Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/generic/gemm_tcopy_4.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
int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
kusano 2b45e8
kusano 2b45e8
  BLASLONG i, j;
kusano 2b45e8
kusano 2b45e8
  FLOAT *a_offset, *a_offset1, *a_offset2, *a_offset3, *a_offset4;
kusano 2b45e8
  FLOAT *b_offset, *b_offset1, *b_offset2, *b_offset3;
kusano 2b45e8
  FLOAT  ctemp1,  ctemp2,  ctemp3,  ctemp4;
kusano 2b45e8
  FLOAT  ctemp5,  ctemp6,  ctemp7,  ctemp8;
kusano 2b45e8
  FLOAT  ctemp9, ctemp10, ctemp11, ctemp12;
kusano 2b45e8
  FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
kusano 2b45e8
kusano 2b45e8
  a_offset   = a;
kusano 2b45e8
  b_offset   = b;
kusano 2b45e8
kusano 2b45e8
  b_offset2  = b + m  * (n & ~3);
kusano 2b45e8
  b_offset3  = b + m  * (n & ~1);
kusano 2b45e8
kusano 2b45e8
  j = (m >> 2);
kusano 2b45e8
  if (j > 0){
kusano 2b45e8
    do{
kusano 2b45e8
      a_offset1  = a_offset;
kusano 2b45e8
      a_offset2  = a_offset1 + lda;
kusano 2b45e8
      a_offset3  = a_offset2 + lda;
kusano 2b45e8
      a_offset4  = a_offset3 + lda;
kusano 2b45e8
      a_offset  += 4 * lda;
kusano 2b45e8
      
kusano 2b45e8
      b_offset1  = b_offset;
kusano 2b45e8
      b_offset  += 16;
kusano 2b45e8
kusano 2b45e8
      i = (n >> 2);
kusano 2b45e8
      if (i > 0){
kusano 2b45e8
	do{
kusano 2b45e8
	  ctemp1  = *(a_offset1 + 0);
kusano 2b45e8
	  ctemp2  = *(a_offset1 + 1);
kusano 2b45e8
	  ctemp3  = *(a_offset1 + 2);
kusano 2b45e8
	  ctemp4  = *(a_offset1 + 3);
kusano 2b45e8
	  
kusano 2b45e8
	  ctemp5  = *(a_offset2 + 0);
kusano 2b45e8
	  ctemp6  = *(a_offset2 + 1);
kusano 2b45e8
	  ctemp7  = *(a_offset2 + 2);
kusano 2b45e8
	  ctemp8  = *(a_offset2 + 3);
kusano 2b45e8
	  
kusano 2b45e8
	  ctemp9  = *(a_offset3 + 0);
kusano 2b45e8
	  ctemp10 = *(a_offset3 + 1);
kusano 2b45e8
	  ctemp11 = *(a_offset3 + 2);
kusano 2b45e8
	  ctemp12 = *(a_offset3 + 3);
kusano 2b45e8
	  
kusano 2b45e8
	  ctemp13 = *(a_offset4 + 0);
kusano 2b45e8
	  ctemp14 = *(a_offset4 + 1);
kusano 2b45e8
	  ctemp15 = *(a_offset4 + 2);
kusano 2b45e8
	  ctemp16 = *(a_offset4 + 3);
kusano 2b45e8
kusano 2b45e8
	  a_offset1 += 4;
kusano 2b45e8
	  a_offset2 += 4;
kusano 2b45e8
	  a_offset3 += 4;
kusano 2b45e8
	  a_offset4 += 4;
kusano 2b45e8
	
kusano 2b45e8
	  *(b_offset1 +  0) = ctemp1;
kusano 2b45e8
	  *(b_offset1 +  1) = ctemp2;
kusano 2b45e8
	  *(b_offset1 +  2) = ctemp3;
kusano 2b45e8
	  *(b_offset1 +  3) = ctemp4;
kusano 2b45e8
	
kusano 2b45e8
	  *(b_offset1 +  4) = ctemp5;
kusano 2b45e8
	  *(b_offset1 +  5) = ctemp6;
kusano 2b45e8
	  *(b_offset1 +  6) = ctemp7;
kusano 2b45e8
	  *(b_offset1 +  7) = ctemp8;
kusano 2b45e8
	  
kusano 2b45e8
	  *(b_offset1 +  8) = ctemp9;
kusano 2b45e8
	  *(b_offset1 +  9) = ctemp10;
kusano 2b45e8
	  *(b_offset1 + 10) = ctemp11;
kusano 2b45e8
	  *(b_offset1 + 11) = ctemp12;
kusano 2b45e8
	  
