Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/power/zgemm_kernel_power6.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
#ifndef __64BIT__
kusano 2b45e8
#define LOAD	lwz
kusano 2b45e8
#else
kusano 2b45e8
#define LOAD	ld
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef __64BIT__
kusano 2b45e8
#define STACKSIZE 320
kusano 2b45e8
#define ALPHA_R 296(SP)
kusano 2b45e8
#define ALPHA_I 304(SP)
kusano 2b45e8
#define FZERO	312(SP)
kusano 2b45e8
#else
kusano 2b45e8
#define STACKSIZE 256
kusano 2b45e8
#define ALPHA_R 224(SP)
kusano 2b45e8
#define ALPHA_I 232(SP)
kusano 2b45e8
#define FZERO	240(SP)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#define	M	r3
kusano 2b45e8
#define	N	r4
kusano 2b45e8
#define	K	r5
kusano 2b45e8
kusano 2b45e8
#ifdef linux
kusano 2b45e8
#ifndef __64BIT__
kusano 2b45e8
#define A	r6
kusano 2b45e8
#define	B	r7
kusano 2b45e8
#define	C	r8
kusano 2b45e8
#define	LDC	r9
kusano 2b45e8
#define OFFSET	r10
kusano 2b45e8
#else
kusano 2b45e8
#define A	r8
kusano 2b45e8
#define	B	r9
kusano 2b45e8
#define	C	r10
kusano 2b45e8
#define	LDC	r6
kusano 2b45e8
#define OFFSET	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 A	r10
kusano 2b45e8
#define	B	r6
kusano 2b45e8
#define	C	r7
kusano 2b45e8
#define	LDC	r8
kusano 2b45e8
#define OFFSET	r9
kusano 2b45e8
#else
kusano 2b45e8
#define A	r8
kusano 2b45e8
#define	B	r9
kusano 2b45e8
#define	C	r10
kusano 2b45e8
#define	LDC	r6
kusano 2b45e8
#define OFFSET	r7
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#define TEMP	r19
kusano 2b45e8
#define KK	r20
kusano 2b45e8
#define	BB	r21
kusano 2b45e8
#define	I	r22
kusano 2b45e8
#define J	r23
kusano 2b45e8
#define AO	r24
kusano 2b45e8
#define	BO	r25
kusano 2b45e8
#define	CO1	r26
kusano 2b45e8
#define CO2	r27
kusano 2b45e8
#define	CO3	r28
kusano 2b45e8
#define CO4	r29
kusano 2b45e8
kusano 2b45e8
#define PREA	r30
kusano 2b45e8
#define PREC	r31
kusano 2b45e8
	
kusano 2b45e8
#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
kusano 2b45e8
#define FMA1	FMADD
kusano 2b45e8
#define FMA2	FMADD
kusano 2b45e8
#define FMA3	FNMSUB
kusano 2b45e8
#define FMA4	FMADD
kusano 2b45e8
#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
kusano 2b45e8
#define FMA1	FMADD
kusano 2b45e8
#define FMA2	FNMSUB
kusano 2b45e8
#define FMA3	FMADD
kusano 2b45e8
#define FMA4	FMADD
kusano 2b45e8
#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
kusano 2b45e8
#define FMA1	FMADD
kusano 2b45e8
#define FMA2	FMADD
kusano 2b45e8
#define FMA3	FMADD
kusano 2b45e8
#define FMA4	FNMSUB
kusano 2b45e8
#else
kusano 2b45e8
#define FMA1	FMADD
kusano 2b45e8
#define FMA2	FNMSUB
kusano 2b45e8
#define FMA3	FNMSUB
kusano 2b45e8
#define FMA4	FNMSUB
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifndef NEEDPARAM
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	PROFCODE
kusano 2b45e8
kusano 2b45e8
	addi	SP, SP, -STACKSIZE
kusano 2b45e8
	li	r0, 0
kusano 2b45e8
kusano 2b45e8
	stfd	f14,    0(SP)
kusano 2b45e8
	stfd	f15,    8(SP)
kusano 2b45e8
	stfd	f16,   16(SP)
kusano 2b45e8
	stfd	f17,   24(SP)
kusano 2b45e8
kusano 2b45e8
	stfd	f18,   32(SP)
kusano 2b45e8
	stfd	f19,   40(SP)
kusano 2b45e8
	stfd	f20,   48(SP)
kusano 2b45e8
	stfd	f21,   56(SP)
kusano 2b45e8
kusano 2b45e8
	stfd	f22,   64(SP)
kusano 2b45e8
	stfd	f23,   72(SP)
kusano 2b45e8
	stfd	f24,   80(SP)
kusano 2b45e8
	stfd	f25,   88(SP)
kusano 2b45e8
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
kusano 2b45e8
	stfd	f30,  128(SP)
kusano 2b45e8
	stfd	f31,  136(SP)
kusano 2b45e8
kusano 2b45e8
#ifdef __64BIT__
kusano 2b45e8
	std	r31,  144(SP)
kusano 2b45e8
	std	r30,  152(SP)
kusano 2b45e8
	std	r29,  160(SP)
kusano 2b45e8
	std	r28,  168(SP)
kusano 2b45e8
	std	r27,  176(SP)
kusano 2b45e8
	std	r26,  184(SP)
kusano 2b45e8
	std	r25,  192(SP)
kusano 2b45e8
	std	r24,  200(SP)
kusano 2b45e8
	std	r23,  208(SP)
kusano 2b45e8
	std	r22,  216(SP)
kusano 2b45e8
	std	r21,  224(SP)
kusano 2b45e8
#ifdef TRMMKERNEL
kusano 2b45e8
	std	r20,  232(SP)
kusano 2b45e8
	std	r19,  240(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#else
kusano 2b45e8
	stw	r31,  144(SP)
kusano 2b45e8
	stw	r30,  148(SP)
kusano 2b45e8
	stw	r29,  152(SP)
kusano 2b45e8
	stw	r28,  156(SP)
kusano 2b45e8
	stw	r27,  160(SP)
kusano 2b45e8
	stw	r26,  164(SP)
kusano 2b45e8
	stw	r25,  168(SP)
kusano 2b45e8
	stw	r24,  172(SP)
kusano 2b45e8
	stw	r23,  176(SP)
kusano 2b45e8
	stw	r22,  180(SP)
kusano 2b45e8
	stw	r21,  184(SP)
kusano 2b45e8
#ifdef TRMMKERNEL
kusano 2b45e8
	stw	r20,  188(SP)
kusano 2b45e8
	stw	r19,  192(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	stfd	f1,  ALPHA_R
kusano 2b45e8
	stfd	f2,  ALPHA_I
kusano 2b45e8
	stw	r0,  FZERO
kusano 2b45e8
kusano 2b45e8
#ifdef linux
kusano 2b45e8
#ifdef __64BIT__
kusano 2b45e8
	ld	LDC,    112 + STACKSIZE(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(_AIX) || defined(__APPLE__)
kusano 2b45e8
#ifdef __64BIT__
kusano 2b45e8
	ld	LDC,    112 + STACKSIZE(SP)
kusano 2b45e8
#else
kusano 2b45e8
#ifdef DOUBLE
kusano 2b45e8
	lwz	B,       56 + STACKSIZE(SP)
kusano 2b45e8
	lwz	C,       60 + STACKSIZE(SP)
kusano 2b45e8
	lwz	LDC,     64 + STACKSIZE(SP)
kusano 2b45e8
#else
kusano 2b45e8
	lwz	LDC,     56 + STACKSIZE(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef TRMMKERNEL
kusano 2b45e8
#if defined(linux) && defined(__64BIT__)
kusano 2b45e8
	ld	OFFSET,  120 + STACKSIZE(SP)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(_AIX) || defined(__APPLE__)
kusano 2b45e8
#ifdef __64BIT__
kusano 2b45e8
	ld	OFFSET,  120 + STACKSIZE(SP)
kusano 2b45e8
#else
kusano 2b45e8
#ifdef DOUBLE
kusano 2b45e8
	lwz	OFFSET,   68 + STACKSIZE(SP)
kusano 2b45e8
#else
kusano 2b45e8
	lwz	OFFSET,   60 + STACKSIZE(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
#if defined(TRMMKERNEL) && !defined(LEFT)
kusano 2b45e8
	neg	KK, OFFSET
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	slwi	LDC, LDC, ZBASE_SHIFT
kusano 2b45e8
	li	PREA,  (16 * 3) * SIZE
kusano 2b45e8
	li	PREC,   3 * SIZE
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, M, 0
kusano 2b45e8
	ble	LL(999)
kusano 2b45e8
	cmpwi	cr0, N, 0
kusano 2b45e8
	ble	LL(999)
kusano 2b45e8
	cmpwi	cr0, K, 0
kusano 2b45e8
	ble	LL(999)
kusano 2b45e8
kusano 2b45e8
	srawi.	J, N,  2
kusano 2b45e8
	ble	LL(30)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(10):
kusano 2b45e8
	mr	CO1, C
kusano 2b45e8
	add	CO2, C,   LDC
kusano 2b45e8
	add	CO3, CO2, LDC
kusano 2b45e8
	add	CO4, CO3, LDC
kusano 2b45e8
	add	C,   CO4, LDC
kusano 2b45e8
kusano 2b45e8
#if defined(TRMMKERNEL) && defined(LEFT)
kusano 2b45e8
	mr	KK, OFFSET
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	slwi	BB, K, ZBASE_SHIFT + 2
kusano 2b45e8
	mr	AO, A
kusano 2b45e8
kusano 2b45e8
	lfs	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
	srawi.	I, M,  1
kusano 2b45e8
	ble	LL(20)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(11):
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f20,  0 * SIZE(B)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f21,  1 * SIZE(B)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f22,  2 * SIZE(B)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
	LFD	f23,  3 * SIZE(B)
kusano 2b45e8
	LFD	f24,  4 * SIZE(B)
kusano 2b45e8
	LFD	f25,  5 * SIZE(B)
kusano 2b45e8
	LFD	f26,  6 * SIZE(B)
kusano 2b45e8
	LFD	f27,  7 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	dcbtst	CO1, PREC
kusano 2b45e8
	dcbtst	CO2, PREC
kusano 2b45e8
	dcbtst	CO3, PREC
kusano 2b45e8
	dcbtst	CO4, PREC
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  K,  3
kusano 2b45e8
	mr	BO,  B
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(15)
kusano 2b45e8
#else
kusano 2b45e8
#if	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f20,  0 * SIZE(B)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f21,  1 * SIZE(B)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f22,  2 * SIZE(B)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
	LFD	f23,  3 * SIZE(B)
kusano 2b45e8
	LFD	f24,  4 * SIZE(B)
kusano 2b45e8
	LFD	f25,  5 * SIZE(B)
kusano 2b45e8
	LFD	f26,  6 * SIZE(B)
kusano 2b45e8
	LFD	f27,  7 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	mr	BO,  B
kusano 2b45e8
#else
kusano 2b45e8
	slwi	r0,   KK, 1 + ZBASE_SHIFT
kusano 2b45e8
	slwi	TEMP, KK, 2 + ZBASE_SHIFT
kusano 2b45e8
	add	AO, AO, r0
kusano 2b45e8
	add	BO, B,  TEMP
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f20,  0 * SIZE(BO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f21,  1 * SIZE(BO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f22,  2 * SIZE(BO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
	LFD	f23,  3 * SIZE(BO)
kusano 2b45e8
	LFD	f24,  4 * SIZE(BO)
kusano 2b45e8
	LFD	f25,  5 * SIZE(BO)
kusano 2b45e8
	LFD	f26,  6 * SIZE(BO)
kusano 2b45e8
	LFD	f27,  7 * SIZE(BO)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	dcbtst	CO1, PREC
kusano 2b45e8
	dcbtst	CO2, PREC
kusano 2b45e8
	dcbtst	CO3, PREC
kusano 2b45e8
	dcbtst	CO4, PREC
kusano 2b45e8
kusano 2b45e8
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#elif defined(LEFT)
kusano 2b45e8
	addi	TEMP, KK, 2
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, KK, 4
kusano 2b45e8
#endif
kusano 2b45e8
	srawi.	TEMP,  TEMP,  3
kusano 2b45e8
	mtspr	CTR, TEMP
kusano 2b45e8
	ble	LL(15)
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(12):
kusano 2b45e8
	dcbt	AO, PREA
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	nop
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
	LFD	f28,  4 * SIZE(AO)
kusano 2b45e8
	LFD	f29,  5 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f30,  6 * SIZE(AO)
kusano 2b45e8
	LFD	f31,  7 * SIZE(AO)
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA1	f6,  f18, f22, f6
kusano 2b45e8
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA2	f7,  f18, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f16, f24, f8
kusano 2b45e8
	FMA1	f10, f18, f24, f10
kusano 2b45e8
	FMA2	f9,  f16, f25, f9
kusano 2b45e8
	FMA2	f11, f18, f25, f11
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f16, f26, f12
kusano 2b45e8
	FMA1	f14, f18, f26, f14
kusano 2b45e8
	FMA2	f13, f16, f27, f13
kusano 2b45e8
	FMA2	f15, f18, f27, f15
kusano 2b45e8
kusano 2b45e8
	FMA4	f1,  f17, f20, f1
kusano 2b45e8
	FMA4	f3,  f19, f20, f3
kusano 2b45e8
	FMA3	f0,  f17, f21, f0
kusano 2b45e8
	FMA3	f2,  f19, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA4	f5,  f17, f22, f5
kusano 2b45e8
	FMA4	f7,  f19, f22, f7
kusano 2b45e8
	FMA3	f4,  f17, f23, f4
kusano 2b45e8
	FMA3	f6,  f19, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  8 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  9 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 10 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 11 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f24, f9
kusano 2b45e8
	FMA4	f11, f19, f24, f11
kusano 2b45e8
	FMA3	f8,  f17, f25, f8
kusano 2b45e8
	FMA3	f10, f19, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f17, f26, f13
kusano 2b45e8
	FMA4	f15, f19, f26, f15
kusano 2b45e8
	FMA3	f12, f17, f27, f12
