extras.c

/**************************************************************************
**
** Copyright (C) 1993 David E. Steward & Zbigniew Leyk, all rights reserved.
**
**			     Meschach Library
** 
** This Meschach Library is provided "as is" without any express 
** or implied warranty of any kind with respect to this software. 
** In particular the authors shall not be liable for any direct, 
** indirect, special, incidental or consequential damages arising 
** in any way from use of the software.
** 
** Everyone is granted permission to copy, modify and redistribute this
** Meschach Library, provided:
**  1.  All copies contain this copyright notice.
**  2.  All modified copies shall carry a notice stating who
**      made the last modification and the date of such modification.
**  3.  No charge is made for this software or works derived from it.  
**      This clause shall not be construed as constraining other software
**      distributed on the same medium as this software, nor is a
**      distribution fee considered a charge.
**
***************************************************************************/


/*
	Memory port routines: MEM_COPY and MEM_ZERO
*/

/* For BSD 4.[23] environments: using bcopy() and bzero() */

#include "machine.h"

#ifndef MEM_COPY
void	MEM_COPY(from,to,len)
char	*from, *to;
int	len;
{
    int		i;

    if ( from < to )
    {
	for ( i = 0; i < len; i++ )
	    *to++ = *from++;
    }
    else
    {
	from += len;	to += len;
	for ( i = 0; i < len; i++ )
	    *(--to) = *(--from);
    }
}
#endif

#ifndef MEM_ZERO
void	MEM_ZERO(ptr,len)
char	*ptr;
int	len;
{
    int		i;

    for ( i = 0; i < len; i++ )
	*(ptr++) = '\0';
}
#endif

/*
	This file contains versions of something approximating the well-known
	BLAS routines in C, suitable for Meschach (hence the `m').
	These are "vanilla" implementations, at least with some consideration
	of the effects of caching and paging, and maybe some loop unrolling
	for register-rich machines
*/

/*
	Organisation of matrices: it is assumed that matrices are represented
	by Real **'s. To keep flexibility, there is also an "initial
	column" parameter j0, so that the actual elements used are
		A[0][j0],   A[0][j0+1],   ..., A[0][j0+n-1]
		A[1][j0],   A[1][j0+1],   ..., A[1][j0+n-1]
		   ..         ..          ...      ..
		A[m-1][j0], A[m-1][j0+1], ..., A[m-1][j0+n-1]
*/

static char	rcsid[] = "$Id: extras.c,v 1.4 1995/06/08 15:13:15 des Exp $";

#include	<math.h>

#define	REGISTER_RICH	1

/* mblar-1 routines */

/* Mscale -- sets x <- alpha.x */
void	Mscale(len,alpha,x)
int	len;
double	alpha;
Real	*x;
{
    register int	i;

    for ( i = 0; i < len; i++ )
	x[i] *= alpha;
}

/* Mswap -- swaps x and y */
void	Mswap(len,x,y)
int	len;
Real	*x, *y;
{
    register int	i;
    register Real	tmp;

    for ( i = 0; i < len; i++ )
    {
	tmp = x[i];
	x[i] = y[i];
	y[i] = tmp;
    }
}

/* Mcopy -- copies x to y */
void	Mcopy(len,x,y)
int	len;
Real	*x, *y;
{
    register int	i;

    for ( i = 0; i < len; i++ )
	y[i] = x[i];
}

/* Maxpy -- y <- y + alpha.x */
void	Maxpy(len,alpha,x,y)
int	len;
double	alpha;
Real	*x, *y;
{
    register int	i, len4;

    /****************************************
    for ( i = 0; i < len; i++ )
	y[i] += alpha*x[i];
    ****************************************/

#ifdef REGISTER_RICH
    len4 = len / 4;
    len  = len % 4;
    for ( i = 0; i < len4; i++ )
    {
	y[4*i]   += alpha*x[4*i];
	y[4*i+1] += alpha*x[4*i+1];
	y[4*i+2] += alpha*x[4*i+2];
	y[4*i+3] += alpha*x[4*i+3];
    }
    x += 4*len4;	y += 4*len4;
#endif
    for ( i = 0; i < len; i++ )
	y[i] += alpha*x[i];
}

