bw / opentoonz

Clone
Source Code
GIT

Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/x86/gemm_kernel_1x4.S

feature08b_assistants feature08f_assistants feature08g_assistants feature08h_assistants feature08j_assistants feature08k_assistants improve_tool_selection master
  1.   9f1fa31ac16c14ed1b6f99e7f2102c22373f895d
  2.   thirdparty
  3.   openblas
  4.   xianyi-OpenBLAS-e6e87a2
  5.   kernel
  6.   x86
  7.   gemm_kernel_1x4.S
Blob Raw
kusano 2b45e8 
/*********************************************************************/
kusano 2b45e8 
/* Copyright 2009, 2010 The University of Texas at Austin.           */
kusano 2b45e8 
/* All rights reserved.                                              */
kusano 2b45e8 
/*                                                                   */
kusano 2b45e8 
/* Redistribution and use in source and binary forms, with or        */
kusano 2b45e8 
/* without modification, are permitted provided that the following   */
kusano 2b45e8 
/* conditions are met:                                               */
kusano 2b45e8 
/*                                                                   */
kusano 2b45e8 
/*   1. Redistributions of source code must retain the above         */
kusano 2b45e8 
/*      copyright notice, this list of conditions and the following  */
kusano 2b45e8 
/*      disclaimer.                                                  */
kusano 2b45e8 
/*                                                                   */
kusano 2b45e8 
/*   2. Redistributions in binary form must reproduce the above      */
kusano 2b45e8 
/*      copyright notice, this list of conditions and the following  */
kusano 2b45e8 
/*      disclaimer in the documentation and/or other materials       */
kusano 2b45e8 
/*      provided with the distribution.                              */
kusano 2b45e8 
/*                                                                   */
kusano 2b45e8 
/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
kusano 2b45e8 
/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
kusano 2b45e8 
/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
kusano 2b45e8 
/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
kusano 2b45e8 
/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
kusano 2b45e8 
/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
kusano 2b45e8 
/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
kusano 2b45e8 
/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
kusano 2b45e8 
/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
kusano 2b45e8 
/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
kusano 2b45e8 
/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
kusano 2b45e8 
/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
kusano 2b45e8 
/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
kusano 2b45e8 
/*    POSSIBILITY OF SUCH DAMAGE.                                    */
kusano 2b45e8 
/*                                                                   */
kusano 2b45e8 
/* The views and conclusions contained in the software and           */
kusano 2b45e8 
/* documentation are those of the authors and should not be          */
kusano 2b45e8 
/* interpreted as representing official policies, either expressed   */
kusano 2b45e8 
/* or implied, of The University of Texas at Austin.                 */
kusano 2b45e8 
/*********************************************************************/
kusano 2b45e8
kusano 2b45e8 
#define ASSEMBLER
kusano 2b45e8 
#include "common.h"
kusano 2b45e8
kusano 2b45e8 