kusano 2b45e8
	FMA3	f14, f19, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 12 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 13 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 14 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 15 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f28, f20, f0
kusano 2b45e8
	FMA1	f2,  f30, f20, f2
kusano 2b45e8
	FMA2	f1,  f28, f21, f1
kusano 2b45e8
	FMA2	f3,  f30, f21, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  8 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  9 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 10 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 11 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f28, f22, f4
kusano 2b45e8
	FMA1	f6,  f30, f22, f6
kusano 2b45e8
	FMA2	f5,  f28, f23, f5
kusano 2b45e8
	FMA2	f7,  f30, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f28, f24, f8
kusano 2b45e8
	FMA1	f10, f30, f24, f10
kusano 2b45e8
	FMA2	f9,  f28, f25, f9
kusano 2b45e8
	FMA2	f11, f30, f25, f11
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f28, f26, f12
kusano 2b45e8
	FMA1	f14, f30, f26, f14
kusano 2b45e8
	FMA2	f13, f28, f27, f13
kusano 2b45e8
	FMA2	f15, f30, f27, f15
kusano 2b45e8
kusano 2b45e8
	FMA4	f1,  f29, f20, f1
kusano 2b45e8
	FMA4	f3,  f31, f20, f3
kusano 2b45e8
	FMA3	f0,  f29, f21, f0
kusano 2b45e8
	FMA3	f2,  f31, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA4	f5,  f29, f22, f5
kusano 2b45e8
	FMA4	f7,  f31, f22, f7
kusano 2b45e8
	FMA3	f4,  f29, f23, f4
kusano 2b45e8
	FMA3	f6,  f31, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 16 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 17 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 18 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 19 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f29, f24, f9
kusano 2b45e8
	FMA4	f11, f31, f24, f11
kusano 2b45e8
	FMA3	f8,  f29, f25, f8
kusano 2b45e8
	FMA3	f10, f31, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f29, f26, f13
kusano 2b45e8
	FMA4	f15, f31, f26, f15
kusano 2b45e8
	FMA3	f12, f29, f27, f12
kusano 2b45e8
	FMA3	f14, f31, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 20 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 21 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 22 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 23 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 12 * SIZE(AO)
kusano 2b45e8
	LFD	f29, 13 * SIZE(AO)
kusano 2b45e8
	LFD	f30, 14 * SIZE(AO)
kusano 2b45e8
	LFD	f31, 15 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA1	f6,  f18, f22, f6
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA2	f7,  f18, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f16, f24, f8
kusano 2b45e8
	FMA1	f10, f18, f24, f10
kusano 2b45e8
	FMA2	f9,  f16, f25, f9
kusano 2b45e8
	FMA2	f11, f18, f25, f11
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f16, f26, f12
kusano 2b45e8
	FMA1	f14, f18, f26, f14
kusano 2b45e8
	FMA2	f13, f16, f27, f13
kusano 2b45e8
	FMA2	f15, f18, f27, f15
kusano 2b45e8
kusano 2b45e8
	FMA4	f1,  f17, f20, f1
kusano 2b45e8
	FMA4	f3,  f19, f20, f3
kusano 2b45e8
	FMA3	f0,  f17, f21, f0
kusano 2b45e8
	FMA3	f2,  f19, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA4	f5,  f17, f22, f5
kusano 2b45e8
	FMA4	f7,  f19, f22, f7
kusano 2b45e8
	FMA3	f4,  f17, f23, f4
kusano 2b45e8
	FMA3	f6,  f19, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 24 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 25 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 26 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 27 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f24, f9
kusano 2b45e8
	FMA4	f11, f19, f24, f11
kusano 2b45e8
	FMA3	f8,  f17, f25, f8
kusano 2b45e8
	FMA3	f10, f19, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f17, f26, f13
kusano 2b45e8
	FMA4	f15, f19, f26, f15
kusano 2b45e8
	FMA3	f12, f17, f27, f12
kusano 2b45e8
	FMA3	f14, f19, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 28 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 29 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 30 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 31 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f28, f20, f0
kusano 2b45e8
	FMA1	f2,  f30, f20, f2
kusano 2b45e8
	FMA2	f1,  f28, f21, f1
kusano 2b45e8
	FMA2	f3,  f30, f21, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 16 * SIZE(AO)
kusano 2b45e8
	LFD	f17, 17 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 18 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 19 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f28, f22, f4
kusano 2b45e8
	FMA1	f6,  f30, f22, f6
kusano 2b45e8
	FMA2	f5,  f28, f23, f5
kusano 2b45e8
	FMA2	f7,  f30, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f28, f24, f8
kusano 2b45e8
	FMA1	f10, f30, f24, f10
kusano 2b45e8
	FMA2	f9,  f28, f25, f9
kusano 2b45e8
	FMA2	f11, f30, f25, f11
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f28, f26, f12
kusano 2b45e8
	FMA1	f14, f30, f26, f14
kusano 2b45e8
	FMA2	f13, f28, f27, f13
kusano 2b45e8
	FMA2	f15, f30, f27, f15
kusano 2b45e8
kusano 2b45e8
	FMA4	f1,  f29, f20, f1
kusano 2b45e8
	FMA4	f3,  f31, f20, f3
kusano 2b45e8
	FMA3	f0,  f29, f21, f0
kusano 2b45e8
	FMA3	f2,  f31, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA4	f5,  f29, f22, f5
kusano 2b45e8
	FMA4	f7,  f31, f22, f7
kusano 2b45e8
	FMA3	f4,  f29, f23, f4
kusano 2b45e8
	FMA3	f6,  f31, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 32 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 33 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 34 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 35 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f29, f24, f9
kusano 2b45e8
	FMA4	f11, f31, f24, f11
kusano 2b45e8
	FMA3	f8,  f29, f25, f8
kusano 2b45e8
	FMA3	f10, f31, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f29, f26, f13
kusano 2b45e8
	FMA4	f15, f31, f26, f15
kusano 2b45e8
	FMA3	f12, f29, f27, f12
kusano 2b45e8
	FMA3	f14, f31, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 36 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 37 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 38 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 39 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 20 * SIZE(AO)
kusano 2b45e8
	LFD	f29, 21 * SIZE(AO)
kusano 2b45e8
	LFD	f30, 22 * SIZE(AO)
kusano 2b45e8
	LFD	f31, 23 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA1	f6,  f18, f22, f6
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA2	f7,  f18, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f16, f24, f8
kusano 2b45e8
	FMA1	f10, f18, f24, f10
kusano 2b45e8
	FMA2	f9,  f16, f25, f9
kusano 2b45e8
	FMA2	f11, f18, f25, f11
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f16, f26, f12
kusano 2b45e8
	FMA1	f14, f18, f26, f14
kusano 2b45e8
	FMA2	f13, f16, f27, f13
kusano 2b45e8
	FMA2	f15, f18, f27, f15
kusano 2b45e8
kusano 2b45e8
	FMA4	f1,  f17, f20, f1
kusano 2b45e8
	FMA4	f3,  f19, f20, f3
kusano 2b45e8
	FMA3	f0,  f17, f21, f0
kusano 2b45e8
	FMA3	f2,  f19, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA4	f5,  f17, f22, f5
kusano 2b45e8
	FMA4	f7,  f19, f22, f7
kusano 2b45e8
	FMA3	f4,  f17, f23, f4
kusano 2b45e8
	FMA3	f6,  f19, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 40 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 41 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 42 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 43 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f24, f9
kusano 2b45e8
	FMA4	f11, f19, f24, f11
kusano 2b45e8
	FMA3	f8,  f17, f25, f8
kusano 2b45e8
	FMA3	f10, f19, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f17, f26, f13
kusano 2b45e8
	FMA4	f15, f19, f26, f15
kusano 2b45e8
	FMA3	f12, f17, f27, f12
kusano 2b45e8
	FMA3	f14, f19, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 44 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 45 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 46 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 47 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f28, f20, f0
kusano 2b45e8
	FMA1	f2,  f30, f20, f2
kusano 2b45e8
	FMA2	f1,  f28, f21, f1
kusano 2b45e8
	FMA2	f3,  f30, f21, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 24 * SIZE(AO)
kusano 2b45e8
	LFD	f17, 25 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 26 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 27 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f28, f22, f4
kusano 2b45e8
	FMA1	f6,  f30, f22, f6
kusano 2b45e8
	FMA2	f5,  f28, f23, f5
kusano 2b45e8
	FMA2	f7,  f30, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f28, f24, f8
kusano 2b45e8
	FMA1	f10, f30, f24, f10
kusano 2b45e8
	FMA2	f9,  f28, f25, f9
kusano 2b45e8
	FMA2	f11, f30, f25, f11
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f28, f26, f12
kusano 2b45e8
	FMA1	f14, f30, f26, f14
kusano 2b45e8
	FMA2	f13, f28, f27, f13
kusano 2b45e8
	FMA2	f15, f30, f27, f15
kusano 2b45e8
kusano 2b45e8
	FMA4	f1,  f29, f20, f1
kusano 2b45e8
	FMA4	f3,  f31, f20, f3
kusano 2b45e8
	FMA3	f0,  f29, f21, f0
kusano 2b45e8
	FMA3	f2,  f31, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA4	f5,  f29, f22, f5
kusano 2b45e8
	FMA4	f7,  f31, f22, f7
kusano 2b45e8
	FMA3	f4,  f29, f23, f4
kusano 2b45e8
	FMA3	f6,  f31, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 48 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 49 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 50 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 51 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f29, f24, f9
kusano 2b45e8
	FMA4	f11, f31, f24, f11
kusano 2b45e8
	FMA3	f8,  f29, f25, f8
kusano 2b45e8
	FMA3	f10, f31, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f29, f26, f13
kusano 2b45e8
	FMA4	f15, f31, f26, f15
kusano 2b45e8
	FMA3	f12, f29, f27, f12
kusano 2b45e8
	FMA3	f14, f31, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 52 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 53 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 54 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 55 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 28 * SIZE(AO)
kusano 2b45e8
	LFD	f29, 29 * SIZE(AO)
kusano 2b45e8
	LFD	f30, 30 * SIZE(AO)
kusano 2b45e8
	LFD	f31, 31 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA1	f6,  f18, f22, f6
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA2	f7,  f18, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f16, f24, f8
kusano 2b45e8
	FMA1	f10, f18, f24, f10
kusano 2b45e8
	FMA2	f9,  f16, f25, f9
kusano 2b45e8
	FMA2	f11, f18, f25, f11
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f16, f26, f12
kusano 2b45e8
	FMA1	f14, f18, f26, f14
kusano 2b45e8
	FMA2	f13, f16, f27, f13
kusano 2b45e8
	FMA2	f15, f18, f27, f15
kusano 2b45e8
kusano 2b45e8
	FMA4	f1,  f17, f20, f1
kusano 2b45e8
	FMA4	f3,  f19, f20, f3
kusano 2b45e8
	FMA3	f0,  f17, f21, f0
