Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/power/symv_U.S

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 M	r3
kusano 2b45e8
#define IS	r4
kusano 2b45e8
#define A	r5
kusano 2b45e8
#define LDA	r6
kusano 2b45e8
#define X	r7
kusano 2b45e8
#define	INCX	r8
kusano 2b45e8
#define	Y	r9
kusano 2b45e8
#define	INCY	r10
kusano 2b45e8
#define BUFFER	r14
kusano 2b45e8
#else
kusano 2b45e8
#define M	r3
kusano 2b45e8
#define IS	r4
kusano 2b45e8
#define A	r6
kusano 2b45e8
#define LDA	r7
kusano 2b45e8
#define X	r8
kusano 2b45e8
#define	INCX	r9
kusano 2b45e8
#define	Y	r10
kusano 2b45e8
#define	INCY	r5
kusano 2b45e8
#define BUFFER	r14
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 M	r3
kusano 2b45e8
#define IS	r4
kusano 2b45e8
#define A	r7
kusano 2b45e8
#define LDA	r8
kusano 2b45e8
#define X	r9
kusano 2b45e8
#define	INCX	r10
kusano 2b45e8
#define	Y	r5
kusano 2b45e8
#define	INCY	r6
kusano 2b45e8
#define BUFFER	r14
kusano 2b45e8
#else
kusano 2b45e8
#define M	r3
kusano 2b45e8
#define IS	r4
kusano 2b45e8
#define A	r6
kusano 2b45e8
#define LDA	r7
kusano 2b45e8
#define X	r8
kusano 2b45e8
#define	INCX	r9
kusano 2b45e8
#define	Y	r10
kusano 2b45e8
#define	INCY	r5
kusano 2b45e8
#define BUFFER	r14
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#define I	r11
kusano 2b45e8
#define	J	r12
kusano 2b45e8
kusano 2b45e8
#define AO1	r15
kusano 2b45e8
#define AO2	r16
kusano 2b45e8
#define AO3	r17
kusano 2b45e8
#define AO4	r18
kusano 2b45e8
#define XX	r19
kusano 2b45e8
#define YY	r20
kusano 2b45e8
#define	NEW_Y	r21
kusano 2b45e8
#define TEMP	r22
kusano 2b45e8
#define	PREA	r24
kusano 2b45e8
kusano 2b45e8
#define y01 f0
kusano 2b45e8
#define y02 f1
kusano 2b45e8
#define y03 f2
kusano 2b45e8
#define y04 f3
kusano 2b45e8
kusano 2b45e8
#define atemp1 f4
kusano 2b45e8
#define atemp2 f5
kusano 2b45e8
#define atemp3 f6
kusano 2b45e8
#define atemp4 f7
kusano 2b45e8
kusano 2b45e8
#define xtemp1 f8
kusano 2b45e8
#define xtemp2 f9
kusano 2b45e8
#define xtemp3 f10
kusano 2b45e8
#define xtemp4 f11
kusano 2b45e8
kusano 2b45e8
#define xsum1  f12
kusano 2b45e8
#define xsum2  f13
kusano 2b45e8
#define xsum3  f14
kusano 2b45e8
#define xsum4  f15
kusano 2b45e8
kusano 2b45e8
#define a1     f16
kusano 2b45e8
#define a2     f17
kusano 2b45e8
#define a3     f18
kusano 2b45e8
#define a4     f19
kusano 2b45e8
#define a5     f20
kusano 2b45e8
#define a6     f21
kusano 2b45e8
#define a7     f22
kusano 2b45e8
#define a8     f23
kusano 2b45e8
#define a9     f24
kusano 2b45e8
#define a10    f25
kusano 2b45e8
#define a11    f26
kusano 2b45e8
#define a12    f27
kusano 2b45e8
#define a13    f28
kusano 2b45e8
#define a14    f29
kusano 2b45e8
#define a15    f30
kusano 2b45e8
#define a16    f31
kusano 2b45e8
kusano 2b45e8
#define alpha  f1
kusano 2b45e8
kusano 2b45e8
#if defined(PPCG4)
kusano 2b45e8
#define PREFETCHSIZE_A  24
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(PPC440) || defined(PPC440FP2)
kusano 2b45e8
#define PREFETCHSIZE_A  24
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef PPC970
kusano 2b45e8
#define PREFETCHSIZE_A  64
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef CELL
kusano 2b45e8
#define PREFETCHSIZE_A  72
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef POWER4
kusano 2b45e8
#define PREFETCHSIZE_A  16
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef POWER5
kusano 2b45e8
#define PREFETCHSIZE_A  96
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef POWER6
kusano 2b45e8
#define PREFETCHSIZE_A  40
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(POWER4) || defined(POWER5) || defined(POWER6) || defined(PPC970)
kusano 2b45e8
#define NOP1
kusano 2b45e8
#define NOP2
kusano 2b45e8
#else
kusano 2b45e8
#define NOP1   mr	LDA, LDA
kusano 2b45e8
#define NOP2   mr	INCX, INCX
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifndef NEEDPARAM
kusano 2b45e8
kusano 2b45e8
#ifndef __64BIT__
kusano 2b45e8
#define STACKSIZE 224
kusano 2b45e8
#define ALPHA     200(SP)
kusano 2b45e8
#define	FZERO	  208(SP)
kusano 2b45e8
#else
kusano 2b45e8
#define STACKSIZE 280
kusano 2b45e8
#define ALPHA     256(SP)
kusano 2b45e8
#define FZERO	  264(SP)
kusano 2b45e8
#endif
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
	stfd	f18,    32(SP)
kusano 2b45e8
	stfd	f19,    40(SP)
kusano 2b45e8
	stfd	f20,    48(SP)
kusano 2b45e8
	stfd	f21,    56(SP)
kusano 2b45e8
	stfd	f22,    64(SP)
kusano 2b45e8
	stfd	f23,    72(SP)
kusano 2b45e8
	stfd	f24,    80(SP)
kusano 2b45e8
	stfd	f25,    88(SP)
kusano 2b45e8
	stfd	f26,    96(SP)
kusano 2b45e8
	stfd	f27,   104(SP)
kusano 2b45e8
	stfd	f28,   112(SP)
kusano 2b45e8
	stfd	f29,   120(SP)
kusano 2b45e8
	stfd	f30,   128(SP)
kusano 2b45e8
	stfd	f31,   136(SP)
kusano 2b45e8
kusano 2b45e8
#ifdef __64BIT__
kusano 2b45e8
	std	r0,    FZERO
kusano 2b45e8
	std	r14,   144(SP)
kusano 2b45e8
	std	r15,   152(SP)
kusano 2b45e8
	std	r16,   160(SP)
kusano 2b45e8
	std	r17,   168(SP)
kusano 2b45e8
	std	r18,   176(SP)
kusano 2b45e8
	std	r19,   184(SP)
kusano 2b45e8
	std	r20,   192(SP)
kusano 2b45e8
	std	r21,   200(SP)
kusano 2b45e8
	std	r22,   208(SP)
kusano 2b45e8
	std	r23,   216(SP)
kusano 2b45e8
	std	r24,   224(SP)
kusano 2b45e8
	std	r25,   232(SP)
kusano 2b45e8
	std	r26,   240(SP)
kusano 2b45e8
	std	r27,   248(SP)
kusano 2b45e8
#else
kusano 2b45e8
	stw	r0,    0 + FZERO
kusano 2b45e8
	stw	r0,    4 + FZERO
kusano 2b45e8
	stw	r14,   144(SP)
kusano 2b45e8
	stw	r15,   148(SP)
kusano 2b45e8
	stw	r16,   152(SP)
kusano 2b45e8
	stw	r17,   156(SP)
kusano 2b45e8
	stw	r18,   160(SP)
kusano 2b45e8
	stw	r19,   164(SP)
kusano 2b45e8
	stw	r20,   168(SP)
kusano 2b45e8
	stw	r21,   172(SP)
kusano 2b45e8
	stw	r22,   176(SP)
kusano 2b45e8
	stw	r23,   180(SP)
kusano 2b45e8
	stw	r24,   184(SP)
kusano 2b45e8
	stw	r25,   188(SP)
kusano 2b45e8
	stw	r26,   192(SP)
kusano 2b45e8
