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, offset;
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,  *ao5,  *ao6,  *ao7,  *ao8;
kusano 2b45e8
  FLOAT *ao9, *ao10, *ao11, *ao12, *ao13, *ao14, *ao15, *ao16;
kusano 2b45e8
kusano 2b45e8
  js = (n >> 4);
kusano 2b45e8
  while (js > 0){
kusano 2b45e8
kusano 2b45e8
    offset = posX - posY;
kusano 2b45e8
    
kusano 2b45e8
    if (offset >   0) ao1  = a + posX +  0 + posY * lda; else ao1  = a + posY + (posX +  0) * lda;
kusano 2b45e8
    if (offset >  -1) ao2  = a + posX +  1 + posY * lda; else ao2  = a + posY + (posX +  1) * lda;
kusano 2b45e8
    if (offset >  -2) ao3  = a + posX +  2 + posY * lda; else ao3  = a + posY + (posX +  2) * lda;
kusano 2b45e8
    if (offset >  -3) ao4  = a + posX +  3 + posY * lda; else ao4  = a + posY + (posX +  3) * lda;
kusano 2b45e8
    if (offset >  -4) ao5  = a + posX +  4 + posY * lda; else ao5  = a + posY + (posX +  4) * lda;
kusano 2b45e8
    if (offset >  -5) ao6  = a + posX +  5 + posY * lda; else ao6  = a + posY + (posX +  5) * lda;
kusano 2b45e8
    if (offset >  -6) ao7  = a + posX +  6 + posY * lda; else ao7  = a + posY + (posX +  6) * lda;
kusano 2b45e8
    if (offset >  -7) ao8  = a + posX +  7 + posY * lda; else ao8  = a + posY + (posX +  7) * lda;
kusano 2b45e8
    if (offset >  -8) ao9  = a + posX +  8 + posY * lda; else ao9  = a + posY + (posX +  8) * lda;
kusano 2b45e8
    if (offset >  -9) ao10 = a + posX +  9 + posY * lda; else ao10 = a + posY + (posX +  9) * lda;
kusano 2b45e8
    if (offset > -10) ao11 = a + posX + 10 + posY * lda; else ao11 = a + posY + (posX + 10) * lda;
kusano 2b45e8
    if (offset > -11) ao12 = a + posX + 11 + posY * lda; else ao12 = a + posY + (posX + 11) * lda;
kusano 2b45e8
    if (offset > -12) ao13 = a + posX + 12 + posY * lda; else ao13 = a + posY + (posX + 12) * lda;
kusano 2b45e8
    if (offset > -13) ao14 = a + posX + 13 + posY * lda; else ao14 = a + posY + (posX + 13) * lda;
kusano 2b45e8
    if (offset > -14) ao15 = a + posX + 14 + posY * lda; else ao15 = a + posY + (posX + 14) * lda;
kusano 2b45e8
    if (offset > -15) ao16 = a + posX + 15 + posY * lda; else ao16 = a + posY + (posX + 15) * lda;
kusano 2b45e8
kusano 2b45e8
    i     = m;
kusano 2b45e8
kusano 2b45e8
    while (i > 0) {
kusano 2b45e8
      data01 = *(ao1  + 0);
kusano 2b45e8
      data02 = *(ao2  + 0);
kusano 2b45e8
      data03 = *(ao3  + 0);
kusano 2b45e8
      data04 = *(ao4  + 0);
kusano 2b45e8
      data05 = *(ao5  + 0);
kusano 2b45e8
      data06 = *(ao6  + 0);
kusano 2b45e8
      data07 = *(ao7  + 0);
kusano 2b45e8
      data08 = *(ao8  + 0);
kusano 2b45e8
      data09 = *(ao9  + 0);
kusano 2b45e8
      data10 = *(ao10 + 0);
kusano 2b45e8
      data11 = *(ao11 + 0);
kusano 2b45e8
      data12 = *(ao12 + 0);
kusano 2b45e8
      data13 = *(ao13 + 0);
kusano 2b45e8
      data14 = *(ao14 + 0);
kusano 2b45e8
      data15 = *(ao15 + 0);
kusano 2b45e8
      data16 = *(ao16 + 0);
kusano 2b45e8
kusano 2b45e8
      if (offset >   0) ao1  += lda; else ao1 ++;
kusano 2b45e8
      if (offset >  -1) ao2  += lda; else ao2 ++;
kusano 2b45e8
      if (offset >  -2) ao3  += lda; else ao3 ++;
kusano 2b45e8
      if (offset >  -3) ao4  += lda; else ao4 ++;
kusano 2b45e8
      if (offset >  -4) ao5  += lda; else ao5 ++;
kusano 2b45e8
      if (offset >  -5) ao6  += lda; else ao6 ++;
kusano 2b45e8
      if (offset >  -6) ao7  += lda; else ao7 ++;
kusano 2b45e8
      if (offset >  -7) ao8  += lda; else ao8 ++;
kusano 2b45e8
      if (offset >  -8) ao9  += lda; else ao9 ++;
kusano 2b45e8
      if (offset >  -9) ao10 += lda; else ao10 ++;
kusano 2b45e8
      if (offset > -10) ao11 += lda; else ao11 ++;
kusano 2b45e8
      if (offset > -11) ao12 += lda; else ao12 ++;
kusano 2b45e8
      if (offset > -12) ao13 += lda; else ao13 ++;
kusano 2b45e8
      if (offset > -13) ao14 += lda; else ao14 ++;
kusano 2b45e8
      if (offset > -14) ao15 += lda; else ao15 ++;
kusano 2b45e8
      if (offset > -15) ao16 += lda; else ao16 ++;
kusano 2b45e8
      
