|
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092 |
- #include "common.h"
-
- int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FLOAT* C,BLASLONG ldc ,BLASLONG offset)
- {
-
- BLASLONG i,j,k;
- FLOAT *C0,*C1,*C2,*C3,*ptrba,*ptrbb;
-
- FLOAT res0_0;
- FLOAT res0_1;
- FLOAT res0_2;
- FLOAT res0_3;
- FLOAT res0_4;
- FLOAT res0_5;
- FLOAT res0_6;
- FLOAT res0_7;
-
- FLOAT res0_8;
- FLOAT res0_9;
- FLOAT res0_10;
- FLOAT res0_11;
- FLOAT res0_12;
- FLOAT res0_13;
- FLOAT res0_14;
- FLOAT res0_15;
-
- FLOAT res1_0;
- FLOAT res1_1;
- FLOAT res1_2;
- FLOAT res1_3;
- FLOAT res1_4;
- FLOAT res1_5;
- FLOAT res1_6;
- FLOAT res1_7;
-
- FLOAT res1_8;
- FLOAT res1_9;
- FLOAT res1_10;
- FLOAT res1_11;
- FLOAT res1_12;
- FLOAT res1_13;
- FLOAT res1_14;
- FLOAT res1_15;
-
- FLOAT res2_0;
- FLOAT res2_1;
- FLOAT res2_2;
- FLOAT res2_3;
- FLOAT res2_4;
- FLOAT res2_5;
- FLOAT res2_6;
- FLOAT res2_7;
-
- FLOAT res2_8;
- FLOAT res2_9;
- FLOAT res2_10;
- FLOAT res2_11;
- FLOAT res2_12;
- FLOAT res2_13;
- FLOAT res2_14;
- FLOAT res2_15;
-
- FLOAT res3_0;
- FLOAT res3_1;
- FLOAT res3_2;
- FLOAT res3_3;
- FLOAT res3_4;
- FLOAT res3_5;
- FLOAT res3_6;
- FLOAT res3_7;
-
- FLOAT res3_8;
- FLOAT res3_9;
- FLOAT res3_10;
- FLOAT res3_11;
- FLOAT res3_12;
- FLOAT res3_13;
- FLOAT res3_14;
- FLOAT res3_15;
-
- FLOAT a0;
- FLOAT a1;
-
- FLOAT b0;
- FLOAT b1;
- FLOAT b2;
- FLOAT b3;
-
- BLASLONG off, temp;
-
- #if !defined(LEFT)
- off = -offset;
- #else
- off = 0;
- #endif
-
- for (j=0; j<bn/4; j+=1)
- {
- C0 = C;
- C1 = C0+ldc;
- C2 = C0+2*ldc;
- C3 = C0+3*ldc;
-
- #if defined(TRMMKERNEL) && defined(LEFT)
- off = offset;
- #endif
-
-
- ptrba = ba;
-
-
- for (i=0; i<bm/16; i+=1)
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*16;
- ptrbb = bb + off*4;
- #endif
-
- res0_0 = 0;
- res0_1 = 0;
- res0_2 = 0;
- res0_3 = 0;
- res0_4 = 0;
- res0_5 = 0;
- res0_6 = 0;
- res0_7 = 0;
-
- res0_8 = 0;
- res0_9 = 0;
- res0_10 = 0;
- res0_11 = 0;
- res0_12 = 0;
- res0_13 = 0;
- res0_14 = 0;
- res0_15 = 0;
-
- res1_0 = 0;
- res1_1 = 0;
- res1_2 = 0;
- res1_3 = 0;
- res1_4 = 0;
- res1_5 = 0;
- res1_6 = 0;
- res1_7 = 0;
-
- res1_8 = 0;
- res1_9 = 0;
- res1_10 = 0;
- res1_11 = 0;
- res1_12 = 0;
- res1_13 = 0;
- res1_14 = 0;
- res1_15 = 0;
-
- res2_0 = 0;
- res2_1 = 0;
- res2_2 = 0;
- res2_3 = 0;
- res2_4 = 0;
- res2_5 = 0;
- res2_6 = 0;
- res2_7 = 0;
-
- res2_8 = 0;
- res2_9 = 0;
- res2_10 = 0;
- res2_11 = 0;
- res2_12 = 0;
- res2_13 = 0;
- res2_14 = 0;
- res2_15 = 0;
-
- res3_0 = 0;
- res3_1 = 0;
- res3_2 = 0;
- res3_3 = 0;
- res3_4 = 0;
- res3_5 = 0;
- res3_6 = 0;
- res3_7 = 0;
-
- res3_8 = 0;
- res3_9 = 0;
- res3_10 = 0;
- res3_11 = 0;
- res3_12 = 0;
- res3_13 = 0;
- res3_14 = 0;
- res3_15 = 0;
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+16; // number of values in A
- #else
- temp = off+4; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
- b2 = ptrbb[2];
- b3 = ptrbb[3];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
- res2_0 += a0*b2;
- res3_0 += a0*b3;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
- res1_1 += a1*b1;
- res2_1 += a1*b2;
- res3_1 += a1*b3;
-
- a0 = ptrba[2];
- res0_2 += a0*b0;
- res1_2 += a0*b1;
- res2_2 += a0*b2;
- res3_2 += a0*b3;
-
- a1 = ptrba[3];
- res0_3 += a1*b0;
- res1_3 += a1*b1;
- res2_3 += a1*b2;
- res3_3 += a1*b3;
-
- a0 = ptrba[4];
- res0_4 += a0*b0;
- res1_4 += a0*b1;
- res2_4 += a0*b2;
- res3_4 += a0*b3;
-
- a1 = ptrba[5];
- res0_5 += a1*b0;
- res1_5 += a1*b1;
- res2_5 += a1*b2;
- res3_5 += a1*b3;
-
- a0 = ptrba[6];
- res0_6 += a0*b0;
- res1_6 += a0*b1;
