C0[15] = res0_15;
  2358. C1[0] = res1_0;
  2359. C1[1] = res1_1;
  2360. C1[2] = res1_2;
  2361. C1[3] = res1_3;
  2362. C1[4] = res1_4;
  2363. C1[5] = res1_5;
  2364. C1[6] = res1_6;
  2365. C1[7] = res1_7;
  2366. C1[8] = res1_8;
  2367. C1[9] = res1_9;
  2368. C1[10] = res1_10;
  2369. C1[11] = res1_11;
  2370. C1[12] = res1_12;
  2371. C1[13] = res1_13;
  2372. C1[14] = res1_14;
  2373. C1[15] = res1_15;
  2374. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2375. temp = bk - off;
  2376. #ifdef LEFT
  2377. temp -= 16; // number of values in A
  2378. #else
  2379. temp -= 2; // number of values in B
  2380. #endif
  2381. ptrba += temp*16;
  2382. ptrbb += temp*2;
  2383. #endif
  2384. #ifdef LEFT
  2385. off += 16; // number of values in A
  2386. #endif
  2387. C0 = C0+16;
  2388. C1 = C1+16;
  2389. }
  2390. if ( bm & 8)
  2391. {
  2392. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2393. ptrbb = bb;
  2394. #else
  2395. ptrba += off*8;
  2396. ptrbb = bb + off*2;
  2397. #endif
  2398. res0_0 = 0;
  2399. res0_1 = 0;
  2400. res0_2 = 0;
  2401. res0_3 = 0;
  2402. res0_4 = 0;
  2403. res0_5 = 0;
  2404. res0_6 = 0;
  2405. res0_7 = 0;
  2406. res1_0 = 0;
  2407. res1_1 = 0;
  2408. res1_2 = 0;
  2409. res1_3 = 0;
  2410. res1_4 = 0;
  2411. res1_5 = 0;
  2412. res1_6 = 0;
  2413. res1_7 = 0;
  2414. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2415. temp = bk-off;
  2416. #elif defined(LEFT)
  2417. temp = off+8; // number of values in A
  2418. #else
  2419. temp = off+2; // number of values in B
  2420. #endif
  2421. for (k=0; k<temp; k++)
  2422. {
  2423. b0 = ptrbb[0];
  2424. b1 = ptrbb[1];
  2425. a0 = ptrba[0];
  2426. res0_0 += a0*b0;
  2427. res1_0 += a0*b1;
  2428. a1 = ptrba[1];
  2429. res0_1 += a1*b0;
  2430. res1_1 += a1*b1;
  2431. a0 = ptrba[2];
  2432. res0_2 += a0*b0;
  2433. res1_2 += a0*b1;
  2434. a1 = ptrba[3];
  2435. res0_3 += a1*b0;
  2436. res1_3 += a1*b1;
  2437. a0 = ptrba[4];
  2438. res0_4 += a0*b0;
  2439. res1_4 += a0*b1;
  2440. a1 = ptrba[5];
  2441. res0_5 += a1*b0;
  2442. res1_5 += a1*b1;
  2443. a0 = ptrba[6];
  2444. res0_6 += a0*b0;
  2445. res1_6 += a0*b1;
  2446. a1 = ptrba[7];
  2447. res0_7 += a1*b0;
  2448. res1_7 += a1*b1;
  2449. ptrba = ptrba+8;
  2450. ptrbb = ptrbb+2;
  2451. }
  2452. res0_0 *= alpha;
  2453. res0_1 *= alpha;
  2454. res0_2 *= alpha;
  2455. res0_3 *= alpha;
  2456. res0_4 *= alpha;
  2457. res0_5 *= alpha;
  2458. res0_6 *= alpha;
  2459. res0_7 *= alpha;
  2460. res1_0 *= alpha;
  2461. res1_1 *= alpha;
  2462. res1_2 *= alpha;
  2463. res1_3 *= alpha;
  2464. res1_4 *= alpha;
  2465. res1_5 *= alpha;
  2466. res1_6 *= alpha;
  2467. res1_7 *= alpha;
  2468. C0[0] = res0_0;
  2469. C0[1] = res0_1;
  2470. C0[2] = res0_2;
  2471. C0[3] = res0_3;
