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
kusano 2b45e8
#include "common.h"
kusano 2b45e8
#include "version.h"
kusano 2b45e8
kusano 2b45e8
#define PREFETCH_SIZE 80
kusano 2b45e8
kusano 2b45e8
#define N	$16
kusano 2b45e8
#define X	$17
kusano 2b45e8
#define INCX	$18
kusano 2b45e8
#define XX	$19
kusano 2b45e8
kusano 2b45e8
#define I	$0
kusano 2b45e8
kusano 2b45e8
#define a0	$f0
kusano 2b45e8
#define a1	$f1
kusano 2b45e8
#define a2	$f10
kusano 2b45e8
#define a3	$f11
kusano 2b45e8
#define t0	$f12
kusano 2b45e8
#define t1	$f13
kusano 2b45e8
#define t2	$f14
kusano 2b45e8
#define t3	$f15
kusano 2b45e8
kusano 2b45e8
#define x0	$f16
kusano 2b45e8
#define x1	$f17
kusano 2b45e8
#define x2	$f18
kusano 2b45e8
#define x3	$f19
kusano 2b45e8
#define x4	$f20
kusano 2b45e8
#define x5	$f21
kusano 2b45e8
#define x6	$f22
kusano 2b45e8
#define x7	$f23
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
kusano 2b45e8
#if defined(EV4) || defined(EV5)
kusano 2b45e8
        .frame $30,16,$26,0
kusano 2b45e8
        .mask 0x4000000,-16
kusano 2b45e8
        ldah	$29,    0($27)		!gpdisp!1
kusano 2b45e8
        lda	$29,    0($29)		!gpdisp!1
kusano 2b45e8
 
kusano 2b45e8
	lda	$sp,  -16($sp)
kusano 2b45e8
	ldq	$27, sqrt($29)		!literal!2
kusano 2b45e8
	stq	$26,    0($sp)
kusano 2b45e8
kusano 2b45e8
	PROFCODE
kusano 2b45e8
	.prologue 1
kusano 2b45e8
#else
kusano 2b45e8
	PROFCODE
kusano 2b45e8
#endif
kusano 2b45e8
	
kusano 2b45e8
	fclr	a0
kusano 2b45e8
	SXADDQ	INCX, 0, INCX
kusano 2b45e8
	fclr	a1
kusano 2b45e8
	ble	N, $L999
kusano 2b45e8
kusano 2b45e8
	fclr	a2
kusano 2b45e8
	cmpeq	INCX, SIZE, $0
kusano 2b45e8
	fclr	a3
kusano 2b45e8
	beq	$0,   $L20
kusano 2b45e8
kusano 2b45e8
	fclr	t0
kusano 2b45e8
	sra	N, 4, I
kusano 2b45e8
	fclr	t1
kusano 2b45e8
	ble	I, $L15
kusano 2b45e8
kusano 2b45e8
	fclr	t2
kusano 2b45e8
	LD	x0,  0 * SIZE(X)
kusano 2b45e8
	fclr	t3
kusano 2b45e8
	LD	x1,  1 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	LD	x2,  2 * SIZE(X)
kusano 2b45e8
	LD	x3,  3 * SIZE(X)
kusano 2b45e8
	LD	x4,  4 * SIZE(X)
kusano 2b45e8
	LD	x5,  5 * SIZE(X)
kusano 2b45e8
	LD	x6,  6 * SIZE(X)
kusano 2b45e8
	LD	x7,  7 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	lda	I, -1(I)
kusano 2b45e8
	ble	I, $L12
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L11:
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	ldl	$31, (PREFETCH_SIZE) * SIZE(X)
kusano 2b45e8
	mult	x0, x0, t0
kusano 2b45e8
	LD	x0,  8 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	mov	X, XX
kusano 2b45e8
	mult	x1, x1, t1
kusano 2b45e8
	LD	x1,  9 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x2, x2, t2
kusano 2b45e8
	LD	x2, 10 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x3, x3, t3
kusano 2b45e8
	LD	x3, 11 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x4, x4, t0
kusano 2b45e8
	LD	x4, 12 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x5, x5, t1
kusano 2b45e8
	LD	x5, 13 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x6, x6, t2
kusano 2b45e8
	LD	x6, 14 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x7, x7, t3
kusano 2b45e8
	LD	x7, 15 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x0, x0, t0
kusano 2b45e8
	LD	x0, 16 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	lda	X,  16 * SIZE(X)
kusano 2b45e8
	mult	x1, x1, t1
kusano 2b45e8
	LD	x1, 17 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x2, x2, t2
kusano 2b45e8
	LD	x2, 18 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x3, x3, t3
kusano 2b45e8
	LD	x3, 19 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x4, x4, t0
kusano 2b45e8
	LD	x4, 20 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	lda	I, -1(I)
kusano 2b45e8
	mult	x5, x5, t1
kusano 2b45e8
	LD	x5, 21 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x6, x6, t2
kusano 2b45e8
	LD	x6, 22 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	mult	x7, x7, t3
kusano 2b45e8
	LD	x7, 23 * SIZE(XX)
kusano 2b45e8
	bgt	I, $L11
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L12:
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	mov	X, XX
kusano 2b45e8
	mult	x0, x0, t0
kusano 2b45e8
	LD	x0,  8 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x1, x1, t1
kusano 2b45e8
	LD	x1,  9 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x2, x2, t2
kusano 2b45e8
	LD	x2, 10 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x3, x3, t3
kusano 2b45e8
	LD	x3, 11 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x4, x4, t0
kusano 2b45e8
	LD	x4, 12 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x5, x5, t1
kusano 2b45e8
	LD	x5, 13 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x6, x6, t2
kusano 2b45e8
	LD	x6, 14 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	lda	X,  16 * SIZE(X)
kusano 2b45e8
	mult	x7, x7, t3
kusano 2b45e8
