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, BLASLONG posX, BLASLONG posY, FLOAT *b){
kusano 2b45e8
kusano 2b45e8
  BLASLONG i, js;
kusano 2b45e8
  BLASLONG X;
kusano 2b45e8
kusano 2b45e8
  FLOAT data01, data02, data03, data04, data05, data06, data07, data08;
kusano 2b45e8
  FLOAT data09, data10, data11, data12, data13, data14, data15, data16;
kusano 2b45e8
  FLOAT *ao1, *ao2, *ao3, *ao4;
kusano 2b45e8
kusano 2b45e8
  js = (n >> 2);
kusano 2b45e8
kusano 2b45e8
  if (js > 0){
kusano 2b45e8
    do {
kusano 2b45e8
      X = posX;
kusano 2b45e8
kusano 2b45e8
      if (posX <= posY) {
kusano 2b45e8
	ao1 = a + posX + (posY + 0) * lda;
kusano 2b45e8
	ao2 = a + posX + (posY + 1) * lda;
kusano 2b45e8
	ao3 = a + posX + (posY + 2) * lda;
kusano 2b45e8
	ao4 = a + posX + (posY + 3) * lda;
kusano 2b45e8
      } else {
kusano 2b45e8
	ao1 = a + posY + (posX + 0) * lda;
kusano 2b45e8
	ao2 = a + posY + (posX + 1) * lda;
kusano 2b45e8
	ao3 = a + posY + (posX + 2) * lda;
kusano 2b45e8
	ao4 = a + posY + (posX + 3) * lda;
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      i = (m >> 2);
kusano 2b45e8
      if (i > 0) {
kusano 2b45e8
	do {
kusano 2b45e8
	  if (X < posY) {
kusano 2b45e8
	    data01 = *(ao1 + 0);
kusano 2b45e8
	    data02 = *(ao1 + 1);
kusano 2b45e8
	    data03 = *(ao1 + 2);
kusano 2b45e8
	    data04 = *(ao1 + 3);
kusano 2b45e8
	    
kusano 2b45e8
	    data05 = *(ao2 + 0);
kusano 2b45e8
	    data06 = *(ao2 + 1);
kusano 2b45e8
	    data07 = *(ao2 + 2);
kusano 2b45e8
	    data08 = *(ao2 + 3);
kusano 2b45e8
	    
kusano 2b45e8
	    data09 = *(ao3 + 0);
kusano 2b45e8
	    data10 = *(ao3 + 1);
kusano 2b45e8
	    data11 = *(ao3 + 2);
kusano 2b45e8
	    data12 = *(ao3 + 3);
kusano 2b45e8
	    
kusano 2b45e8
	    data13 = *(ao4 + 0);
kusano 2b45e8
	    data14 = *(ao4 + 1);
kusano 2b45e8
	    data15 = *(ao4 + 2);
kusano 2b45e8
	    data16 = *(ao4 + 3);
kusano 2b45e8
	  
kusano 2b45e8
	    b[ 0] = data01;
kusano 2b45e8
	    b[ 1] = data05;
kusano 2b45e8
	    b[ 2] = data09;
kusano 2b45e8
	    b[ 3] = data13;
kusano 2b45e8
	    b[ 4] = data02;
kusano 2b45e8
	    b[ 5] = data06;
kusano 2b45e8
	    b[ 6] = data10;
kusano 2b45e8
	    b[ 7] = data14;
kusano 2b45e8
	    
kusano 2b45e8
	    b[ 8] = data03;
kusano 2b45e8
	    b[ 9] = data07;
kusano 2b45e8
	    b[10] = data11;
kusano 2b45e8
	    b[11] = data15;
kusano 2b45e8
	    b[12] = data04;
kusano 2b45e8
	    b[13] = data08;
kusano 2b45e8
	    b[14] = data12;
kusano 2b45e8
	    b[15] = data16;
kusano 2b45e8
	    
kusano 2b45e8
	    ao1 += 4;
kusano 2b45e8
	    ao2 += 4;
kusano 2b45e8
	    ao3 += 4;
kusano 2b45e8
	    ao4 += 4;
kusano 2b45e8
	    b += 16;
kusano 2b45e8
	  } else 
kusano 2b45e8
	  if (X > posY) {
kusano 2b45e8
	    ao1 += 4 * lda;
kusano 2b45e8
	    ao2 += 4 * lda;
kusano 2b45e8
	    ao3 += 4 * lda;
kusano 2b45e8
	    ao4 += 4 * lda;
kusano 2b45e8
kusano 2b45e8
	    b   += 16;
kusano 2b45e8
	    } else {
kusano 2b45e8
#ifdef UNIT
kusano 2b45e8
	      data05 = *(ao2 + 0);
kusano 2b45e8
	    