  2472. C0[4] = res0_4;
  2473. C0[5] = res0_5;
  2474. C0[6] = res0_6;
  2475. C0[7] = res0_7;
  2476. C1[0] = res1_0;
  2477. C1[1] = res1_1;
  2478. C1[2] = res1_2;
  2479. C1[3] = res1_3;
  2480. C1[4] = res1_4;
  2481. C1[5] = res1_5;
  2482. C1[6] = res1_6;
  2483. C1[7] = res1_7;
  2484. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2485. temp = bk - off;
  2486. #ifdef LEFT
  2487. temp -= 8; // number of values in A
  2488. #else
  2489. temp -= 2; // number of values in B
  2490. #endif
  2491. ptrba += temp*8;
  2492. ptrbb += temp*2;
  2493. #endif
  2494. #ifdef LEFT
  2495. off += 8; // number of values in A
  2496. #endif
  2497. C0 = C0+8;
  2498. C1 = C1+8;
  2499. }
  2500. if ( bm & 4 )
  2501. {
  2502. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2503. ptrbb = bb;
  2504. #else
  2505. ptrba += off*4;
  2506. ptrbb = bb + off*2;
  2507. #endif
  2508. res0_0 = 0;
  2509. res0_1 = 0;
  2510. res0_2 = 0;
  2511. res0_3 = 0;
  2512. res1_0 = 0;
  2513. res1_1 = 0;
  2514. res1_2 = 0;
  2515. res1_3 = 0;
  2516. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2517. temp = bk-off;
  2518. #elif defined(LEFT)
  2519. temp = off+4; // number of values in A
  2520. #else
  2521. temp = off+2; // number of values in B
  2522. #endif
  2523. for (k=0; k<temp; k++)
  2524. {
  2525. b0 = ptrbb[0];
  2526. b1 = ptrbb[1];
  2527. a0 = ptrba[0];
  2528. res0_0 += a0*b0;
  2529. res1_0 += a0*b1;
  2530. a1 = ptrba[1];
  2531. res0_1 += a1*b0;
  2532. res1_1 += a1*b1;
  2533. a0 = ptrba[2];
  2534. res0_2 += a0*b0;
  2535. res1_2 += a0*b1;
  2536. a1 = ptrba[3];
  2537. res0_3 += a1*b0;
  2538. res1_3 += a1*b1;
  2539. ptrba = ptrba+4;
  2540. ptrbb = ptrbb+2;
  2541. }
  2542. res0_0 *= alpha;
  2543. res0_1 *= alpha;
  2544. res0_2 *= alpha;
  2545. res0_3 *= alpha;
  2546. res1_0 *= alpha;
  2547. res1_1 *= alpha;
  2548. res1_2 *= alpha;
  2549. res1_3 *= alpha;
  2550. C0[0] = res0_0;
  2551. C0[1] = res0_1;
  2552. C0[2] = res0_2;
  2553. C0[3] = res0_3;
  2554. C1[0] = res1_0;
  2555. C1[1] = res1_1;
  2556. C1[2] = res1_2;
  2557. C1[3] = res1_3;
  2558. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2559. temp = bk - off;
  2560. #ifdef LEFT
  2561. temp -= 4; // number of values in A
  2562. #else
  2563. temp -= 2; // number of values in B
  2564. #endif
  2565. ptrba += temp*4;
  2566. ptrbb += temp*2;
  2567. #endif
  2568. #ifdef LEFT
  2569. off += 4; // number of values in A
  2570. #endif
  2571. C0 = C0+4;
  2572. C1 = C1+4;
  2573. }
  2574. if ( bm & 2 )
  2575. {
  2576. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2577. ptrbb = bb;
  2578. #else
  2579. ptrba += off*2;
  2580. ptrbb = bb + off*2;
  2581. #endif
  2582. res0_0 = 0;
  2583. res0_1 = 0;
  2584. res1_0 = 0;
  2585. res1_1 = 0;
