Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/alpha/zgemm_beta.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
#include "version.h"
kusano 2b45e8
kusano 2b45e8
	.set noat
kusano 2b45e8
	.set noreorder
kusano 2b45e8
.text
kusano 2b45e8
	.align 5
kusano 2b45e8
	.globl CNAME
kusano 2b45e8
	.ent CNAME
kusano 2b45e8
CNAME:
kusano 2b45e8
	.frame	$sp, 0, $26, 0
kusano 2b45e8
kusano 2b45e8
#ifdef PROFILE
kusano 2b45e8
	ldgp	$gp, 0($27)
kusano 2b45e8
	lda	$28, _mcount
kusano 2b45e8
	jsr	$28, ($28), _mcount
kusano 2b45e8
	.prologue 1
kusano 2b45e8
#else
kusano 2b45e8
	.prologue 0
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	ldq	$18,   24($sp)
kusano 2b45e8
	ble	$16, $End
kusano 2b45e8
	ldl	$19,   32($sp)
kusano 2b45e8
	ble	$17, $End
kusano 2b45e8
kusano 2b45e8
	addq	$19, $19, $19
kusano 2b45e8
	fbne	$f19,$Main
kusano 2b45e8
	fbne	$f20,$Main
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L13:
kusano 2b45e8
	mov	$18, $1
kusano 2b45e8
	lda	$17,  -1($17)
kusano 2b45e8
	SXADDQ	$19, $18, $18
kusano 2b45e8
	mov	$16, $2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L12:
kusano 2b45e8
	ST	$f31, 0*SIZE($1)
kusano 2b45e8
	ST	$f31, 1*SIZE($1)
kusano 2b45e8
	lda	$2,  -1($2)
kusano 2b45e8
	lda	$1,   2*SIZE($1)
kusano 2b45e8
	bgt	$2, $L12
kusano 2b45e8
	bgt	$17,$L13
kusano 2b45e8
	clr	$0
kusano 2b45e8
	ret
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
/* Main Routine */
kusano 2b45e8
$Main:
kusano 2b45e8
	sra	$16, 1, $2		# $2 = (m >> 1)
kusano 2b45e8
	mov	$18, $1			# c_offset = c
kusano 2b45e8
	lda	$17,  -1($17)		# n --
kusano 2b45e8
	SXADDQ	$19, $18, $18		# c += ldc
kusano 2b45e8
	beq	$2, $L18
kusano 2b45e8
kusano 2b45e8
	LD	$f14, 0*SIZE($1)
kusano 2b45e8
	LD	$f15, 1*SIZE($1)
kusano 2b45e8
	LD	$f24, 2*SIZE($1)
kusano 2b45e8
	LD	$f25, 3*SIZE($1)
kusano 2b45e8
	lda	$2,    -1($2)		# $2 --
kusano 2b45e8
	ble	$2, $L19
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
$L23:
kusano 2b45e8
	MUL	$f19, $f14, $f10
kusano 2b45e8
	lds	$f31,  9*SIZE($1)
kusano 2b45e8
	MUL	$f20, $f15, $f11
kusano 2b45e8
	lda	$2,    -1($2)
kusano 2b45e8
kusano 2b45e8
	MUL	$f19, $f15, $f12
kusano 2b45e8
	LD	$f15, 5*SIZE($1)
kusano 2b45e8
	MUL	$f20, $f14, $f13
kusano 2b45e8
	LD	$f14, 4*SIZE($1)
kusano 2b45e8
kusano 2b45e8
	MUL	$f19, $f24, $f16
kusano 2b45e8
	unop
kusano 2b45e8
	MUL	$f20, $f25, $f17
kusano 2b45e8
	unop
kusano 2b45e8
kusano 2b45e8
	MUL	$f19, $f25, $f18
kusano 2b45e8
	LD	$f25, 7*SIZE($1)
kusano 2b45e8
	SUB	$f10, $f11, $f22
kusano 2b45e8
	unop
kusano 2b45e8
kusano 2b45e8
	MUL	$f20, $f24, $f21
kusano 2b45e8
	LD	$f24, 6*SIZE($1)
kusano 2b45e8
	ADD	$f12, $f13, $f23
kusano 2b45e8
	lda	$1,   4*SIZE($1)
kusano 2b45e8
kusano 2b45e8
	SUB	$f16, $f17, $f26
kusano 2b45e8
	ADD	$f18, $f21, $f27
kusano 2b45e8
	ST	$f22,-4*SIZE($1)
kusano 2b45e8
	ST	$f23,-3*SIZE($1)
kusano 2b45e8
kusano 2b45e8
	ST	$f26,-2*SIZE($1)
kusano 2b45e8
	ST	$f27,-1*SIZE($1)
kusano 2b45e8
	unop
kusano 2b45e8
	bgt	$2,$L23
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L19:
kusano 2b45e8
	MUL	$f19, $f14, $f10
kusano 2b45e8
	MUL	$f20, $f15, $f11
kusano 2b45e8
	MUL	$f19, $f15, $f12
kusano 2b45e8
	MUL	$f20, $f14, $f13
kusano 2b45e8
kusano 2b45e8
	MUL	$f19, $f24, $f16
kusano 2b45e8
	MUL	$f20, $f25, $f17
kusano 2b45e8
	MUL	$f19, $f25, $f18
kusano 2b45e8
	MUL	$f20, $f24, $f21
kusano 2b45e8
kusano 2b45e8
	SUB	$f10, $f11, $f22
kusano 2b45e8
	ADD	$f12, $f13, $f23
kusano 2b45e8
	SUB	$f16, $f17, $f26
kusano 2b45e8
	ADD	$f18, $f21, $f27
kusano 2b45e8
	lda	$1,   4*SIZE($1)
kusano 2b45e8
kusano 2b45e8
	ST	$f22, -4*SIZE($1)
kusano 2b45e8
	ST	$f23, -3*SIZE($1)
kusano 2b45e8
	ST	$f26, -2*SIZE($1)
kusano 2b45e8
	ST	$f27, -1*SIZE($1)
kusano 2b45e8
kusano 2b45e8
	blbs	$16, $L18
kusano 2b45e8
	bgt	$17, $Main
kusano 2b45e8
	clr	$0
kusano 2b45e8
	ret
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$L18:
kusano 2b45e8
	LD	$f14, 0*SIZE($1)
kusano 2b45e8
	LD	$f15, 1*SIZE($1)
kusano 2b45e8
	MUL	$f19, $f15, $f13
kusano 2b45e8
	MUL	$f20, $f14, $f10
kusano 2b45e8
kusano 2b45e8
	MUL	$f19, $f14, $f12
kusano 2b45e8
	MUL	$f20, $f15, $f11
kusano 2b45e8
	ADD	$f13, $f10, $f26
kusano 2b45e8
	SUB	$f12, $f11, $f27
kusano 2b45e8
kusano 2b45e8
	ST	$f26, 1*SIZE($1)
kusano 2b45e8
	ST	$f27, 0*SIZE($1)
kusano 2b45e8
	lda	$1,   2*SIZE($1)
kusano 2b45e8
	bgt	$17, $Main
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
$End:
kusano 2b45e8
	clr	$0
kusano 2b45e8
	ret
kusano 2b45e8
	.ident	VERSION
kusano 2b45e8
	.end CNAME