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;
kusano 2b45e8
  FLOAT *boffset;
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
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
  j = (n >> 4);
kusano 2b45e8
  if (j > 0){
kusano 2b45e8
    do{
kusano 2b45e8
      aoffset1  = aoffset;
kusano 2b45e8
      aoffset2  = aoffset + lda;
kusano 2b45e8
      aoffset += 16;
kusano 2b45e8
      
kusano 2b45e8
      i = (m >> 1);
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
	  ctemp09 = *(aoffset1 +  8);
kusano 2b45e8
	  ctemp10 = *(aoffset1 +  9);
kusano 2b45e8
	  ctemp11 = *(aoffset1 + 10);
kusano 2b45e8
	  ctemp12 = *(aoffset1 + 11);
kusano 2b45e8
	  ctemp13 = *(aoffset1 + 12);
kusano 2b45e8
	  ctemp14 = *(aoffset1 + 13);
kusano 2b45e8
	  ctemp15 = *(aoffset1 + 14);
kusano 2b45e8
	  ctemp16 = *(aoffset1 + 15);
kusano 2b45e8
kusano 2b45e8
	  ctemp17 = *(aoffset2 +  0);
kusano 2b45e8
	  ctemp18 = *(aoffset2 +  1);
kusano 2b45e8
	  ctemp19 = *(aoffset2 +  2);
kusano 2b45e8
	  ctemp20 = *(aoffset2 +  3);
kusano 2b45e8
	  ctemp21 = *(aoffset2 +  4);
kusano 2b45e8
	  ctemp22 = *(aoffset2 +  5);
kusano 2b45e8
	  ctemp23 = *(aoffset2 +  6);
kusano 2b45e8
	  ctemp24 = *(aoffset2 +  7);
kusano 2b45e8
	  ctemp25 = *(aoffset2 +  8);
kusano 2b45e8
	  ctemp26 = *(aoffset2 +  9);
kusano 2b45e8
	  ctemp27 = *(aoffset2 + 10);
kusano 2b45e8
	  ctemp28 = *(aoffset2 + 11);
kusano 2b45e8
	  ctemp29 = *(aoffset2 + 12);
kusano 2b45e8
	  ctemp30 = *(aoffset2 + 13);
kusano 2b45e8
	  ctemp31 = *(aoffset2 + 14);
kusano 2b45e8
	  ctemp32 = *(aoffset2 + 15);
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
	  *(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
	  *(boffset + 16) = -ctemp17;
kusano 2b45e8
	  *(boffset + 17) = -ctemp18;
kusano 2b45e8
	  *(boffset + 18) = -ctemp19;
kusano 2b45e8
	  *(boffset + 19) = -ctemp20;
kusano 2b45e8
	  *(boffset + 20) = -ctemp21;
kusano 2b45e8
	  *(boffset + 21) = -ctemp22;
kusano 2b45e8
	  *(boffset + 22) = -ctemp23;
kusano 2b45e8
	  *(boffset + 23) = -ctemp24;
kusano 2b45e8
kusano 2b45e8
	  *(boffset + 24) = -ctemp25;
kusano 2b45e8
	  *(boffset + 25) = -ctemp26;
kusano 2b45e8
	  *(boffset + 26) = -ctemp27;
kusano 2b45e8
	  *(boffset + 27) = -ctemp28;
kusano 2b45e8
	  *(boffset + 28) = -ctemp29;
kusano 2b45e8
	  *(boffset + 29) = -ctemp30;
kusano 2b45e8
	  *(boffset + 30) = -ctemp31;
kusano 2b45e8
	  *(boffset + 31) = -ctemp32;
kusano 2b45e8
kusano 2b45e8
	  aoffset1 +=  2 * lda;
kusano 2b45e8
	  aoffset2 +=  2 * lda;
kusano 2b45e8
	  boffset   += 32;
kusano 2b45e8
kusano 2b45e8
	  i --;
kusano 2b45e8
	}while(i > 0);
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      if (m & 1){
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
	ctemp09 = *(aoffset1 +  8);
kusano 2b45e8
	ctemp10 = *(aoffset1 +  9);
kusano 2b45e8
	ctemp11 = *(aoffset1 + 10);
kusano 2b45e8
	ctemp12 = *(aoffset1 + 11);
kusano 2b45e8
	ctemp13 = *(aoffset1 + 12);
kusano 2b45e8
	ctemp14 = *(aoffset1 + 13);
kusano 2b45e8
	ctemp15 = *(aoffset1 + 14);
kusano 2b45e8
	ctemp16 = *(aoffset1 + 15);
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
	*(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
	boffset   += 16;
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      j--;
kusano 2b45e8
    }while(j > 0);
kusano 2b45e8
  } /* end of if(j > 0) */
kusano 2b45e8
kusano 2b45e8
  if (n & 8){
kusano 2b45e8
    aoffset1  = aoffset;
kusano 2b45e8
    aoffset2  = aoffset + lda;
kusano 2b45e8
    aoffset += 8;
kusano 2b45e8
    
kusano 2b45e8
    i = (m >> 1);
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
	  