	stw	r27,   196(SP)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef linux
kusano 2b45e8
#ifndef __64BIT__
kusano 2b45e8
	lwz	BUFFER,   56 + STACKSIZE(SP)
kusano 2b45e8
#else
kusano 2b45e8
	ld	INCY,    112 + STACKSIZE(SP)
kusano 2b45e8
	ld	BUFFER,  120 + STACKSIZE(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(_AIX) || defined(__APPLE__)
kusano 2b45e8
#ifndef __64BIT__
kusano 2b45e8
#ifdef DOUBLE
kusano 2b45e8
	lwz	Y,       56 + STACKSIZE(SP)
kusano 2b45e8
	lwz	INCY,    60 + STACKSIZE(SP)
kusano 2b45e8
	lwz	BUFFER,  64 + STACKSIZE(SP)
kusano 2b45e8
#else
kusano 2b45e8
	lwz	  INCY,  56 + STACKSIZE(SP)
kusano 2b45e8
	lwz	BUFFER,  60 + STACKSIZE(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#else
kusano 2b45e8
	ld	INCY,    112 + STACKSIZE(SP)
kusano 2b45e8
	ld	BUFFER,  120 + STACKSIZE(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	STFD	alpha, ALPHA
kusano 2b45e8
kusano 2b45e8
	slwi	LDA,  LDA,  BASE_SHIFT
kusano 2b45e8
	slwi	INCX, INCX, BASE_SHIFT
kusano 2b45e8
	slwi	INCY, INCY, BASE_SHIFT
kusano 2b45e8
kusano 2b45e8
	li	PREA, PREFETCHSIZE_A * SIZE
kusano 2b45e8
	sub	IS, M, IS
kusano 2b45e8
	
kusano 2b45e8
	cmpwi	cr0, M, 0
kusano 2b45e8
	ble-	LL(999)
kusano 2b45e8
kusano 2b45e8
	mullw	TEMP, IS, LDA
kusano 2b45e8
	add	A, A, TEMP
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, INCX, SIZE
kusano 2b45e8
	beq	LL(05)
kusano 2b45e8
kusano 2b45e8
	mr	XX, X
kusano 2b45e8
	mr	X, BUFFER
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, M, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(03)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(01):
kusano 2b45e8
	LFD	a1, 0 * SIZE(XX)
kusano 2b45e8
	add	XX, XX, INCX
kusano 2b45e8
	LFD	a2, 0 * SIZE(XX)
kusano 2b45e8
	add	XX, XX, INCX
kusano 2b45e8
	LFD	a3, 0 * SIZE(XX)
kusano 2b45e8
	add	XX, XX, INCX
kusano 2b45e8
	LFD	a4, 0 * SIZE(XX)
kusano 2b45e8
	add	XX, XX, INCX
kusano 2b45e8
	LFD	a5, 0 * SIZE(XX)
kusano 2b45e8
	add	XX, XX, INCX
kusano 2b45e8
	LFD	a6, 0 * SIZE(XX)
kusano 2b45e8
	add	XX, XX, INCX
kusano 2b45e8
	LFD	a7, 0 * SIZE(XX)
kusano 2b45e8
	add	XX, XX, INCX
kusano 2b45e8
	LFD	a8, 0 * SIZE(XX)
kusano 2b45e8
	add	XX, XX, INCX
kusano 2b45e8
kusano 2b45e8
	dcbt	XX, PREA
kusano 2b45e8
	dcbtst	BUFFER, PREA
kusano 2b45e8
kusano 2b45e8
	STFD	a1, 0 * SIZE(BUFFER)
kusano 2b45e8
	STFD	a2, 1 * SIZE(BUFFER)
kusano 2b45e8
	STFD	a3, 2 * SIZE(BUFFER)
kusano 2b45e8
	STFD	a4, 3 * SIZE(BUFFER)
kusano 2b45e8
	STFD	a5, 4 * SIZE(BUFFER)
kusano 2b45e8
	STFD	a6, 5 * SIZE(BUFFER)
kusano 2b45e8
	STFD	a7, 6 * SIZE(BUFFER)
kusano 2b45e8
	STFD	a8, 7 * SIZE(BUFFER)
kusano 2b45e8
kusano 2b45e8
	addi	BUFFER, BUFFER, 8 * SIZE
kusano 2b45e8
	bdnz	LL(01)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(03):
kusano 2b45e8
	andi.	r0, M, 7
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(05)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(04):
kusano 2b45e8
	LFD	a1, 0 * SIZE(XX)
kusano 2b45e8
	add	XX, XX, INCX
kusano 2b45e8
kusano 2b45e8
	STFD	a1, 0 * SIZE(BUFFER)
kusano 2b45e8
	addi	BUFFER, BUFFER, 1 * SIZE
kusano 2b45e8
	bdnz	LL(04)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(05):
kusano 2b45e8
	mr	NEW_Y, Y
kusano 2b45e8
	lfd	f0, FZERO
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, INCY, SIZE
kusano 2b45e8
	beq	LL(10)
kusano 2b45e8
kusano 2b45e8
	mr	NEW_Y, BUFFER
kusano 2b45e8
kusano 2b45e8
	addi	r0, M,  7
kusano 2b45e8
	srawi.	r0, r0, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(06):
kusano 2b45e8
	STFD	f0, 0 * SIZE(BUFFER)
kusano 2b45e8
	STFD	f0, 1 * SIZE(BUFFER)
kusano 2b45e8
	STFD	f0, 2 * SIZE(BUFFER)
kusano 2b45e8
	STFD	f0, 3 * SIZE(BUFFER)
kusano 2b45e8
	STFD	f0, 4 * SIZE(BUFFER)
kusano 2b45e8
	STFD	f0, 5 * SIZE(BUFFER)
kusano 2b45e8
	STFD	f0, 6 * SIZE(BUFFER)
kusano 2b45e8
	STFD	f0, 7 * SIZE(BUFFER)
kusano 2b45e8
	addi	BUFFER, BUFFER, 8 * SIZE
kusano 2b45e8
	bdnz	LL(06)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(10):
kusano 2b45e8
	addi	TEMP, IS, 4
kusano 2b45e8
	cmpw	cr0, TEMP, M
kusano 2b45e8
	bgt	LL(20)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(11):
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	add	AO2, A,   LDA
kusano 2b45e8
	add	AO3, AO2, LDA
kusano 2b45e8
	add	AO4, AO3, LDA
kusano 2b45e8
	add	A,   AO4, LDA
kusano 2b45e8
kusano 2b45e8
	slwi	TEMP,  IS,  BASE_SHIFT
kusano 2b45e8
	add	TEMP, X, TEMP
kusano 2b45e8
kusano 2b45e8
	LFD	a16, ALPHA 
kusano 2b45e8
	lfd	xsum1, FZERO
kusano 2b45e8
kusano 2b45e8
	LFD	atemp1, 0 * SIZE(TEMP)
kusano 2b45e8
	LFD	atemp2, 1 * SIZE(TEMP)
kusano 2b45e8
	LFD	atemp3, 2 * SIZE(TEMP)
kusano 2b45e8
	LFD	atemp4, 3 * SIZE(TEMP)
kusano 2b45e8
kusano 2b45e8
	LFD	xtemp1,  0 * SIZE(X)
kusano 2b45e8
	LFD	xtemp2,  1 * SIZE(X)
kusano 2b45e8
	LFD	xtemp3,  2 * SIZE(X)
kusano 2b45e8
	LFD	xtemp4,  3 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	LFD	y01,  0 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	y02,  1 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	y03,  2 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	y04,  3 * SIZE(NEW_Y)
kusano 2b45e8
kusano 2b45e8
	LFD	a1,  0 * SIZE(AO1)
kusano 2b45e8
	FMUL	atemp1, a16, atemp1
kusano 2b45e8
	LFD	a2,  1 * SIZE(AO1)
kusano 2b45e8
	FMUL	atemp2, a16, atemp2
kusano 2b45e8
	LFD	a3,  2 * SIZE(AO1)
kusano 2b45e8
	FMUL	atemp3, a16, atemp3
kusano 2b45e8
	LFD	a4,  3 * SIZE(AO1)
kusano 2b45e8
	FMUL	atemp4, a16, atemp4
kusano 2b45e8
kusano 2b45e8
	LFD	a5,  0 * SIZE(AO2)
kusano 2b45e8
	fmr	xsum2, xsum1
kusano 2b45e8
	LFD	a6,  1 * SIZE(AO2)
kusano 2b45e8
	fmr	xsum3, xsum1
kusano 2b45e8
	LFD	a7,  2 * SIZE(AO2)
kusano 2b45e8
	fmr	xsum4, xsum1
kusano 2b45e8
	LFD	a8,  3 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	LFD	a9,  0 * SIZE(AO3)
kusano 2b45e8
	LFD	a10, 1 * SIZE(AO3)
kusano 2b45e8
	LFD	a11, 2 * SIZE(AO3)
kusano 2b45e8
	LFD	a12, 3 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	LFD	a13, 0 * SIZE(AO4)
kusano 2b45e8
