kusano 2b45e8
/*********************************************************************/
kusano 2b45e8
/* Copyright 2009, 2010 The University of Texas at Austin.           */
kusano 2b45e8
/* All rights reserved.                                              */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/* Redistribution and use in source and binary forms, with or        */
kusano 2b45e8
/* without modification, are permitted provided that the following   */
kusano 2b45e8
/* conditions are met:                                               */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/*   1. Redistributions of source code must retain the above         */
kusano 2b45e8
/*      copyright notice, this list of conditions and the following  */
kusano 2b45e8
/*      disclaimer.                                                  */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/*   2. Redistributions in binary form must reproduce the above      */
kusano 2b45e8
/*      copyright notice, this list of conditions and the following  */
kusano 2b45e8
/*      disclaimer in the documentation and/or other materials       */
kusano 2b45e8
/*      provided with the distribution.                              */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
kusano 2b45e8
/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
kusano 2b45e8
/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
kusano 2b45e8
/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
kusano 2b45e8
/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
kusano 2b45e8
/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
kusano 2b45e8
/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
kusano 2b45e8
/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
kusano 2b45e8
/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
kusano 2b45e8
/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
kusano 2b45e8
/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
kusano 2b45e8
/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
kusano 2b45e8
/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
kusano 2b45e8
/*    POSSIBILITY OF SUCH DAMAGE.                                    */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/* The views and conclusions contained in the software and           */
kusano 2b45e8
/* documentation are those of the authors and should not be          */
kusano 2b45e8
/* interpreted as representing official policies, either expressed   */
kusano 2b45e8
/* or implied, of The University of Texas at Austin.                 */
kusano 2b45e8
/*********************************************************************/
kusano 2b45e8
kusano 2b45e8
#include <stdio.h></stdio.h>
kusano 2b45e8
#include "common.h"
kusano 2b45e8
kusano 2b45e8
int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
kusano 2b45e8
kusano 2b45e8
  BLASLONG i, j;
kusano 2b45e8
kusano 2b45e8
  FLOAT *aoffset;
kusano 2b45e8
  FLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
kusano 2b45e8
  FLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
kusano 2b45e8
kusano 2b45e8
  FLOAT *boffset,  *boffset1, *boffset2, *boffset3, *boffset4;
kusano 2b45e8
kusano 2b45e8
  FLOAT ctemp01, ctemp02, ctemp03, ctemp04;
kusano 2b45e8
  FLOAT ctemp05, ctemp06, ctemp07, ctemp08;
kusano 2b45e8
  FLOAT ctemp09, ctemp10, ctemp11, ctemp12;
kusano 2b45e8
  FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
kusano 2b45e8
  FLOAT ctemp17, ctemp18, ctemp19, ctemp20;
kusano 2b45e8
  FLOAT ctemp21, ctemp22, ctemp23, ctemp24;
kusano 2b45e8
  FLOAT ctemp25, ctemp26, ctemp27, ctemp28;
kusano 2b45e8
  FLOAT ctemp29, ctemp30, ctemp31, ctemp32;
kusano 2b45e8
  FLOAT ctemp33, ctemp34, ctemp35, ctemp36;
kusano 2b45e8
  FLOAT ctemp37, ctemp38, ctemp39, ctemp40;
kusano 2b45e8
  FLOAT ctemp41, ctemp42, ctemp43, ctemp44;
kusano 2b45e8
  FLOAT ctemp45, ctemp46, ctemp47, ctemp48;
kusano 2b45e8
  FLOAT ctemp49, ctemp50, ctemp51, ctemp52;
kusano 2b45e8
  FLOAT ctemp53, ctemp54, ctemp55, ctemp56;
kusano 2b45e8
  FLOAT ctemp57, ctemp58, ctemp59, ctemp60;
kusano 2b45e8
  FLOAT ctemp61, ctemp62, ctemp63, ctemp64;
kusano 2b45e8
kusano 2b45e8
  aoffset   = a;
kusano 2b45e8
  boffset   = b;
kusano 2b45e8
kusano 2b45e8
#if 0
kusano 2b45e8
  fprintf(stderr, "M = %d N = %d\n", m, n);
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
  boffset2  = b + m  * (n & ~7);
kusano 2b45e8
  boffset3  = b + m  * (n & ~3);
kusano 2b45e8
  boffset4  = b + m  * (n & ~1);
kusano 2b45e8
kusano 2b45e8
  j = (m >> 3);
kusano 2b45e8
  if (j > 0){
kusano 2b45e8
    do{
kusano 2b45e8
      aoffset1  = aoffset;
kusano 2b45e8
      aoffset2  = aoffset1 + lda;
kusano 2b45e8
      aoffset3  = aoffset2 + lda;
kusano 2b45e8
      aoffset4  = aoffset3 + lda;
kusano 2b45e8
      aoffset5  = aoffset4 + lda;
kusano 2b45e8
      aoffset6  = aoffset5 + lda;
kusano 2b45e8
      aoffset7  = aoffset6 + lda;
kusano 2b45e8
      aoffset8  = aoffset7 + lda;
kusano 2b45e8
      aoffset += 8 * lda;
kusano 2b45e8
kusano 2b45e8
      boffset1  = boffset;
kusano 2b45e8
      boffset  += 64;
kusano 2b45e8
kusano 2b45e8
      i = (n >> 3);
kusano 2b45e8
      if (i > 0){
kusano 2b45e8
	do{
kusano 2b45e8
	  ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
	  ctemp02 = *(aoffset1 + 1);
kusano 2b45e8
	  ctemp03 = *(aoffset1 + 2);
kusano 2b45e8
	  ctemp04 = *(aoffset1 + 3);
kusano 2b45e8
	  ctemp05 = *(aoffset1 + 4);
kusano 2b45e8
	  ctemp06 = *(aoffset1 + 5);
kusano 2b45e8
	  ctemp07 = *(aoffset1 + 6);
kusano 2b45e8
	  ctemp08 = *(aoffset1 + 7);
kusano 2b45e8
	  aoffset1 += 8;
kusano 2b45e8
	  