kusano 2b45e8
	FMA3	f2,  f19, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA4	f5,  f17, f22, f5
kusano 2b45e8
	FMA4	f7,  f19, f22, f7
kusano 2b45e8
	FMA3	f4,  f17, f23, f4
kusano 2b45e8
	FMA3	f6,  f19, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 56 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 57 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 58 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 59 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f24, f9
kusano 2b45e8
	FMA4	f11, f19, f24, f11
kusano 2b45e8
	FMA3	f8,  f17, f25, f8
kusano 2b45e8
	FMA3	f10, f19, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f17, f26, f13
kusano 2b45e8
	FMA4	f15, f19, f26, f15
kusano 2b45e8
	FMA3	f12, f17, f27, f12
kusano 2b45e8
	FMA3	f14, f19, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 60 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 61 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 62 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 63 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f28, f20, f0
kusano 2b45e8
	FMA1	f2,  f30, f20, f2
kusano 2b45e8
	FMA2	f1,  f28, f21, f1
kusano 2b45e8
	FMA2	f3,  f30, f21, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 32 * SIZE(AO)
kusano 2b45e8
	LFD	f17, 33 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 34 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 35 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f28, f22, f4
kusano 2b45e8
	FMA1	f6,  f30, f22, f6
kusano 2b45e8
	FMA2	f5,  f28, f23, f5
kusano 2b45e8
	FMA2	f7,  f30, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f28, f24, f8
kusano 2b45e8
	FMA1	f10, f30, f24, f10
kusano 2b45e8
	FMA2	f9,  f28, f25, f9
kusano 2b45e8
	FMA2	f11, f30, f25, f11
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f28, f26, f12
kusano 2b45e8
	FMA1	f14, f30, f26, f14
kusano 2b45e8
	FMA2	f13, f28, f27, f13
kusano 2b45e8
	FMA2	f15, f30, f27, f15
kusano 2b45e8
kusano 2b45e8
	FMA4	f1,  f29, f20, f1
kusano 2b45e8
	FMA4	f3,  f31, f20, f3
kusano 2b45e8
	FMA3	f0,  f29, f21, f0
kusano 2b45e8
	FMA3	f2,  f31, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA4	f5,  f29, f22, f5
kusano 2b45e8
	FMA4	f7,  f31, f22, f7
kusano 2b45e8
	FMA3	f4,  f29, f23, f4
kusano 2b45e8
	FMA3	f6,  f31, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 64 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 65 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 66 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 67 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f29, f24, f9
kusano 2b45e8
	FMA4	f11, f31, f24, f11
kusano 2b45e8
	FMA3	f8,  f29, f25, f8
kusano 2b45e8
	FMA3	f10, f31, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f29, f26, f13
kusano 2b45e8
	FMA4	f15, f31, f26, f15
kusano 2b45e8
	FMA3	f12, f29, f27, f12
kusano 2b45e8
	FMA3	f14, f31, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 68 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 69 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 70 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 71 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO, AO, 32 * SIZE
kusano 2b45e8
	addi	BO, BO, 64 * SIZE
kusano 2b45e8
	bdnz	LL(12)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(15):
kusano 2b45e8
	lfd	f30,  ALPHA_R
kusano 2b45e8
	lfd	f31,  ALPHA_I
kusano 2b45e8
kusano 2b45e8
	dcbtst	B, BB
kusano 2b45e8
	addi	BB, BB, 16 * SIZE
kusano 2b45e8
	dcbtst	B, BB
kusano 2b45e8
	addi	BB, BB, 16 * SIZE
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	andi.	r0,  K,  7
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(18)
kusano 2b45e8
#else
kusano 2b45e8
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#elif defined(LEFT)
kusano 2b45e8
	addi	TEMP, KK, 2
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, KK, 4
kusano 2b45e8
#endif
kusano 2b45e8
	andi.	TEMP,  TEMP,  7
kusano 2b45e8
	mtspr	CTR, TEMP
kusano 2b45e8
	ble	LL(18)
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(16):
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA1	f6,  f18, f22, f6
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA2	f7,  f18, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f16, f24, f8
kusano 2b45e8
	FMA1	f10, f18, f24, f10
kusano 2b45e8
	FMA2	f9,  f16, f25, f9
kusano 2b45e8
	FMA2	f11, f18, f25, f11
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f16, f26, f12
kusano 2b45e8
	FMA1	f14, f18, f26, f14
kusano 2b45e8
	FMA2	f13, f16, f27, f13
kusano 2b45e8
	FMA2	f15, f18, f27, f15
kusano 2b45e8
kusano 2b45e8
	FMA4	f1,  f17, f20, f1
kusano 2b45e8
	FMA4	f3,  f19, f20, f3
kusano 2b45e8
	FMA3	f0,  f17, f21, f0
kusano 2b45e8
	FMA3	f2,  f19, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA4	f5,  f17, f22, f5
kusano 2b45e8
	FMA4	f7,  f19, f22, f7
kusano 2b45e8
	FMA3	f4,  f17, f23, f4
kusano 2b45e8
	FMA3	f6,  f19, f23, f6
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f24, f9
kusano 2b45e8
	FMA4	f11, f19, f24, f11
kusano 2b45e8
	FMA3	f8,  f17, f25, f8
kusano 2b45e8
	FMA3	f10, f19, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f17, f26, f13
kusano 2b45e8
	FMA4	f15, f19, f26, f15
kusano 2b45e8
	FMA3	f12, f17, f27, f12
kusano 2b45e8
	FMA3	f14, f19, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  4 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  5 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  6 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  7 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  8 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  9 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 10 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 11 * SIZE(BO)
kusano 2b45e8
	LFD	f24, 12 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 13 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 14 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 15 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO, AO,  4 * SIZE
kusano 2b45e8
	addi	BO, BO,  8 * SIZE
kusano 2b45e8
	bdnz	LL(16)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(18):
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 0 * SIZE(CO1)
kusano 2b45e8
	LFD	f17, 1 * SIZE(CO1)
kusano 2b45e8
	LFD	f18, 2 * SIZE(CO1)
kusano 2b45e8
	LFD	f19, 3 * SIZE(CO1)
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 0 * SIZE(CO2)
kusano 2b45e8
	LFD	f21, 1 * SIZE(CO2)
kusano 2b45e8
	LFD	f22, 2 * SIZE(CO2)
kusano 2b45e8
	LFD	f23, 3 * SIZE(CO2)
kusano 2b45e8
kusano 2b45e8
	FNMSUB	f24, f31, f1, f16
kusano 2b45e8
	FMADD	f25, f31, f0, f17
kusano 2b45e8
	FNMSUB	f26, f31, f3, f18
kusano 2b45e8
	FMADD	f27, f31, f2, f19
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 0 * SIZE(CO3)
kusano 2b45e8
	LFD	f17, 1 * SIZE(CO3)
kusano 2b45e8
	LFD	f18, 2 * SIZE(CO3)
kusano 2b45e8
	LFD	f19, 3 * SIZE(CO3)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f30, f0, f24
kusano 2b45e8
	FMADD	f1,  f30, f1, f25
kusano 2b45e8
	FMADD	f2,  f30, f2, f26
kusano 2b45e8
	FMADD	f3,  f30, f3, f27
kusano 2b45e8
kusano 2b45e8
	FNMSUB	f24, f31, f5, f20
kusano 2b45e8
	FMADD	f25, f31, f4, f21
kusano 2b45e8
	FNMSUB	f26, f31, f7, f22
kusano 2b45e8
	FMADD	f27, f31, f6, f23
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 0 * SIZE(CO4)
kusano 2b45e8
	LFD	f21, 1 * SIZE(CO4)
kusano 2b45e8
	LFD	f22, 2 * SIZE(CO4)
kusano 2b45e8
	LFD	f23, 3 * SIZE(CO4)
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f30, f4,  f24
kusano 2b45e8
	FMADD	f5,  f30, f5,  f25
kusano 2b45e8
	FMADD	f6,  f30, f6,  f26
kusano 2b45e8
	FMADD	f7,  f30, f7,  f27
kusano 2b45e8
kusano 2b45e8
	FNMSUB	f24, f31, f9,  f16
kusano 2b45e8
	FMADD	f25, f31, f8,  f17
kusano 2b45e8
	FNMSUB	f26, f31, f11, f18
kusano 2b45e8
	FMADD	f27, f31, f10, f19
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f30, f8,  f24
kusano 2b45e8
	FMADD	f9,  f30, f9,  f25
kusano 2b45e8
	FMADD	f10, f30, f10, f26
kusano 2b45e8
	FMADD	f11, f30, f11, f27
kusano 2b45e8
kusano 2b45e8
	FNMSUB	f24, f31, f13, f20
kusano 2b45e8
	FMADD	f25, f31, f12, f21
kusano 2b45e8
	FNMSUB	f26, f31, f15, f22
kusano 2b45e8
	FMADD	f27, f31, f14, f23
kusano 2b45e8
kusano 2b45e8
	FMADD	f12,  f30, f12, f24
kusano 2b45e8
	FMADD	f13,  f30, f13, f25
kusano 2b45e8
	FMADD	f14,  f30, f14, f26
kusano 2b45e8
	FMADD	f15,  f30, f15, f27
kusano 2b45e8
kusano 2b45e8
#else
kusano 2b45e8
kusano 2b45e8
	FMUL	f16, f31, f1
kusano 2b45e8
	FMUL	f17, f31, f0
kusano 2b45e8
	FMUL	f18, f31, f3
kusano 2b45e8
	FMUL	f19, f31, f2
kusano 2b45e8
kusano 2b45e8
	FMUL	f20, f31, f5
kusano 2b45e8
	FMUL	f21, f31, f4
kusano 2b45e8
	FMUL	f22, f31, f7
kusano 2b45e8
	FMUL	f23, f31, f6
kusano 2b45e8
kusano 2b45e8
	FMSUB	f0,  f30, f0, f16
kusano 2b45e8
	FMADD	f1,  f30, f1, f17
kusano 2b45e8
	FMADD	f2,  f30, f2, f18
kusano 2b45e8
	FMADD	f3,  f30, f3, f19
kusano 2b45e8
kusano 2b45e8
	FMSUB	f4,  f30, f4, f20
kusano 2b45e8
	FMADD	f5,  f30, f5, f21
kusano 2b45e8
	FMADD	f6,  f30, f6, f22
kusano 2b45e8
	FMADD	f7,  f30, f7, f23
kusano 2b45e8
kusano 2b45e8
	FMUL	f16, f31, f9
kusano 2b45e8
	FMUL	f17, f31, f8
kusano 2b45e8
	FMUL	f18, f31, f11
kusano 2b45e8
	FMUL	f19, f31, f10
kusano 2b45e8
kusano 2b45e8
	FMUL	f20, f31, f13
kusano 2b45e8
	FMUL	f21, f31, f12
kusano 2b45e8
	FMUL	f22, f31, f15
kusano 2b45e8
	FMUL	f23, f31, f14
kusano 2b45e8
kusano 2b45e8
	FMSUB	f8,  f30, f8,  f16
kusano 2b45e8
	FMADD	f9,  f30, f9,  f17
kusano 2b45e8
	FMADD	f10, f30, f10, f18
kusano 2b45e8
	FMADD	f11, f30, f11, f19
kusano 2b45e8
kusano 2b45e8
	FMSUB	f12, f30, f12, f20
kusano 2b45e8
	FMADD	f13, f30, f13, f21
kusano 2b45e8
	FMADD	f14, f30, f14, f22
kusano 2b45e8
	FMADD	f15, f30, f15, f23
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	STFD	f0,  0 * SIZE(CO1)
kusano 2b45e8
	STFD	f1,  1 * SIZE(CO1)
kusano 2b45e8
	STFD	f2,  2 * SIZE(CO1)
kusano 2b45e8
	STFD	f3,  3 * SIZE(CO1)
kusano 2b45e8
kusano 2b45e8
	lfs	f0,  FZERO
kusano 2b45e8
 	fmr	f1,  f0
kusano 2b45e8
	fmr	f2,  f0
kusano 2b45e8
	fmr	f3,  f0
kusano 2b45e8
kusano 2b45e8
	STFD	f4,  0 * SIZE(CO2)
kusano 2b45e8
	STFD	f5,  1 * SIZE(CO2)
kusano 2b45e8
	STFD	f6,  2 * SIZE(CO2)
kusano 2b45e8
	STFD	f7,  3 * SIZE(CO2)
kusano 2b45e8
kusano 2b45e8
	fmr	f4,  f0
kusano 2b45e8
	fmr	f5,  f0
kusano 2b45e8
	fmr	f6,  f0
kusano 2b45e8
	fmr	f7,  f0
kusano 2b45e8
kusano 2b45e8
	STFD	f8,  0 * SIZE(CO3)
kusano 2b45e8
	STFD	f9,  1 * SIZE(CO3)
kusano 2b45e8
	STFD	f10, 2 * SIZE(CO3)
kusano 2b45e8
	STFD	f11, 3 * SIZE(CO3)
kusano 2b45e8
kusano 2b45e8
	fmr	f8,  f0
kusano 2b45e8
	fmr	f9,  f0
kusano 2b45e8
	fmr	f10, f0
kusano 2b45e8
	fmr	f11, f0
kusano 2b45e8
kusano 2b45e8
	STFD	f12, 0 * SIZE(CO4)
kusano 2b45e8
	STFD	f13, 1 * SIZE(CO4)
kusano 2b45e8
	STFD	f14, 2 * SIZE(CO4)
kusano 2b45e8
	STFD	f15, 3 * SIZE(CO4)
kusano 2b45e8
kusano 2b45e8
	fmr	f12, f0
kusano 2b45e8
	fmr	f13, f0
kusano 2b45e8
	fmr	f14, f0
kusano 2b45e8
	fmr	f15, f0
kusano 2b45e8
kusano 2b45e8
	addi	CO1, CO1, 4 * SIZE
kusano 2b45e8
	addi	CO2, CO2, 4 * SIZE
kusano 2b45e8
	addi	CO3, CO3, 4 * SIZE
kusano 2b45e8
	addi	CO4, CO4, 4 * SIZE
kusano 2b45e8
	
kusano 2b45e8