	LFD	a14, 1 * SIZE(AO4)
kusano 2b45e8
	LFD	a15, 2 * SIZE(AO4)
kusano 2b45e8
	LFD	a16, 3 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	mr	XX, X
kusano 2b45e8
	mr	YY, NEW_Y
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  IS, 4
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(14)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(12):
kusano 2b45e8
	FMADD	xsum1, xtemp1, a1,  xsum1
kusano 2b45e8
	DCBT(AO1, PREA)
kusano 2b45e8
	FMADD	y01, atemp1, a1,  y01
kusano 2b45e8
	LFD	a1,  4 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp1, a5,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp1, a2,  y02
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp1, a9,  xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp1, a3,  y03
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp1, a13, xsum4
kusano 2b45e8
	LFD	xtemp1,  4 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp1, a4,  y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp2, a2,  xsum1
kusano 2b45e8
	LFD	a2,  5 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp2, a5,  y01
kusano 2b45e8
	LFD	a5,  4 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp2, a6,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp2, a6,  y02
kusano 2b45e8
	LFD	a6,  5 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp2, a10, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp2, a7,  y03
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp2, a14, xsum4
kusano 2b45e8
	LFD	xtemp2,  5 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp2, a8,  y04
kusano 2b45e8
#	DCBT(X, PREX)
kusano 2b45e8
	NOP2
kusano 2b45e8
	
kusano 2b45e8
	FMADD	xsum1, xtemp3, a3,  xsum1
kusano 2b45e8
	LFD	a3,  6 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp3, a9,  y01
kusano 2b45e8
	LFD	a9,  4 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp3, a7,  xsum2
kusano 2b45e8
	LFD	a7,  6 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp3, a10, y02
kusano 2b45e8
	LFD	a10, 5 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp3, a11, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp3, a11, y03
kusano 2b45e8
	LFD	a11, 6 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp3, a15, xsum4
kusano 2b45e8
	LFD	xtemp3,  6 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp3, a12, y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp4, a4,  xsum1
kusano 2b45e8
	LFD	a4,  7 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp4, a13, y01
kusano 2b45e8
	LFD	a13, 4 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp4, a8,  xsum2
kusano 2b45e8
	LFD	a8,  7 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp4, a14, y02
kusano 2b45e8
	LFD	a14, 5 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp4, a12, xsum3
kusano 2b45e8
	LFD	a12, 7 * SIZE(AO3)
kusano 2b45e8
	FMADD	y03, atemp4, a15, y03
kusano 2b45e8
	LFD	a15, 6 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp4, a16, xsum4
kusano 2b45e8
	LFD	xtemp4,  7 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp4, a16, y04
kusano 2b45e8
	LFD	a16, 7 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	STFD	y01,  0 * SIZE(YY)
kusano 2b45e8
	LFD	y01,  4 * SIZE(YY)
kusano 2b45e8
	STFD	y02,  1 * SIZE(YY)
kusano 2b45e8
	LFD	y02,  5 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	STFD	y03,  2 * SIZE(YY)
kusano 2b45e8
	LFD	y03,  6 * SIZE(YY)
kusano 2b45e8
	STFD	y04,  3 * SIZE(YY)
kusano 2b45e8
	LFD	y04,  7 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp1, a1,  xsum1
kusano 2b45e8
	DCBT(AO2, PREA)
kusano 2b45e8
	FMADD	y01, atemp1, a1,  y01
kusano 2b45e8
	LFD	a1,  8 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp1, a5,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp1, a2,  y02
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp1, a9,  xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp1, a3,  y03
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp1, a13, xsum4
kusano 2b45e8
	LFD	xtemp1,  8 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp1, a4,  y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp2, a2,  xsum1
kusano 2b45e8
	LFD	a2,  9 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp2, a5,  y01
kusano 2b45e8
	LFD	a5,  8 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp2, a6,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp2, a6,  y02
kusano 2b45e8
	LFD	a6,  9 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp2, a10, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp2, a7,  y03
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp2, a14, xsum4
kusano 2b45e8
	LFD	xtemp2,  9 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp2, a8,  y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp3, a3,  xsum1
kusano 2b45e8
	LFD	a3, 10 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp3, a9,  y01
kusano 2b45e8
	LFD	a9,  8 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp3, a7,  xsum2
kusano 2b45e8
	LFD	a7, 10 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp3, a10, y02
kusano 2b45e8
	LFD	a10, 9 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp3, a11, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp3, a11, y03
kusano 2b45e8
	LFD	a11, 10 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp3, a15, xsum4
kusano 2b45e8
	LFD	xtemp3, 10 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp3, a12, y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp4, a4,  xsum1
kusano 2b45e8
	LFD	a4, 11 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp4, a13, y01
kusano 2b45e8
	LFD	a13, 8 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp4, a8,  xsum2
kusano 2b45e8
	LFD	a8, 11 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp4, a14, y02
kusano 2b45e8
	LFD	a14, 9 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp4, a12, xsum3
kusano 2b45e8
	LFD	a12, 11 * SIZE(AO3)
kusano 2b45e8
	FMADD	y03, atemp4, a15, y03
kusano 2b45e8