kusano 2b45e8
	  ctemp09 = *(aoffset2 + 0);
kusano 2b45e8
	  ctemp10 = *(aoffset2 + 1);
kusano 2b45e8
	  ctemp11 = *(aoffset2 + 2);
kusano 2b45e8
	  ctemp12 = *(aoffset2 + 3);
kusano 2b45e8
	  ctemp13 = *(aoffset2 + 4);
kusano 2b45e8
	  ctemp14 = *(aoffset2 + 5);
kusano 2b45e8
	  ctemp15 = *(aoffset2 + 6);
kusano 2b45e8
	  ctemp16 = *(aoffset2 + 7);
kusano 2b45e8
	  aoffset2 += 8;
kusano 2b45e8
	  
kusano 2b45e8
	  ctemp17 = *(aoffset3 + 0);
kusano 2b45e8
	  ctemp18 = *(aoffset3 + 1);
kusano 2b45e8
	  ctemp19 = *(aoffset3 + 2);
kusano 2b45e8
	  ctemp20 = *(aoffset3 + 3);
kusano 2b45e8
	  ctemp21 = *(aoffset3 + 4);
kusano 2b45e8
	  ctemp22 = *(aoffset3 + 5);
kusano 2b45e8
	  ctemp23 = *(aoffset3 + 6);
kusano 2b45e8
	  ctemp24 = *(aoffset3 + 7);
kusano 2b45e8
	  aoffset3 += 8;
kusano 2b45e8
	  
kusano 2b45e8
	  ctemp25 = *(aoffset4 + 0);
kusano 2b45e8
	  ctemp26 = *(aoffset4 + 1);
kusano 2b45e8
	  ctemp27 = *(aoffset4 + 2);
kusano 2b45e8
	  ctemp28 = *(aoffset4 + 3);
kusano 2b45e8
	  ctemp29 = *(aoffset4 + 4);
kusano 2b45e8
	  ctemp30 = *(aoffset4 + 5);
kusano 2b45e8
	  ctemp31 = *(aoffset4 + 6);
kusano 2b45e8
	  ctemp32 = *(aoffset4 + 7);
kusano 2b45e8
	  aoffset4 += 8;
kusano 2b45e8
kusano 2b45e8
	  ctemp33 = *(aoffset5 + 0);
kusano 2b45e8
	  ctemp34 = *(aoffset5 + 1);
kusano 2b45e8
	  ctemp35 = *(aoffset5 + 2);
kusano 2b45e8
	  ctemp36 = *(aoffset5 + 3);
kusano 2b45e8
	  ctemp37 = *(aoffset5 + 4);
kusano 2b45e8
	  ctemp38 = *(aoffset5 + 5);
kusano 2b45e8
	  ctemp39 = *(aoffset5 + 6);
kusano 2b45e8
	  ctemp40 = *(aoffset5 + 7);
kusano 2b45e8
	  aoffset5 += 8;
kusano 2b45e8
	  
kusano 2b45e8
	  ctemp41 = *(aoffset6 + 0);
kusano 2b45e8
	  ctemp42 = *(aoffset6 + 1);
kusano 2b45e8
	  ctemp43 = *(aoffset6 + 2);
kusano 2b45e8
	  ctemp44 = *(aoffset6 + 3);
kusano 2b45e8
	  ctemp45 = *(aoffset6 + 4);
kusano 2b45e8
	  ctemp46 = *(aoffset6 + 5);
kusano 2b45e8
	  ctemp47 = *(aoffset6 + 6);
kusano 2b45e8
	  ctemp48 = *(aoffset6 + 7);
kusano 2b45e8
	  aoffset6 += 8;
kusano 2b45e8
	  
kusano 2b45e8
	  ctemp49 = *(aoffset7 + 0);
kusano 2b45e8
	  ctemp50 = *(aoffset7 + 1);
kusano 2b45e8
	  ctemp51 = *(aoffset7 + 2);
kusano 2b45e8
	  ctemp52 = *(aoffset7 + 3);
kusano 2b45e8
	  ctemp53 = *(aoffset7 + 4);
kusano 2b45e8
	  ctemp54 = *(aoffset7 + 5);
kusano 2b45e8
	  ctemp55 = *(aoffset7 + 6);
kusano 2b45e8
	  ctemp56 = *(aoffset7 + 7);
kusano 2b45e8
	  aoffset7 += 8;
kusano 2b45e8
	  
kusano 2b45e8
	  ctemp57 = *(aoffset8 + 0);
kusano 2b45e8
	  ctemp58 = *(aoffset8 + 1);
kusano 2b45e8
	  ctemp59 = *(aoffset8 + 2);
kusano 2b45e8
	  ctemp60 = *(aoffset8 + 3);
kusano 2b45e8
	  ctemp61 = *(aoffset8 + 4);
kusano 2b45e8
	  ctemp62 = *(aoffset8 + 5);
kusano 2b45e8
	  ctemp63 = *(aoffset8 + 6);
kusano 2b45e8
	  ctemp64 = *(aoffset8 + 7);
kusano 2b45e8
	  aoffset8 += 8;
kusano 2b45e8
	
kusano 2b45e8
	  *(boffset1 +  0) = ctemp01;
kusano 2b45e8
	  *(boffset1 +  1) = ctemp02;
kusano 2b45e8
	  *(boffset1 +  2) = ctemp03;
kusano 2b45e8
	  *(boffset1 +  3) = ctemp04;
kusano 2b45e8
	  *(boffset1 +  4) = ctemp05;
kusano 2b45e8
	  *(boffset1 +  5) = ctemp06;
kusano 2b45e8
	  *(boffset1 +  6) = ctemp07;
kusano 2b45e8
	  *(boffset1 +  7) = ctemp08;
kusano 2b45e8
	  
