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.

zgemm_kernel_power6.S 54 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667266826692670267126722673267426752676267726782679268026812682268326842685268626872688268926902691269226932694269526962697269826992700270127022703270427052706270727082709271027112712271327142715271627172718271927202721272227232724272527262727272827292730273127322733273427352736273727382739274027412742274327442745274627472748274927502751275227532754275527562757275827592760276127622763276427652766276727682769277027712772277327742775277627772778277927802781278227832784278527862787278827892790279127922793279427952796279727982799280028012802280328042805280628072808280928102811281228132814281528162817281828192820282128222823282428252826282728282829283028312832283328342835283628372838283928402841284228432844284528462847284828492850285128522853285428552856285728582859286028612862286328642865286628672868286928702871287228732874287528762877287828792880288128822883288428852886288728882889289028912892289328942895289628972898289929002901290229032904
  1. /*********************************************************************/
  2. /* Copyright 2009, 2010 The University of Texas at Austin. */
  3. /* All rights reserved. */
  4. /* */
  5. /* Redistribution and use in source and binary forms, with or */
  6. /* without modification, are permitted provided that the following */
  7. /* conditions are met: */
  8. /* */
  9. /* 1. Redistributions of source code must retain the above */
  10. /* copyright notice, this list of conditions and the following */
  11. /* disclaimer. */
  12. /* */
  13. /* 2. Redistributions in binary form must reproduce the above */
  14. /* copyright notice, this list of conditions and the following */
  15. /* disclaimer in the documentation and/or other materials */
  16. /* provided with the distribution. */
  17. /* */
  18. /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
  19. /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
  20. /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
  21. /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
  22. /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
  23. /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
  24. /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
  25. /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
  26. /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
  27. /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
  28. /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
  29. /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
  30. /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
  31. /* POSSIBILITY OF SUCH DAMAGE. */
  32. /* */
  33. /* The views and conclusions contained in the software and */
  34. /* documentation are those of the authors and should not be */
  35. /* interpreted as representing official policies, either expressed */
  36. /* or implied, of The University of Texas at Austin. */
  37. /*********************************************************************/
  38. #define ASSEMBLER
  39. #include "common.h"
  40. #ifndef __64BIT__
  41. #define LOAD lwz
  42. #else
  43. #define LOAD ld
  44. #endif
  45. #ifdef __64BIT__
  46. #define STACKSIZE 320
  47. #define ALPHA_R 296(SP)
  48. #define ALPHA_I 304(SP)
  49. #define FZERO 312(SP)
  50. #else
  51. #define STACKSIZE 256
  52. #define ALPHA_R 224(SP)
  53. #define ALPHA_I 232(SP)
  54. #define FZERO 240(SP)
  55. #endif
  56. #define M r3
  57. #define N r4
  58. #define K r5
  59. #if defined(linux) || defined(__FreeBSD__)
  60. #ifndef __64BIT__
  61. #define A r6
  62. #define B r7
  63. #define C r8
  64. #define LDC r9
  65. #define OFFSET r10
  66. #else
  67. #define A r8
  68. #define B r9
  69. #define C r10
  70. #define LDC r6
  71. #define OFFSET r7
  72. #endif
  73. #endif
  74. #if defined(_AIX) || defined(__APPLE__)
  75. #if !defined(__64BIT__) && defined(DOUBLE)
  76. #define A r10
  77. #define B r6
  78. #define C r7
  79. #define LDC r8
  80. #define OFFSET r9
  81. #else
  82. #define A r8
  83. #define B r9
  84. #define C r10
  85. #define LDC r6
  86. #define OFFSET r7
  87. #endif
  88. #endif
  89. #define TEMP r19
  90. #define KK r20
  91. #define BB r21
  92. #define I r22
  93. #define J r23
  94. #define AO r24
  95. #define BO r25
  96. #define CO1 r26
  97. #define CO2 r27
  98. #define CO3 r28
  99. #define CO4 r29
  100. #define PREA r30
  101. #define PREC r31
  102. #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
  103. #define FMA1 FMADD
  104. #define FMA2 FMADD
  105. #define FMA3 FNMSUB
  106. #define FMA4 FMADD
  107. #elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
  108. #define FMA1 FMADD
  109. #define FMA2 FNMSUB
  110. #define FMA3 FMADD
  111. #define FMA4 FMADD
  112. #elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
  113. #define FMA1 FMADD
  114. #define FMA2 FMADD
  115. #define FMA3 FMADD
  116. #define FMA4 FNMSUB
  117. #else
  118. #define FMA1 FMADD
  119. #define FMA2 FNMSUB
  120. #define FMA3 FNMSUB
  121. #define FMA4 FNMSUB
  122. #endif
  123. #ifndef NEEDPARAM
  124. PROLOGUE
  125. PROFCODE
  126. addi SP, SP, -STACKSIZE
  127. li r0, 0
  128. stfd f14, 0(SP)
  129. stfd f15, 8(SP)
  130. stfd f16, 16(SP)
  131. stfd f17, 24(SP)
  132. stfd f18, 32(SP)
  133. stfd f19, 40(SP)
  134. stfd f20, 48(SP)
  135. stfd f21, 56(SP)
  136. stfd f22, 64(SP)
  137. stfd f23, 72(SP)
  138. stfd f24, 80(SP)
  139. stfd f25, 88(SP)
  140. stfd f26, 96(SP)
  141. stfd f27, 104(SP)
  142. stfd f28, 112(SP)
  143. stfd f29, 120(SP)
  144. stfd f30, 128(SP)
  145. stfd f31, 136(SP)
  146. #ifdef __64BIT__
  147. std r31, 144(SP)
  148. std r30, 152(SP)
  149. std r29, 160(SP)
  150. std r28, 168(SP)
  151. std r27, 176(SP)
  152. std r26, 184(SP)
  153. std r25, 192(SP)
  154. std r24, 200(SP)
  155. std r23, 208(SP)
  156. std r22, 216(SP)
  157. std r21, 224(SP)
  158. #ifdef TRMMKERNEL
  159. std r20, 232(SP)
  160. std r19, 240(SP)
  161. #endif
  162. #else
  163. stw r31, 144(SP)
  164. stw r30, 148(SP)
  165. stw r29, 152(SP)
  166. stw r28, 156(SP)
  167. stw r27, 160(SP)
  168. stw r26, 164(SP)
  169. stw r25, 168(SP)
  170. stw r24, 172(SP)
  171. stw r23, 176(SP)
  172. stw r22, 180(SP)
  173. stw r21, 184(SP)
  174. #ifdef TRMMKERNEL
  175. stw r20, 188(SP)
  176. stw r19, 192(SP)
  177. #endif
  178. #endif
  179. stfd f1, ALPHA_R
  180. stfd f2, ALPHA_I
  181. stw r0, FZERO
  182. #if defined(linux) || defined(__FreeBSD__)
  183. #ifdef __64BIT__
  184. ld LDC, FRAMESLOT(0) + STACKSIZE(SP)
  185. #endif
  186. #endif
  187. #if defined(_AIX) || defined(__APPLE__)
  188. #ifdef __64BIT__
  189. ld LDC, FRAMESLOT(0) + STACKSIZE(SP)
  190. #else
  191. #ifdef DOUBLE
  192. lwz B, FRAMESLOT(0) + STACKSIZE(SP)
  193. lwz C, FRAMESLOT(1) + STACKSIZE(SP)
  194. lwz LDC, FRAMESLOT(2) + STACKSIZE(SP)
  195. #else
  196. lwz LDC, FRAMESLOT(0) + STACKSIZE(SP)
  197. #endif
  198. #endif
  199. #endif
  200. #ifdef TRMMKERNEL
  201. #if (defined(linux) || defined(__FreeBSD__)) && defined(__64BIT__)
  202. ld OFFSET, FRAMESLOT(1) + STACKSIZE(SP)
  203. #endif
  204. #if defined(_AIX) || defined(__APPLE__)
  205. #ifdef __64BIT__
  206. ld OFFSET, FRAMESLOT(1) + STACKSIZE(SP)
  207. #else
  208. #ifdef DOUBLE
  209. lwz OFFSET, FRAMESLOT(3) + STACKSIZE(SP)
  210. #else
  211. lwz OFFSET, FRAMESLOT(1) + STACKSIZE(SP)
  212. #endif
  213. #endif
  214. #endif
  215. #if defined(TRMMKERNEL) && !defined(LEFT)
  216. neg KK, OFFSET
  217. #endif
  218. #endif
  219. slwi LDC, LDC, ZBASE_SHIFT
  220. li PREA, (16 * 3) * SIZE
  221. li PREC, 3 * SIZE
  222. cmpwi cr0, M, 0
  223. ble LL(999)
  224. cmpwi cr0, N, 0
  225. ble LL(999)
  226. cmpwi cr0, K, 0
  227. ble LL(999)
  228. srawi. J, N, 2
  229. ble LL(30)
  230. .align 4
  231. LL(10):
  232. mr CO1, C
  233. add CO2, C, LDC
  234. add CO3, CO2, LDC
  235. add CO4, CO3, LDC
  236. add C, CO4, LDC
  237. #if defined(TRMMKERNEL) && defined(LEFT)
  238. mr KK, OFFSET
  239. #endif
  240. slwi BB, K, ZBASE_SHIFT + 2
  241. mr AO, A
  242. lfs f0, FZERO
  243. fmr f1, f0
  244. fmr f2, f0
  245. fmr f3, f0
  246. fmr f4, f0
  247. fmr f5, f0
  248. fmr f6, f0
  249. fmr f7, f0
  250. fmr f8, f0
  251. fmr f9, f0
  252. fmr f10, f0
  253. fmr f11, f0
  254. fmr f12, f0
  255. fmr f13, f0
  256. fmr f14, f0
  257. fmr f15, f0
  258. srawi. I, M, 1
