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 "common.h"
kusano 2b45e8
kusano 2b45e8
#ifndef UNIT
kusano 2b45e8
#define INV(a) (ONE / (a))
kusano 2b45e8
#else
kusano 2b45e8
#define INV(a) (ONE)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT *b){
kusano 2b45e8
kusano 2b45e8
  BLASLONG i, ii, j, jj, k;
kusano 2b45e8
kusano 2b45e8
  FLOAT *a1, *a2,  *a3,  *a4,  *a5,  *a6,  *a7,  *a8;
kusano 2b45e8
  FLOAT *a9, *a10, *a11, *a12, *a13, *a14, *a15, *a16;
kusano 2b45e8
kusano 2b45e8
  
kusano 2b45e8
  jj = offset;
kusano 2b45e8
kusano 2b45e8
  j = (n >> 4);
kusano 2b45e8
  while (j > 0){
kusano 2b45e8
kusano 2b45e8
    a1  = a +  0 * lda;
kusano 2b45e8
    a2  = a +  1 * lda;
kusano 2b45e8
    a3  = a +  2 * lda;
kusano 2b45e8
    a4  = a +  3 * lda;
kusano 2b45e8
    a5  = a +  4 * lda;
kusano 2b45e8
    a6  = a +  5 * lda;
kusano 2b45e8
    a7  = a +  6 * lda;
kusano 2b45e8
    a8  = a +  7 * lda;
kusano 2b45e8
    a9  = a +  8 * lda;
kusano 2b45e8
    a10 = a +  9 * lda;
kusano 2b45e8
    a11 = a + 10 * lda;
kusano 2b45e8
    a12 = a + 11 * lda;
kusano 2b45e8
    a13 = a + 12 * lda;
kusano 2b45e8
    a14 = a + 13 * lda;
kusano 2b45e8
    a15 = a + 14 * lda;
kusano 2b45e8
    a16 = a + 15 * lda;
kusano 2b45e8
kusano 2b45e8
    a += 16 * lda;
kusano 2b45e8
    ii = 0;
kusano 2b45e8
kusano 2b45e8
    for (i = 0; i < m; i++) {
kusano 2b45e8
      
kusano 2b45e8
      if ((ii >= jj ) && (ii - jj < 16)) {
kusano 2b45e8
	*(b +  ii - jj) = INV(*(a1 + (ii - jj) * lda));
kusano 2b45e8
	for (k = ii - jj + 1; k < 16; k ++) {
kusano 2b45e8
	  *(b +  k) = *(a1 +  k * lda);
kusano 2b45e8
	}
kusano 2b45e8
      }
kusano 2b45e8
      
kusano 2b45e8
      if (ii - jj < 0) {
kusano 2b45e8
	*(b +  0) = *(a1  +  0);
kusano 2b45e8
	*(b +  1) = *(a2  +  0);
kusano 2b45e8
	*(b +  2) = *(a3  +  0);
kusano 2b45e8
	*(b +  3) = *(a4  +  0);
kusano 2b45e8
	*(b +  4) = *(a5  +  0);
kusano 2b45e8
	*(b +  5) = *(a6  +  0);
kusano 2b45e8
	*(b +  6) = *(a7  +  0);
kusano 2b45e8
	*(b +  7) = *(a8  +  0);
kusano 2b45e8
	*(b +  8) = *(a9  +  0);
kusano 2b45e8
	*(b +  9) = *(a10 +  0);
kusano 2b45e8
	*(b + 10) = *(a11 +  0);
kusano 2b45e8
	*(b + 11) = *(a12 +  0);
kusano 2b45e8
	*(b + 12) = *(a13 +  0);
kusano 2b45e8
	*(b + 13) = *(a14 +  0);
kusano 2b45e8
	*(b + 14) = *(a15 +  0);
kusano 2b45e8
	*(b + 15) = *(a16 +  0);
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      a1  ++;
kusano 2b45e8
      a2  ++;
kusano 2b45e8
      a3  ++;
kusano 2b45e8
      a4  ++;
kusano 2b45e8
      a5  ++;
kusano 2b45e8
      a6  ++;
kusano 2b45e8
      a7  ++;
kusano 2b45e8
      a8  ++;
kusano 2b45e8
      a9  ++;
kusano 2b45e8
      a10 ++;
kusano 2b45e8
      a11 ++;
kusano 2b45e8
      a12 ++;
kusano 2b45e8
      a13 ++;
kusano 2b45e8
      a14 ++;
kusano 2b45e8
      a15 ++;
kusano 2b45e8
      a16 ++;
kusano 2b45e8
      b  += 16;
kusano 2b45e8
      ii ++;
kusano 2b45e8
    }
kusano 2b45e8
kusano 2b45e8
    jj += 16;
kusano 2b45e8
    j --;
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (n & 8) {
kusano 2b45e8
kusano 2b45e8
    a1  = a +  0 * lda;
kusano 2b45e8
    a2  = a +  1 * lda;
kusano 2b45e8
    a3  = a +  2 * lda;
kusano 2b45e8
    a4  = a +  3 * lda;
kusano 2b45e8
    a5  = a +  4 * lda;
kusano 2b45e8
    a6  = a +  5 * lda;
kusano 2b45e8
    a7  = a +  6 * lda;
kusano 2b45e8
    a8  = a +  7 * lda;
kusano 2b45e8
    a += 8 * lda;
kusano 2b45e8
    ii = 0;
kusano 2b45e8
kusano 2b45e8
    for (i = 0; i < m; i++) {
kusano 2b45e8
      