kusano 2b45e8
	      data09 = *(ao3 + 0);
kusano 2b45e8
	      data10 = *(ao3 + 1);
kusano 2b45e8
	      
kusano 2b45e8
	      data13 = *(ao4 + 0);
kusano 2b45e8
	      data14 = *(ao4 + 1);
kusano 2b45e8
	      data15 = *(ao4 + 2);
kusano 2b45e8
	      
kusano 2b45e8
	      b[ 0] = ONE;
kusano 2b45e8
	      b[ 1] = data05;
kusano 2b45e8
	      b[ 2] = data09;
kusano 2b45e8
	      b[ 3] = data13;
kusano 2b45e8
kusano 2b45e8
	      b[ 4] = ZERO;
kusano 2b45e8
	      b[ 5] = ONE;
kusano 2b45e8
	      b[ 6] = data10;
kusano 2b45e8
	      b[ 7] = data14;
kusano 2b45e8
	      
kusano 2b45e8
	      b[ 8] = ZERO;
kusano 2b45e8
	      b[ 9] = ZERO;
kusano 2b45e8
	      b[10] = ONE;
kusano 2b45e8
	      b[11] = data15;
kusano 2b45e8
kusano 2b45e8
	      b[12] = ZERO;
kusano 2b45e8
	      b[13] = ZERO;
kusano 2b45e8
	      b[14] = ZERO;
kusano 2b45e8
	      b[15] = ONE;
kusano 2b45e8
#else
kusano 2b45e8
	      data01 = *(ao1 + 0);
kusano 2b45e8
	      
kusano 2b45e8
	      data05 = *(ao2 + 0);
kusano 2b45e8
	      data06 = *(ao2 + 1);
kusano 2b45e8
	      
kusano 2b45e8
	      data09 = *(ao3 + 0);
kusano 2b45e8
	      data10 = *(ao3 + 1);
kusano 2b45e8
	      data11 = *(ao3 + 2);
kusano 2b45e8
	      
kusano 2b45e8
	      data13 = *(ao4 + 0);
kusano 2b45e8
	      data14 = *(ao4 + 1);
kusano 2b45e8
	      data15 = *(ao4 + 2);
kusano 2b45e8
	      data16 = *(ao4 + 3);
kusano 2b45e8
	      
kusano 2b45e8
	      b[ 0] = data01;
kusano 2b45e8
	      b[ 1] = data05;
kusano 2b45e8
	      b[ 2] = data09;
kusano 2b45e8
	      b[ 3] = data13;
kusano 2b45e8
kusano 2b45e8
	      b[ 4] = ZERO;
kusano 2b45e8
	      b[ 5] = data06;
kusano 2b45e8
	      b[ 6] = data10;
kusano 2b45e8
	      b[ 7] = data14;
kusano 2b45e8
	      
kusano 2b45e8
	      b[ 8] = ZERO;
kusano 2b45e8
	      b[ 9] = ZERO;
kusano 2b45e8
	      b[10] = data11;
kusano 2b45e8
	      b[11] = data15;
kusano 2b45e8
kusano 2b45e8
	      b[12] = ZERO;
kusano 2b45e8
	      b[13] = ZERO;
kusano 2b45e8
	      b[14] = ZERO;
kusano 2b45e8
	      b[15] = data16;
kusano 2b45e8
#endif
kusano 2b45e8
	      ao1 += 4;
kusano 2b45e8
	      ao2 += 4;
kusano 2b45e8
	      ao3 += 4;
kusano 2b45e8
	      ao4 += 4;
kusano 2b45e8
     
kusano 2b45e8
	      b += 16;
kusano 2b45e8
	    }
kusano 2b45e8
kusano 2b45e8
	  X += 4;
kusano 2b45e8
	  i --;
kusano 2b45e8
	} while (i > 0);
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      i = (m & 3);
kusano 2b45e8
      if (i) {
kusano 2b45e8
      
