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
#if defined(DOUBLE) && !defined(__64BIT__)
kusano 2b45e8
#define N	%i0
kusano 2b45e8
#define X	%i5
kusano 2b45e8
#define INCX	%i1
kusano 2b45e8
#define Y	%i2
kusano 2b45e8
#define INCY	%i3
kusano 2b45e8
#define I	%i4
kusano 2b45e8
#else
kusano 2b45e8
#define N	%i0
kusano 2b45e8
#define X	%i4
kusano 2b45e8
#define INCX	%i5
kusano 2b45e8
#define Y	%i1
kusano 2b45e8
#define INCY	%i2
kusano 2b45e8
#define I	%i3
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#define XX	%l0
kusano 2b45e8
#define YY	%l1
kusano 2b45e8
kusano 2b45e8
#ifdef DOUBLE
kusano 2b45e8
#define a1	%f0
kusano 2b45e8
#define a2	%f2
kusano 2b45e8
#define a3	%f4
kusano 2b45e8
#define a4	%f6
kusano 2b45e8
#define a5	%f8
kusano 2b45e8
#define a6	%f10
kusano 2b45e8
#define a7	%f12
kusano 2b45e8
#define a8	%f14
kusano 2b45e8
#define b1	%f16
kusano 2b45e8
#define b2	%f18
kusano 2b45e8
#define b3	%f20
kusano 2b45e8
#define b4	%f22
kusano 2b45e8
#define b5	%f24
kusano 2b45e8
#define b6	%f26
kusano 2b45e8
#define b7	%f28
kusano 2b45e8
#define b8	%f30
kusano 2b45e8
#else
kusano 2b45e8
#define a1	%f0
kusano 2b45e8
#define a2	%f1
kusano 2b45e8
#define a3	%f2
kusano 2b45e8
#define a4	%f3
kusano 2b45e8
#define a5	%f4
kusano 2b45e8
#define a6	%f5
kusano 2b45e8
#define a7	%f6
kusano 2b45e8
#define a8	%f7
kusano 2b45e8
#define b1	%f8
kusano 2b45e8
#define b2	%f9
kusano 2b45e8
#define b3	%f10
kusano 2b45e8
#define b4	%f11
kusano 2b45e8
#define b5	%f12
kusano 2b45e8
#define b6	%f13
kusano 2b45e8
#define b7	%f14
kusano 2b45e8
#define b8	%f15
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef DOUBLE
kusano 2b45e8
#define PREFETCHSIZE 128
kusano 2b45e8
#else
kusano 2b45e8
#define PREFETCHSIZE 256
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	SAVESP
kusano 2b45e8
kusano 2b45e8
#ifndef __64BIT__
kusano 2b45e8
#ifdef DOUBLE
kusano 2b45e8
	ld	[%sp + STACK_START + 28], INCX
kusano 2b45e8
	ld	[%sp + STACK_START + 32], Y
kusano 2b45e8
	ld	[%sp + STACK_START + 36], INCY
kusano 2b45e8
#else
kusano 2b45e8
	ld	[%sp+  STACK_START + 28], Y
kusano 2b45e8
	ld	[%sp+  STACK_START + 32], INCY
kusano 2b45e8
#endif
kusano 2b45e8
#else
kusano 2b45e8
	ldx	[%sp+  STACK_START + 56], Y
kusano 2b45e8
	ldx	[%sp+  STACK_START + 64], INCY
kusano 2b45e8
#endif
kusano 2b45e8
	
kusano 2b45e8
	sll	INCX, BASE_SHIFT, INCX
kusano 2b45e8
	sll	INCY, BASE_SHIFT, INCY
kusano 2b45e8
kusano 2b45e8
	cmp	INCX, SIZE
