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	%i0
kusano 2b45e8
#define X	%i1
kusano 2b45e8
#define INCX	%i2
kusano 2b45e8
#define Y	%i3
kusano 2b45e8
#define INCY	%i4
kusano 2b45e8
#define I	%i5
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
#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
#endif
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	SAVESP
kusano 2b45e8
kusano 2b45e8
	sll	INCX, ZBASE_SHIFT, INCX
kusano 2b45e8
	sll	INCY, ZBASE_SHIFT, INCY
kusano 2b45e8
kusano 2b45e8
	cmp	INCX, 2 * SIZE
kusano 2b45e8
	bne	.LL50
kusano 2b45e8
	nop
kusano 2b45e8
	cmp	INCY, 2 * SIZE
kusano 2b45e8
	bne	.LL50
kusano 2b45e8
	nop
kusano 2b45e8
kusano 2b45e8
	sra	N, 2, I
kusano 2b45e8
	cmp	I, 0
kusano 2b45e8
	ble,pn	%icc, .LL15
kusano 2b45e8
	nop
kusano 2b45e8
kusano 2b45e8
#define PREFETCHSIZE 32
kusano 2b45e8
kusano 2b45e8
.LL11:
kusano 2b45e8
	prefetch [X  + PREFETCHSIZE * SIZE], 0
kusano 2b45e8
	prefetch [Y  + PREFETCHSIZE * SIZE], 0
kusano 2b45e8
kusano 2b45e8
	LDF	[X +  0 * SIZE], a1
kusano 2b45e8
	LDF	[X +  1 * SIZE], a2
kusano 2b45e8
	LDF	[X +  2 * SIZE], a3
kusano 2b45e8
	LDF	[X +  3 * SIZE], a4
kusano 2b45e8
	LDF	[X +  4 * SIZE], a5
kusano 2b45e8
	LDF	[X +  5 * SIZE], a6
kusano 2b45e8
	LDF	[X +  6 * SIZE], a7
kusano 2b45e8
	LDF	[X +  7 * SIZE], a8
kusano 2b45e8
kusano 2b45e8
	STF	a1, [Y +  0 * SIZE]
kusano 2b45e8
	add	I, -1, I
kusano 2b45e8
	STF	a2, [Y +  1 * SIZE]
kusano 2b45e8
	cmp	I, 0
kusano 2b45e8
	STF	a3, [Y +  2 * SIZE]
kusano 2b45e8
	add	X, 8 * SIZE, X
kusano 2b45e8
	STF	a4, [Y +  3 * SIZE]
kusano 2b45e8
	STF	a5, [Y +  4 * SIZE]
kusano 2b45e8
	STF	a6, [Y +  5 * SIZE]
kusano 2b45e8
	STF	a7, [Y +  6 * SIZE]
kusano 2b45e8
	STF	a8, [Y +  7 * SIZE]
kusano 2b45e8
kusano 2b45e8
	bg,pt	%icc, .LL11
kusano 2b45e8
	add	Y, 8 * SIZE, Y
kusano 2b45e8
kusano 2b45e8
.LL15:
kusano 2b45e8
	and	N, 3, 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
	LDF	[X +  1 * SIZE], a2
kusano 2b45e8
	add	I, -1, I
kusano 2b45e8
	cmp	I, 0
kusano 2b45e8
	STF	a1, [Y +  0 * SIZE]
kusano 2b45e8
	add	X, 2 * SIZE, X
kusano 2b45e8
	STF	a2, [Y +  1 * SIZE]
kusano 2b45e8
	bg,pt	%icc, .LL16
kusano 2b45e8
	add	Y, 2 * SIZE, Y
kusano 2b45e8
kusano 2b45e8
.LL19:
kusano 2b45e8
	return	%i7 + 8
kusano 2b45e8
	clr	%g0
kusano 2b45e8
kusano 2b45e8
.LL50:
kusano 2b45e8
	sra	N, 2, I
kusano 2b45e8
	cmp	I, 0
kusano 2b45e8
	ble,pn	%icc, .LL55
kusano 2b45e8
	nop
kusano 2b45e8
kusano 2b45e8
.LL51:
kusano 2b45e8
	LDF	[X +  0 * SIZE], a1
kusano 2b45e8
	LDF	[X +  1 * SIZE], a2
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	LDF	[X +  0 * SIZE], a3
kusano 2b45e8
	LDF	[X +  1 * SIZE], a4
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	LDF	[X +  0 * SIZE], a5
kusano 2b45e8
	LDF	[X +  1 * SIZE], a6
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	LDF	[X +  0 * SIZE], a7
kusano 2b45e8
	LDF	[X +  1 * SIZE], a8
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
kusano 2b45e8
	STF	a1, [Y +  0 * SIZE]
kusano 2b45e8
	add	I, -1, I
kusano 2b45e8
	STF	a2, [Y +  1 * SIZE]
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
	cmp	I, 0
kusano 2b45e8
	STF	a3, [Y +  0 * SIZE]
kusano 2b45e8
	STF	a4, [Y +  1 * SIZE]
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
	STF	a5, [Y +  0 * SIZE]
kusano 2b45e8
	STF	a6, [Y +  1 * SIZE]
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
	STF	a7, [Y +  0 * SIZE]
kusano 2b45e8
	STF	a8, [Y +  1 * SIZE]
kusano 2b45e8
kusano 2b45e8
	bg,pt	%icc, .LL51
kusano 2b45e8
	add	Y, INCY, Y
kusano 2b45e8
kusano 2b45e8
.LL55:
kusano 2b45e8
	and	N, 3, 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	[X +  1 * SIZE], a2
kusano 2b45e8
	add	I, -1, I
kusano 2b45e8
	cmp	I, 0
kusano 2b45e8
	add	X, INCX, X
kusano 2b45e8
	STF	a1, [Y +  0 * SIZE]
kusano 2b45e8
	STF	a2, [Y +  1 * 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