- res2_6 += a0*b2;
- res3_6 += a0*b3;
-
- a1 = ptrba[7];
- res0_7 += a1*b0;
- res1_7 += a1*b1;
- res2_7 += a1*b2;
- res3_7 += a1*b3;
-
- a0 = ptrba[8];
- res0_8 += a0*b0;
- res1_8 += a0*b1;
- res2_8 += a0*b2;
- res3_8 += a0*b3;
-
- a1 = ptrba[9];
- res0_9 += a1*b0;
- res1_9 += a1*b1;
- res2_9 += a1*b2;
- res3_9 += a1*b3;
-
- a0 = ptrba[10];
- res0_10 += a0*b0;
- res1_10 += a0*b1;
- res2_10 += a0*b2;
- res3_10 += a0*b3;
-
- a1 = ptrba[11];
- res0_11 += a1*b0;
- res1_11 += a1*b1;
- res2_11 += a1*b2;
- res3_11 += a1*b3;
-
- a0 = ptrba[12];
- res0_12 += a0*b0;
- res1_12 += a0*b1;
- res2_12 += a0*b2;
- res3_12 += a0*b3;
-
- a1 = ptrba[13];
- res0_13 += a1*b0;
- res1_13 += a1*b1;
- res2_13 += a1*b2;
- res3_13 += a1*b3;
-
- a0 = ptrba[14];
- res0_14 += a0*b0;
- res1_14 += a0*b1;
- res2_14 += a0*b2;
- res3_14 += a0*b3;
-
- a1 = ptrba[15];
- res0_15 += a1*b0;
- res1_15 += a1*b1;
- res2_15 += a1*b2;
- res3_15 += a1*b3;
-
-
- ptrba = ptrba+16;
- ptrbb = ptrbb+4;
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
- res0_2 *= alpha;
- res0_3 *= alpha;
- res0_4 *= alpha;
- res0_5 *= alpha;
- res0_6 *= alpha;
- res0_7 *= alpha;
-
- res0_8 *= alpha;
- res0_9 *= alpha;
- res0_10 *= alpha;
- res0_11 *= alpha;
- res0_12 *= alpha;
- res0_13 *= alpha;
- res0_14 *= alpha;
- res0_15 *= alpha;
-
- res1_0 *= alpha;
- res1_1 *= alpha;
- res1_2 *= alpha;
- res1_3 *= alpha;
- res1_4 *= alpha;
- res1_5 *= alpha;
- res1_6 *= alpha;
- res1_7 *= alpha;
-
- res1_8 *= alpha;
- res1_9 *= alpha;
- res1_10 *= alpha;
- res1_11 *= alpha;
- res1_12 *= alpha;
- res1_13 *= alpha;
- res1_14 *= alpha;
- res1_15 *= alpha;
-
- res2_0 *= alpha;
- res2_1 *= alpha;
- res2_2 *= alpha;
- res2_3 *= alpha;
- res2_4 *= alpha;
- res2_5 *= alpha;
- res2_6 *= alpha;
- res2_7 *= alpha;
-
- res2_8 *= alpha;
- res2_9 *= alpha;
- res2_10 *= alpha;
- res2_11 *= alpha;
- res2_12 *= alpha;
- res2_13 *= alpha;
- res2_14 *= alpha;
- res2_15 *= alpha;
-
- res3_0 *= alpha;
- res3_1 *= alpha;
- res3_2 *= alpha;
- res3_3 *= alpha;
- res3_4 *= alpha;
- res3_5 *= alpha;
- res3_6 *= alpha;
- res3_7 *= alpha;
-
- res3_8 *= alpha;
- res3_9 *= alpha;
- res3_10 *= alpha;
- res3_11 *= alpha;
- res3_12 *= alpha;
- res3_13 *= alpha;
- res3_14 *= alpha;
- res3_15 *= alpha;
-
- C0[0] = res0_0;
- C0[1] = res0_1;
- C0[2] = res0_2;
- C0[3] = res0_3;
- C0[4] = res0_4;
- C0[5] = res0_5;
- C0[6] = res0_6;
- C0[7] = res0_7;
-
- C0[8] = res0_8;
- C0[9] = res0_9;
- C0[10] = res0_10;
- C0[11] = res0_11;
- C0[12] = res0_12;
- C0[13] = res0_13;
- C0[14] = res0_14;
- C0[15] = res0_15;
-
- C1[0] = res1_0;
- C1[1] = res1_1;
- C1[2] = res1_2;
- C1[3] = res1_3;
- C1[4] = res1_4;
- C1[5] = res1_5;
- C1[6] = res1_6;
- C1[7] = res1_7;
-
- C1[8] = res1_8;
- C1[9] = res1_9;
- C1[10] = res1_10;
- C1[11] = res1_11;
- C1[12] = res1_12;
- C1[13] = res1_13;
- C1[14] = res1_14;
- C1[15] = res1_15;
-
- C2[0] = res2_0;
- C2[1] = res2_1;
- C2[2] = res2_2;
- C2[3] = res2_3;
- C2[4] = res2_4;
- C2[5] = res2_5;
- C2[6] = res2_6;
- C2[7] = res2_7;
-
- C2[8] = res2_8;
- C2[9] = res2_9;
- C2[10] = res2_10;
- C2[11] = res2_11;
- C2[12] = res2_12;
- C2[13] = res2_13;
- C2[14] = res2_14;
- C2[15] = res2_15;
-
- C3[0] = res3_0;
- C3[1] = res3_1;
- C3[2] = res3_2;
- C3[3] = res3_3;
- C3[4] = res3_4;
- C3[5] = res3_5;
- C3[6] = res3_6;
- C3[7] = res3_7;
-
- C3[8] = res3_8;
- C3[9] = res3_9;
- C3[10] = res3_10;
- C3[11] = res3_11;
- C3[12] = res3_12;
- C3[13] = res3_13;
- C3[14] = res3_14;
- C3[15] = res3_15;
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 16; // number of values in A