  2586. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2587. temp = bk-off;
  2588. #elif defined(LEFT)
  2589. temp = off+2; // number of values in A
  2590. #else
  2591. temp = off+2; // number of values in B
  2592. #endif
  2593. for (k=0; k<temp; k++)
  2594. {
  2595. b0 = ptrbb[0];
  2596. b1 = ptrbb[1];
  2597. a0 = ptrba[0];
  2598. res0_0 += a0*b0;
  2599. res1_0 += a0*b1;
  2600. a1 = ptrba[1];
  2601. res0_1 += a1*b0;
  2602. res1_1 += a1*b1;
  2603. ptrba = ptrba+2;
  2604. ptrbb = ptrbb+2;
  2605. }
  2606. res0_0 *= alpha;
  2607. res0_1 *= alpha;
  2608. res1_0 *= alpha;
  2609. res1_1 *= alpha;
  2610. C0[0] = res0_0;
  2611. C0[1] = res0_1;
  2612. C1[0] = res1_0;
  2613. C1[1] = res1_1;
  2614. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2615. temp = bk - off;
  2616. #ifdef LEFT
  2617. temp -= 2; // number of values in A
  2618. #else
  2619. temp -= 2; // number of values in B
  2620. #endif
  2621. ptrba += temp*2;
  2622. ptrbb += temp*2;
  2623. #endif
  2624. #ifdef LEFT
  2625. off += 2; // number of values in A
  2626. #endif
  2627. C0 = C0+2;
  2628. C1 = C1+2;
  2629. }
  2630. if ( bm & 1 )
  2631. {
  2632. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2633. ptrbb = bb;
  2634. #else
  2635. ptrba += off*1;
  2636. ptrbb = bb + off*2;
  2637. #endif
  2638. res0_0 = 0;
  2639. res1_0 = 0;
  2640. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2641. temp = bk-off;
  2642. #elif defined(LEFT)
  2643. temp = off+1; // number of values in A
  2644. #else
  2645. temp = off+2; // number of values in B
  2646. #endif
  2647. for (k=0; k<temp; k++)
  2648. {
  2649. b0 = ptrbb[0];
  2650. b1 = ptrbb[1];
  2651. a0 = ptrba[0];
  2652. res0_0 += a0*b0;
  2653. res1_0 += a0*b1;
  2654. ptrba = ptrba+1;
  2655. ptrbb = ptrbb+2;
  2656. }
  2657. res0_0 *= alpha;
  2658. res1_0 *= alpha;
  2659. C0[0] = res0_0;
  2660. C1[0] = res1_0;
  2661. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2662. temp = bk - off;
  2663. #ifdef LEFT
  2664. temp -= 1; // number of values in A
  2665. #else
  2666. temp -= 2; // number of values in B
  2667. #endif
  2668. ptrba += temp*1;
  2669. ptrbb += temp*2;
  2670. #endif
  2671. #ifdef LEFT
  2672. off += 1; // number of values in A
  2673. #endif
  2674. C0 = C0+1;
  2675. C1 = C1+1;
  2676. }
  2677. #if defined(TRMMKERNEL) && !defined(LEFT)
  2678. off += 2;
  2679. #endif
  2680. k = (bk<<1);
  2681. bb = bb+k;
  2682. i = (ldc<<1);
  2683. C = C+i;
  2684. }
  2685. for (j=0; j<(bn&1); j+=1)
  2686. {
  2687. C0 = C;
  2688. #if defined(TRMMKERNEL) && defined(LEFT)
  2689. off = offset;
  2690. #endif
  2691. ptrba = ba;
  2692. for (i=0; i<bm/16; i+=1)
  2693. {
  2694. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2695. ptrbb = bb;
  2696. #else
  2697. ptrba += off*16;