kusano 2b45e8
	if (X < posY) {
kusano 2b45e8
kusano 2b45e8
	  if (m & 2) {
kusano 2b45e8
	    data01 = *(ao1 + 0);
kusano 2b45e8
	    data02 = *(ao1 + 1);
kusano 2b45e8
	    data03 = *(ao2 + 0);
kusano 2b45e8
	    data04 = *(ao2 + 1);
kusano 2b45e8
	    data05 = *(ao3 + 0);
kusano 2b45e8
	    data06 = *(ao3 + 1);
kusano 2b45e8
	    data07 = *(ao4 + 0);
kusano 2b45e8
	    data08 = *(ao4 + 1);
kusano 2b45e8
kusano 2b45e8
	    b[ 0] = data01;
kusano 2b45e8
	    b[ 1] = data03;
kusano 2b45e8
	    b[ 2] = data05;
kusano 2b45e8
	    b[ 3] = data07;
kusano 2b45e8
	    b[ 4] = data02;
kusano 2b45e8
	    b[ 5] = data04;
kusano 2b45e8
	    b[ 6] = data06;
kusano 2b45e8
	    b[ 7] = data08;
kusano 2b45e8
	  
kusano 2b45e8
	    ao1 += 2;
kusano 2b45e8
	    ao2 += 2;
kusano 2b45e8
	    ao3 += 2;
kusano 2b45e8
	    ao4 += 2;
kusano 2b45e8
	    b += 8;
kusano 2b45e8
	  }
kusano 2b45e8
	  
kusano 2b45e8
	  if (m & 1) {
kusano 2b45e8
	    data01 = *(ao1 + 0);
kusano 2b45e8
	    data03 = *(ao2 + 0);
kusano 2b45e8
	    data05 = *(ao3 + 0);
kusano 2b45e8
	    data07 = *(ao4 + 0);
kusano 2b45e8
kusano 2b45e8
	    b[ 0] = data01;
kusano 2b45e8
	    b[ 1] = data03;
kusano 2b45e8
	    b[ 2] = data05;
kusano 2b45e8
	    b[ 3] = data07;
kusano 2b45e8
kusano 2b45e8
	    ao1 += 1;
kusano 2b45e8
	    ao2 += 1;
kusano 2b45e8
	    ao3 += 1;
kusano 2b45e8
	    ao4 += 1;
kusano 2b45e8
	    b += 4;
kusano 2b45e8
	  }
kusano 2b45e8
	  
kusano 2b45e8
	} else 
kusano 2b45e8
	  if (X > posY) {
kusano 2b45e8
	    if (m & 2) {
kusano 2b45e8
	      ao1 += 2 * lda;
kusano 2b45e8
	      ao2 += 2 * lda;
kusano 2b45e8
	      b   += 8;
kusano 2b45e8
	    }
kusano 2b45e8
	    
kusano 2b45e8
	    if (m & 1) {
kusano 2b45e8
	      ao1 += lda;
kusano 2b45e8
	      b += 4;
kusano 2b45e8
	    }
kusano 2b45e8
    
kusano 2b45e8
	  } else {
kusano 2b45e8
#ifdef UNIT
kusano 2b45e8
	    data05 = *(ao2 + 0);
kusano 2b45e8
	    data09 = *(ao3 + 0);
kusano 2b45e8
	    data13 = *(ao4 + 0);
kusano 2b45e8
kusano 2b45e8
	    if (i >= 2) {
kusano 2b45e8
	      data10 = *(ao3 + 1);
kusano 2b45e8
	      data14 = *(ao4 + 1);
kusano 2b45e8
	    }
kusano 2b45e8
kusano 2b45e8
	    if (i >= 3) {
kusano 2b45e8
	      data15 = *(ao4 + 2);
kusano 2b45e8
	    }
kusano 2b45e8
	    
kusano 2b45e8
	    b[ 0] = ONE;
kusano 2b45e8
	    b[ 1] = data05;
kusano 2b45e8
	    b[ 2] = data09;
kusano 2b45e8
	    b[ 3] = data13;
kusano 2b45e8
	    b += 4;
kusano 2b45e8
	      
kusano 2b45e8
	    if(i >= 2) {
kusano 2b45e8
	      b[ 0] = ZERO;
kusano 2b45e8
	      b[ 1] = ONE;
kusano 2b45e8
	      b[ 2] = data10;
kusano 2b45e8
	      b[ 3] = data14;
kusano 2b45e8
	      b += 4;
kusano 2b45e8
	    }
kusano 2b45e8
	      
