Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/generic/laswp_ncopy_2.c

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
#define PREFETCHSIZE 12
kusano 2b45e8
kusano 2b45e8
#define a2	(a1 + 1)
kusano 2b45e8
#define a4	(a3 + 1)
kusano 2b45e8
kusano 2b45e8
int CNAME(BLASLONG n, BLASLONG k1, BLASLONG k2, FLOAT *a, BLASLONG lda, blasint *ipiv, FLOAT *buffer){
kusano 2b45e8
kusano 2b45e8
  BLASLONG i, j, ip1, ip2;
kusano 2b45e8
  blasint *piv;
kusano 2b45e8
  FLOAT *a1, *a3;
kusano 2b45e8
  FLOAT *b1, *b2, *b3, *b4;
kusano 2b45e8
  FLOAT A1, A2, B1, B2, A3, A4, B3, B4;
kusano 2b45e8
kusano 2b45e8
  a--;
kusano 2b45e8
  k1 --;
kusano 2b45e8
kusano 2b45e8
 ipiv += k1;
kusano 2b45e8
kusano 2b45e8
  if (n  <= 0) return 0;
kusano 2b45e8
      
kusano 2b45e8
  j = (n >> 1);
kusano 2b45e8
  if (j > 0) {
kusano 2b45e8
    do {
kusano 2b45e8
      piv = ipiv;
kusano 2b45e8
      
kusano 2b45e8
      a1 = a + k1 + 1;
kusano 2b45e8
      a3 = a1 + 1 * lda;
kusano 2b45e8
      
kusano 2b45e8
      ip1 = *(piv + 0);
kusano 2b45e8
      ip2 = *(piv + 1);
kusano 2b45e8
      piv += 2;
kusano 2b45e8
      
kusano 2b45e8
      b1 = a + ip1;
kusano 2b45e8
      b2 = a + ip2;
kusano 2b45e8
      
kusano 2b45e8
      b3 = b1 + 1 * lda;
kusano 2b45e8
      b4 = b2 + 1 * lda;
kusano 2b45e8
      
kusano 2b45e8
      i = ((k2 - k1) >> 1);
kusano 2b45e8
      
kusano 2b45e8
      if (i > 0) {
kusano 2b45e8
	do {
kusano 2b45e8
kusano 2b45e8
#ifdef __GNUC__
kusano 2b45e8
	  __builtin_prefetch(a1 + PREFETCHSIZE, 0, 0);
kusano 2b45e8
	  __builtin_prefetch(a3 + PREFETCHSIZE, 0, 0);
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	  A1 = *a1;
kusano 2b45e8
	  A2 = *a2;
kusano 2b45e8
	  A3 = *a3;
kusano 2b45e8
	  A4 = *a4;
kusano 2b45e8
	  
kusano 2b45e8
	  B1 = *b1;
kusano 2b45e8
	  B2 = *b2;
kusano 2b45e8
	  B3 = *b3;
kusano 2b45e8
	  B4 = *b4;
kusano 2b45e8
	  
kusano 2b45e8
	  ip1 = *(piv + 0);
kusano 2b45e8
	  ip2 = *(piv + 1);
kusano 2b45e8
	  piv += 2;
kusano 2b45e8
	  
kusano 2b45e8
	  if (b1 == a1) {
kusano 2b45e8
	    if (b2 == a2) {
kusano 2b45e8
	      *(buffer + 0) = A1;
kusano 2b45e8
	      *(buffer + 1) = A3;
kusano 2b45e8
	      *(buffer + 2) = A2;
kusano 2b45e8
	      *(buffer + 3) = A4;
kusano 2b45e8
	    } else {
kusano 2b45e8
	      *(buffer + 0) = A1;
kusano 2b45e8
	      *(buffer + 1) = A3;
kusano 2b45e8
	      *(buffer + 2) = B2;
kusano 2b45e8
	      *(buffer + 3) = B4;
kusano 2b45e8
	      