  2698. ptrbb = bb + off*1;
  2699. #endif
  2700. res0_0 = 0;
  2701. res0_1 = 0;
  2702. res0_2 = 0;
  2703. res0_3 = 0;
  2704. res0_4 = 0;
  2705. res0_5 = 0;
  2706. res0_6 = 0;
  2707. res0_7 = 0;
  2708. res0_8 = 0;
  2709. res0_9 = 0;
  2710. res0_10 = 0;
  2711. res0_11 = 0;
  2712. res0_12 = 0;
  2713. res0_13 = 0;
  2714. res0_14 = 0;
  2715. res0_15 = 0;
  2716. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2717. temp = bk-off;
  2718. #elif defined(LEFT)
  2719. temp = off+16; // number of values in A
  2720. #else
  2721. temp = off+1; // number of values in B
  2722. #endif
  2723. for (k=0; k<temp; k++)
  2724. {
  2725. b0 = ptrbb[0];
  2726. a0 = ptrba[0];
  2727. res0_0 += a0*b0;
  2728. a1 = ptrba[1];
  2729. res0_1 += a1*b0;
  2730. a0 = ptrba[2];
  2731. res0_2 += a0*b0;
  2732. a1 = ptrba[3];
  2733. res0_3 += a1*b0;
  2734. a0 = ptrba[4];
  2735. res0_4 += a0*b0;
  2736. a1 = ptrba[5];
  2737. res0_5 += a1*b0;
  2738. a0 = ptrba[6];
  2739. res0_6 += a0*b0;
  2740. a1 = ptrba[7];
  2741. res0_7 += a1*b0;
  2742. a0 = ptrba[8];
  2743. res0_8 += a0*b0;
  2744. a1 = ptrba[9];
  2745. res0_9 += a1*b0;
  2746. a0 = ptrba[10];
  2747. res0_10 += a0*b0;
  2748. a1 = ptrba[11];
  2749. res0_11 += a1*b0;
  2750. a0 = ptrba[12];
  2751. res0_12 += a0*b0;
  2752. a1 = ptrba[13];
  2753. res0_13 += a1*b0;
  2754. a0 = ptrba[14];
  2755. res0_14 += a0*b0;
  2756. a1 = ptrba[15];
  2757. res0_15 += a1*b0;
  2758. ptrba = ptrba+16;
  2759. ptrbb = ptrbb+1;
  2760. }
  2761. res0_0 *= alpha;
  2762. res0_1 *= alpha;
  2763. res0_2 *= alpha;
  2764. res0_3 *= alpha;
  2765. res0_4 *= alpha;
  2766. res0_5 *= alpha;
  2767. res0_6 *= alpha;
  2768. res0_7 *= alpha;
  2769. res0_8 *= alpha;
  2770. res0_9 *= alpha;
  2771. res0_10 *= alpha;
  2772. res0_11 *= alpha;
  2773. res0_12 *= alpha;
  2774. res0_13 *= alpha;
  2775. res0_14 *= alpha;
  2776. res0_15 *= alpha;
  2777. C0[0] = res0_0;
  2778. C0[1] = res0_1;
  2779. C0[2] = res0_2;
  2780. C0[3] = res0_3;
  2781. C0[4] = res0_4;
  2782. C0[5] = res0_5;
  2783. C0[6] = res0_6;
  2784. C0[7] = res0_7;
  2785. C0[8] = res0_8;
  2786. C0[9] = res0_9;
  2787. C0[10] = res0_10;
  2788. C0[11] = res0_11;
  2789. C0[12] = res0_12;
  2790. C0[13] = res0_13;
  2791. C0[14] = res0_14;
  2792. C0[15] = res0_15;
  2793. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2794. temp = bk - off;
  2795. #ifdef LEFT
  2796. temp -= 16; // number of values in A
  2797. #else
  2798. temp -= 1; // number of values in B
  2799. #endif
  2800. ptrba += temp*16;
  2801. ptrbb += temp*1;
  2802. #endif
  2803. #ifdef LEFT
  2804. off += 16; // number of values in A
  2805. #endif
  2806. C0 = C0+16;
  2807. }
  2808. if ( bm & 8 )
  2809. {
  2810. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2811. ptrbb = bb;
