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
#if defined(OPTERON) || defined(BARCELONA)
kusano 2b45e8
#define PREFETCH	prefetch
kusano 2b45e8
#define PREFETCHW	prefetchw
kusano 2b45e8
#else
kusano 2b45e8
#define PREFETCH	prefetcht0
kusano 2b45e8
#define PREFETCHW	prefetcht0
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#define PREFETCHSIZE (5 + 4 * 10)
kusano 2b45e8
#define STACK	16
kusano 2b45e8
#define ARGS	16
kusano 2b45e8
	
kusano 2b45e8
#define J	 0 + STACK(%esp)
kusano 2b45e8
#define KK	 4 + STACK(%esp)
kusano 2b45e8
#define KKK	 8 + STACK(%esp)
kusano 2b45e8
kusano 2b45e8
#define M	 4 + STACK + ARGS(%esp)
kusano 2b45e8
#define N	 8 + STACK + ARGS(%esp)
kusano 2b45e8
#define K	12 + STACK + ARGS(%esp)
kusano 2b45e8
#define ALPHA	16 + STACK + ARGS(%esp)
kusano 2b45e8
#define A	32 + STACK + ARGS(%esp)
kusano 2b45e8
#define ARG_B	36 + STACK + ARGS(%esp)
kusano 2b45e8
#define C	40 + STACK + ARGS(%esp)
kusano 2b45e8
#define ARG_LDC	44 + STACK + ARGS(%esp)
kusano 2b45e8
#define OFFSET	48 + STACK + ARGS(%esp)
kusano 2b45e8
kusano 2b45e8
#define I	%esi
kusano 2b45e8
#define B	%ebx
kusano 2b45e8
#define CO	%edi
kusano 2b45e8
#define AO	%edx
kusano 2b45e8
#define BO	%ecx
kusano 2b45e8
#define LDC	%ebp
kusano 2b45e8
kusano 2b45e8
#define PREFETCH_OFFSET 48
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
kusano 2b45e8
	subl	$ARGS, %esp	# Generate Stack Frame
kusano 2b45e8
kusano 2b45e8
	pushl	%ebp
kusano 2b45e8
	pushl	%edi
kusano 2b45e8
	pushl	%esi
kusano 2b45e8
	pushl	%ebx
kusano 2b45e8
kusano 2b45e8
	PROFCODE
kusano 2b45e8
kusano 2b45e8
#if defined(TRMMKERNEL) && !defined(LEFT)
kusano 2b45e8
	movl	OFFSET, %eax
kusano 2b45e8
	negl	%eax
kusano 2b45e8
	movl	%eax, KK
kusano 2b45e8
#endif
kusano 2b45e8
	
kusano 2b45e8
	movl	ARG_LDC, LDC
kusano 2b45e8
	movl	ARG_B,   B
kusano 2b45e8
kusano 2b45e8
	addl	$8 * SIZE, A
kusano 2b45e8
	addl	$8 * SIZE, B
kusano 2b45e8
	
kusano 2b45e8
	sall	$BASE_SHIFT, LDC
kusano 2b45e8
kusano 2b45e8
	movl	N,   %eax
kusano 2b45e8
	sarl	$1,  %eax
kusano 2b45e8
	movl	%eax, J
kusano 2b45e8
	je	.L30
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L01:
kusano 2b45e8
#if defined(TRMMKERNEL) && defined(LEFT)
kusano 2b45e8
	movl	OFFSET, %eax
kusano 2b45e8
	movl	%eax, KK
kusano 2b45e8
#endif	
kusano 2b45e8
kusano 2b45e8
	movl	A, AO
kusano 2b45e8
kusano 2b45e8
	movl	C, CO
kusano 2b45e8
	lea	(, LDC, 2), %eax
kusano 2b45e8
	addl	%eax, C
kusano 2b45e8
kusano 2b45e8
	movl	M,  I
kusano 2b45e8
	sarl	$1, I
kusano 2b45e8
	je	.L20
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L11:
kusano 2b45e8
#if !defined(TRMMKERNEL) || \
kusano 2b45e8
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	movl	B, BO
kusano 2b45e8
#else
kusano 2b45e8
	movl	KK,   %eax
kusano 2b45e8
	sall	$BASE_SHIFT, %eax
kusano 2b45e8
	leal	(AO, %eax, 2), AO
kusano 2b45e8
	leal	(B,  %eax, 2), BO
kusano 2b45e8
#endif	
kusano 2b45e8
kusano 2b45e8
	fldz
kusano 2b45e8
	fldz
kusano 2b45e8
	fldz
kusano 2b45e8
	fldz
kusano 2b45e8
kusano 2b45e8
#if   defined(HAVE_3DNOW)
kusano 2b45e8
	prefetchw	2 * SIZE(CO)
kusano 2b45e8
 	prefetchw	2 * SIZE(CO, LDC, 1)
kusano 2b45e8
#elif defined(HAVE_SSE)
kusano 2b45e8
	prefetchnta	2 * SIZE(CO)
kusano 2b45e8
 	prefetchnta	2 * SIZE(CO, LDC, 1)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
	subl	KK, %eax
kusano 2b45e8
	movl	%eax, KKK	
kusano 2b45e8
#else
kusano 2b45e8
	movl	KK, %eax
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addl	$2, %eax
kusano 2b45e8
#else
kusano 2b45e8
	addl	$2, %eax
kusano 2b45e8
#endif
kusano 2b45e8
	movl	%eax, KKK
kusano 2b45e8
#endif
kusano 2b45e8
	sarl	$2, %eax
kusano 2b45e8
 	je	.L15
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L12:
kusano 2b45e8
