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
#define ASSEMBLER
kusano 2b45e8
#include "common.h"
kusano 2b45e8
kusano 2b45e8
#define N	$4
kusano 2b45e8
#define	X	$5
kusano 2b45e8
#define INCX	$6
kusano 2b45e8
#define Y	$7
kusano 2b45e8
#define INCY	$8
kusano 2b45e8
	
kusano 2b45e8
#define XX	$9
kusano 2b45e8
#define YY	$10
kusano 2b45e8
kusano 2b45e8
#define C	$f17
kusano 2b45e8
#define S	$f18
kusano 2b45e8
kusano 2b45e8
#define I	$2
kusano 2b45e8
#define TEMP	$3
kusano 2b45e8
kusano 2b45e8
#define a1	$f4
kusano 2b45e8
#define a2	$f5
kusano 2b45e8
#define a3	$f6
kusano 2b45e8
#define a4	$f7
kusano 2b45e8
kusano 2b45e8
#define b1	$f8
kusano 2b45e8
#define b2	$f9
kusano 2b45e8
#define b3	$f10
kusano 2b45e8
#define b4	$f11
kusano 2b45e8
kusano 2b45e8
#define t1	$f0
kusano 2b45e8
#define t2	$f1
kusano 2b45e8
#define t3	$f2
kusano 2b45e8
#define t4	$f3
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	
kusano 2b45e8
	dsll	INCX, INCX, BASE_SHIFT
kusano 2b45e8
	li	TEMP, SIZE
kusano 2b45e8
kusano 2b45e8
	blez	N, .L999
kusano 2b45e8
	dsll	INCY, INCY, BASE_SHIFT
kusano 2b45e8
kusano 2b45e8
	bne	INCX, TEMP, .L20
kusano 2b45e8
	dsra	I, N, 2
kusano 2b45e8
kusano 2b45e8
	bne	INCY, TEMP, .L20
kusano 2b45e8
	NOP
kusano 2b45e8
kusano 2b45e8
	blez	I, .L15
kusano 2b45e8
	daddiu	I, I, -1
kusano 2b45e8
kusano 2b45e8
	LD	a1,  0 * SIZE(X)
kusano 2b45e8
	LD	b1,  0 * SIZE(Y)
kusano 2b45e8
	LD	a2,  1 * SIZE(X)
kusano 2b45e8
	LD	b2,  1 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	LD	a3,  2 * SIZE(X)
kusano 2b45e8
	LD	b3,  2 * SIZE(Y)
kusano 2b45e8
	MUL	t1, S, b1
kusano 2b45e8
kusano 2b45e8
	LD	a4,  3 * SIZE(X)
kusano 2b45e8
	MUL	t2, C, b1
kusano 2b45e8
	LD	b4,  3 * SIZE(Y)
kusano 2b45e8
	MUL	t3, S, b2
kusano 2b45e8
kusano 2b45e8
	blez	I, .L13
kusano 2b45e8
	MUL	t4, C, b2
kusano 2b45e8
	.align 3
kusano 2b45e8
kusano 2b45e8
.L12:
kusano 2b45e8
	MADD	t1, t1, C, a1
kusano 2b45e8
	LD	b1,  4 * SIZE(Y)
kusano 2b45e8
	NMSUB	t2, t2, S, a1
kusano 2b45e8
	LD	a1,  4 * SIZE(X)
kusano 2b45e8
	MADD	t3, t3, C, a2
kusano 2b45e8
	LD	b2,  5 * SIZE(Y)
kusano 2b45e8
	NMSUB	t4, t4, S, a2
kusano 2b45e8
	LD	a2,  5 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	ST	t1,  0 * SIZE(X)
kusano 2b45e8
	MUL	t1, S, b3
kusano 2b45e8
	ST	t2,  0 * SIZE(Y)
kusano 2b45e8
	MUL	t2, C, b3
kusano 2b45e8
	ST	t3,  1 * SIZE(X)
kusano 2b45e8