kusano 2b45e8
	    if (i >= 3) {
kusano 2b45e8
	      b[ 0] = ZERO;
kusano 2b45e8
	      b[ 1] = ZERO;
kusano 2b45e8
	      b[ 2] = ONE;
kusano 2b45e8
	      b[ 3] = data15;
kusano 2b45e8
	      b += 4;
kusano 2b45e8
	    }
kusano 2b45e8
#else
kusano 2b45e8
	    data01 = *(ao1 + 0);
kusano 2b45e8
	    data05 = *(ao2 + 0);
kusano 2b45e8
	    data09 = *(ao3 + 0);
kusano 2b45e8
	    data13 = *(ao4 + 0);
kusano 2b45e8
kusano 2b45e8
	    if (i >= 2) {
kusano 2b45e8
	      data06 = *(ao2 + 1);
kusano 2b45e8
	      data10 = *(ao3 + 1);
kusano 2b45e8
	      data14 = *(ao4 + 1);
kusano 2b45e8
	    }
kusano 2b45e8
kusano 2b45e8
	    if (i >= 3) {
kusano 2b45e8
	      data11 = *(ao3 + 2);
kusano 2b45e8
	      data15 = *(ao4 + 2);
kusano 2b45e8
	    }
kusano 2b45e8
	    
kusano 2b45e8
	    b[ 0] = data01;
kusano 2b45e8
	    b[ 1] = data05;
kusano 2b45e8
	    b[ 2] = data09;
kusano 2b45e8
	    b[ 3] = data13;
kusano 2b45e8
	    b += 4;
kusano 2b45e8
	      
kusano 2b45e8
	    if(i >= 2) {
kusano 2b45e8
	      b[ 0] = ZERO;
kusano 2b45e8
	      b[ 1] = data06;
kusano 2b45e8
	      b[ 2] = data10;
kusano 2b45e8
	      b[ 3] = data14;
kusano 2b45e8
	      b += 4;
kusano 2b45e8
	    }
kusano 2b45e8
	      
kusano 2b45e8
	    if (i >= 3) {
kusano 2b45e8
	      b[ 0] = ZERO;
kusano 2b45e8
	      b[ 1] = ZERO;
kusano 2b45e8
	      b[ 2] = data11;
kusano 2b45e8
	      b[ 3] = data15;
kusano 2b45e8
	      b += 4;
kusano 2b45e8
	    }
kusano 2b45e8
#endif
kusano 2b45e8
	  }
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      posY += 4;
kusano 2b45e8
      js --;
kusano 2b45e8
    } while (js > 0);
kusano 2b45e8
  } /* End of main loop */
kusano 2b45e8
kusano 2b45e8
  if (n & 2){
kusano 2b45e8
      X = posX;
kusano 2b45e8
kusano 2b45e8
      if (posX <= posY) {
kusano 2b45e8
	ao1 = a + posX + (posY + 0) * lda;
kusano 2b45e8
	ao2 = a + posX + (posY + 1) * lda;
kusano 2b45e8
      } else {
kusano 2b45e8
	ao1 = a + posY + (posX + 0) * lda;
kusano 2b45e8
	ao2 = a + posY + (posX + 1) * lda;
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      i = (m >> 1);
kusano 2b45e8
      if (i > 0) {
kusano 2b45e8
	do {
kusano 2b45e8
	  if (X < posY) {
kusano 2b45e8
	    data01 = *(ao1 + 0);
kusano 2b45e8
	    data02 = *(ao1 + 1);
kusano 2b45e8
	    data05 = *(ao2 + 0);
kusano 2b45e8
	    data06 = *(ao2 + 1);
kusano 2b45e8
	    
kusano 2b45e8
	    b[ 0] = data01;
kusano 2b45e8
	    b[ 1] = data05;
kusano 2b45e8
	    b[ 2] = data02;
kusano 2b45e8
	    b[ 3] = data06;
kusano 2b45e8
	    