/* Mdot -- returns x'.y */
double	Mdot(len,x,y)
int	len;
Real	*x, *y;
{
    register int	i, len4;
    register Real	sum;

#ifndef REGISTER_RICH
    sum = 0.0;
#endif

#ifdef REGISTER_RICH
    register Real	sum0, sum1, sum2, sum3;
    
    sum0 = sum1 = sum2 = sum3 = 0.0;
    
    len4 = len / 4;
    len  = len % 4;
    
    for ( i = 0; i < len4; i++ )
    {
	sum0 += x[4*i  ]*y[4*i  ];
	sum1 += x[4*i+1]*y[4*i+1];
	sum2 += x[4*i+2]*y[4*i+2];
	sum3 += x[4*i+3]*y[4*i+3];
    }
    sum = sum0 + sum1 + sum2 + sum3;
    x += 4*len4;	y += 4*len4;
#endif

    for ( i = 0; i < len; i++ )
	sum += x[i]*y[i];

    return sum;
}

#ifndef ABS
#define	ABS(x)	((x) >= 0 ? (x) : -(x))
#endif

/* Mnorminf -- returns ||x||_inf */
double	Mnorminf(len,x)
int	len;
Real	*x;
{
    register int	i;
    register Real	tmp, max_val;

    max_val = 0.0;
    for ( i = 0; i < len; i++ )
    {
	tmp = ABS(x[i]);
	if ( max_val < tmp )
	    max_val = tmp;
    }

    return max_val;
}

/* Mnorm1 -- returns ||x||_1 */
double	Mnorm1(len,x)
int	len;
Real	*x;
{
    register int	i;
    register Real	sum;

    sum = 0.0;
    for ( i = 0; i < len; i++ )
	sum += ABS(x[i]);

    return sum;
}

/* Mnorm2 -- returns ||x||_2 */
double	Mnorm2(len,x)
int	len;
Real	*x;
{
    register int	i;
    register Real	norm, invnorm, sum, tmp;

    norm = Mnorminf(len,x);
    if ( norm == 0.0 )
	return 0.0;
    invnorm = 1.0/norm;
    sum = 0.0;
    for ( i = 0; i < len; i++ )
    {
	tmp = x[i]*invnorm;
	sum += tmp*tmp;
    }

    return sum/invnorm;
}

/* mblar-2 routines */

/* Mmv -- y <- alpha.A.x + beta.y */
void	Mmv(m,n,alpha,A,j0,x,beta,y)
int	m, n, j0;
double	alpha, beta;
Real	**A, *x, *y;
{
    register int	i, j, m4, n4;
    register Real	sum0, sum1, sum2, sum3, tmp0, tmp1, tmp2, tmp3;
    register Real	*dp0, *dp1, *dp2, *dp3;

    /****************************************
    for ( i = 0; i < m; i++ )
	y[i] += alpha*Mdot(n,&(A[i][j0]),x);
    ****************************************/

    m4 = n4 = 0;

#ifdef REGISTER_RICH
    m4 = m / 4;
    m  = m % 4;
    n4 = n / 4;
    n  = n % 4;

    for ( i = 0; i < m4; i++ )
    {
	sum0 = sum1 = sum2 = sum3 = 0.0;
	dp0 = &(A[4*i  ][j0]);
	dp1 = &(A[4*i+1][j0]);
	dp2 = &(A[4*i+2][j0]);
	dp3 = &(A[4*i+3][j0]);