kusano 2b45e8
	bne	.LL50
kusano 2b45e8
	nop
kusano 2b45e8
	cmp	INCY, SIZE
kusano 2b45e8
	bne	.LL50
kusano 2b45e8
	nop
kusano 2b45e8
kusano 2b45e8
	sra	N, 3, I
kusano 2b45e8
	cmp	I, 0
kusano 2b45e8
	ble,pn	%icc, .LL15
kusano 2b45e8
	nop
kusano 2b45e8
kusano 2b45e8
	LDF	[X +  0 * SIZE], a1
kusano 2b45e8
	LDF	[Y +  0 * SIZE], b1
kusano 2b45e8
	LDF	[X +  1 * SIZE], a2
kusano 2b45e8
	LDF	[Y +  1 * SIZE], b2
kusano 2b45e8
	LDF	[X +  2 * SIZE], a3
kusano 2b45e8
	LDF	[Y +  2 * SIZE], b3
kusano 2b45e8
	LDF	[X +  3 * SIZE], a4
kusano 2b45e8
	LDF	[Y +  3 * SIZE], b4
kusano 2b45e8
	LDF	[X +  4 * SIZE], a5
kusano 2b45e8
	LDF	[Y +  4 * SIZE], b5
kusano 2b45e8
	LDF	[X +  5 * SIZE], a6
kusano 2b45e8
	LDF	[Y +  5 * SIZE], b6
kusano 2b45e8
	LDF	[X +  6 * SIZE], a7
kusano 2b45e8
	LDF	[Y +  6 * SIZE], b7
kusano 2b45e8
	LDF	[X +  7 * SIZE], a8
kusano 2b45e8
	LDF	[Y +  7 * SIZE], b8
kusano 2b45e8
kusano 2b45e8
	deccc	I
kusano 2b45e8
	ble,pn	%icc, .LL12
kusano 2b45e8
	nop
kusano 2b45e8
kusano 2b45e8
.LL11:
kusano 2b45e8
	prefetch [X  + PREFETCHSIZE * SIZE], 0
kusano 2b45e8
	deccc	I
kusano 2b45e8
kusano 2b45e8
	STF	a1, [Y +  0 * SIZE]
kusano 2b45e8
	LDF	[X +  8 * SIZE], a1
kusano 2b45e8
	STF	b1, [X +  0 * SIZE]
kusano 2b45e8
	LDF	[Y +  8 * SIZE], b1
kusano 2b45e8
kusano 2b45e8
	STF	a2, [Y +  1 * SIZE]
kusano 2b45e8
	LDF	[X +  9 * SIZE], a2
kusano 2b45e8
	STF	b2, [X +  1 * SIZE]
kusano 2b45e8
	LDF	[Y +  9 * SIZE], b2
kusano 2b45e8
kusano 2b45e8
	STF	a3, [Y +  2 * SIZE]
kusano 2b45e8
	LDF	[X + 10 * SIZE], a3
kusano 2b45e8
	STF	b3, [X +  2 * SIZE]
kusano 2b45e8
	LDF	[Y + 10 * SIZE], b3
kusano 2b45e8
kusano 2b45e8
	STF	a4, [Y +  3 * SIZE]
kusano 2b45e8
	LDF	[X + 11 * SIZE], a4
kusano 2b45e8
	STF	b4, [X +  3 * SIZE]
kusano 2b45e8
	LDF	[Y + 11 * SIZE], b4
kusano 2b45e8
kusano 2b45e8
	prefetch [Y  + PREFETCHSIZE * SIZE], 0
kusano 2b45e8
	add	X, 8 * SIZE, X
kusano 2b45e8
kusano 2b45e8
	STF	a5, [Y +  4 * SIZE]
kusano 2b45e8
	LDF	[X +  4 * SIZE], a5
kusano 2b45e8
	STF	b5, [X -  4 * SIZE]
kusano 2b45e8
	LDF	[Y + 12 * SIZE], b5
kusano 2b45e8
kusano 2b45e8
	STF	a6, [Y +  5 * SIZE]
kusano 2b45e8
	LDF	[X +  5 * SIZE], a6
kusano 2b45e8
	STF	b6, [X -  3 * SIZE]
kusano 2b45e8
	LDF	[Y + 13 * SIZE], b6
kusano 2b45e8
kusano 2b45e8
	STF	a7, [Y +  6 * SIZE]
kusano 2b45e8
	LDF	[X +  6 * SIZE], a7
kusano 2b45e8
	STF	b7, [X -  2 * SIZE]
kusano 2b45e8
	LDF	[Y + 14 * SIZE], b7
kusano 2b45e8
kusano 2b45e8
	STF	a8, [Y +  7 * SIZE]
