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 P 2048
kusano 2b45e8
kusano 2b45e8
#ifndef __64BIT__
kusano 2b45e8
#define STACKSIZE 224
kusano 2b45e8
#else
kusano 2b45e8
#define STACKSIZE 304
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef linux
kusano 2b45e8
#ifndef __64BIT__
kusano 2b45e8
#define M	r3
kusano 2b45e8
#define	N	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
#else
kusano 2b45e8
#define M	r3
kusano 2b45e8
#define	N	r4
kusano 2b45e8
#define A	r8
kusano 2b45e8
#define LDA	r9
kusano 2b45e8
#define X	r10
kusano 2b45e8
#define	INCX	r5
kusano 2b45e8
#define	Y	r6
kusano 2b45e8
#define	INCY	r7
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	N	r4
kusano 2b45e8
#define A	r10
kusano 2b45e8
#define LDA	r5
kusano 2b45e8
#define X	r6
kusano 2b45e8
#define	INCX	r7
kusano 2b45e8
#define	Y	r8
kusano 2b45e8
#define	INCY	r9
kusano 2b45e8
#else
kusano 2b45e8
#define M	r3
kusano 2b45e8
#define	N	r4
kusano 2b45e8
#define A	r8
kusano 2b45e8
#define LDA	r9
kusano 2b45e8
#define X	r10
kusano 2b45e8
#define	INCX	r5
kusano 2b45e8
#define	Y	r6
kusano 2b45e8
#define	INCY	r7
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#define	BUFFER	r11
kusano 2b45e8
#define	XP	r12
kusano 2b45e8
#define	MIN_N	r14
kusano 2b45e8
#define	J	r15
kusano 2b45e8
#define CO	r16
kusano 2b45e8
#define	BO	r17
kusano 2b45e8
#define	PLDA_M	r18
kusano 2b45e8
#define	AO1	r19
kusano 2b45e8
#define	AO2	r20
kusano 2b45e8
#define	AO3	r21
kusano 2b45e8
#define	AO4	r22
kusano 2b45e8
#define IS	r23
kusano 2b45e8
#define	PREA	r24
kusano 2b45e8
#define	PREC	r25
kusano 2b45e8
kusano 2b45e8
#define	Y1	r23   /* dummy; should be same as gemv_n.S */
kusano 2b45e8
#define Y2	r24   /* dummy; should be same as gemv_n.S */
kusano 2b45e8
kusano 2b45e8
#if defined(PPCG4)
kusano 2b45e8
#define PREFETCHSIZE_A  34
kusano 2b45e8
#define PREFETCHSIZE_C  16
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(PPC440) || defined(PPC440FP2)
kusano 2b45e8
#define PREFETCHSIZE_A  34
kusano 2b45e8
#define PREFETCHSIZE_C  16
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef PPC970
kusano 2b45e8
#define PREFETCHSIZE_A  56
kusano 2b45e8
#define PREFETCHSIZE_C  16
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef CELL
kusano 2b45e8
#define PREFETCHSIZE_A  56
kusano 2b45e8
#define PREFETCHSIZE_C  16
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef POWER4
kusano 2b45e8
#define PREFETCHSIZE_A  34
kusano 2b45e8
#define PREFETCHSIZE_C  16
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef POWER5
kusano 2b45e8
#define PREFETCHSIZE_A  40
kusano 2b45e8
#define PREFETCHSIZE_C   8
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef POWER6
kusano 2b45e8
#define PREFETCHSIZE_A  24
kusano 2b45e8
#define PREFETCHSIZE_C   8
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if !(defined(CONJ) && defined(XCONJ))
kusano 2b45e8
#define FMADDR FMADD
kusano 2b45e8
#define FMSUBR FNMSUB
kusano 2b45e8
#else
kusano 2b45e8
#define FMADDR FNMSUB
kusano 2b45e8
#define FMSUBR FMADD
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifndef NEEDPARAM
kusano 2b45e8
kusano 2b45e8
#ifndef __64BIT__
kusano 2b45e8
#define FZERO	200(SP)
kusano 2b45e8
#define ALPHA_R 208(SP)
kusano 2b45e8
#define ALPHA_I 216(SP)
kusano 2b45e8
#else
kusano 2b45e8
#define FZERO	256(SP)
kusano 2b45e8
#define ALPHA_R 264(SP)
kusano 2b45e8
#define ALPHA_I 272(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	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	r0,    FZERO
kusano 2b45e8
#else
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	r0,    FZERO
kusano 2b45e8
	stw	r0,    4 + FZERO
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef linux
kusano 2b45e8
#ifndef __64BIT__
kusano 2b45e8
	lwz	INCY,	  8 + STACKSIZE(SP)
kusano 2b45e8
	lwz	BUFFER,  12 + STACKSIZE(SP)
kusano 2b45e8
#else
kusano 2b45e8
	ld	INCX,    112 + STACKSIZE(SP)
kusano 2b45e8
	ld	Y,       120 + STACKSIZE(SP)
kusano 2b45e8
	ld	INCY,    128 + STACKSIZE(SP)
kusano 2b45e8
	ld	BUFFER,  136 + 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	LDA,     56 + STACKSIZE(SP)
kusano 2b45e8
	lwz	X,       60 + STACKSIZE(SP)
kusano 2b45e8
	lwz	INCX,    64 + STACKSIZE(SP)
kusano 2b45e8
	lwz	Y,       68 + STACKSIZE(SP)
kusano 2b45e8
	lwz	INCY,    72 + STACKSIZE(SP)
kusano 2b45e8
	lwz	BUFFER,  76 + STACKSIZE(SP)
kusano 2b45e8
#else
kusano 2b45e8
	lwz	INCX,    56 + STACKSIZE(SP)
kusano 2b45e8
	lwz	Y,       60 + STACKSIZE(SP)
kusano 2b45e8
	lwz	INCY,    64 + STACKSIZE(SP)
kusano 2b45e8
	lwz	BUFFER,  68 + STACKSIZE(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#else
kusano 2b45e8
	ld	INCX,    112 + STACKSIZE(SP)
kusano 2b45e8
	ld	Y,       120 + STACKSIZE(SP)
kusano 2b45e8
	ld	INCY,    128 + STACKSIZE(SP)
kusano 2b45e8
	ld	BUFFER,  136 + STACKSIZE(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	stfd	f1, ALPHA_R
kusano 2b45e8
	stfd	f2, ALPHA_I
kusano 2b45e8
	
kusano 2b45e8
	mullw	PLDA_M, LDA, N
kusano 2b45e8
	li	XP,  P
kusano 2b45e8
	subf	PLDA_M, XP, PLDA_M
kusano 2b45e8
	slwi	PLDA_M, PLDA_M, ZBASE_SHIFT
kusano 2b45e8
kusano 2b45e8
	slwi	LDA,  LDA,  ZBASE_SHIFT
kusano 2b45e8
	slwi	INCX, INCX, ZBASE_SHIFT
kusano 2b45e8
	slwi	INCY, INCY, ZBASE_SHIFT
kusano 2b45e8
kusano 2b45e8
	li	IS,  0
kusano 2b45e8
kusano 2b45e8
	li	PREA, PREFETCHSIZE_A * SIZE
kusano 2b45e8
	li	PREC, PREFETCHSIZE_C * SIZE
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, M, 0
kusano 2b45e8
	ble	LL(End)
kusano 2b45e8
	cmpwi	cr0, N, 0
kusano 2b45e8
	ble	LL(End)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(ISLoop):
kusano 2b45e8
	subf	MIN_N, IS, M
kusano 2b45e8
	slwi	r0, IS, ZBASE_SHIFT
kusano 2b45e8
	cmpi	cr0, 0, MIN_N, P
kusano 2b45e8
	ble+	LL(min_nP)
kusano 2b45e8
	li	MIN_N, P
kusano 2b45e8
LL(min_nP):
kusano 2b45e8
	add	XP, X,  r0
kusano 2b45e8
	cmpwi	cr0, INCX, 2 * SIZE
kusano 2b45e8
	beq	LL(Main)
kusano 2b45e8
kusano 2b45e8
	mr	XP, BUFFER
kusano 2b45e8
	addi	CO, BUFFER, -SIZE
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, MIN_N, 2