#define STACK	16
kusano 2b45e8 
#define ARGS	16
kusano 2b45e8
kusano 2b45e8 
#define J	 0 + STACK(%esp)
kusano 2b45e8 
#define I	 4 + STACK(%esp)
kusano 2b45e8 
#define KK	 8 + STACK(%esp)
kusano 2b45e8 
#define KKK	12 + 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 
#ifdef DOUBLE
kusano 2b45e8 
#define STACK_A	24 + STACK + ARGS(%esp)
kusano 2b45e8 
#define STACK_B	28 + STACK + ARGS(%esp)
kusano 2b45e8 
#define C	32 + STACK + ARGS(%esp)
kusano 2b45e8 
#define STACK_LDC	36 + STACK + ARGS(%esp)
kusano 2b45e8 
#define OFFSET	40 + STACK + ARGS(%esp)
kusano 2b45e8 
#else
kusano 2b45e8 
#define STACK_A	20 + STACK + ARGS(%esp)
kusano 2b45e8 
#define STACK_B	24 + STACK + ARGS(%esp)
kusano 2b45e8 
#define C	28 + STACK + ARGS(%esp)
kusano 2b45e8 
#define STACK_LDC	32 + STACK + ARGS(%esp)
kusano 2b45e8 
#define OFFSET	36 + STACK + ARGS(%esp)
kusano 2b45e8 
#endif
kusano 2b45e8
kusano 2b45e8 
#define A	%edx
kusano 2b45e8 
#define B	%ecx
kusano 2b45e8 
#define BB	%ebx
kusano 2b45e8 
#define LDC	%ebp
kusano 2b45e8 
#define BX	%esi
kusano 2b45e8
kusano 2b45e8 
#define PREFETCHSIZE (8 * 5 + 4)
kusano 2b45e8
kusano 2b45e8 
#define AOFFSET   1
kusano 2b45e8 
#define BOFFSET  -7
kusano 2b45e8
kusano 2b45e8 
#ifdef HAVE_3DNOW
kusano 2b45e8 
#define PREFETCH	prefetch
kusano 2b45e8 
#else
kusano 2b45e8 
#define PREFETCH	prefetcht0
kusano 2b45e8 
#endif
kusano 2b45e8
kusano 2b45e8 
#define KERNEL \
kusano 2b45e8 
	PREFETCH	PREFETCHSIZE * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	-15 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	-14 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	-13 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	-15 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	-12 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	-11 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	-10 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	 -9 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	-14 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	 -8 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	 -7 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	 -6 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	 -5 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	-13 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	 -4 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	 -3 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	 -2 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	 -1 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	-12 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	  0 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	  1 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	  2 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	  3 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	-11 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	  4 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	  5 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	  6 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	  7 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	-10 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	  8 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	  9 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	 10 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	 11 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	 -9 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	 12 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	 13 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	 14 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	 15 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	  8 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	fxch	%st(1);\