- #else
- temp -= 4; // number of values in B
- #endif
- ptrba += temp*16;
- ptrbb += temp*4;
- #endif
-
- #ifdef LEFT
- off += 16; // number of values in A
- #endif
-
- C0 = C0+16;
- C1 = C1+16;
- C2 = C2+16;
- C3 = C3+16;
- }
-
-
- if ( bm & 8)
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*8;
- ptrbb = bb + off*4;
- #endif
-
- res0_0 = 0;
- res0_1 = 0;
- res0_2 = 0;
- res0_3 = 0;
- res0_4 = 0;
- res0_5 = 0;
- res0_6 = 0;
- res0_7 = 0;
-
- res1_0 = 0;
- res1_1 = 0;
- res1_2 = 0;
- res1_3 = 0;
- res1_4 = 0;
- res1_5 = 0;
- res1_6 = 0;
- res1_7 = 0;
-
- res2_0 = 0;
- res2_1 = 0;
- res2_2 = 0;
- res2_3 = 0;
- res2_4 = 0;
- res2_5 = 0;
- res2_6 = 0;
- res2_7 = 0;
-
- res3_0 = 0;
- res3_1 = 0;
- res3_2 = 0;
- res3_3 = 0;
- res3_4 = 0;
- res3_5 = 0;
- res3_6 = 0;
- res3_7 = 0;
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+8; // number of values in A
- #else
- temp = off+4; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
- b2 = ptrbb[2];
- b3 = ptrbb[3];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
- res2_0 += a0*b2;
- res3_0 += a0*b3;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
- res1_1 += a1*b1;
- res2_1 += a1*b2;
- res3_1 += a1*b3;
-
- a0 = ptrba[2];
- res0_2 += a0*b0;
- res1_2 += a0*b1;
- res2_2 += a0*b2;
- res3_2 += a0*b3;
-
- a1 = ptrba[3];
- res0_3 += a1*b0;
- res1_3 += a1*b1;
- res2_3 += a1*b2;
- res3_3 += a1*b3;
-
- a0 = ptrba[4];
- res0_4 += a0*b0;
- res1_4 += a0*b1;
- res2_4 += a0*b2;
- res3_4 += a0*b3;
-
- a1 = ptrba[5];
- res0_5 += a1*b0;
- res1_5 += a1*b1;
- res2_5 += a1*b2;
- res3_5 += a1*b3;
-
- a0 = ptrba[6];
- res0_6 += a0*b0;
- res1_6 += a0*b1;
- res2_6 += a0*b2;
- res3_6 += a0*b3;
-
- a1 = ptrba[7];
- res0_7 += a1*b0;
- res1_7 += a1*b1;
- res2_7 += a1*b2;
- res3_7 += a1*b3;
-
- ptrba = ptrba+8;
- ptrbb = ptrbb+4;
-
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
- res0_2 *= alpha;
- res0_3 *= alpha;
- res0_4 *= alpha;
- res0_5 *= alpha;
- res0_6 *= alpha;
- res0_7 *= alpha;
-
- res1_0 *= alpha;
- res1_1 *= alpha;
- res1_2 *= alpha;
- res1_3 *= alpha;
- res1_4 *= alpha;
- res1_5 *= alpha;
- res1_6 *= alpha;
- res1_7 *= alpha;
-
- res2_0 *= alpha;
- res2_1 *= alpha;
- res2_2 *= alpha;
- res2_3 *= alpha;
- res2_4 *= alpha;
- res2_5 *= alpha;
- res2_6 *= alpha;
- res2_7 *= alpha;
-
- res3_0 *= alpha;
- res3_1 *= alpha;
- res3_2 *= alpha;
- res3_3 *= alpha;
- res3_4 *= alpha;
- res3_5 *= alpha;
- res3_6 *= alpha;
- res3_7 *= alpha;
-
- C0[0] = res0_0;
- C0[1] = res0_1;
- C0[2] = res0_2;
- C0[3] = res0_3;
- C0[4] = res0_4;
- C0[5] = res0_5;
- C0[6] = res0_6;
- C0[7] = res0_7;
-
- C1[0] = res1_0;
- C1[1] = res1_1;
- C1[2] = res1_2;
- C1[3] = res1_3;
- C1[4] = res1_4;
- C1[5] = res1_5;
- C1[6] = res1_6;
- C1[7] = res1_7;
-
- C2[0] = res2_0;
- C2[1] = res2_1;
- C2[2] = res2_2;
- C2[3] = res2_3;
- C2[4] = res2_4;
- C2[5] = res2_5;
- C2[6] = res2_6;
- C2[7] = res2_7;
-
- C3[0] = res3_0;
- C3[1] = res3_1;
- C3[2] = res3_2;
- C3[3] = res3_3;
- C3[4] = res3_4;
- C3[5] = res3_5;
- C3[6] = res3_6;
- C3[7] = res3_7;
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 8; // number of values in A
- #else
- temp -= 4; // number of values in B
- #endif
- ptrba += temp*8;
- ptrbb += temp*4;
- #endif
-
- #ifdef LEFT
- off += 8; // number of values in A
- #endif
-
- C0 = C0+8;
- C1 = C1+8;
- C2 = C2+8;
- C3 = C3+8;
- }
-
- if ( bm & 4 )
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*4;
- ptrbb = bb + off*4;
- #endif
-
- res0_0 = 0;
- res0_1 = 0;
- res0_2 = 0;
- res0_3 = 0;
-
- res1_0 = 0;
- res1_1 = 0;
- res1_2 = 0;
- res1_3 = 0;
-