	LFD	a15, 10 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp4, a16, xsum4
kusano 2b45e8
	LFD	xtemp4, 11 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp4, a16, y04
kusano 2b45e8
	LFD	a16, 11 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	STFD	y01,  4 * SIZE(YY)
kusano 2b45e8
	LFD	y01,  8 * SIZE(YY)
kusano 2b45e8
	STFD	y02,  5 * SIZE(YY)
kusano 2b45e8
	LFD	y02,  9 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	STFD	y03,  6 * SIZE(YY)
kusano 2b45e8
	LFD	y03, 10 * SIZE(YY)
kusano 2b45e8
	STFD	y04,  7 * SIZE(YY)
kusano 2b45e8
	LFD	y04, 11 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp1, a1,  xsum1
kusano 2b45e8
	DCBT(AO3, PREA)
kusano 2b45e8
	FMADD	y01, atemp1, a1,  y01
kusano 2b45e8
	LFD	a1, 12 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp1, a5,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp1, a2,  y02
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp1, a9,  xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp1, a3,  y03
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp1, a13, xsum4
kusano 2b45e8
	LFD	xtemp1, 12 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp1, a4,  y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp2, a2,  xsum1
kusano 2b45e8
	LFD	a2, 13 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp2, a5,  y01
kusano 2b45e8
	LFD	a5, 12 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp2, a6,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp2, a6,  y02
kusano 2b45e8
	LFD	a6, 13 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp2, a10, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp2, a7,  y03
kusano 2b45e8
#	DCBT(Y1, PREY)
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp2, a14, xsum4
kusano 2b45e8
	LFD	xtemp2, 13 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp2, a8,  y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp3, a3,  xsum1
kusano 2b45e8
	LFD	a3, 14 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp3, a9,  y01
kusano 2b45e8
	LFD	a9, 12 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp3, a7,  xsum2
kusano 2b45e8
	LFD	a7, 14 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp3, a10, y02
kusano 2b45e8
	LFD	a10,13 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp3, a11, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp3, a11, y03
kusano 2b45e8
	LFD	a11, 14 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp3, a15, xsum4
kusano 2b45e8
	LFD	xtemp3, 14 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp3, a12, y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp4, a4,  xsum1
kusano 2b45e8
	LFD	a4, 15 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp4, a13, y01
kusano 2b45e8
	LFD	a13,12 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp4, a8,  xsum2
kusano 2b45e8
	LFD	a8, 15 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp4, a14, y02
kusano 2b45e8
	LFD	a14, 13 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp4, a12, xsum3
kusano 2b45e8
	LFD	a12, 15 * SIZE(AO3)
kusano 2b45e8
	FMADD	y03, atemp4, a15, y03
kusano 2b45e8
	LFD	a15, 14 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp4, a16, xsum4
kusano 2b45e8
	LFD	xtemp4,  15 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp4, a16, y04
kusano 2b45e8
	LFD	a16, 15 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	STFD	y01,  8 * SIZE(YY)
kusano 2b45e8
	LFD	y01, 12 * SIZE(YY)
kusano 2b45e8
	STFD	y02,  9 * SIZE(YY)
kusano 2b45e8
	LFD	y02, 13 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	STFD	y03, 10 * SIZE(YY)
kusano 2b45e8
	LFD	y03, 14 * SIZE(YY)
kusano 2b45e8
	STFD	y04, 11 * SIZE(YY)
kusano 2b45e8
	LFD	y04, 15 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp1, a1,  xsum1
kusano 2b45e8
	DCBT(AO4, PREA)
kusano 2b45e8
	FMADD	y01, atemp1, a1,  y01
kusano 2b45e8
	LFD	a1, 16 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp1, a5,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp1, a2,  y02
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp1, a9,  xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp1, a3,  y03
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp1, a13, xsum4
kusano 2b45e8
	LFD	xtemp1, 16 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp1, a4,  y04
kusano 2b45e8
	addi	YY, YY, 16 * SIZE
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp2, a2,  xsum1
kusano 2b45e8
	LFD	a2, 17 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp2, a5,  y01
kusano 2b45e8
	LFD	a5, 16 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp2, a6,  xsum2
kusano 2b45e8
	addi	AO3, AO3, 16 * SIZE
kusano 2b45e8
	FMADD	y02, atemp2, a6,  y02
kusano 2b45e8
	LFD	a6, 17 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp2, a10, xsum3
kusano 2b45e8
	addi	AO1, AO1, 16 * SIZE
kusano 2b45e8
	FMADD	y03, atemp2, a7,  y03
kusano 2b45e8
	addi	AO2, AO2, 16 * SIZE
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp2, a14, xsum4
kusano 2b45e8
	LFD	xtemp2, 17 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp2, a8,  y04
kusano 2b45e8
	addi	AO4, AO4, 16 * SIZE
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp3, a3,  xsum1
kusano 2b45e8
	LFD	a3,  2 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp3, a9,  y01
kusano 2b45e8
	LFD	a9,  0 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp3, a7,  xsum2
kusano 2b45e8
	LFD	a7,  2 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp3, a10, y02
kusano 2b45e8
	LFD	a10,  1 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp3, a11, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp3, a11, y03
kusano 2b45e8
	LFD	a11,  2 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp3, a15, xsum4
kusano 2b45e8
	LFD	xtemp3, 18 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp3, a12, y04
kusano 2b45e8
	addi	XX, XX, 16 * SIZE
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp4, a4,  xsum1
kusano 2b45e8
