|
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402 |
- #include "common.h"
- #include <stdbool.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,*C4,*C5,*C6,*C7,*ptrba,*ptrbb;
-
- FLOAT res0_0;
- FLOAT res0_1;
- FLOAT res0_2;
- FLOAT res0_3;
-
- FLOAT res1_0;
- FLOAT res1_1;
- FLOAT res1_2;
- FLOAT res1_3;
-
- FLOAT res2_0;
- FLOAT res2_1;
- FLOAT res2_2;
- FLOAT res2_3;
-
- FLOAT res3_0;
- FLOAT res3_1;
- FLOAT res3_2;
- FLOAT res3_3;
-
- FLOAT res4_0;
- FLOAT res4_1;
- FLOAT res4_2;
- FLOAT res4_3;
-
- FLOAT res5_0;
- FLOAT res5_1;
- FLOAT res5_2;
- FLOAT res5_3;
-
- FLOAT res6_0;
- FLOAT res6_1;
- FLOAT res6_2;
- FLOAT res6_3;
-
- FLOAT res7_0;
- FLOAT res7_1;
- FLOAT res7_2;
- FLOAT res7_3;
-
- FLOAT a0;
- FLOAT a1;
-
- FLOAT b0;
- FLOAT b1;
- FLOAT b2;
- FLOAT b3;
- FLOAT b4;
- FLOAT b5;
- FLOAT b6;
- FLOAT b7;
-
- BLASLONG off, temp;
-
- bool left;
- bool transposed;
- bool backwards;
-
- #ifdef LEFT
- left = true;
- #else
- left = false;
- #endif
-
- #ifdef TRANSA
- transposed = true;
- #else
- transposed = false;
- #endif
-
- backwards = left != transposed;
-
- if (!left) {
- off = -offset;
- }
-
-
- for (j=0; j<bn/8; j+=1) // do blocks of the Mx8 loops
- {
- C0 = C;
- C1 = C0+ldc;
- C2 = C1+ldc;
- C3 = C2+ldc;
- C4 = C3+ldc;
- C5 = C4+ldc;
- C6 = C5+ldc;
- C7 = C6+ldc;
-
-
- if (left) {
- off = offset;
- }
-
- ptrba = ba;
-
- for (i=0; i<bm/4; i+=1) // do blocks of 4x4
- {
-
- ptrbb = bb;
- if (backwards)
- {
- ptrba += off*4; // number of values in A
- ptrbb += off*8; // number of values in B
- }
-
- 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;
-
- res4_0 = 0;
- res4_1 = 0;
- res4_2 = 0;
- res4_3 = 0;
-
- res5_0 = 0;
- res5_1 = 0;
- res5_2 = 0;
- res5_3 = 0;
-
- res6_0 = 0;
- res6_1 = 0;
- res6_2 = 0;
- res6_3 = 0;
-
- res7_0 = 0;
- res7_1 = 0;
- res7_2 = 0;
- res7_3 = 0;
-
- temp = backwards ? bk-off :
- left ? off + 4 : // number of values in A
- off + 8; // number of values in B
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
- b2 = ptrbb[2];
- b3 = ptrbb[3];
- b4 = ptrbb[4];
- b5 = ptrbb[5];
- b6 = ptrbb[6];
- b7 = ptrbb[7];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
- res2_0 += a0*b2;
- res3_0 += a0*b3;
- res4_0 += a0*b4;
- res5_0 += a0*b5;
- res6_0 += a0*b6;
- res7_0 += a0*b7;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
- res1_1 += a1*b1;
- res2_1 += a1*b2;
- res3_1 += a1*b3;
- res4_1 += a1*b4;
- res5_1 += a1*b5;
- res6_1 += a1*b6;
- res7_1 += a1*b7;
-
- a0 = ptrba[2];
- res0_2 += a0*b0;
- res1_2 += a0*b1;
- res2_2 += a0*b2;
- res3_2 += a0*b3;
- res4_2 += a0*b4;
- res5_2 += a0*b5;
- res6_2 += a0*b6;
- res7_2 += a0*b7;
-
- a1 = ptrba[3];
- res0_3 += a1*b0;
- res1_3 += a1*b1;
- res2_3 += a1*b2;
- res3_3 += a1*b3;
- res4_3 += a1*b4;
- res5_3 += a1*b5;
- res6_3 += a1*b6;
- res7_3 += a1*b7;
-
- ptrba = ptrba+4;
- ptrbb = ptrbb+8;
- }
-
- 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;
-
- res4_0 *= alpha;
- res4_1 *= alpha;
- res4_2 *= alpha;
- res4_3 *= alpha;
-
- res5_0 *= alpha;
- res5_1 *= alpha;
- res5_2 *= alpha;
- res5_3 *= alpha;
-