- res2_0 = 0;
- res2_1 = 0;
- res2_2 = 0;
- res2_3 = 0;
-
- res3_0 = 0;
- res3_1 = 0;
- res3_2 = 0;
- res3_3 = 0;
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+4; // number of values in A
- #else
- temp = off+4; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
- b2 = ptrbb[2];
- b3 = ptrbb[3];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
- res2_0 += a0*b2;
- res3_0 += a0*b3;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
- res1_1 += a1*b1;
- res2_1 += a1*b2;
- res3_1 += a1*b3;
-
- a0 = ptrba[2];
- res0_2 += a0*b0;
- res1_2 += a0*b1;
- res2_2 += a0*b2;
- res3_2 += a0*b3;
-
- a1 = ptrba[3];
- res0_3 += a1*b0;
- res1_3 += a1*b1;
- res2_3 += a1*b2;
- res3_3 += a1*b3;
-
- ptrba = ptrba+4;
- ptrbb = ptrbb+4;
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
- res0_2 *= alpha;
- res0_3 *= alpha;
-
- res1_0 *= alpha;
- res1_1 *= alpha;
- res1_2 *= alpha;
- res1_3 *= alpha;
-
- res2_0 *= alpha;
- res2_1 *= alpha;
- res2_2 *= alpha;
- res2_3 *= alpha;
-
- res3_0 *= alpha;
- res3_1 *= alpha;
- res3_2 *= alpha;
- res3_3 *= alpha;
-
- C0[0] = res0_0;
- C0[1] = res0_1;
- C0[2] = res0_2;
- C0[3] = res0_3;
-
- C1[0] = res1_0;
- C1[1] = res1_1;
- C1[2] = res1_2;
- C1[3] = res1_3;
-
-
- C2[0] = res2_0;
- C2[1] = res2_1;
- C2[2] = res2_2;
- C2[3] = res2_3;
-
- C3[0] = res3_0;
- C3[1] = res3_1;
- C3[2] = res3_2;
- C3[3] = res3_3;
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 4; // number of values in A
- #else
- temp -= 4; // number of values in B
- #endif
- ptrba += temp*4;
- ptrbb += temp*4;
- #endif
-
- #ifdef LEFT
- off += 4; // number of values in A
- #endif
-
- C0 = C0+4;
- C1 = C1+4;
- C2 = C2+4;
- C3 = C3+4;
- }
-
- if ( bm & 2 )
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*2;
- ptrbb = bb + off*4;
- #endif
-
- res0_0 = 0;
- res0_1 = 0;
-
- res1_0 = 0;
- res1_1 = 0;
-
- res2_0 = 0;
- res2_1 = 0;
-
- res3_0 = 0;
- res3_1 = 0;
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+2; // number of values in A
- #else
- temp = off+4; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
- b2 = ptrbb[2];
- b3 = ptrbb[3];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
- res2_0 += a0*b2;
- res3_0 += a0*b3;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
- res1_1 += a1*b1;
- res2_1 += a1*b2;
- res3_1 += a1*b3;
-
- ptrba = ptrba+2;
- ptrbb = ptrbb+4;
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
-
- res1_0 *= alpha;
- res1_1 *= alpha;
-
- res2_0 *= alpha;
- res2_1 *= alpha;
-
- res3_0 *= alpha;
- res3_1 *= alpha;
-
- C0[0] = res0_0;
- C0[1] = res0_1;
-
- C1[0] = res1_0;
- C1[1] = res1_1;
-
- C2[0] = res2_0;
- C2[1] = res2_1;
-
- C3[0] = res3_0;
- C3[1] = res3_1;
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 2; // number of values in A
- #else
- temp -= 4; // number of values in B
- #endif
- ptrba += temp*2;
- ptrbb += temp*4;
- #endif
-
- #ifdef LEFT
- off += 2; // number of values in A
- #endif
-
- C0 = C0+2;
- C1 = C1+2;
- C2 = C2+2;
- C3 = C3+2;
- }
-
- if ( bm & 1 )
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*1;
- ptrbb = bb + off*4;
- #endif
-
- res0_0 = 0;
- res1_0 = 0;
- res2_0 = 0;
- res3_0 = 0;
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+1; // number of values in A
- #else
- temp = off+4; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
- b2 = ptrbb[2];
- b3 = ptrbb[3];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
- res2_0 += a0*b2;
- res3_0 += a0*b3;
-
- ptrba = ptrba+1;
- ptrbb = ptrbb+4;
- }
- res0_0 *= alpha;
- res1_0 *= alpha;
- res2_0 *= alpha;
- res3_0 *= alpha;
-
- C0[0] = res0_0;