kusano 2b45e8
	  *(b_offset1 + 12) = ctemp13;
kusano 2b45e8
	  *(b_offset1 + 13) = ctemp14;
kusano 2b45e8
	  *(b_offset1 + 14) = ctemp15;
kusano 2b45e8
	  *(b_offset1 + 15) = ctemp16;
kusano 2b45e8
	  
kusano 2b45e8
	  b_offset1 += m * 4;
kusano 2b45e8
	  i --;
kusano 2b45e8
	}while(i > 0);
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      if (n & 2) {
kusano 2b45e8
	ctemp1  = *(a_offset1 + 0);
kusano 2b45e8
	ctemp2  = *(a_offset1 + 1);
kusano 2b45e8
kusano 2b45e8
	ctemp3  = *(a_offset2 + 0);
kusano 2b45e8
	ctemp4  = *(a_offset2 + 1);
kusano 2b45e8
	
kusano 2b45e8
	ctemp5  = *(a_offset3 + 0);
kusano 2b45e8
	ctemp6  = *(a_offset3 + 1);
kusano 2b45e8
	
kusano 2b45e8
	ctemp7  = *(a_offset4 + 0);
kusano 2b45e8
	ctemp8  = *(a_offset4 + 1);
kusano 2b45e8
	
kusano 2b45e8
	a_offset1 += 2;
kusano 2b45e8
	a_offset2 += 2;
kusano 2b45e8
	a_offset3 += 2;
kusano 2b45e8
	a_offset4 += 2;
kusano 2b45e8
	
kusano 2b45e8
	*(b_offset2 +  0) = ctemp1;
kusano 2b45e8
	*(b_offset2 +  1) = ctemp2;
kusano 2b45e8
	*(b_offset2 +  2) = ctemp3;
kusano 2b45e8
	*(b_offset2 +  3) = ctemp4;
kusano 2b45e8
	
kusano 2b45e8
	*(b_offset2 +  4) = ctemp5;
kusano 2b45e8
	*(b_offset2 +  5) = ctemp6;
kusano 2b45e8
	*(b_offset2 +  6) = ctemp7;
kusano 2b45e8
	*(b_offset2 +  7) = ctemp8;
kusano 2b45e8
	
kusano 2b45e8
	b_offset2 += 8;
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      if (n & 1) {
kusano 2b45e8
	ctemp1  = *(a_offset1 + 0);
kusano 2b45e8
	ctemp2  = *(a_offset2 + 0);
kusano 2b45e8
	ctemp3  = *(a_offset3 + 0);
kusano 2b45e8
	ctemp4  = *(a_offset4 + 0);
kusano 2b45e8
	
kusano 2b45e8
	*(b_offset3 +  0) = ctemp1;
kusano 2b45e8
	*(b_offset3 +  1) = ctemp2;
kusano 2b45e8
	*(b_offset3 +  2) = ctemp3;
kusano 2b45e8
	*(b_offset3 +  3) = ctemp4;
kusano 2b45e8
	
kusano 2b45e8
	b_offset3 += 4;
kusano 2b45e8
      }
kusano 2b45e8
kusano 2b45e8
      j--;
kusano 2b45e8
    }while(j > 0);
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (m & 2){
kusano 2b45e8
    a_offset1  = a_offset;
kusano 2b45e8
    a_offset2  = a_offset1 + lda;
kusano 2b45e8
    a_offset  += 2 * lda;
kusano 2b45e8
    
kusano 2b45e8
    b_offset1  = b_offset;
kusano 2b45e8
    b_offset  += 8;
kusano 2b45e8
    
kusano 2b45e8
    i = (n >> 2);
kusano 2b45e8
    if (i > 0){
kusano 2b45e8
      do{
kusano 2b45e8
	ctemp1  = *(a_offset1 + 0);
kusano 2b45e8
	ctemp2  = *(a_offset1 + 1);
kusano 2b45e8
	ctemp3  = *(a_offset1 + 2);
kusano 2b45e8
	ctemp4  = *(a_offset1 + 3);
kusano 2b45e8
	
kusano 2b45e8
	ctemp5  = *(a_offset2 + 0);
kusano 2b45e8
	ctemp6  = *(a_offset2 + 1);
kusano 2b45e8
	ctemp7  = *(a_offset2 + 2);
kusano 2b45e8
	ctemp8  = *(a_offset2 + 3);
kusano 2b45e8
	