  259. ble LL(20)
  260. .align 4
  261. LL(11):
  262. #ifndef TRMMKERNEL
  263. LFD f16, 0 * SIZE(AO)
  264. LFD f20, 0 * SIZE(B)
  265. LFD f17, 1 * SIZE(AO)
  266. LFD f21, 1 * SIZE(B)
  267. LFD f18, 2 * SIZE(AO)
  268. LFD f22, 2 * SIZE(B)
  269. LFD f19, 3 * SIZE(AO)
  270. LFD f23, 3 * SIZE(B)
  271. LFD f24, 4 * SIZE(B)
  272. LFD f25, 5 * SIZE(B)
  273. LFD f26, 6 * SIZE(B)
  274. LFD f27, 7 * SIZE(B)
  275. dcbtst CO1, PREC
  276. dcbtst CO2, PREC
  277. dcbtst CO3, PREC
  278. dcbtst CO4, PREC
  279. srawi. r0, K, 3
  280. mr BO, B
  281. mtspr CTR, r0
  282. ble LL(15)
  283. #else
  284. #if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
  285. (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
  286. LFD f16, 0 * SIZE(AO)
  287. LFD f20, 0 * SIZE(B)
  288. LFD f17, 1 * SIZE(AO)
  289. LFD f21, 1 * SIZE(B)
  290. LFD f18, 2 * SIZE(AO)
  291. LFD f22, 2 * SIZE(B)
  292. LFD f19, 3 * SIZE(AO)
  293. LFD f23, 3 * SIZE(B)
  294. LFD f24, 4 * SIZE(B)
  295. LFD f25, 5 * SIZE(B)
  296. LFD f26, 6 * SIZE(B)
  297. LFD f27, 7 * SIZE(B)
  298. mr BO, B
  299. #else
  300. slwi r0, KK, 1 + ZBASE_SHIFT
  301. slwi TEMP, KK, 2 + ZBASE_SHIFT
  302. add AO, AO, r0
  303. add BO, B, TEMP
  304. LFD f16, 0 * SIZE(AO)
  305. LFD f20, 0 * SIZE(BO)
  306. LFD f17, 1 * SIZE(AO)
  307. LFD f21, 1 * SIZE(BO)
  308. LFD f18, 2 * SIZE(AO)
  309. LFD f22, 2 * SIZE(BO)
  310. LFD f19, 3 * SIZE(AO)
  311. LFD f23, 3 * SIZE(BO)
  312. LFD f24, 4 * SIZE(BO)
  313. LFD f25, 5 * SIZE(BO)
  314. LFD f26, 6 * SIZE(BO)
  315. LFD f27, 7 * SIZE(BO)
  316. #endif
  317. dcbtst CO1, PREC
  318. dcbtst CO2, PREC
  319. dcbtst CO3, PREC
  320. dcbtst CO4, PREC
  321. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  322. sub TEMP, K, KK
  323. #elif defined(LEFT)
  324. addi TEMP, KK, 2
  325. #else
  326. addi TEMP, KK, 4
  327. #endif
  328. srawi. TEMP, TEMP, 3
  329. mtspr CTR, TEMP
  330. ble LL(15)
  331. #endif
  332. .align 4
  333. LL(12):
  334. dcbt AO, PREA
  335. FMA1 f0, f16, f20, f0
  336. nop
  337. FMA1 f2, f18, f20, f2
  338. FMA2 f1, f16, f21, f1
  339. FMA2 f3, f18, f21, f3
  340. LFD f28, 4 * SIZE(AO)
  341. LFD f29, 5 * SIZE(AO)
  342. LFD f30, 6 * SIZE(AO)
  343. LFD f31, 7 * SIZE(AO)
  344. FMA1 f4, f16, f22, f4
  345. FMA1 f6, f18, f22, f6
  346. FMA2 f5, f16, f23, f5
  347. FMA2 f7, f18, f23, f7
  348. FMA1 f8, f16, f24, f8
  349. FMA1 f10, f18, f24, f10
  350. FMA2 f9, f16, f25, f9
  351. FMA2 f11, f18, f25, f11
  352. FMA1 f12, f16, f26, f12
  353. FMA1 f14, f18, f26, f14
  354. FMA2 f13, f16, f27, f13
  355. FMA2 f15, f18, f27, f15
  356. FMA4 f1, f17, f20, f1
  357. FMA4 f3, f19, f20, f3
  358. FMA3 f0, f17, f21, f0
  359. FMA3 f2, f19, f21, f2
  360. FMA4 f5, f17, f22, f5
  361. FMA4 f7, f19, f22, f7
  362. FMA3 f4, f17, f23, f4
  363. FMA3 f6, f19, f23, f6
  364. LFD f20, 8 * SIZE(BO)
  365. LFD f21, 9 * SIZE(BO)
  366. LFD f22, 10 * SIZE(BO)
  367. LFD f23, 11 * SIZE(BO)
  368. FMA4 f9, f17, f24, f9
  369. FMA4 f11, f19, f24, f11
  370. FMA3 f8, f17, f25, f8
  371. FMA3 f10, f19, f25, f10
  372. FMA4 f13, f17, f26, f13
  373. FMA4 f15, f19, f26, f15
  374. FMA3 f12, f17, f27, f12
  375. FMA3 f14, f19, f27, f14
  376. LFD f24, 12 * SIZE(BO)
  377. LFD f25, 13 * SIZE(BO)
  378. LFD f26, 14 * SIZE(BO)
  379. LFD f27, 15 * SIZE(BO)
  380. FMA1 f0, f28, f20, f0
  381. FMA1 f2, f30, f20, f2
  382. FMA2 f1, f28, f21, f1
  383. FMA2 f3, f30, f21, f3
  384. LFD f16, 8 * SIZE(AO)
  385. LFD f17, 9 * SIZE(AO)
  386. LFD f18, 10 * SIZE(AO)
  387. LFD f19, 11 * SIZE(AO)
  388. FMA1 f4, f28, f22, f4
  389. FMA1 f6, f30, f22, f6
  390. FMA2 f5, f28, f23, f5
  391. FMA2 f7, f30, f23, f7
  392. FMA1 f8, f28, f24, f8
  393. FMA1 f10, f30, f24, f10
  394. FMA2 f9, f28, f25, f9
  395. FMA2 f11, f30, f25, f11
  396. FMA1 f12, f28, f26, f12
  397. FMA1 f14, f30, f26, f14
  398. FMA2 f13, f28, f27, f13
  399. FMA2 f15, f30, f27, f15
  400. FMA4 f1, f29, f20, f1
  401. FMA4 f3, f31, f20, f3
  402. FMA3 f0, f29, f21, f0
  403. FMA3 f2, f31, f21, f2
  404. FMA4 f5, f29, f22, f5
  405. FMA4 f7, f31, f22, f7
  406. FMA3 f4, f29, f23, f4
  407. FMA3 f6, f31, f23, f6
  408. LFD f20, 16 * SIZE(BO)
  409. LFD f21, 17 * SIZE(BO)
  410. LFD f22, 18 * SIZE(BO)
  411. LFD f23, 19 * SIZE(BO)
  412. FMA4 f9, f29, f24, f9
  413. FMA4 f11, f31, f24, f11
  414. FMA3 f8, f29, f25, f8
  415. FMA3 f10, f31, f25, f10
  416. FMA4 f13, f29, f26, f13
  417. FMA4 f15, f31, f26, f15
  418. FMA3 f12, f29, f27, f12
  419. FMA3 f14, f31, f27, f14
  420. LFD f24, 20 * SIZE(BO)
  421. LFD f25, 21 * SIZE(BO)
  422. LFD f26, 22 * SIZE(BO)
  423. LFD f27, 23 * SIZE(BO)
  424. FMA1 f0, f16, f20, f0
  425. FMA1 f2, f18, f20, f2
  426. FMA2 f1, f16, f21, f1
  427. FMA2 f3, f18, f21, f3
  428. LFD f28, 12 * SIZE(AO)
  429. LFD f29, 13 * SIZE(AO)
  430. LFD f30, 14 * SIZE(AO)
  431. LFD f31, 15 * SIZE(AO)
  432. FMA1 f4, f16, f22, f4
  433. FMA1 f6, f18, f22, f6
  434. FMA2 f5, f16, f23, f5
  435. FMA2 f7, f18, f23, f7
  436. FMA1 f8, f16, f24, f8
  437. FMA1 f10, f18, f24, f10
  438. FMA2 f9, f16, f25, f9
  439. FMA2 f11, f18, f25, f11
  440. FMA1 f12, f16, f26, f12
  441. FMA1 f14, f18, f26, f14
  442. FMA2 f13, f16, f27, f13
  443. FMA2 f15, f18, f27, f15
  444. FMA4 f1, f17, f20, f1
  445. FMA4 f3, f19, f20, f3
  446. FMA3 f0, f17, f21, f0
  447. FMA3 f2, f19, f21, f2
  448. FMA4 f5, f17, f22, f5
  449. FMA4 f7, f19, f22, f7
  450. FMA3 f4, f17, f23, f4
  451. FMA3 f6, f19, f23, f6
  452. LFD f20, 24 * SIZE(BO)
  453. LFD f21, 25 * SIZE(BO)
  454. LFD f22, 26 * SIZE(BO)
  455. LFD f23, 27 * SIZE(BO)
  456. FMA4 f9, f17, f24, f9
  457. FMA4 f11, f19, f24, f11
  458. FMA3 f8, f17, f25, f8
  459. FMA3 f10, f19, f25, f10
  460. FMA4 f13, f17, f26, f13
  461. FMA4 f15, f19, f26, f15
  462. FMA3 f12, f17, f27, f12
  463. FMA3 f14, f19, f27, f14
  464. LFD f24, 28 * SIZE(BO)
  465. LFD f25, 29 * SIZE(BO)
  466. LFD f26, 30 * SIZE(BO)
  467. LFD f27, 31 * SIZE(BO)
  468. FMA1 f0, f28, f20, f0
  469. FMA1 f2, f30, f20, f2
  470. FMA2 f1, f28, f21, f1
  471. FMA2 f3, f30, f21, f3
  472. LFD f16, 16 * SIZE(AO)
  473. LFD f17, 17 * SIZE(AO)
  474. LFD f18, 18 * SIZE(AO)
  475. LFD f19, 19 * SIZE(AO)
  476. FMA1 f4, f28, f22, f4
  477. FMA1 f6, f30, f22, f6
  478. FMA2 f5, f28, f23, f5
  479. FMA2 f7, f30, f23, f7
  480. FMA1 f8, f28, f24, f8
  481. FMA1 f10, f30, f24, f10
  482. FMA2 f9, f28, f25, f9
  483. FMA2 f11, f30, f25, f11
  484. FMA1 f12, f28, f26, f12
  485. FMA1 f14, f30, f26, f14
  486. FMA2 f13, f28, f27, f13
  487. FMA2 f15, f30, f27, f15
  488. FMA4 f1, f29, f20, f1
  489. FMA4 f3, f31, f20, f3
  490. FMA3 f0, f29, f21, f0
  491. FMA3 f2, f31, f21, f2
  492. FMA4 f5, f29, f22, f5
  493. FMA4 f7, f31, f22, f7
  494. FMA3 f4, f29, f23, f4
  495. FMA3 f6, f31, f23, f6
  496. LFD f20, 32 * SIZE(BO)
  497. LFD f21, 33 * SIZE(BO)
  498. LFD f22, 34 * SIZE(BO)
  499. LFD f23, 35 * SIZE(BO)
  500. FMA4 f9, f29, f24, f9
  501. FMA4 f11, f31, f24, f11
  502. FMA3 f8, f29, f25, f8
  503. FMA3 f10, f31, f25, f10
  504. FMA4 f13, f29, f26, f13
  505. FMA4 f15, f31, f26, f15
  506. FMA3 f12, f29, f27, f12
  507. FMA3 f14, f31, f27, f14
  508. LFD f24, 36 * SIZE(BO)
  509. LFD f25, 37 * SIZE(BO)
  510. LFD f26, 38 * SIZE(BO)
  511. LFD f27, 39 * SIZE(BO)
  512. FMA1 f0, f16, f20, f0
  513. FMA1 f2, f18, f20, f2
  514. FMA2 f1, f16, f21, f1
  515. FMA2 f3, f18, f21, f3
  516. LFD f28, 20 * SIZE(AO)
  517. LFD f29, 21 * SIZE(AO)
  518. LFD f30, 22 * SIZE(AO)
  519. LFD f31, 23 * SIZE(AO)
  520. FMA1 f4, f16, f22, f4
  521. FMA1 f6, f18, f22, f6
  522. FMA2 f5, f16, f23, f5
  523. FMA2 f7, f18, f23, f7
  524. FMA1 f8, f16, f24, f8
  525. FMA1 f10, f18, f24, f10
  526. FMA2 f9, f16, f25, f9
  527. FMA2 f11, f18, f25, f11
  528. FMA1 f12, f16, f26, f12
  529. FMA1 f14, f18, f26, f14
  530. FMA2 f13, f16, f27, f13
  531. FMA2 f15, f18, f27, f15
  532. FMA4 f1, f17, f20, f1
  533. FMA4 f3, f19, f20, f3
  534. FMA3 f0, f17, f21, f0
  535. FMA3 f2, f19, f21, f2
  536. FMA4 f5, f17, f22, f5
  537. FMA4 f7, f19, f22, f7
  538. FMA3 f4, f17, f23, f4
  539. FMA3 f6, f19, f23, f6
  540. LFD f20, 40 * SIZE(BO)
  541. LFD f21, 41 * SIZE(BO)
  542. LFD f22, 42 * SIZE(BO)
  543. LFD f23, 43 * SIZE(BO)
  544. FMA4 f9, f17, f24, f9
  545. FMA4 f11, f19, f24, f11
  546. FMA3 f8, f17, f25, f8
  547. FMA3 f10, f19, f25, f10
  548. FMA4 f13, f17, f26, f13
  549. FMA4 f15, f19, f26, f15
  550. FMA3 f12, f17, f27, f12
  551. FMA3 f14, f19, f27, f14
  552. LFD f24, 44 * SIZE(BO)