- C1[0] = res1_0;
- C2[0] = res2_0;
- C3[0] = res3_0;
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 1; // number of values in A
- #else
- temp -= 4; // number of values in B
- #endif
- ptrba += temp*1;
- ptrbb += temp*4;
- #endif
-
- #ifdef LEFT
- off += 1; // number of values in A
- #endif
-
- C0 = C0+1;
- C1 = C1+1;
- C2 = C2+1;
- C3 = C3+1;
-
- }
-
-
- #if defined(TRMMKERNEL) && !defined(LEFT)
- off += 4;
- #endif
-
- k = (bk<<2);
- bb = bb+k;
- i = (ldc<<2);
- C = C+i;
- }
-
-
- if(bn&2)
- {
- C0 = C;
- C1 = C0+ldc;
-
- #if defined(TRMMKERNEL) && defined(LEFT)
- off = offset;
- #endif
-
-
- ptrba = ba;
-
-
- for (i=0; i<bm/16; i+=1)
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*16;
- ptrbb = bb + off*2;
- #endif
-
- res0_0 = 0;
- res0_1 = 0;
- res0_2 = 0;
- res0_3 = 0;
- res0_4 = 0;
- res0_5 = 0;
- res0_6 = 0;
- res0_7 = 0;
-
- res0_8 = 0;
- res0_9 = 0;
- res0_10 = 0;
- res0_11 = 0;
- res0_12 = 0;
- res0_13 = 0;
- res0_14 = 0;
- res0_15 = 0;
-
- res1_0 = 0;
- res1_1 = 0;
- res1_2 = 0;
- res1_3 = 0;
- res1_4 = 0;
- res1_5 = 0;
- res1_6 = 0;
- res1_7 = 0;
-
- res1_8 = 0;
- res1_9 = 0;
- res1_10 = 0;
- res1_11 = 0;
- res1_12 = 0;
- res1_13 = 0;
- res1_14 = 0;
- res1_15 = 0;
-
-
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+16; // number of values in A
- #else
- temp = off+2; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
- res1_1 += a1*b1;
-
- a0 = ptrba[2];
- res0_2 += a0*b0;
- res1_2 += a0*b1;
-
- a1 = ptrba[3];
- res0_3 += a1*b0;
- res1_3 += a1*b1;
-
- a0 = ptrba[4];
- res0_4 += a0*b0;
- res1_4 += a0*b1;
-
- a1 = ptrba[5];
- res0_5 += a1*b0;
- res1_5 += a1*b1;
-
- a0 = ptrba[6];
- res0_6 += a0*b0;
- res1_6 += a0*b1;
-
- a1 = ptrba[7];
- res0_7 += a1*b0;
- res1_7 += a1*b1;
-
- a0 = ptrba[8];
- res0_8 += a0*b0;
- res1_8 += a0*b1;
-
- a1 = ptrba[9];
- res0_9 += a1*b0;
- res1_9 += a1*b1;
-
- a0 = ptrba[10];
- res0_10 += a0*b0;
- res1_10 += a0*b1;
-
- a1 = ptrba[11];
- res0_11 += a1*b0;
- res1_11 += a1*b1;
-
- a0 = ptrba[12];
- res0_12 += a0*b0;
- res1_12 += a0*b1;
-
- a1 = ptrba[13];
- res0_13 += a1*b0;
- res1_13 += a1*b1;
-
- a0 = ptrba[14];
- res0_14 += a0*b0;
- res1_14 += a0*b1;
-
- a1 = ptrba[15];
- res0_15 += a1*b0;
- res1_15 += a1*b1;
-
-
- ptrba = ptrba+16;
- ptrbb = ptrbb+2;
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
- res0_2 *= alpha;
- res0_3 *= alpha;
- res0_4 *= alpha;
- res0_5 *= alpha;
- res0_6 *= alpha;
- res0_7 *= alpha;
-
- res0_8 *= alpha;
- res0_9 *= alpha;
- res0_10 *= alpha;
- res0_11 *= alpha;
- res0_12 *= alpha;
- res0_13 *= alpha;
- res0_14 *= alpha;
- res0_15 *= alpha;
-
- res1_0 *= alpha;
- res1_1 *= alpha;
- res1_2 *= alpha;
- res1_3 *= alpha;
- res1_4 *= alpha;
- res1_5 *= alpha;
- res1_6 *= alpha;
- res1_7 *= alpha;
-
- res1_8 *= alpha;
- res1_9 *= alpha;
- res1_10 *= alpha;
- res1_11 *= alpha;
- res1_12 *= alpha;
- res1_13 *= alpha;
- res1_14 *= alpha;
- res1_15 *= alpha;
-
- C0[0] = res0_0;
- C0[1] = res0_1;
- C0[2] = res0_2;
- C0[3] = res0_3;
- C0[4] = res0_4;
- C0[5] = res0_5;
- C0[6] = res0_6;
- C0[7] = res0_7;
-
- C0[8] = res0_8;
- C0[9] = res0_9;
- C0[10] = res0_10;
- C0[11] = res0_11;
- C0[12] = res0_12;
- C0[13] = res0_13;
- C0[14] = res0_14;
- C0[15] = res0_15;
-
- C1[0] = res1_0;
- C1[1] = res1_1;
- C1[2] = res1_2;
- C1[3] = res1_3;
- C1[4] = res1_4;
- C1[5] = res1_5;
- C1[6] = res1_6;
- C1[7] = res1_7;
-
- C1[8] = res1_8;
- C1[9] = res1_9;
- C1[10] = res1_10;
- C1[11] = res1_11;
- C1[12] = res1_12;
- C1[13] = res1_13;
- C1[14] = res1_14;
- C1[15] = res1_15;
-