	LFD	a4,  3 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp4, a13, y01
kusano 2b45e8
	LFD	a13,  0 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp4, a8,  xsum2
kusano 2b45e8
	LFD	a8,  3 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp4, a14, y02
kusano 2b45e8
	LFD	a14,  1 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp4, a12, xsum3
kusano 2b45e8
	LFD	a12,  3 * SIZE(AO3)
kusano 2b45e8
	FMADD	y03, atemp4, a15, y03
kusano 2b45e8
	LFD	a15,  2 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp4, a16, xsum4
kusano 2b45e8
	LFD	xtemp4,  3 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp4, a16, y04
kusano 2b45e8
	LFD	a16,  3 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	STFD	y01, -4 * SIZE(YY)
kusano 2b45e8
	LFD	y01,  0 * SIZE(YY)
kusano 2b45e8
	STFD	y02, -3 * SIZE(YY)
kusano 2b45e8
	LFD	y02,  1 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	STFD	y03, -2 * SIZE(YY)
kusano 2b45e8
	LFD	y03,  2 * SIZE(YY)
kusano 2b45e8
	STFD	y04, -1 * SIZE(YY)
kusano 2b45e8
	LFD	y04,  3 * SIZE(YY)
kusano 2b45e8
	bdnz	LL(12)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(14):
kusano 2b45e8
	andi.	r0,  IS, 8
kusano 2b45e8
	ble	LL(15)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp1, a1,  xsum1
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y01, atemp1, a1,  y01
kusano 2b45e8
	LFD	a1,  4 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp1, a5,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp1, a2,  y02
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp1, a9,  xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp1, a3,  y03
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp1, a13, xsum4
kusano 2b45e8
	LFD	xtemp1,  4 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp1, a4,  y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp2, a2,  xsum1
kusano 2b45e8
	LFD	a2,  5 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp2, a5,  y01
kusano 2b45e8
	LFD	a5,  4 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp2, a6,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp2, a6,  y02
kusano 2b45e8
	LFD	a6,  5 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp2, a10, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp2, a7,  y03
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp2, a14, xsum4
kusano 2b45e8
	LFD	xtemp2,  5 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp2, a8,  y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp3, a3,  xsum1
kusano 2b45e8
	LFD	a3,  6 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp3, a9,  y01
kusano 2b45e8
	LFD	a9,  4 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp3, a7,  xsum2
kusano 2b45e8
	LFD	a7,  6 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp3, a10, y02
kusano 2b45e8
	LFD	a10, 5 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp3, a11, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp3, a11, y03
kusano 2b45e8
	LFD	a11, 6 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp3, a15, xsum4
kusano 2b45e8
	LFD	xtemp3,  6 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp3, a12, y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp4, a4,  xsum1
kusano 2b45e8
	LFD	a4,  7 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp4, a13, y01
kusano 2b45e8
	LFD	a13, 4 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp4, a8,  xsum2
kusano 2b45e8
	LFD	a8,  7 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp4, a14, y02
kusano 2b45e8
	LFD	a14, 5 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp4, a12, xsum3
kusano 2b45e8
	LFD	a12, 7 * SIZE(AO3)
kusano 2b45e8
	FMADD	y03, atemp4, a15, y03
kusano 2b45e8
	LFD	a15, 6 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp4, a16, xsum4
kusano 2b45e8
	LFD	xtemp4,  7 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp4, a16, y04
kusano 2b45e8
	LFD	a16, 7 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	STFD	y01,  0 * SIZE(YY)
kusano 2b45e8
	LFD	y01,  4 * SIZE(YY)
kusano 2b45e8
	STFD	y02,  1 * SIZE(YY)
kusano 2b45e8
	LFD	y02,  5 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	STFD	y03,  2 * SIZE(YY)
kusano 2b45e8
	LFD	y03,  6 * SIZE(YY)
kusano 2b45e8
	STFD	y04,  3 * SIZE(YY)
kusano 2b45e8
	LFD	y04,  7 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp1, a1,  xsum1
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y01, atemp1, a1,  y01
kusano 2b45e8
	LFD	a1,  8 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp1, a5,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp1, a2,  y02
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp1, a9,  xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp1, a3,  y03
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp1, a13, xsum4
kusano 2b45e8
	LFD	xtemp1,  8 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp1, a4,  y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp2, a2,  xsum1
kusano 2b45e8
	LFD	a2,  9 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp2, a5,  y01
kusano 2b45e8
	LFD	a5,  8 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp2, a6,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp2, a6,  y02
kusano 2b45e8
	LFD	a6,  9 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp2, a10, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp2, a7,  y03
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp2, a14, xsum4
kusano 2b45e8
	LFD	xtemp2,  9 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp2, a8,  y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp3, a3,  xsum1
kusano 2b45e8
	LFD	a3, 10 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp3, a9,  y01
kusano 2b45e8
	LFD	a9,  8 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp3, a7,  xsum2
kusano 2b45e8
	LFD	a7, 10 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp3, a10, y02
kusano 2b45e8