  2812. #else
  2813. ptrba += off*8;
  2814. ptrbb = bb + off*1;
  2815. #endif
  2816. res0_0 = 0;
  2817. res0_1 = 0;
  2818. res0_2 = 0;
  2819. res0_3 = 0;
  2820. res0_4 = 0;
  2821. res0_5 = 0;
  2822. res0_6 = 0;
  2823. res0_7 = 0;
  2824. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2825. temp = bk-off;
  2826. #elif defined(LEFT)
  2827. temp = off+8; // number of values in A
  2828. #else
  2829. temp = off+1; // number of values in B
  2830. #endif
  2831. for (k=0; k<temp; k++)
  2832. {
  2833. b0 = ptrbb[0];
  2834. a0 = ptrba[0];
  2835. res0_0 += a0*b0;
  2836. a1 = ptrba[1];
  2837. res0_1 += a1*b0;
  2838. a0 = ptrba[2];
  2839. res0_2 += a0*b0;
  2840. a1 = ptrba[3];
  2841. res0_3 += a1*b0;
  2842. a0 = ptrba[4];
  2843. res0_4 += a0*b0;
  2844. a1 = ptrba[5];
  2845. res0_5 += a1*b0;
  2846. a0 = ptrba[6];
  2847. res0_6 += a0*b0;
  2848. a1 = ptrba[7];
  2849. res0_7 += a1*b0;
  2850. ptrba = ptrba+8;
  2851. ptrbb = ptrbb+1;
  2852. }
  2853. res0_0 *= alpha;
  2854. res0_1 *= alpha;
  2855. res0_2 *= alpha;
  2856. res0_3 *= alpha;
  2857. res0_4 *= alpha;
  2858. res0_5 *= alpha;
  2859. res0_6 *= alpha;
  2860. res0_7 *= alpha;
  2861. C0[0] = res0_0;
  2862. C0[1] = res0_1;
  2863. C0[2] = res0_2;
  2864. C0[3] = res0_3;
  2865. C0[4] = res0_4;
  2866. C0[5] = res0_5;
  2867. C0[6] = res0_6;
  2868. C0[7] = res0_7;
  2869. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2870. temp = bk - off;
  2871. #ifdef LEFT
  2872. temp -= 8; // number of values in A
  2873. #else
  2874. temp -= 1; // number of values in B
  2875. #endif
  2876. ptrba += temp*8;
  2877. ptrbb += temp*1;
  2878. #endif
  2879. #ifdef LEFT
  2880. off += 8; // number of values in A
  2881. #endif
  2882. C0 = C0+8;
  2883. }
  2884. if ( bm & 4 )
  2885. {
  2886. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2887. ptrbb = bb;
  2888. #else
  2889. ptrba += off*4;
  2890. ptrbb = bb + off*1;
  2891. #endif
  2892. res0_0 = 0;
  2893. res0_1 = 0;
  2894. res0_2 = 0;
  2895. res0_3 = 0;
  2896. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2897. temp = bk-off;
  2898. #elif defined(LEFT)
  2899. temp = off+4; // number of values in A
  2900. #else
  2901. temp = off+1; // number of values in B
  2902. #endif
  2903. for (k=0; k<temp; k++)
  2904. {
  2905. b0 = ptrbb[0];
  2906. a0 = ptrba[0];
  2907. res0_0 += a0*b0;
  2908. a1 = ptrba[1];
  2909. res0_1 += a1*b0;
  2910. a0 = ptrba[2];
  2911. res0_2 += a0*b0;
  2912. a1 = ptrba[3];
  2913. res0_3 += a1*b0;
  2914. ptrba = ptrba+4;
  2915. ptrbb = ptrbb+1;
  2916. }
  2917. res0_0 *= alpha;
  2918. res0_1 *= alpha;
  2919. res0_2 *= alpha;
  2920. res0_3 *= alpha;
  2921. C0[0] = res0_0;
  2922. C0[1] = res0_1;
  2923. C0[2] = res0_2;
  2924. C0[3] = res0_3;