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 16; // number of values in A
- #else
- temp -= 2; // number of values in B
- #endif
- ptrba += temp*16;
- ptrbb += temp*2;
- #endif
-
- #ifdef LEFT
- off += 16; // number of values in A
- #endif
-
- C0 = C0+16;
- C1 = C1+16;
- }
-
-
-
-
- if ( bm & 8)
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*8;
- ptrbb = bb + off*2;
- #endif
-
- res0_0 = 0;
- res0_1 = 0;
- res0_2 = 0;
- res0_3 = 0;
- res0_4 = 0;
- res0_5 = 0;
- res0_6 = 0;
- res0_7 = 0;
-
- res1_0 = 0;
- res1_1 = 0;
- res1_2 = 0;
- res1_3 = 0;
- res1_4 = 0;
- res1_5 = 0;
- res1_6 = 0;
- res1_7 = 0;
-
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+8; // number of values in A
- #else
- temp = off+2; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
- res1_1 += a1*b1;
-
- a0 = ptrba[2];
- res0_2 += a0*b0;
- res1_2 += a0*b1;
-
- a1 = ptrba[3];
- res0_3 += a1*b0;
- res1_3 += a1*b1;
-
- a0 = ptrba[4];
- res0_4 += a0*b0;
- res1_4 += a0*b1;
-
- a1 = ptrba[5];
- res0_5 += a1*b0;
- res1_5 += a1*b1;
-
- a0 = ptrba[6];
- res0_6 += a0*b0;
- res1_6 += a0*b1;
-
- a1 = ptrba[7];
- res0_7 += a1*b0;
- res1_7 += a1*b1;
-
- ptrba = ptrba+8;
- ptrbb = ptrbb+2;
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
- res0_2 *= alpha;
- res0_3 *= alpha;
- res0_4 *= alpha;
- res0_5 *= alpha;
- res0_6 *= alpha;
- res0_7 *= alpha;
-
- res1_0 *= alpha;
- res1_1 *= alpha;
- res1_2 *= alpha;
- res1_3 *= alpha;
- res1_4 *= alpha;
- res1_5 *= alpha;
- res1_6 *= alpha;
- res1_7 *= alpha;
-
- C0[0] = res0_0;
- C0[1] = res0_1;
- C0[2] = res0_2;
- C0[3] = res0_3;
- C0[4] = res0_4;
- C0[5] = res0_5;
- C0[6] = res0_6;
- C0[7] = res0_7;
-
- C1[0] = res1_0;
- C1[1] = res1_1;
- C1[2] = res1_2;
- C1[3] = res1_3;
- C1[4] = res1_4;
- C1[5] = res1_5;
- C1[6] = res1_6;
- C1[7] = res1_7;
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 8; // number of values in A
- #else
- temp -= 2; // number of values in B
- #endif
- ptrba += temp*8;
- ptrbb += temp*2;
- #endif
-
- #ifdef LEFT
- off += 8; // number of values in A
- #endif
-
- C0 = C0+8;
- C1 = C1+8;
- }
-
- if ( bm & 4 )
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*4;
- ptrbb = bb + off*2;
- #endif
-
- res0_0 = 0;
- res0_1 = 0;
- res0_2 = 0;
- res0_3 = 0;
-
- res1_0 = 0;
- res1_1 = 0;
- res1_2 = 0;
- res1_3 = 0;
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+4; // number of values in A
- #else
- temp = off+2; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
- res1_1 += a1*b1;
-
- a0 = ptrba[2];
- res0_2 += a0*b0;
- res1_2 += a0*b1;
-
- a1 = ptrba[3];
- res0_3 += a1*b0;
- res1_3 += a1*b1;
-
- ptrba = ptrba+4;
- ptrbb = ptrbb+2;
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
- res0_2 *= alpha;
- res0_3 *= alpha;
-
- res1_0 *= alpha;
- res1_1 *= alpha;
- res1_2 *= alpha;
- res1_3 *= alpha;
-
- C0[0] = res0_0;
- C0[1] = res0_1;
- C0[2] = res0_2;
- C0[3] = res0_3;
-
- C1[0] = res1_0;
- C1[1] = res1_1;
- C1[2] = res1_2;
- C1[3] = res1_3;
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 4; // number of values in A
- #else
- temp -= 2; // number of values in B
- #endif
- ptrba += temp*4;
- ptrbb += temp*2;
- #endif
-
- #ifdef LEFT
- off += 4; // number of values in A
- #endif
-
- C0 = C0+4;
- C1 = C1+4;
-
- }
-
- if ( bm & 2 )
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*2;
- ptrbb = bb + off*2;
- #endif
-
- res0_0 = 0;
- res0_1 = 0;
-
- res1_0 = 0;
- res1_1 = 0;
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+2; // number of values in A
- #else
- temp = off+2; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
- res1_1 += a1*b1;
-