	LFD	a10, 9 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp3, a11, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp3, a11, y03
kusano 2b45e8
	LFD	a11, 10 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp3, a15, xsum4
kusano 2b45e8
	LFD	xtemp3, 10 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp3, a12, y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp4, a4,  xsum1
kusano 2b45e8
	LFD	a4, 11 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp4, a13, y01
kusano 2b45e8
	LFD	a13, 8 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp4, a8,  xsum2
kusano 2b45e8
	LFD	a8, 11 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp4, a14, y02
kusano 2b45e8
	LFD	a14, 9 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp4, a12, xsum3
kusano 2b45e8
	LFD	a12, 11 * SIZE(AO3)
kusano 2b45e8
	FMADD	y03, atemp4, a15, y03
kusano 2b45e8
	LFD	a15, 10 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp4, a16, xsum4
kusano 2b45e8
	LFD	xtemp4, 11 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp4, a16, y04
kusano 2b45e8
	LFD	a16, 11 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	addi	AO1, AO1, 8 * SIZE
kusano 2b45e8
	addi	AO2, AO2, 8 * SIZE
kusano 2b45e8
	addi	AO3, AO3, 8 * SIZE
kusano 2b45e8
	addi	AO4, AO4, 8 * SIZE
kusano 2b45e8
kusano 2b45e8
	STFD	y01,  4 * SIZE(YY)
kusano 2b45e8
	LFD	y01,  8 * SIZE(YY)
kusano 2b45e8
	STFD	y02,  5 * SIZE(YY)
kusano 2b45e8
	LFD	y02,  9 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	STFD	y03,  6 * SIZE(YY)
kusano 2b45e8
	LFD	y03, 10 * SIZE(YY)
kusano 2b45e8
	STFD	y04,  7 * SIZE(YY)
kusano 2b45e8
	LFD	y04, 11 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	addi	XX, XX, 8 * SIZE
kusano 2b45e8
	addi	YY, YY, 8 * SIZE
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(15):
kusano 2b45e8
	andi.	r0,  IS, 4
kusano 2b45e8
	ble	LL(18)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp1, a1,  xsum1
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y01, atemp1, a1,  y01
kusano 2b45e8
	LFD	a1,  4 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp1, a5,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp1, a2,  y02
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp1, a9,  xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp1, a3,  y03
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp1, a13, xsum4
kusano 2b45e8
	LFD	xtemp1,  4 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp1, a4,  y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp2, a2,  xsum1
kusano 2b45e8
	LFD	a2,  5 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp2, a5,  y01
kusano 2b45e8
	LFD	a5,  4 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp2, a6,  xsum2
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y02, atemp2, a6,  y02
kusano 2b45e8
	LFD	a6,  5 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp2, a10, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp2, a7,  y03
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp2, a14, xsum4
kusano 2b45e8
	LFD	xtemp2,  5 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp2, a8,  y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp3, a3,  xsum1
kusano 2b45e8
	LFD	a3,  6 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp3, a9,  y01
kusano 2b45e8
	LFD	a9,  4 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp3, a7,  xsum2
kusano 2b45e8
	LFD	a7,  6 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp3, a10, y02
kusano 2b45e8
	LFD	a10, 5 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp3, a11, xsum3
kusano 2b45e8
	NOP1
kusano 2b45e8
	FMADD	y03, atemp3, a11, y03
kusano 2b45e8
	LFD	a11, 6 * SIZE(AO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp3, a15, xsum4
kusano 2b45e8
	LFD	xtemp3,  6 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp3, a12, y04
kusano 2b45e8
	NOP2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp4, a4,  xsum1
kusano 2b45e8
	LFD	a4,  7 * SIZE(AO1)
kusano 2b45e8
	FMADD	y01, atemp4, a13, y01
kusano 2b45e8
	LFD	a13, 4 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum2, xtemp4, a8,  xsum2
kusano 2b45e8
	LFD	a8,  7 * SIZE(AO2)
kusano 2b45e8
	FMADD	y02, atemp4, a14, y02
kusano 2b45e8
	LFD	a14, 5 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum3, xtemp4, a12, xsum3
kusano 2b45e8
	LFD	a12, 7 * SIZE(AO3)
kusano 2b45e8
	FMADD	y03, atemp4, a15, y03
kusano 2b45e8
	LFD	a15, 6 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum4, xtemp4, a16, xsum4
kusano 2b45e8
	LFD	xtemp4,  7 * SIZE(XX)
kusano 2b45e8
	FMADD	y04, atemp4, a16, y04
kusano 2b45e8
	LFD	a16, 7 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	addi	AO1, AO1, 4 * SIZE
kusano 2b45e8
	addi	AO2, AO2, 4 * SIZE
kusano 2b45e8
	addi	AO3, AO3, 4 * SIZE
kusano 2b45e8
	addi	AO4, AO4, 4 * SIZE
kusano 2b45e8
kusano 2b45e8
	STFD	y01,  0 * SIZE(YY)
kusano 2b45e8
	LFD	y01,  4 * SIZE(YY)
kusano 2b45e8
	STFD	y02,  1 * SIZE(YY)
kusano 2b45e8
	LFD	y02,  5 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	STFD	y03,  2 * SIZE(YY)
kusano 2b45e8
	LFD	y03,  6 * SIZE(YY)
kusano 2b45e8
	STFD	y04,  3 * SIZE(YY)
kusano 2b45e8
	LFD	y04,  7 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	addi	XX, XX, 4 * SIZE
kusano 2b45e8
	addi	YY, YY, 4 * SIZE
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(18):
kusano 2b45e8
	LFD	xtemp1, ALPHA 
kusano 2b45e8
kusano 2b45e8
	FMUL	xsum1, xtemp1, xsum1
kusano 2b45e8
	FMUL	xsum2, xtemp1, xsum2
kusano 2b45e8
	FMUL	xsum3, xtemp1, xsum3
kusano 2b45e8
	FMUL	xsum4, xtemp1, xsum4
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, atemp1, a1,  xsum1
kusano 2b45e8
	FMADD	xsum2, atemp1, a5,  xsum2
kusano 2b45e8
	FMADD	xsum3, atemp1, a9,  xsum3
kusano 2b45e8
	FMADD	xsum4, atemp1, a13, xsum4
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, atemp2, a5,  xsum1
kusano 2b45e8
	FMADD	xsum2, atemp2, a6,  xsum2
kusano 2b45e8
	FMADD	xsum3, atemp2, a10, xsum3
kusano 2b45e8
	FMADD	xsum4, atemp2, a14, xsum4
