|
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875 |
- #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,*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 a0;
- FLOAT a1;
-
- FLOAT b0;
- FLOAT b1;
- FLOAT b2;
- FLOAT b3;
-
- 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/4; j+=1) // 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;
- }
|