rockbox/apps/eq_arm.S

90 lines
3.5 KiB
ArmAsm
Raw Normal View History

/***************************************************************************
* __________ __ ___.
* Open \______ \ ____ ____ | | _\_ |__ _______ ___
* Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ /
* Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < <
* Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \
* \/ \/ \/ \/ \/
* $Id$
*
* Copyright (C) 2006-2007 Thom Johansen
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* as published by the Free Software Foundation; either version 2
* of the License, or (at your option) any later version.
*
* This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY
* KIND, either express or implied.
*
****************************************************************************/
#include "config.h"
/* uncomment this to make filtering calculate lower bits after shifting.
* without this, "shift" of the lower bits will be lost here.
*/
/* #define HIGH_PRECISION */
/*
* void eq_filter(int32_t **x, struct eqfilter *f, unsigned num,
* unsigned channels, unsigned shift)
*/
#if CONFIG_CPU == PP5002
.section .icode,"ax",%progbits
#else
.text
#endif
.global eq_filter
eq_filter:
ldr r12, [sp] @ get shift parameter
stmdb sp!, { r0-r11, lr } @ save all params and clobbered regs
ldmia r1!, { r4-r8 } @ load coefs
mov r10, r1 @ loop prelude expects filter struct addr in r10
.filterloop:
ldr r9, [sp] @ get pointer to this channels data
add r0, r9, #4
str r0, [sp] @ save back pointer to next channels data
ldr r9, [r9] @ r9 = x[]
ldr r14, [sp, #8] @ r14 = numsamples
ldmia r10, { r0-r3 } @ load history, r10 should be filter struct addr
str r10, [sp, #4] @ save it for loop end
/* r0-r3 = history, r4-r8 = coefs, r9 = x[], r10..r11 = accumulator,
* r12 = shift amount, r14 = number of samples.
*/
.loop:
/* Direct form 1 filtering code.
* y[n] = b0*x[i] + b1*x[i - 1] + b2*x[i - 2] + a1*y[i - 1] + a2*y[i - 2],
* where y[] is output and x[] is input. This is performed out of order to
* reuse registers, we're pretty short on regs.
*/
smull r10, r11, r6, r1 @ acc = b2*x[i - 2]
mov r1, r0 @ fix input history
smlal r10, r11, r5, r0 @ acc += b1*x[i - 1]
ldr r0, [r9] @ load input and fix history in same operation
smlal r10, r11, r4, r0 @ acc += b0*x[i]
smlal r10, r11, r7, r2 @ acc += a1*y[i - 1]
smlal r10, r11, r8, r3 @ acc += a2*y[i - 2]
mov r3, r2 @ fix output history
mov r2, r11, asl r12 @ get upper part of result and shift left
#ifdef HIGH_PRECISION
rsb r11, r12, #32 @ get shift amount for lower part
orr r2, r2, r10, lsr r11 @ then mix in correctly shifted lower part
#endif
str r2, [r9], #4 @ save result
subs r14, r14, #1 @ are we done with this channel?
bne .loop
ldr r10, [sp, #4] @ load filter struct pointer
stmia r10!, { r0-r3 } @ save back history
ldr r11, [sp, #12] @ load number of channels
subs r11, r11, #1 @ all channels processed?
strne r11, [sp, #12]
bne .filterloop
add sp, sp, #16 @ compensate for temp storage
ldmpc regs=r4-r11