  553. LFD f25, 45 * SIZE(BO)
  554. LFD f26, 46 * SIZE(BO)
  555. LFD f27, 47 * SIZE(BO)
  556. FMA1 f0, f28, f20, f0
  557. FMA1 f2, f30, f20, f2
  558. FMA2 f1, f28, f21, f1
  559. FMA2 f3, f30, f21, f3
  560. LFD f16, 24 * SIZE(AO)
  561. LFD f17, 25 * SIZE(AO)
  562. LFD f18, 26 * SIZE(AO)
  563. LFD f19, 27 * SIZE(AO)
  564. FMA1 f4, f28, f22, f4
  565. FMA1 f6, f30, f22, f6
  566. FMA2 f5, f28, f23, f5
  567. FMA2 f7, f30, f23, f7
  568. FMA1 f8, f28, f24, f8
  569. FMA1 f10, f30, f24, f10
  570. FMA2 f9, f28, f25, f9
  571. FMA2 f11, f30, f25, f11
  572. FMA1 f12, f28, f26, f12
  573. FMA1 f14, f30, f26, f14
  574. FMA2 f13, f28, f27, f13
  575. FMA2 f15, f30, f27, f15
  576. FMA4 f1, f29, f20, f1
  577. FMA4 f3, f31, f20, f3
  578. FMA3 f0, f29, f21, f0
  579. FMA3 f2, f31, f21, f2
  580. FMA4 f5, f29, f22, f5
  581. FMA4 f7, f31, f22, f7
  582. FMA3 f4, f29, f23, f4
  583. FMA3 f6, f31, f23, f6
  584. LFD f20, 48 * SIZE(BO)
  585. LFD f21, 49 * SIZE(BO)
  586. LFD f22, 50 * SIZE(BO)
  587. LFD f23, 51 * SIZE(BO)
  588. FMA4 f9, f29, f24, f9
  589. FMA4 f11, f31, f24, f11
  590. FMA3 f8, f29, f25, f8
  591. FMA3 f10, f31, f25, f10
  592. FMA4 f13, f29, f26, f13
  593. FMA4 f15, f31, f26, f15
  594. FMA3 f12, f29, f27, f12
  595. FMA3 f14, f31, f27, f14
  596. LFD f24, 52 * SIZE(BO)
  597. LFD f25, 53 * SIZE(BO)
  598. LFD f26, 54 * SIZE(BO)
  599. LFD f27, 55 * SIZE(BO)
  600. FMA1 f0, f16, f20, f0
  601. FMA1 f2, f18, f20, f2
  602. FMA2 f1, f16, f21, f1
  603. FMA2 f3, f18, f21, f3
  604. LFD f28, 28 * SIZE(AO)
  605. LFD f29, 29 * SIZE(AO)
  606. LFD f30, 30 * SIZE(AO)
  607. LFD f31, 31 * SIZE(AO)
  608. FMA1 f4, f16, f22, f4
  609. FMA1 f6, f18, f22, f6
  610. FMA2 f5, f16, f23, f5
  611. FMA2 f7, f18, f23, f7
  612. FMA1 f8, f16, f24, f8
  613. FMA1 f10, f18, f24, f10
  614. FMA2 f9, f16, f25, f9
  615. FMA2 f11, f18, f25, f11
  616. FMA1 f12, f16, f26, f12
  617. FMA1 f14, f18, f26, f14
  618. FMA2 f13, f16, f27, f13
  619. FMA2 f15, f18, f27, f15
  620. FMA4 f1, f17, f20, f1
  621. FMA4 f3, f19, f20, f3
  622. FMA3 f0, f17, f21, f0
  623. FMA3 f2, f19, f21, f2
  624. FMA4 f5, f17, f22, f5
  625. FMA4 f7, f19, f22, f7
  626. FMA3 f4, f17, f23, f4
  627. FMA3 f6, f19, f23, f6
  628. LFD f20, 56 * SIZE(BO)
  629. LFD f21, 57 * SIZE(BO)
  630. LFD f22, 58 * SIZE(BO)
  631. LFD f23, 59 * SIZE(BO)
  632. FMA4 f9, f17, f24, f9
  633. FMA4 f11, f19, f24, f11
  634. FMA3 f8, f17, f25, f8
  635. FMA3 f10, f19, f25, f10
  636. FMA4 f13, f17, f26, f13
  637. FMA4 f15, f19, f26, f15
  638. FMA3 f12, f17, f27, f12
  639. FMA3 f14, f19, f27, f14
  640. LFD f24, 60 * SIZE(BO)
  641. LFD f25, 61 * SIZE(BO)
  642. LFD f26, 62 * SIZE(BO)
  643. LFD f27, 63 * SIZE(BO)
  644. FMA1 f0, f28, f20, f0
  645. FMA1 f2, f30, f20, f2
  646. FMA2 f1, f28, f21, f1
  647. FMA2 f3, f30, f21, f3
  648. LFD f16, 32 * SIZE(AO)
  649. LFD f17, 33 * SIZE(AO)
  650. LFD f18, 34 * SIZE(AO)
  651. LFD f19, 35 * SIZE(AO)
  652. FMA1 f4, f28, f22, f4
  653. FMA1 f6, f30, f22, f6
  654. FMA2 f5, f28, f23, f5
  655. FMA2 f7, f30, f23, f7
  656. FMA1 f8, f28, f24, f8
  657. FMA1 f10, f30, f24, f10
  658. FMA2 f9, f28, f25, f9
  659. FMA2 f11, f30, f25, f11
  660. FMA1 f12, f28, f26, f12
  661. FMA1 f14, f30, f26, f14
  662. FMA2 f13, f28, f27, f13
  663. FMA2 f15, f30, f27, f15
  664. FMA4 f1, f29, f20, f1
  665. FMA4 f3, f31, f20, f3
  666. FMA3 f0, f29, f21, f0
  667. FMA3 f2, f31, f21, f2
  668. FMA4 f5, f29, f22, f5
  669. FMA4 f7, f31, f22, f7
  670. FMA3 f4, f29, f23, f4
  671. FMA3 f6, f31, f23, f6
  672. LFD f20, 64 * SIZE(BO)
  673. LFD f21, 65 * SIZE(BO)
  674. LFD f22, 66 * SIZE(BO)
  675. LFD f23, 67 * SIZE(BO)
  676. FMA4 f9, f29, f24, f9
  677. FMA4 f11, f31, f24, f11
  678. FMA3 f8, f29, f25, f8
  679. FMA3 f10, f31, f25, f10
  680. FMA4 f13, f29, f26, f13
  681. FMA4 f15, f31, f26, f15
  682. FMA3 f12, f29, f27, f12
  683. FMA3 f14, f31, f27, f14
  684. LFD f24, 68 * SIZE(BO)
  685. LFD f25, 69 * SIZE(BO)
  686. LFD f26, 70 * SIZE(BO)
  687. LFD f27, 71 * SIZE(BO)
  688. addi AO, AO, 32 * SIZE
  689. addi BO, BO, 64 * SIZE
  690. bdnz LL(12)
  691. .align 4
  692. LL(15):
  693. lfd f30, ALPHA_R
  694. lfd f31, ALPHA_I
  695. dcbtst B, BB
  696. addi BB, BB, 16 * SIZE
  697. dcbtst B, BB
  698. addi BB, BB, 16 * SIZE
  699. #ifndef TRMMKERNEL
  700. andi. r0, K, 7
  701. mtspr CTR, r0
  702. ble LL(18)
  703. #else
  704. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  705. sub TEMP, K, KK
  706. #elif defined(LEFT)
  707. addi TEMP, KK, 2
  708. #else
  709. addi TEMP, KK, 4
  710. #endif
  711. andi. TEMP, TEMP, 7
  712. mtspr CTR, TEMP
  713. ble LL(18)
  714. #endif
  715. .align 4
  716. LL(16):
  717. FMA1 f0, f16, f20, f0
  718. FMA1 f2, f18, f20, f2
  719. FMA2 f1, f16, f21, f1
  720. FMA2 f3, f18, f21, f3
  721. FMA1 f4, f16, f22, f4
  722. FMA1 f6, f18, f22, f6
  723. FMA2 f5, f16, f23, f5
  724. FMA2 f7, f18, f23, f7
  725. FMA1 f8, f16, f24, f8
  726. FMA1 f10, f18, f24, f10
  727. FMA2 f9, f16, f25, f9
  728. FMA2 f11, f18, f25, f11
  729. FMA1 f12, f16, f26, f12
  730. FMA1 f14, f18, f26, f14
  731. FMA2 f13, f16, f27, f13
  732. FMA2 f15, f18, f27, f15
  733. FMA4 f1, f17, f20, f1
  734. FMA4 f3, f19, f20, f3
  735. FMA3 f0, f17, f21, f0
  736. FMA3 f2, f19, f21, f2
  737. FMA4 f5, f17, f22, f5
  738. FMA4 f7, f19, f22, f7
  739. FMA3 f4, f17, f23, f4
  740. FMA3 f6, f19, f23, f6
  741. FMA4 f9, f17, f24, f9
  742. FMA4 f11, f19, f24, f11
  743. FMA3 f8, f17, f25, f8
  744. FMA3 f10, f19, f25, f10
  745. FMA4 f13, f17, f26, f13
  746. FMA4 f15, f19, f26, f15
  747. FMA3 f12, f17, f27, f12
  748. FMA3 f14, f19, f27, f14
  749. LFD f16, 4 * SIZE(AO)
  750. LFD f17, 5 * SIZE(AO)
  751. LFD f18, 6 * SIZE(AO)
  752. LFD f19, 7 * SIZE(AO)
  753. LFD f20, 8 * SIZE(BO)
  754. LFD f21, 9 * SIZE(BO)
  755. LFD f22, 10 * SIZE(BO)
  756. LFD f23, 11 * SIZE(BO)
  757. LFD f24, 12 * SIZE(BO)
  758. LFD f25, 13 * SIZE(BO)
  759. LFD f26, 14 * SIZE(BO)
  760. LFD f27, 15 * SIZE(BO)
  761. addi AO, AO, 4 * SIZE
  762. addi BO, BO, 8 * SIZE
  763. bdnz LL(16)
  764. .align 4
  765. LL(18):
  766. #ifndef TRMMKERNEL
  767. LFD f16, 0 * SIZE(CO1)
  768. LFD f17, 1 * SIZE(CO1)
  769. LFD f18, 2 * SIZE(CO1)
  770. LFD f19, 3 * SIZE(CO1)
  771. LFD f20, 0 * SIZE(CO2)
  772. LFD f21, 1 * SIZE(CO2)
  773. LFD f22, 2 * SIZE(CO2)
  774. LFD f23, 3 * SIZE(CO2)
  775. FNMSUB f24, f31, f1, f16
  776. FMADD f25, f31, f0, f17
  777. FNMSUB f26, f31, f3, f18
  778. FMADD f27, f31, f2, f19
  779. LFD f16, 0 * SIZE(CO3)
  780. LFD f17, 1 * SIZE(CO3)
  781. LFD f18, 2 * SIZE(CO3)
  782. LFD f19, 3 * SIZE(CO3)
  783. FMADD f0, f30, f0, f24
  784. FMADD f1, f30, f1, f25
  785. FMADD f2, f30, f2, f26
  786. FMADD f3, f30, f3, f27
  787. FNMSUB f24, f31, f5, f20
  788. FMADD f25, f31, f4, f21
  789. FNMSUB f26, f31, f7, f22
  790. FMADD f27, f31, f6, f23
  791. LFD f20, 0 * SIZE(CO4)
  792. LFD f21, 1 * SIZE(CO4)
  793. LFD f22, 2 * SIZE(CO4)
  794. LFD f23, 3 * SIZE(CO4)
  795. FMADD f4, f30, f4, f24
  796. FMADD f5, f30, f5, f25
  797. FMADD f6, f30, f6, f26
  798. FMADD f7, f30, f7, f27
  799. FNMSUB f24, f31, f9, f16
  800. FMADD f25, f31, f8, f17
  801. FNMSUB f26, f31, f11, f18
  802. FMADD f27, f31, f10, f19
  803. FMADD f8, f30, f8, f24
  804. FMADD f9, f30, f9, f25
  805. FMADD f10, f30, f10, f26
  806. FMADD f11, f30, f11, f27
  807. FNMSUB f24, f31, f13, f20
  808. FMADD f25, f31, f12, f21
  809. FNMSUB f26, f31, f15, f22
  810. FMADD f27, f31, f14, f23
  811. FMADD f12, f30, f12, f24
  812. FMADD f13, f30, f13, f25
  813. FMADD f14, f30, f14, f26
  814. FMADD f15, f30, f15, f27
  815. #else
  816. FMUL f16, f31, f1
  817. FMUL f17, f31, f0
  818. FMUL f18, f31, f3
  819. FMUL f19, f31, f2
  820. FMUL f20, f31, f5
  821. FMUL f21, f31, f4
  822. FMUL f22, f31, f7
  823. FMUL f23, f31, f6
  824. FMSUB f0, f30, f0, f16
  825. FMADD f1, f30, f1, f17
  826. FMADD f2, f30, f2, f18
  827. FMADD f3, f30, f3, f19
  828. FMSUB f4, f30, f4, f20
  829. FMADD f5, f30, f5, f21
  830. FMADD f6, f30, f6, f22
  831. FMADD f7, f30, f7, f23
  832. FMUL f16, f31, f9
  833. FMUL f17, f31, f8
  834. FMUL f18, f31, f11
  835. FMUL f19, f31, f10
  836. FMUL f20, f31, f13
  837. FMUL f21, f31, f12
  838. FMUL f22, f31, f15
  839. FMUL f23, f31, f14
  840. FMSUB f8, f30, f8, f16
  841. FMADD f9, f30, f9, f17
  842. FMADD f10, f30, f10, f18
  843. FMADD f11, f30, f11, f19
  844. FMSUB f12, f30, f12, f20
  845. FMADD f13, f30, f13, f21