	MUL	t3, S, b4
kusano 2b45e8
	ST	t4,  1 * SIZE(Y)
kusano 2b45e8
	MUL	t4, C, b4
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
	MADD	t1, t1, C, a3
kusano 2b45e8
	LD	b3,  6 * SIZE(Y)
kusano 2b45e8
	NMSUB	t2, t2, S, a3
kusano 2b45e8
	LD	a3,  6 * SIZE(X)
kusano 2b45e8
	MADD	t3, t3, C, a4
kusano 2b45e8
	LD	b4,  7 * SIZE(Y)
kusano 2b45e8
	NMSUB	t4, t4, S, a4
kusano 2b45e8
	LD	a4,  7 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	ST	t1,  2 * SIZE(X)
kusano 2b45e8
	MUL	t1, S, b1
kusano 2b45e8
	ST	t2,  2 * SIZE(Y)
kusano 2b45e8
	MUL	t2, C, b1
kusano 2b45e8
	ST	t3,  3 * SIZE(X)
kusano 2b45e8
	MUL	t3, S, b2
kusano 2b45e8
	ST	t4,  3 * SIZE(Y)
kusano 2b45e8
	MUL	t4, C, b2
kusano 2b45e8
kusano 2b45e8
	daddiu	I, I, -1
kusano 2b45e8
	daddiu	X, X, 4 * SIZE
kusano 2b45e8
kusano 2b45e8
	bgtz	I, .L12
kusano 2b45e8
	daddiu	Y, Y, 4 * SIZE
kusano 2b45e8
	.align 3
kusano 2b45e8
kusano 2b45e8
.L13:
kusano 2b45e8
	MADD	t1, t1, C, a1
kusano 2b45e8
	NMSUB	t2, t2, S, a1
kusano 2b45e8
	MADD	t3, t3, C, a2
kusano 2b45e8
	NMSUB	t4, t4, S, a2
kusano 2b45e8
kusano 2b45e8
	ST	t1,  0 * SIZE(X)
kusano 2b45e8
	MUL	t1, S, b3
kusano 2b45e8
	ST	t2,  0 * SIZE(Y)
kusano 2b45e8
	MUL	t2, C, b3
kusano 2b45e8
	ST	t3,  1 * SIZE(X)
kusano 2b45e8
	MUL	t3, S, b4
kusano 2b45e8
	ST	t4,  1 * SIZE(Y)
kusano 2b45e8
	MUL	t4, C, b4
kusano 2b45e8
kusano 2b45e8
	MADD	t1, t1, C, a3
kusano 2b45e8
	NMSUB	t2, t2, S, a3
kusano 2b45e8
	MADD	t3, t3, C, a4
kusano 2b45e8
	daddiu	X, X, 4 * SIZE
kusano 2b45e8
	NMSUB	t4, t4, S, a4
kusano 2b45e8
	daddiu	Y, Y, 4 * SIZE
kusano 2b45e8
kusano 2b45e8
	ST	t1, -2 * SIZE(X)
kusano 2b45e8
	ST	t2, -2 * SIZE(Y)
kusano 2b45e8
	ST	t3, -1 * SIZE(X)
kusano 2b45e8
	ST	t4, -1 * SIZE(Y)
kusano 2b45e8
	.align 3
kusano 2b45e8
kusano 2b45e8
.L15:
kusano 2b45e8
	andi	I,  N, 3
kusano 2b45e8
kusano 2b45e8
	blez	I, .L999
kusano 2b45e8
	NOP
kusano 2b45e8
	.align	3
kusano 2b45e8
kusano 2b45e8
.L16:
kusano 2b45e8
	LD	a1,  0 * SIZE(X)
kusano 2b45e8
	LD	b1,  0 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	MUL	t1, S, b1
kusano 2b45e8
	MUL	t2, C, b1
kusano 2b45e8
kusano 2b45e8
	MADD	t1, t1, C, a1
kusano 2b45e8
	NMSUB	t2, t2, S, a1
kusano 2b45e8
kusano 2b45e8
	ST	t1,  0 * SIZE(X)
kusano 2b45e8
	ST	t2,  0 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	daddiu I, I, -1
kusano 2b45e8
kusano 2b45e8
	daddiu	X, X, SIZE
kusano 2b45e8