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(CopyRemain)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(CopyKernel):
kusano 2b45e8
	LFD	f0, 0 * SIZE(X)
kusano 2b45e8
	LFD	f1, 1 * SIZE(X)
kusano 2b45e8
	add	X, X, INCX
kusano 2b45e8
	LFD	f2, 0 * SIZE(X)
kusano 2b45e8
	LFD	f3, 1 * SIZE(X)
kusano 2b45e8
	add	X, X, INCX
kusano 2b45e8
	LFD	f4, 0 * SIZE(X)
kusano 2b45e8
	LFD	f5, 1 * SIZE(X)
kusano 2b45e8
	add	X, X, INCX
kusano 2b45e8
	LFD	f6, 0 * SIZE(X)
kusano 2b45e8
	LFD	f7, 1 * SIZE(X)
kusano 2b45e8
	add	X, X, INCX
kusano 2b45e8
kusano 2b45e8
	STFD	f0,  1 * SIZE(CO)
kusano 2b45e8
	STFD	f1,  2 * SIZE(CO)
kusano 2b45e8
	STFD	f2,  3 * SIZE(CO)
kusano 2b45e8
	STFD	f3,  4 * SIZE(CO)
kusano 2b45e8
	STFD	f4,  5 * SIZE(CO)
kusano 2b45e8
	STFD	f5,  6 * SIZE(CO)
kusano 2b45e8
	STFD	f6,  7 * SIZE(CO)
kusano 2b45e8
	STFDU	f7,  8 * SIZE(CO)
kusano 2b45e8
	bdnz	LL(CopyKernel)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(CopyRemain):
kusano 2b45e8
	andi.	r0, MIN_N, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(Main)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(CopySub):
kusano 2b45e8
	LFD	f0, 0 * SIZE(X)
kusano 2b45e8
	LFD	f1, 1 * SIZE(X)
kusano 2b45e8
	add	X, X, INCX
kusano 2b45e8
	STFD	f0,  1 * SIZE(CO)
kusano 2b45e8
	STFDU	f1,  2 * SIZE(CO)
kusano 2b45e8
	bdnz	LL(CopySub)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(Main):
kusano 2b45e8
	mr	CO, Y
kusano 2b45e8
	addi	XP, XP, -SIZE
kusano 2b45e8
	srawi.	J, N, 2
kusano 2b45e8
	ble	LL(Remain)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(MainHead):
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
	mr     BO, XP
kusano 2b45e8
kusano 2b45e8
	lfd	 f0,  FZERO
kusano 2b45e8
	fmr	 f1,  f0
kusano 2b45e8
	fmr	 f2,  f0
kusano 2b45e8
	fmr	 f3,  f0
kusano 2b45e8
	fmr	 f4,  f0
kusano 2b45e8
	fmr	 f5,  f0
kusano 2b45e8
	fmr	 f6,  f0
kusano 2b45e8
	fmr	 f7,  f0
kusano 2b45e8
	fmr	 f8,  f0
kusano 2b45e8
	fmr	 f9,  f0
kusano 2b45e8
	fmr	 f10, f0
kusano 2b45e8
	fmr	 f11, f0
kusano 2b45e8
	fmr	 f12, f0
kusano 2b45e8
	fmr	 f13, f0
kusano 2b45e8
	fmr	 f14, f0
kusano 2b45e8
	fmr	 f15, f0
kusano 2b45e8
kusano 2b45e8
	dcbtst	 PREC, CO
kusano 2b45e8
	srawi.	r0,  MIN_N, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(MainN3)
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 0 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 1 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 0 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 1 * SIZE(AO2)
kusano 2b45e8
	LFD	f20, 0 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 1 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 0 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 1 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 1 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 2 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 3 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 4 * SIZE(BO)
kusano 2b45e8
	LFD	f28, 5 * SIZE(BO)
kusano 2b45e8
	LFD	f29, 6 * SIZE(BO)
kusano 2b45e8
	LFD	f30, 7 * SIZE(BO)
kusano 2b45e8
	LFD	f31, 8 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	bdz	LL(MainKernelSkip)
kusano 2b45e8
	.align 5
kusano 2b45e8
kusano 2b45e8
LL(MainKernel):
kusano 2b45e8
	FMADD	f0,  f16,  f24, f0
kusano 2b45e8
	FMADD	f1,  f16,  f25, f1
kusano 2b45e8
	FMADD	f2,  f17,  f24, f2
kusano 2b45e8
	FMADD	f3,  f17,  f25, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f24, f4
kusano 2b45e8
	FMADD	f5,  f18,  f25, f5
kusano 2b45e8
	FMADD	f6,  f19,  f24, f6
kusano 2b45e8
	FMADD	f7,  f19,  f25, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 2 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 3 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 2 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 3 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f24, f8
kusano 2b45e8
	FMADD	f9,  f20,  f25, f9
kusano 2b45e8
	FMADD	f10, f21,  f24, f10
kusano 2b45e8
	FMADD	f11, f21,  f25, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f24, f12
kusano 2b45e8
	FMADD	f13, f22,  f25, f13
kusano 2b45e8
	FMADD	f14, f23,  f24, f14
kusano 2b45e8
	FMADD	f15, f23,  f25, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 2 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 3 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 2 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 3 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f26, f0
kusano 2b45e8
	FMADD	f1,  f16,  f27, f1
kusano 2b45e8
	FMADD	f2,  f17,  f26, f2
kusano 2b45e8
	FMADD	f3,  f17,  f27, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f26, f4
kusano 2b45e8
	FMADD	f5,  f18,  f27, f5
kusano 2b45e8
	FMADD	f6,  f19,  f26, f6
kusano 2b45e8
	FMADD	f7,  f19,  f27, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 4 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 5 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 4 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 5 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f26, f8
kusano 2b45e8
	FMADD	f9,  f20,  f27, f9
kusano 2b45e8
	FMADD	f10, f21,  f26, f10
kusano 2b45e8
	FMADD	f11, f21,  f27, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f26, f12
kusano 2b45e8
	FMADD	f13, f22,  f27, f13
kusano 2b45e8
	FMADD	f14, f23,  f26, f14
kusano 2b45e8
	FMADD	f15, f23,  f27, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 4 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 5 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 4 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 5 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	LFD	f24,  9 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 10 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 11 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 12 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f28, f0
kusano 2b45e8
	FMADD	f1,  f16,  f29, f1
kusano 2b45e8
	FMADD	f2,  f17,  f28, f2
kusano 2b45e8
	FMADD	f3,  f17,  f29, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f28, f4
kusano 2b45e8
	FMADD	f5,  f18,  f29, f5
kusano 2b45e8