kusano 2b45e8
	LDF	[X +  7 * SIZE], a8
kusano 2b45e8
	STF	b8, [X -  1 * SIZE]
kusano 2b45e8
	LDF	[Y + 15 * SIZE], b8
kusano 2b45e8
kusano 2b45e8
	bg,pt	%icc, .LL11
kusano 2b45e8
	add	Y, 8 * SIZE, Y
kusano 2b45e8
kusano 2b45e8
.LL12:
kusano 2b45e8
	STF	a1, [Y +  0 * SIZE]
kusano 2b45e8
	STF	b1, [X +  0 * SIZE]
kusano 2b45e8
	STF	a2, [Y +  1 * SIZE]
kusano 2b45e8
	STF	b2, [X +  1 * SIZE]
kusano 2b45e8
	STF	a3, [Y +  2 * SIZE]
kusano 2b45e8
	STF	b3, [X +  2 * SIZE]
kusano 2b45e8
	STF	a4, [Y +  3 * SIZE]
kusano 2b45e8
	STF	b4, [X +  3 * SIZE]
kusano 2b45e8
	STF	a5, [Y +  4 * SIZE]
kusano 2b45e8
	STF	b5, [X +  4 * SIZE]
kusano 2b45e8
	STF	a6, [Y +  5 * SIZE]
kusano 2b45e8
	STF	b6, [X +  5 * SIZE]
kusano 2b45e8
	STF	a7, [Y +  6 * SIZE]
kusano 2b45e8
	STF	b7, [X +  6 * SIZE]
kusano 2b45e8
	STF	a8, [Y +  7 * SIZE]
kusano 2b45e8
	STF	b8, [X +  7 * SIZE]
kusano 2b45e8
	add	X, 8 * SIZE, X
kusano 2b45e8
	add	Y, 8 * SIZE, Y
kusano 2b45e8
kusano 2b45e8
.LL15:
kusano 2b45e8
	and	N, 7, I
kusano 2b45e8
	cmp	I,  0
kusano 2b45e8
	ble,a,pn %icc, .LL19
kusano 2b45e8
	nop
kusano 2b45e8
kusano 2b45e8
.LL16:
kusano 2b45e8
	LDF	[X +  0 * SIZE], a1
kusano 2b45e8
	add	I, -1, I
kusano 2b45e8
	LDF	[Y +  0 * SIZE], b1
kusano 2b45e8
	cmp	I, 0
kusano 2b45e8
	STF	a1, [Y +  0 * SIZE]
kusano 2b45e8
	add	Y, 1 * SIZE, Y
kusano 2b45e8
	STF	b1, [X +  0 * SIZE]
kusano 2b45e8
	bg,pt	%icc, .LL16
kusano 2b45e8
	add	X, 1 * SIZE, X
kusano 2b45e8
kusano 2b45e8
.LL19:
kusano 2b45e8
	return	%i7 + 8
kusano 2b45e8
	clr	%g0
kusano 2b45e8
kusano 2b45e8
.LL50:
kusano 2b45e8
	sra	N, 3, I
kusano 2b45e8
	mov	X, XX
kusano 2b45e8
	cmp	I, 0
kusano 2b45e8
	ble,pn	%icc, .LL55
kusano 2b45e8
	mov	Y, YY
kusano 2b45e8
kusano 2b45e8
.LL51:
kusano 2b45e8
	LDF	[X +  0 * SIZE], a1
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	LDF	[Y +  0 * SIZE], b1
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
	LDF	[X +  0 * SIZE], a2
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	LDF	[Y +  0 * SIZE], b2
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
	LDF	[X +  0 * SIZE], a3
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	LDF	[Y +  0 * SIZE], b3
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
	LDF	[X +  0 * SIZE], a4
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	LDF	[Y +  0 * SIZE], b4
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
	LDF	[X +  0 * SIZE], a5
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	LDF	[Y +  0 * SIZE], b5