	PREFETCH	(PREFETCHSIZE + 0) * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(BO)
kusano 2b45e8
	fld	 %st(1)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -7 * SIZE(BO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	 
kusano 2b45e8
	FLD	 -7 * SIZE(AO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	faddp	 %st, %st(6)
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -6 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -6 * SIZE(BO)
kusano 2b45e8
	fld	 %st(1)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -5 * SIZE(BO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	 
kusano 2b45e8
	FLD	 -5 * SIZE(AO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	faddp	 %st, %st(6)
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	PREFETCH	(PREFETCHSIZE + 4) * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -4 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -4 * SIZE(BO)
kusano 2b45e8
	fld	 %st(1)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -3 * SIZE(BO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	 
kusano 2b45e8
	FLD	 -3 * SIZE(AO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	faddp	 %st, %st(6)
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -2 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -2 * SIZE(BO)
kusano 2b45e8
	fld	 %st(1)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -1 * SIZE(BO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	 
kusano 2b45e8
	FLD	 -1 * SIZE(AO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	faddp	 %st, %st(6)
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	addl	$8 * SIZE,AO
kusano 2b45e8
	addl	$8 * SIZE,BO
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jne	.L12
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L15:
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
#else
kusano 2b45e8
	movl	KKK, %eax
kusano 2b45e8
#endif
kusano 2b45e8
	and	$3,  %eax
kusano 2b45e8
	je	.L18
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L16:
kusano 2b45e8
	FLD	 -8 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(BO)
kusano 2b45e8
	fld	 %st(1)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -7 * SIZE(BO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	 
kusano 2b45e8
	FLD	 -7 * SIZE(AO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	faddp	 %st, %st(6)
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
	faddp	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	addl	$2 * SIZE,AO
kusano 2b45e8
	addl	$2 * SIZE,BO
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jne	 .L16
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L18:
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	FLD	ALPHA
kusano 2b45e8
kusano 2b45e8
	fmul	%st, %st(1)
kusano 2b45e8
	fmul	%st, %st(2)
kusano 2b45e8
	fmul	%st, %st(3)
kusano 2b45e8
	fmulp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(CO)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(CO)
kusano 2b45e8
kusano 2b45e8
	FLD	1 * SIZE(CO)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	1 * SIZE(CO)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(CO, LDC)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(CO, LDC)
kusano 2b45e8
kusano 2b45e8
	FLD	1 * SIZE(CO, LDC)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	1 * SIZE(CO, LDC)
kusano 2b45e8
#else
kusano 2b45e8
	FST	0 * SIZE(CO)
kusano 2b45e8
	FST	1 * SIZE(CO)
kusano 2b45e8
	FST	0 * SIZE(CO, LDC)
kusano 2b45e8
	FST	1 * SIZE(CO, LDC)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
	subl	KKK, %eax
kusano 2b45e8
	sall	$BASE_SHIFT, %eax
kusano 2b45e8
	leal	(AO, %eax, 2), AO
kusano 2b45e8
	leal	(BO, %eax, 2), BO
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(TRMMKERNEL) && defined(LEFT)