kusano 2b45e8 
	FLD	 16 * SIZE + BOFFSET(B, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	-15 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	PREFETCH	(PREFETCHSIZE + 8) * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	-14 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	-13 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	 -7 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	-12 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	-11 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	-10 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	 -9 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	 -6 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	 -8 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	 -7 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	 -6 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	 -5 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	 -5 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	 -4 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	 -3 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	 -2 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	 -1 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	 -4 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	  0 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	  1 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	  2 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	  3 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	 -3 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	  4 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	  5 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	  6 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	  7 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	 -2 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	  8 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	  9 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	 10 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	 11 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	 -1 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	FLD	 12 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(4);\
kusano 2b45e8 
	FLD	 13 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(5);\
kusano 2b45e8 
	FLD	 14 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	fmul	%st(1), %st;\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FMUL	 15 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	faddp	%st, %st(6);\
kusano 2b45e8 
	FLD	 16 * SIZE + AOFFSET(A, %eax, 1);\
kusano 2b45e8 
	fxch	%st(2);\
kusano 2b45e8 
	FLD	 16 * SIZE + BOFFSET(BB, %eax, 4);\
kusano 2b45e8 
	subl	 $-16 * SIZE, %eax
kusano 2b45e8
kusano 2b45e8 
/*
kusano 2b45e8
kusano 2b45e8 
  A hint of scheduling is received from following URL
kusano 2b45e8
kusano 2b45e8 
  http://www.netlib.org/atlas/atlas-comm/msg00260.html
kusano 2b45e8
kusano 2b45e8 
*/
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	STACK_LDC, LDC
kusano 2b45e8 
	leal	(, LDC, SIZE), LDC
kusano 2b45e8
kusano 2b45e8 
	subl	$(AOFFSET - 16 * SIZE), STACK_A
kusano 2b45e8 
	subl	$(BOFFSET - 16 * SIZE), STACK_B
kusano 2b45e8
kusano 2b45e8 
	movl	M, %eax
kusano 2b45e8 
	testl	%eax, %eax
kusano 2b45e8 
	jle	.L999
kusano 2b45e8
kusano 2b45e8 
	movl	N, %eax
kusano 2b45e8 