- res6_0 *= alpha;
- res6_1 *= alpha;
- res6_2 *= alpha;
- res6_3 *= alpha;
-
- res7_0 *= alpha;
- res7_1 *= alpha;
- res7_2 *= alpha;
- res7_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;
-
- C4[0] = res4_0;
- C4[1] = res4_1;
- C4[2] = res4_2;
- C4[3] = res4_3;
-
- C5[0] = res5_0;
- C5[1] = res5_1;
- C5[2] = res5_2;
- C5[3] = res5_3;
-
- C6[0] = res6_0;
- C6[1] = res6_1;
- C6[2] = res6_2;
- C6[3] = res6_3;
-
- C7[0] = res7_0;
- C7[1] = res7_1;
- C7[2] = res7_2;
- C7[3] = res7_3;
-
- if (!backwards) {
- temp = bk-off;
- temp = left ? temp - 4 : // number of values in A
- temp - 8; // number of values in B
-
- ptrba += temp*4; // number of values in A
- ptrbb += temp*8; // number of values in B
- }
- #ifdef LEFT
- off += 4; // number of values in A
- #endif
-
- C0 = C0+4;
- C1 = C1+4;
- C2 = C2+4;
- C3 = C3+4;
- C4 = C4+4;
- C5 = C5+4;
- C6 = C6+4;
- C7 = C7+4;
-
- }
-
- if ( bm & 2 ) // do any 2x4 loop
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*2;
- ptrbb = bb + off*8;
- #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;
-
- res4_0 = 0;
- res4_1 = 0;
-
- res5_0 = 0;
- res5_1 = 0;
-
- res6_0 = 0;
- res6_1 = 0;
-
- res7_0 = 0;
- res7_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+8; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
- b2 = ptrbb[2];
- b3 = ptrbb[3];
- b4 = ptrbb[4];
- b5 = ptrbb[5];
- b6 = ptrbb[6];
- b7 = ptrbb[7];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
- res2_0 += a0*b2;
- res3_0 += a0*b3;
- res4_0 += a0*b4;
- res5_0 += a0*b5;
- res6_0 += a0*b6;
- res7_0 += a0*b7;
-
- a1 = ptrba[1];
- res0_1 += a1*b0;
- res1_1 += a1*b1;
- res2_1 += a1*b2;
- res3_1 += a1*b3;
- res4_1 += a1*b4;
- res5_1 += a1*b5;
- res6_1 += a1*b6;
- res7_1 += a1*b7;
-
- ptrba = ptrba+2;
- ptrbb = ptrbb+8;
- }
-
- res0_0 *= alpha;
- res0_1 *= alpha;
-
- res1_0 *= alpha;
- res1_1 *= alpha;
-
- res2_0 *= alpha;
- res2_1 *= alpha;
-
- res3_0 *= alpha;
- res3_1 *= alpha;
-
- res4_0 *= alpha;
- res4_1 *= alpha;
-
- res5_0 *= alpha;
- res5_1 *= alpha;
-
- res6_0 *= alpha;
- res6_1 *= alpha;
-
- res7_0 *= alpha;
- res7_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;
-
- C4[0] = res4_0;
- C4[1] = res4_1;
-
- C5[0] = res5_0;
- C5[1] = res5_1;
-
- C6[0] = res6_0;
- C6[1] = res6_1;
-
- C7[0] = res7_0;
- C7[1] = res7_1;
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 2; // number of values in A
- #else
- temp -= 8; // number of values in B
- #endif
- ptrba += temp*2;
- ptrbb += temp*8;
- #endif
-
- #ifdef LEFT
- off += 2; // number of values in A
- #endif
-
- C0 = C0+2;
- C1 = C1+2;
- C2 = C2+2;
- C3 = C3+2;
- C4 = C4+2;
- C5 = C5+2;
- C6 = C6+2;
- C7 = C7+2;
-
- }
-
- if ( bm & 1 ) // do any 1x4 loop
- {
-
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- ptrbb = bb;
- #else
- ptrba += off*1;
- ptrbb = bb + off*8;
- #endif
-
- res0_0 = 0;
- res1_0 = 0;
- res2_0 = 0;
- res3_0 = 0;
- res4_0 = 0;
- res5_0 = 0;
- res6_0 = 0;
- res7_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+8; // number of values in B
- #endif
-
- for (k=0; k<temp; k++)
- {
- b0 = ptrbb[0];