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, atemp3, a9,  xsum1
kusano 2b45e8
	FMADD	xsum2, atemp3, a10, xsum2
kusano 2b45e8
	FMADD	xsum3, atemp3, a11, xsum3
kusano 2b45e8
	FMADD	xsum4, atemp3, a15, xsum4
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, atemp4, a13, xsum1
kusano 2b45e8
	FMADD	xsum2, atemp4, a14, xsum2
kusano 2b45e8
	FMADD	xsum3, atemp4, a15, xsum3
kusano 2b45e8
	FMADD	xsum4, atemp4, a16, xsum4
kusano 2b45e8
kusano 2b45e8
	FADD	y01, y01, xsum1
kusano 2b45e8
	FADD	y02, y02, xsum2
kusano 2b45e8
	FADD	y03, y03, xsum3
kusano 2b45e8
	FADD	y04, y04, xsum4
kusano 2b45e8
kusano 2b45e8
	STFD	y01,  0 * SIZE(YY)
kusano 2b45e8
	STFD	y02,  1 * SIZE(YY)
kusano 2b45e8
	STFD	y03,  2 * SIZE(YY)
kusano 2b45e8
	STFD	y04,  3 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	addi	TEMP, IS, 8
kusano 2b45e8
	addi	IS,   IS, 4
kusano 2b45e8
	cmpw	cr0, TEMP, M
kusano 2b45e8
	ble	LL(11)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(20):
kusano 2b45e8
	andi.	TEMP, M, 2
kusano 2b45e8
	ble	LL(30)
kusano 2b45e8
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	add	AO2, A,   LDA
kusano 2b45e8
	add	A,   AO2, LDA
kusano 2b45e8
kusano 2b45e8
	slwi	TEMP,  IS,  BASE_SHIFT
kusano 2b45e8
	add	TEMP, X, TEMP
kusano 2b45e8
kusano 2b45e8
	LFD	atemp1, 0 * SIZE(TEMP)
kusano 2b45e8
	LFD	atemp2, 1 * SIZE(TEMP)
kusano 2b45e8
kusano 2b45e8
	LFD	a1, ALPHA 
kusano 2b45e8
kusano 2b45e8
	FMUL	atemp1, a1, atemp1
kusano 2b45e8
	FMUL	atemp2, a1, atemp2
kusano 2b45e8
kusano 2b45e8
	lfd	xsum1, FZERO
kusano 2b45e8
	fmr	xsum2, xsum1
kusano 2b45e8
kusano 2b45e8
	mr	XX, X
kusano 2b45e8
	mr	YY, NEW_Y
kusano 2b45e8
kusano 2b45e8
	LFD	xtemp1,  0 * SIZE(XX)
kusano 2b45e8
	LFD	xtemp2,  1 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	LFD	y01,  0 * SIZE(YY)
kusano 2b45e8
	LFD	y02,  1 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	LFD	a1,  0 * SIZE(AO1)
kusano 2b45e8
	LFD	a2,  1 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	LFD	a5,  0 * SIZE(AO2)
kusano 2b45e8
	LFD	a6,  1 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  IS, 1
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(28)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(22):
kusano 2b45e8
	FMADD	xsum1, xtemp1, a1,  xsum1
kusano 2b45e8
	FMADD	xsum2, xtemp1, a5,  xsum2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, xtemp2, a2,  xsum1
kusano 2b45e8
	FMADD	xsum2, xtemp2, a6,  xsum2
kusano 2b45e8
kusano 2b45e8
	FMADD	y01, atemp1, a1,  y01
kusano 2b45e8
	FMADD	y02, atemp1, a2,  y02
kusano 2b45e8
	FMADD	y01, atemp2, a5,  y01
kusano 2b45e8
	FMADD	y02, atemp2, a6,  y02
kusano 2b45e8
kusano 2b45e8
	LFD	xtemp1,  2 * SIZE(XX)
kusano 2b45e8
	LFD	xtemp2,  3 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	LFD	a1,  2 * SIZE(AO1)
kusano 2b45e8
	LFD	a2,  3 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	LFD	a5,  2 * SIZE(AO2)
kusano 2b45e8
	LFD	a6,  3 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	STFD	y01,  0 * SIZE(YY)
kusano 2b45e8
	STFD	y02,  1 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	LFD	y01,  2 * SIZE(YY)
kusano 2b45e8
	LFD	y02,  3 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	addi	AO1, AO1, 2 * SIZE
kusano 2b45e8
	addi	AO2, AO2, 2 * SIZE
kusano 2b45e8
kusano 2b45e8
	addi	XX, XX, 2 * SIZE
kusano 2b45e8
	addi	YY, YY, 2 * SIZE
kusano 2b45e8
kusano 2b45e8
	bdnz	LL(22)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(28):
kusano 2b45e8
	LFD	xtemp1, ALPHA 
kusano 2b45e8
kusano 2b45e8
	FMUL	xsum1, xtemp1, xsum1
kusano 2b45e8
	FMUL	xsum2, xtemp1, xsum2
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, atemp1, a1,  xsum1
kusano 2b45e8
	FMADD	xsum2, atemp1, a5,  xsum2
kusano 2b45e8
	FMADD	xsum1, atemp2, a5,  xsum1
kusano 2b45e8
	FMADD	xsum2, atemp2, a6,  xsum2
kusano 2b45e8
kusano 2b45e8
	FADD	y01, y01, xsum1
kusano 2b45e8
	FADD	y02, y02, xsum2
kusano 2b45e8
kusano 2b45e8
	STFD	y01,  0 * SIZE(YY)
kusano 2b45e8
	STFD	y02,  1 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	addi	IS, IS, 2
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
LL(30):
kusano 2b45e8
	andi.	TEMP, M, 1
kusano 2b45e8
	ble	LL(990)
kusano 2b45e8
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
kusano 2b45e8
	slwi	TEMP,  IS,  BASE_SHIFT
kusano 2b45e8
	add	TEMP, X, TEMP
kusano 2b45e8
kusano 2b45e8
	LFD	atemp1, 0 * SIZE(TEMP)
kusano 2b45e8
kusano 2b45e8
	LFD	a1, ALPHA 
kusano 2b45e8
kusano 2b45e8
	FMUL	atemp1, a1, atemp1
kusano 2b45e8
kusano 2b45e8
	lfd	xsum1, FZERO
kusano 2b45e8
kusano 2b45e8
	mr	XX, X
kusano 2b45e8
	mr	YY, NEW_Y
kusano 2b45e8
kusano 2b45e8
	LFD	xtemp1,  0 * SIZE(XX)
kusano 2b45e8
	LFD	y01,  0 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	LFD	a1,  0 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	mtspr	CTR, IS
kusano 2b45e8
	cmpwi	cr0, IS, 0
kusano 2b45e8
	ble	LL(38)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(32):
kusano 2b45e8
	FMADD	xsum1, xtemp1, a1,  xsum1
kusano 2b45e8
kusano 2b45e8
	FMADD	y01, atemp1, a1,  y01
kusano 2b45e8
kusano 2b45e8
	LFD	xtemp1,  1 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	LFD	a1,   1 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	STFD	y01,  0 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	LFD	y01,  1 * SIZE(YY)
kusano 2b45e8
kusano 2b45e8
	addi	AO1, AO1, 1 * SIZE
kusano 2b45e8
kusano 2b45e8
	addi	XX, XX, 1 * SIZE
kusano 2b45e8
	addi	YY, YY, 1 * SIZE
kusano 2b45e8
kusano 2b45e8
	bdnz	LL(32)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(38):
kusano 2b45e8
	LFD	xtemp1, ALPHA 
kusano 2b45e8
kusano 2b45e8
	FMUL	xsum1, xtemp1, xsum1
kusano 2b45e8
kusano 2b45e8
	FMADD	xsum1, atemp1, a1,  xsum1
kusano 2b45e8
kusano 2b45e8
	FADD	y01, y01, xsum1
kusano 2b45e8
kusano 2b45e8
	STFD	y01,  0 * SIZE(YY)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(990):
kusano 2b45e8
	cmpwi	cr0, INCY, SIZE
kusano 2b45e8
	beq	LL(999)
kusano 2b45e8