kusano 2b45e8
	    ao1 += 2;
kusano 2b45e8
	    ao2 += 2;
kusano 2b45e8
	    b += 4;
kusano 2b45e8
kusano 2b45e8
	  } else 
kusano 2b45e8
	    if (X > posY) {
kusano 2b45e8
	      ao1 += 2 * lda;
kusano 2b45e8
	      ao2 += 2 * lda;
kusano 2b45e8
	      b += 4;
kusano 2b45e8
kusano 2b45e8
	    } else {
kusano 2b45e8
#ifdef UNIT
kusano 2b45e8
	      data05 = *(ao2 + 0);
kusano 2b45e8
kusano 2b45e8
	      b[ 0] = ONE;
kusano 2b45e8
	      b[ 1] = data05;
kusano 2b45e8
	      b[ 2] = ZERO;
kusano 2b45e8
	      b[ 3] = ONE;
kusano 2b45e8
#else
kusano 2b45e8
	      data01 = *(ao1 + 0);
kusano 2b45e8
	      data05 = *(ao2 + 0);
kusano 2b45e8
	      data06 = *(ao2 + 1);
kusano 2b45e8
kusano 2b45e8
	      b[ 0] = data01;
kusano 2b45e8
	      b[ 1] = data05;
kusano 2b45e8
	      b[ 2] = ZERO;
kusano 2b45e8
	      b[ 3] = data06;
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	      ao1 += 2 * lda;
kusano 2b45e8
	      ao2 += 2 * lda;
kusano 2b45e8
	      
kusano 2b45e8
	      b += 4;
kusano 2b45e8
	    }
kusano 2b45e8
kusano 2b45e8
	  X += 2;
kusano 2b45e8
	  i --;
kusano 2b45e8
	} while (i > 0);
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      i = (m & 1);
kusano 2b45e8
      if (i) {
kusano 2b45e8
      
kusano 2b45e8
	if (X < posY) {
kusano 2b45e8
	  data01 = *(ao1 + 0);
kusano 2b45e8
	  data05 = *(ao2 + 0);
kusano 2b45e8
kusano 2b45e8
	  b[ 0] = data01;
kusano 2b45e8
	  b[ 1] = data05;
kusano 2b45e8
	  ao1 += 1;
kusano 2b45e8
	  ao2 += 1;
kusano 2b45e8
	  b += 2;
kusano 2b45e8
	} else 
kusano 2b45e8
	  if (X > posY) {
kusano 2b45e8
	    ao1 += lda;
kusano 2b45e8
	    ao2 += lda;
kusano 2b45e8
	    b += 2;
kusano 2b45e8
	  } else {
kusano 2b45e8
#ifdef UNIT
kusano 2b45e8
	    data05 = *(ao2 + 0);
kusano 2b45e8
	    b[ 0] = ONE;
kusano 2b45e8
	    b[ 1] = data05;
kusano 2b45e8
#else
kusano 2b45e8
	    data01 = *(ao1 + 0);
kusano 2b45e8
	    data05 = *(ao2 + 0);
kusano 2b45e8
kusano 2b45e8
	    b[ 0] = data01;
kusano 2b45e8
	    b[ 1] = data05;
kusano 2b45e8
#endif
kusano 2b45e8
	    ao1 += lda;
kusano 2b45e8
	    ao2 += lda;
kusano 2b45e8
	    b += 2;
kusano 2b45e8
	  }
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      posY += 2;
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (n & 1){
kusano 2b45e8
      X = posX;
kusano 2b45e8
kusano 2b45e8
      if (posX <= posY) {
kusano 2b45e8
	ao1 = a + posX + (posY + 0) * lda;
kusano 2b45e8
      } else {
kusano 2b45e8
	ao1 = a + posY + (posX + 0) * lda;
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      i = m;
kusano 2b45e8
      if (m > 0) {
kusano 2b45e8
	do {
kusano 2b45e8
	  if (X < posY) {
kusano 2b45e8
	    data01 = *(ao1 + 0);
kusano 2b45e8
	    b[ 0] = data01;
kusano 2b45e8
	    ao1 += 1;
kusano 2b45e8
	    b += 1;
kusano 2b45e8
	  } else
kusano 2b45e8
	    if (X > posY)  {
kusano 2b45e8
	      ao1 += lda;
kusano 2b45e8
	      b += 1;
kusano 2b45e8
	    } else {
kusano 2b45e8
#ifdef UNIT
kusano 2b45e8
	      b[ 0] = ONE;
kusano 2b45e8
#else
kusano 2b45e8
	      data01 = *(ao1 + 0);
kusano 2b45e8
	      b[ 0] = data01;
kusano 2b45e8
#endif
kusano 2b45e8
	      ao1 += lda;
kusano 2b45e8
	      b += 1;
kusano 2b45e8
	    }
kusano 2b45e8
	  
kusano 2b45e8
	  X += 1;
kusano 2b45e8
	  i --;
kusano 2b45e8
	} while (i > 0);
kusano 2b45e8
      }
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  return 0;
kusano 2b45e8
}