	testl	%eax, %eax
kusano 2b45e8 
	jle	.L999
kusano 2b45e8
kusano 2b45e8 
	movl	K, %eax
kusano 2b45e8 
	testl	%eax, %eax
kusano 2b45e8 
	jle	.L999
kusano 2b45e8
kusano 2b45e8 
	movl	N,   %eax
kusano 2b45e8 
	sarl	$2,  %eax
kusano 2b45e8 
	movl	%eax, J
kusano 2b45e8 
	je	.L20
kusano 2b45e8 
	ALIGN_3
kusano 2b45e8
kusano 2b45e8 
.L11:
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	STACK_A, A
kusano 2b45e8 
	movl	STACK_B, B
kusano 2b45e8 
	movl	C, %edi
kusano 2b45e8
kusano 2b45e8 
	movl	K, BX
kusano 2b45e8 
	sall	$BASE_SHIFT + 2, BX
kusano 2b45e8 
	addl	B, BX
kusano 2b45e8
kusano 2b45e8 
	movl	M, %eax
kusano 2b45e8 
	movl	%eax, I
kusano 2b45e8 
	ALIGN_3
kusano 2b45e8
kusano 2b45e8 
.L14:
kusano 2b45e8 
	prefetchnta	-16 * SIZE + BOFFSET(BX)
kusano 2b45e8 
	subl	$-8 * SIZE, BX
kusano 2b45e8
kusano 2b45e8 
	movl	STACK_B, B
kusano 2b45e8
kusano 2b45e8 
#if !defined(TRMMKERNEL) || \
kusano 2b45e8 
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8 
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8 
#else
kusano 2b45e8 
	movl	KK,   %eax
kusano 2b45e8 
	leal	(, %eax, SIZE), %eax
kusano 2b45e8 
	leal	(A, %eax, 1), A
kusano 2b45e8 
	leal	(B, %eax, 4), B
kusano 2b45e8 
#endif
kusano 2b45e8
kusano 2b45e8 
	leal	(%edi, LDC, 2), %eax
kusano 2b45e8
kusano 2b45e8 
	fldz
kusano 2b45e8 
	fldz
kusano 2b45e8 
	fldz
kusano 2b45e8 
	fldz
kusano 2b45e8
kusano 2b45e8 
	FLD	   0 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	  -8 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	 -16 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	 -16 * SIZE + BOFFSET(B)
kusano 2b45e8
kusano 2b45e8 
#ifdef HAVE_3DNOW
kusano 2b45e8 
	prefetchw	1 * SIZE(%edi)
kusano 2b45e8 
	prefetchw	2 * SIZE(%edi, LDC)
kusano 2b45e8 
	prefetchw	1 * SIZE(%eax)
kusano 2b45e8 
	prefetchw	2 * SIZE(%eax, LDC)
kusano 2b45e8 
#elif defined(HAVE_SSE)
kusano 2b45e8 
	prefetcht0	1 * SIZE(%edi)
kusano 2b45e8 
	prefetcht0	2 * SIZE(%edi, LDC)
kusano 2b45e8 
	prefetcht0	1 * SIZE(%eax)
kusano 2b45e8 
	prefetcht0	2 * SIZE(%eax, LDC)
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	$1, %eax
kusano 2b45e8 
#else
kusano 2b45e8 
	addl	$4, %eax
kusano 2b45e8 
#endif
kusano 2b45e8 
	movl	%eax, KKK
kusano 2b45e8 
#endif
kusano 2b45e8
kusano 2b45e8 
	andl	$-16, %eax
kusano 2b45e8
kusano 2b45e8 
	leal	(, %eax, SIZE), %eax
kusano 2b45e8 
	leal	(A, %eax, 1), A
kusano 2b45e8 
	leal	32 * SIZE(B, %eax, 4), BB
kusano 2b45e8 
	leal	(B, %eax, 4), B
kusano 2b45e8 
	negl	%eax
kusano 2b45e8 
	NOBRANCH
kusano 2b45e8 
 	je	.L16
kusano 2b45e8 
	ALIGN_4
kusano 2b45e8
kusano 2b45e8 
.L15:
kusano 2b45e8 
	KERNEL
kusano 2b45e8 
	jge	.L16
kusano 2b45e8 
	KERNEL
kusano 2b45e8 
	jge	.L16
kusano 2b45e8 
	KERNEL
kusano 2b45e8 
	jge	.L16
kusano 2b45e8 
	KERNEL
kusano 2b45e8 
	jl	.L15
kusano 2b45e8 
	ALIGN_4
kusano 2b45e8
kusano 2b45e8 
.L16:
kusano 2b45e8 
#ifndef TRMMKERNEL
kusano 2b45e8 
	movl	K, %eax
kusano 2b45e8 
#else
kusano 2b45e8 
	movl	KKK, %eax
kusano 2b45e8 
#endif
kusano 2b45e8 
	and	$15, %eax
kusano 2b45e8 
	je	.L19
kusano 2b45e8 
	ALIGN_4
kusano 2b45e8
kusano 2b45e8 
.L17:
kusano 2b45e8 
	fmul	%st(1), %st
kusano 2b45e8 
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8 
	FLD	-15 * SIZE + BOFFSET(B)
kusano 2b45e8 
	fmul	%st(1), %st
kusano 2b45e8 
	faddp	%st, %st(5)
kusano 2b45e8
kusano 2b45e8 
	FLD	-14 * SIZE + BOFFSET(B)
kusano 2b45e8 
	fmul	%st(1), %st
kusano 2b45e8 
	faddp	%st, %st(6)
kusano 2b45e8
kusano 2b45e8 
	FMUL	-13 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(6)
kusano 2b45e8 
	FLD	-15 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	-12 * SIZE + BOFFSET(B)