	daddiu	Y, Y, SIZE
kusano 2b45e8
kusano 2b45e8
	bgtz	I, .L16
kusano 2b45e8
	NOP
kusano 2b45e8
	j	.L999
kusano 2b45e8
	NOP
kusano 2b45e8
	.align 3
kusano 2b45e8
kusano 2b45e8
.L20:
kusano 2b45e8
	move	XX, X
kusano 2b45e8
	move	YY, Y
kusano 2b45e8
kusano 2b45e8
	blez	I, .L25
kusano 2b45e8
	daddiu	I, I, -1
kusano 2b45e8
kusano 2b45e8
	LD	a1,  0 * SIZE(X)
kusano 2b45e8
	dadd	X, X, INCX
kusano 2b45e8
	LD	b1,  0 * SIZE(Y)
kusano 2b45e8
	dadd	Y, Y, INCY
kusano 2b45e8
kusano 2b45e8
	LD	a2,  0 * SIZE(X)
kusano 2b45e8
	dadd	X, X, INCX
kusano 2b45e8
	LD	b2,  0 * SIZE(Y)
kusano 2b45e8
	dadd	Y, Y, INCY
kusano 2b45e8
kusano 2b45e8
	LD	a3,  0 * SIZE(X)
kusano 2b45e8
	dadd	X, X, INCX
kusano 2b45e8
	LD	b3,  0 * SIZE(Y)
kusano 2b45e8
	dadd	Y, Y, INCY
kusano 2b45e8
kusano 2b45e8
	MUL	t1, S, b1
kusano 2b45e8
kusano 2b45e8
	LD	a4,  0 * SIZE(X)
kusano 2b45e8
	dadd	X, X, INCX
kusano 2b45e8
	MUL	t2, C, b1
kusano 2b45e8
	LD	b4,  0 * SIZE(Y)
kusano 2b45e8
	dadd	Y, Y, INCY
kusano 2b45e8
kusano 2b45e8
	MUL	t3, S, b2
kusano 2b45e8
	blez	I, .L23
kusano 2b45e8
	MUL	t4, C, b2
kusano 2b45e8
	.align 3
kusano 2b45e8
kusano 2b45e8
.L22:
kusano 2b45e8
	MADD	t1, t1, C, a1
kusano 2b45e8
	LD	b1,  0 * SIZE(Y)
kusano 2b45e8
	dadd	Y, Y, INCY
kusano 2b45e8
	NMSUB	t2, t2, S, a1
kusano 2b45e8
	LD	a1,  0 * SIZE(X)
kusano 2b45e8
	dadd	X, X, INCX
kusano 2b45e8
	MADD	t3, t3, C, a2
kusano 2b45e8
	LD	b2,  0 * SIZE(Y)
kusano 2b45e8
	dadd	Y, Y, INCY
kusano 2b45e8
	NMSUB	t4, t4, S, a2
kusano 2b45e8
	LD	a2,  0 * SIZE(X)
kusano 2b45e8
	dadd	X, X, INCX
kusano 2b45e8
kusano 2b45e8
	ST	t1,  0 * SIZE(XX)
kusano 2b45e8
	dadd	XX, XX, INCX
kusano 2b45e8
	MUL	t1, S, b3
kusano 2b45e8
	ST	t2,  0 * SIZE(YY)
kusano 2b45e8
	dadd	YY, YY, INCY
kusano 2b45e8
	MUL	t2, C, b3
kusano 2b45e8
	ST	t3,  0 * SIZE(XX)
kusano 2b45e8
	dadd	XX, XX, INCX
kusano 2b45e8
	MUL	t3, S, b4
kusano 2b45e8
	ST	t4,  0 * SIZE(YY)
kusano 2b45e8
	dadd	YY, YY, INCY
kusano 2b45e8
	MUL	t4, C, b4
kusano 2b45e8
kusano 2b45e8
	MADD	t1, t1, C, a3
kusano 2b45e8
	LD	b3,  0 * SIZE(Y)
kusano 2b45e8
	dadd	Y, Y, INCY
kusano 2b45e8
	NMSUB	t2, t2, S, a3
kusano 2b45e8
	LD	a3,  0 * SIZE(X)
kusano 2b45e8
	dadd	X, X, INCX
kusano 2b45e8
	MADD	t3, t3, C, a4
kusano 2b45e8
	LD	b4,  0 * SIZE(Y)
kusano 2b45e8
	dadd	Y, Y, INCY
kusano 2b45e8
	NMSUB	t4, t4, S, a4
kusano 2b45e8
	LD	a4,  0 * SIZE(X)
kusano 2b45e8
	dadd	X, X, INCX