  2925. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2926. temp = bk - off;
  2927. #ifdef LEFT
  2928. temp -= 4; // number of values in A
  2929. #else
  2930. temp -= 1; // number of values in B
  2931. #endif
  2932. ptrba += temp*4;
  2933. ptrbb += temp*1;
  2934. #endif
  2935. #ifdef LEFT
  2936. off += 4; // number of values in A
  2937. #endif
  2938. C0 = C0+4;
  2939. }
  2940. if ( bm & 2 )
  2941. {
  2942. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2943. ptrbb = bb;
  2944. #else
  2945. ptrba += off*2;
  2946. ptrbb = bb + off*1;
  2947. #endif
  2948. res0_0 = 0;
  2949. res0_1 = 0;
  2950. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2951. temp = bk-off;
  2952. #elif defined(LEFT)
  2953. temp = off+2; // number of values in A
  2954. #else
  2955. temp = off+1; // number of values in B
  2956. #endif
  2957. for (k=0; k<temp; k++)
  2958. {
  2959. b0 = ptrbb[0];
  2960. a0 = ptrba[0];
  2961. res0_0 += a0*b0;
  2962. a1 = ptrba[1];
  2963. res0_1 += a1*b0;
  2964. ptrba = ptrba+2;
  2965. ptrbb = ptrbb+1;
  2966. }
  2967. res0_0 *= alpha;
  2968. res0_1 *= alpha;
  2969. C0[0] = res0_0;
  2970. C0[1] = res0_1;
  2971. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2972. temp = bk - off;
  2973. #ifdef LEFT
  2974. temp -= 2; // number of values in A
  2975. #else
  2976. temp -= 1; // number of values in B
  2977. #endif
  2978. ptrba += temp*2;
  2979. ptrbb += temp*1;
  2980. #endif
  2981. #ifdef LEFT
  2982. off += 2; // number of values in A
  2983. #endif
  2984. C0 = C0+2;
  2985. }
  2986. if ( bm & 1 )
  2987. {
  2988. #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  2989. ptrbb = bb;
  2990. #else
  2991. ptrba += off*1;
  2992. ptrbb = bb + off*1;
  2993. #endif
  2994. res0_0 = 0;
  2995. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2996. temp = bk-off;
  2997. #elif defined(LEFT)
  2998. temp = off+1; // number of values in A
  2999. #else
  3000. temp = off+1; // number of values in B
  3001. #endif
  3002. for (k=0; k<temp; k++)
  3003. {
  3004. b0 = ptrbb[0];
  3005. a0 = ptrba[0];
  3006. res0_0 += a0*b0;
  3007. ptrba = ptrba+1;
  3008. ptrbb = ptrbb+1;
  3009. }
  3010. res0_0 *= alpha;
  3011. C0[0] = res0_0;
  3012. #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
  3013. temp = bk - off;
  3014. #ifdef LEFT
  3015. temp -= 1; // number of values in A
  3016. #else
  3017. temp -= 1; // number of values in B
  3018. #endif
  3019. ptrba += temp*1;
  3020. ptrbb += temp*1;
  3021. #endif
  3022. #ifdef LEFT
  3023. off += 1; // number of values in A
  3024. #endif
  3025. C0 = C0+1;
  3026. }
  3027. #if defined(TRMMKERNEL) && !defined(LEFT)
  3028. off += 1;
  3029. #endif
  3030. k = (bk<<0);
  3031. bb = bb+k;
  3032. C = C+ldc;
  3033. }
  3034. return 0;
  3035. }