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
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
	*(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 * lda;
kusano 2b45e8
	aoffset2 +=  2 * lda;
kusano 2b45e8
	boffset   += 16;
kusano 2b45e8
	
kusano 2b45e8
	i --;
kusano 2b45e8
      }while(i > 0);
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
    if (m & 1){
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
	
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
      boffset   += 8;
kusano 2b45e8
    }
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (n & 4){
kusano 2b45e8
    aoffset1  = aoffset;
kusano 2b45e8
    aoffset2  = aoffset + lda;
kusano 2b45e8
    aoffset += 4;
kusano 2b45e8
    
kusano 2b45e8
    i = (m >> 1);
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
kusano 2b45e8
	ctemp05 = *(aoffset2 +  0);
kusano 2b45e8
	ctemp06 = *(aoffset2 +  1);
kusano 2b45e8
	ctemp07 = *(aoffset2 +  2);
kusano 2b45e8
	ctemp08 = *(aoffset2 +  3);
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 * lda;
kusano 2b45e8
	aoffset2 +=  2 * lda;
kusano 2b45e8
	boffset   += 8;
kusano 2b45e8
	
kusano 2b45e8
	i --;
kusano 2b45e8
      }while(i > 0);
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
    if (m & 1){
kusano 2b45e8
      ctemp01 = *(aoffset1 +  0);
kusano 2b45e8
      ctemp02 = *(aoffset1 +  1);
kusano 2b45e8
      ctemp03 = *(aoffset1 +  2);
kusano 2b45e8
      ctemp04 = *(aoffset1 +  3);
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
      boffset   += 4;
kusano 2b45e8
    }
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (n & 2){
kusano 2b45e8
    aoffset1  = aoffset;
kusano 2b45e8
    aoffset2  = aoffset + lda;
kusano 2b45e8
    aoffset += 2;
kusano 2b45e8
    
kusano 2b45e8
    i = (m >> 1);
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 * lda;
kusano 2b45e8
	aoffset2 +=  2 * lda;
kusano 2b45e8
	boffset   += 4;
kusano 2b45e8
	
kusano 2b45e8
	i --;
kusano 2b45e8
      }while(i > 0);
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
    if (m & 1){
kusano 2b45e8
      ctemp01 = *(aoffset1 +  0);
kusano 2b45e8
      ctemp02 = *(aoffset1 +  1);
kusano 2b45e8
kusano 2b45e8
      *(boffset +  0) = -ctemp01;
kusano 2b45e8
      *(boffset +  1) = -ctemp02;
kusano 2b45e8
      boffset   += 2;
kusano 2b45e8
    }
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (n & 1){
kusano 2b45e8
    aoffset1  = aoffset;
kusano 2b45e8
    aoffset2  = aoffset + lda;
kusano 2b45e8
    
kusano 2b45e8
    i = (m >> 1);
kusano 2b45e8
    if (i > 0){
kusano 2b45e8
      do{
kusano 2b45e8
	ctemp01 = *(aoffset1 +  0);
kusano 2b45e8
	ctemp02 = *(aoffset2 +  0);
kusano 2b45e8
kusano 2b45e8
	*(boffset +  0) = -ctemp01;
kusano 2b45e8
	*(boffset +  1) = -ctemp02;
kusano 2b45e8
kusano 2b45e8
	aoffset1 +=  2 * lda;
kusano 2b45e8
	aoffset2 +=  2 * lda;
kusano 2b45e8
	boffset   += 2;
kusano 2b45e8
	
kusano 2b45e8
	i --;
kusano 2b45e8
      }while(i > 0);
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
    if (m & 1){
kusano 2b45e8
      ctemp01 = *(aoffset1 +  0);
kusano 2b45e8
      *(boffset +  0) = -ctemp01;
kusano 2b45e8
      boffset   += 1;
kusano 2b45e8
    }
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  return 0;
kusano 2b45e8
}