#ifdef TRMMKERNEL
kusano 2b45e8
#if ( defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
    (!defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addi	TEMP, TEMP, -2
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, TEMP, -4
kusano 2b45e8
#endif
kusano 2b45e8
	slwi	r0,   TEMP, 1 + ZBASE_SHIFT
kusano 2b45e8
	slwi	TEMP, TEMP, 2 + ZBASE_SHIFT
kusano 2b45e8
	add	AO, AO, r0
kusano 2b45e8
	add	BO, BO, TEMP
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addi	KK, KK, 2
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	addic.	I, I, -1
kusano 2b45e8
	bgt	LL(11)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(20):
kusano 2b45e8
	andi.	I,  M,  1
kusano 2b45e8
	ble	LL(29)
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(B)
kusano 2b45e8
	LFD	f21,  1 * SIZE(B)
kusano 2b45e8
	LFD	f22,  2 * SIZE(B)
kusano 2b45e8
	LFD	f23,  3 * SIZE(B)
kusano 2b45e8
	LFD	f24,  4 * SIZE(B)
kusano 2b45e8
	LFD	f25,  5 * SIZE(B)
kusano 2b45e8
	LFD	f26,  6 * SIZE(B)
kusano 2b45e8
	LFD	f27,  7 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  K,  2
kusano 2b45e8
	mr	BO,  B
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(25)
kusano 2b45e8
#else
kusano 2b45e8
#if	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(B)
kusano 2b45e8
	LFD	f21,  1 * SIZE(B)
kusano 2b45e8
	LFD	f22,  2 * SIZE(B)
kusano 2b45e8
	LFD	f23,  3 * SIZE(B)
kusano 2b45e8
	LFD	f24,  4 * SIZE(B)
kusano 2b45e8
	LFD	f25,  5 * SIZE(B)
kusano 2b45e8
	LFD	f26,  6 * SIZE(B)
kusano 2b45e8
	LFD	f27,  7 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	mr	BO,  B
kusano 2b45e8
#else
kusano 2b45e8
	slwi	r0,   KK, 0 + ZBASE_SHIFT
kusano 2b45e8
	slwi	TEMP, KK, 2 + ZBASE_SHIFT
kusano 2b45e8
	add	AO, AO, r0
kusano 2b45e8
	add	BO, B,  TEMP
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  1 * SIZE(BO)
kusano 2b45e8
	LFD	f22,  2 * SIZE(BO)
kusano 2b45e8
	LFD	f23,  3 * SIZE(BO)
kusano 2b45e8
	LFD	f24,  4 * SIZE(BO)
kusano 2b45e8
	LFD	f25,  5 * SIZE(BO)
kusano 2b45e8
	LFD	f26,  6 * SIZE(BO)
kusano 2b45e8
	LFD	f27,  7 * SIZE(BO)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#elif defined(LEFT)
kusano 2b45e8
	addi	TEMP, KK, 1
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, KK, 4
kusano 2b45e8
#endif
kusano 2b45e8
	srawi.	TEMP,  TEMP,  2
kusano 2b45e8
	mtspr	CTR, TEMP
kusano 2b45e8
	ble	LL(25)
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(22):
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA4	f3,  f17, f20, f3
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA3	f2,  f17, f21, f2
kusano 2b45e8
kusano 2b45e8
	LFD	f28,  4 * SIZE(AO)
kusano 2b45e8
	LFD	f29,  5 * SIZE(AO)
kusano 2b45e8
	LFD	f30,  6 * SIZE(AO)
kusano 2b45e8
	LFD	f31,  7 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA4	f7,  f17, f22, f7
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA3	f6,  f17, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  8 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  9 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 10 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 11 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f16, f24, f8
kusano 2b45e8
	FMA4	f11, f17, f24, f11
kusano 2b45e8
	FMA2	f9,  f16, f25, f9
kusano 2b45e8
	FMA3	f10, f17, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f16, f26, f12
kusano 2b45e8
	FMA4	f15, f17, f26, f15
kusano 2b45e8
	FMA2	f13, f16, f27, f13
kusano 2b45e8
	FMA3	f14, f17, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 12 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 13 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 14 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 15 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f18, f20, f0
kusano 2b45e8
	FMA4	f3,  f19, f20, f3
kusano 2b45e8
	FMA2	f1,  f18, f21, f1
kusano 2b45e8
	FMA3	f2,  f19, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f18, f22, f4
kusano 2b45e8
	FMA4	f7,  f19, f22, f7
kusano 2b45e8
	FMA2	f5,  f18, f23, f5
kusano 2b45e8
	FMA3	f6,  f19, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 16 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 17 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 18 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 19 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f18, f24, f8
kusano 2b45e8
	FMA4	f11, f19, f24, f11
kusano 2b45e8
	FMA2	f9,  f18, f25, f9
kusano 2b45e8
	FMA3	f10, f19, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f18, f26, f12
kusano 2b45e8
	FMA4	f15, f19, f26, f15
kusano 2b45e8
	FMA2	f13, f18, f27, f13
kusano 2b45e8
	FMA3	f14, f19, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 20 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 21 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 22 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 23 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f28, f20, f0
kusano 2b45e8
	FMA4	f3,  f29, f20, f3
kusano 2b45e8
	FMA2	f1,  f28, f21, f1
kusano 2b45e8
	FMA3	f2,  f29, f21, f2
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  8 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  9 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 10 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 11 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f28, f22, f4
kusano 2b45e8
	FMA4	f7,  f29, f22, f7
kusano 2b45e8
	FMA2	f5,  f28, f23, f5
kusano 2b45e8
	FMA3	f6,  f29, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 24 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 25 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 26 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 27 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f28, f24, f8
kusano 2b45e8
	FMA4	f11, f29, f24, f11
kusano 2b45e8
	FMA2	f9,  f28, f25, f9
kusano 2b45e8
	FMA3	f10, f29, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f28, f26, f12
kusano 2b45e8
	FMA4	f15, f29, f26, f15
kusano 2b45e8
	FMA2	f13, f28, f27, f13
kusano 2b45e8
	FMA3	f14, f29, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 28 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 29 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 30 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 31 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f30, f20, f0
kusano 2b45e8
	FMA4	f3,  f31, f20, f3
kusano 2b45e8
	FMA2	f1,  f30, f21, f1
kusano 2b45e8
	FMA3	f2,  f31, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f30, f22, f4
kusano 2b45e8
	FMA4	f7,  f31, f22, f7
kusano 2b45e8
	FMA2	f5,  f30, f23, f5
kusano 2b45e8
	FMA3	f6,  f31, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 32 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 33 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 34 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 35 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f30, f24, f8
kusano 2b45e8
	FMA4	f11, f31, f24, f11
kusano 2b45e8
	FMA2	f9,  f30, f25, f9
kusano 2b45e8
	FMA3	f10, f31, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f30, f26, f12
kusano 2b45e8
	FMA4	f15, f31, f26, f15
kusano 2b45e8
	FMA2	f13, f30, f27, f13
kusano 2b45e8
	FMA3	f14, f31, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 36 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 37 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 38 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 39 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO, AO,  8 * SIZE
kusano 2b45e8
	addi	BO, BO, 32 * SIZE
kusano 2b45e8
kusano 2b45e8
	bdnz	LL(22)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(25):
kusano 2b45e8
	lfd	f30,  ALPHA_R
kusano 2b45e8
	lfd	f31,  ALPHA_I
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	andi.	r0,  K,  3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(28)
kusano 2b45e8
#else
kusano 2b45e8
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#elif defined(LEFT)
kusano 2b45e8
	addi	TEMP, KK, 1
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, KK, 4
kusano 2b45e8
#endif
kusano 2b45e8
	andi.	TEMP,  TEMP,  3
kusano 2b45e8
	mtspr	CTR, TEMP
kusano 2b45e8
	ble	LL(28)
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(26):
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA4	f3,  f17, f20, f3
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA3	f2,  f17, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA4	f7,  f17, f22, f7
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA3	f6,  f17, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  8 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  9 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 10 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 11 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f8,  f16, f24, f8
kusano 2b45e8
	FMA4	f11, f17, f24, f11
kusano 2b45e8
	FMA2	f9,  f16, f25, f9
kusano 2b45e8
	FMA3	f10, f17, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA1	f12, f16, f26, f12
kusano 2b45e8
	FMA4	f15, f17, f26, f15
kusano 2b45e8
	FMA2	f13, f16, f27, f13
kusano 2b45e8
	FMA3	f14, f17, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 12 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 13 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 14 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 15 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO, AO,  2 * SIZE
kusano 2b45e8
	addi	BO, BO,  8 * SIZE
kusano 2b45e8
	bdnz	LL(26)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(28):
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 0 * SIZE(CO1)
kusano 2b45e8
	LFD	f17, 1 * SIZE(CO1)
kusano 2b45e8
	LFD	f18, 0 * SIZE(CO2)
kusano 2b45e8
	LFD	f19, 1 * SIZE(CO2)
kusano 2b45e8
kusano 2b45e8
	FADD	f0,  f0,  f2
kusano 2b45e8
	FADD	f1,  f1,  f3
kusano 2b45e8
	FADD	f4,  f4,  f6
kusano 2b45e8
	FADD	f5,  f5,  f7
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 0 * SIZE(CO3)
kusano 2b45e8
	LFD	f21, 1 * SIZE(CO3)
kusano 2b45e8
	LFD	f22, 0 * SIZE(CO4)
kusano 2b45e8
	LFD	f23, 1 * SIZE(CO4)
kusano 2b45e8
kusano 2b45e8
	FADD	f8,  f8,  f10
kusano 2b45e8
	FADD	f9,  f9,  f11
kusano 2b45e8
	FADD	f12, f12, f14
kusano 2b45e8
	FADD	f13, f13, f15
kusano 2b45e8
kusano 2b45e8
	FNMSUB	f24, f31, f1, f16
kusano 2b45e8
	FMADD	f25, f31, f0, f17
kusano 2b45e8
	FNMSUB	f26, f31, f5, f18
kusano 2b45e8
	FMADD	f27, f31, f4, f19
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f30, f0, f24
kusano 2b45e8
	FMADD	f1,  f30, f1, f25
kusano 2b45e8
	FMADD	f4,  f30, f4,  f26
kusano 2b45e8
	FMADD	f5,  f30, f5,  f27
kusano 2b45e8
kusano 2b45e8
	FNMSUB	f24, f31, f9,  f20
kusano 2b45e8
	FMADD	f25, f31, f8,  f21
kusano 2b45e8
	FNMSUB	f26, f31, f13, f22
kusano 2b45e8
	FMADD	f27, f31, f12, f23
kusano 2b45e8
kusano 2b45e8
	FMADD	f8,  f30, f8,  f24
kusano 2b45e8
	FMADD	f9,  f30, f9,  f25
kusano 2b45e8
	FMADD	f12,  f30, f12, f26
kusano 2b45e8
	FMADD	f13,  f30, f13, f27
kusano 2b45e8
kusano 2b45e8
#else
kusano 2b45e8
	FADD	f0,  f0,  f2
kusano 2b45e8
	FADD	f1,  f1,  f3
kusano 2b45e8
	FADD	f4,  f4,  f6
kusano 2b45e8
	FADD	f5,  f5,  f7
kusano 2b45e8
kusano 2b45e8