  846. FMADD f14, f30, f14, f22
  847. FMADD f15, f30, f15, f23
  848. #endif
  849. STFD f0, 0 * SIZE(CO1)
  850. STFD f1, 1 * SIZE(CO1)
  851. STFD f2, 2 * SIZE(CO1)
  852. STFD f3, 3 * SIZE(CO1)
  853. lfs f0, FZERO
  854. fmr f1, f0
  855. fmr f2, f0
  856. fmr f3, f0
  857. STFD f4, 0 * SIZE(CO2)
  858. STFD f5, 1 * SIZE(CO2)
  859. STFD f6, 2 * SIZE(CO2)
  860. STFD f7, 3 * SIZE(CO2)
  861. fmr f4, f0
  862. fmr f5, f0
  863. fmr f6, f0
  864. fmr f7, f0
  865. STFD f8, 0 * SIZE(CO3)
  866. STFD f9, 1 * SIZE(CO3)
  867. STFD f10, 2 * SIZE(CO3)
  868. STFD f11, 3 * SIZE(CO3)
  869. fmr f8, f0
  870. fmr f9, f0
  871. fmr f10, f0
  872. fmr f11, f0
  873. STFD f12, 0 * SIZE(CO4)
  874. STFD f13, 1 * SIZE(CO4)
  875. STFD f14, 2 * SIZE(CO4)
  876. STFD f15, 3 * SIZE(CO4)
  877. fmr f12, f0
  878. fmr f13, f0
  879. fmr f14, f0
  880. fmr f15, f0
  881. addi CO1, CO1, 4 * SIZE
  882. addi CO2, CO2, 4 * SIZE
  883. addi CO3, CO3, 4 * SIZE
  884. addi CO4, CO4, 4 * SIZE
  885. #ifdef TRMMKERNEL
  886. #if ( defined(LEFT) && defined(TRANSA)) || \
  887. (!defined(LEFT) && !defined(TRANSA))
  888. sub TEMP, K, KK
  889. #ifdef LEFT
  890. addi TEMP, TEMP, -2
  891. #else
  892. addi TEMP, TEMP, -4
  893. #endif
  894. slwi r0, TEMP, 1 + ZBASE_SHIFT
  895. slwi TEMP, TEMP, 2 + ZBASE_SHIFT
  896. add AO, AO, r0
  897. add BO, BO, TEMP
  898. #endif
  899. #ifdef LEFT
  900. addi KK, KK, 2
  901. #endif
  902. #endif
  903. addic. I, I, -1
  904. bgt LL(11)
  905. .align 4
  906. LL(20):
  907. andi. I, M, 1
  908. ble LL(29)
  909. #ifndef TRMMKERNEL
  910. LFD f16, 0 * SIZE(AO)
  911. LFD f17, 1 * SIZE(AO)
  912. LFD f18, 2 * SIZE(AO)
  913. LFD f19, 3 * SIZE(AO)
  914. LFD f20, 0 * SIZE(B)
  915. LFD f21, 1 * SIZE(B)
  916. LFD f22, 2 * SIZE(B)
  917. LFD f23, 3 * SIZE(B)
  918. LFD f24, 4 * SIZE(B)
  919. LFD f25, 5 * SIZE(B)
  920. LFD f26, 6 * SIZE(B)
  921. LFD f27, 7 * SIZE(B)
  922. srawi. r0, K, 2
  923. mr BO, B
  924. mtspr CTR, r0
  925. ble LL(25)
  926. #else
  927. #if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
  928. (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
  929. LFD f16, 0 * SIZE(AO)
  930. LFD f17, 1 * SIZE(AO)
  931. LFD f18, 2 * SIZE(AO)
  932. LFD f19, 3 * SIZE(AO)
  933. LFD f20, 0 * SIZE(B)
  934. LFD f21, 1 * SIZE(B)
  935. LFD f22, 2 * SIZE(B)
  936. LFD f23, 3 * SIZE(B)
  937. LFD f24, 4 * SIZE(B)
  938. LFD f25, 5 * SIZE(B)
  939. LFD f26, 6 * SIZE(B)
  940. LFD f27, 7 * SIZE(B)
  941. mr BO, B
  942. #else
  943. slwi r0, KK, 0 + ZBASE_SHIFT
  944. slwi TEMP, KK, 2 + ZBASE_SHIFT
  945. add AO, AO, r0
  946. add BO, B, TEMP
  947. LFD f16, 0 * SIZE(AO)
  948. LFD f17, 1 * SIZE(AO)
  949. LFD f18, 2 * SIZE(AO)
  950. LFD f19, 3 * SIZE(AO)
  951. LFD f20, 0 * SIZE(BO)
  952. LFD f21, 1 * SIZE(BO)
  953. LFD f22, 2 * SIZE(BO)
  954. LFD f23, 3 * SIZE(BO)
  955. LFD f24, 4 * SIZE(BO)
  956. LFD f25, 5 * SIZE(BO)
  957. LFD f26, 6 * SIZE(BO)
  958. LFD f27, 7 * SIZE(BO)
  959. #endif
  960. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  961. sub TEMP, K, KK
  962. #elif defined(LEFT)
  963. addi TEMP, KK, 1
  964. #else
  965. addi TEMP, KK, 4
  966. #endif
  967. srawi. TEMP, TEMP, 2
  968. mtspr CTR, TEMP
  969. ble LL(25)
  970. #endif
  971. .align 4
  972. LL(22):
  973. FMA1 f0, f16, f20, f0
  974. FMA2 f1, f16, f21, f1
  975. FMA4 f1, f17, f20, f1
  976. FMA3 f0, f17, f21, f0
  977. LFD f28, 4 * SIZE(AO)
  978. LFD f29, 5 * SIZE(AO)
  979. LFD f30, 6 * SIZE(AO)
  980. LFD f31, 7 * SIZE(AO)
  981. FMA1 f4, f16, f22, f4
  982. FMA2 f5, f16, f23, f5
  983. FMA4 f5, f17, f22, f5
  984. FMA3 f4, f17, f23, f4
  985. LFD f20, 8 * SIZE(BO)
  986. LFD f21, 9 * SIZE(BO)
  987. LFD f22, 10 * SIZE(BO)
  988. LFD f23, 11 * SIZE(BO)
  989. FMA1 f8, f16, f24, f8
  990. FMA2 f9, f16, f25, f9
  991. FMA4 f9, f17, f24, f9
  992. FMA3 f8, f17, f25, f8
  993. FMA1 f12, f16, f26, f12
  994. FMA2 f13, f16, f27, f13
  995. FMA4 f13, f17, f26, f13
  996. FMA3 f12, f17, f27, f12
  997. LFD f24, 12 * SIZE(BO)
  998. LFD f25, 13 * SIZE(BO)
  999. LFD f26, 14 * SIZE(BO)
  1000. LFD f27, 15 * SIZE(BO)
  1001. FMA1 f0, f18, f20, f0
  1002. FMA2 f1, f18, f21, f1
  1003. FMA4 f1, f19, f20, f1
  1004. FMA3 f0, f19, f21, f0
  1005. FMA1 f4, f18, f22, f4
  1006. FMA2 f5, f18, f23, f5
  1007. FMA4 f5, f19, f22, f5
  1008. FMA3 f4, f19, f23, f4
  1009. LFD f20, 16 * SIZE(BO)
  1010. LFD f21, 17 * SIZE(BO)
  1011. LFD f22, 18 * SIZE(BO)
  1012. LFD f23, 19 * SIZE(BO)
  1013. FMA1 f8, f18, f24, f8
  1014. FMA2 f9, f18, f25, f9
  1015. FMA4 f9, f19, f24, f9
  1016. FMA3 f8, f19, f25, f8
  1017. FMA1 f12, f18, f26, f12
  1018. FMA2 f13, f18, f27, f13
  1019. FMA4 f13, f19, f26, f13
  1020. FMA3 f12, f19, f27, f12
  1021. LFD f24, 20 * SIZE(BO)
  1022. LFD f25, 21 * SIZE(BO)
  1023. LFD f26, 22 * SIZE(BO)
  1024. LFD f27, 23 * SIZE(BO)
  1025. FMA1 f0, f28, f20, f0
  1026. FMA2 f1, f28, f21, f1
  1027. FMA4 f1, f29, f20, f1
  1028. FMA3 f0, f29, f21, f0
  1029. LFD f16, 8 * SIZE(AO)
  1030. LFD f17, 9 * SIZE(AO)
  1031. LFD f18, 10 * SIZE(AO)
  1032. LFD f19, 11 * SIZE(AO)
  1033. FMA1 f4, f28, f22, f4
  1034. FMA2 f5, f28, f23, f5
  1035. FMA4 f5, f29, f22, f5
  1036. FMA3 f4, f29, f23, f4
  1037. LFD f20, 24 * SIZE(BO)
  1038. LFD f21, 25 * SIZE(BO)
  1039. LFD f22, 26 * SIZE(BO)
  1040. LFD f23, 27 * SIZE(BO)
  1041. FMA1 f8, f28, f24, f8
  1042. FMA2 f9, f28, f25, f9
  1043. FMA4 f9, f29, f24, f9
  1044. FMA3 f8, f29, f25, f8
  1045. FMA1 f12, f28, f26, f12
  1046. FMA2 f13, f28, f27, f13
  1047. FMA4 f13, f29, f26, f13
  1048. FMA3 f12, f29, f27, f12
  1049. LFD f24, 28 * SIZE(BO)
  1050. LFD f25, 29 * SIZE(BO)
  1051. LFD f26, 30 * SIZE(BO)
  1052. LFD f27, 31 * SIZE(BO)
  1053. FMA1 f0, f30, f20, f0
  1054. FMA2 f1, f30, f21, f1
  1055. FMA4 f1, f31, f20, f1
  1056. FMA3 f0, f31, f21, f0
  1057. FMA1 f4, f30, f22, f4
  1058. FMA2 f5, f30, f23, f5
  1059. FMA4 f5, f31, f22, f5
  1060. FMA3 f4, f31, f23, f4
  1061. LFD f20, 32 * SIZE(BO)
  1062. LFD f21, 33 * SIZE(BO)
  1063. LFD f22, 34 * SIZE(BO)
  1064. LFD f23, 35 * SIZE(BO)
  1065. FMA1 f8, f30, f24, f8
  1066. FMA2 f9, f30, f25, f9
  1067. FMA4 f9, f31, f24, f9
  1068. FMA3 f8, f31, f25, f8
  1069. FMA1 f12, f30, f26, f12
  1070. FMA2 f13, f30, f27, f13
  1071. FMA4 f13, f31, f26, f13
  1072. FMA3 f12, f31, f27, f12
  1073. LFD f24, 36 * SIZE(BO)
  1074. LFD f25, 37 * SIZE(BO)
  1075. LFD f26, 38 * SIZE(BO)
  1076. LFD f27, 39 * SIZE(BO)
  1077. addi AO, AO, 8 * SIZE
  1078. addi BO, BO, 32 * SIZE
  1079. bdnz LL(22)
  1080. .align 4
  1081. LL(25):
  1082. lfd f30, ALPHA_R
  1083. lfd f31, ALPHA_I
  1084. #ifndef TRMMKERNEL
  1085. andi. r0, K, 3
  1086. mtspr CTR, r0
  1087. ble LL(28)
  1088. #else
  1089. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  1090. sub TEMP, K, KK
  1091. #elif defined(LEFT)
  1092. addi TEMP, KK, 1
  1093. #else
  1094. addi TEMP, KK, 4
  1095. #endif
  1096. andi. TEMP, TEMP, 3
  1097. mtspr CTR, TEMP
  1098. ble LL(28)
  1099. #endif
  1100. .align 4
  1101. LL(26):
  1102. FMA1 f0, f16, f20, f0
  1103. FMA2 f1, f16, f21, f1
  1104. FMA4 f1, f17, f20, f1
  1105. FMA3 f0, f17, f21, f0
  1106. FMA1 f4, f16, f22, f4
  1107. FMA2 f5, f16, f23, f5
  1108. FMA4 f5, f17, f22, f5
  1109. FMA3 f4, f17, f23, f4
  1110. LFD f20, 8 * SIZE(BO)
  1111. LFD f21, 9 * SIZE(BO)
  1112. LFD f22, 10 * SIZE(BO)
  1113. LFD f23, 11 * SIZE(BO)
  1114. FMA1 f8, f16, f24, f8
  1115. FMA2 f9, f16, f25, f9
  1116. FMA4 f9, f17, f24, f9
  1117. FMA3 f8, f17, f25, f8
  1118. FMA1 f12, f16, f26, f12
  1119. FMA2 f13, f16, f27, f13
  1120. FMA4 f13, f17, f26, f13
  1121. FMA3 f12, f17, f27, f12
  1122. LFD f16, 2 * SIZE(AO)
  1123. LFD f17, 3 * SIZE(AO)
  1124. LFD f24, 12 * SIZE(BO)
  1125. LFD f25, 13 * SIZE(BO)
  1126. LFD f26, 14 * SIZE(BO)
  1127. LFD f27, 15 * SIZE(BO)
  1128. addi AO, AO, 2 * SIZE
  1129. addi BO, BO, 8 * SIZE
  1130. bdnz LL(26)
  1131. .align 4
  1132. LL(28):
  1133. #ifndef TRMMKERNEL
  1134. LFD f16, 0 * SIZE(CO1)
  1135. LFD f17, 1 * SIZE(CO1)
  1136. LFD f18, 0 * SIZE(CO2)
  1137. LFD f19, 1 * SIZE(CO2)
  1138. LFD f20, 0 * SIZE(CO3)
  1139. LFD f21, 1 * SIZE(CO3)
  1140. LFD f22, 0 * SIZE(CO4)
  1141. LFD f23, 1 * SIZE(CO4)
  1142. fmr f2, f0
  1143. fmr f3, f1
  1144. fmr f6, f4
  1145. fmr f7, f5
  1146. FMADD f24, f30, f0, f16
  1147. FMADD f25, f30, f1, f17
  1148. FMADD f26, f30, f4, f18
  1149. FMADD f27, f30, f5, f19
  1150. FNMSUB f0, f31, f3, f24
  1151. FMADD f1, f31, f2, f25