	for ( j = 0; j < n4; j++ )
	{
	    tmp0 = x[4*j  ];
	    tmp1 = x[4*j+1];
	    tmp2 = x[4*j+2];
	    tmp3 = x[4*j+3];
	    sum0 = sum0 + dp0[j]*tmp0 + dp0[j+1]*tmp1 +
		dp0[j+2]*tmp2 + dp0[j+3]*tmp3;
	    sum1 = sum1 + dp1[j]*tmp0 + dp1[j+1]*tmp1 +
		dp1[j+2]*tmp2 + dp1[j+3]*tmp3;
	    sum2 = sum2 + dp2[j]*tmp0 + dp2[j+1]*tmp1 +
		dp2[j+2]*tmp2 + dp2[j+3]*tmp3;
	    sum3 = sum3 + dp3[j]*tmp0 + dp3[j+1]*tmp2 +
		dp3[j+2]*tmp2 + dp3[j+3]*tmp3;
	}
	for ( j = 0; j < n; j++ )
	{
	    sum0 += dp0[4*n4+j]*x[4*n4+j];
	    sum1 += dp1[4*n4+j]*x[4*n4+j];
	    sum2 += dp2[4*n4+j]*x[4*n4+j];
	    sum3 += dp3[4*n4+j]*x[4*n4+j];
	}
	y[4*i  ] = beta*y[4*i  ] + alpha*sum0;
	y[4*i+1] = beta*y[4*i+1] + alpha*sum1;
	y[4*i+2] = beta*y[4*i+2] + alpha*sum2;
	y[4*i+3] = beta*y[4*i+3] + alpha*sum3;
    }
#endif

    for ( i = 0; i < m; i++ )
	y[4*m4+i] = beta*y[i] + alpha*Mdot(4*n4+n,&(A[4*m4+i][j0]),x);
}

/* Mvm -- y <- alpha.A^T.x + beta.y */
void	Mvm(m,n,alpha,A,j0,x,beta,y)
int	m, n, j0;
double	alpha, beta;
Real	**A, *x, *y;
{
    register int	i, j, m4, n2;
    register Real	*Aref;
    register Real 	tmp;

#ifdef REGISTER_RICH
    register Real	*Aref0, *Aref1;
    register Real	tmp0, tmp1;
    register Real	yval0, yval1, yval2, yval3;
#endif

    if ( beta != 1.0 )
	Mscale(m,beta,y);
    /****************************************
    for ( j = 0; j < n; j++ )
	Maxpy(m,alpha*x[j],&(A[j][j0]),y);
    ****************************************/
    m4 = n2 = 0;

    m4 = m / 4;
    m  = m % 4;
#ifdef REGISTER_RICH
    n2 = n / 2;
    n  = n % 2;

    for ( j = 0; j < n2; j++ )
    {
	tmp0 = alpha*x[2*j];
	tmp1 = alpha*x[2*j+1];
	Aref0 = &(A[2*j  ][j0]);
	Aref1 = &(A[2*j+1][j0]);
	for ( i = 0; i < m4; i++ )
	{
	    yval0 = y[4*i  ] + tmp0*Aref0[4*i  ];
	    yval1 = y[4*i+1] + tmp0*Aref0[4*i+1];
	    yval2 = y[4*i+2] + tmp0*Aref0[4*i+2];
	    yval3 = y[4*i+3] + tmp0*Aref0[4*i+3];
	    y[4*i  ] = yval0 + tmp1*Aref1[4*i  ];
	    y[4*i+1] = yval1 + tmp1*Aref1[4*i+1];
	    y[4*i+2] = yval2 + tmp1*Aref1[4*i+2];
	    y[4*i+3] = yval3 + tmp1*Aref1[4*i+3];
	}
	y += 4*m4;	Aref0 += 4*m4;	Aref1 += 4*m4;
	for ( i = 0; i < m; i++ )
	    y[i] += tmp0*Aref0[i] + tmp1*Aref1[i];
    }
#endif

    for ( j = 0; j < n; j++ )
    {
	tmp = alpha*x[2*n2+j];
	Aref = &(A[2*n2+j][j0]);
	for ( i = 0; i < m4; i++ )
	{
	    y[4*i  ] += tmp*Aref[4*i  ];
	    y[4*i+1] += tmp*Aref[4*i+1];
	    y[4*i+2] += tmp*Aref[4*i+2];
	    y[4*i+3] += tmp*Aref[4*i+3];
	}
	y += 4*m4;	Aref += 4*m4;
	for ( i = 0; i < m; i++ )
	    y[i] += tmp*Aref[i];
    }
}

