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
  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;
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
kusano 2b45e8
  aoffset = a;
kusano 2b45e8
  boffset = b;
kusano 2b45e8
  lda *= 2;
kusano 2b45e8
  
kusano 2b45e8
  j = (n >> 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
      i = m;
kusano 2b45e8
      if (i > 0){
kusano 2b45e8
	do{
kusano 2b45e8
	  ctemp01 = *(aoffset1 +  0);
kusano 2b45e8
	  ctemp02 = *(aoffset1 +  1);
kusano 2b45e8
	  ctemp03 = *(aoffset2 +  0);
kusano 2b45e8
	  ctemp04 = *(aoffset2 +  1);
kusano 2b45e8
	  ctemp05 = *(aoffset3 +  0);
kusano 2b45e8
	  ctemp06 = *(aoffset3 +  1);
kusano 2b45e8
	  ctemp07 = *(aoffset4 +  0);
kusano 2b45e8
	  ctemp08 = *(aoffset4 +  1);
kusano 2b45e8
	  ctemp09 = *(aoffset5 +  0);
kusano 2b45e8
	  ctemp10 = *(aoffset5 +  1);
kusano 2b45e8
	  ctemp11 = *(aoffset6 +  0);
kusano 2b45e8
	  ctemp12 = *(aoffset6 +  1);
kusano 2b45e8
	  ctemp13 = *(aoffset7 +  0);
kusano 2b45e8
	  ctemp14 = *(aoffset7 +  1);
kusano 2b45e8
	  ctemp15 = *(aoffset8 +  0);
kusano 2b45e8
	  ctemp16 = *(aoffset8 +  1);
kusano 2b45e8
	  
kusano 2b45e8
	  *(boffset +  0) = ctemp01;
kusano 2b45e8
	  *(boffset +  1) = ctemp02;
kusano 2b45e8
	  *(boffset +  2) = ctemp03;
kusano 2b45e8
	  *(boffset +  3) = ctemp04;
kusano 2b45e8
	  *(boffset +  4) = ctemp05;
kusano 2b45e8
	  *(boffset +  5) = ctemp06;
kusano 2b45e8
	  *(boffset +  6) = ctemp07;
kusano 2b45e8
	  *(boffset +  7) = ctemp08;
kusano 2b45e8
	  *(boffset +  8) = ctemp09;
kusano 2b45e8
	  *(boffset +  9) = ctemp10;
kusano 2b45e8
	  *(boffset + 10) = ctemp11;
kusano 2b45e8
	  *(boffset + 11) = ctemp12;
kusano 2b45e8
	  *(boffset + 12) = ctemp13;
kusano 2b45e8
	  *(boffset + 13) = ctemp14;
kusano 2b45e8
	  *(boffset + 14) = ctemp15;
kusano 2b45e8
	  *(boffset + 15) = ctemp16;
kusano 2b45e8
kusano 2b45e8
	  aoffset1 += 2;
kusano 2b45e8
	  aoffset2 += 2;
kusano 2b45e8
	  aoffset3 += 2;
kusano 2b45e8
	  aoffset4 += 2;
kusano 2b45e8
	  aoffset5 += 2;
kusano 2b45e8
	  aoffset6 += 2;
kusano 2b45e8
	  aoffset7 += 2;
kusano 2b45e8
	  aoffset8 += 2;
kusano 2b45e8
	  
kusano 2b45e8
	  boffset += 16;
kusano 2b45e8
	  i --;
kusano 2b45e8
	}while(i > 0);
kusano 2b45e8
      }
kusano 2b45e8
      j--;
kusano 2b45e8
    }while(j > 0);
kusano 2b45e8
  } /* end of if(j > 0) */
kusano 2b45e8
  
