Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/power/snrm2_ppc440.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
#define N	r3
kusano 2b45e8
#define X	r4
kusano 2b45e8
#define INCX	r5	
kusano 2b45e8
kusano 2b45e8
#define PRE	r8
kusano 2b45e8
kusano 2b45e8
#define FZERO	144(SP)
kusano 2b45e8
#define FONE	148(SP)
kusano 2b45e8
#define C1	152(SP)
kusano 2b45e8
#define C2	156(SP)
kusano 2b45e8
kusano 2b45e8
#define STACKSIZE 160
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	PROFCODE
kusano 2b45e8
kusano 2b45e8
	addi	SP, SP, -STACKSIZE
kusano 2b45e8
kusano 2b45e8
	li	r10,   0
kusano 2b45e8
	lis	r11,   0x3f80
kusano 2b45e8
	lis	r6,    0x3f00
kusano 2b45e8
	lis	r7,    0x4040
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
	stw	r10,  FZERO
kusano 2b45e8
	stw	r11,  FONE
kusano 2b45e8
	stw	r6,   C1
kusano 2b45e8
	stw	r7,   C2
kusano 2b45e8
kusano 2b45e8
	lfs	f1,   FZERO
kusano 2b45e8
kusano 2b45e8
#ifdef F_INTERFACE
kusano 2b45e8
	LDINT	N,    0(N)
kusano 2b45e8
	LDINT	INCX, 0(INCX)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	slwi	INCX, INCX, BASE_SHIFT
kusano 2b45e8
	li	PRE, 3 * 16 * SIZE 
kusano 2b45e8
kusano 2b45e8
	sub	X, X, INCX
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, N, 0
kusano 2b45e8
	ble-	LL(999)
kusano 2b45e8
	cmpwi	cr0, INCX, 0
kusano 2b45e8
	ble-	LL(999)
kusano 2b45e8
kusano 2b45e8
	fmr	f0,  f1
kusano 2b45e8
	fmr	f2,  f1
kusano 2b45e8
	fmr	f3,  f1
kusano 2b45e8
	fmr	f4,  f1
kusano 2b45e8
	fmr	f5,  f1
kusano 2b45e8
	fmr	f6,  f1
kusano 2b45e8
	fmr	f7,  f1
kusano 2b45e8
	fmr	f8,  f1
kusano 2b45e8
	fmr	f9,  f1
kusano 2b45e8
	fmr	f10, f1
kusano 2b45e8
	fmr	f11, f1
kusano 2b45e8
	fmr	f12, f1
kusano 2b45e8
	fmr	f13, f1
kusano 2b45e8
	fmr	f14, f1
kusano 2b45e8
	fmr	f15, f1
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, N, 4
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	beq-	cr0, LL(50)
kusano 2b45e8
kusano 2b45e8
	LFDUX	f16, X, INCX
kusano 2b45e8
	LFDUX	f17, X, INCX
kusano 2b45e8
	LFDUX	f18, X, INCX
kusano 2b45e8
	LFDUX	f19, X, INCX
kusano 2b45e8
	LFDUX	f20, X, INCX
kusano 2b45e8
	LFDUX	f21, X, INCX
kusano 2b45e8
	LFDUX	f22, X, INCX
kusano 2b45e8
	LFDUX	f23, X, INCX
kusano 2b45e8
kusano 2b45e8
	LFDUX	f24, X, INCX
kusano 2b45e8
	LFDUX	f25, X, INCX
kusano 2b45e8
	LFDUX	f26, X, INCX
kusano 2b45e8
	LFDUX	f27, X, INCX
kusano 2b45e8
kusano 2b45e8
	LFDUX	f28, X, INCX
kusano 2b45e8
	LFDUX	f29, X, INCX
kusano 2b45e8
	LFDUX	f30, X, INCX
kusano 2b45e8
	LFDUX	f31, X, INCX
kusano 2b45e8
	bdz	LL(20)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(10):
kusano 2b45e8
	fmadd	f0,  f16, f16, f0
kusano 2b45e8
	LFDUX	f16, X, INCX
kusano 2b45e8
	fmadd	f1,  f17, f17, f1
kusano 2b45e8
	LFDUX	f17, X, INCX
kusano 2b45e8
	fmadd	f2,  f18, f18, f2
kusano 2b45e8
	LFDUX	f18, X, INCX
kusano 2b45e8
	fmadd	f3,  f19, f19, f3
kusano 2b45e8
	LFDUX	f19, X, INCX
kusano 2b45e8
kusano 2b45e8
#ifdef PPCG4
kusano 2b45e8
	dcbt	X, PRE
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	fmadd	f4,  f20, f20, f4
kusano 2b45e8
	LFDUX	f20, X, INCX
kusano 2b45e8
	fmadd	f5,  f21, f21, f5
kusano 2b45e8
	LFDUX	f21, X, INCX
kusano 2b45e8
	fmadd	f6,  f22, f22, f6
kusano 2b45e8
	LFDUX	f22, X, INCX
kusano 2b45e8
	fmadd	f7,  f23, f23, f7
kusano 2b45e8
	LFDUX	f23, X, INCX
kusano 2b45e8