kusano 2b45e8
	  *(boffset1 +  8) = ctemp09;
kusano 2b45e8
	  *(boffset1 +  9) = ctemp10;
kusano 2b45e8
	  *(boffset1 + 10) = ctemp11;
kusano 2b45e8
	  *(boffset1 + 11) = ctemp12;
kusano 2b45e8
	  *(boffset1 + 12) = ctemp13;
kusano 2b45e8
	  *(boffset1 + 13) = ctemp14;
kusano 2b45e8
	  *(boffset1 + 14) = ctemp15;
kusano 2b45e8
	  *(boffset1 + 15) = ctemp16;
kusano 2b45e8
kusano 2b45e8
	  *(boffset1 + 16) = ctemp17;
kusano 2b45e8
	  *(boffset1 + 17) = ctemp18;
kusano 2b45e8
	  *(boffset1 + 18) = ctemp19;
kusano 2b45e8
	  *(boffset1 + 19) = ctemp20;
kusano 2b45e8
	  *(boffset1 + 20) = ctemp21;
kusano 2b45e8
	  *(boffset1 + 21) = ctemp22;
kusano 2b45e8
	  *(boffset1 + 22) = ctemp23;
kusano 2b45e8
	  *(boffset1 + 23) = ctemp24;
kusano 2b45e8
	  
kusano 2b45e8
	  *(boffset1 + 24) = ctemp25;
kusano 2b45e8
	  *(boffset1 + 25) = ctemp26;
kusano 2b45e8
	  *(boffset1 + 26) = ctemp27;
kusano 2b45e8
	  *(boffset1 + 27) = ctemp28;
kusano 2b45e8
	  *(boffset1 + 28) = ctemp29;
kusano 2b45e8
	  *(boffset1 + 29) = ctemp30;
kusano 2b45e8
	  *(boffset1 + 30) = ctemp31;
kusano 2b45e8
	  *(boffset1 + 31) = ctemp32;
kusano 2b45e8
kusano 2b45e8
	  *(boffset1 + 32) = ctemp33;
kusano 2b45e8
	  *(boffset1 + 33) = ctemp34;
kusano 2b45e8
	  *(boffset1 + 34) = ctemp35;
kusano 2b45e8
	  *(boffset1 + 35) = ctemp36;
kusano 2b45e8
	  *(boffset1 + 36) = ctemp37;
kusano 2b45e8
	  *(boffset1 + 37) = ctemp38;
kusano 2b45e8
	  *(boffset1 + 38) = ctemp39;
kusano 2b45e8
	  *(boffset1 + 39) = ctemp40;
kusano 2b45e8
	  
kusano 2b45e8
	  *(boffset1 + 40) = ctemp41;
kusano 2b45e8
	  *(boffset1 + 41) = ctemp42;
kusano 2b45e8
	  *(boffset1 + 42) = ctemp43;
kusano 2b45e8
	  *(boffset1 + 43) = ctemp44;
kusano 2b45e8
	  *(boffset1 + 44) = ctemp45;
kusano 2b45e8
	  *(boffset1 + 45) = ctemp46;
kusano 2b45e8
	  *(boffset1 + 46) = ctemp47;
kusano 2b45e8
	  *(boffset1 + 47) = ctemp48;
kusano 2b45e8
kusano 2b45e8
	  *(boffset1 + 48) = ctemp49;
kusano 2b45e8
	  *(boffset1 + 49) = ctemp50;
kusano 2b45e8
	  *(boffset1 + 50) = ctemp51;
kusano 2b45e8
	  *(boffset1 + 51) = ctemp52;
kusano 2b45e8
	  *(boffset1 + 52) = ctemp53;
kusano 2b45e8
	  *(boffset1 + 53) = ctemp54;
kusano 2b45e8
	  *(boffset1 + 54) = ctemp55;
kusano 2b45e8
	  *(boffset1 + 55) = ctemp56;
kusano 2b45e8
	  
kusano 2b45e8
	  *(boffset1 + 56) = ctemp57;
kusano 2b45e8
	  *(boffset1 + 57) = ctemp58;
kusano 2b45e8
	  *(boffset1 + 58) = ctemp59;
kusano 2b45e8
	  *(boffset1 + 59) = ctemp60;
kusano 2b45e8
	  *(boffset1 + 60) = ctemp61;
kusano 2b45e8
	  *(boffset1 + 61) = ctemp62;
kusano 2b45e8
	  *(boffset1 + 62) = ctemp63;
kusano 2b45e8
	  *(boffset1 + 63) = ctemp64;
kusano 2b45e8
	  
kusano 2b45e8
	  boffset1 += m * 8;
kusano 2b45e8
	  i --;
kusano 2b45e8
	}while(i > 0);
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      if (n & 4){
kusano 2b45e8
	ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
	ctemp02 = *(aoffset1 + 1);
kusano 2b45e8
	ctemp03 = *(aoffset1 + 2);
kusano 2b45e8
	ctemp04 = *(aoffset1 + 3);
kusano 2b45e8
	aoffset1 += 4;
kusano 2b45e8
	
kusano 2b45e8
	ctemp05 = *(aoffset2 + 0);
kusano 2b45e8
	ctemp06 = *(aoffset2 + 1);
kusano 2b45e8
	ctemp07 = *(aoffset2 + 2);
kusano 2b45e8
	ctemp08 = *(aoffset2 + 3);
kusano 2b45e8
	aoffset2 += 4;
kusano 2b45e8
	
kusano 2b45e8
	ctemp09 = *(aoffset3 + 0);
kusano 2b45e8
	ctemp10 = *(aoffset3 + 1);
kusano 2b45e8
	ctemp11 = *(aoffset3 + 2);
kusano 2b45e8
	ctemp12 = *(aoffset3 + 3);
kusano 2b45e8
	aoffset3 += 4;
kusano 2b45e8
	
kusano 2b45e8
	ctemp13 = *(aoffset4 + 0);
kusano 2b45e8
	ctemp14 = *(aoffset4 + 1);
kusano 2b45e8
	ctemp15 = *(aoffset4 + 2);
kusano 2b45e8
	ctemp16 = *(aoffset4 + 3);
kusano 2b45e8
	aoffset4 += 4;
kusano 2b45e8
	