	FMADD	f6,  f19,  f28, f6
kusano 2b45e8
	FMADD	f7,  f19,  f29, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 6 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 7 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 6 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 7 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f28, f8
kusano 2b45e8
	FMADD	f9,  f20,  f29, f9
kusano 2b45e8
	FMADD	f10, f21,  f28, f10
kusano 2b45e8
	FMADD	f11, f21,  f29, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f28, f12
kusano 2b45e8
	FMADD	f13, f22,  f29, f13
kusano 2b45e8
	FMADD	f14, f23,  f28, f14
kusano 2b45e8
	FMADD	f15, f23,  f29, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 6 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 7 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 6 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 7 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f30, f0
kusano 2b45e8
	FMADD	f1,  f16,  f31, f1
kusano 2b45e8
	FMADD	f2,  f17,  f30, f2
kusano 2b45e8
	FMADD	f3,  f17,  f31, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f30, f4
kusano 2b45e8
	FMADD	f5,  f18,  f31, f5
kusano 2b45e8
	FMADD	f6,  f19,  f30, f6
kusano 2b45e8
	FMADD	f7,  f19,  f31, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 8 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 9 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 8 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 9 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f30, f8
kusano 2b45e8
	FMADD	f9,  f20,  f31, f9
kusano 2b45e8
	FMADD	f10, f21,  f30, f10
kusano 2b45e8
	FMADD	f11, f21,  f31, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f30, f12
kusano 2b45e8
	FMADD	f13, f22,  f31, f13
kusano 2b45e8
	FMADD	f14, f23,  f30, f14
kusano 2b45e8
	FMADD	f15, f23,  f31, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 8 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 9 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 8 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 9 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 13 * SIZE(BO)
kusano 2b45e8
	LFD	f29, 14 * SIZE(BO)
kusano 2b45e8
	LFD	f30, 15 * SIZE(BO)
kusano 2b45e8
	LFD	f31, 16 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f24, f0
kusano 2b45e8
	FMADD	f1,  f16,  f25, f1
kusano 2b45e8
	FMADD	f2,  f17,  f24, f2
kusano 2b45e8
	FMADD	f3,  f17,  f25, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f24, f4
kusano 2b45e8
	FMADD	f5,  f18,  f25, f5
kusano 2b45e8
	FMADD	f6,  f19,  f24, f6
kusano 2b45e8
	FMADD	f7,  f19,  f25, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 10 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 11 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 10 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 11 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f24, f8
kusano 2b45e8
	FMADD	f9,  f20,  f25, f9
kusano 2b45e8
	FMADD	f10, f21,  f24, f10
kusano 2b45e8
	FMADD	f11, f21,  f25, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f24, f12
kusano 2b45e8
	FMADD	f13, f22,  f25, f13
kusano 2b45e8
	FMADD	f14, f23,  f24, f14
kusano 2b45e8
	FMADD	f15, f23,  f25, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 10 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 11 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 10 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 11 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f26, f0
kusano 2b45e8
	FMADD	f1,  f16,  f27, f1
kusano 2b45e8
	FMADD	f2,  f17,  f26, f2
kusano 2b45e8
	FMADD	f3,  f17,  f27, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f26, f4
kusano 2b45e8
	FMADD	f5,  f18,  f27, f5
kusano 2b45e8
	FMADD	f6,  f19,  f26, f6
kusano 2b45e8
	FMADD	f7,  f19,  f27, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 12 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 13 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 12 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 13 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f26, f8
kusano 2b45e8
	FMADD	f9,  f20,  f27, f9
kusano 2b45e8
	FMADD	f10, f21,  f26, f10
kusano 2b45e8
	FMADD	f11, f21,  f27, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f26, f12
kusano 2b45e8
	FMADD	f13, f22,  f27, f13
kusano 2b45e8
	FMADD	f14, f23,  f26, f14
kusano 2b45e8
	FMADD	f15, f23,  f27, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 12 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 13 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 12 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 13 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 17 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 18 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 19 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 20 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f28, f0
kusano 2b45e8
	FMADD	f1,  f16,  f29, f1
kusano 2b45e8
	FMADD	f2,  f17,  f28, f2
kusano 2b45e8
	FMADD	f3,  f17,  f29, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f28, f4
kusano 2b45e8
	FMADD	f5,  f18,  f29, f5
kusano 2b45e8
	FMADD	f6,  f19,  f28, f6
kusano 2b45e8
	FMADD	f7,  f19,  f29, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 14 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 15 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 14 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 15 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f28, f8
kusano 2b45e8
	FMADD	f9,  f20,  f29, f9
kusano 2b45e8
	FMADD	f10, f21,  f28, f10
kusano 2b45e8
	FMADD	f11, f21,  f29, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f28, f12
kusano 2b45e8
	FMADD	f13, f22,  f29, f13
kusano 2b45e8
	FMADD	f14, f23,  f28, f14
kusano 2b45e8
	FMADD	f15, f23,  f29, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 14 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 15 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 14 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 15 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f30, f0