  1152. FNMSUB f4, f31, f7, f26
  1153. FMADD f5, f31, f6, f27
  1154. fmr f10, f8
  1155. fmr f11, f9
  1156. fmr f14, f12
  1157. fmr f15, f13
  1158. FMADD f24, f30, f8, f20
  1159. FMADD f25, f30, f9, f21
  1160. FMADD f26, f30, f12, f22
  1161. FMADD f27, f30, f13, f23
  1162. FNMSUB f8, f31, f11, f24
  1163. FMADD f9, f31, f10, f25
  1164. FNMSUB f12, f31, f15, f26
  1165. FMADD f13, f31, f14, f27
  1166. #else
  1167. FMUL f16, f31, f1
  1168. FMUL f17, f31, f0
  1169. FMUL f18, f31, f5
  1170. FMUL f19, f31, f4
  1171. FMSUB f0, f30, f0, f16
  1172. FMADD f1, f30, f1, f17
  1173. FMSUB f4, f30, f4, f18
  1174. FMADD f5, f30, f5, f19
  1175. FMUL f20, f31, f9
  1176. FMUL f21, f31, f8
  1177. FMUL f22, f31, f13
  1178. FMUL f23, f31, f12
  1179. FMSUB f8, f30, f8, f20
  1180. FMADD f9, f30, f9, f21
  1181. FMSUB f12, f30, f12, f22
  1182. FMADD f13, f30, f13, f23
  1183. #endif
  1184. STFD f0, 0 * SIZE(CO1)
  1185. STFD f1, 1 * SIZE(CO1)
  1186. STFD f4, 0 * SIZE(CO2)
  1187. STFD f5, 1 * SIZE(CO2)
  1188. STFD f8, 0 * SIZE(CO3)
  1189. STFD f9, 1 * SIZE(CO3)
  1190. STFD f12, 0 * SIZE(CO4)
  1191. STFD f13, 1 * SIZE(CO4)
  1192. addi CO1, CO1, 2 * SIZE
  1193. addi CO2, CO2, 2 * SIZE
  1194. addi CO3, CO3, 2 * SIZE
  1195. addi CO4, CO4, 2 * SIZE
  1196. #ifdef TRMMKERNEL
  1197. #if ( defined(LEFT) && defined(TRANSA)) || \
  1198. (!defined(LEFT) && !defined(TRANSA))
  1199. sub TEMP, K, KK
  1200. #ifdef LEFT
  1201. addi TEMP, TEMP, -1
  1202. #else
  1203. addi TEMP, TEMP, -4
  1204. #endif
  1205. slwi r0, TEMP, 0 + ZBASE_SHIFT
  1206. slwi TEMP, TEMP, 2 + ZBASE_SHIFT
  1207. add AO, AO, r0
  1208. add BO, BO, TEMP
  1209. #endif
  1210. #ifdef LEFT
  1211. addi KK, KK, 1
  1212. #endif
  1213. #endif
  1214. .align 4
  1215. LL(29):
  1216. #if defined(TRMMKERNEL) && !defined(LEFT)
  1217. addi KK, KK, 4
  1218. #endif
  1219. mr B, BO
  1220. addic. J, J, -1
  1221. bgt LL(10)
  1222. .align 4
  1223. LL(30):
  1224. andi. J, N, 2
  1225. ble LL(50)
  1226. mr CO1, C
  1227. add CO2, C, LDC
  1228. add C, CO2, LDC
  1229. #if defined(TRMMKERNEL) && defined(LEFT)
  1230. mr KK, OFFSET
  1231. #endif
  1232. slwi BB, K, ZBASE_SHIFT + 1
  1233. mr AO, A
  1234. lfs f0, FZERO
  1235. fmr f1, f0
  1236. fmr f2, f0
  1237. fmr f3, f0
  1238. fmr f4, f0
  1239. fmr f5, f0
  1240. fmr f6, f0
  1241. fmr f7, f0
  1242. fmr f8, f0
  1243. fmr f9, f0
  1244. fmr f10, f0
  1245. fmr f11, f0
  1246. fmr f12, f0
  1247. fmr f13, f0
  1248. fmr f14, f0
  1249. fmr f15, f0
  1250. srawi. I, M, 1
  1251. ble LL(40)
  1252. .align 4
  1253. LL(31):
  1254. #ifndef TRMMKERNEL
  1255. LFD f16, 0 * SIZE(AO)
  1256. LFD f17, 1 * SIZE(AO)
  1257. LFD f18, 2 * SIZE(AO)
  1258. LFD f19, 3 * SIZE(AO)
  1259. LFD f20, 0 * SIZE(B)
  1260. LFD f21, 1 * SIZE(B)
  1261. LFD f22, 2 * SIZE(B)
  1262. LFD f23, 3 * SIZE(B)
  1263. LFD f24, 4 * SIZE(B)
  1264. LFD f25, 5 * SIZE(B)
  1265. LFD f26, 6 * SIZE(B)
  1266. LFD f27, 7 * SIZE(B)
  1267. dcbtst CO1, PREC
  1268. dcbtst CO2, PREC
  1269. srawi. r0, K, 3
  1270. mr BO, B
  1271. mtspr CTR, r0
  1272. ble LL(35)
  1273. #else
  1274. #if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
  1275. (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
  1276. LFD f16, 0 * SIZE(AO)
  1277. LFD f17, 1 * SIZE(AO)
  1278. LFD f18, 2 * SIZE(AO)
  1279. LFD f19, 3 * SIZE(AO)
  1280. LFD f20, 0 * SIZE(B)
  1281. LFD f21, 1 * SIZE(B)
  1282. LFD f22, 2 * SIZE(B)
  1283. LFD f23, 3 * SIZE(B)
  1284. LFD f24, 4 * SIZE(B)
  1285. LFD f25, 5 * SIZE(B)
  1286. LFD f26, 6 * SIZE(B)
  1287. LFD f27, 7 * SIZE(B)
  1288. mr BO, B
  1289. #else
  1290. slwi r0, KK, 1 + ZBASE_SHIFT
  1291. slwi TEMP, KK, 1 + ZBASE_SHIFT
  1292. add AO, AO, r0
  1293. add BO, B, TEMP
  1294. LFD f16, 0 * SIZE(AO)
  1295. LFD f17, 1 * SIZE(AO)
  1296. LFD f18, 2 * SIZE(AO)
  1297. LFD f19, 3 * SIZE(AO)
  1298. LFD f20, 0 * SIZE(BO)
  1299. LFD f21, 1 * SIZE(BO)
  1300. LFD f22, 2 * SIZE(BO)
  1301. LFD f23, 3 * SIZE(BO)
  1302. LFD f24, 4 * SIZE(BO)
  1303. LFD f25, 5 * SIZE(BO)
  1304. LFD f26, 6 * SIZE(BO)
  1305. LFD f27, 7 * SIZE(BO)
  1306. #endif
  1307. dcbtst CO1, PREC
  1308. dcbtst CO2, PREC
  1309. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  1310. sub TEMP, K, KK
  1311. #elif defined(LEFT)
  1312. addi TEMP, KK, 2
  1313. #else
  1314. addi TEMP, KK, 2
  1315. #endif
  1316. srawi. TEMP, TEMP, 3
  1317. mtspr CTR, TEMP
  1318. ble LL(35)
  1319. #endif
  1320. .align 4
  1321. LL(32):
  1322. dcbt AO, PREA
  1323. dcbtst BO, PREA
  1324. FMA1 f0, f16, f20, f0
  1325. FMA1 f2, f18, f20, f2
  1326. FMA2 f1, f16, f21, f1
  1327. FMA2 f3, f18, f21, f3
  1328. LFD f28, 4 * SIZE(AO)
  1329. LFD f29, 5 * SIZE(AO)
  1330. LFD f30, 6 * SIZE(AO)
  1331. LFD f31, 7 * SIZE(AO)
  1332. FMA1 f4, f16, f22, f4
  1333. FMA1 f6, f18, f22, f6
  1334. FMA2 f5, f16, f23, f5
  1335. FMA2 f7, f18, f23, f7
  1336. FMA4 f1, f17, f20, f1
  1337. FMA4 f3, f19, f20, f3
  1338. FMA3 f0, f17, f21, f0
  1339. FMA3 f2, f19, f21, f2
  1340. FMA4 f5, f17, f22, f5
  1341. FMA4 f7, f19, f22, f7
  1342. FMA3 f4, f17, f23, f4
  1343. FMA3 f6, f19, f23, f6
  1344. LFD f20, 8 * SIZE(BO)
  1345. LFD f21, 9 * SIZE(BO)
  1346. LFD f22, 10 * SIZE(BO)
  1347. LFD f23, 11 * SIZE(BO)
  1348. FMA1 f0, f28, f24, f0
  1349. FMA1 f2, f30, f24, f2
  1350. FMA2 f1, f28, f25, f1
  1351. FMA2 f3, f30, f25, f3
  1352. LFD f16, 8 * SIZE(AO)
  1353. LFD f17, 9 * SIZE(AO)
  1354. LFD f18, 10 * SIZE(AO)
  1355. LFD f19, 11 * SIZE(AO)
  1356. FMA1 f4, f28, f26, f4
  1357. FMA1 f6, f30, f26, f6
  1358. FMA2 f5, f28, f27, f5
  1359. FMA2 f7, f30, f27, f7
  1360. FMA4 f1, f29, f24, f1
  1361. FMA4 f3, f31, f24, f3
  1362. FMA3 f0, f29, f25, f0
  1363. FMA3 f2, f31, f25, f2
  1364. FMA4 f5, f29, f26, f5
  1365. FMA4 f7, f31, f26, f7
  1366. FMA3 f4, f29, f27, f4
  1367. FMA3 f6, f31, f27, f6
  1368. LFD f24, 12 * SIZE(BO)
  1369. LFD f25, 13 * SIZE(BO)
  1370. LFD f26, 14 * SIZE(BO)
  1371. LFD f27, 15 * SIZE(BO)
  1372. FMA1 f0, f16, f20, f0
  1373. FMA1 f2, f18, f20, f2
  1374. FMA2 f1, f16, f21, f1
  1375. FMA2 f3, f18, f21, f3
  1376. LFD f28, 12 * SIZE(AO)
  1377. LFD f29, 13 * SIZE(AO)
  1378. LFD f30, 14 * SIZE(AO)
  1379. LFD f31, 15 * SIZE(AO)
  1380. FMA1 f4, f16, f22, f4
  1381. FMA1 f6, f18, f22, f6
  1382. FMA2 f5, f16, f23, f5
  1383. FMA2 f7, f18, f23, f7
  1384. FMA4 f1, f17, f20, f1
  1385. FMA4 f3, f19, f20, f3
  1386. FMA3 f0, f17, f21, f0
  1387. FMA3 f2, f19, f21, f2
  1388. FMA4 f5, f17, f22, f5
  1389. FMA4 f7, f19, f22, f7
  1390. FMA3 f4, f17, f23, f4
  1391. FMA3 f6, f19, f23, f6
  1392. LFD f20, 16 * SIZE(BO)
  1393. LFD f21, 17 * SIZE(BO)
  1394. LFD f22, 18 * SIZE(BO)
  1395. LFD f23, 19 * SIZE(BO)
  1396. FMA1 f0, f28, f24, f0
  1397. FMA1 f2, f30, f24, f2
  1398. FMA2 f1, f28, f25, f1
  1399. FMA2 f3, f30, f25, f3
  1400. LFD f16, 16 * SIZE(AO)
  1401. LFD f17, 17 * SIZE(AO)
  1402. LFD f18, 18 * SIZE(AO)
  1403. LFD f19, 19 * SIZE(AO)
  1404. FMA1 f4, f28, f26, f4
  1405. FMA1 f6, f30, f26, f6
  1406. FMA2 f5, f28, f27, f5
  1407. FMA2 f7, f30, f27, f7
  1408. FMA4 f1, f29, f24, f1
  1409. FMA4 f3, f31, f24, f3
  1410. FMA3 f0, f29, f25, f0
  1411. FMA3 f2, f31, f25, f2
  1412. FMA4 f5, f29, f26, f5
  1413. FMA4 f7, f31, f26, f7
  1414. FMA3 f4, f29, f27, f4
  1415. FMA3 f6, f31, f27, f6
  1416. LFD f24, 20 * SIZE(BO)
  1417. LFD f25, 21 * SIZE(BO)
  1418. LFD f26, 22 * SIZE(BO)
  1419. LFD f27, 23 * SIZE(BO)
  1420. FMA1 f0, f16, f20, f0
  1421. FMA1 f2, f18, f20, f2
  1422. FMA2 f1, f16, f21, f1
  1423. FMA2 f3, f18, f21, f3
  1424. LFD f28, 20 * SIZE(AO)
  1425. LFD f29, 21 * SIZE(AO)
  1426. LFD f30, 22 * SIZE(AO)
  1427. LFD f31, 23 * SIZE(AO)
  1428. FMA1 f4, f16, f22, f4
  1429. FMA1 f6, f18, f22, f6
  1430. FMA2 f5, f16, f23, f5
  1431. FMA2 f7, f18, f23, f7
  1432. FMA4 f1, f17, f20, f1
  1433. FMA4 f3, f19, f20, f3
  1434. FMA3 f0, f17, f21, f0
  1435. FMA3 f2, f19, f21, f2
  1436. FMA4 f5, f17, f22, f5
  1437. FMA4 f7, f19, f22, f7
  1438. FMA3 f4, f17, f23, f4
  1439. FMA3 f6, f19, f23, f6
  1440. LFD f20, 24 * SIZE(BO)
  1441. LFD f21, 25 * SIZE(BO)
  1442. LFD f22, 26 * SIZE(BO)
  1443. LFD f23, 27 * SIZE(BO)
  1444. FMA1 f0, f28, f24, f0
  1445. FMA1 f2, f30, f24, f2
  1446. FMA2 f1, f28, f25, f1
  1447. FMA2 f3, f30, f25, f3
  1448. LFD f16, 24 * SIZE(AO)
  1449. LFD f17, 25 * SIZE(AO)
  1450. LFD f18, 26 * SIZE(AO)
  1451. LFD f19, 27 * SIZE(AO)
  1452. FMA1 f4, f28, f26, f4
  1453. FMA1 f6, f30, f26, f6