kusano 2b45e8
  if (n & 4){
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
    i = m;
kusano 2b45e8
    if (i > 0){
kusano 2b45e8
      do{
kusano 2b45e8
	ctemp01 = *(aoffset1 +  0);
kusano 2b45e8
	ctemp02 = *(aoffset1 +  1);
kusano 2b45e8
	ctemp03 = *(aoffset2 +  0);
kusano 2b45e8
	ctemp04 = *(aoffset2 +  1);
kusano 2b45e8
	ctemp05 = *(aoffset3 +  0);
kusano 2b45e8
	ctemp06 = *(aoffset3 +  1);
kusano 2b45e8
	ctemp07 = *(aoffset4 +  0);
kusano 2b45e8
	ctemp08 = *(aoffset4 +  1);
kusano 2b45e8
	
kusano 2b45e8
	*(boffset +  0) = ctemp01;
kusano 2b45e8
	*(boffset +  1) = ctemp02;
kusano 2b45e8
	*(boffset +  2) = ctemp03;
kusano 2b45e8
	*(boffset +  3) = ctemp04;
kusano 2b45e8
	*(boffset +  4) = ctemp05;
kusano 2b45e8
	*(boffset +  5) = ctemp06;
kusano 2b45e8
	*(boffset +  6) = ctemp07;
kusano 2b45e8
	*(boffset +  7) = ctemp08;
kusano 2b45e8
kusano 2b45e8
	aoffset1 += 2;
kusano 2b45e8
	aoffset2 += 2;
kusano 2b45e8
	aoffset3 += 2;
kusano 2b45e8
	aoffset4 += 2;
kusano 2b45e8
kusano 2b45e8
	boffset += 8;
kusano 2b45e8
	i --;
kusano 2b45e8
      }while(i > 0);
kusano 2b45e8
    }
kusano 2b45e8
  } /* end of if(j > 0) */
kusano 2b45e8
kusano 2b45e8
  if (n & 2){
kusano 2b45e8
    aoffset1  = aoffset;
kusano 2b45e8
    aoffset2  = aoffset1 + lda;
kusano 2b45e8
    aoffset += 2 * lda;
kusano 2b45e8
    
kusano 2b45e8
    i = m;
kusano 2b45e8
    if (i > 0){
kusano 2b45e8
      do{
kusano 2b45e8
	ctemp01 = *(aoffset1 +  0);
kusano 2b45e8
	ctemp02 = *(aoffset1 +  1);
kusano 2b45e8
	ctemp03 = *(aoffset2 +  0);
kusano 2b45e8
	ctemp04 = *(aoffset2 +  1);
kusano 2b45e8
	
kusano 2b45e8
	*(boffset +  0) = ctemp01;
kusano 2b45e8
	*(boffset +  1) = ctemp02;
kusano 2b45e8
	*(boffset +  2) = ctemp03;
kusano 2b45e8
	*(boffset +  3) = ctemp04;
kusano 2b45e8
	
kusano 2b45e8
	aoffset1 +=  2;
kusano 2b45e8
	aoffset2 +=  2;
kusano 2b45e8
	boffset  +=  4;
kusano 2b45e8
	i --;
kusano 2b45e8
      }while(i > 0);
kusano 2b45e8
    }
kusano 2b45e8
   
kusano 2b45e8
  } /* end of if(j > 0) */
kusano 2b45e8
kusano 2b45e8
  if (n & 1){
kusano 2b45e8
    aoffset1  = aoffset;
kusano 2b45e8
    
kusano 2b45e8
    i = m;
kusano 2b45e8
    if (i > 0){
kusano 2b45e8
      do{
kusano 2b45e8
	ctemp01 = *(aoffset1 +  0);
kusano 2b45e8
	ctemp02 = *(aoffset1 +  1);
kusano 2b45e8
kusano 2b45e8
	*(boffset +  0) = ctemp01;
kusano 2b45e8
	*(boffset +  1) = ctemp02;
kusano 2b45e8
kusano 2b45e8
	aoffset1 += 2;
kusano 2b45e8
	boffset  += 2;
kusano 2b45e8
	i --;
kusano 2b45e8
      }while(i > 0);
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
  } /* end of if(j > 0) */
kusano 2b45e8
kusano 2b45e8
  return 0;
kusano 2b45e8
}