- ptrba = ptrba+2;
- ptrbb = ptrbb+2;
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
-
- res1_0 *= alpha;
- res1_1 *= alpha;
-
- C0[0] = res0_0;
- C0[1] = res0_1;
-
- C1[0] = res1_0;
- C1[1] = res1_1;
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 2; // number of values in A
- #else
- temp -= 2; // number of values in B
- #endif
- ptrba += temp*2;
- ptrbb += temp*2;
- #endif
-
- #ifdef LEFT
- off += 2; // number of values in A
- #endif
-
- C0 = C0+2;
- C1 = C1+2;
-
- }
-
- if ( bm & 1 )
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*1;
- ptrbb = bb + off*2;
- #endif
-
- res0_0 = 0;
-
- res1_0 = 0;
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+1; // number of values in A
- #else
- temp = off+2; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
-
- ptrba = ptrba+1;
- ptrbb = ptrbb+2;
- }
-
- res0_0 *= alpha;
-
- res1_0 *= alpha;
-
- C0[0] = res0_0;
-
- C1[0] = res1_0;
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 1; // number of values in A
- #else
- temp -= 2; // number of values in B
- #endif
- ptrba += temp*1;
- ptrbb += temp*2;
- #endif
-
- #ifdef LEFT
- off += 1; // number of values in A
- #endif
-
- C0 = C0+1;
- C1 = C1+1;
-
- }
-
-
- #if defined(TRMMKERNEL) && !defined(LEFT)
- off += 2;
- #endif
-
- k = (bk<<1);
- bb = bb+k;
- i = (ldc<<1);
- C = C+i;
- }
-
-
- for (j=0; j<(bn&1); j+=1)
- {
- C0 = C;
-
- #if defined(TRMMKERNEL) && defined(LEFT)
- off = offset;
- #endif
-
- ptrba = ba;
-
-
- for (i=0; i<bm/16; i+=1)
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*16;
- ptrbb = bb + off*1;
- #endif
-
- res0_0 = 0;
- res0_1 = 0;
- res0_2 = 0;
- res0_3 = 0;
- res0_4 = 0;
- res0_5 = 0;
- res0_6 = 0;
- res0_7 = 0;
-
- res0_8 = 0;
- res0_9 = 0;
- res0_10 = 0;
- res0_11 = 0;
- res0_12 = 0;
- res0_13 = 0;
- res0_14 = 0;
- res0_15 = 0;
-
-
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+16; // number of values in A
- #else
- temp = off+1; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
-
- a0 = ptrba[2];
- res0_2 += a0*b0;
-
- a1 = ptrba[3];
- res0_3 += a1*b0;
-
- a0 = ptrba[4];
- res0_4 += a0*b0;
-
- a1 = ptrba[5];
- res0_5 += a1*b0;
-
- a0 = ptrba[6];
- res0_6 += a0*b0;
-
- a1 = ptrba[7];
- res0_7 += a1*b0;
-
- a0 = ptrba[8];
- res0_8 += a0*b0;
-
- a1 = ptrba[9];
- res0_9 += a1*b0;
-
- a0 = ptrba[10];
- res0_10 += a0*b0;
-
- a1 = ptrba[11];
- res0_11 += a1*b0;
-
- a0 = ptrba[12];
- res0_12 += a0*b0;
-
- a1 = ptrba[13];
- res0_13 += a1*b0;
-
- a0 = ptrba[14];
- res0_14 += a0*b0;
-
- a1 = ptrba[15];
- res0_15 += a1*b0;
-
-
- ptrba = ptrba+16;
- ptrbb = ptrbb+1;
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
- res0_2 *= alpha;
- res0_3 *= alpha;
- res0_4 *= alpha;
- res0_5 *= alpha;
- res0_6 *= alpha;
- res0_7 *= alpha;
-
- res0_8 *= alpha;
- res0_9 *= alpha;
- res0_10 *= alpha;
- res0_11 *= alpha;
- res0_12 *= alpha;
- res0_13 *= alpha;
- res0_14 *= alpha;
- res0_15 *= alpha;
-
- C0[0] = res0_0;
- C0[1] = res0_1;
- C0[2] = res0_2;
- C0[3] = res0_3;
- C0[4] = res0_4;
- C0[5] = res0_5;
- C0[6] = res0_6;
- C0[7] = res0_7;
-
- C0[8] = res0_8;
- C0[9] = res0_9;
- C0[10] = res0_10;
- C0[11] = res0_11;
- C0[12] = res0_12;
- C0[13] = res0_13;
- C0[14] = res0_14;
- C0[15] = res0_15;
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 16; // number of values in A
- #else
- temp -= 1; // number of values in B
- #endif
- ptrba += temp*16;
- ptrbb += temp*1;
- #endif
-
- #ifdef LEFT
- off += 16; // number of values in A
- #endif
-
- C0 = C0+16;
- }
-
-
-
-
- if ( bm & 8 )
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*8;