kusano 2b45e8
	ctemp17 = *(aoffset5 + 0);
kusano 2b45e8
	ctemp18 = *(aoffset5 + 1);
kusano 2b45e8
	ctemp19 = *(aoffset5 + 2);
kusano 2b45e8
	ctemp20 = *(aoffset5 + 3);
kusano 2b45e8
	aoffset5 += 4;
kusano 2b45e8
	
kusano 2b45e8
	ctemp21 = *(aoffset6 + 0);
kusano 2b45e8
	ctemp22 = *(aoffset6 + 1);
kusano 2b45e8
	ctemp23 = *(aoffset6 + 2);
kusano 2b45e8
	ctemp24 = *(aoffset6 + 3);
kusano 2b45e8
	aoffset6 += 4;
kusano 2b45e8
	
kusano 2b45e8
	ctemp25 = *(aoffset7 + 0);
kusano 2b45e8
	ctemp26 = *(aoffset7 + 1);
kusano 2b45e8
	ctemp27 = *(aoffset7 + 2);
kusano 2b45e8
	ctemp28 = *(aoffset7 + 3);
kusano 2b45e8
	aoffset7 += 4;
kusano 2b45e8
	
kusano 2b45e8
	ctemp29 = *(aoffset8 + 0);
kusano 2b45e8
	ctemp30 = *(aoffset8 + 1);
kusano 2b45e8
	ctemp31 = *(aoffset8 + 2);
kusano 2b45e8
	ctemp32 = *(aoffset8 + 3);
kusano 2b45e8
	aoffset8 += 4;
kusano 2b45e8
	
kusano 2b45e8
	*(boffset2 +  0) = ctemp01;
kusano 2b45e8
	*(boffset2 +  1) = ctemp02;
kusano 2b45e8
	*(boffset2 +  2) = ctemp03;
kusano 2b45e8
	*(boffset2 +  3) = ctemp04;
kusano 2b45e8
	*(boffset2 +  4) = ctemp05;
kusano 2b45e8
	*(boffset2 +  5) = ctemp06;
kusano 2b45e8
	*(boffset2 +  6) = ctemp07;
kusano 2b45e8
	*(boffset2 +  7) = ctemp08;
kusano 2b45e8
	*(boffset2 +  8) = ctemp09;
kusano 2b45e8
	*(boffset2 +  9) = ctemp10;
kusano 2b45e8
	*(boffset2 + 10) = ctemp11;
kusano 2b45e8
	*(boffset2 + 11) = ctemp12;
kusano 2b45e8
	*(boffset2 + 12) = ctemp13;
kusano 2b45e8
	*(boffset2 + 13) = ctemp14;
kusano 2b45e8
	*(boffset2 + 14) = ctemp15;
kusano 2b45e8
	*(boffset2 + 15) = ctemp16;
kusano 2b45e8
kusano 2b45e8
	*(boffset2 + 16) = ctemp17;
kusano 2b45e8
	*(boffset2 + 17) = ctemp18;
kusano 2b45e8
	*(boffset2 + 18) = ctemp19;
kusano 2b45e8
	*(boffset2 + 19) = ctemp20;
kusano 2b45e8
	*(boffset2 + 20) = ctemp21;
kusano 2b45e8
	*(boffset2 + 21) = ctemp22;
kusano 2b45e8
	*(boffset2 + 22) = ctemp23;
kusano 2b45e8
	*(boffset2 + 23) = ctemp24;
kusano 2b45e8
	*(boffset2 + 24) = ctemp25;
kusano 2b45e8
	*(boffset2 + 25) = ctemp26;
kusano 2b45e8
	*(boffset2 + 26) = ctemp27;
kusano 2b45e8
	*(boffset2 + 27) = ctemp28;
kusano 2b45e8
	*(boffset2 + 28) = ctemp29;
kusano 2b45e8
	*(boffset2 + 29) = ctemp30;
kusano 2b45e8
	*(boffset2 + 30) = ctemp31;
kusano 2b45e8
	*(boffset2 + 31) = ctemp32;
kusano 2b45e8
kusano 2b45e8
	boffset2 += 32;
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      if (n & 2){
kusano 2b45e8
	ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
	ctemp02 = *(aoffset1 + 1);
kusano 2b45e8
	aoffset1 += 2;
kusano 2b45e8
	
kusano 2b45e8
	ctemp03 = *(aoffset2 + 0);
kusano 2b45e8
	ctemp04 = *(aoffset2 + 1);
kusano 2b45e8
	aoffset2 += 2;
kusano 2b45e8
	
kusano 2b45e8
	ctemp05 = *(aoffset3 + 0);
kusano 2b45e8
	ctemp06 = *(aoffset3 + 1);
kusano 2b45e8
	aoffset3 += 2;
kusano 2b45e8
	
kusano 2b45e8
	ctemp07 = *(aoffset4 + 0);
kusano 2b45e8
	ctemp08 = *(aoffset4 + 1);
kusano 2b45e8
	aoffset4 += 2;
kusano 2b45e8
	
kusano 2b45e8
	ctemp09 = *(aoffset5 + 0);
kusano 2b45e8
	ctemp10 = *(aoffset5 + 1);
kusano 2b45e8
	aoffset5 += 2;
kusano 2b45e8
	
kusano 2b45e8
	ctemp11 = *(aoffset6 + 0);
kusano 2b45e8
	ctemp12 = *(aoffset6 + 1);
kusano 2b45e8
	aoffset6 += 2;
kusano 2b45e8
	
kusano 2b45e8
	ctemp13 = *(aoffset7 + 0);
kusano 2b45e8
	ctemp14 = *(aoffset7 + 1);
kusano 2b45e8
	aoffset7 += 2;
kusano 2b45e8
	
kusano 2b45e8
	ctemp15 = *(aoffset8 + 0);
kusano 2b45e8
	ctemp16 = *(aoffset8 + 1);
kusano 2b45e8
	aoffset8 += 2;
kusano 2b45e8
	