kusano 2b45e8
	FMADD	f1,  f16,  f31, f1
kusano 2b45e8
	FMADD	f2,  f17,  f30, f2
kusano 2b45e8
	FMADD	f3,  f17,  f31, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f30, f4
kusano 2b45e8
	FMADD	f5,  f18,  f31, f5
kusano 2b45e8
	FMADD	f6,  f19,  f30, f6
kusano 2b45e8
	FMADD	f7,  f19,  f31, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 16 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 17 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 16 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 17 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	addi	AO1, AO1, 16 * SIZE
kusano 2b45e8
	addi	AO2, AO2, 16 * SIZE
kusano 2b45e8
	DCBT(AO1, PREA)
kusano 2b45e8
	DCBT(AO2, PREA)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f30, f8
kusano 2b45e8
	FMADD	f9,  f20,  f31, f9
kusano 2b45e8
	FMADD	f10, f21,  f30, f10
kusano 2b45e8
	FMADD	f11, f21,  f31, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f30, f12
kusano 2b45e8
	FMADD	f13, f22,  f31, f13
kusano 2b45e8
	FMADD	f14, f23,  f30, f14
kusano 2b45e8
	FMADD	f15, f23,  f31, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 16 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 17 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 16 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 17 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 21 * SIZE(BO)
kusano 2b45e8
	LFD	f29, 22 * SIZE(BO)
kusano 2b45e8
	LFD	f30, 23 * SIZE(BO)
kusano 2b45e8
	LFD	f31, 24 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO3, AO3, 16 * SIZE
kusano 2b45e8
	addi	AO4, AO4, 16 * SIZE
kusano 2b45e8
	DCBT(AO3, PREA)
kusano 2b45e8
	DCBT(AO4, PREA)
kusano 2b45e8
kusano 2b45e8
	addi	BO, BO, 16 * SIZE
kusano 2b45e8
 	bdnz	LL(MainKernel)
kusano 2b45e8
	.align 4	
kusano 2b45e8
kusano 2b45e8
LL(MainKernelSkip):
kusano 2b45e8
	FMADD	f0,  f16,  f24, f0
kusano 2b45e8
	FMADD	f1,  f16,  f25, f1
kusano 2b45e8
	FMADD	f2,  f17,  f24, f2
kusano 2b45e8
	FMADD	f3,  f17,  f25, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f24, f4
kusano 2b45e8
	FMADD	f5,  f18,  f25, f5
kusano 2b45e8
	FMADD	f6,  f19,  f24, f6
kusano 2b45e8
	FMADD	f7,  f19,  f25, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 2 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 3 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 2 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 3 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f24, f8
kusano 2b45e8
	FMADD	f9,  f20,  f25, f9
kusano 2b45e8
	FMADD	f10, f21,  f24, f10
kusano 2b45e8
	FMADD	f11, f21,  f25, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f24, f12
kusano 2b45e8
	FMADD	f13, f22,  f25, f13
kusano 2b45e8
	FMADD	f14, f23,  f24, f14
kusano 2b45e8
	FMADD	f15, f23,  f25, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 2 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 3 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 2 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 3 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f26, f0
kusano 2b45e8
	FMADD	f1,  f16,  f27, f1
kusano 2b45e8
	FMADD	f2,  f17,  f26, f2
kusano 2b45e8
	FMADD	f3,  f17,  f27, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f26, f4
kusano 2b45e8
	FMADD	f5,  f18,  f27, f5
kusano 2b45e8
	FMADD	f6,  f19,  f26, f6
kusano 2b45e8
	FMADD	f7,  f19,  f27, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 4 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 5 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 4 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 5 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f26, f8
kusano 2b45e8
	FMADD	f9,  f20,  f27, f9
kusano 2b45e8
	FMADD	f10, f21,  f26, f10
kusano 2b45e8
	FMADD	f11, f21,  f27, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f26, f12
kusano 2b45e8
	FMADD	f13, f22,  f27, f13
kusano 2b45e8
	FMADD	f14, f23,  f26, f14
kusano 2b45e8
	FMADD	f15, f23,  f27, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 4 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 5 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 4 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 5 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f28, f0
kusano 2b45e8
	FMADD	f1,  f16,  f29, f1
kusano 2b45e8
	FMADD	f2,  f17,  f28, f2
kusano 2b45e8
	FMADD	f3,  f17,  f29, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f28, f4
kusano 2b45e8
	FMADD	f5,  f18,  f29, f5
kusano 2b45e8
	FMADD	f6,  f19,  f28, f6
kusano 2b45e8
	FMADD	f7,  f19,  f29, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 6 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 7 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 6 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 7 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f28, f8
kusano 2b45e8
	FMADD	f9,  f20,  f29, f9
kusano 2b45e8
	FMADD	f10, f21,  f28, f10
kusano 2b45e8
	FMADD	f11, f21,  f29, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f28, f12
kusano 2b45e8
	FMADD	f13, f22,  f29, f13
kusano 2b45e8
	FMADD	f14, f23,  f28, f14
kusano 2b45e8
	FMADD	f15, f23,  f29, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 6 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 7 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 6 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 7 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f30, f0
kusano 2b45e8
	FMADD	f1,  f16,  f31, f1
kusano 2b45e8
	FMADD	f2,  f17,  f30, f2
kusano 2b45e8
	FMADD	f3,  f17,  f31, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f30, f4
kusano 2b45e8
	FMADD	f5,  f18,  f31, f5
kusano 2b45e8
	FMADD	f6,  f19,  f30, f6
kusano 2b45e8
	FMADD	f7,  f19,  f31, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 8 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 9 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 8 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 9 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f30, f8
kusano 2b45e8
	FMADD	f9,  f20,  f31, f9
kusano 2b45e8
	FMADD	f10, f21,  f30, f10