kusano 2b45e8
	a_offset1 += 4;
kusano 2b45e8
	a_offset2 += 4;
kusano 2b45e8
	
kusano 2b45e8
	*(b_offset1 +  0) = ctemp1;
kusano 2b45e8
	*(b_offset1 +  1) = ctemp2;
kusano 2b45e8
	*(b_offset1 +  2) = ctemp3;
kusano 2b45e8
	*(b_offset1 +  3) = ctemp4;
kusano 2b45e8
	
kusano 2b45e8
	*(b_offset1 +  4) = ctemp5;
kusano 2b45e8
	*(b_offset1 +  5) = ctemp6;
kusano 2b45e8
	*(b_offset1 +  6) = ctemp7;
kusano 2b45e8
	*(b_offset1 +  7) = ctemp8;
kusano 2b45e8
kusano 2b45e8
	b_offset1 += m * 4;
kusano 2b45e8
	i --;
kusano 2b45e8
      }while(i > 0);
kusano 2b45e8
    }
kusano 2b45e8
kusano 2b45e8
    if (n & 2) {
kusano 2b45e8
      ctemp1  = *(a_offset1 + 0);
kusano 2b45e8
      ctemp2  = *(a_offset1 + 1);
kusano 2b45e8
      
kusano 2b45e8
      ctemp3  = *(a_offset2 + 0);
kusano 2b45e8
      ctemp4  = *(a_offset2 + 1);
kusano 2b45e8
      
kusano 2b45e8
      a_offset1 += 2;
kusano 2b45e8
      a_offset2 += 2;
kusano 2b45e8
      
kusano 2b45e8
      *(b_offset2 +  0) = ctemp1;
kusano 2b45e8
      *(b_offset2 +  1) = ctemp2;
kusano 2b45e8
      *(b_offset2 +  2) = ctemp3;
kusano 2b45e8
      *(b_offset2 +  3) = ctemp4;
kusano 2b45e8
      
kusano 2b45e8
      b_offset2 += 4;
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
    if (n & 1) {
kusano 2b45e8
      ctemp1  = *(a_offset1 + 0);
kusano 2b45e8
      ctemp2  = *(a_offset2 + 0);
kusano 2b45e8
      
kusano 2b45e8
      *(b_offset3 +  0) = ctemp1;
kusano 2b45e8
      *(b_offset3 +  1) = ctemp2;
kusano 2b45e8
      b_offset3 += 2;
kusano 2b45e8
    }
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  if (m & 1){
kusano 2b45e8
    a_offset1  = a_offset;
kusano 2b45e8
    b_offset1  = b_offset;
kusano 2b45e8
    
kusano 2b45e8
    i = (n >> 2);
kusano 2b45e8
    if (i > 0){
kusano 2b45e8
      do{
kusano 2b45e8
	ctemp1  = *(a_offset1 + 0);
kusano 2b45e8
	ctemp2  = *(a_offset1 + 1);
kusano 2b45e8
	ctemp3  = *(a_offset1 + 2);
kusano 2b45e8
	ctemp4  = *(a_offset1 + 3);
kusano 2b45e8
	
kusano 2b45e8
	a_offset1 += 4;
kusano 2b45e8
	
kusano 2b45e8
	*(b_offset1 +  0) = ctemp1;
kusano 2b45e8
	*(b_offset1 +  1) = ctemp2;
kusano 2b45e8
	*(b_offset1 +  2) = ctemp3;
kusano 2b45e8
	*(b_offset1 +  3) = ctemp4;
kusano 2b45e8
kusano 2b45e8
	b_offset1 += 4 * m;
kusano 2b45e8
kusano 2b45e8
	i --;
kusano 2b45e8
      }while(i > 0);
kusano 2b45e8
    }
kusano 2b45e8
kusano 2b45e8
    if (n & 2) {
kusano 2b45e8
      ctemp1  = *(a_offset1 + 0);
kusano 2b45e8
      ctemp2  = *(a_offset1 + 1);
kusano 2b45e8
      a_offset1 += 2;
kusano 2b45e8
      
kusano 2b45e8
      *(b_offset2 +  0) = ctemp1;
kusano 2b45e8
      *(b_offset2 +  1) = ctemp2;
kusano 2b45e8
    }
kusano 2b45e8
    
kusano 2b45e8
    if (n & 1) {
kusano 2b45e8
      ctemp1  = *(a_offset1 + 0);
kusano 2b45e8
      *(b_offset3 +  0) = ctemp1;
kusano 2b45e8
    }
kusano 2b45e8
  }
kusano 2b45e8
kusano 2b45e8
  return 0;
kusano 2b45e8
}