- ptrbb = bb + off*1;
- #endif
-
- res0_0 = 0;
- res0_1 = 0;
- res0_2 = 0;
- res0_3 = 0;
- res0_4 = 0;
- res0_5 = 0;
- res0_6 = 0;
- res0_7 = 0;
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+8; // number of values in A
- #else
- temp = off+1; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
-
- a0 = ptrba[2];
- res0_2 += a0*b0;
-
- a1 = ptrba[3];
- res0_3 += a1*b0;
-
- a0 = ptrba[4];
- res0_4 += a0*b0;
-
- a1 = ptrba[5];
- res0_5 += a1*b0;
-
- a0 = ptrba[6];
- res0_6 += a0*b0;
-
- a1 = ptrba[7];
- res0_7 += a1*b0;
-
- ptrba = ptrba+8;
- ptrbb = ptrbb+1;
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
- res0_2 *= alpha;
- res0_3 *= alpha;
- res0_4 *= alpha;
- res0_5 *= alpha;
- res0_6 *= alpha;
- res0_7 *= alpha;
-
- C0[0] = res0_0;
- C0[1] = res0_1;
- C0[2] = res0_2;
- C0[3] = res0_3;
- C0[4] = res0_4;
- C0[5] = res0_5;
- C0[6] = res0_6;
- C0[7] = res0_7;
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 8; // number of values in A
- #else
- temp -= 1; // number of values in B
- #endif
- ptrba += temp*8;
- ptrbb += temp*1;
- #endif
-
- #ifdef LEFT
- off += 8; // number of values in A
- #endif
-
- C0 = C0+8;
- }
-
- if ( bm & 4 )
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*4;
- ptrbb = bb + off*1;
- #endif
-
- res0_0 = 0;
- res0_1 = 0;
- res0_2 = 0;
- res0_3 = 0;
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+4; // number of values in A
- #else
- temp = off+1; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
-
- a0 = ptrba[2];
- res0_2 += a0*b0;
-
- a1 = ptrba[3];
- res0_3 += a1*b0;
-
- ptrba = ptrba+4;
- ptrbb = ptrbb+1;
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
- res0_2 *= alpha;
- res0_3 *= alpha;
-
- C0[0] = res0_0;
- C0[1] = res0_1;
- C0[2] = res0_2;
- C0[3] = res0_3;
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 4; // number of values in A
- #else
- temp -= 1; // number of values in B
- #endif
- ptrba += temp*4;
- ptrbb += temp*1;
- #endif
-
- #ifdef LEFT
- off += 4; // number of values in A
- #endif
-
- C0 = C0+4;
-
- }
-
- if ( bm & 2 )
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*2;
- ptrbb = bb + off*1;
- #endif
-
- res0_0 = 0;
- res0_1 = 0;
-
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+2; // number of values in A
- #else
- temp = off+1; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
-
- ptrba = ptrba+2;
- ptrbb = ptrbb+1;
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
-
- C0[0] = res0_0;
- C0[1] = res0_1;
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 2; // number of values in A
- #else
- temp -= 1; // number of values in B
- #endif
- ptrba += temp*2;
- ptrbb += temp*1;
- #endif
-
- #ifdef LEFT
- off += 2; // number of values in A
- #endif
-
- C0 = C0+2;
-
- }
-
- if ( bm & 1 )
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*1;
- ptrbb = bb + off*1;
- #endif
-
- res0_0 = 0;
-
-
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- temp = bk-off;
- #elif defined(LEFT)
- temp = off+1; // number of values in A
- #else
- temp = off+1; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
-
- ptrba = ptrba+1;
- ptrbb = ptrbb+1;
- }
-
- res0_0 *= alpha;
-
- C0[0] = res0_0;
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 1; // number of values in A
- #else
- temp -= 1; // number of values in B
- #endif
- ptrba += temp*1;
- ptrbb += temp*1;
- #endif
-
- #ifdef LEFT
- off += 1; // number of values in A
- #endif
-
- C0 = C0+1;
-
- }
-
-
-
- #if defined(TRMMKERNEL) && !defined(LEFT)
- off += 1;
- #endif
-
- k = (bk<<0);
- bb = bb+k;
- C = C+ldc;
- }
- return 0;
- }
|