kusano 2b45e8
	FMADD	f11, f21,  f31, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f30, f12
kusano 2b45e8
	FMADD	f13, f22,  f31, f13
kusano 2b45e8
	FMADD	f14, f23,  f30, f14
kusano 2b45e8
	FMADD	f15, f23,  f31, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 8 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 9 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 8 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 9 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	LFD	f24,  9 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 10 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 11 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 12 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 13 * SIZE(BO)
kusano 2b45e8
	LFD	f29, 14 * SIZE(BO)
kusano 2b45e8
	LFD	f30, 15 * SIZE(BO)
kusano 2b45e8
	LFDU	f31, 16 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f24, f0
kusano 2b45e8
	FMADD	f1,  f16,  f25, f1
kusano 2b45e8
	FMADD	f2,  f17,  f24, f2
kusano 2b45e8
	FMADD	f3,  f17,  f25, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f24, f4
kusano 2b45e8
	FMADD	f5,  f18,  f25, f5
kusano 2b45e8
	FMADD	f6,  f19,  f24, f6
kusano 2b45e8
	FMADD	f7,  f19,  f25, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 10 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 11 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 10 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 11 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f24, f8
kusano 2b45e8
	FMADD	f9,  f20,  f25, f9
kusano 2b45e8
	FMADD	f10, f21,  f24, f10
kusano 2b45e8
	FMADD	f11, f21,  f25, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f24, f12
kusano 2b45e8
	FMADD	f13, f22,  f25, f13
kusano 2b45e8
	FMADD	f14, f23,  f24, f14
kusano 2b45e8
	FMADD	f15, f23,  f25, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 10 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 11 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 10 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 11 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f26, f0
kusano 2b45e8
	FMADD	f1,  f16,  f27, f1
kusano 2b45e8
	FMADD	f2,  f17,  f26, f2
kusano 2b45e8
	FMADD	f3,  f17,  f27, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f26, f4
kusano 2b45e8
	FMADD	f5,  f18,  f27, f5
kusano 2b45e8
	FMADD	f6,  f19,  f26, f6
kusano 2b45e8
	FMADD	f7,  f19,  f27, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 12 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 13 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 12 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 13 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f26, f8
kusano 2b45e8
	FMADD	f9,  f20,  f27, f9
kusano 2b45e8
	FMADD	f10, f21,  f26, f10
kusano 2b45e8
	FMADD	f11, f21,  f27, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f26, f12
kusano 2b45e8
	FMADD	f13, f22,  f27, f13
kusano 2b45e8
	FMADD	f14, f23,  f26, f14
kusano 2b45e8
	FMADD	f15, f23,  f27, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 12 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 13 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 12 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 13 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f28, f0
kusano 2b45e8
	FMADD	f1,  f16,  f29, f1
kusano 2b45e8
	FMADD	f2,  f17,  f28, f2
kusano 2b45e8
	FMADD	f3,  f17,  f29, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f28, f4
kusano 2b45e8
	FMADD	f5,  f18,  f29, f5
kusano 2b45e8
	FMADD	f6,  f19,  f28, f6
kusano 2b45e8
	FMADD	f7,  f19,  f29, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 14 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 15 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 14 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 15 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f28, f8
kusano 2b45e8
	FMADD	f9,  f20,  f29, f9
kusano 2b45e8
	FMADD	f10, f21,  f28, f10
kusano 2b45e8
	FMADD	f11, f21,  f29, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f28, f12
kusano 2b45e8
	FMADD	f13, f22,  f29, f13
kusano 2b45e8
	FMADD	f14, f23,  f28, f14
kusano 2b45e8
	FMADD	f15, f23,  f29, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 14 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 15 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 14 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 15 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	addi	AO1, AO1, 16 * SIZE
kusano 2b45e8
	addi	AO2, AO2, 16 * SIZE
kusano 2b45e8
	addi	AO3, AO3, 16 * SIZE
kusano 2b45e8
	addi	AO4, AO4, 16 * SIZE
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f30, f0
kusano 2b45e8
	FMADD	f1,  f16,  f31, f1
kusano 2b45e8
	FMADD	f2,  f17,  f30, f2
kusano 2b45e8
	FMADD	f3,  f17,  f31, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f30, f4
kusano 2b45e8
	FMADD	f5,  f18,  f31, f5
kusano 2b45e8
	FMADD	f6,  f19,  f30, f6
kusano 2b45e8
	FMADD	f7,  f19,  f31, f7
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f30, f8
kusano 2b45e8
	FMADD	f9,  f20,  f31, f9
kusano 2b45e8
	FMADD	f10, f21,  f30, f10
kusano 2b45e8
	FMADD	f11, f21,  f31, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f30, f12
kusano 2b45e8
	FMADD	f13, f22,  f31, f13
kusano 2b45e8
	FMADD	f14, f23,  f30, f14
kusano 2b45e8
	FMADD	f15, f23,  f31, f15
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(MainN3):
kusano 2b45e8
	andi.	r0, MIN_N, 7
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(MainFinish)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 0 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 1 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 0 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 1 * SIZE(AO2)
kusano 2b45e8
	LFD	f20, 0 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 1 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 0 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 1 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 1 * SIZE(BO)
kusano 2b45e8