kusano 2b45e8
      b[ 0] = data01;
kusano 2b45e8
      b[ 1] = data02;
kusano 2b45e8
      b[ 2] = data03;
kusano 2b45e8
      b[ 3] = data04;
kusano 2b45e8
      b[ 4] = data05;
kusano 2b45e8
      b[ 5] = data06;
kusano 2b45e8
      b[ 6] = data07;
kusano 2b45e8
      b[ 7] = data08;
kusano 2b45e8
      b[ 8] = data09;
kusano 2b45e8
      b[ 9] = data10;
kusano 2b45e8
      b[10] = data11;
kusano 2b45e8
      b[11] = data12;
kusano 2b45e8
      b[12] = data13;
kusano 2b45e8
      b[13] = data14;
kusano 2b45e8
      b[14] = data15;
kusano 2b45e8
      b[15] = data16;
kusano 2b45e8
kusano 2b45e8
      b += 16;
kusano 2b45e8
kusano 2b45e8
      offset --;
kusano 2b45e8
      i --;
kusano 2b45e8
    }
kusano 2b45e8
kusano 2b45e8
    posX += 16;
kusano 2b45e8
    js --;
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (n & 8) {
kusano 2b45e8
kusano 2b45e8
    offset = posX - posY;
kusano 2b45e8
    
kusano 2b45e8
    if (offset >  0) ao1 = a + posX + 0 + posY * lda; else ao1 = a + posY + (posX + 0) * lda;
kusano 2b45e8
    if (offset > -1) ao2 = a + posX + 1 + posY * lda; else ao2 = a + posY + (posX + 1) * lda;
kusano 2b45e8
    if (offset > -2) ao3 = a + posX + 2 + posY * lda; else ao3 = a + posY + (posX + 2) * lda;
kusano 2b45e8
    if (offset > -3) ao4 = a + posX + 3 + posY * lda; else ao4 = a + posY + (posX + 3) * lda;
kusano 2b45e8
    if (offset > -4) ao5 = a + posX + 4 + posY * lda; else ao5 = a + posY + (posX + 4) * lda;
kusano 2b45e8
    if (offset > -5) ao6 = a + posX + 5 + posY * lda; else ao6 = a + posY + (posX + 5) * lda;
kusano 2b45e8
    if (offset > -6) ao7 = a + posX + 6 + posY * lda; else ao7 = a + posY + (posX + 6) * lda;
kusano 2b45e8
    if (offset > -7) ao8 = a + posX + 7 + posY * lda; else ao8 = a + posY + (posX + 7) * lda;
kusano 2b45e8
kusano 2b45e8
    i     = m;
kusano 2b45e8
kusano 2b45e8
    while (i > 0) {
kusano 2b45e8
      data01 = *(ao1 + 0);
kusano 2b45e8
      data02 = *(ao2 + 0);
kusano 2b45e8
      data03 = *(ao3 + 0);
kusano 2b45e8
      data04 = *(ao4 + 0);
kusano 2b45e8
      data05 = *(ao5 + 0);
kusano 2b45e8
      data06 = *(ao6 + 0);
kusano 2b45e8
      data07 = *(ao7 + 0);
kusano 2b45e8
      data08 = *(ao8 + 0);
kusano 2b45e8
      
kusano 2b45e8
      if (offset >   0) ao1 += lda; else ao1 ++;
kusano 2b45e8
      if (offset >  -1) ao2 += lda; else ao2 ++;
kusano 2b45e8
      if (offset >  -2) ao3 += lda; else ao3 ++;
kusano 2b45e8
      if (offset >  -3) ao4 += lda; else ao4 ++;
kusano 2b45e8
      if (offset >  -4) ao5 += lda; else ao5 ++;
kusano 2b45e8
      if (offset >  -5) ao6 += lda; else ao6 ++;
kusano 2b45e8
      if (offset >  -6) ao7 += lda; else ao7 ++;
kusano 2b45e8
      if (offset >  -7) ao8 += lda; else ao8 ++;
kusano 2b45e8
kusano 2b45e8
      b[ 0] = data01;
kusano 2b45e8
      b[ 1] = data02;
kusano 2b45e8
      b[ 2] = data03;
kusano 2b45e8
      b[ 3] = data04;
kusano 2b45e8
      b[ 4] = data05;
kusano 2b45e8
      b[ 5] = data06;
kusano 2b45e8
      b[ 6] = data07;
kusano 2b45e8
      b[ 7] = data08;
kusano 2b45e8
kusano 2b45e8
      b += 8;
kusano 2b45e8
kusano 2b45e8
      offset --;
kusano 2b45e8
      i --;
kusano 2b45e8
    }
kusano 2b45e8
kusano 2b45e8
    posX += 8;
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (n & 4) {
kusano 2b45e8
    offset = posX - posY;
kusano 2b45e8
    