kusano 2b45e8
kusano 2b45e8 
	addl	$1 * SIZE,A
kusano 2b45e8 
	addl	$4 * SIZE,B
kusano 2b45e8
kusano 2b45e8 
	decl	%eax
kusano 2b45e8 
	jne	 .L17
kusano 2b45e8 
	ALIGN_4
kusano 2b45e8
kusano 2b45e8 
.L19:
kusano 2b45e8 
	ffreep	%st(0)
kusano 2b45e8 
	ffreep	%st(0)
kusano 2b45e8 
	ffreep	%st(0)
kusano 2b45e8 
	ffreep	%st(0)
kusano 2b45e8
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 
	leal	(%edi, LDC, 2), %eax
kusano 2b45e8
kusano 2b45e8 
#ifndef TRMMKERNEL
kusano 2b45e8 
	FADD	(%edi)
kusano 2b45e8 
	FST	(%edi)
kusano 2b45e8 
	FADD	(%edi,LDC)
kusano 2b45e8 
	FST	(%edi,LDC)
kusano 2b45e8 
	FADD	(%eax)
kusano 2b45e8 
	FST	(%eax)
kusano 2b45e8 
	FADD	(%eax,LDC)
kusano 2b45e8 
	FST	(%eax,LDC)
kusano 2b45e8 
#else
kusano 2b45e8 
	FST	(%edi)
kusano 2b45e8 
	FST	(%edi,LDC)
kusano 2b45e8 
	FST	(%eax)
kusano 2b45e8 
	FST	(%eax,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 
	leal	(,%eax, SIZE), %eax
kusano 2b45e8 
	leal	(A, %eax, 1), A
kusano 2b45e8 
	leal	(B, %eax, 4), B
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, %edi
kusano 2b45e8 
	decl	I
kusano 2b45e8 
	jne	.L14
kusano 2b45e8
kusano 2b45e8 
#if defined(TRMMKERNEL) && !defined(LEFT)
kusano 2b45e8 
	addl	$4, KK
kusano 2b45e8 
#endif
kusano 2b45e8
kusano 2b45e8 
	leal	(, LDC, 4), %eax
kusano 2b45e8 
	addl	%eax, C
kusano 2b45e8 
	movl	B, STACK_B
kusano 2b45e8 
	decl	J
kusano 2b45e8 
	jne	.L11
kusano 2b45e8 
	ALIGN_4
kusano 2b45e8
kusano 2b45e8 
.L20:
kusano 2b45e8 
	movl	N,   %eax
kusano 2b45e8 
	andl	$2,  %eax
kusano 2b45e8 
	je	.L30
kusano 2b45e8 
	ALIGN_3
kusano 2b45e8
kusano 2b45e8 
.L21:
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	STACK_A, A
kusano 2b45e8 
	movl	STACK_B, B
kusano 2b45e8 
	movl	C, %edi
kusano 2b45e8
kusano 2b45e8 
	movl	M, %eax
kusano 2b45e8 
	movl	%eax, I
kusano 2b45e8 
	ALIGN_3
kusano 2b45e8
kusano 2b45e8 
.L24:
kusano 2b45e8 
	movl	STACK_B, B
kusano 2b45e8
kusano 2b45e8 
#if !defined(TRMMKERNEL) || \
kusano 2b45e8 
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8 
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8 
#else
kusano 2b45e8 
	movl	KK,   %eax
kusano 2b45e8 
	leal	(, %eax, SIZE), %eax
kusano 2b45e8 
	leal	(A, %eax, 1), A
kusano 2b45e8 
	leal	(B, %eax, 2), B
kusano 2b45e8 
#endif
kusano 2b45e8
kusano 2b45e8 
	fldz
kusano 2b45e8 
	fldz
kusano 2b45e8 
	fldz
kusano 2b45e8 
	fldz
kusano 2b45e8
kusano 2b45e8 
	FLD	 -16 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	 -16 * SIZE + BOFFSET(B)
kusano 2b45e8
kusano 2b45e8 
	prefetchw	1 * SIZE(%edi)
kusano 2b45e8 
	prefetchw	1 * SIZE(%edi, LDC)
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	$3, %eax
kusano 2b45e8 
 	je	.L26
kusano 2b45e8 
	ALIGN_3
kusano 2b45e8
kusano 2b45e8 
.L25:
kusano 2b45e8 
	fmul	%st(1), %st
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FMUL	-15 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FLD	-15 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	-14 * SIZE + BOFFSET(B)
kusano 2b45e8
kusano 2b45e8 
	fmul	%st(1), %st
kusano 2b45e8 
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8 
	FMUL	-13 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8 
	FLD	-14 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	-12 * SIZE + BOFFSET(B)
kusano 2b45e8
kusano 2b45e8 
	fmul	%st(1), %st
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FMUL	-11 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FLD	-13 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	-10 * SIZE + BOFFSET(B)
kusano 2b45e8
kusano 2b45e8 
	fmul	%st(1), %st
kusano 2b45e8 