	LFDU	f25, 2 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO1, AO1, 2 * SIZE
kusano 2b45e8
	addi	AO2, AO2, 2 * SIZE
kusano 2b45e8
	addi	AO3, AO3, 2 * SIZE
kusano 2b45e8
	addi	AO4, AO4, 2 * SIZE
kusano 2b45e8
kusano 2b45e8
	bdz	LL(MainN3KernelSkip)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(MainN3Kernel):
kusano 2b45e8
	FMADD	f0,  f16,  f24, f0
kusano 2b45e8
	FMADD	f1,  f16,  f25, f1
kusano 2b45e8
	FMADD	f2,  f17,  f24, f2
kusano 2b45e8
	FMADD	f3,  f17,  f25, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f24, f4
kusano 2b45e8
	FMADD	f5,  f18,  f25, f5
kusano 2b45e8
	FMADD	f6,  f19,  f24, f6
kusano 2b45e8
	FMADD	f7,  f19,  f25, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 0 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 1 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 0 * SIZE(AO2)
kusano 2b45e8
	LFD	f19, 1 * SIZE(AO2)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f24, f8
kusano 2b45e8
	FMADD	f9,  f20,  f25, f9
kusano 2b45e8
	FMADD	f10, f21,  f24, f10
kusano 2b45e8
	FMADD	f11, f21,  f25, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f24, f12
kusano 2b45e8
	FMADD	f13, f22,  f25, f13
kusano 2b45e8
	FMADD	f14, f23,  f24, f14
kusano 2b45e8
	FMADD	f15, f23,  f25, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 0 * SIZE(AO3)
kusano 2b45e8
	LFD	f21, 1 * SIZE(AO3)
kusano 2b45e8
	LFD	f22, 0 * SIZE(AO4)
kusano 2b45e8
	LFD	f23, 1 * SIZE(AO4)
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 1 * SIZE(BO)
kusano 2b45e8
	LFDU	f25, 2 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO1, AO1, 2 * SIZE
kusano 2b45e8
	addi	AO2, AO2, 2 * SIZE
kusano 2b45e8
	addi	AO3, AO3, 2 * SIZE
kusano 2b45e8
	addi	AO4, AO4, 2 * SIZE
kusano 2b45e8
kusano 2b45e8
	bdnz	LL(MainN3Kernel)
kusano 2b45e8
	.align 4	
kusano 2b45e8
kusano 2b45e8
LL(MainN3KernelSkip):
kusano 2b45e8
	FMADD	f0,  f16,  f24, f0
kusano 2b45e8
	FMADD	f1,  f16,  f25, f1
kusano 2b45e8
	FMADD	f2,  f17,  f24, f2
kusano 2b45e8
	FMADD	f3,  f17,  f25, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f24, f4
kusano 2b45e8
	FMADD	f5,  f18,  f25, f5
kusano 2b45e8
	FMADD	f6,  f19,  f24, f6
kusano 2b45e8
	FMADD	f7,  f19,  f25, f7
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f24, f8
kusano 2b45e8
	FMADD	f9,  f20,  f25, f9
kusano 2b45e8
	FMADD	f10, f21,  f24, f10
kusano 2b45e8
	FMADD	f11, f21,  f25, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f24, f12
kusano 2b45e8
	FMADD	f13, f22,  f25, f13
kusano 2b45e8
	FMADD	f14, f23,  f24, f14
kusano 2b45e8
	FMADD	f15, f23,  f25, f15
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(MainFinish):
kusano 2b45e8
	lfd	f30, ALPHA_R
kusano 2b45e8
	lfd	f31, ALPHA_I
kusano 2b45e8
kusano 2b45e8
#ifndef XCONJ
kusano 2b45e8
#ifndef CONJ
kusano 2b45e8
	FSUB	f0,  f0,  f3
kusano 2b45e8
	FADD	f1,  f1,  f2
kusano 2b45e8
	FSUB	f4,  f4,  f7
kusano 2b45e8
	FADD	f5,  f5,  f6
kusano 2b45e8
	FSUB	f8,  f8,  f11
kusano 2b45e8
	FADD	f9,  f9,  f10
kusano 2b45e8
	FSUB	f12, f12, f15
kusano 2b45e8
	FADD	f13, f13, f14
kusano 2b45e8
#else
kusano 2b45e8
	FADD	f0,  f0,  f3
kusano 2b45e8
	FSUB	f1,  f1,  f2
kusano 2b45e8
	FADD	f4,  f4,  f7
kusano 2b45e8
	FSUB	f5,  f5,  f6
kusano 2b45e8
	FADD	f8,  f8,  f11
kusano 2b45e8
	FSUB	f9,  f9,  f10
kusano 2b45e8
	FADD	f12, f12, f15
kusano 2b45e8
	FSUB	f13, f13, f14
kusano 2b45e8
#endif
kusano 2b45e8
#else
kusano 2b45e8
#ifndef CONJ
kusano 2b45e8
	FADD	f0,  f0,  f3
kusano 2b45e8
	FSUB	f1,  f2,  f1
kusano 2b45e8
	FADD	f4,  f4,  f7
kusano 2b45e8
	FSUB	f5,  f6,  f5
kusano 2b45e8
	FADD	f8,  f8,  f11
kusano 2b45e8
	FSUB	f9,  f10, f9
kusano 2b45e8
	FADD	f12, f12, f15
kusano 2b45e8
	FSUB	f13, f14, f13
kusano 2b45e8
#else
kusano 2b45e8
	FSUB	f0,  f0,  f3
kusano 2b45e8
	FADD	f1,  f1,  f2
kusano 2b45e8
	FSUB	f4,  f4,  f7
kusano 2b45e8
	FADD	f5,  f5,  f6
kusano 2b45e8
	FSUB	f8,  f8,  f11
kusano 2b45e8
	FADD	f9,  f9,  f10
kusano 2b45e8
	FSUB	f12, f12, f15
kusano 2b45e8
	FADD	f13, f13, f14
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	mr	BO, CO
kusano 2b45e8
	cmpwi	cr0, INCY, 2 * SIZE
kusano 2b45e8
	bne	LL(FinishN1)
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(CO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(CO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(CO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(CO)
kusano 2b45e8
	LFD	f20,  4 * SIZE(CO)
kusano 2b45e8
	LFD	f21,  5 * SIZE(CO)
kusano 2b45e8
	LFD	f22,  6 * SIZE(CO)
kusano 2b45e8
	LFD	f23,  7 * SIZE(CO)
kusano 2b45e8
kusano 2b45e8
	FMADD	f16, f30, f0,  f16
kusano 2b45e8
	FMADDR	f17, f30, f1,  f17
kusano 2b45e8
	FMADD	f18, f30, f4,  f18
kusano 2b45e8
	FMADDR	f19, f30, f5,  f19
kusano 2b45e8
kusano 2b45e8
	FMADD	f20, f30, f8,  f20
kusano 2b45e8
	FMADDR	f21, f30, f9,  f21
kusano 2b45e8
	FMADD	f22, f30, f12, f22
kusano 2b45e8
	FMADDR	f23, f30, f13, f23
kusano 2b45e8
kusano 2b45e8
	FMSUBR	f16, f31, f1,  f16
kusano 2b45e8
	FMADD	f17, f31, f0,  f17
kusano 2b45e8
	FMSUBR	f18, f31, f5,  f18
kusano 2b45e8
	FMADD	f19, f31, f4,  f19
kusano 2b45e8
kusano 2b45e8
	FMSUBR	f20, f31, f9,  f20
kusano 2b45e8
	FMADD	f21, f31, f8,  f21
kusano 2b45e8
	FMSUBR	f22, f31, f13, f22
kusano 2b45e8
	FMADD	f23, f31, f12, f23
kusano 2b45e8
kusano 2b45e8
	STFD	f16,  0 * SIZE(CO)
kusano 2b45e8
	STFD	f17,  1 * SIZE(CO)
kusano 2b45e8
	STFD	f18,  2 * SIZE(CO)
kusano 2b45e8
	STFD	f19,  3 * SIZE(CO)
kusano 2b45e8
kusano 2b45e8
	STFD	f20,  4 * SIZE(CO)
kusano 2b45e8
	STFD	f21,  5 * SIZE(CO)
kusano 2b45e8
	STFD	f22,  6 * SIZE(CO)
kusano 2b45e8
	STFD	f23,  7 * SIZE(CO)
kusano 2b45e8
kusano 2b45e8
	addi	CO, CO, 8 * SIZE
kusano 2b45e8
kusano 2b45e8
	addi	J, J, -1
kusano 2b45e8
	cmpwi	cr0, J, 0
kusano 2b45e8
	bgt	LL(MainHead)
kusano 2b45e8
	b	LL(Remain)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(FinishN1):
kusano 2b45e8
	LFD	f16,  0 * SIZE(CO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(CO)
kusano 2b45e8
	add	CO, CO, INCY
kusano 2b45e8
kusano 2b45e8
	LFD	f18,  0 * SIZE(CO)
kusano 2b45e8
	LFD	f19,  1 * SIZE(CO)
kusano 2b45e8
	add	CO, CO, INCY
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(CO)