  1454. FMA2 f5, f28, f27, f5
  1455. FMA2 f7, f30, f27, f7
  1456. FMA4 f1, f29, f24, f1
  1457. FMA4 f3, f31, f24, f3
  1458. FMA3 f0, f29, f25, f0
  1459. FMA3 f2, f31, f25, f2
  1460. FMA4 f5, f29, f26, f5
  1461. FMA4 f7, f31, f26, f7
  1462. FMA3 f4, f29, f27, f4
  1463. FMA3 f6, f31, f27, f6
  1464. LFD f24, 28 * SIZE(BO)
  1465. LFD f25, 29 * SIZE(BO)
  1466. LFD f26, 30 * SIZE(BO)
  1467. LFD f27, 31 * SIZE(BO)
  1468. FMA1 f0, f16, f20, f0
  1469. FMA1 f2, f18, f20, f2
  1470. FMA2 f1, f16, f21, f1
  1471. FMA2 f3, f18, f21, f3
  1472. LFD f28, 28 * SIZE(AO)
  1473. LFD f29, 29 * SIZE(AO)
  1474. LFD f30, 30 * SIZE(AO)
  1475. LFD f31, 31 * SIZE(AO)
  1476. FMA1 f4, f16, f22, f4
  1477. FMA1 f6, f18, f22, f6
  1478. FMA2 f5, f16, f23, f5
  1479. FMA2 f7, f18, f23, f7
  1480. FMA4 f1, f17, f20, f1
  1481. FMA4 f3, f19, f20, f3
  1482. FMA3 f0, f17, f21, f0
  1483. FMA3 f2, f19, f21, f2
  1484. FMA4 f5, f17, f22, f5
  1485. FMA4 f7, f19, f22, f7
  1486. FMA3 f4, f17, f23, f4
  1487. FMA3 f6, f19, f23, f6
  1488. LFD f20, 32 * SIZE(BO)
  1489. LFD f21, 33 * SIZE(BO)
  1490. LFD f22, 34 * SIZE(BO)
  1491. LFD f23, 35 * SIZE(BO)
  1492. FMA1 f0, f28, f24, f0
  1493. FMA1 f2, f30, f24, f2
  1494. FMA2 f1, f28, f25, f1
  1495. FMA2 f3, f30, f25, f3
  1496. LFD f16, 32 * SIZE(AO)
  1497. LFD f17, 33 * SIZE(AO)
  1498. LFD f18, 34 * SIZE(AO)
  1499. LFD f19, 35 * SIZE(AO)
  1500. FMA1 f4, f28, f26, f4
  1501. FMA1 f6, f30, f26, f6
  1502. FMA2 f5, f28, f27, f5
  1503. FMA2 f7, f30, f27, f7
  1504. FMA4 f1, f29, f24, f1
  1505. FMA4 f3, f31, f24, f3
  1506. FMA3 f0, f29, f25, f0
  1507. FMA3 f2, f31, f25, f2
  1508. FMA4 f5, f29, f26, f5
  1509. FMA4 f7, f31, f26, f7
  1510. FMA3 f4, f29, f27, f4
  1511. FMA3 f6, f31, f27, f6
  1512. LFD f24, 36 * SIZE(BO)
  1513. LFD f25, 37 * SIZE(BO)
  1514. LFD f26, 38 * SIZE(BO)
  1515. LFD f27, 39 * SIZE(BO)
  1516. addi AO, AO, 32 * SIZE
  1517. addi BO, BO, 32 * SIZE
  1518. bdnz LL(32)
  1519. .align 4
  1520. LL(35):
  1521. lfd f30, ALPHA_R
  1522. lfd f31, ALPHA_I
  1523. dcbtst B, BB
  1524. addi BB, BB, 16 * SIZE
  1525. #ifndef TRMMKERNEL
  1526. andi. r0, K, 7
  1527. mtspr CTR, r0
  1528. ble LL(38)
  1529. #else
  1530. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  1531. sub TEMP, K, KK
  1532. #elif defined(LEFT)
  1533. addi TEMP, KK, 2
  1534. #else
  1535. addi TEMP, KK, 2
  1536. #endif
  1537. andi. TEMP, TEMP, 7
  1538. mtspr CTR, TEMP
  1539. ble LL(38)
  1540. #endif
  1541. .align 4
  1542. LL(36):
  1543. FMA1 f0, f16, f20, f0
  1544. FMA1 f2, f18, f20, f2
  1545. FMA2 f1, f16, f21, f1
  1546. FMA2 f3, f18, f21, f3
  1547. FMA1 f4, f16, f22, f4
  1548. FMA1 f6, f18, f22, f6
  1549. FMA2 f5, f16, f23, f5
  1550. FMA2 f7, f18, f23, f7
  1551. FMA4 f1, f17, f20, f1
  1552. FMA4 f3, f19, f20, f3
  1553. FMA3 f0, f17, f21, f0
  1554. FMA3 f2, f19, f21, f2
  1555. LFD f16, 4 * SIZE(AO)
  1556. LFD f18, 6 * SIZE(AO)
  1557. LFD f20, 4 * SIZE(BO)
  1558. LFD f21, 5 * SIZE(BO)
  1559. FMA4 f5, f17, f22, f5
  1560. FMA4 f7, f19, f22, f7
  1561. FMA3 f4, f17, f23, f4
  1562. FMA3 f6, f19, f23, f6
  1563. LFD f17, 5 * SIZE(AO)
  1564. LFD f19, 7 * SIZE(AO)
  1565. LFD f22, 6 * SIZE(BO)
  1566. LFD f23, 7 * SIZE(BO)
  1567. addi AO, AO, 4 * SIZE
  1568. addi BO, BO, 4 * SIZE
  1569. bdnz LL(36)
  1570. .align 4
  1571. LL(38):
  1572. #ifndef TRMMKERNEL
  1573. LFD f16, 0 * SIZE(CO1)
  1574. LFD f17, 1 * SIZE(CO1)
  1575. LFD f18, 2 * SIZE(CO1)
  1576. LFD f19, 3 * SIZE(CO1)
  1577. LFD f20, 0 * SIZE(CO2)
  1578. LFD f21, 1 * SIZE(CO2)
  1579. LFD f22, 2 * SIZE(CO2)
  1580. LFD f23, 3 * SIZE(CO2)
  1581. fmr f8, f0
  1582. fmr f9, f1
  1583. fmr f10, f2
  1584. fmr f11, f3
  1585. FMADD f24, f30, f0, f16
  1586. FMADD f25, f30, f1, f17
  1587. FMADD f26, f30, f2, f18
  1588. FMADD f27, f30, f3, f19
  1589. FNMSUB f0, f31, f9, f24
  1590. FMADD f1, f31, f8, f25
  1591. FNMSUB f2, f31, f11, f26
  1592. FMADD f3, f31, f10, f27
  1593. fmr f12, f4
  1594. fmr f13, f5
  1595. fmr f14, f6
  1596. fmr f15, f7
  1597. FMADD f24, f30, f4, f20
  1598. FMADD f25, f30, f5, f21
  1599. FMADD f26, f30, f6, f22
  1600. FMADD f27, f30, f7, f23
  1601. FNMSUB f4, f31, f13, f24
  1602. FMADD f5, f31, f12, f25
  1603. FNMSUB f6, f31, f15, f26
  1604. FMADD f7, f31, f14, f27
  1605. #else
  1606. FMUL f16, f31, f1
  1607. FMUL f17, f31, f0
  1608. FMUL f18, f31, f3
  1609. FMUL f19, f31, f2
  1610. FMUL f20, f31, f5
  1611. FMUL f21, f31, f4
  1612. FMUL f22, f31, f7
  1613. FMUL f23, f31, f6
  1614. FMSUB f0, f30, f0, f16
  1615. FMADD f1, f30, f1, f17
  1616. FMADD f2, f30, f2, f18
  1617. FMADD f3, f30, f3, f19
  1618. FMSUB f4, f30, f4, f20
  1619. FMADD f5, f30, f5, f21
  1620. FMADD f6, f30, f6, f22
  1621. FMADD f7, f30, f7, f23
  1622. #endif
  1623. STFD f0, 0 * SIZE(CO1)
  1624. STFD f1, 1 * SIZE(CO1)
  1625. STFD f2, 2 * SIZE(CO1)
  1626. STFD f3, 3 * SIZE(CO1)
  1627. lfs f0, FZERO
  1628. fmr f1, f0
  1629. fmr f2, f0
  1630. fmr f3, f0
  1631. STFD f4, 0 * SIZE(CO2)
  1632. STFD f5, 1 * SIZE(CO2)
  1633. STFD f6, 2 * SIZE(CO2)
  1634. STFD f7, 3 * SIZE(CO2)
  1635. fmr f4, f0
  1636. fmr f5, f0
  1637. fmr f6, f0
  1638. fmr f7, f0
  1639. fmr f8, f0
  1640. fmr f9, f0
  1641. fmr f10, f0
  1642. fmr f11, f0
  1643. fmr f12, f0
  1644. fmr f13, f0
  1645. fmr f14, f0
  1646. fmr f15, f0
  1647. addi CO1, CO1, 4 * SIZE
  1648. addi CO2, CO2, 4 * SIZE
  1649. #ifdef TRMMKERNEL
  1650. #if ( defined(LEFT) && defined(TRANSA)) || \
  1651. (!defined(LEFT) && !defined(TRANSA))
  1652. sub TEMP, K, KK
  1653. #ifdef LEFT
  1654. addi TEMP, TEMP, -2
  1655. #else
  1656. addi TEMP, TEMP, -2
  1657. #endif
  1658. slwi r0, TEMP, 1 + ZBASE_SHIFT
  1659. slwi TEMP, TEMP, 1 + ZBASE_SHIFT
  1660. add AO, AO, r0
  1661. add BO, BO, TEMP
  1662. #endif
  1663. #ifdef LEFT
  1664. addi KK, KK, 2
  1665. #endif
  1666. #endif
  1667. addic. I, I, -1
  1668. bgt LL(31)
  1669. .align 4
  1670. LL(40):
  1671. andi. I, M, 1
  1672. ble LL(49)
  1673. #ifndef TRMMKERNEL
  1674. LFD f16, 0 * SIZE(AO)
  1675. LFD f17, 1 * SIZE(AO)
  1676. LFD f20, 0 * SIZE(B)
  1677. LFD f21, 1 * SIZE(B)
  1678. LFD f22, 2 * SIZE(B)
  1679. LFD f23, 3 * SIZE(B)
  1680. srawi. r0, K, 2
  1681. mr BO, B
  1682. mtspr CTR, r0
  1683. ble LL(45)
  1684. #else
  1685. #if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
  1686. (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
  1687. LFD f16, 0 * SIZE(AO)
  1688. LFD f17, 1 * SIZE(AO)
  1689. LFD f20, 0 * SIZE(B)
  1690. LFD f21, 1 * SIZE(B)
  1691. LFD f22, 2 * SIZE(B)
  1692. LFD f23, 3 * SIZE(B)
  1693. mr BO, B
  1694. #else
  1695. slwi r0, KK, 0 + ZBASE_SHIFT
  1696. slwi TEMP, KK, 1 + ZBASE_SHIFT
  1697. add AO, AO, r0
  1698. add BO, B, TEMP
  1699. LFD f16, 0 * SIZE(AO)
  1700. LFD f17, 1 * SIZE(AO)
  1701. LFD f20, 0 * SIZE(BO)
  1702. LFD f21, 1 * SIZE(BO)
  1703. LFD f22, 2 * SIZE(BO)
  1704. LFD f23, 3 * SIZE(BO)
  1705. #endif
  1706. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  1707. sub TEMP, K, KK
  1708. #elif defined(LEFT)
  1709. addi TEMP, KK, 1
  1710. #else
  1711. addi TEMP, KK, 2
  1712. #endif
  1713. srawi. TEMP, TEMP, 2
  1714. mtspr CTR, TEMP
  1715. ble LL(45)
  1716. #endif
  1717. .align 4
  1718. LL(42):
  1719. FMA1 f0, f16, f20, f0
  1720. FMA2 f1, f16, f21, f1
  1721. FMA4 f1, f17, f20, f1
  1722. FMA3 f0, f17, f21, f0
  1723. FMA1 f4, f16, f22, f4
  1724. FMA2 f5, f16, f23, f5
  1725. FMA4 f5, f17, f22, f5
  1726. FMA3 f4, f17, f23, f4
  1727. LFD f16, 2 * SIZE(AO)
  1728. LFD f17, 3 * SIZE(AO)
  1729. LFD f20, 4 * SIZE(BO)
  1730. LFD f21, 5 * SIZE(BO)
  1731. LFD f22, 6 * SIZE(BO)
  1732. LFD f23, 7 * SIZE(BO)
  1733. FMA1 f0, f16, f20, f0
  1734. FMA2 f1, f16, f21, f1
  1735. FMA4 f1, f17, f20, f1
  1736. FMA3 f0, f17, f21, f0
  1737. FMA1 f4, f16, f22, f4
  1738. FMA2 f5, f16, f23, f5
  1739. FMA4 f5, f17, f22, f5
  1740. FMA3 f4, f17, f23, f4
  1741. LFD f16, 4 * SIZE(AO)
  1742. LFD f17, 5 * SIZE(AO)
  1743. LFD f20, 8 * SIZE(BO)
  1744. LFD f21, 9 * SIZE(BO)
  1745. LFD f22, 10 * SIZE(BO)
  1746. LFD f23, 11 * SIZE(BO)
  1747. FMA1 f0, f16, f20, f0
  1748. FMA2 f1, f16, f21, f1
  1749. FMA4 f1, f17, f20, f1
  1750. FMA3 f0, f17, f21, f0
  1751. FMA1 f4, f16, f22, f4
  1752. FMA2 f5, f16, f23, f5
  1753. FMA4 f5, f17, f22, f5
  1754. FMA3 f4, f17, f23, f4
  1755. LFD f16, 6 * SIZE(AO)
  1756. LFD f17, 7 * SIZE(AO)
  1757. LFD f20, 12 * SIZE(BO)
  1758. LFD f21, 13 * SIZE(BO)