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8 
	FMUL	 -9 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8 
	FLD	-12 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	 -8 * SIZE + BOFFSET(B)
kusano 2b45e8
kusano 2b45e8 
	fmul	%st(1), %st
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FMUL	 -7 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FLD	-11 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	 -6 * SIZE + BOFFSET(B)
kusano 2b45e8
kusano 2b45e8 
	fmul	%st(1), %st
kusano 2b45e8 
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8 
	FMUL	 -5 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8 
	FLD	-10 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	 -4 * SIZE + BOFFSET(B)
kusano 2b45e8
kusano 2b45e8 
	fmul	%st(1), %st
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FMUL	 -3 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FLD	 -9 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	 -2 * SIZE + BOFFSET(B)
kusano 2b45e8
kusano 2b45e8 
	fmul	%st(1), %st
kusano 2b45e8 
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8 
	FMUL	 -1 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8 
	FLD	 -8 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	  0 * SIZE + BOFFSET(B)
kusano 2b45e8
kusano 2b45e8 
	addl	$  8 * SIZE, A
kusano 2b45e8 
	subl	$-16 * SIZE, B
kusano 2b45e8
kusano 2b45e8 
	decl	%eax
kusano 2b45e8 
	jne	.L25
kusano 2b45e8 
	ALIGN_4
kusano 2b45e8
kusano 2b45e8 
.L26:
kusano 2b45e8 
#ifndef TRMMKERNEL
kusano 2b45e8 
	movl	K, %eax
kusano 2b45e8 
#else
kusano 2b45e8 
	movl	KKK, %eax
kusano 2b45e8 
#endif
kusano 2b45e8 
	and	$7, %eax
kusano 2b45e8 
	je	.L29
kusano 2b45e8 
	ALIGN_4
kusano 2b45e8
kusano 2b45e8 
.L27:
kusano 2b45e8 
	fmul	%st(1), %st
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FMUL	-15 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FLD	 -15 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FLD	 -14 * SIZE + BOFFSET(B)
kusano 2b45e8
kusano 2b45e8 
	addl	$1 * SIZE,A
kusano 2b45e8 
	addl	$2 * SIZE,B
kusano 2b45e8
kusano 2b45e8 
	decl	%eax
kusano 2b45e8 
	jne	 .L27
kusano 2b45e8 
	ALIGN_4
kusano 2b45e8
kusano 2b45e8 
.L29:
kusano 2b45e8 
	ffreep	%st(0)
kusano 2b45e8 
	ffreep	%st(0)
kusano 2b45e8
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FLD	ALPHA
kusano 2b45e8
kusano 2b45e8 
	fmul	%st, %st(1)
kusano 2b45e8 
	fmulp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
#ifndef TRMMKERNEL
kusano 2b45e8 
	FADD	(%edi)
kusano 2b45e8 
	FST	(%edi)
kusano 2b45e8 
	FADD	(%edi,LDC)
kusano 2b45e8 
	FST	(%edi,LDC)
kusano 2b45e8 
#else
kusano 2b45e8 
	FST	(%edi)
kusano 2b45e8 
	FST	(%edi,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 
	leal	(,%eax, SIZE), %eax
kusano 2b45e8 
	leal	(A, %eax, 1), A
kusano 2b45e8 
	leal	(B, %eax, 2), B
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, %edi
kusano 2b45e8 
	decl	I
kusano 2b45e8 
	jne	.L24
kusano 2b45e8
kusano 2b45e8 
#if defined(TRMMKERNEL) && !defined(LEFT)
kusano 2b45e8 
	addl	$2, KK
kusano 2b45e8 
#endif
kusano 2b45e8
kusano 2b45e8 
	leal	(, LDC, 2), %eax
kusano 2b45e8 
	addl	%eax, C
kusano 2b45e8 
	movl	B, STACK_B
kusano 2b45e8 
	ALIGN_4
kusano 2b45e8
kusano 2b45e8 
.L30:
kusano 2b45e8 
	movl	N,   %eax
kusano 2b45e8 
	andl	$1,  %eax
kusano 2b45e8 
	je	.L999
kusano 2b45e8 
	ALIGN_3
kusano 2b45e8
kusano 2b45e8 
.L31:
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	STACK_A, A
kusano 2b45e8 
	movl	STACK_B, B
kusano 2b45e8 
	movl	C, %edi
kusano 2b45e8
kusano 2b45e8 
	movl	M, %eax
kusano 2b45e8 
	movl	%eax, I
kusano 2b45e8 
	ALIGN_3
kusano 2b45e8
kusano 2b45e8 