	FMUL	f16, f31, f1
kusano 2b45e8
	FMUL	f17, f31, f0
kusano 2b45e8
	FMUL	f18, f31, f5
kusano 2b45e8
	FMUL	f19, f31, f4
kusano 2b45e8
kusano 2b45e8
	FMSUB	f0,  f30, f0, f16
kusano 2b45e8
	FMADD	f1,  f30, f1, f17
kusano 2b45e8
	FMSUB	f4,  f30, f4, f18
kusano 2b45e8
	FMADD	f5,  f30, f5, f19
kusano 2b45e8
kusano 2b45e8
	FADD	f8,  f8,  f10
kusano 2b45e8
	FADD	f9,  f9,  f11
kusano 2b45e8
	FADD	f12, f12, f14
kusano 2b45e8
	FADD	f13, f13, f15
kusano 2b45e8
kusano 2b45e8
	FMUL	f20, f31, f9
kusano 2b45e8
	FMUL	f21, f31, f8
kusano 2b45e8
	FMUL	f22, f31, f13
kusano 2b45e8
	FMUL	f23, f31, f12
kusano 2b45e8
kusano 2b45e8
	FMSUB	f8,  f30, f8,  f20
kusano 2b45e8
	FMADD	f9,  f30, f9,  f21
kusano 2b45e8
	FMSUB	f12, f30, f12, f22
kusano 2b45e8
	FMADD	f13, f30, f13, f23
kusano 2b45e8
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	STFD	f0,  0 * SIZE(CO1)
kusano 2b45e8
	STFD	f1,  1 * SIZE(CO1)
kusano 2b45e8
	STFD	f4,  0 * SIZE(CO2)
kusano 2b45e8
	STFD	f5,  1 * SIZE(CO2)
kusano 2b45e8
kusano 2b45e8
	STFD	f8,  0 * SIZE(CO3)
kusano 2b45e8
	STFD	f9,  1 * SIZE(CO3)
kusano 2b45e8
	STFD	f12, 0 * SIZE(CO4)
kusano 2b45e8
	STFD	f13, 1 * SIZE(CO4)
kusano 2b45e8
kusano 2b45e8
	addi	CO1, CO1, 2 * SIZE
kusano 2b45e8
	addi	CO2, CO2, 2 * SIZE
kusano 2b45e8
	addi	CO3, CO3, 2 * SIZE
kusano 2b45e8
	addi	CO4, CO4, 2 * SIZE
kusano 2b45e8
	
kusano 2b45e8
#ifdef TRMMKERNEL
kusano 2b45e8
#if ( defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
    (!defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addi	TEMP, TEMP, -1
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, TEMP, -4
kusano 2b45e8
#endif
kusano 2b45e8
	slwi	r0,   TEMP, 0 + ZBASE_SHIFT
kusano 2b45e8
	slwi	TEMP, TEMP, 2 + ZBASE_SHIFT
kusano 2b45e8
	add	AO, AO, r0
kusano 2b45e8
	add	BO, BO, TEMP
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addi	KK, KK, 1
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(29):
kusano 2b45e8
#if defined(TRMMKERNEL) && !defined(LEFT)
kusano 2b45e8
	addi	KK, KK, 4
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	mr	B,  BO
kusano 2b45e8
kusano 2b45e8
	addic.	J, J, -1
kusano 2b45e8
	bgt	LL(10)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(30):
kusano 2b45e8
	andi.	J, N,  2
kusano 2b45e8
	ble	LL(50)
kusano 2b45e8
kusano 2b45e8
	mr	CO1, C
kusano 2b45e8
	add	CO2, C,   LDC
kusano 2b45e8
	add	C,   CO2, LDC
kusano 2b45e8
kusano 2b45e8
#if defined(TRMMKERNEL) && defined(LEFT)
kusano 2b45e8
	mr	KK, OFFSET
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	slwi	BB, K, ZBASE_SHIFT + 1
kusano 2b45e8
	mr	AO, A
kusano 2b45e8
kusano 2b45e8
	lfs	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
	srawi.	I, M,  1
kusano 2b45e8
	ble	LL(40)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(31):
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(B)
kusano 2b45e8
	LFD	f21,  1 * SIZE(B)
kusano 2b45e8
	LFD	f22,  2 * SIZE(B)
kusano 2b45e8
	LFD	f23,  3 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	LFD	f24,  4 * SIZE(B)
kusano 2b45e8
	LFD	f25,  5 * SIZE(B)
kusano 2b45e8
	LFD	f26,  6 * SIZE(B)
kusano 2b45e8
	LFD	f27,  7 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	dcbtst	CO1, PREC
kusano 2b45e8
	dcbtst	CO2, PREC
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  K,  3
kusano 2b45e8
	mr	BO,  B
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(35)
kusano 2b45e8
#else
kusano 2b45e8
#if	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(B)
kusano 2b45e8
	LFD	f21,  1 * SIZE(B)
kusano 2b45e8
	LFD	f22,  2 * SIZE(B)
kusano 2b45e8
	LFD	f23,  3 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	LFD	f24,  4 * SIZE(B)
kusano 2b45e8
	LFD	f25,  5 * SIZE(B)
kusano 2b45e8
	LFD	f26,  6 * SIZE(B)
kusano 2b45e8
	LFD	f27,  7 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	mr	BO,  B
kusano 2b45e8
#else
kusano 2b45e8
	slwi	r0,   KK, 1 + ZBASE_SHIFT
kusano 2b45e8
	slwi	TEMP, KK, 1 + ZBASE_SHIFT
kusano 2b45e8
	add	AO, AO, r0
kusano 2b45e8
	add	BO, B,  TEMP
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  1 * SIZE(BO)
kusano 2b45e8
	LFD	f22,  2 * SIZE(BO)
kusano 2b45e8
	LFD	f23,  3 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	LFD	f24,  4 * SIZE(BO)
kusano 2b45e8
	LFD	f25,  5 * SIZE(BO)
kusano 2b45e8
	LFD	f26,  6 * SIZE(BO)
kusano 2b45e8
	LFD	f27,  7 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	dcbtst	CO1, PREC
kusano 2b45e8
	dcbtst	CO2, PREC
kusano 2b45e8
kusano 2b45e8
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#elif defined(LEFT)
kusano 2b45e8
	addi	TEMP, KK, 2
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, KK, 2
kusano 2b45e8
#endif
kusano 2b45e8
	srawi.	TEMP,  TEMP,  3
kusano 2b45e8
	mtspr	CTR, TEMP
kusano 2b45e8
	ble	LL(35)
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(32):
kusano 2b45e8
	dcbt	AO, PREA
kusano 2b45e8
	dcbtst	BO, PREA
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f28,  4 * SIZE(AO)
kusano 2b45e8
	LFD	f29,  5 * SIZE(AO)
kusano 2b45e8
	LFD	f30,  6 * SIZE(AO)
kusano 2b45e8
	LFD	f31,  7 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA1	f6,  f18, f22, f6
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA2	f7,  f18, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f20, f9
kusano 2b45e8
	FMA4	f11, f19, f20, f11
kusano 2b45e8
	FMA3	f8,  f17, f21, f8
kusano 2b45e8
	FMA3	f10, f19, f21, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f17, f22, f13
kusano 2b45e8
	FMA4	f15, f19, f22, f15
kusano 2b45e8
	FMA3	f12, f17, f23, f12
kusano 2b45e8
	FMA3	f14, f19, f23, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  8 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  9 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 10 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 11 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f28, f24, f0
kusano 2b45e8
	FMA1	f2,  f30, f24, f2
kusano 2b45e8
	FMA2	f1,  f28, f25, f1
kusano 2b45e8
	FMA2	f3,  f30, f25, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  8 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  9 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 10 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 11 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f28, f26, f4
kusano 2b45e8
	FMA1	f6,  f30, f26, f6
kusano 2b45e8
	FMA2	f5,  f28, f27, f5
kusano 2b45e8
	FMA2	f7,  f30, f27, f7
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f29, f24, f9
kusano 2b45e8
	FMA4	f11, f31, f24, f11
kusano 2b45e8
	FMA3	f8,  f29, f25, f8
kusano 2b45e8
	FMA3	f10, f31, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f29, f26, f13
kusano 2b45e8
	FMA4	f15, f31, f26, f15
kusano 2b45e8
	FMA3	f12, f29, f27, f12
kusano 2b45e8
	FMA3	f14, f31, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 12 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 13 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 14 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 15 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 12 * SIZE(AO)
kusano 2b45e8
	LFD	f29, 13 * SIZE(AO)
kusano 2b45e8
	LFD	f30, 14 * SIZE(AO)
kusano 2b45e8
	LFD	f31, 15 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA1	f6,  f18, f22, f6
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA2	f7,  f18, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f20, f9
kusano 2b45e8
	FMA4	f11, f19, f20, f11
kusano 2b45e8
	FMA3	f8,  f17, f21, f8
kusano 2b45e8
	FMA3	f10, f19, f21, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f17, f22, f13
kusano 2b45e8
	FMA4	f15, f19, f22, f15
kusano 2b45e8
	FMA3	f12, f17, f23, f12
kusano 2b45e8
	FMA3	f14, f19, f23, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 16 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 17 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 18 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 19 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f28, f24, f0
kusano 2b45e8
	FMA1	f2,  f30, f24, f2
kusano 2b45e8
	FMA2	f1,  f28, f25, f1
kusano 2b45e8
	FMA2	f3,  f30, f25, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 16 * SIZE(AO)
kusano 2b45e8
	LFD	f17, 17 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 18 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 19 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f28, f26, f4
kusano 2b45e8
	FMA1	f6,  f30, f26, f6
kusano 2b45e8
	FMA2	f5,  f28, f27, f5
kusano 2b45e8
	FMA2	f7,  f30, f27, f7
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f29, f24, f9
kusano 2b45e8
	FMA4	f11, f31, f24, f11
kusano 2b45e8
	FMA3	f8,  f29, f25, f8
kusano 2b45e8
	FMA3	f10, f31, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f29, f26, f13
kusano 2b45e8
	FMA4	f15, f31, f26, f15
kusano 2b45e8
	FMA3	f12, f29, f27, f12
kusano 2b45e8
	FMA3	f14, f31, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 20 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 21 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 22 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 23 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 20 * SIZE(AO)
kusano 2b45e8
	LFD	f29, 21 * SIZE(AO)
kusano 2b45e8
	LFD	f30, 22 * SIZE(AO)
kusano 2b45e8
	LFD	f31, 23 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA1	f6,  f18, f22, f6
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA2	f7,  f18, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f20, f9
kusano 2b45e8
	FMA4	f11, f19, f20, f11
kusano 2b45e8
	FMA3	f8,  f17, f21, f8
kusano 2b45e8
	FMA3	f10, f19, f21, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f17, f22, f13
kusano 2b45e8
	FMA4	f15, f19, f22, f15
kusano 2b45e8
	FMA3	f12, f17, f23, f12
kusano 2b45e8
	FMA3	f14, f19, f23, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 24 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 25 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 26 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 27 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f28, f24, f0
kusano 2b45e8
	FMA1	f2,  f30, f24, f2
kusano 2b45e8
	FMA2	f1,  f28, f25, f1
kusano 2b45e8
	FMA2	f3,  f30, f25, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 24 * SIZE(AO)
kusano 2b45e8
	LFD	f17, 25 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 26 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 27 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f28, f26, f4
kusano 2b45e8
	FMA1	f6,  f30, f26, f6
kusano 2b45e8
	FMA2	f5,  f28, f27, f5
kusano 2b45e8
	FMA2	f7,  f30, f27, f7
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f29, f24, f9
kusano 2b45e8
	FMA4	f11, f31, f24, f11
kusano 2b45e8
	FMA3	f8,  f29, f25, f8
kusano 2b45e8
	FMA3	f10, f31, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f29, f26, f13
kusano 2b45e8
	FMA4	f15, f31, f26, f15
kusano 2b45e8
	FMA3	f12, f29, f27, f12
kusano 2b45e8
	FMA3	f14, f31, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 28 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 29 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 30 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 31 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f28, 28 * SIZE(AO)