kusano 2b45e8
	*(boffset3 +  0) = ctemp01;
kusano 2b45e8
	*(boffset3 +  1) = ctemp02;
kusano 2b45e8
	*(boffset3 +  2) = ctemp03;
kusano 2b45e8
	*(boffset3 +  3) = ctemp04;
kusano 2b45e8
	*(boffset3 +  4) = ctemp05;
kusano 2b45e8
	*(boffset3 +  5) = ctemp06;
kusano 2b45e8
	*(boffset3 +  6) = ctemp07;
kusano 2b45e8
	*(boffset3 +  7) = ctemp08;
kusano 2b45e8
	*(boffset3 +  8) = ctemp09;
kusano 2b45e8
	*(boffset3 +  9) = ctemp10;
kusano 2b45e8
	*(boffset3 + 10) = ctemp11;
kusano 2b45e8
	*(boffset3 + 11) = ctemp12;
kusano 2b45e8
	*(boffset3 + 12) = ctemp13;
kusano 2b45e8
	*(boffset3 + 13) = ctemp14;
kusano 2b45e8
	*(boffset3 + 14) = ctemp15;
kusano 2b45e8
	*(boffset3 + 15) = ctemp16;
kusano 2b45e8
	boffset3 += 16;
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      if (n & 1){
kusano 2b45e8
	ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
	aoffset1 ++;
kusano 2b45e8
	ctemp02 = *(aoffset2 + 0);
kusano 2b45e8
	aoffset2 ++;
kusano 2b45e8
	ctemp03 = *(aoffset3 + 0);
kusano 2b45e8
	aoffset3 ++;
kusano 2b45e8
	ctemp04 = *(aoffset4 + 0);
kusano 2b45e8
	aoffset4 ++;
kusano 2b45e8
	ctemp05 = *(aoffset5 + 0);
kusano 2b45e8
	aoffset5 ++;
kusano 2b45e8
	ctemp06 = *(aoffset6 + 0);
kusano 2b45e8
	aoffset6 ++;
kusano 2b45e8
	ctemp07 = *(aoffset7 + 0);
kusano 2b45e8
	aoffset7 ++;
kusano 2b45e8
	ctemp08 = *(aoffset8 + 0);
kusano 2b45e8
	aoffset8 ++;
kusano 2b45e8
	
kusano 2b45e8
	*(boffset4 +  0) = ctemp01;
kusano 2b45e8
	*(boffset4 +  1) = ctemp02;
kusano 2b45e8
	*(boffset4 +  2) = ctemp03;
kusano 2b45e8
	*(boffset4 +  3) = ctemp04;
kusano 2b45e8
	*(boffset4 +  4) = ctemp05;
kusano 2b45e8
	*(boffset4 +  5) = ctemp06;
kusano 2b45e8
	*(boffset4 +  6) = ctemp07;
kusano 2b45e8
	*(boffset4 +  7) = ctemp08;
kusano 2b45e8
	boffset4 += 8;
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      j--;
kusano 2b45e8
    }while(j > 0);
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (m & 4){
kusano 2b45e8
kusano 2b45e8
    aoffset1  = aoffset;
kusano 2b45e8
    aoffset2  = aoffset1 + lda;
kusano 2b45e8
    aoffset3  = aoffset2 + lda;
kusano 2b45e8
    aoffset4  = aoffset3 + lda;
kusano 2b45e8
    aoffset += 4 * lda;
kusano 2b45e8
    
kusano 2b45e8
    boffset1  = boffset;
kusano 2b45e8
    boffset  += 32;
kusano 2b45e8
    
kusano 2b45e8
    i = (n >> 3);
kusano 2b45e8
    if (i > 0){
kusano 2b45e8
kusano 2b45e8
      do{
kusano 2b45e8
	ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
	ctemp02 = *(aoffset1 + 1);
kusano 2b45e8
	ctemp03 = *(aoffset1 + 2);
kusano 2b45e8
	ctemp04 = *(aoffset1 + 3);
kusano 2b45e8
	ctemp05 = *(aoffset1 + 4);
kusano 2b45e8
	ctemp06 = *(aoffset1 + 5);
kusano 2b45e8
	ctemp07 = *(aoffset1 + 6);
kusano 2b45e8
	ctemp08 = *(aoffset1 + 7);
kusano 2b45e8
	aoffset1 += 8;
kusano 2b45e8
	
kusano 2b45e8
	ctemp09 = *(aoffset2 + 0);
kusano 2b45e8
	ctemp10 = *(aoffset2 + 1);
kusano 2b45e8
	ctemp11 = *(aoffset2 + 2);
kusano 2b45e8
	ctemp12 = *(aoffset2 + 3);
kusano 2b45e8
	ctemp13 = *(aoffset2 + 4);
kusano 2b45e8
	ctemp14 = *(aoffset2 + 5);
kusano 2b45e8
	ctemp15 = *(aoffset2 + 6);
kusano 2b45e8
	ctemp16 = *(aoffset2 + 7);
kusano 2b45e8
	aoffset2 += 8;
kusano 2b45e8
	
kusano 2b45e8
	ctemp17 = *(aoffset3 + 0);
kusano 2b45e8
	ctemp18 = *(aoffset3 + 1);
kusano 2b45e8
	ctemp19 = *(aoffset3 + 2);
kusano 2b45e8
	ctemp20 = *(aoffset3 + 3);
kusano 2b45e8
	ctemp21 = *(aoffset3 + 4);
kusano 2b45e8
	ctemp22 = *(aoffset3 + 5);
kusano 2b45e8
	ctemp23 = *(aoffset3 + 6);
kusano 2b45e8
	ctemp24 = *(aoffset3 + 7);
kusano 2b45e8
	aoffset3 += 8;
kusano 2b45e8
	