kusano 2b45e8
	addl	$2, KK
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	addl	$2 * SIZE, CO
kusano 2b45e8
	decl	I
kusano 2b45e8
	jne	.L11
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L20:
kusano 2b45e8
	movl	 M, %eax
kusano 2b45e8
	andl	$1, %eax
kusano 2b45e8
	je	.L29
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L21:
kusano 2b45e8
#if !defined(TRMMKERNEL) || \
kusano 2b45e8
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	movl	B, BO
kusano 2b45e8
#else
kusano 2b45e8
	movl	KK,   %eax
kusano 2b45e8
	sall	$BASE_SHIFT, %eax
kusano 2b45e8
	leal	(AO, %eax, 1), AO
kusano 2b45e8
	leal	( B, %eax, 2), BO
kusano 2b45e8
#endif	
kusano 2b45e8
kusano 2b45e8
	fldz
kusano 2b45e8
	fldz
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
	subl	KK, %eax
kusano 2b45e8
	movl	%eax, KKK	
kusano 2b45e8
#else
kusano 2b45e8
	movl	KK, %eax
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addl	$1, %eax
kusano 2b45e8
#else
kusano 2b45e8
	addl	$2, %eax
kusano 2b45e8
#endif
kusano 2b45e8
	movl	%eax, KKK
kusano 2b45e8
#endif
kusano 2b45e8
	sarl	$2, %eax
kusano 2b45e8
 	je	.L25
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L22:
kusano 2b45e8
	PREFETCH	(PREFETCHSIZE + 0) * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(BO)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -7 * SIZE(BO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -7 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -6 * SIZE(BO)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -5 * SIZE(BO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -6 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -4 * SIZE(BO)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -3 * SIZE(BO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -5 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -2 * SIZE(BO)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -1 * SIZE(BO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	addl	$4 * SIZE,AO
kusano 2b45e8
	addl	$8 * SIZE,BO
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jne	.L22
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L25:
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
#else
kusano 2b45e8
	movl	KKK, %eax
kusano 2b45e8
#endif
kusano 2b45e8
	and	$3,  %eax
kusano 2b45e8
	je	.L28
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L26:
kusano 2b45e8
	FLD	 -8 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(BO)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -7 * SIZE(BO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	addl	$1 * SIZE,AO
kusano 2b45e8
	addl	$2 * SIZE,BO
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jne	 .L26
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L28:
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	FLD	ALPHA
kusano 2b45e8
kusano 2b45e8
	fmul	%st, %st(1)
kusano 2b45e8
	fmulp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(CO)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(CO)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(CO, LDC)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(CO, LDC)
kusano 2b45e8
#else
kusano 2b45e8
	FST	0 * SIZE(CO)
kusano 2b45e8
	FST	0 * SIZE(CO, LDC)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
	subl	KKK, %eax
kusano 2b45e8
	sall	$BASE_SHIFT, %eax
kusano 2b45e8
	leal	(AO, %eax, 1), AO
kusano 2b45e8
	leal	(BO, %eax, 2), BO
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(TRMMKERNEL) && defined(LEFT)
kusano 2b45e8
	addl	$1, KK
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	addl	$1 * SIZE, CO
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L29:
kusano 2b45e8
#if defined(TRMMKERNEL) && !defined(LEFT)
kusano 2b45e8