/* Mupdate -- A <- A + alpha.x.y^T */
void	Mupdate(m,n,alpha,x,y,A,j0)
int	m, n, j0;
double	alpha;
Real	**A, *x, *y;
{
    register int	i, j, n4;
    register Real	*Aref;
    register Real 	tmp;

    /****************************************
    for ( i = 0; i < m; i++ )
	Maxpy(n,alpha*x[i],y,&(A[i][j0]));
    ****************************************/

    n4 = n / 4;
    n  = n % 4;
    for ( i = 0; i < m; i++ )
    {
	tmp = alpha*x[i];
	Aref = &(A[i][j0]);
	for ( j = 0; j < n4; j++ )
	{
	    Aref[4*j  ] += tmp*y[4*j  ];
	    Aref[4*j+1] += tmp*y[4*j+1];
	    Aref[4*j+2] += tmp*y[4*j+2];
	    Aref[4*j+3] += tmp*y[4*j+3];
	}
	Aref += 4*n4;	y += 4*n4;
	for ( j = 0; j < n; j++ )
	    Aref[j] += tmp*y[j];
    }
}

/* mblar-3 routines */

/* Mmm -- C <- C + alpha.A.B */
void	Mmm(m,n,p,alpha,A,Aj0,B,Bj0,C,Cj0)
int	m, n, p;	/* C is m x n */
double  alpha;
Real	**A, **B, **C;
int	Aj0, Bj0, Cj0;
{
    register int	i, j, k;
    /* register Real	tmp, sum; */

    /****************************************
    for ( i = 0; i < m; i++ )
	for ( k = 0; k < p; k++ )
	    Maxpy(n,alpha*A[i][Aj0+k],&(B[k][Bj0]),&(C[i][Cj0]));
    ****************************************/
    for ( i = 0; i < m; i++ )
	Mvm(p,n,alpha,B,Bj0,&(A[i][Aj0]),1.0,&(C[i][Cj0]));
}

/* Mmtrm -- C <- C + alpha.A^T.B */
void	Mmtrm(m,n,p,alpha,A,Aj0,B,Bj0,C,Cj0)
int	m, n, p;	/* C is m x n */
double  alpha;
Real	**A, **B, **C;
int	Aj0, Bj0, Cj0;
{
    register int	i, j, k;

    /****************************************
    for ( i = 0; i < m; i++ )
	for ( k = 0; k < p; k++ )
	    Maxpy(n,alpha*A[k][Aj0+i],&(B[k][Bj0]),&(C[i][Cj0]));
    ****************************************/
    for ( k = 0; k < p; k++ )
	Mupdate(m,n,alpha,&(A[k][Aj0]),&(B[k][Bj0]),C,Cj0);
}

/* Mmmtr -- C <- C + alpha.A.B^T */
void	Mmmtr(m,n,p,alpha,A,Aj0,B,Bj0,C,Cj0)
int	m, n, p;	/* C is m x n */
double  alpha;
Real	**A, **B, **C;
int	Aj0, Bj0, Cj0;
{
    register int	i, j, k;

    /****************************************
    for ( i = 0; i < m; i++ )
	for ( j = 0; j < n; j++ )
	    C[i][Cj0+j] += alpha*Mdot(p,&(A[i][Aj0]),&(B[j][Bj0]));
    ****************************************/
    for ( i = 0; i < m; i++ )
	Mmv(n,p,alpha,B,Bj0,&(A[i][Aj0]),1.0,&(C[i][Cj0]));
}

/* Mmtrmtr -- C <- C + alpha.A^T.B^T */
void	Mmtrmtr(m,n,p,alpha,A,Aj0,B,Bj0,C,Cj0)
int	m, n, p;	/* C is m x n */
double  alpha;
Real	**A, **B, **C;
int	Aj0, Bj0, Cj0;
{
    register int	i, j, k;

    for ( i = 0; i < m; i++ )
	for ( j = 0; j < n; j++ )
	    for ( k = 0; k < p; k++ )
		C[i][Cj0+j] += A[i][Aj0+k]*B[k][Bj0+j];
}