kusano 2b45e8
      if ((ii >= jj ) && (ii - jj < 8)) {
kusano 2b45e8
	*(b +  ii - jj) = INV(*(a1 + (ii - jj) * lda));
kusano 2b45e8
	for (k = ii - jj + 1; k < 8; k ++) {
kusano 2b45e8
	  *(b +  k) = *(a1 +  k * lda);
kusano 2b45e8
	}
kusano 2b45e8
      }
kusano 2b45e8
      
kusano 2b45e8
      if (ii - jj < 0) {
kusano 2b45e8
	*(b +  0) = *(a1  +  0);
kusano 2b45e8
	*(b +  1) = *(a2  +  0);
kusano 2b45e8
	*(b +  2) = *(a3  +  0);
kusano 2b45e8
	*(b +  3) = *(a4  +  0);
kusano 2b45e8
	*(b +  4) = *(a5  +  0);
kusano 2b45e8
	*(b +  5) = *(a6  +  0);
kusano 2b45e8
	*(b +  6) = *(a7  +  0);
kusano 2b45e8
	*(b +  7) = *(a8  +  0);
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      a1  ++;
kusano 2b45e8
      a2  ++;
kusano 2b45e8
      a3  ++;
kusano 2b45e8
      a4  ++;
kusano 2b45e8
      a5  ++;
kusano 2b45e8
      a6  ++;
kusano 2b45e8
      a7  ++;
kusano 2b45e8
      a8  ++;
kusano 2b45e8
      b  += 8;
kusano 2b45e8
      ii ++;
kusano 2b45e8
    }
kusano 2b45e8
kusano 2b45e8
    jj += 8;
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (n & 4) {
kusano 2b45e8
kusano 2b45e8
    a1  = a +  0 * lda;
kusano 2b45e8
    a2  = a +  1 * lda;
kusano 2b45e8
    a3  = a +  2 * lda;
kusano 2b45e8
    a4  = a +  3 * lda;
kusano 2b45e8
    a += 4 * lda;
kusano 2b45e8
    ii = 0;
kusano 2b45e8
kusano 2b45e8
    for (i = 0; i < m; i++) {
kusano 2b45e8
      
kusano 2b45e8
      if ((ii >= jj ) && (ii - jj < 4)) {
kusano 2b45e8
	*(b +  ii - jj) = INV(*(a1 + (ii - jj) * lda));
kusano 2b45e8
	for (k = ii - jj + 1; k < 4; k ++) {
kusano 2b45e8
	  *(b +  k) = *(a1 +  k * lda);
kusano 2b45e8
	}
kusano 2b45e8
      }
kusano 2b45e8
      
kusano 2b45e8
      if (ii - jj < 0) {
kusano 2b45e8
	*(b +  0) = *(a1  +  0);
kusano 2b45e8
	*(b +  1) = *(a2  +  0);
kusano 2b45e8
	*(b +  2) = *(a3  +  0);
kusano 2b45e8
	*(b +  3) = *(a4  +  0);
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      a1  ++;
kusano 2b45e8
      a2  ++;
kusano 2b45e8
      a3  ++;
kusano 2b45e8
      a4  ++;
kusano 2b45e8
      b  += 4;
kusano 2b45e8
      ii ++;
kusano 2b45e8
    }
kusano 2b45e8
kusano 2b45e8
    jj += 4;
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (n & 2) {
kusano 2b45e8
kusano 2b45e8
    a1  = a +  0 * lda;
kusano 2b45e8
    a2  = a +  1 * lda;
kusano 2b45e8
    a += 2 * lda;
kusano 2b45e8
    ii = 0;
kusano 2b45e8
kusano 2b45e8
    for (i = 0; i < m; i++) {
kusano 2b45e8
      
kusano 2b45e8
      if ((ii >= jj ) && (ii - jj < 2)) {
kusano 2b45e8
	*(b +  ii - jj) = INV(*(a1 + (ii - jj) * lda));
kusano 2b45e8
	for (k = ii - jj + 1; k < 2; k ++) {
kusano 2b45e8
	  *(b +  k) = *(a1 +  k * lda);
kusano 2b45e8
	}
kusano 2b45e8
      }
kusano 2b45e8
      
kusano 2b45e8
      if (ii - jj < 0) {
kusano 2b45e8
	*(b +  0) = *(a1  +  0);
kusano 2b45e8
	*(b +  1) = *(a2  +  0);
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      a1  ++;
kusano 2b45e8
      a2  ++;
kusano 2b45e8
      b  += 2;
kusano 2b45e8
      ii ++;
kusano 2b45e8
    }
kusano 2b45e8
kusano 2b45e8
    jj += 2;
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (n & 1) {
kusano 2b45e8
kusano 2b45e8
    a1  = a +  0 * lda;
kusano 2b45e8
    ii = 0;
kusano 2b45e8
kusano 2b45e8
    for (i = 0; i < m; i++) {
kusano 2b45e8
      
kusano 2b45e8
      if ((ii >= jj ) && (ii - jj < 1)) {
kusano 2b45e8
	*(b +  ii - jj) = INV(*(a1 + (ii - jj) * lda));
kusano 2b45e8
	for (k = ii - jj + 1; k < 1; k ++) {
kusano 2b45e8
	  *(b +  k) = *(a1 +  k * lda);
kusano 2b45e8
	}
kusano 2b45e8
      }
kusano 2b45e8
      
kusano 2b45e8
      if (ii - jj < 0) {
kusano 2b45e8
	*(b +  0) = *(a1  +  0);
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      a1  ++;
kusano 2b45e8
      b  += 1;
kusano 2b45e8
      ii ++;
kusano 2b45e8
    }
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  return 0;
kusano 2b45e8
}