.L34:
kusano 2b45e8 
	movl	STACK_B, B
kusano 2b45e8
kusano 2b45e8 
#if !defined(TRMMKERNEL) || \
kusano 2b45e8 
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
kusano 2b45e8 
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
kusano 2b45e8 
#else
kusano 2b45e8 
	movl	KK,   %eax
kusano 2b45e8 
	leal	(, %eax, SIZE), %eax
kusano 2b45e8 
	leal	(A, %eax, 1), A
kusano 2b45e8 
	leal	(B, %eax, 1), B
kusano 2b45e8 
#endif
kusano 2b45e8
kusano 2b45e8 
	fldz
kusano 2b45e8 
	fldz
kusano 2b45e8 
	fldz
kusano 2b45e8 
	fldz
kusano 2b45e8
kusano 2b45e8 
	prefetchw	1 * SIZE(%edi)
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	$3, %eax
kusano 2b45e8 
 	je	.L36
kusano 2b45e8 
	ALIGN_3
kusano 2b45e8
kusano 2b45e8 
.L35:
kusano 2b45e8 
	FLD	-16 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FMUL	-16 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(1)
kusano 2b45e8
kusano 2b45e8 
	FLD	-15 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FMUL	-15 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FLD	-14 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FMUL	-14 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(3)
kusano 2b45e8
kusano 2b45e8 
	FLD	-13 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FMUL	-13 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8 
	FLD	-12 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FMUL	-12 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(1)
kusano 2b45e8
kusano 2b45e8 
	FLD	-11 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FMUL	-11 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8 
	FLD	-10 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FMUL	-10 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(3)
kusano 2b45e8
kusano 2b45e8 
	FLD	 -9 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FMUL	 -9 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8 
	addl	$8 * SIZE, A
kusano 2b45e8 
	addl	$8 * SIZE, B
kusano 2b45e8
kusano 2b45e8 
	decl	%eax
kusano 2b45e8 
	jne	.L35
kusano 2b45e8 
	ALIGN_4
kusano 2b45e8
kusano 2b45e8 
.L36:
kusano 2b45e8 
#ifndef TRMMKERNEL
kusano 2b45e8 
	movl	K, %eax
kusano 2b45e8 
#else
kusano 2b45e8 
	movl	KKK, %eax
kusano 2b45e8 
#endif
kusano 2b45e8 
	and	$7, %eax
kusano 2b45e8 
	je	.L39
kusano 2b45e8 
	ALIGN_4
kusano 2b45e8
kusano 2b45e8 
.L37:
kusano 2b45e8 
	FLD	-16 * SIZE + AOFFSET(A)
kusano 2b45e8 
	FMUL	-16 * SIZE + BOFFSET(B)
kusano 2b45e8 
	faddp	%st, %st(1)
kusano 2b45e8
kusano 2b45e8 
	addl	$1 * SIZE,A
kusano 2b45e8 
	addl	$1 * SIZE,B
kusano 2b45e8 
	decl	%eax
kusano 2b45e8 
	jne	 .L37
kusano 2b45e8 
	ALIGN_4
kusano 2b45e8
kusano 2b45e8 
.L39:
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8 
	faddp	%st, %st(2)
kusano 2b45e8 
	faddp	%st, %st(1)
kusano 2b45e8
kusano 2b45e8 
	FMUL	ALPHA
kusano 2b45e8
kusano 2b45e8 
#ifndef TRMMKERNEL
kusano 2b45e8 
	FADD	(%edi)
kusano 2b45e8 
	FST	(%edi)
kusano 2b45e8 
#else
kusano 2b45e8 
	FST	(%edi)
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 
	leal	(,%eax, SIZE), %eax
kusano 2b45e8 
	leal	(A, %eax, 1), A
kusano 2b45e8 
	leal	(B, %eax, 1), B
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, %edi
kusano 2b45e8 
	decl	I
kusano 2b45e8 
	jne	.L34
kusano 2b45e8
kusano 2b45e8 
#if defined(TRMMKERNEL) && !defined(LEFT)
kusano 2b45e8 
	addl	$1, KK
kusano 2b45e8 
#endif
kusano 2b45e8
kusano 2b45e8 
	addl	LDC, C
kusano 2b45e8 
	movl	B, STACK_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

Powered by Pagure 5.7.4

SSH Hostkey/Fingerprint | Documentation | About

© 2014-2019 Red Hat, Inc. and others.

Logo by Pavel Lashutin