kusano 2b45e8
	LFD	f29, 29 * SIZE(AO)
kusano 2b45e8
	LFD	f30, 30 * SIZE(AO)
kusano 2b45e8
	LFD	f31, 31 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA1	f6,  f18, f22, f6
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA2	f7,  f18, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f20, f9
kusano 2b45e8
	FMA4	f11, f19, f20, f11
kusano 2b45e8
	FMA3	f8,  f17, f21, f8
kusano 2b45e8
	FMA3	f10, f19, f21, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f17, f22, f13
kusano 2b45e8
	FMA4	f15, f19, f22, f15
kusano 2b45e8
	FMA3	f12, f17, f23, f12
kusano 2b45e8
	FMA3	f14, f19, f23, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 32 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 33 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 34 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 35 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f28, f24, f0
kusano 2b45e8
	FMA1	f2,  f30, f24, f2
kusano 2b45e8
	FMA2	f1,  f28, f25, f1
kusano 2b45e8
	FMA2	f3,  f30, f25, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 32 * SIZE(AO)
kusano 2b45e8
	LFD	f17, 33 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 34 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 35 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f28, f26, f4
kusano 2b45e8
	FMA1	f6,  f30, f26, f6
kusano 2b45e8
	FMA2	f5,  f28, f27, f5
kusano 2b45e8
	FMA2	f7,  f30, f27, f7
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f29, f24, f9
kusano 2b45e8
	FMA4	f11, f31, f24, f11
kusano 2b45e8
	FMA3	f8,  f29, f25, f8
kusano 2b45e8
	FMA3	f10, f31, f25, f10
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f29, f26, f13
kusano 2b45e8
	FMA4	f15, f31, f26, f15
kusano 2b45e8
	FMA3	f12, f29, f27, f12
kusano 2b45e8
	FMA3	f14, f31, f27, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f24, 36 * SIZE(BO)
kusano 2b45e8
	LFD	f25, 37 * SIZE(BO)
kusano 2b45e8
	LFD	f26, 38 * SIZE(BO)
kusano 2b45e8
	LFD	f27, 39 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO, AO, 32 * SIZE
kusano 2b45e8
	addi	BO, BO, 32 * SIZE
kusano 2b45e8
kusano 2b45e8
	bdnz	LL(32)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(35):
kusano 2b45e8
	lfd	f30,  ALPHA_R
kusano 2b45e8
	lfd	f31,  ALPHA_I
kusano 2b45e8
kusano 2b45e8
	dcbtst	B, BB
kusano 2b45e8
	addi	BB, BB, 16 * SIZE
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	andi.	r0,  K,  7
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(38)
kusano 2b45e8
#else
kusano 2b45e8
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#elif defined(LEFT)
kusano 2b45e8
	addi	TEMP, KK, 2
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, KK, 2
kusano 2b45e8
#endif
kusano 2b45e8
	andi.	TEMP,  TEMP,  7
kusano 2b45e8
	mtspr	CTR, TEMP
kusano 2b45e8
	ble	LL(38)
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(36):
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA1	f6,  f18, f22, f6
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA2	f7,  f18, f23, f7
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f20, f9
kusano 2b45e8
	FMA4	f11, f19, f20, f11
kusano 2b45e8
	FMA3	f8,  f17, f21, f8
kusano 2b45e8
	FMA3	f10, f19, f21, f10
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  4 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  6 * SIZE(AO)
kusano 2b45e8
	LFD	f20,  4 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  5 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA4	f13, f17, f22, f13
kusano 2b45e8
	FMA4	f15, f19, f22, f15
kusano 2b45e8
	FMA3	f12, f17, f23, f12
kusano 2b45e8
	FMA3	f14, f19, f23, f14
kusano 2b45e8
kusano 2b45e8
	LFD	f17,  5 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  7 * SIZE(AO)
kusano 2b45e8
	LFD	f22,  6 * SIZE(BO)
kusano 2b45e8
	LFD	f23,  7 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO, AO,  4 * SIZE
kusano 2b45e8
	addi	BO, BO,  4 * SIZE
kusano 2b45e8
	bdnz	LL(36)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(38):
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 0 * SIZE(CO1)
kusano 2b45e8
	LFD	f17, 1 * SIZE(CO1)
kusano 2b45e8
	LFD	f18, 2 * SIZE(CO1)
kusano 2b45e8
	LFD	f19, 3 * SIZE(CO1)
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
	LFD	f20, 0 * SIZE(CO2)
kusano 2b45e8
	LFD	f21, 1 * SIZE(CO2)
kusano 2b45e8
	LFD	f22, 2 * SIZE(CO2)
kusano 2b45e8
	LFD	f23, 3 * SIZE(CO2)
kusano 2b45e8
kusano 2b45e8
	FADD	f4, f4, f12
kusano 2b45e8
	FADD	f5, f5, f13
kusano 2b45e8
	FADD	f6, f6, f14
kusano 2b45e8
	FADD	f7, f7, f15
kusano 2b45e8
kusano 2b45e8
	FNMSUB	f24, f31, f1, f16
kusano 2b45e8
	FMADD	f25, f31, f0, f17
kusano 2b45e8
	FNMSUB	f26, f31, f3, f18
kusano 2b45e8
	FMADD	f27, f31, f2, f19
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f30, f0, f24
kusano 2b45e8
	FMADD	f1,  f30, f1, f25
kusano 2b45e8
	FMADD	f2,  f30, f2, f26
kusano 2b45e8
	FMADD	f3,  f30, f3, f27
kusano 2b45e8
kusano 2b45e8
	FNMSUB	f24, f31, f5, f20
kusano 2b45e8
	FMADD	f25, f31, f4, f21
kusano 2b45e8
	FNMSUB	f26, f31, f7, f22
kusano 2b45e8
	FMADD	f27, f31, f6, f23
kusano 2b45e8
kusano 2b45e8
	FMADD	f4,  f30, f4,  f24
kusano 2b45e8
	FMADD	f5,  f30, f5,  f25
kusano 2b45e8
	FMADD	f6,  f30, f6,  f26
kusano 2b45e8
	FMADD	f7,  f30, f7,  f27
kusano 2b45e8
kusano 2b45e8
#else
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
	FADD	f4, f4, f12
kusano 2b45e8
	FADD	f5, f5, f13
kusano 2b45e8
	FADD	f6, f6, f14
kusano 2b45e8
	FADD	f7, f7, f15
kusano 2b45e8
kusano 2b45e8
	FMUL	f16, f31, f1
kusano 2b45e8
	FMUL	f17, f31, f0
kusano 2b45e8
	FMUL	f18, f31, f3
kusano 2b45e8
	FMUL	f19, f31, f2
kusano 2b45e8
kusano 2b45e8
	FMUL	f20, f31, f5
kusano 2b45e8
	FMUL	f21, f31, f4
kusano 2b45e8
	FMUL	f22, f31, f7
kusano 2b45e8
	FMUL	f23, f31, f6
kusano 2b45e8
kusano 2b45e8
	FMSUB	f0,  f30, f0, f16
kusano 2b45e8
	FMADD	f1,  f30, f1, f17
kusano 2b45e8
	FMADD	f2,  f30, f2, f18
kusano 2b45e8
	FMADD	f3,  f30, f3, f19
kusano 2b45e8
kusano 2b45e8
	FMSUB	f4,  f30, f4, f20
kusano 2b45e8
	FMADD	f5,  f30, f5, f21
kusano 2b45e8
	FMADD	f6,  f30, f6, f22
kusano 2b45e8
	FMADD	f7,  f30, f7, f23
kusano 2b45e8
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	STFD	f0,  0 * SIZE(CO1)
kusano 2b45e8
	STFD	f1,  1 * SIZE(CO1)
kusano 2b45e8
	STFD	f2,  2 * SIZE(CO1)
kusano 2b45e8
	STFD	f3,  3 * SIZE(CO1)
kusano 2b45e8
kusano 2b45e8
	lfs	f0,  FZERO
kusano 2b45e8
 	fmr	f1,  f0
kusano 2b45e8
	fmr	f2,  f0
kusano 2b45e8
	fmr	f3,  f0
kusano 2b45e8
kusano 2b45e8
	STFD	f4,  0 * SIZE(CO2)
kusano 2b45e8
	STFD	f5,  1 * SIZE(CO2)
kusano 2b45e8
	STFD	f6,  2 * SIZE(CO2)
kusano 2b45e8
	STFD	f7,  3 * SIZE(CO2)
kusano 2b45e8
kusano 2b45e8
	fmr	f4,  f0
kusano 2b45e8
	fmr	f5,  f0
kusano 2b45e8
	fmr	f6,  f0
kusano 2b45e8
	fmr	f7,  f0
kusano 2b45e8
kusano 2b45e8
	fmr	f8,  f0
kusano 2b45e8
	fmr	f9,  f0
kusano 2b45e8
	fmr	f10, f0
kusano 2b45e8
	fmr	f11, f0
kusano 2b45e8
kusano 2b45e8
	fmr	f12, f0
kusano 2b45e8
	fmr	f13, f0
kusano 2b45e8
	fmr	f14, f0
kusano 2b45e8
	fmr	f15, f0
kusano 2b45e8
kusano 2b45e8
	addi	CO1, CO1, 4 * SIZE
kusano 2b45e8
	addi	CO2, CO2, 4 * SIZE
kusano 2b45e8
	
kusano 2b45e8
#ifdef TRMMKERNEL
kusano 2b45e8
#if ( defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
    (!defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addi	TEMP, TEMP, -2
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, TEMP, -2
kusano 2b45e8
#endif
kusano 2b45e8
	slwi	r0,   TEMP, 1 + ZBASE_SHIFT
kusano 2b45e8
	slwi	TEMP, TEMP, 1 + ZBASE_SHIFT
kusano 2b45e8
	add	AO, AO, r0
kusano 2b45e8
	add	BO, BO, TEMP
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addi	KK, KK, 2
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	addic.	I, I, -1
kusano 2b45e8
	bgt	LL(31)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(40):
kusano 2b45e8
	andi.	I,  M,  1
kusano 2b45e8
	ble	LL(49)
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(B)
kusano 2b45e8
	LFD	f21,  1 * SIZE(B)
kusano 2b45e8
	LFD	f22,  2 * SIZE(B)
kusano 2b45e8
	LFD	f23,  3 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  K,  2
kusano 2b45e8
	mr	BO,  B
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(45)
kusano 2b45e8
#else
kusano 2b45e8
#if	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(B)
kusano 2b45e8
	LFD	f21,  1 * SIZE(B)
kusano 2b45e8
	LFD	f22,  2 * SIZE(B)
kusano 2b45e8
	LFD	f23,  3 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	mr	BO,  B
kusano 2b45e8
#else
kusano 2b45e8
	slwi	r0,   KK, 0 + ZBASE_SHIFT
kusano 2b45e8
	slwi	TEMP, KK, 1 + ZBASE_SHIFT
kusano 2b45e8
	add	AO, AO, r0
kusano 2b45e8
	add	BO, B,  TEMP
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  1 * SIZE(BO)
kusano 2b45e8
	LFD	f22,  2 * SIZE(BO)
kusano 2b45e8
	LFD	f23,  3 * SIZE(BO)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#elif defined(LEFT)
kusano 2b45e8
	addi	TEMP, KK, 1
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, KK, 2
kusano 2b45e8
#endif
kusano 2b45e8
	srawi.	TEMP,  TEMP,  2
kusano 2b45e8
	mtspr	CTR, TEMP
kusano 2b45e8
	ble	LL(45)
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(42):
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA4	f3,  f17, f20, f3
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA3	f2,  f17, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA4	f7,  f17, f22, f7
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA3	f6,  f17, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  4 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  5 * SIZE(BO)
kusano 2b45e8
	LFD	f22,  6 * SIZE(BO)
kusano 2b45e8
	LFD	f23,  7 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA4	f3,  f17, f20, f3
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA3	f2,  f17, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA4	f7,  f17, f22, f7
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA3	f6,  f17, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  4 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  5 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  8 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  9 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 10 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 11 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA4	f3,  f17, f20, f3
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA3	f2,  f17, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA4	f7,  f17, f22, f7
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA3	f6,  f17, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  6 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  7 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 12 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 13 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 14 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 15 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA4	f3,  f17, f20, f3
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA3	f2,  f17, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA4	f7,  f17, f22, f7