  1759. LFD f22, 14 * SIZE(BO)
  1760. LFD f23, 15 * SIZE(BO)
  1761. FMA1 f0, f16, f20, f0
  1762. FMA2 f1, f16, f21, f1
  1763. FMA4 f1, f17, f20, f1
  1764. FMA3 f0, f17, f21, f0
  1765. FMA1 f4, f16, f22, f4
  1766. FMA2 f5, f16, f23, f5
  1767. FMA4 f5, f17, f22, f5
  1768. FMA3 f4, f17, f23, f4
  1769. LFD f16, 8 * SIZE(AO)
  1770. LFD f17, 9 * SIZE(AO)
  1771. LFD f20, 16 * SIZE(BO)
  1772. LFD f21, 17 * SIZE(BO)
  1773. LFD f22, 18 * SIZE(BO)
  1774. LFD f23, 19 * SIZE(BO)
  1775. addi AO, AO, 8 * SIZE
  1776. addi BO, BO, 16 * SIZE
  1777. bdnz LL(42)
  1778. .align 4
  1779. LL(45):
  1780. lfd f30, ALPHA_R
  1781. lfd f31, ALPHA_I
  1782. #ifndef TRMMKERNEL
  1783. andi. r0, K, 3
  1784. mtspr CTR, r0
  1785. ble LL(48)
  1786. #else
  1787. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  1788. sub TEMP, K, KK
  1789. #elif defined(LEFT)
  1790. addi TEMP, KK, 1
  1791. #else
  1792. addi TEMP, KK, 2
  1793. #endif
  1794. andi. TEMP, TEMP, 3
  1795. mtspr CTR, TEMP
  1796. ble LL(48)
  1797. #endif
  1798. .align 4
  1799. LL(46):
  1800. FMA1 f0, f16, f20, f0
  1801. FMA2 f1, f16, f21, f1
  1802. FMA4 f1, f17, f20, f1
  1803. FMA3 f0, f17, f21, f0
  1804. FMA1 f4, f16, f22, f4
  1805. FMA2 f5, f16, f23, f5
  1806. FMA4 f5, f17, f22, f5
  1807. FMA3 f4, f17, f23, f4
  1808. LFD f16, 2 * SIZE(AO)
  1809. LFD f17, 3 * SIZE(AO)
  1810. LFD f20, 4 * SIZE(BO)
  1811. LFD f21, 5 * SIZE(BO)
  1812. LFD f22, 6 * SIZE(BO)
  1813. LFD f23, 7 * SIZE(BO)
  1814. addi AO, AO, 2 * SIZE
  1815. addi BO, BO, 4 * SIZE
  1816. bdnz LL(46)
  1817. .align 4
  1818. LL(48):
  1819. #ifndef TRMMKERNEL
  1820. LFD f16, 0 * SIZE(CO1)
  1821. LFD f17, 1 * SIZE(CO1)
  1822. LFD f20, 0 * SIZE(CO2)
  1823. LFD f21, 1 * SIZE(CO2)
  1824. fmr f2, f0
  1825. fmr f3, f1
  1826. fmr f6, f4
  1827. fmr f7, f5
  1828. FMADD f24, f30, f0, f16
  1829. FMADD f25, f30, f1, f17
  1830. FMADD f26, f30, f4, f20
  1831. FMADD f27, f30, f5, f21
  1832. FNMSUB f0, f31, f3, f24
  1833. FMADD f1, f31, f2, f25
  1834. FNMSUB f4, f31, f7, f26
  1835. FMADD f5, f31, f6, f27
  1836. #else
  1837. FMUL f16, f31, f1
  1838. FMUL f17, f31, f0
  1839. FMUL f18, f31, f5
  1840. FMUL f19, f31, f4
  1841. FMSUB f0, f30, f0, f16
  1842. FMADD f1, f30, f1, f17
  1843. FMSUB f4, f30, f4, f18
  1844. FMADD f5, f30, f5, f19
  1845. #endif
  1846. STFD f0, 0 * SIZE(CO1)
  1847. STFD f1, 1 * SIZE(CO1)
  1848. STFD f4, 0 * SIZE(CO2)
  1849. STFD f5, 1 * SIZE(CO2)
  1850. addi CO1, CO1, 2 * SIZE
  1851. addi CO2, CO2, 2 * SIZE
  1852. #ifdef TRMMKERNEL
  1853. #if ( defined(LEFT) && defined(TRANSA)) || \
  1854. (!defined(LEFT) && !defined(TRANSA))
  1855. sub TEMP, K, KK
  1856. #ifdef LEFT
  1857. addi TEMP, TEMP, -1
  1858. #else
  1859. addi TEMP, TEMP, -2
  1860. #endif
  1861. slwi r0, TEMP, 0 + ZBASE_SHIFT
  1862. slwi TEMP, TEMP, 1 + ZBASE_SHIFT
  1863. add AO, AO, r0
  1864. add BO, BO, TEMP
  1865. #endif
  1866. #ifdef LEFT
  1867. addi KK, KK, 1
  1868. #endif
  1869. #endif
  1870. .align 4
  1871. LL(49):
  1872. #if defined(TRMMKERNEL) && !defined(LEFT)
  1873. addi KK, KK, 2
  1874. #endif
  1875. mr B, BO
  1876. .align 4
  1877. LL(50):
  1878. andi. J, N, 1
  1879. ble LL(999)
  1880. mr CO1, C
  1881. add C, CO1, LDC
  1882. #if defined(TRMMKERNEL) && defined(LEFT)
  1883. mr KK, OFFSET
  1884. #endif
  1885. mr AO, A
  1886. lfs f0, FZERO
  1887. fmr f1, f0
  1888. fmr f2, f0
  1889. fmr f3, f0
  1890. fmr f8, f0
  1891. fmr f9, f0
  1892. fmr f10, f0
  1893. fmr f11, f0
  1894. srawi. I, M, 1
  1895. ble LL(60)
  1896. .align 4
  1897. LL(51):
  1898. #ifndef TRMMKERNEL
  1899. LFD f16, 0 * SIZE(AO)
  1900. LFD f17, 1 * SIZE(AO)
  1901. LFD f18, 2 * SIZE(AO)
  1902. LFD f19, 3 * SIZE(AO)
  1903. LFD f20, 0 * SIZE(B)
  1904. LFD f21, 1 * SIZE(B)
  1905. LFD f22, 2 * SIZE(B)
  1906. LFD f23, 3 * SIZE(B)
  1907. dcbtst CO1, PREC
  1908. srawi. r0, K, 3
  1909. mr BO, B
  1910. mtspr CTR, r0
  1911. ble LL(55)
  1912. #else
  1913. #if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
  1914. (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
  1915. LFD f16, 0 * SIZE(AO)
  1916. LFD f17, 1 * SIZE(AO)
  1917. LFD f18, 2 * SIZE(AO)
  1918. LFD f19, 3 * SIZE(AO)
  1919. LFD f20, 0 * SIZE(B)
  1920. LFD f21, 1 * SIZE(B)
  1921. LFD f22, 2 * SIZE(B)
  1922. LFD f23, 3 * SIZE(B)
  1923. mr BO, B
  1924. #else
  1925. slwi r0, KK, 1 + ZBASE_SHIFT
  1926. slwi TEMP, KK, 0 + ZBASE_SHIFT
  1927. add AO, AO, r0
  1928. add BO, B, TEMP
  1929. LFD f16, 0 * SIZE(AO)
  1930. LFD f17, 1 * SIZE(AO)
  1931. LFD f18, 2 * SIZE(AO)
  1932. LFD f19, 3 * SIZE(AO)
  1933. LFD f20, 0 * SIZE(BO)
  1934. LFD f21, 1 * SIZE(BO)
  1935. LFD f22, 2 * SIZE(BO)
  1936. LFD f23, 3 * SIZE(BO)
  1937. #endif
  1938. dcbtst CO1, PREC
  1939. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  1940. sub TEMP, K, KK
  1941. #elif defined(LEFT)
  1942. addi TEMP, KK, 2
  1943. #else
  1944. addi TEMP, KK, 1
  1945. #endif
  1946. srawi. TEMP, TEMP, 3
  1947. mtspr CTR, TEMP
  1948. ble LL(55)
  1949. #endif
  1950. .align 4
  1951. LL(52):
  1952. dcbt AO, PREA
  1953. dcbtst BO, PREA
  1954. FMA1 f0, f16, f20, f0
  1955. FMA1 f2, f18, f20, f2
  1956. FMA2 f1, f16, f21, f1
  1957. FMA2 f3, f18, f21, f3
  1958. FMA4 f1, f17, f20, f1
  1959. FMA4 f3, f19, f20, f3
  1960. FMA3 f0, f17, f21, f0
  1961. FMA3 f2, f19, f21, f2
  1962. LFD f16, 4 * SIZE(AO)
  1963. LFD f17, 5 * SIZE(AO)
  1964. LFD f18, 6 * SIZE(AO)
  1965. LFD f19, 7 * SIZE(AO)
  1966. FMA1 f0, f16, f22, f0
  1967. FMA1 f2, f18, f22, f2
  1968. FMA2 f1, f16, f23, f1
  1969. FMA2 f3, f18, f23, f3
  1970. FMA4 f1, f17, f22, f1
  1971. FMA4 f3, f19, f22, f3
  1972. FMA3 f0, f17, f23, f0
  1973. FMA3 f2, f19, f23, f2
  1974. LFD f16, 8 * SIZE(AO)
  1975. LFD f17, 9 * SIZE(AO)
  1976. LFD f18, 10 * SIZE(AO)
  1977. LFD f19, 11 * SIZE(AO)
  1978. LFD f20, 4 * SIZE(BO)
  1979. LFD f21, 5 * SIZE(BO)
  1980. LFD f22, 6 * SIZE(BO)
  1981. LFD f23, 7 * SIZE(BO)
  1982. FMA1 f0, f16, f20, f0
  1983. FMA1 f2, f18, f20, f2
  1984. FMA2 f1, f16, f21, f1
  1985. FMA2 f3, f18, f21, f3
  1986. FMA4 f1, f17, f20, f1
  1987. FMA4 f3, f19, f20, f3
  1988. FMA3 f0, f17, f21, f0
  1989. FMA3 f2, f19, f21, f2
  1990. LFD f16, 12 * SIZE(AO)
  1991. LFD f17, 13 * SIZE(AO)
  1992. LFD f18, 14 * SIZE(AO)
  1993. LFD f19, 15 * SIZE(AO)
  1994. FMA1 f0, f16, f22, f0
  1995. FMA1 f2, f18, f22, f2
  1996. FMA2 f1, f16, f23, f1
  1997. FMA2 f3, f18, f23, f3
  1998. FMA4 f1, f17, f22, f1
  1999. FMA4 f3, f19, f22, f3
  2000. FMA3 f0, f17, f23, f0
  2001. FMA3 f2, f19, f23, f2
  2002. LFD f16, 16 * SIZE(AO)
  2003. LFD f17, 17 * SIZE(AO)
  2004. LFD f18, 18 * SIZE(AO)
  2005. LFD f19, 19 * SIZE(AO)
  2006. LFD f20, 8 * SIZE(BO)
  2007. LFD f21, 9 * SIZE(BO)
  2008. LFD f22, 10 * SIZE(BO)
  2009. LFD f23, 11 * SIZE(BO)
  2010. FMA1 f0, f16, f20, f0
  2011. FMA1 f2, f18, f20, f2
  2012. FMA2 f1, f16, f21, f1
  2013. FMA2 f3, f18, f21, f3
  2014. FMA4 f1, f17, f20, f1
  2015. FMA4 f3, f19, f20, f3
  2016. FMA3 f0, f17, f21, f0
  2017. FMA3 f2, f19, f21, f2
  2018. LFD f16, 20 * SIZE(AO)
  2019. LFD f17, 21 * SIZE(AO)
  2020. LFD f18, 22 * SIZE(AO)
  2021. LFD f19, 23 * SIZE(AO)
  2022. FMA1 f0, f16, f22, f0
  2023. FMA1 f2, f18, f22, f2
  2024. FMA2 f1, f16, f23, f1
  2025. FMA2 f3, f18, f23, f3
  2026. FMA4 f1, f17, f22, f1
  2027. FMA4 f3, f19, f22, f3
  2028. FMA3 f0, f17, f23, f0
  2029. FMA3 f2, f19, f23, f2
  2030. LFD f16, 24 * SIZE(AO)
  2031. LFD f17, 25 * SIZE(AO)
  2032. LFD f18, 26 * SIZE(AO)
  2033. LFD f19, 27 * SIZE(AO)
  2034. LFD f20, 12 * SIZE(BO)
  2035. LFD f21, 13 * SIZE(BO)
  2036. LFD f22, 14 * SIZE(BO)
  2037. LFD f23, 15 * SIZE(BO)
  2038. FMA1 f0, f16, f20, f0
  2039. FMA1 f2, f18, f20, f2
  2040. FMA2 f1, f16, f21, f1
  2041. FMA2 f3, f18, f21, f3
  2042. FMA4 f1, f17, f20, f1
  2043. FMA4 f3, f19, f20, f3
  2044. FMA3 f0, f17, f21, f0
  2045. FMA3 f2, f19, f21, f2
  2046. LFD f16, 28 * SIZE(AO)
  2047. LFD f17, 29 * SIZE(AO)
  2048. LFD f18, 30 * SIZE(AO)
  2049. LFD f19, 31 * SIZE(AO)
  2050. FMA1 f0, f16, f22, f0
  2051. FMA1 f2, f18, f22, f2
  2052. FMA2 f1, f16, f23, f1
  2053. FMA2 f3, f18, f23, f3
  2054. FMA4 f1, f17, f22, f1
  2055. FMA4 f3, f19, f22, f3
  2056. FMA3 f0, f17, f23, f0
  2057. FMA3 f2, f19, f23, f2
  2058. LFD f16, 32 * SIZE(AO)
  2059. LFD f17, 33 * SIZE(AO)
  2060. LFD f18, 34 * SIZE(AO)
  2061. LFD f19, 35 * SIZE(AO)
  2062. LFD f20, 16 * SIZE(BO)
  2063. LFD f21, 17 * SIZE(BO)