kusano 2b45e8
	LFD	f21,  1 * SIZE(CO)
kusano 2b45e8
	add	CO, CO, INCY
kusano 2b45e8
kusano 2b45e8
	LFD	f22,  0 * SIZE(CO)
kusano 2b45e8
	LFD	f23,  1 * SIZE(CO)
kusano 2b45e8
	add	CO, CO, INCY
kusano 2b45e8
kusano 2b45e8
	FMADD	f16, f30, f0,  f16
kusano 2b45e8
	FMADDR	f17, f30, f1,  f17
kusano 2b45e8
	FMADD	f18, f30, f4,  f18
kusano 2b45e8
	FMADDR	f19, f30, f5,  f19
kusano 2b45e8
kusano 2b45e8
	FMADD	f20, f30, f8,  f20
kusano 2b45e8
	FMADDR	f21, f30, f9,  f21
kusano 2b45e8
	FMADD	f22, f30, f12, f22
kusano 2b45e8
	FMADDR	f23, f30, f13, f23
kusano 2b45e8
kusano 2b45e8
	FMSUBR	f16, f31, f1,  f16
kusano 2b45e8
	FMADD	f17, f31, f0,  f17
kusano 2b45e8
	FMSUBR	f18, f31, f5,  f18
kusano 2b45e8
	FMADD	f19, f31, f4,  f19
kusano 2b45e8
kusano 2b45e8
	FMSUBR	f20, f31, f9,  f20
kusano 2b45e8
	FMADD	f21, f31, f8,  f21
kusano 2b45e8
	FMSUBR	f22, f31, f13, f22
kusano 2b45e8
	FMADD	f23, f31, f12, f23
kusano 2b45e8
kusano 2b45e8
	STFD	f16,  0 * SIZE(BO)
kusano 2b45e8
	STFD	f17,  1 * SIZE(BO)
kusano 2b45e8
	add	BO, BO, INCY
kusano 2b45e8
	STFD	f18,  0 * SIZE(BO)
kusano 2b45e8
	STFD	f19,  1 * SIZE(BO)
kusano 2b45e8
	add	BO, BO, INCY
kusano 2b45e8
kusano 2b45e8
	STFD	f20,  0 * SIZE(BO)
kusano 2b45e8
	STFD	f21,  1 * SIZE(BO)
kusano 2b45e8
	add	BO, BO, INCY
kusano 2b45e8
	STFD	f22,  0 * SIZE(BO)
kusano 2b45e8
	STFD	f23,  1 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	J, J, -1
kusano 2b45e8
	cmpwi	cr0, J, 0
kusano 2b45e8
	bgt	LL(MainHead)
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
LL(Remain):
kusano 2b45e8
	andi.	J, N, 3
kusano 2b45e8
	ble	LL(ISEnd)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(RemainHead):
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	add	A,   A,  LDA
kusano 2b45e8
	mr	BO,  XP
kusano 2b45e8
	lfd	f0, FZERO
kusano 2b45e8
kusano 2b45e8
	fmr	 f1,  f0
kusano 2b45e8
	fmr	 f2,  f0
kusano 2b45e8
	fmr	 f3,  f0
kusano 2b45e8
	fmr	 f4,  f0
kusano 2b45e8
	fmr	 f5,  f0
kusano 2b45e8
	fmr	 f6,  f0
kusano 2b45e8
	fmr	 f7,  f0
kusano 2b45e8
	fmr	 f8,  f0
kusano 2b45e8
	fmr	 f9,  f0
kusano 2b45e8
	fmr	 f10, f0
kusano 2b45e8
	fmr	 f11, f0
kusano 2b45e8
	fmr	 f12, f0
kusano 2b45e8
	fmr	 f13, f0
kusano 2b45e8
	fmr	 f14, f0
kusano 2b45e8
	fmr	 f15, f0
kusano 2b45e8
kusano 2b45e8
	srawi.	r0 , MIN_N, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(RemainN3)
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 0 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 1 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 2 * SIZE(AO1)
kusano 2b45e8
	LFD	f19, 3 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 4 * SIZE(AO1)
kusano 2b45e8
	LFD	f21, 5 * SIZE(AO1)
kusano 2b45e8
	LFD	f22, 6 * SIZE(AO1)
kusano 2b45e8
	LFD	f23, 7 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 1 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 2 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 3 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 4 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 5 * SIZE(BO)
kusano 2b45e8
	LFD	f29, 6 * SIZE(BO)
kusano 2b45e8
	LFD	f30, 7 * SIZE(BO)
kusano 2b45e8
	LFD	f31, 8 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	bdz	LL(RemainKernelSkip)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(RemainKernel):
kusano 2b45e8
	FMADD	f0,  f16,  f24, f0
kusano 2b45e8
	FMADD	f1,  f16,  f25, f1
kusano 2b45e8
	FMADD	f2,  f17,  f24, f2
kusano 2b45e8
	FMADD	f3,  f17,  f25, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f26, f4
kusano 2b45e8
	FMADD	f5,  f18,  f27, f5
kusano 2b45e8
	FMADD	f6,  f19,  f26, f6
kusano 2b45e8
	FMADD	f7,  f19,  f27, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  8 * SIZE(AO1)
kusano 2b45e8
	LFD	f17,  9 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 10 * SIZE(AO1)
kusano 2b45e8
	LFD	f19, 11 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	LFD	f24,  9 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 10 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 11 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 12 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f28, f8
kusano 2b45e8
	FMADD	f9,  f20,  f29, f9
kusano 2b45e8
	FMADD	f10, f21,  f28, f10
kusano 2b45e8
	FMADD	f11, f21,  f29, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f30, f12
kusano 2b45e8
	FMADD	f13, f22,  f31, f13
kusano 2b45e8
	FMADD	f14, f23,  f30, f14
kusano 2b45e8
	FMADD	f15, f23,  f31, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 12 * SIZE(AO1)
kusano 2b45e8
	LFD	f21, 13 * SIZE(AO1)
kusano 2b45e8
	LFD	f22, 14 * SIZE(AO1)
kusano 2b45e8
	LFD	f23, 15 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 13 * SIZE(BO)
kusano 2b45e8
	LFD	f29, 14 * SIZE(BO)
kusano 2b45e8
	LFD	f30, 15 * SIZE(BO)
kusano 2b45e8
	LFD	f31, 16 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f24, f0
kusano 2b45e8
	FMADD	f1,  f16,  f25, f1
kusano 2b45e8
	FMADD	f2,  f17,  f24, f2
kusano 2b45e8
	FMADD	f3,  f17,  f25, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f26, f4
kusano 2b45e8
	FMADD	f5,  f18,  f27, f5
kusano 2b45e8
	FMADD	f6,  f19,  f26, f6
kusano 2b45e8
	FMADD	f7,  f19,  f27, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 16 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 17 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 18 * SIZE(AO1)
kusano 2b45e8
	LFD	f19, 19 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 17 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 18 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 19 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 20 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f28, f8
kusano 2b45e8
	FMADD	f9,  f20,  f29, f9
kusano 2b45e8
	FMADD	f10, f21,  f28, f10
kusano 2b45e8
	FMADD	f11, f21,  f29, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f30, f12
kusano 2b45e8
	FMADD	f13, f22,  f31, f13
kusano 2b45e8
	FMADD	f14, f23,  f30, f14
kusano 2b45e8
	FMADD	f15, f23,  f31, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 20 * SIZE(AO1)
kusano 2b45e8
	LFD	f21, 21 * SIZE(AO1)
kusano 2b45e8
	LFD	f22, 22 * SIZE(AO1)