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA3	f6,  f17, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  8 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  9 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 16 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 17 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 18 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 19 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO, AO,  8 * SIZE
kusano 2b45e8
	addi	BO, BO, 16 * SIZE
kusano 2b45e8
kusano 2b45e8
	bdnz	LL(42)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(45):
kusano 2b45e8
	lfd	f30,  ALPHA_R
kusano 2b45e8
	lfd	f31,  ALPHA_I
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	andi.	r0,  K,  3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(48)
kusano 2b45e8
#else
kusano 2b45e8
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#elif defined(LEFT)
kusano 2b45e8
	addi	TEMP, KK, 1
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, KK, 2
kusano 2b45e8
#endif
kusano 2b45e8
	andi.	TEMP,  TEMP,  3
kusano 2b45e8
	mtspr	CTR, TEMP
kusano 2b45e8
	ble	LL(48)
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(46):
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA4	f3,  f17, f20, f3
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA3	f2,  f17, f21, f2
kusano 2b45e8
kusano 2b45e8
	FMA1	f4,  f16, f22, f4
kusano 2b45e8
	FMA4	f7,  f17, f22, f7
kusano 2b45e8
	FMA2	f5,  f16, f23, f5
kusano 2b45e8
	FMA3	f6,  f17, f23, f6
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  4 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  5 * SIZE(BO)
kusano 2b45e8
	LFD	f22,  6 * SIZE(BO)
kusano 2b45e8
	LFD	f23,  7 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO, AO,  2 * SIZE
kusano 2b45e8
	addi	BO, BO,  4 * SIZE
kusano 2b45e8
	bdnz	LL(46)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(48):
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	LFD	f16, 0 * SIZE(CO1)
kusano 2b45e8
	LFD	f17, 1 * SIZE(CO1)
kusano 2b45e8
	LFD	f20, 0 * SIZE(CO2)
kusano 2b45e8
	LFD	f21, 1 * SIZE(CO2)
kusano 2b45e8
kusano 2b45e8
	FADD	f0,  f0,  f2
kusano 2b45e8
	FADD	f1,  f1,  f3
kusano 2b45e8
	FADD	f4,  f4,  f6
kusano 2b45e8
	FADD	f5,  f5,  f7
kusano 2b45e8
kusano 2b45e8
	FNMSUB	f24, f31, f1, f16
kusano 2b45e8
	FMADD	f25, f31, f0, f17
kusano 2b45e8
	FNMSUB	f26, f31, f5, f20
kusano 2b45e8
	FMADD	f27, f31, f4, f21
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f30, f0, f24
kusano 2b45e8
	FMADD	f1,  f30, f1, f25
kusano 2b45e8
	FMADD	f4,  f30, f4,  f26
kusano 2b45e8
	FMADD	f5,  f30, f5,  f27
kusano 2b45e8
kusano 2b45e8
#else
kusano 2b45e8
	FADD	f0,  f0,  f2
kusano 2b45e8
	FADD	f1,  f1,  f3
kusano 2b45e8
	FADD	f4,  f4,  f6
kusano 2b45e8
	FADD	f5,  f5,  f7
kusano 2b45e8
kusano 2b45e8
	FMUL	f16, f31, f1
kusano 2b45e8
	FMUL	f17, f31, f0
kusano 2b45e8
	FMUL	f18, f31, f5
kusano 2b45e8
	FMUL	f19, f31, f4
kusano 2b45e8
kusano 2b45e8
	FMSUB	f0,  f30, f0, f16
kusano 2b45e8
	FMADD	f1,  f30, f1, f17
kusano 2b45e8
	FMSUB	f4,  f30, f4, f18
kusano 2b45e8
	FMADD	f5,  f30, f5, f19
kusano 2b45e8
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	STFD	f0,  0 * SIZE(CO1)
kusano 2b45e8
	STFD	f1,  1 * SIZE(CO1)
kusano 2b45e8
	STFD	f4,  0 * SIZE(CO2)
kusano 2b45e8
	STFD	f5,  1 * SIZE(CO2)
kusano 2b45e8
kusano 2b45e8
	addi	CO1, CO1, 2 * SIZE
kusano 2b45e8
	addi	CO2, CO2, 2 * SIZE
kusano 2b45e8
	
kusano 2b45e8
#ifdef TRMMKERNEL
kusano 2b45e8
#if ( defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
    (!defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addi	TEMP, TEMP, -1
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, TEMP, -2
kusano 2b45e8
#endif
kusano 2b45e8
	slwi	r0,   TEMP, 0 + ZBASE_SHIFT
kusano 2b45e8
	slwi	TEMP, TEMP, 1 + ZBASE_SHIFT
kusano 2b45e8
	add	AO, AO, r0
kusano 2b45e8
	add	BO, BO, TEMP
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addi	KK, KK, 1
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(49):
kusano 2b45e8
#if defined(TRMMKERNEL) && !defined(LEFT)
kusano 2b45e8
	addi	KK, KK, 2
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	mr	B,  BO
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(50):
kusano 2b45e8
	andi.	J, N,  1
kusano 2b45e8
	ble	LL(999)
kusano 2b45e8
kusano 2b45e8
	mr	CO1, C
kusano 2b45e8
	add	C,   CO1, LDC
kusano 2b45e8
kusano 2b45e8
#if defined(TRMMKERNEL) && defined(LEFT)
kusano 2b45e8
	mr	KK, OFFSET
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	mr	AO, A
kusano 2b45e8
kusano 2b45e8
	lfs	f0, FZERO
kusano 2b45e8
 	fmr	f1,  f0
kusano 2b45e8
	fmr	f2,  f0
kusano 2b45e8
	fmr	f3,  f0
kusano 2b45e8
	fmr	f8,  f0
kusano 2b45e8
	fmr	f9,  f0
kusano 2b45e8
	fmr	f10, f0
kusano 2b45e8
	fmr	f11, f0
kusano 2b45e8
kusano 2b45e8
	srawi.	I, M,  1
kusano 2b45e8
	ble	LL(60)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(51):
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(B)
kusano 2b45e8
	LFD	f21,  1 * SIZE(B)
kusano 2b45e8
	LFD	f22,  2 * SIZE(B)
kusano 2b45e8
	LFD	f23,  3 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	dcbtst	CO1, PREC
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  K,  3
kusano 2b45e8
	mr	BO,  B
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(55)
kusano 2b45e8
#else
kusano 2b45e8
#if	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(B)
kusano 2b45e8
	LFD	f21,  1 * SIZE(B)
kusano 2b45e8
	LFD	f22,  2 * SIZE(B)
kusano 2b45e8
	LFD	f23,  3 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	mr	BO,  B
kusano 2b45e8
#else
kusano 2b45e8
	slwi	r0,   KK, 1 + ZBASE_SHIFT
kusano 2b45e8
	slwi	TEMP, KK, 0 + ZBASE_SHIFT
kusano 2b45e8
	add	AO, AO, r0
kusano 2b45e8
	add	BO, B,  TEMP
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  1 * SIZE(BO)
kusano 2b45e8
	LFD	f22,  2 * SIZE(BO)
kusano 2b45e8
	LFD	f23,  3 * SIZE(BO)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	dcbtst	CO1, PREC
kusano 2b45e8
kusano 2b45e8
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#elif defined(LEFT)
kusano 2b45e8
	addi	TEMP, KK, 2
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, KK, 1
kusano 2b45e8
#endif
kusano 2b45e8
	srawi.	TEMP,  TEMP,  3
kusano 2b45e8
	mtspr	CTR, TEMP
kusano 2b45e8
	ble	LL(55)
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(52):
kusano 2b45e8
	dcbt	AO, PREA
kusano 2b45e8
	dcbtst	BO, PREA
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f20, f9
kusano 2b45e8
	FMA4	f11, f19, f20, f11
kusano 2b45e8
	FMA3	f8,  f17, f21, f8
kusano 2b45e8
	FMA3	f10, f19, f21, f10
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  4 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  5 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  6 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  7 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f22, f0
kusano 2b45e8
	FMA1	f2,  f18, f22, f2
kusano 2b45e8
	FMA2	f1,  f16, f23, f1
kusano 2b45e8
	FMA2	f3,  f18, f23, f3
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f22, f9
kusano 2b45e8
	FMA4	f11, f19, f22, f11
kusano 2b45e8
	FMA3	f8,  f17, f23, f8
kusano 2b45e8
	FMA3	f10, f19, f23, f10
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  8 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  9 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 10 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 11 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  4 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  5 * SIZE(BO)
kusano 2b45e8
	LFD	f22,  6 * SIZE(BO)
kusano 2b45e8
	LFD	f23,  7 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f20, f9
kusano 2b45e8
	FMA4	f11, f19, f20, f11
kusano 2b45e8
	FMA3	f8,  f17, f21, f8
kusano 2b45e8
	FMA3	f10, f19, f21, f10
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 12 * SIZE(AO)
kusano 2b45e8
	LFD	f17, 13 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 14 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 15 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f22, f0
kusano 2b45e8
	FMA1	f2,  f18, f22, f2
kusano 2b45e8
	FMA2	f1,  f16, f23, f1
kusano 2b45e8
	FMA2	f3,  f18, f23, f3
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f22, f9
kusano 2b45e8
	FMA4	f11, f19, f22, f11
kusano 2b45e8
	FMA3	f8,  f17, f23, f8
kusano 2b45e8
	FMA3	f10, f19, f23, f10
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 16 * SIZE(AO)
kusano 2b45e8
	LFD	f17, 17 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 18 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 19 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  8 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  9 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 10 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 11 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f20, f9
kusano 2b45e8
	FMA4	f11, f19, f20, f11
kusano 2b45e8
	FMA3	f8,  f17, f21, f8
kusano 2b45e8
	FMA3	f10, f19, f21, f10
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 20 * SIZE(AO)
kusano 2b45e8
	LFD	f17, 21 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 22 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 23 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f22, f0
kusano 2b45e8
	FMA1	f2,  f18, f22, f2
kusano 2b45e8
	FMA2	f1,  f16, f23, f1
kusano 2b45e8
	FMA2	f3,  f18, f23, f3
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f22, f9
kusano 2b45e8
	FMA4	f11, f19, f22, f11
kusano 2b45e8
	FMA3	f8,  f17, f23, f8
kusano 2b45e8
	FMA3	f10, f19, f23, f10
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 24 * SIZE(AO)
kusano 2b45e8
	LFD	f17, 25 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 26 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 27 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 12 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 13 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 14 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 15 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f20, f9
kusano 2b45e8
	FMA4	f11, f19, f20, f11
kusano 2b45e8
	FMA3	f8,  f17, f21, f8
kusano 2b45e8
	FMA3	f10, f19, f21, f10
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 28 * SIZE(AO)
kusano 2b45e8
	LFD	f17, 29 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 30 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 31 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f22, f0
kusano 2b45e8
	FMA1	f2,  f18, f22, f2
kusano 2b45e8
	FMA2	f1,  f16, f23, f1
kusano 2b45e8
	FMA2	f3,  f18, f23, f3
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f22, f9
kusano 2b45e8
	FMA4	f11, f19, f22, f11
kusano 2b45e8
	FMA3	f8,  f17, f23, f8
kusano 2b45e8
	FMA3	f10, f19, f23, f10
kusano 2b45e8
kusano 2b45e8
	LFD	f16, 32 * SIZE(AO)
kusano 2b45e8
	LFD	f17, 33 * SIZE(AO)
kusano 2b45e8
	LFD	f18, 34 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 35 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20, 16 * SIZE(BO)
kusano 2b45e8
	LFD	f21, 17 * SIZE(BO)
kusano 2b45e8
	LFD	f22, 18 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 19 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO, AO, 32 * SIZE
kusano 2b45e8
	addi	BO, BO, 16 * SIZE