  2064. LFD f22, 18 * SIZE(BO)
  2065. LFD f23, 19 * SIZE(BO)
  2066. addi AO, AO, 32 * SIZE
  2067. addi BO, BO, 16 * SIZE
  2068. bdnz LL(52)
  2069. .align 4
  2070. LL(55):
  2071. lfd f30, ALPHA_R
  2072. lfd f31, ALPHA_I
  2073. #ifndef TRMMKERNEL
  2074. andi. r0, K, 7
  2075. mtspr CTR, r0
  2076. ble LL(58)
  2077. #else
  2078. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2079. sub TEMP, K, KK
  2080. #elif defined(LEFT)
  2081. addi TEMP, KK, 2
  2082. #else
  2083. addi TEMP, KK, 1
  2084. #endif
  2085. andi. TEMP, TEMP, 7
  2086. mtspr CTR, TEMP
  2087. ble LL(58)
  2088. #endif
  2089. .align 4
  2090. LL(56):
  2091. FMA1 f0, f16, f20, f0
  2092. FMA1 f2, f18, f20, f2
  2093. FMA2 f1, f16, f21, f1
  2094. FMA2 f3, f18, f21, f3
  2095. LFD f16, 4 * SIZE(AO)
  2096. LFD f18, 6 * SIZE(AO)
  2097. FMA4 f1, f17, f20, f1
  2098. FMA4 f3, f19, f20, f3
  2099. FMA3 f0, f17, f21, f0
  2100. FMA3 f2, f19, f21, f2
  2101. LFD f17, 5 * SIZE(AO)
  2102. LFD f19, 7 * SIZE(AO)
  2103. LFD f20, 2 * SIZE(BO)
  2104. LFD f21, 3 * SIZE(BO)
  2105. addi AO, AO, 4 * SIZE
  2106. addi BO, BO, 2 * SIZE
  2107. bdnz LL(56)
  2108. .align 4
  2109. LL(58):
  2110. #ifndef TRMMKERNEL
  2111. LFD f16, 0 * SIZE(CO1)
  2112. LFD f17, 1 * SIZE(CO1)
  2113. LFD f18, 2 * SIZE(CO1)
  2114. LFD f19, 3 * SIZE(CO1)
  2115. fmr f8, f0
  2116. fmr f9, f1
  2117. fmr f10, f2
  2118. fmr f11, f3
  2119. FMADD f24, f30, f0, f16
  2120. FMADD f25, f30, f1, f17
  2121. FMADD f26, f30, f2, f18
  2122. FMADD f27, f30, f3, f19
  2123. FNMSUB f0, f31, f9, f24
  2124. FMADD f1, f31, f8, f25
  2125. FNMSUB f2, f31, f11, f26
  2126. FMADD f3, f31, f10, f27
  2127. #else
  2128. FMUL f16, f31, f1
  2129. FMUL f17, f31, f0
  2130. FMUL f18, f31, f3
  2131. FMUL f19, f31, f2
  2132. FMSUB f0, f30, f0, f16
  2133. FMADD f1, f30, f1, f17
  2134. FMADD f2, f30, f2, f18
  2135. FMADD f3, f30, f3, f19
  2136. #endif
  2137. STFD f0, 0 * SIZE(CO1)
  2138. STFD f1, 1 * SIZE(CO1)
  2139. STFD f2, 2 * SIZE(CO1)
  2140. STFD f3, 3 * SIZE(CO1)
  2141. lfs f0, FZERO
  2142. fmr f1, f0
  2143. fmr f2, f0
  2144. fmr f3, f0
  2145. fmr f8, f0
  2146. fmr f9, f0
  2147. fmr f10, f0
  2148. fmr f11, f0
  2149. addi CO1, CO1, 4 * SIZE
  2150. #ifdef TRMMKERNEL
  2151. #if ( defined(LEFT) && defined(TRANSA)) || \
  2152. (!defined(LEFT) && !defined(TRANSA))
  2153. sub TEMP, K, KK
  2154. #ifdef LEFT
  2155. addi TEMP, TEMP, -2
  2156. #else
  2157. addi TEMP, TEMP, -1
  2158. #endif
  2159. slwi r0, TEMP, 1 + ZBASE_SHIFT
  2160. slwi TEMP, TEMP, 0 + ZBASE_SHIFT
  2161. add AO, AO, r0
  2162. add BO, BO, TEMP
  2163. #endif
  2164. #ifdef LEFT
  2165. addi KK, KK, 2
  2166. #endif
  2167. #endif
  2168. addic. I, I, -1
  2169. bgt LL(51)
  2170. .align 4
  2171. LL(60):
  2172. andi. I, M, 1
  2173. ble LL(999)
  2174. #ifndef TRMMKERNEL
  2175. LFD f16, 0 * SIZE(AO)
  2176. LFD f17, 1 * SIZE(AO)
  2177. LFD f18, 2 * SIZE(AO)
  2178. LFD f19, 3 * SIZE(AO)
  2179. LFD f20, 0 * SIZE(B)
  2180. LFD f21, 1 * SIZE(B)
  2181. LFD f22, 2 * SIZE(B)
  2182. LFD f23, 3 * SIZE(B)
  2183. srawi. r0, K, 2
  2184. mr BO, B
  2185. mtspr CTR, r0
  2186. ble LL(65)
  2187. #else
  2188. #if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
  2189. (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
  2190. LFD f16, 0 * SIZE(AO)
  2191. LFD f17, 1 * SIZE(AO)
  2192. LFD f18, 2 * SIZE(AO)
  2193. LFD f19, 3 * SIZE(AO)
  2194. LFD f20, 0 * SIZE(B)
  2195. LFD f21, 1 * SIZE(B)
  2196. LFD f22, 2 * SIZE(B)
  2197. LFD f23, 3 * SIZE(B)
  2198. mr BO, B
  2199. #else
  2200. slwi r0, KK, 0 + ZBASE_SHIFT
  2201. slwi TEMP, KK, 0 + ZBASE_SHIFT
  2202. add AO, AO, r0
  2203. add BO, B, TEMP
  2204. LFD f16, 0 * SIZE(AO)
  2205. LFD f17, 1 * SIZE(AO)
  2206. LFD f18, 2 * SIZE(AO)
  2207. LFD f19, 3 * SIZE(AO)
  2208. LFD f20, 0 * SIZE(BO)
  2209. LFD f21, 1 * SIZE(BO)
  2210. LFD f22, 2 * SIZE(BO)
  2211. LFD f23, 3 * SIZE(BO)
  2212. #endif
  2213. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2214. sub TEMP, K, KK
  2215. #elif defined(LEFT)
  2216. addi TEMP, KK, 1
  2217. #else
  2218. addi TEMP, KK, 1
  2219. #endif
  2220. srawi. TEMP, TEMP, 2
  2221. mtspr CTR, TEMP
  2222. ble LL(65)
  2223. #endif
  2224. .align 4
  2225. LL(62):
  2226. FMA1 f0, f16, f20, f0
  2227. FMA2 f1, f16, f21, f1
  2228. FMA4 f1, f17, f20, f1
  2229. FMA3 f0, f17, f21, f0
  2230. LFD f16, 4 * SIZE(AO)
  2231. LFD f17, 5 * SIZE(AO)
  2232. LFD f20, 4 * SIZE(BO)
  2233. LFD f21, 5 * SIZE(BO)
  2234. FMA1 f0, f18, f22, f0
  2235. FMA2 f1, f18, f23, f1
  2236. FMA4 f1, f19, f22, f1
  2237. FMA3 f0, f19, f23, f0
  2238. LFD f18, 6 * SIZE(AO)
  2239. LFD f19, 7 * SIZE(AO)
  2240. LFD f22, 6 * SIZE(BO)
  2241. LFD f23, 7 * SIZE(BO)
  2242. FMA1 f0, f16, f20, f0
  2243. FMA2 f1, f16, f21, f1
  2244. FMA4 f1, f17, f20, f1
  2245. FMA3 f0, f17, f21, f0
  2246. LFD f16, 8 * SIZE(AO)
  2247. LFD f17, 9 * SIZE(AO)
  2248. LFD f20, 8 * SIZE(BO)
  2249. LFD f21, 9 * SIZE(BO)
  2250. FMA1 f0, f18, f22, f0
  2251. FMA2 f1, f18, f23, f1
  2252. FMA4 f1, f19, f22, f1
  2253. FMA3 f0, f19, f23, f0
  2254. LFD f18, 10 * SIZE(AO)
  2255. LFD f19, 11 * SIZE(AO)
  2256. LFD f22, 10 * SIZE(BO)
  2257. LFD f23, 11 * SIZE(BO)
  2258. addi AO, AO, 8 * SIZE
  2259. addi BO, BO, 8 * SIZE
  2260. bdnz LL(62)
  2261. .align 4
  2262. LL(65):
  2263. lfd f30, ALPHA_R
  2264. lfd f31, ALPHA_I
  2265. #ifndef TRMMKERNEL
  2266. andi. r0, K, 3
  2267. mtspr CTR, r0
  2268. ble LL(68)
  2269. #else
  2270. #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
  2271. sub TEMP, K, KK
  2272. #elif defined(LEFT)
  2273. addi TEMP, KK, 1
  2274. #else
  2275. addi TEMP, KK, 1
  2276. #endif
  2277. andi. TEMP, TEMP, 3
  2278. mtspr CTR, TEMP
  2279. ble LL(68)
  2280. #endif
  2281. .align 4
  2282. LL(66):
  2283. FMA1 f0, f16, f20, f0
  2284. FMA2 f1, f16, f21, f1
  2285. LFD f16, 2 * SIZE(AO)
  2286. FMA4 f1, f17, f20, f1
  2287. LFD f20, 2 * SIZE(BO)
  2288. FMA3 f0, f17, f21, f0
  2289. LFD f17, 3 * SIZE(AO)
  2290. LFD f21, 3 * SIZE(BO)
  2291. addi AO, AO, 2 * SIZE
  2292. addi BO, BO, 2 * SIZE
  2293. bdnz LL(66)
  2294. .align 4
  2295. LL(68):
  2296. #ifndef TRMMKERNEL
  2297. LFD f16, 0 * SIZE(CO1)
  2298. LFD f17, 1 * SIZE(CO1)
  2299. fmr f2, f0
  2300. fmr f3, f1
  2301. FMADD f24, f30, f0, f16
  2302. FMADD f25, f30, f1, f17
  2303. FNMSUB f0, f31, f3, f24
  2304. FMADD f1, f31, f2, f25
  2305. #else
  2306. FMUL f16, f31, f1
  2307. FMUL f17, f31, f0
  2308. FMSUB f0, f30, f0, f16
  2309. FMADD f1, f30, f1, f17
  2310. #endif
  2311. STFD f0, 0 * SIZE(CO1)
  2312. STFD f1, 1 * SIZE(CO1)
  2313. addi CO1, CO1, 2 * SIZE
  2314. #ifdef TRMMKERNEL
  2315. #if ( defined(LEFT) && defined(TRANSA)) || \
  2316. (!defined(LEFT) && !defined(TRANSA))
  2317. sub TEMP, K, KK
  2318. #ifdef LEFT
  2319. addi TEMP, TEMP, -1
  2320. #else
  2321. addi TEMP, TEMP, -1
  2322. #endif
  2323. slwi r0, TEMP, 0 + ZBASE_SHIFT
  2324. slwi TEMP, TEMP, 0 + ZBASE_SHIFT
  2325. add AO, AO, r0
  2326. add BO, BO, TEMP
  2327. #endif
  2328. #ifdef LEFT
  2329. addi KK, KK, 1
  2330. #endif
  2331. #endif
  2332. .align 4
  2333. LL(999):
  2334. addi r3, 0, 0
  2335. lfd f14, 0(SP)
  2336. lfd f15, 8(SP)
  2337. lfd f16, 16(SP)
  2338. lfd f17, 24(SP)
  2339. lfd f18, 32(SP)
  2340. lfd f19, 40(SP)
  2341. lfd f20, 48(SP)
  2342. lfd f21, 56(SP)
  2343. lfd f22, 64(SP)
  2344. lfd f23, 72(SP)
  2345. lfd f24, 80(SP)
  2346. lfd f25, 88(SP)
  2347. lfd f26, 96(SP)
  2348. lfd f27, 104(SP)
  2349. lfd f28, 112(SP)
  2350. lfd f29, 120(SP)
  2351. lfd f30, 128(SP)
  2352. lfd f31, 136(SP)
  2353. #ifdef __64BIT__
  2354. ld r31, 144(SP)
  2355. ld r30, 152(SP)
  2356. ld r29, 160(SP)
  2357. ld r28, 168(SP)
  2358. ld r27, 176(SP)
  2359. ld r26, 184(SP)
  2360. ld r25, 192(SP)
  2361. ld r24, 200(SP)
  2362. ld r23, 208(SP)
  2363. ld r22, 216(SP)
  2364. ld r21, 224(SP)
  2365. #ifdef TRMMKERNEL
  2366. ld r20, 232(SP)
  2367. ld r19, 240(SP)
  2368. #endif
  2369. #else
  2370. lwz r31, 144(SP)
  2371. lwz r30, 148(SP)
  2372. lwz r29, 152(SP)
  2373. lwz r28, 156(SP)
  2374. lwz r27, 160(SP)
  2375. lwz r26, 164(SP)
  2376. lwz r25, 168(SP)
  2377. lwz r24, 172(SP)
  2378. lwz r23, 176(SP)
  2379. lwz r22, 180(SP)
  2380. lwz r21, 184(SP)
  2381. #ifdef TRMMKERNEL
  2382. lwz r20, 188(SP)
  2383. lwz r19, 192(SP)
  2384. #endif
  2385. #endif
  2386. addi SP, SP, STACKSIZE
  2387. blr
  2388. EPILOGUE
  2389. #endif