rockbox/apps/codecs/libmad/synth_full_arm.S

343 lines
8.5 KiB
ArmAsm

/***************************************************************************
* __________ __ ___.
* Open \______ \ ____ ____ | | _\_ |__ _______ ___
* Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ /
* Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < <
* Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \
* \/ \/ \/ \/ \/
* $Id$
*
* Copyright (C) 2007 by Tomasz Malesinski
*
* All files in this archive are subject to the GNU General Public License.
* See the file COPYING in the source tree root for full license agreement.
*
* This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY
* KIND, either express or implied.
*
****************************************************************************/
#include "config.h"
/* This performs slower in IRAM on PP502x and there is no space in
mpegplayer on the PP5002 */
#if defined(CPU_PP502x) || (CONFIG_CPU == PP5002 && defined(MPEGPLAYER))
.section .text,"ax",%progbits
#else
.section .icode,"ax",%progbits
#endif
.global synth_full1
.global synth_full2
;; r0 = pcm
;; r1 = fo
;; r2 = fe
;; r3 = D0ptr
;; r4 = D1ptr
synth_full1:
stmdb sp!, {r4, r5, r6, r7, r8, r9, r10, r11, r12, lr}
ldr r4, [sp, #40]
ldr r5, =synth_full_sp
str sp, [r5]
mov r5, #15
add r2, r2, #32
.l:
add r3, r3, #128
add r4, r4, #128
ldmia r1!, {r10, r11, r12, lr}
ldr r7, [r3, #4]
smull r6, r7, r10, r7
ldr r9, [r4, #120]
smull r8, r9, r10, r9
ldr r10, [r3, #60]
smlal r6, r7, r11, r10
ldr r10, [r3, #52]
smlal r6, r7, r12, r10
ldr r10, [r3, #44]
smlal r6, r7, lr, r10
ldr r10, [r4, #64]
smlal r8, r9, r11, r10
ldr r10, [r4, #72]
smlal r8, r9, r12, r10
ldr r10, [r4, #80]
smlal r8, r9, lr, r10
ldmia r1!, {r11, r12, sp, lr}
ldr r10, [r3, #36]
smlal r6, r7, r11, r10
ldr r10, [r3, #28]
smlal r6, r7, r12, r10
ldr r10, [r3, #20]
smlal r6, r7, sp, r10
ldr r10, [r3, #12]
smlal r6, r7, lr, r10
ldr r10, [r4, #88]
smlal r8, r9, r11, r10
ldr r10, [r4, #96]
smlal r8, r9, r12, r10
ldr r10, [r4, #104]
smlal r8, r9, sp, r10
ldr r10, [r4, #112]
smlal r8, r9, lr, r10
rsbs r6, r6, #0
rsc r7, r7, #0
ldmia r2!, {r11, r12, sp, lr}
ldr r10, [r3, #0]
smlal r6, r7, r11, r10
ldr r10, [r3, #56]
smlal r6, r7, r12, r10
ldr r10, [r3, #48]
smlal r6, r7, sp, r10
ldr r10, [r3, #40]
smlal r6, r7, lr, r10
ldr r10, [r4, #60]
smlal r8, r9, r11, r10
ldr r10, [r4, #68]
smlal r8, r9, r12, r10
ldr r10, [r4, #76]
smlal r8, r9, sp, r10
ldr r10, [r4, #84]
smlal r8, r9, lr, r10
ldmia r2!, {r11, r12, sp, lr}
ldr r10, [r3, #32]
smlal r6, r7, r11, r10
ldr r10, [r3, #24]
smlal r6, r7, r12, r10
ldr r10, [r3, #16]
smlal r6, r7, sp, r10
ldr r10, [r3, #8]
smlal r6, r7, lr, r10
ldr r10, [r4, #92]
smlal r8, r9, r11, r10
ldr r10, [r4, #100]
smlal r8, r9, r12, r10
ldr r10, [r4, #108]
smlal r8, r9, sp, r10
ldr r10, [r4, #116]
smlal r8, r9, lr, r10
movs r6, r6, lsr #16
adc r6, r6, r7, lsl #16
str r6, [r0, -r5, lsl #2]
movs r8, r8, lsr #16
adc r8, r8, r9, lsl #16
str r8, [r0, r5, lsl #2]
subs r5, r5, #1
bne .l
ldr r5, =synth_full_sp
ldr sp, [r5]
ldmia sp!, {r4, r5, r6, r7, r8, r9, r10, r11, r12, pc}
synth_full2:
stmdb sp!, {r4, r5, r6, r7, r8, r9, r10, r11, r12, lr}
ldr r4, [sp, #40]
ldr r5, =synth_full_sp
str sp, [r5]
mov r5, #15
add r2, r2, #32
.l2:
add r3, r3, #128
add r4, r4, #128
ldmia r1!, {r10, r11, r12, lr}