kusano 2b45e8
	mr	YY, Y
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, M, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(995)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(991):
kusano 2b45e8
	LFD	f0,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	LFD	f1,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	LFD	f2,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	LFD	f3,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	LFD	f4,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	LFD	f5,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	LFD	f6,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	LFD	f7,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
kusano 2b45e8
	LFD	f8,   0 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	f9,   1 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	f10,  2 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	f11,  3 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	f12,  4 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	f13,  5 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	f14,  6 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	f15,  7 * SIZE(NEW_Y)
kusano 2b45e8
	addi	NEW_Y, NEW_Y, 8 * SIZE
kusano 2b45e8
kusano 2b45e8
	FADD	f8,  f8,  f0
kusano 2b45e8
	FADD	f9,  f9,  f1
kusano 2b45e8
	FADD	f10, f10, f2
kusano 2b45e8
	FADD	f11, f11, f3
kusano 2b45e8
	FADD	f12, f12, f4
kusano 2b45e8
	FADD	f13, f13, f5
kusano 2b45e8
	FADD	f14, f14, f6
kusano 2b45e8
	FADD	f15, f15, f7
kusano 2b45e8
kusano 2b45e8
	STFD	f8,  0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	STFD	f9,  0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	STFD	f10, 0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	STFD	f11, 0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	STFD	f12, 0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	STFD	f13, 0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	STFD	f14, 0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	STFD	f15, 0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	bdnz	LL(991)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(995):
kusano 2b45e8
	andi.	J, M, 4
kusano 2b45e8
	ble	LL(996)
kusano 2b45e8
kusano 2b45e8
	LFD	f0,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	LFD	f1,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	LFD	f2,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	LFD	f3,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
kusano 2b45e8
	LFD	f8,   0 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	f9,   1 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	f10,  2 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	f11,  3 * SIZE(NEW_Y)
kusano 2b45e8
	addi	NEW_Y, NEW_Y, 4 * SIZE
kusano 2b45e8
kusano 2b45e8
	FADD	f8,  f8,  f0
kusano 2b45e8
	FADD	f9,  f9,  f1
kusano 2b45e8
	FADD	f10, f10, f2
kusano 2b45e8
	FADD	f11, f11, f3
kusano 2b45e8
kusano 2b45e8
	STFD	f8,  0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	STFD	f9,  0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	STFD	f10, 0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	STFD	f11, 0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(996):
kusano 2b45e8
	andi.	J, M, 2
kusano 2b45e8
	ble	LL(997)
kusano 2b45e8
kusano 2b45e8
	LFD	f0,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	LFD	f1,  0 * SIZE(Y)
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
kusano 2b45e8
	LFD	f8,   0 * SIZE(NEW_Y)
kusano 2b45e8
	LFD	f9,   1 * SIZE(NEW_Y)
kusano 2b45e8
	addi	NEW_Y, NEW_Y, 2 * SIZE
kusano 2b45e8
kusano 2b45e8
	FADD	f8,  f8,  f0
kusano 2b45e8
	FADD	f9,  f9,  f1
kusano 2b45e8
kusano 2b45e8
	STFD	f8,  0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	STFD	f9,  0 * SIZE(YY)
kusano 2b45e8
	add	YY, YY, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(997):
kusano 2b45e8
	andi.	J, M, 1
kusano 2b45e8
	ble	LL(999)
kusano 2b45e8
kusano 2b45e8
	LFD	f0,  0 * SIZE(Y)
kusano 2b45e8
	LFD	f8,   0 * SIZE(NEW_Y)
kusano 2b45e8
kusano 2b45e8
	FADD	f8,  f8,  f0
kusano 2b45e8
kusano 2b45e8
	STFD	f8,  0 * SIZE(YY)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(999):
kusano 2b45e8
	li	r3, 0
kusano 2b45e8
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
	lfd	f18,    32(SP)
kusano 2b45e8
	lfd	f19,    40(SP)
kusano 2b45e8
	lfd	f20,    48(SP)
kusano 2b45e8
	lfd	f21,    56(SP)
kusano 2b45e8
	lfd	f22,    64(SP)
kusano 2b45e8
	lfd	f23,    72(SP)
kusano 2b45e8
	lfd	f24,    80(SP)
kusano 2b45e8
	lfd	f25,    88(SP)
kusano 2b45e8
	lfd	f26,    96(SP)
kusano 2b45e8
	lfd	f27,   104(SP)
kusano 2b45e8
	lfd	f28,   112(SP)
kusano 2b45e8
	lfd	f29,   120(SP)
kusano 2b45e8
	lfd	f30,   128(SP)
kusano 2b45e8
	lfd	f31,   136(SP)
kusano 2b45e8
kusano 2b45e8
#ifdef __64BIT__
kusano 2b45e8
	ld	r14,   144(SP)
kusano 2b45e8
	ld	r15,   152(SP)
kusano 2b45e8
	ld	r16,   160(SP)
kusano 2b45e8
	ld	r17,   168(SP)
kusano 2b45e8
	ld	r18,   176(SP)
kusano 2b45e8
	ld	r19,   184(SP)
kusano 2b45e8
	ld	r20,   192(SP)
kusano 2b45e8
	ld	r21,   200(SP)
kusano 2b45e8
	ld	r22,   208(SP)
kusano 2b45e8
	ld	r23,   216(SP)
kusano 2b45e8
	ld	r24,   224(SP)
kusano 2b45e8
	ld	r25,   232(SP)
kusano 2b45e8
	ld	r26,   240(SP)
kusano 2b45e8
	ld	r27,   248(SP)
kusano 2b45e8
#else
kusano 2b45e8
	lwz	r14,   144(SP)
kusano 2b45e8
	lwz	r15,   148(SP)
kusano 2b45e8
	lwz	r16,   152(SP)
kusano 2b45e8
	lwz	r17,   156(SP)
kusano 2b45e8
	lwz	r18,   160(SP)
kusano 2b45e8
	lwz	r19,   164(SP)
kusano 2b45e8
	lwz	r20,   168(SP)
kusano 2b45e8
	lwz	r21,   172(SP)
kusano 2b45e8
	lwz	r22,   176(SP)
kusano 2b45e8
	lwz	r23,   180(SP)
kusano 2b45e8
	lwz	r24,   184(SP)
kusano 2b45e8
	lwz	r25,   188(SP)
kusano 2b45e8
	lwz	r26,   192(SP)
kusano 2b45e8
	lwz	r27,   196(SP)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	addi	SP, SP, STACKSIZE
kusano 2b45e8
	blr
kusano 2b45e8
kusano 2b45e8
	EPILOGUE
kusano 2b45e8
#endif