kusano 2b45e8
	fmadd	f8,  f24, f24, f8
kusano 2b45e8
	LFDUX	f24, X, INCX
kusano 2b45e8
	fmadd	f9,  f25, f25, f9
kusano 2b45e8
	LFDUX	f25, X, INCX
kusano 2b45e8
	fmadd	f10, f26, f26, f10
kusano 2b45e8
	LFDUX	f26, X, INCX
kusano 2b45e8
	fmadd	f11, f27, f27, f11
kusano 2b45e8
	LFDUX	f27, X, INCX
kusano 2b45e8
kusano 2b45e8
#ifdef PPCG4
kusano 2b45e8
	dcbt	X, PRE
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	fmadd	f12, f28, f28, f12
kusano 2b45e8
	LFDUX	f28, X, INCX
kusano 2b45e8
	fmadd	f13, f29, f29, f13
kusano 2b45e8
	LFDUX	f29, X, INCX
kusano 2b45e8
	fmadd	f14, f30, f30, f14
kusano 2b45e8
	LFDUX	f30, X, INCX
kusano 2b45e8
	fmadd	f15, f31, f31, f15
kusano 2b45e8
	LFDUX	f31, X, INCX
kusano 2b45e8
kusano 2b45e8
	bdnz	LL(10)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(20):
kusano 2b45e8
	fmadd	f0,  f16, f16, f0
kusano 2b45e8
	fmadd	f1,  f17, f17, f1
kusano 2b45e8
	fmadd	f2,  f18, f18, f2
kusano 2b45e8
	fmadd	f3,  f19, f19, f3
kusano 2b45e8
kusano 2b45e8
	fmadd	f4,  f20, f20, f4
kusano 2b45e8
	fmadd	f5,  f21, f21, f5
kusano 2b45e8
	fmadd	f6,  f22, f22, f6
kusano 2b45e8
	fmadd	f7,  f23, f23, f7
kusano 2b45e8
kusano 2b45e8
	fmadd	f8,  f24, f24, f8
kusano 2b45e8
	fmadd	f9,  f25, f25, f9
kusano 2b45e8
	fmadd	f10, f26, f26, f10
kusano 2b45e8
	fmadd	f11, f27, f27, f11
kusano 2b45e8
kusano 2b45e8
	fmadd	f12, f28, f28, f12
kusano 2b45e8
	fmadd	f13, f29, f29, f13
kusano 2b45e8
	fmadd	f14, f30, f30, f14
kusano 2b45e8
	fmadd	f15, f31, f31, f15
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(50):
kusano 2b45e8
	andi.	r0,  N, 15
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	beq-	cr0, LL(70)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(60):
kusano 2b45e8
	LFDUX	f16, X, INCX
kusano 2b45e8
	fmadd	f0,  f16, f16, f0
kusano 2b45e8
	bdnz	LL(60)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(70):
kusano 2b45e8
	fadd   f0,  f0,  f1
kusano 2b45e8
	fadd   f2,  f2,  f3
kusano 2b45e8
	fadd   f4,  f4,  f5
kusano 2b45e8
	fadd   f6,  f6,  f7
kusano 2b45e8
kusano 2b45e8
	fadd   f8,  f8,  f9
kusano 2b45e8
	fadd   f10, f10, f11
kusano 2b45e8
	fadd   f12, f12, f13
kusano 2b45e8
	fadd   f14, f14, f15
kusano 2b45e8
kusano 2b45e8
	fadd   f0,  f0,  f2
kusano 2b45e8
	fadd   f4,  f4,  f6
kusano 2b45e8
	fadd   f8,  f8,  f10
kusano 2b45e8
	fadd   f12, f12, f14
kusano 2b45e8
kusano 2b45e8
	fadd   f0,  f0,  f4
kusano 2b45e8
	fadd   f8,  f8,  f12
kusano 2b45e8
kusano 2b45e8
	fadd   f1,  f0,  f8
kusano 2b45e8
	lfs    f4,   FZERO
kusano 2b45e8
kusano 2b45e8
	fcmpu	cr0, f1, f4
kusano 2b45e8
	beq	cr0, LL(999)
kusano 2b45e8
kusano 2b45e8
	frsqrte f0, f1
kusano 2b45e8
	lfs	f8, C1
kusano 2b45e8
	lfs	f9, C2
kusano 2b45e8
kusano 2b45e8
	fmul	f2, f1, f0
kusano 2b45e8
	fadd	f7, f8, f8
kusano 2b45e8
	fmul	f3, f0, f8
kusano 2b45e8
	fnmsub	f4, f2, f0, f9
kusano 2b45e8
	fmul	f0, f3, f4
kusano 2b45e8
kusano 2b45e8
	fmul	f5, f1, f0
kusano 2b45e8
	fmul	f2, f5, f8
kusano 2b45e8
	fnmsub	f3, f5, f0, f7
kusano 2b45e8
	fmadd	f1, f2, f3, f5
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(999):
kusano 2b45e8
	lfd	f14,    0(SP)
kusano 2b45e8
	lfd	f15,    8(SP)
kusano 2b45e8
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
	addi	SP, SP, STACKSIZE
kusano 2b45e8
	blr
kusano 2b45e8
kusano 2b45e8
	EPILOGUE