kusano 2b45e8
	      *b2 = A2;
kusano 2b45e8
	      *b4 = A4;
kusano 2b45e8
	    }
kusano 2b45e8
	} else 
kusano 2b45e8
	  if (b1 == a2) {
kusano 2b45e8
	    if (b2 == a2) {
kusano 2b45e8
	      *(buffer + 0) = A2;
kusano 2b45e8
	      *(buffer + 1) = A4;
kusano 2b45e8
	      *(buffer + 2) = A1;
kusano 2b45e8
	      *(buffer + 3) = A3;
kusano 2b45e8
	    } else {
kusano 2b45e8
	      *(buffer + 0) = A2;
kusano 2b45e8
	      *(buffer + 1) = A4;
kusano 2b45e8
	      *(buffer + 2) = B2;
kusano 2b45e8
	      *(buffer + 3) = B4;
kusano 2b45e8
	      *b2 = A1;
kusano 2b45e8
	      *b4 = A3;
kusano 2b45e8
	    }
kusano 2b45e8
	  } else {
kusano 2b45e8
	    if (b2 == a2) {
kusano 2b45e8
	      *(buffer + 0) = B1;
kusano 2b45e8
	      *(buffer + 1) = B3;
kusano 2b45e8
	      *(buffer + 2) = A2;
kusano 2b45e8
	      *(buffer + 3) = A4;
kusano 2b45e8
	      *b1 = A1;
kusano 2b45e8
	      *b3 = A3;
kusano 2b45e8
	    } else 
kusano 2b45e8
	      if (b2 == b1) {
kusano 2b45e8
		*(buffer + 0) = B1;
kusano 2b45e8
		*(buffer + 1) = B3;
kusano 2b45e8
		*(buffer + 2) = A1;
kusano 2b45e8
		*(buffer + 3) = A3;
kusano 2b45e8
		*b1 = A2;
kusano 2b45e8
		*b3 = A4;
kusano 2b45e8
	      } else {
kusano 2b45e8
		*(buffer + 0) = B1;
kusano 2b45e8
		*(buffer + 1) = B3;
kusano 2b45e8
		*(buffer + 2) = B2;
kusano 2b45e8
		*(buffer + 3) = B4;
kusano 2b45e8
		*b1 = A1;
kusano 2b45e8
		*b2 = A2;
kusano 2b45e8
		*b3 = A3;
kusano 2b45e8
		*b4 = A4;
kusano 2b45e8
	      }
kusano 2b45e8
	  }
kusano 2b45e8
	  
kusano 2b45e8
	  buffer += 4;
kusano 2b45e8
	  
kusano 2b45e8
	  b1 = a + ip1;
kusano 2b45e8
	  b2 = a + ip2;
kusano 2b45e8
	  
kusano 2b45e8
	  b3 = b1 + 1 * lda;
kusano 2b45e8
	  b4 = b2 + 1 * lda;
kusano 2b45e8
	  
kusano 2b45e8
	  a1 += 2;
kusano 2b45e8
	  a3 += 2;
kusano 2b45e8
	  
kusano 2b45e8
	  i --;
kusano 2b45e8
	} while (i > 0);
kusano 2b45e8
      }
kusano 2b45e8
      
kusano 2b45e8
      i = ((k2 - k1) & 1);
kusano 2b45e8
      
kusano 2b45e8
      if (i > 0) {
kusano 2b45e8
	A1 = *a1;
kusano 2b45e8
	B1 = *b1;
kusano 2b45e8
	A3 = *a3;
kusano 2b45e8
	B3 = *b3;
kusano 2b45e8
	
kusano 2b45e8
	if (a1 == b1) {
kusano 2b45e8
	  *(buffer + 0) = A1;
kusano 2b45e8
	  *(buffer + 1) = A3;
kusano 2b45e8
	} else {
kusano 2b45e8
	  *(buffer + 0) = B1;
kusano 2b45e8
	  *(buffer + 1) = B3;
kusano 2b45e8
	  *b1 = A1;
kusano 2b45e8
	  *b3 = A3;
kusano 2b45e8
	}
kusano 2b45e8
	buffer += 2;
kusano 2b45e8
      }
kusano 2b45e8
      
kusano 2b45e8
      a += 2 * lda;
kusano 2b45e8
      j --;
kusano 2b45e8
    } while (j > 0);
kusano 2b45e8
  }
kusano 2b45e8
  