kusano 2b45e8
    if (offset >  0) ao1 = a + posX + 0 + posY * lda; else ao1 = a + posY + (posX + 0) * lda;
kusano 2b45e8
    if (offset > -1) ao2 = a + posX + 1 + posY * lda; else ao2 = a + posY + (posX + 1) * lda;
kusano 2b45e8
    if (offset > -2) ao3 = a + posX + 2 + posY * lda; else ao3 = a + posY + (posX + 2) * lda;
kusano 2b45e8
    if (offset > -3) ao4 = a + posX + 3 + posY * lda; else ao4 = a + posY + (posX + 3) * lda;
kusano 2b45e8
kusano 2b45e8
    i     = m;
kusano 2b45e8
kusano 2b45e8
    while (i > 0) {
kusano 2b45e8
      data01 = *(ao1 + 0);
kusano 2b45e8
      data02 = *(ao2 + 0);
kusano 2b45e8
      data03 = *(ao3 + 0);
kusano 2b45e8
      data04 = *(ao4 + 0);
kusano 2b45e8
      
kusano 2b45e8
      if (offset >   0) ao1 += lda; else ao1 ++;
kusano 2b45e8
      if (offset >  -1) ao2 += lda; else ao2 ++;
kusano 2b45e8
      if (offset >  -2) ao3 += lda; else ao3 ++;
kusano 2b45e8
      if (offset >  -3) ao4 += lda; else ao4 ++;
kusano 2b45e8
kusano 2b45e8
      b[ 0] = data01;
kusano 2b45e8
      b[ 1] = data02;
kusano 2b45e8
      b[ 2] = data03;
kusano 2b45e8
      b[ 3] = data04;
kusano 2b45e8
kusano 2b45e8
      b += 4;
kusano 2b45e8
kusano 2b45e8
      offset --;
kusano 2b45e8
      i --;
kusano 2b45e8
    }
kusano 2b45e8
kusano 2b45e8
    posX += 4;
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (n & 2) {
kusano 2b45e8
    offset = posX - posY;
kusano 2b45e8
    
kusano 2b45e8
    if (offset >  0) ao1 = a + posX + 0 + posY * lda; else ao1 = a + posY + (posX + 0) * lda;
kusano 2b45e8
    if (offset > -1) ao2 = a + posX + 1 + posY * lda; else ao2 = a + posY + (posX + 1) * lda;
kusano 2b45e8
kusano 2b45e8
    i     = m;
kusano 2b45e8
kusano 2b45e8
    while (i > 0) {
kusano 2b45e8
      data01 = *(ao1 + 0);
kusano 2b45e8
      data02 = *(ao2 + 0);
kusano 2b45e8
      
kusano 2b45e8
      if (offset >   0) ao1 += lda; else ao1 ++;
kusano 2b45e8
      if (offset >  -1) ao2 += lda; else ao2 ++;
kusano 2b45e8
kusano 2b45e8
      b[ 0] = data01;
kusano 2b45e8
      b[ 1] = data02;
kusano 2b45e8
kusano 2b45e8
      b += 2;
kusano 2b45e8
kusano 2b45e8
      offset --;
kusano 2b45e8
      i --;
kusano 2b45e8
    }
kusano 2b45e8
kusano 2b45e8
    posX += 2;
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (n & 1) {
kusano 2b45e8
    offset = posX - posY;
kusano 2b45e8
    
kusano 2b45e8
    if (offset >  0) ao1 = a + posX + 0 + posY * lda; else ao1 = a + posY + (posX + 0) * lda;
kusano 2b45e8
kusano 2b45e8
    i     = m;
kusano 2b45e8
kusano 2b45e8
    while (i > 0) {
kusano 2b45e8
      data01 = *(ao1 + 0);
kusano 2b45e8
      
kusano 2b45e8
      if (offset >   0) ao1 += lda; else ao1 ++;
kusano 2b45e8
kusano 2b45e8
      b[ 0] = data01;
kusano 2b45e8
kusano 2b45e8
      b ++;
kusano 2b45e8
kusano 2b45e8
      offset --;
kusano 2b45e8
      i --;
kusano 2b45e8
    }
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  return 0;
kusano 2b45e8
}