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
#ifdef linux
kusano 2b45e8
#ifndef __64BIT__
kusano 2b45e8
#define N	r3
kusano 2b45e8
#define X	r6
kusano 2b45e8
#define INCX	r7	
kusano 2b45e8
#define Y	r8
kusano 2b45e8
#define INCY	r9
kusano 2b45e8
#define	YY	r5
kusano 2b45e8
#define PRE	r4
kusano 2b45e8
#else
kusano 2b45e8
#define N	r3
kusano 2b45e8
#define X	r7
kusano 2b45e8
#define INCX	r8	
kusano 2b45e8
#define Y	r9
kusano 2b45e8
#define INCY	r10
kusano 2b45e8
#define	YY	r5
kusano 2b45e8
#define PRE	r4
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(_AIX) || defined(__APPLE__)
kusano 2b45e8
#if !defined(__64BIT__) && defined(DOUBLE)
kusano 2b45e8
#define N	r3
kusano 2b45e8
#define X	r8
kusano 2b45e8
#define INCX	r9	
kusano 2b45e8
#define Y	r10
kusano 2b45e8
#define INCY	r4
kusano 2b45e8
#define	YY	r6
kusano 2b45e8
#define PRE	r5
kusano 2b45e8
#else
kusano 2b45e8
#define N	r3
kusano 2b45e8
#define X	r7
kusano 2b45e8
#define INCX	r8	
kusano 2b45e8
#define Y	r9
kusano 2b45e8
#define INCY	r10
kusano 2b45e8
#define	YY	r5
kusano 2b45e8
#define PRE	r4
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#define ALPHA	f24
kusano 2b45e8
kusano 2b45e8
#define STACKSIZE 96
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	PROFCODE
kusano 2b45e8
kusano 2b45e8
	addi	SP, SP, -STACKSIZE
kusano 2b45e8
	li	r0,   0
kusano 2b45e8
kusano 2b45e8
	stfd	f14,    0(SP)
kusano 2b45e8
	stfd	f15,    8(SP)
kusano 2b45e8
	stfd	f16,   16(SP)
kusano 2b45e8
	stfd	f17,   24(SP)
kusano 2b45e8
kusano 2b45e8
	stfd	f18,   32(SP)
kusano 2b45e8
	stfd	f19,   40(SP)
kusano 2b45e8
	stfd	f20,   48(SP)
kusano 2b45e8
	stfd	f21,   56(SP)
kusano 2b45e8
kusano 2b45e8
	stfd	f22,   64(SP)
kusano 2b45e8
	stfd	f23,   72(SP)
kusano 2b45e8
	stfd	f24,   80(SP)
kusano 2b45e8
kusano 2b45e8
#if (defined(_AIX) || defined(__APPLE__)) && !defined(__64BIT__) && defined(DOUBLE)
kusano 2b45e8
	lwz	INCY,    56 + STACKSIZE(SP)
kusano 2b45e8
#endif
kusano 2b45e8
	
kusano 2b45e8
	fmr	ALPHA, f1
kusano 2b45e8
	slwi	INCX, INCX, BASE_SHIFT
kusano 2b45e8
	slwi	INCY, INCY, BASE_SHIFT
kusano 2b45e8
	li	PRE, 2 * 16 * SIZE
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, N, 0
kusano 2b45e8
	ble-	LL(999)
kusano 2b45e8
kusano 2b45e8
	sub	X, X, INCX
kusano 2b45e8
	sub	Y, Y, INCY
kusano 2b45e8
	mr	YY, Y
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, N, 4