	addl	$2, KK
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movl	BO, B
kusano 2b45e8
	decl	J
kusano 2b45e8
	jne	.L01
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L30:
kusano 2b45e8
	movl	N,  %eax
kusano 2b45e8
	testl	$1, %eax	
kusano 2b45e8
	je	.L999
kusano 2b45e8
kusano 2b45e8
#if defined(TRMMKERNEL) && defined(LEFT)
kusano 2b45e8
	movl	OFFSET, %eax
kusano 2b45e8
	movl	%eax, KK
kusano 2b45e8
#endif	
kusano 2b45e8
kusano 2b45e8
	movl	A, AO
kusano 2b45e8
kusano 2b45e8
	movl	C, CO
kusano 2b45e8
	addl	LDC, C
kusano 2b45e8
kusano 2b45e8
	movl	M,  I
kusano 2b45e8
	sarl	$1, I
kusano 2b45e8
	je	.L40
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L31:
kusano 2b45e8
#if !defined(TRMMKERNEL) || \
kusano 2b45e8
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	movl	B, BO
kusano 2b45e8
#else
kusano 2b45e8
	movl	KK,   %eax
kusano 2b45e8
	sall	$BASE_SHIFT, %eax
kusano 2b45e8
	leal	(AO, %eax, 2), AO
kusano 2b45e8
	leal	( B, %eax, 1), BO
kusano 2b45e8
#endif	
kusano 2b45e8
kusano 2b45e8
	fldz
kusano 2b45e8
	fldz
kusano 2b45e8
kusano 2b45e8
#if   defined(HAVE_3DNOW)
kusano 2b45e8
	prefetchw	2 * SIZE(CO)
kusano 2b45e8
#elif defined(HAVE_SSE)
kusano 2b45e8
	prefetchnta	2 * SIZE(CO)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
	subl	KK, %eax
kusano 2b45e8
	movl	%eax, KKK	
kusano 2b45e8
#else
kusano 2b45e8
	movl	KK, %eax
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addl	$2, %eax
kusano 2b45e8
#else
kusano 2b45e8
	addl	$1, %eax
kusano 2b45e8
#endif
kusano 2b45e8
	movl	%eax, KKK
kusano 2b45e8
#endif
kusano 2b45e8
	sarl	$2, %eax
kusano 2b45e8
 	je	.L35
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L32:
kusano 2b45e8
	PREFETCH	(PREFETCHSIZE + 0) * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(BO)
kusano 2b45e8
	FLD	 -8 * SIZE(AO)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -7 * SIZE(AO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -7 * SIZE(BO)
kusano 2b45e8
	FLD	 -6 * SIZE(AO)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -5 * SIZE(AO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -6 * SIZE(BO)
kusano 2b45e8
	FLD	 -4 * SIZE(AO)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -3 * SIZE(AO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -5 * SIZE(BO)
kusano 2b45e8
	FLD	 -2 * SIZE(AO)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -1 * SIZE(AO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	addl	$8 * SIZE,AO
kusano 2b45e8
	addl	$4 * SIZE,BO
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jne	.L32
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L35:
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
#else
kusano 2b45e8
	movl	KKK, %eax
kusano 2b45e8
#endif
kusano 2b45e8
	and	$3,  %eax
kusano 2b45e8
	je	.L38
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L36:
kusano 2b45e8
	FLD	 -8 * SIZE(BO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(AO)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 -7 * SIZE(AO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(2)
kusano 2b45e8
kusano 2b45e8
	addl	$2 * SIZE,AO
kusano 2b45e8
	addl	$1 * SIZE,BO
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jne	 .L36
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L38:
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	FLD	ALPHA
kusano 2b45e8
kusano 2b45e8
	fmul	%st, %st(1)
kusano 2b45e8
	fmulp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(CO)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(CO)
kusano 2b45e8
kusano 2b45e8
	FLD	1 * SIZE(CO)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	1 * SIZE(CO)
kusano 2b45e8
#else
kusano 2b45e8
	FST	0 * SIZE(CO)
kusano 2b45e8
	FST	1 * SIZE(CO)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