kusano 2b45e8
  if (n & 1) {
kusano 2b45e8
    piv = ipiv;
kusano 2b45e8
kusano 2b45e8
    a1 = a + k1 + 1;
kusano 2b45e8
    
kusano 2b45e8
    ip1 = *(piv + 0);
kusano 2b45e8
    ip2 = *(piv + 1);
kusano 2b45e8
    piv += 2;
kusano 2b45e8
    
kusano 2b45e8
    b1 = a + ip1;
kusano 2b45e8
    b2 = a + ip2;
kusano 2b45e8
    
kusano 2b45e8
    i = ((k2 - k1) >> 1);
kusano 2b45e8
    
kusano 2b45e8
    if (i > 0) {
kusano 2b45e8
      do {
kusano 2b45e8
	A1 = *a1;
kusano 2b45e8
	A2 = *a2;
kusano 2b45e8
	B1 = *b1;
kusano 2b45e8
	B2 = *b2;
kusano 2b45e8
	
kusano 2b45e8
	ip1 = *(piv + 0);
kusano 2b45e8
	ip2 = *(piv + 1);
kusano 2b45e8
	piv += 2;
kusano 2b45e8
	
kusano 2b45e8
	if (b1 == a1) {
kusano 2b45e8
	    if (b2 == a2) {
kusano 2b45e8
	      *(buffer + 0) = A1;
kusano 2b45e8
	      *(buffer + 1) = A2;
kusano 2b45e8
	    } else {
kusano 2b45e8
	      *(buffer + 0) = A1;
kusano 2b45e8
	      *(buffer + 1) = B2;
kusano 2b45e8
kusano 2b45e8
	      *b2 = A2;
kusano 2b45e8
	    }
kusano 2b45e8
	} else 
kusano 2b45e8
	  if (b1 == a2) {
kusano 2b45e8
	      if (b2 == a2) {
kusano 2b45e8
		*(buffer + 0) = A2;
kusano 2b45e8
		*(buffer + 1) = A1;
kusano 2b45e8
	      } else {
kusano 2b45e8
		*(buffer + 0) = A2;
kusano 2b45e8
		*(buffer + 1) = B2;
kusano 2b45e8
		*b2 = A1;
kusano 2b45e8
	      }
kusano 2b45e8
	  } else {
kusano 2b45e8
	      if (b2 == a2) {
kusano 2b45e8
		*(buffer + 0) = B1;
kusano 2b45e8
		*(buffer + 1) = A2;
kusano 2b45e8
		*b1 = A1;
kusano 2b45e8
	      } else 
kusano 2b45e8
		if (b2 == b1) {
kusano 2b45e8
		  *(buffer + 0) = B1;
kusano 2b45e8
		  *(buffer + 1) = A1;
kusano 2b45e8
		  *b1 = A2;
kusano 2b45e8
		} else {
kusano 2b45e8
		  *(buffer + 0) = B1;
kusano 2b45e8
		  *(buffer + 1) = B2;
kusano 2b45e8
		  *b1 = A1;
kusano 2b45e8
		  *b2 = A2;
kusano 2b45e8
		}
kusano 2b45e8
	  }
kusano 2b45e8
	
kusano 2b45e8
	buffer += 2;
kusano 2b45e8
kusano 2b45e8
	b1 = a + ip1;
kusano 2b45e8
	b2 = a + ip2;
kusano 2b45e8
	
kusano 2b45e8
	a1 += 2;
kusano 2b45e8
kusano 2b45e8
	i --;
kusano 2b45e8
      } while (i > 0);
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
    i = ((k2 - k1) & 1);
kusano 2b45e8
    
kusano 2b45e8
    if (i > 0) {
kusano 2b45e8
      A1 = *a1;
kusano 2b45e8
      B1 = *b1;
kusano 2b45e8
kusano 2b45e8
      if (a1 == b1) {
kusano 2b45e8
	*(buffer + 0) = A1;
kusano 2b45e8
      } else {
kusano 2b45e8
	*(buffer + 0) = B1;
kusano 2b45e8
	*b1 = A1;
kusano 2b45e8
      }
kusano 2b45e8
    }
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  return 0;
kusano 2b45e8
} 
kusano 2b45e8