kusano 2b45e8
	mtspr	CTR,  r0
kusano 2b45e8
	beq-	LL(150)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
	LFDUX	f0,    X, INCX
kusano 2b45e8
	LFDUX	f1,    X, INCX
kusano 2b45e8
	LFDUX	f2,    X, INCX
kusano 2b45e8
	LFDUX	f3,    X, INCX
kusano 2b45e8
kusano 2b45e8
	LFDUX	f8,    Y, INCY
kusano 2b45e8
	LFDUX	f9,    Y, INCY
kusano 2b45e8
	LFDUX	f10,   Y, INCY
kusano 2b45e8
	LFDUX	f11,   Y, INCY
kusano 2b45e8
kusano 2b45e8
	LFDUX	f4,    X, INCX
kusano 2b45e8
	LFDUX	f5,    X, INCX
kusano 2b45e8
	LFDUX	f6,    X, INCX
kusano 2b45e8
	LFDUX	f7,    X, INCX
kusano 2b45e8
kusano 2b45e8
	LFDUX	f12,   Y, INCY
kusano 2b45e8
	LFDUX	f13,   Y, INCY
kusano 2b45e8
	LFDUX	f14,   Y, INCY
kusano 2b45e8
	LFDUX	f15,   Y, INCY
kusano 2b45e8
	bdz	LL(120)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(110):
kusano 2b45e8
	FMADD	f16,  ALPHA, f0, f8
kusano 2b45e8
	LFDUX	f0,    X, INCX
kusano 2b45e8
	LFDUX	f8,    Y, INCY
kusano 2b45e8
#ifdef PPCG4
kusano 2b45e8
	dcbt	X, PRE
kusano 2b45e8
#endif
kusano 2b45e8
	FMADD	f17,  ALPHA, f1, f9
kusano 2b45e8
	LFDUX	f1,    X, INCX
kusano 2b45e8
	LFDUX	f9,    Y, INCY
kusano 2b45e8
	FMADD	f18,  ALPHA, f2, f10
kusano 2b45e8
	LFDUX	f2,    X, INCX
kusano 2b45e8
	LFDUX	f10,   Y, INCY
kusano 2b45e8
#ifdef PPCG4
kusano 2b45e8
	dcbtst	Y, PRE
kusano 2b45e8
#endif
kusano 2b45e8
	FMADD	f19,  ALPHA, f3, f11
kusano 2b45e8
	LFDUX	f3,    X, INCX
kusano 2b45e8
	LFDUX	f11,   Y, INCY
kusano 2b45e8
kusano 2b45e8
	FMADD	f20,  ALPHA, f4, f12
kusano 2b45e8
	LFDUX	f4,    X, INCX
kusano 2b45e8
	LFDUX	f12,   Y, INCY
kusano 2b45e8
#if defined(PPCG4) && defined(DOUBLE)
kusano 2b45e8
	dcbt	X, PRE
kusano 2b45e8
#endif
kusano 2b45e8
	FMADD	f21,  ALPHA, f5, f13
kusano 2b45e8
	LFDUX	f5,    X, INCX
kusano 2b45e8
	LFDUX	f13,   Y, INCY
kusano 2b45e8
	FMADD	f22,  ALPHA, f6, f14
kusano 2b45e8
	LFDUX	f6,    X, INCX
kusano 2b45e8
 	LFDUX	f14,   Y, INCY
kusano 2b45e8
#if defined(PPCG4) && defined(DOUBLE)
kusano 2b45e8
	dcbtst	Y, PRE
kusano 2b45e8
#endif
kusano 2b45e8
	FMADD	f23,  ALPHA, f7, f15
kusano 2b45e8
	LFDUX	f7,    X, INCX
kusano 2b45e8
	LFDUX	f15,   Y, INCY
kusano 2b45e8
kusano 2b45e8
	STFDUX	f16,   YY, INCY
kusano 2b45e8
	STFDUX	f17,   YY, INCY
kusano 2b45e8
	STFDUX	f18,   YY, INCY
kusano 2b45e8
	STFDUX	f19,   YY, INCY
kusano 2b45e8
kusano 2b45e8
	FMADD	f16,  ALPHA, f0, f8
kusano 2b45e8
	LFDUX	f0,    X, INCX