	subl	KKK, %eax
kusano 2b45e8
	sall	$BASE_SHIFT, %eax
kusano 2b45e8
	leal	(AO, %eax, 2), AO
kusano 2b45e8
	leal	(BO, %eax, 1), BO
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(TRMMKERNEL) && defined(LEFT)
kusano 2b45e8
	addl	$2, KK
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	addl	$2 * SIZE, CO
kusano 2b45e8
	decl	I
kusano 2b45e8
	jne	.L31
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L40:
kusano 2b45e8
	movl	 M, %eax
kusano 2b45e8
	andl	$1, %eax
kusano 2b45e8
	je	.L49
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L41:
kusano 2b45e8
#if !defined(TRMMKERNEL) || \
kusano 2b45e8
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	movl	B, BO
kusano 2b45e8
#else
kusano 2b45e8
	movl	KK,   %eax
kusano 2b45e8
	sall	$BASE_SHIFT, %eax
kusano 2b45e8
	leal	(AO, %eax, 1), AO
kusano 2b45e8
	leal	( B, %eax, 1), BO
kusano 2b45e8
#endif	
kusano 2b45e8
kusano 2b45e8
	fldz
kusano 2b45e8
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
	subl	KK, %eax
kusano 2b45e8
	movl	%eax, KKK	
kusano 2b45e8
#else
kusano 2b45e8
	movl	KK, %eax
kusano 2b45e8
#ifdef LEFT
kusano 2b45e8
	addl	$1, %eax
kusano 2b45e8
#else
kusano 2b45e8
	addl	$1, %eax
kusano 2b45e8
#endif
kusano 2b45e8
	movl	%eax, KKK
kusano 2b45e8
#endif
kusano 2b45e8
	sarl	$2, %eax
kusano 2b45e8
 	je	.L45
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L42:
kusano 2b45e8
	PREFETCH	(PREFETCHSIZE + 0) * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(AO)
kusano 2b45e8
	FLD	 -8 * SIZE(BO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	FLD	 -7 * SIZE(AO)
kusano 2b45e8
	FLD	 -7 * SIZE(BO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	FLD	 -6 * SIZE(AO)
kusano 2b45e8
	FLD	 -6 * SIZE(BO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	FLD	 -5 * SIZE(AO)
kusano 2b45e8
	FLD	 -5 * SIZE(BO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	addl	$4 * SIZE,AO
kusano 2b45e8
	addl	$4 * SIZE,BO
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jne	.L42
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L45:
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
#else
kusano 2b45e8
	movl	KKK, %eax
kusano 2b45e8
#endif
kusano 2b45e8
	and	$3,  %eax
kusano 2b45e8
	je	.L48
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L46:
kusano 2b45e8
	FLD	 -8 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(BO)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	addl	$1 * SIZE,AO
kusano 2b45e8
	addl	$1 * SIZE,BO
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jne	 .L46
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L48:
kusano 2b45e8
#ifndef TRMMKERNEL
kusano 2b45e8
	FLD	ALPHA
kusano 2b45e8
kusano 2b45e8
	fmulp	%st, %st(1)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(CO)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(CO)
kusano 2b45e8
#else
kusano 2b45e8
	FST	0 * SIZE(CO)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8
	movl	K, %eax
kusano 2b45e8
	subl	KKK, %eax
kusano 2b45e8
	sall	$BASE_SHIFT, %eax
kusano 2b45e8
	leal	(AO, %eax, 1), AO
kusano 2b45e8
	leal	(BO, %eax, 1), BO
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(TRMMKERNEL) && defined(LEFT)
kusano 2b45e8
	addl	$1, KK
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	addl	$1 * SIZE, CO
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L49:
kusano 2b45e8
#if defined(TRMMKERNEL) && !defined(LEFT)
kusano 2b45e8
	addl	$1, KK
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movl	BO, B
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L999:
kusano 2b45e8
	popl	%ebx
kusano 2b45e8
	popl	%esi
kusano 2b45e8
	popl	%edi
kusano 2b45e8
	popl	%ebp
kusano 2b45e8
	addl	$ARGS, %esp
kusano 2b45e8
	ret
kusano 2b45e8
kusano 2b45e8
	EPILOGUE