kusano 2b45e8
kusano 2b45e8
	bdnz	LL(52)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(55):
kusano 2b45e8
	lfd	f30,  ALPHA_R
kusano 2b45e8
	lfd	f31,  ALPHA_I
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	andi.	r0,  K,  7
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(58)
kusano 2b45e8
#else
kusano 2b45e8
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#elif defined(LEFT)
kusano 2b45e8
	addi	TEMP, KK, 2
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, KK, 1
kusano 2b45e8
#endif
kusano 2b45e8
	andi.	TEMP,  TEMP,  7
kusano 2b45e8
	mtspr	CTR, TEMP
kusano 2b45e8
	ble	LL(58)
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(56):
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA1	f2,  f18, f20, f2
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA2	f3,  f18, f21, f3
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  4 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  6 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FMA4	f9,  f17, f20, f9
kusano 2b45e8
	FMA4	f11, f19, f20, f11
kusano 2b45e8
	FMA3	f8,  f17, f21, f8
kusano 2b45e8
	FMA3	f10, f19, f21, f10
kusano 2b45e8
kusano 2b45e8
	LFD	f17,  5 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  7 * SIZE(AO)
kusano 2b45e8
	LFD	f20,  2 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  3 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO, AO,  4 * SIZE
kusano 2b45e8
	addi	BO, BO,  2 * SIZE
kusano 2b45e8
	bdnz	LL(56)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(58):
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	LFD	f16, 0 * SIZE(CO1)
kusano 2b45e8
	LFD	f17, 1 * SIZE(CO1)
kusano 2b45e8
	LFD	f18, 2 * SIZE(CO1)
kusano 2b45e8
	LFD	f19, 3 * SIZE(CO1)
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
	FNMSUB	f24, f31, f1, f16
kusano 2b45e8
	FMADD	f25, f31, f0, f17
kusano 2b45e8
	FNMSUB	f26, f31, f3, f18
kusano 2b45e8
	FMADD	f27, f31, f2, f19
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f30, f0, f24
kusano 2b45e8
	FMADD	f1,  f30, f1, f25
kusano 2b45e8
	FMADD	f2,  f30, f2, f26
kusano 2b45e8
	FMADD	f3,  f30, f3, f27
kusano 2b45e8
kusano 2b45e8
#else
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
	FMUL	f16, f31, f1
kusano 2b45e8
	FMUL	f17, f31, f0
kusano 2b45e8
	FMUL	f18, f31, f3
kusano 2b45e8
	FMUL	f19, f31, f2
kusano 2b45e8
kusano 2b45e8
	FMSUB	f0,  f30, f0, f16
kusano 2b45e8
	FMADD	f1,  f30, f1, f17
kusano 2b45e8
	FMADD	f2,  f30, f2, f18
kusano 2b45e8
	FMADD	f3,  f30, f3, f19
kusano 2b45e8
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	STFD	f0,  0 * SIZE(CO1)
kusano 2b45e8
	STFD	f1,  1 * SIZE(CO1)
kusano 2b45e8
	STFD	f2,  2 * SIZE(CO1)
kusano 2b45e8
	STFD	f3,  3 * SIZE(CO1)
kusano 2b45e8
kusano 2b45e8
	lfs	f0,  FZERO
kusano 2b45e8
 	fmr	f1,  f0
kusano 2b45e8
	fmr	f2,  f0
kusano 2b45e8
	fmr	f3,  f0
kusano 2b45e8
kusano 2b45e8
	fmr	f8,  f0
kusano 2b45e8
	fmr	f9,  f0
kusano 2b45e8
	fmr	f10, f0
kusano 2b45e8
	fmr	f11, f0
kusano 2b45e8
kusano 2b45e8
	addi	CO1, CO1, 4 * SIZE
kusano 2b45e8
	
kusano 2b45e8
#ifdef TRMMKERNEL
kusano 2b45e8
#if ( defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
    (!defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addi	TEMP, TEMP, -2
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, TEMP, -1
kusano 2b45e8
#endif
kusano 2b45e8
	slwi	r0,   TEMP, 1 + ZBASE_SHIFT
kusano 2b45e8
	slwi	TEMP, TEMP, 0 + ZBASE_SHIFT
kusano 2b45e8
	add	AO, AO, r0
kusano 2b45e8
	add	BO, BO, TEMP
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addi	KK, KK, 2
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	addic.	I, I, -1
kusano 2b45e8
	bgt	LL(51)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(60):
kusano 2b45e8
	andi.	I,  M,  1
kusano 2b45e8
	ble	LL(999)
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(B)
kusano 2b45e8
	LFD	f21,  1 * SIZE(B)
kusano 2b45e8
	LFD	f22,  2 * SIZE(B)
kusano 2b45e8
	LFD	f23,  3 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  K,  2
kusano 2b45e8
	mr	BO,  B
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(65)
kusano 2b45e8
#else
kusano 2b45e8
#if	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(B)
kusano 2b45e8
	LFD	f21,  1 * SIZE(B)
kusano 2b45e8
	LFD	f22,  2 * SIZE(B)
kusano 2b45e8
	LFD	f23,  3 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	mr	BO,  B
kusano 2b45e8
#else
kusano 2b45e8
	slwi	r0,   KK, 0 + ZBASE_SHIFT
kusano 2b45e8
	slwi	TEMP, KK, 0 + ZBASE_SHIFT
kusano 2b45e8
	add	AO, AO, r0
kusano 2b45e8
	add	BO, B,  TEMP
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  0 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  1 * SIZE(AO)
kusano 2b45e8
	LFD	f18,  2 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,  0 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  1 * SIZE(BO)
kusano 2b45e8
	LFD	f22,  2 * SIZE(BO)
kusano 2b45e8
	LFD	f23,  3 * SIZE(BO)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#elif defined(LEFT)
kusano 2b45e8
	addi	TEMP, KK, 1
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, KK, 1
kusano 2b45e8
#endif
kusano 2b45e8
	srawi.	TEMP,  TEMP,  2
kusano 2b45e8
	mtspr	CTR, TEMP
kusano 2b45e8
	ble	LL(65)
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(62):
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA4	f3,  f17, f20, f3
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA3	f2,  f17, f21, f2
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  4 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  5 * SIZE(AO)
kusano 2b45e8
	LFD	f20,  4 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  5 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f18, f22, f0
kusano 2b45e8
	FMA4	f3,  f19, f22, f3
kusano 2b45e8
	FMA2	f1,  f18, f23, f1
kusano 2b45e8
	FMA3	f2,  f19, f23, f2
kusano 2b45e8
kusano 2b45e8
	LFD	f18,  6 * SIZE(AO)
kusano 2b45e8
	LFD	f19,  7 * SIZE(AO)
kusano 2b45e8
	LFD	f22,  6 * SIZE(BO)
kusano 2b45e8
	LFD	f23,  7 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA4	f3,  f17, f20, f3
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	FMA3	f2,  f17, f21, f2
kusano 2b45e8
kusano 2b45e8
	LFD	f16,  8 * SIZE(AO)
kusano 2b45e8
	LFD	f17,  9 * SIZE(AO)
kusano 2b45e8
	LFD	f20,  8 * SIZE(BO)
kusano 2b45e8
	LFD	f21,  9 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FMA1	f0,  f18, f22, f0
kusano 2b45e8
	FMA4	f3,  f19, f22, f3
kusano 2b45e8
	FMA2	f1,  f18, f23, f1
kusano 2b45e8
	FMA3	f2,  f19, f23, f2
kusano 2b45e8
kusano 2b45e8
	LFD	f18, 10 * SIZE(AO)
kusano 2b45e8
	LFD	f19, 11 * SIZE(AO)
kusano 2b45e8
	LFD	f22, 10 * SIZE(BO)
kusano 2b45e8
	LFD	f23, 11 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	addi	AO, AO,  8 * SIZE
kusano 2b45e8
	addi	BO, BO,  8 * SIZE
kusano 2b45e8
	bdnz	LL(62)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(65):
kusano 2b45e8
	lfd	f30,  ALPHA_R
kusano 2b45e8
	lfd	f31,  ALPHA_I
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	andi.	r0,  K,  3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(68)
kusano 2b45e8
#else
kusano 2b45e8
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#elif defined(LEFT)
kusano 2b45e8
	addi	TEMP, KK, 1
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, KK, 1
kusano 2b45e8
#endif
kusano 2b45e8
	andi.	TEMP,  TEMP,  3
kusano 2b45e8
	mtspr	CTR, TEMP
kusano 2b45e8
	ble	LL(68)
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(66):
kusano 2b45e8
	FMA1	f0,  f16, f20, f0
kusano 2b45e8
	FMA4	f3,  f17, f20, f3
kusano 2b45e8
	LFD	f20,  2 * SIZE(BO)
kusano 2b45e8
	FMA2	f1,  f16, f21, f1
kusano 2b45e8
	LFD	f16,  2 * SIZE(AO)
kusano 2b45e8
	FMA3	f2,  f17, f21, f2
kusano 2b45e8
	LFD	f17,  3 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	LFD	f21,  3 * SIZE(BO)
kusano 2b45e8
	addi	AO, AO,  2 * SIZE
kusano 2b45e8
	addi	BO, BO,  2 * SIZE
kusano 2b45e8
	bdnz	LL(66)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(68):
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	LFD	f16, 0 * SIZE(CO1)
kusano 2b45e8
	LFD	f17, 1 * SIZE(CO1)
kusano 2b45e8
kusano 2b45e8
	FADD	f0,  f0,  f2
kusano 2b45e8
	FADD	f1,  f1,  f3
kusano 2b45e8
kusano 2b45e8
	FNMSUB	f24, f31, f1, f16
kusano 2b45e8
	FMADD	f25, f31, f0, f17
kusano 2b45e8
kusano 2b45e8
	FMADD	f0,  f30, f0, f24
kusano 2b45e8
	FMADD	f1,  f30, f1, f25
kusano 2b45e8
kusano 2b45e8
#else
kusano 2b45e8
kusano 2b45e8
	FADD	f0,  f0,  f2
kusano 2b45e8
	FADD	f1,  f1,  f3
kusano 2b45e8
kusano 2b45e8
	FMUL	f16, f31, f1
kusano 2b45e8
	FMUL	f17, f31, f0
kusano 2b45e8
kusano 2b45e8
	FMSUB	f0,  f30, f0, f16
kusano 2b45e8
	FMADD	f1,  f30, f1, f17
kusano 2b45e8
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	STFD	f0,  0 * SIZE(CO1)
kusano 2b45e8
	STFD	f1,  1 * SIZE(CO1)
kusano 2b45e8
kusano 2b45e8
	addi	CO1, CO1, 2 * SIZE
kusano 2b45e8
	
kusano 2b45e8
#ifdef TRMMKERNEL
kusano 2b45e8
#if ( defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
    (!defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	sub	TEMP, K, KK
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addi	TEMP, TEMP, -1
kusano 2b45e8
#else
kusano 2b45e8
	addi	TEMP, TEMP, -1
kusano 2b45e8
#endif
kusano 2b45e8
	slwi	r0,   TEMP, 0 + ZBASE_SHIFT
kusano 2b45e8
	slwi	TEMP, TEMP, 0 + ZBASE_SHIFT
kusano 2b45e8
	add	AO, AO, r0
kusano 2b45e8
	add	BO, BO, TEMP
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addi	KK, KK, 1
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(999):
kusano 2b45e8
	addi	r3, 0, 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
kusano 2b45e8
	lfd	f18,   32(SP)
kusano 2b45e8
	lfd	f19,   40(SP)
kusano 2b45e8
	lfd	f20,   48(SP)
kusano 2b45e8
	lfd	f21,   56(SP)
kusano 2b45e8
kusano 2b45e8
	lfd	f22,   64(SP)
kusano 2b45e8
	lfd	f23,   72(SP)
kusano 2b45e8
	lfd	f24,   80(SP)
kusano 2b45e8
	lfd	f25,   88(SP)
kusano 2b45e8
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
kusano 2b45e8
	lfd	f30,  128(SP)
kusano 2b45e8
	lfd	f31,  136(SP)
kusano 2b45e8
kusano 2b45e8
#ifdef __64BIT__
kusano 2b45e8
	ld	r31,  144(SP)
kusano 2b45e8
	ld	r30,  152(SP)
kusano 2b45e8
	ld	r29,  160(SP)
kusano 2b45e8
	ld	r28,  168(SP)
kusano 2b45e8
	ld	r27,  176(SP)
kusano 2b45e8
	ld	r26,  184(SP)
kusano 2b45e8
	ld	r25,  192(SP)
kusano 2b45e8
	ld	r24,  200(SP)
kusano 2b45e8
	ld	r23,  208(SP)
kusano 2b45e8
	ld	r22,  216(SP)
kusano 2b45e8
	ld	r21,  224(SP)
kusano 2b45e8
#ifdef TRMMKERNEL
kusano 2b45e8
	ld	r20,  232(SP)
kusano 2b45e8
	ld	r19,  240(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#else
kusano 2b45e8
	lwz	r31,  144(SP)
kusano 2b45e8
	lwz	r30,  148(SP)
kusano 2b45e8
	lwz	r29,  152(SP)
kusano 2b45e8
	lwz	r28,  156(SP)
kusano 2b45e8
	lwz	r27,  160(SP)
kusano 2b45e8
	lwz	r26,  164(SP)
kusano 2b45e8
	lwz	r25,  168(SP)
kusano 2b45e8
	lwz	r24,  172(SP)
kusano 2b45e8
	lwz	r23,  176(SP)
kusano 2b45e8
	lwz	r22,  180(SP)
kusano 2b45e8
	lwz	r21,  184(SP)
kusano 2b45e8
#ifdef TRMMKERNEL
kusano 2b45e8
	lwz	r20,  188(SP)
kusano 2b45e8
	lwz	r19,  192(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	addi	SP, SP, STACKSIZE
kusano 2b45e8
kusano 2b45e8
	blr
kusano 2b45e8
kusano 2b45e8
	EPILOGUE
kusano 2b45e8
#endif