kusano 2b45e8
	LFDUX	f8,    Y, INCY
kusano 2b45e8
#ifdef PPCG4
kusano 2b45e8
	dcbt	X, PRE
kusano 2b45e8
#endif
kusano 2b45e8
	FMADD	f17,  ALPHA, f1, f9
kusano 2b45e8
	LFDUX	f1,    X, INCX
kusano 2b45e8
	LFDUX	f9,    Y, INCY
kusano 2b45e8
	FMADD	f18,  ALPHA, f2, f10
kusano 2b45e8
	LFDUX	f2,    X, INCX
kusano 2b45e8
	LFDUX	f10,   Y, INCY
kusano 2b45e8
#ifdef PPCG4
kusano 2b45e8
	dcbtst	Y, PRE
kusano 2b45e8
#endif
kusano 2b45e8
	FMADD	f19,  ALPHA, f3, f11
kusano 2b45e8
	LFDUX	f3,    X, INCX
kusano 2b45e8
	LFDUX	f11,   Y, INCY
kusano 2b45e8
kusano 2b45e8
	STFDUX	f20,   YY, INCY
kusano 2b45e8
	STFDUX	f21,   YY, INCY
kusano 2b45e8
	STFDUX	f22,   YY, INCY
kusano 2b45e8
	STFDUX	f23,   YY, INCY
kusano 2b45e8
kusano 2b45e8
	FMADD	f20,  ALPHA, f4, f12
kusano 2b45e8
	LFDUX	f4,    X, INCX
kusano 2b45e8
	LFDUX	f12,   Y, INCY
kusano 2b45e8
#if defined(PPCG4) && defined(DOUBLE)
kusano 2b45e8
	dcbt	X, PRE
kusano 2b45e8
#endif
kusano 2b45e8
	FMADD	f21,  ALPHA, f5, f13
kusano 2b45e8
	LFDUX	f5,    X, INCX
kusano 2b45e8
	LFDUX	f13,   Y, INCY
kusano 2b45e8
	FMADD	f22,  ALPHA, f6, f14
kusano 2b45e8
	LFDUX	f6,    X, INCX
kusano 2b45e8
	LFDUX	f14,   Y, INCY
kusano 2b45e8
#if defined(PPCG4) && defined(DOUBLE)
kusano 2b45e8
	dcbtst	Y, PRE
kusano 2b45e8
#endif
kusano 2b45e8
	FMADD	f23,  ALPHA, f7, f15
kusano 2b45e8
	LFDUX	f7,    X, INCX
kusano 2b45e8
	LFDUX	f15,   Y, INCY
kusano 2b45e8
kusano 2b45e8
	STFDUX	f16,   YY, INCY
kusano 2b45e8
	STFDUX	f17,   YY, INCY
kusano 2b45e8
	STFDUX	f18,   YY, INCY
kusano 2b45e8
	STFDUX	f19,   YY, INCY
kusano 2b45e8
kusano 2b45e8
	STFDUX	f20,   YY, INCY
kusano 2b45e8
	STFDUX	f21,   YY, INCY
kusano 2b45e8
	STFDUX	f22,   YY, INCY
kusano 2b45e8
	STFDUX	f23,   YY, INCY
kusano 2b45e8
	bdnz	LL(110)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(120):
kusano 2b45e8
	FMADD	f16,  ALPHA, f0, f8
kusano 2b45e8
	LFDUX	f0,    X, INCX
kusano 2b45e8
	LFDUX	f8,    Y, INCY
kusano 2b45e8
	FMADD	f17,  ALPHA, f1, f9
kusano 2b45e8
	LFDUX	f1,    X, INCX
kusano 2b45e8
	LFDUX	f9,    Y, INCY
kusano 2b45e8
	FMADD	f18,  ALPHA, f2, f10
kusano 2b45e8
	LFDUX	f2,    X, INCX
kusano 2b45e8
	LFDUX	f10,   Y, INCY
kusano 2b45e8
	FMADD	f19,  ALPHA, f3, f11
kusano 2b45e8
	LFDUX	f3,    X, INCX
kusano 2b45e8
	LFDUX	f11,   Y, INCY
kusano 2b45e8
kusano 2b45e8
	FMADD	f20,  ALPHA, f4, f12
kusano 2b45e8
	LFDUX	f4,    X, INCX