kusano 2b45e8
	ctemp25 = *(aoffset4 + 0);
kusano 2b45e8
	ctemp26 = *(aoffset4 + 1);
kusano 2b45e8
	ctemp27 = *(aoffset4 + 2);
kusano 2b45e8
	ctemp28 = *(aoffset4 + 3);
kusano 2b45e8
	ctemp29 = *(aoffset4 + 4);
kusano 2b45e8
	ctemp30 = *(aoffset4 + 5);
kusano 2b45e8
	ctemp31 = *(aoffset4 + 6);
kusano 2b45e8
	ctemp32 = *(aoffset4 + 7);
kusano 2b45e8
	aoffset4 += 8;
kusano 2b45e8
	
kusano 2b45e8
	*(boffset1 +  0) = ctemp01;
kusano 2b45e8
	*(boffset1 +  1) = ctemp02;
kusano 2b45e8
	*(boffset1 +  2) = ctemp03;
kusano 2b45e8
	*(boffset1 +  3) = ctemp04;
kusano 2b45e8
	*(boffset1 +  4) = ctemp05;
kusano 2b45e8
	*(boffset1 +  5) = ctemp06;
kusano 2b45e8
	*(boffset1 +  6) = ctemp07;
kusano 2b45e8
	*(boffset1 +  7) = ctemp08;
kusano 2b45e8
	
kusano 2b45e8
	*(boffset1 +  8) = ctemp09;
kusano 2b45e8
	*(boffset1 +  9) = ctemp10;
kusano 2b45e8
	*(boffset1 + 10) = ctemp11;
kusano 2b45e8
	*(boffset1 + 11) = ctemp12;
kusano 2b45e8
	*(boffset1 + 12) = ctemp13;
kusano 2b45e8
	*(boffset1 + 13) = ctemp14;
kusano 2b45e8
	*(boffset1 + 14) = ctemp15;
kusano 2b45e8
	*(boffset1 + 15) = ctemp16;
kusano 2b45e8
	
kusano 2b45e8
	*(boffset1 + 16) = ctemp17;
kusano 2b45e8
	*(boffset1 + 17) = ctemp18;
kusano 2b45e8
	*(boffset1 + 18) = ctemp19;
kusano 2b45e8
	*(boffset1 + 19) = ctemp20;
kusano 2b45e8
	*(boffset1 + 20) = ctemp21;
kusano 2b45e8
	*(boffset1 + 21) = ctemp22;
kusano 2b45e8
	*(boffset1 + 22) = ctemp23;
kusano 2b45e8
	*(boffset1 + 23) = ctemp24;
kusano 2b45e8
kusano 2b45e8
	*(boffset1 + 24) = ctemp25;
kusano 2b45e8
	*(boffset1 + 25) = ctemp26;
kusano 2b45e8
	*(boffset1 + 26) = ctemp27;
kusano 2b45e8
	*(boffset1 + 27) = ctemp28;
kusano 2b45e8
	*(boffset1 + 28) = ctemp29;
kusano 2b45e8
	*(boffset1 + 29) = ctemp30;
kusano 2b45e8
	*(boffset1 + 30) = ctemp31;
kusano 2b45e8
	*(boffset1 + 31) = ctemp32;
kusano 2b45e8
kusano 2b45e8
	boffset1 += 8 * m;
kusano 2b45e8
	i --;
kusano 2b45e8
      }while(i > 0);
kusano 2b45e8
    }
kusano 2b45e8
kusano 2b45e8
    if (n & 4) {
kusano 2b45e8
      ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
      ctemp02 = *(aoffset1 + 1);
kusano 2b45e8
      ctemp03 = *(aoffset1 + 2);
kusano 2b45e8
      ctemp04 = *(aoffset1 + 3);
kusano 2b45e8
      aoffset1 += 4;
kusano 2b45e8
      
kusano 2b45e8
      ctemp05 = *(aoffset2 + 0);
kusano 2b45e8
      ctemp06 = *(aoffset2 + 1);
kusano 2b45e8
      ctemp07 = *(aoffset2 + 2);
kusano 2b45e8
      ctemp08 = *(aoffset2 + 3);
kusano 2b45e8
      aoffset2 += 4;
kusano 2b45e8
      
kusano 2b45e8
      ctemp09 = *(aoffset3 + 0);
kusano 2b45e8
      ctemp10 = *(aoffset3 + 1);
kusano 2b45e8
      ctemp11 = *(aoffset3 + 2);
kusano 2b45e8
      ctemp12 = *(aoffset3 + 3);
kusano 2b45e8
      aoffset3 += 4;
kusano 2b45e8
      
kusano 2b45e8
      ctemp13 = *(aoffset4 + 0);
kusano 2b45e8
      ctemp14 = *(aoffset4 + 1);
kusano 2b45e8
      ctemp15 = *(aoffset4 + 2);
kusano 2b45e8
      ctemp16 = *(aoffset4 + 3);
kusano 2b45e8
      aoffset4 += 4;
kusano 2b45e8
      
kusano 2b45e8
      *(boffset2 +  0) = ctemp01;
kusano 2b45e8
      *(boffset2 +  1) = ctemp02;
kusano 2b45e8
      *(boffset2 +  2) = ctemp03;
kusano 2b45e8
      *(boffset2 +  3) = ctemp04;
kusano 2b45e8
      *(boffset2 +  4) = ctemp05;
kusano 2b45e8
      *(boffset2 +  5) = ctemp06;
kusano 2b45e8
      *(boffset2 +  6) = ctemp07;
kusano 2b45e8
      *(boffset2 +  7) = ctemp08;
kusano 2b45e8
      
kusano 2b45e8
      *(boffset2 +  8) = ctemp09;
kusano 2b45e8
      *(boffset2 +  9) = ctemp10;
kusano 2b45e8
      *(boffset2 + 10) = ctemp11;
kusano 2b45e8
      *(boffset2 + 11) = ctemp12;
kusano 2b45e8
      *(boffset2 + 12) = ctemp13;
kusano 2b45e8
      *(boffset2 + 13) = ctemp14;
kusano 2b45e8
      *(boffset2 + 14) = ctemp15;
kusano 2b45e8
      *(boffset2 + 15) = ctemp16;
kusano 2b45e8
      boffset2 += 16;
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
    if (n & 2){
kusano 2b45e8
      ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
      ctemp02 = *(aoffset1 + 1);
kusano 2b45e8
      aoffset1 += 2;
kusano 2b45e8
      