- b1 = ptrbb[1];
- b2 = ptrbb[2];
- b3 = ptrbb[3];
- b4 = ptrbb[4];
- b5 = ptrbb[5];
- b6 = ptrbb[6];
- b7 = ptrbb[7];
-
- a0 = ptrba[0];
- res0_0 += a0*b0;
- res1_0 += a0*b1;
- res2_0 += a0*b2;
- res3_0 += a0*b3;
- res4_0 += a0*b4;
- res5_0 += a0*b5;
- res6_0 += a0*b6;
- res7_0 += a0*b7;
-
- ptrba = ptrba+1;
- ptrbb = ptrbb+8;
- }
-
- res0_0 *= alpha;
-
- res1_0 *= alpha;
-
- res2_0 *= alpha;
-
- res3_0 *= alpha;
- res4_0 *= alpha;
- res5_0 *= alpha;
- res6_0 *= alpha;
- res7_0 *= alpha;
-
- C0[0] = res0_0;
-
- C1[0] = res1_0;
-
- C2[0] = res2_0;
-
- C3[0] = res3_0;
- C4[0] = res4_0;
- C5[0] = res5_0;
- C6[0] = res6_0;
- C7[0] = res7_0;
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- temp = bk - off;
- #ifdef LEFT
- temp -= 1; // number of values in A
- #else
- temp -= 8; // number of values in B
- #endif
- ptrba += temp*1;
- ptrbb += temp*8;
- #endif
-
- #ifdef LEFT
- off += 1; // number of values in A
- #endif
-
- C0 = C0+1;
- C1 = C1+1;
- C2 = C2+1;
- C3 = C3+1;
- C4 = C4+1;
- C5 = C5+1;
- C6 = C6+1;
- C7 = C7+1;
-
- }
-
-
- #if defined(TRMMKERNEL) && !defined(LEFT)
- off += 8;
- #endif
-
- k = (bk<<3);
- bb = bb+k;
- i = (ldc<<3);
- C = C+i;
- }
-
-
-
- for (j=0; j<(bn&4); j+=4) // do blocks of the Mx4 loops
- {
- C0 = C;
- C1 = C0+ldc;
- C2 = C1+ldc;
- C3 = C2+ldc;
-
-
- if (left) {
- off = offset;
- }
-
- ptrba = ba;
-
- for (i=0; i<bm/4; i+=1) // do blocks of 4x4
- {
-
- ptrbb = bb;
- if (backwards)
- {
- ptrba += off*4; // number of values in A
- ptrbb += off*4; // number of values in B
- }
-
- 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;
-
- temp = backwards ? bk-off :
- left ? off + 4 : // number of values in A
- off + 4; // number of values in B
-
- 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 (!backwards) {
- temp = bk-off;
- temp = left ? temp - 4 : // number of values in A
- temp - 4; // number of values in B
-
- ptrba += temp*4; // number of values in A
- ptrbb += temp*4; // number of values in B
- }
- #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 ) // do any 2x4 loop
- {
-
- #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 ) // do any 1x4 loop
- {
-
- #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;
- }
-
-
-
- for (j=0; j<(bn&2); j+=2) // do the Mx2 loops
- {
- C0 = C;
- C1 = C0+ldc;
-
- #if defined(TRMMKERNEL) && defined(LEFT)
- off = offset;
- #endif
-
-
- ptrba = ba;
-
- for (i=0; i<bm/4; i+=1) // do blocks of 4x2
- {
-
- #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 ) // do any 2x2 loop
- {
-
- #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 ) // do any 1x2 loop
- {
-
- #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) // do the Mx1 loops
- {
- C0 = C;
-
- #if defined(TRMMKERNEL) && defined(LEFT)
- off = offset;
- #endif
-
- ptrba = ba;
-
- for (i=0; i<bm/4; i+=1) // do blocks of 4x1 loops
- {
-
- #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 ) // do any 2x1 loop
- {
-
- #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 ) // do any 1x1 loop
- {
-
- #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;
- }
|