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
	LDF	[X +  0 * SIZE], a6
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	LDF	[Y +  0 * SIZE], b6
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
	LDF	[X +  0 * SIZE], a7
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	LDF	[Y +  0 * SIZE], b7
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
	LDF	[X +  0 * SIZE], a8
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	LDF	[Y +  0 * SIZE], b8
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
kusano 2b45e8
	STF	a1, [YY +  0 * SIZE]
kusano 2b45e8
	add	I, -1, I
kusano 2b45e8
	add	YY, INCY, YY
kusano 2b45e8
	STF	b1, [XX +  0 * SIZE]
kusano 2b45e8
	cmp	I, 0
kusano 2b45e8
	add	XX, INCX, XX
kusano 2b45e8
	STF	a2, [YY +  0 * SIZE]
kusano 2b45e8
	add	YY, INCY, YY
kusano 2b45e8
	STF	b2, [XX +  0 * SIZE]
kusano 2b45e8
	add	XX, INCX, XX
kusano 2b45e8
	STF	a3, [YY +  0 * SIZE]
kusano 2b45e8
	add	YY, INCY, YY
kusano 2b45e8
	STF	b3, [XX +  0 * SIZE]
kusano 2b45e8
	add	XX, INCX, XX
kusano 2b45e8
	STF	a4, [YY +  0 * SIZE]
kusano 2b45e8
	add	YY, INCY, YY
kusano 2b45e8
	STF	b4, [XX +  0 * SIZE]
kusano 2b45e8
	add	XX, INCX, XX
kusano 2b45e8
	STF	a5, [YY +  0 * SIZE]
kusano 2b45e8
	add	YY, INCY, YY
kusano 2b45e8
	STF	b5, [XX +  0 * SIZE]
kusano 2b45e8
	add	XX, INCX, XX
kusano 2b45e8
	STF	a6, [YY +  0 * SIZE]
kusano 2b45e8
	add	YY, INCY, YY
kusano 2b45e8
	STF	b6, [XX +  0 * SIZE]
kusano 2b45e8
	add	XX, INCX, XX
kusano 2b45e8
	STF	a7, [YY +  0 * SIZE]
kusano 2b45e8
	add	YY, INCY, YY
kusano 2b45e8
	STF	b7, [XX +  0 * SIZE]
kusano 2b45e8
	add	XX, INCX, XX
kusano 2b45e8
	STF	a8, [YY +  0 * SIZE]
kusano 2b45e8
	add	YY, INCY, YY
kusano 2b45e8
	STF	b8, [XX +  0 * SIZE]
kusano 2b45e8
kusano 2b45e8
	bg,pt	%icc, .LL51
kusano 2b45e8
	add	XX, INCX, XX
kusano 2b45e8
kusano 2b45e8
.LL55:
kusano 2b45e8
	and	N, 7, I
kusano 2b45e8
	cmp	I,  0
kusano 2b45e8
	ble,a,pn %icc, .LL59
kusano 2b45e8
	nop
kusano 2b45e8
kusano 2b45e8
.LL56:
kusano 2b45e8
	LDF	[X +  0 * SIZE], a1
kusano 2b45e8
	LDF	[Y +  0 * SIZE], b1
kusano 2b45e8
	add	I, -1, I
kusano 2b45e8
	cmp	I, 0
kusano 2b45e8
	STF	b1, [X +  0 * SIZE]
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	STF	a1, [Y +  0 * SIZE]
kusano 2b45e8
	bg,pt	%icc, .LL56
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
kusano 2b45e8
.LL59:
kusano 2b45e8
	return	%i7 + 8
kusano 2b45e8
	clr	%o0
kusano 2b45e8
kusano 2b45e8
	EPILOGUE