kusano 2b45e8
	LFDUX	f12,   Y, INCY
kusano 2b45e8
	FMADD	f21,  ALPHA, f5, f13
kusano 2b45e8
	LFDUX	f5,    X, INCX
kusano 2b45e8
	LFDUX	f13,   Y, INCY
kusano 2b45e8
	FMADD	f22,  ALPHA, f6, f14
kusano 2b45e8
	LFDUX	f6,    X, INCX
kusano 2b45e8
	LFDUX	f14,   Y, INCY
kusano 2b45e8
	FMADD	f23,  ALPHA, f7, f15
kusano 2b45e8
	LFDUX	f7,    X, INCX
kusano 2b45e8
	LFDUX	f15,   Y, INCY
kusano 2b45e8
kusano 2b45e8
	STFDUX	f16,   YY, INCY
kusano 2b45e8
	STFDUX	f17,   YY, INCY
kusano 2b45e8
	STFDUX	f18,   YY, INCY
kusano 2b45e8
	STFDUX	f19,   YY, INCY
kusano 2b45e8
kusano 2b45e8
	FMADD	f16,  ALPHA, f0, f8
kusano 2b45e8
	STFDUX	f20,   YY, INCY
kusano 2b45e8
	FMADD	f17,  ALPHA, f1, f9
kusano 2b45e8
	STFDUX	f21,   YY, INCY
kusano 2b45e8
	FMADD	f18,  ALPHA, f2, f10
kusano 2b45e8
	STFDUX	f22,   YY, INCY
kusano 2b45e8
	FMADD	f19,  ALPHA, f3, f11
kusano 2b45e8
	STFDUX	f23,   YY, INCY
kusano 2b45e8
kusano 2b45e8
	FMADD	f20,  ALPHA, f4, f12
kusano 2b45e8
	STFDUX	f16,   YY, INCY
kusano 2b45e8
	FMADD	f21,  ALPHA, f5, f13
kusano 2b45e8
	STFDUX	f17,   YY, INCY
kusano 2b45e8
	FMADD	f22,  ALPHA, f6, f14
kusano 2b45e8
	STFDUX	f18,   YY, INCY
kusano 2b45e8
	FMADD	f23,  ALPHA, f7, f15
kusano 2b45e8
	STFDUX	f19,   YY, INCY
kusano 2b45e8
kusano 2b45e8
	STFDUX	f20,   YY, INCY
kusano 2b45e8
	STFDUX	f21,   YY, INCY
kusano 2b45e8
	STFDUX	f22,   YY, INCY
kusano 2b45e8
	STFDUX	f23,   YY, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(150):
kusano 2b45e8
	andi.	r0,  N, 15
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	beq	LL(999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(160):
kusano 2b45e8
	LFDUX	f0,   X, INCX
kusano 2b45e8
	LFDUX	f8,   Y, INCY
kusano 2b45e8
kusano 2b45e8
	FMADD	f16,  ALPHA, f0, f8
kusano 2b45e8
kusano 2b45e8
	STFDUX	f16,  YY, INCY
kusano 2b45e8
	bdnz	LL(160)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(999):
kusano 2b45e8
	lfd	f14,    0(SP)
kusano 2b45e8
	lfd	f15,    8(SP)
kusano 2b45e8
	lfd	f16,   16(SP)
kusano 2b45e8
	lfd	f17,   24(SP)
kusano 2b45e8
kusano 2b45e8
	lfd	f18,   32(SP)
kusano 2b45e8
	lfd	f19,   40(SP)
kusano 2b45e8
	lfd	f20,   48(SP)
kusano 2b45e8
	lfd	f21,   56(SP)
kusano 2b45e8
kusano 2b45e8
	lfd	f22,   64(SP)
kusano 2b45e8
	lfd	f23,   72(SP)
kusano 2b45e8
	lfd	f24,   80(SP)
kusano 2b45e8
	addi	SP, SP, STACKSIZE
kusano 2b45e8
	blr
kusano 2b45e8
kusano 2b45e8
	EPILOGUE