kusano 2b45e8
      ctemp03 = *(aoffset2 + 0);
kusano 2b45e8
      ctemp04 = *(aoffset2 + 1);
kusano 2b45e8
      aoffset2 += 2;
kusano 2b45e8
      
kusano 2b45e8
      ctemp05 = *(aoffset3 + 0);
kusano 2b45e8
      ctemp06 = *(aoffset3 + 1);
kusano 2b45e8
      aoffset3 += 2;
kusano 2b45e8
      
kusano 2b45e8
      ctemp07 = *(aoffset4 + 0);
kusano 2b45e8
      ctemp08 = *(aoffset4 + 1);
kusano 2b45e8
      aoffset4 += 2;
kusano 2b45e8
      
kusano 2b45e8
      *(boffset3 +  0) = ctemp01;
kusano 2b45e8
      *(boffset3 +  1) = ctemp02;
kusano 2b45e8
      *(boffset3 +  2) = ctemp03;
kusano 2b45e8
      *(boffset3 +  3) = ctemp04;
kusano 2b45e8
      *(boffset3 +  4) = ctemp05;
kusano 2b45e8
      *(boffset3 +  5) = ctemp06;
kusano 2b45e8
      *(boffset3 +  6) = ctemp07;
kusano 2b45e8
      *(boffset3 +  7) = ctemp08;
kusano 2b45e8
      boffset3 += 8;
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
    if (n & 1){
kusano 2b45e8
      ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
      aoffset1 ++;
kusano 2b45e8
      ctemp02 = *(aoffset2 + 0);
kusano 2b45e8
      aoffset2 ++;
kusano 2b45e8
      ctemp03 = *(aoffset3 + 0);
kusano 2b45e8
      aoffset3 ++;
kusano 2b45e8
      ctemp04 = *(aoffset4 + 0);
kusano 2b45e8
      aoffset4 ++;
kusano 2b45e8
      
kusano 2b45e8
      *(boffset4 +  0) = ctemp01;
kusano 2b45e8
      *(boffset4 +  1) = ctemp02;
kusano 2b45e8
      *(boffset4 +  2) = ctemp03;
kusano 2b45e8
      *(boffset4 +  3) = ctemp04;
kusano 2b45e8
      boffset4 += 4;
kusano 2b45e8
    }
kusano 2b45e8
  }
kusano 2b45e8
  
kusano 2b45e8
  if (m & 2){
kusano 2b45e8
    aoffset1  = aoffset;
kusano 2b45e8
    aoffset2  = aoffset1 + lda;
kusano 2b45e8
    aoffset += 2 * lda;
kusano 2b45e8
    
kusano 2b45e8
    boffset1  = boffset;
kusano 2b45e8
    boffset  += 16;
kusano 2b45e8
    
kusano 2b45e8
    i = (n >> 3);
kusano 2b45e8
    if (i > 0){
kusano 2b45e8
      do{
kusano 2b45e8
	ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
	ctemp02 = *(aoffset1 + 1);
kusano 2b45e8
	ctemp03 = *(aoffset1 + 2);
kusano 2b45e8
	ctemp04 = *(aoffset1 + 3);
kusano 2b45e8
	ctemp05 = *(aoffset1 + 4);
kusano 2b45e8
	ctemp06 = *(aoffset1 + 5);
kusano 2b45e8
	ctemp07 = *(aoffset1 + 6);
kusano 2b45e8
	ctemp08 = *(aoffset1 + 7);
kusano 2b45e8
	aoffset1 += 8;
kusano 2b45e8
	
kusano 2b45e8
	ctemp09 = *(aoffset2 + 0);
kusano 2b45e8
	ctemp10 = *(aoffset2 + 1);
kusano 2b45e8
	ctemp11 = *(aoffset2 + 2);
kusano 2b45e8
	ctemp12 = *(aoffset2 + 3);
kusano 2b45e8
	ctemp13 = *(aoffset2 + 4);
kusano 2b45e8
	ctemp14 = *(aoffset2 + 5);
kusano 2b45e8
	ctemp15 = *(aoffset2 + 6);
kusano 2b45e8
	ctemp16 = *(aoffset2 + 7);
kusano 2b45e8
	aoffset2 += 8;
kusano 2b45e8
	
kusano 2b45e8
	*(boffset1 +  0) = ctemp01;
kusano 2b45e8
	*(boffset1 +  1) = ctemp02;
kusano 2b45e8
	*(boffset1 +  2) = ctemp03;
kusano 2b45e8
	*(boffset1 +  3) = ctemp04;
kusano 2b45e8
	*(boffset1 +  4) = ctemp05;
kusano 2b45e8
	*(boffset1 +  5) = ctemp06;
kusano 2b45e8
	*(boffset1 +  6) = ctemp07;
kusano 2b45e8
	*(boffset1 +  7) = ctemp08;
kusano 2b45e8
	
kusano 2b45e8
	*(boffset1 +  8) = ctemp09;
kusano 2b45e8
	*(boffset1 +  9) = ctemp10;
kusano 2b45e8
	*(boffset1 + 10) = ctemp11;
kusano 2b45e8
	*(boffset1 + 11) = ctemp12;
kusano 2b45e8
	*(boffset1 + 12) = ctemp13;
kusano 2b45e8
	*(boffset1 + 13) = ctemp14;
kusano 2b45e8
	*(boffset1 + 14) = ctemp15;
kusano 2b45e8
	*(boffset1 + 15) = ctemp16;
kusano 2b45e8
	
kusano 2b45e8
	boffset1 += 8 * m;
kusano 2b45e8
	i --;
kusano 2b45e8
      }while(i > 0);
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
    if (n & 4){
kusano 2b45e8
      ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
      ctemp02 = *(aoffset1 + 1);
kusano 2b45e8
      ctemp03 = *(aoffset1 + 2);
kusano 2b45e8
      ctemp04 = *(aoffset1 + 3);
kusano 2b45e8
      aoffset1 += 4;
kusano 2b45e8
      