kusano 2b45e8
kusano 2b45e8
	ST	t1,  0 * SIZE(XX)
kusano 2b45e8
	dadd	XX, XX, INCX
kusano 2b45e8
	MUL	t1, S, b1
kusano 2b45e8
	ST	t2,  0 * SIZE(YY)
kusano 2b45e8
	dadd	YY, YY, INCY
kusano 2b45e8
	MUL	t2, C, b1
kusano 2b45e8
	ST	t3,  0 * SIZE(XX)
kusano 2b45e8
	dadd	XX, XX, INCX
kusano 2b45e8
	MUL	t3, S, b2
kusano 2b45e8
	ST	t4,  0 * SIZE(YY)
kusano 2b45e8
	MUL	t4, C, b2
kusano 2b45e8
	daddiu	I, I, -1
kusano 2b45e8
kusano 2b45e8
	bgtz	I, .L22
kusano 2b45e8
	dadd	YY, YY, INCY
kusano 2b45e8
	.align 3
kusano 2b45e8
kusano 2b45e8
.L23:
kusano 2b45e8
	MADD	t1, t1, C, a1
kusano 2b45e8
	NMSUB	t2, t2, S, a1
kusano 2b45e8
	MADD	t3, t3, C, a2
kusano 2b45e8
	NMSUB	t4, t4, S, a2
kusano 2b45e8
kusano 2b45e8
	ST	t1,  0 * SIZE(XX)
kusano 2b45e8
	dadd	XX, XX, INCX
kusano 2b45e8
	MUL	t1, S, b3
kusano 2b45e8
	ST	t2,  0 * SIZE(YY)
kusano 2b45e8
	dadd	YY, YY, INCY
kusano 2b45e8
	MUL	t2, C, b3
kusano 2b45e8
	ST	t3,  0 * SIZE(XX)
kusano 2b45e8
	dadd	XX, XX, INCX
kusano 2b45e8
	MUL	t3, S, b4
kusano 2b45e8
	ST	t4,  0 * SIZE(YY)
kusano 2b45e8
	dadd	YY, YY, INCY
kusano 2b45e8
	MUL	t4, C, b4
kusano 2b45e8
kusano 2b45e8
	MADD	t1, t1, C, a3
kusano 2b45e8
	NMSUB	t2, t2, S, a3
kusano 2b45e8
	MADD	t3, t3, C, a4
kusano 2b45e8
	NMSUB	t4, t4, S, a4
kusano 2b45e8
kusano 2b45e8
	ST	t1,  0 * SIZE(XX)
kusano 2b45e8
	dadd	XX, XX, INCX
kusano 2b45e8
	ST	t2,  0 * SIZE(YY)
kusano 2b45e8
	dadd	YY, YY, INCY
kusano 2b45e8
	ST	t3,  0 * SIZE(XX)
kusano 2b45e8
	dadd	XX, XX, INCX
kusano 2b45e8
	ST	t4,  0 * SIZE(YY)
kusano 2b45e8
	dadd	YY, YY, INCY
kusano 2b45e8
	.align 3
kusano 2b45e8
kusano 2b45e8
.L25:
kusano 2b45e8
	andi	I,  N, 3
kusano 2b45e8
kusano 2b45e8
	blez	I, .L999
kusano 2b45e8
	NOP
kusano 2b45e8
	.align	3
kusano 2b45e8
kusano 2b45e8
.L26:
kusano 2b45e8
	LD	a1,  0 * SIZE(X)
kusano 2b45e8
	LD	b1,  0 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	MUL	t1, S, b1
kusano 2b45e8
	MUL	t2, C, b1
kusano 2b45e8
kusano 2b45e8
	MADD	t1, t1, C, a1
kusano 2b45e8
	daddiu	I, I, -1
kusano 2b45e8
	NMSUB	t2, t2, S, a1
kusano 2b45e8
kusano 2b45e8
	ST	t1,  0 * SIZE(X)
kusano 2b45e8
	ST	t2,  0 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	dadd	X, X, INCX
kusano 2b45e8
	bgtz	I, .L26
kusano 2b45e8
	dadd	Y, Y, INCY
kusano 2b45e8
	.align 3
kusano 2b45e8
kusano 2b45e8
.L999:
kusano 2b45e8
	j	$31
kusano 2b45e8
	NOP
kusano 2b45e8
kusano 2b45e8
	EPILOGUE