	LD	x7, 15 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	mult	x0, x0, t0
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	mult	x1, x1, t1
kusano 2b45e8
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	mult	x2, x2, t2
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	mult	x3, x3, t3
kusano 2b45e8
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	mult	x4, x4, t0
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	mult	x5, x5, t1
kusano 2b45e8
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	mult	x6, x6, t2
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	mult	x7, x7, t3
kusano 2b45e8
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L15:
kusano 2b45e8
	and	N, 15, I
kusano 2b45e8
	ble	I, $L998
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L16:
kusano 2b45e8
	LD	x0,  0 * SIZE(X)
kusano 2b45e8
	lda	X,   1 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	mult	x0, x0, t0
kusano 2b45e8
kusano 2b45e8
	lda	I, -1(I)
kusano 2b45e8
	bgt	I, $L16
kusano 2b45e8
	bsr	$31, $L998
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L20:
kusano 2b45e8
	fclr	t0
kusano 2b45e8
	sra	N, 3, I
kusano 2b45e8
	fclr	t1
kusano 2b45e8
	ble	I, $L25
kusano 2b45e8
kusano 2b45e8
	fclr	t2
kusano 2b45e8
	fclr	t3
kusano 2b45e8
kusano 2b45e8
	LD	x0,  0 * SIZE(X)
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
	LD	x1,  0 * SIZE(X)
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
	LD	x2,  0 * SIZE(X)
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
	LD	x3,  0 * SIZE(X)
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	LD	x4,  0 * SIZE(X)
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
	LD	x5,  0 * SIZE(X)
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
	LD	x6,  0 * SIZE(X)
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	lda	I, -1(I)
kusano 2b45e8
	ble	I, $L22
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L21:
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	LD	x7,  0 * SIZE(X)
kusano 2b45e8
	mult	x0, x0, t0
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	LD	x0,  0 * SIZE(X)
kusano 2b45e8
	mult	x1, x1, t1
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	LD	x1,  0 * SIZE(X)
kusano 2b45e8
	mult	x2, x2, t2
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	LD	x2,  0 * SIZE(X)
kusano 2b45e8
	mult	x3, x3, t3
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	LD	x3,  0 * SIZE(X)
kusano 2b45e8
	mult	x4, x4, t0
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	LD	x4,  0 * SIZE(X)
kusano 2b45e8
	mult	x5, x5, t1
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	LD	x5,  0 * SIZE(X)
kusano 2b45e8
	mult	x6, x6, t2
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	LD	x6,  0 * SIZE(X)
kusano 2b45e8
	mult	x7, x7, t3
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	lda	I, -1(I)
kusano 2b45e8
	bgt	I, $L21
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L22:
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	LD	x7,  0 * SIZE(X)
kusano 2b45e8
	mult	x0, x0, t0
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	unop
kusano 2b45e8
	mult	x1, x1, t1
kusano 2b45e8
	unop
kusano 2b45e8
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	mult	x2, x2, t2
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	mult	x3, x3, t3
kusano 2b45e8
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	mult	x4, x4, t0
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	mult	x5, x5, t1
kusano 2b45e8
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	mult	x6, x6, t2
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	mult	x7, x7, t3
kusano 2b45e8
kusano 2b45e8
	addt	a1, t1, a1
kusano 2b45e8
	addt	a2, t2, a2
kusano 2b45e8
	addt	a3, t3, a3
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L25:
kusano 2b45e8
	and	N, 7, I
kusano 2b45e8
	ble	I, $L998
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L26:
kusano 2b45e8
	LD	x0,  0 * SIZE(X)
kusano 2b45e8
	addq	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
	mult	x0, x0, t0
kusano 2b45e8
kusano 2b45e8
	lda	I, -1(I)
kusano 2b45e8
	bgt	I, $L26
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
$L998:
kusano 2b45e8
	addt	a0, t0, a0
kusano 2b45e8
kusano 2b45e8
	addt	a0, a1, a0
kusano 2b45e8
	addt	a2, a3, a2
kusano 2b45e8
kusano 2b45e8
#if defined(EV4) || defined(EV5)
kusano 2b45e8
	addt	a0, a2, $f16
kusano 2b45e8
	jsr	$26, ($27), sqrt	!lituse_jsr!2
kusano 2b45e8
kusano 2b45e8
        ldah	$29,   0($26)		!gpdisp!3
kusano 2b45e8
        lda	$29,   0($29)		!gpdisp!3
kusano 2b45e8
#else
kusano 2b45e8
	addt	a0, a2, a0
kusano 2b45e8
	sqrtt	a0, a0
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L999:
kusano 2b45e8
#if defined(EV4) || defined(EV5)
kusano 2b45e8
	ldq	$26,   0($sp)
kusano 2b45e8
	lda	$sp,  16($sp)
kusano 2b45e8
#endif
kusano 2b45e8
	ret
kusano 2b45e8
	EPILOGUE