kusano 2b45e8
      ctemp05 = *(aoffset2 + 0);
kusano 2b45e8
      ctemp06 = *(aoffset2 + 1);
kusano 2b45e8
      ctemp07 = *(aoffset2 + 2);
kusano 2b45e8
      ctemp08 = *(aoffset2 + 3);
kusano 2b45e8
      aoffset2 += 4;
kusano 2b45e8
      
kusano 2b45e8
      *(boffset2 +  0) = ctemp01;
kusano 2b45e8
      *(boffset2 +  1) = ctemp02;
kusano 2b45e8
      *(boffset2 +  2) = ctemp03;
kusano 2b45e8
      *(boffset2 +  3) = ctemp04;
kusano 2b45e8
      *(boffset2 +  4) = ctemp05;
kusano 2b45e8
      *(boffset2 +  5) = ctemp06;
kusano 2b45e8
      *(boffset2 +  6) = ctemp07;
kusano 2b45e8
      *(boffset2 +  7) = ctemp08;
kusano 2b45e8
      boffset2 += 8;
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
    if (n & 2){
kusano 2b45e8
      ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
      ctemp02 = *(aoffset1 + 1);
kusano 2b45e8
      aoffset1 += 2;
kusano 2b45e8
      
kusano 2b45e8
      ctemp03 = *(aoffset2 + 0);
kusano 2b45e8
      ctemp04 = *(aoffset2 + 1);
kusano 2b45e8
      aoffset2 += 2;
kusano 2b45e8
      
kusano 2b45e8
      *(boffset3 +  0) = ctemp01;
kusano 2b45e8
      *(boffset3 +  1) = ctemp02;
kusano 2b45e8
      *(boffset3 +  2) = ctemp03;
kusano 2b45e8
      *(boffset3 +  3) = ctemp04;
kusano 2b45e8
      boffset3 += 4;
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
    if (n & 1){
kusano 2b45e8
      ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
      aoffset1 ++;
kusano 2b45e8
      ctemp02 = *(aoffset2 + 0);
kusano 2b45e8
      aoffset2 ++;
kusano 2b45e8
      
kusano 2b45e8
      *(boffset4 +  0) = ctemp01;
kusano 2b45e8
      *(boffset4 +  1) = ctemp02;
kusano 2b45e8
      boffset4 += 2;
kusano 2b45e8
    }
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (m & 1){
kusano 2b45e8
    aoffset1  = aoffset;
kusano 2b45e8
    aoffset += lda;
kusano 2b45e8
    
kusano 2b45e8
    boffset1  = boffset;
kusano 2b45e8
    boffset  += 8;
kusano 2b45e8
    
kusano 2b45e8
    i = (n >> 3);
kusano 2b45e8
    if (i > 0){
kusano 2b45e8
      do{
kusano 2b45e8
	ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
	ctemp02 = *(aoffset1 + 1);
kusano 2b45e8
	ctemp03 = *(aoffset1 + 2);
kusano 2b45e8
	ctemp04 = *(aoffset1 + 3);
kusano 2b45e8
	ctemp05 = *(aoffset1 + 4);
kusano 2b45e8
	ctemp06 = *(aoffset1 + 5);
kusano 2b45e8
	ctemp07 = *(aoffset1 + 6);
kusano 2b45e8
	ctemp08 = *(aoffset1 + 7);
kusano 2b45e8
	aoffset1 += 8;
kusano 2b45e8
	
kusano 2b45e8
	*(boffset1 +  0) = ctemp01;
kusano 2b45e8
	*(boffset1 +  1) = ctemp02;
kusano 2b45e8
	*(boffset1 +  2) = ctemp03;
kusano 2b45e8
	*(boffset1 +  3) = ctemp04;
kusano 2b45e8
	*(boffset1 +  4) = ctemp05;
kusano 2b45e8
	*(boffset1 +  5) = ctemp06;
kusano 2b45e8
	*(boffset1 +  6) = ctemp07;
kusano 2b45e8
	*(boffset1 +  7) = ctemp08;
kusano 2b45e8
	
kusano 2b45e8
	boffset1 += 8 * m;
kusano 2b45e8
	 i --;
kusano 2b45e8
       }while(i > 0);
kusano 2b45e8
     }
kusano 2b45e8
kusano 2b45e8
     if (n & 4){
kusano 2b45e8
       ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
       ctemp02 = *(aoffset1 + 1);
kusano 2b45e8
       ctemp03 = *(aoffset1 + 2);
kusano 2b45e8
       ctemp04 = *(aoffset1 + 3);
kusano 2b45e8
       aoffset1 += 4;
kusano 2b45e8
kusano 2b45e8
       *(boffset2 +  0) = ctemp01;
kusano 2b45e8
       *(boffset2 +  1) = ctemp02;
kusano 2b45e8
       *(boffset2 +  2) = ctemp03;
kusano 2b45e8
       *(boffset2 +  3) = ctemp04;
kusano 2b45e8
       boffset2 += 4;
kusano 2b45e8
     }
kusano 2b45e8
kusano 2b45e8
     if (n & 2){
kusano 2b45e8
       ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
       ctemp02 = *(aoffset1 + 1);
kusano 2b45e8
       aoffset1 += 2;
kusano 2b45e8
kusano 2b45e8
       *(boffset3 +  0) = ctemp01;
kusano 2b45e8
       *(boffset3 +  1) = ctemp02;
kusano 2b45e8
       boffset3 += 2;
kusano 2b45e8
     }
kusano 2b45e8
     
kusano 2b45e8
     if (n & 1){
kusano 2b45e8
       ctemp01 = *(aoffset1 + 0);
kusano 2b45e8
       aoffset1 ++;
kusano 2b45e8
      *(boffset4 +  0) = ctemp01;
kusano 2b45e8
      boffset4 ++;
kusano 2b45e8
    }
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  return 0;
kusano 2b45e8
}