kusano 2b45e8
	LFD	f23, 23 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 21 * SIZE(BO)
kusano 2b45e8
	LFD	f29, 22 * SIZE(BO)
kusano 2b45e8
	LFD	f30, 23 * SIZE(BO)
kusano 2b45e8
	LFD	f31, 24 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO1, AO1, 16 * SIZE
kusano 2b45e8
	addi	BO,  BO,  16 * SIZE
kusano 2b45e8
kusano 2b45e8
	DCBT(AO1, PREA)
kusano 2b45e8
kusano 2b45e8
	bdnz	LL(RemainKernel)
kusano 2b45e8
	.align 4	
kusano 2b45e8
kusano 2b45e8
LL(RemainKernelSkip):
kusano 2b45e8
	FMADD	f0,  f16,  f24, f0
kusano 2b45e8
	FMADD	f1,  f16,  f25, f1
kusano 2b45e8
	FMADD	f2,  f17,  f24, f2
kusano 2b45e8
	FMADD	f3,  f17,  f25, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f26, f4
kusano 2b45e8
	FMADD	f5,  f18,  f27, f5
kusano 2b45e8
	FMADD	f6,  f19,  f26, f6
kusano 2b45e8
	FMADD	f7,  f19,  f27, f7
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  8 * SIZE(AO1)
kusano 2b45e8
	LFD	f17,  9 * SIZE(AO1)
kusano 2b45e8
	LFD	f18, 10 * SIZE(AO1)
kusano 2b45e8
	LFD	f19, 11 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	LFD	f24,  9 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 10 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 11 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 12 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f28, f8
kusano 2b45e8
	FMADD	f9,  f20,  f29, f9
kusano 2b45e8
	FMADD	f10, f21,  f28, f10
kusano 2b45e8
	FMADD	f11, f21,  f29, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f30, f12
kusano 2b45e8
	FMADD	f13, f22,  f31, f13
kusano 2b45e8
	FMADD	f14, f23,  f30, f14
kusano 2b45e8
	FMADD	f15, f23,  f31, f15
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 12 * SIZE(AO1)
kusano 2b45e8
	LFD	f21, 13 * SIZE(AO1)
kusano 2b45e8
	LFD	f22, 14 * SIZE(AO1)
kusano 2b45e8
	LFD	f23, 15 * SIZE(AO1)
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 13 * SIZE(BO)
kusano 2b45e8
	LFD	f29, 14 * SIZE(BO)
kusano 2b45e8
	LFD	f30, 15 * SIZE(BO)
kusano 2b45e8
	LFDU	f31, 16 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f16,  f24, f0
kusano 2b45e8
	FMADD	f1,  f16,  f25, f1
kusano 2b45e8
	FMADD	f2,  f17,  f24, f2
kusano 2b45e8
	FMADD	f3,  f17,  f25, f3
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f18,  f26, f4
kusano 2b45e8
	FMADD	f5,  f18,  f27, f5
kusano 2b45e8
	FMADD	f6,  f19,  f26, f6
kusano 2b45e8
	FMADD	f7,  f19,  f27, f7
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f20,  f28, f8
kusano 2b45e8
	FMADD	f9,  f20,  f29, f9
kusano 2b45e8
	FMADD	f10, f21,  f28, f10
kusano 2b45e8
	FMADD	f11, f21,  f29, f11
kusano 2b45e8
kusano 2b45e8
	FMADD	f12, f22,  f30, f12
kusano 2b45e8
	FMADD	f13, f22,  f31, f13
kusano 2b45e8
	FMADD	f14, f23,  f30, f14
kusano 2b45e8
	FMADD	f15, f23,  f31, f15
kusano 2b45e8
kusano 2b45e8
	addi	AO1, AO1, 16 * SIZE
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(RemainN3):
kusano 2b45e8
	andi.	r0,  MIN_N, 7
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(RemainFinish)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 0 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 1 * SIZE(AO1)
kusano 2b45e8
	LFD	f24, 1 * SIZE(BO)
kusano 2b45e8
	LFDU	f25, 2 * SIZE(BO)
kusano 2b45e8
	addi	AO1, AO1, 2 * SIZE
kusano 2b45e8
	bdz	LL(RemainN3KernelSkip)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(RemainN3Kernel):
kusano 2b45e8
	FMADD	f0,  f16,  f24, f0
kusano 2b45e8
	FMADD	f1,  f16,  f25, f1
kusano 2b45e8
	FMADD	f2,  f17,  f24, f2
kusano 2b45e8
	FMADD	f3,  f17,  f25, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 0 * SIZE(AO1)
kusano 2b45e8
	LFD	f17, 1 * SIZE(AO1)
kusano 2b45e8
	LFD	f24, 1 * SIZE(BO)
kusano 2b45e8
	LFDU	f25, 2 * SIZE(BO)
kusano 2b45e8
	addi	AO1, AO1, 2 * SIZE
kusano 2b45e8
	bdnz	LL(RemainN3Kernel)
kusano 2b45e8
	.align 4	
kusano 2b45e8
kusano 2b45e8
LL(RemainN3KernelSkip):
kusano 2b45e8
	FMADD	f0,  f16,  f24, f0
kusano 2b45e8
	FMADD	f1,  f16,  f25, f1
kusano 2b45e8
	FMADD	f2,  f17,  f24, f2
kusano 2b45e8
	FMADD	f3,  f17,  f25, f3
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(RemainFinish):
kusano 2b45e8
	lfd	f30, ALPHA_R
kusano 2b45e8
	lfd	f31, ALPHA_I
kusano 2b45e8
	LFD	f16,  0 * SIZE(CO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(CO)
kusano 2b45e8
kusano 2b45e8
	FADD	f0, f0, f4
kusano 2b45e8
	FADD	f1, f1, f5
kusano 2b45e8
	FADD	f2, f2, f6
kusano 2b45e8
	FADD	f3, f3, f7
kusano 2b45e8
kusano 2b45e8
	FADD	f8,  f8,  f12
kusano 2b45e8
	FADD	f9,  f9,  f13
kusano 2b45e8
	FADD	f10, f10, f14
kusano 2b45e8
	FADD	f11, f11, f15
kusano 2b45e8
kusano 2b45e8
	FADD	f0, f0, f8
kusano 2b45e8
	FADD	f1, f1, f9
kusano 2b45e8
	FADD	f2, f2, f10
kusano 2b45e8
	FADD	f3, f3, f11
kusano 2b45e8
kusano 2b45e8
#ifndef XCONJ
kusano 2b45e8
#ifndef CONJ
kusano 2b45e8
	FSUB	f0,  f0,  f3
kusano 2b45e8
	FADD	f1,  f1,  f2
kusano 2b45e8
#else
kusano 2b45e8
	FADD	f0,  f0,  f3
kusano 2b45e8
	FSUB	f1,  f1,  f2
kusano 2b45e8
#endif
kusano 2b45e8
#else
kusano 2b45e8
#ifndef CONJ
kusano 2b45e8
	FADD	f0,  f0,  f3
kusano 2b45e8
	FSUB	f1,  f2,  f1
kusano 2b45e8
#else
kusano 2b45e8
	FSUB	f0,  f0,  f3
kusano 2b45e8
	FADD	f1,  f1,  f2
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	FMADD	f16, f30, f0,  f16
kusano 2b45e8
	FMADDR	f17, f30, f1,  f17
kusano 2b45e8
	FMSUBR	f16, f31, f1,  f16
kusano 2b45e8
	FMADD	f17, f31, f0,  f17
kusano 2b45e8
kusano 2b45e8
	STFD	f16,  0 * SIZE(CO)
kusano 2b45e8
	STFD	f17,  1 * SIZE(CO)
kusano 2b45e8
	add	CO, CO, INCY
kusano 2b45e8
kusano 2b45e8
	addi	J, J, -1
kusano 2b45e8
	cmpi	cr0, 0, J, 0
kusano 2b45e8
	bgt	LL(RemainHead)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(ISEnd):
kusano 2b45e8
	subf	A, PLDA_M, A
kusano 2b45e8
	addi	IS, IS, P
kusano 2b45e8
kusano 2b45e8
	cmp	cr0, 0, IS, M
kusano 2b45e8
	blt	LL(ISLoop)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(End):
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
#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
#endif
kusano 2b45e8
kusano 2b45e8
	addi	SP, SP, STACKSIZE
kusano 2b45e8
kusano 2b45e8
	blr
kusano 2b45e8
kusano 2b45e8
	EPILOGUE
kusano 2b45e8
kusano 2b45e8
#endif