ldr r7, [r3, #0]
smull r6, r7, r10, r7
ldr r9, [r4, #60]
smull r8, r9, r10, r9
ldr r10, [r3, #56]
smlal r6, r7, r11, r10
ldr r10, [r3, #48]
smlal r6, r7, r12, r10
ldr r10, [r3, #40]
smlal r6, r7, lr, r10
ldr r10, [r4, #68]
smlal r8, r9, r11, r10
ldr r10, [r4, #76]
smlal r8, r9, r12, r10
ldr r10, [r4, #84]
smlal r8, r9, lr, r10
ldmia r1!, {r11, r12, sp, lr}
ldr r10, [r3, #32]
smlal r6, r7, r11, r10
ldr r10, [r3, #24]
smlal r6, r7, r12, r10
ldr r10, [r3, #16]
smlal r6, r7, sp, r10
ldr r10, [r3, #8]
smlal r6, r7, lr, r10
ldr r10, [r4, #92]
smlal r8, r9, r11, r10
ldr r10, [r4, #100]
smlal r8, r9, r12, r10
ldr r10, [r4, #108]
smlal r8, r9, sp, r10
ldr r10, [r4, #116]
smlal r8, r9, lr, r10
rsbs r6, r6, #0
rsc r7, r7, #0
ldmia r2!, {r11, r12, sp, lr}
ldr r10, [r3, #4]
smlal r6, r7, r11, r10
ldr r10, [r3, #60]
smlal r6, r7, r12, r10
ldr r10, [r3, #52]
smlal r6, r7, sp, r10
ldr r10, [r3, #44]
smlal r6, r7, lr, r10
ldr r10, [r4, #120]
smlal r8, r9, r11, r10
ldr r10, [r4, #64]
smlal r8, r9, r12, r10
ldr r10, [r4, #72]
smlal r8, r9, sp, r10
ldr r10, [r4, #80]
smlal r8, r9, lr, r10
ldmia r2!, {r11, r12, sp, lr}
ldr r10, [r3, #36]
smlal r6, r7, r11, r10
ldr r10, [r3, #28]
smlal r6, r7, r12, r10
ldr r10, [r3, #20]
smlal r6, r7, sp, r10
ldr r10, [r3, #12]
smlal r6, r7, lr, r10
ldr r10, [r4, #88]
smlal r8, r9, r11, r10
ldr r10, [r4, #96]
smlal r8, r9, r12, r10
ldr r10, [r4, #104]
smlal r8, r9, sp, r10
ldr r10, [r4, #112]
smlal r8, r9, lr, r10
movs r6, r6, lsr #16
adc r6, r6, r7, lsl #16
str r6, [r0, -r5, lsl #2]
movs r8, r8, lsr #16
adc r8, r8, r9, lsl #16
str r8, [r0, r5, lsl #2]
subs r5, r5, #1
bne .l2
ldr r5, =synth_full_sp
ldr sp, [r5]
ldmia sp!, {r4, r5, r6, r7, r8, r9, r10, r11, r12, pc}
.global III_aliasreduce
III_aliasreduce:
stmdb sp!, {r4, r5, r6, r7, r8, r9, r10, r11, r12, lr}
add r1, r0, r1, lsl #2
add r0, r0, #72
.arl1:
mov r2, #8
mov r3, r0 @ a
mov r4, r0 @ b
ldr r5, =csa @ cs/ca
.arl2:
ldmdb r3, {r6, r12}
ldmia r4, {r7, lr}
ldmia r5!, {r8, r9}
smull r10, r11, r7, r8
smlal r10, r11, r12, r9
movs r10, r10, lsr #28
adc r10, r10, r11, lsl #4
rsb r7, r7, #0
smull r11, r8, r12, r8
smlal r11, r8, r7, r9
movs r11, r11, lsr #28
adc r11, r11, r8, lsl #4
ldmia r5!, {r8, r9}
smull r12, r7, lr, r8
smlal r12, r7, r6, r9
movs r12, r12, lsr #28
adc r12, r12, r7, lsl #4
stmia r4!, {r10, r12}
rsb lr, lr, #0
smull r7, r10, r6, r8
smlal r7, r10, lr, r9
movs r7, r7, lsr #28
adc r7, r7, r10, lsl #4
stmdb r3!, {r7, r11}
subs r2, r2, #2
bne .arl2
add r0, r0, #72
cmp r0, r1
blo .arl1
ldmia sp!, {r4, r5, r6, r7, r8, r9, r10, r11, r12, pc}
csa:
.word +0x0db84a81
.word -0x083b5fe7
.word +0x0e1b9d7f
.word -0x078c36d2
.word +0x0f31adcf
.word -0x05039814
.word +0x0fbba815
.word -0x02e91dd1
.word +0x0feda417
.word -0x0183603a
.word +0x0ffc8fc8
.word -0x00a7cb87
.word +0x0fff964c
.word -0x003a2847
.word +0x0ffff8d3
.word -0x000f27b4
.global III_overlap
III_overlap:
stmdb sp!, {r4, r5, r6, r7, r8, lr}
add r2, r2, r3, lsl #2
mov r3, #6
.ol:
ldmia r0!, {r4, r5, r6}
ldmia r1!, {r7, r8, lr}
add r4, r4, r7
add r5, r5, r8
add r6, r6, lr
str r4, [r2], #128
str r5, [r2], #128
str r6, [r2], #128
subs r3, r3, #1
bne .ol
sub r1, r1, #72
ldmia r0!, {r4, r5, r6, r7, r8, lr}
stmia r1!, {r4, r5, r6, r7, r8, lr}
ldmia r0!, {r4, r5, r6, r7, r8, lr}
stmia r1!, {r4, r5, r6, r7, r8, lr}
ldmia r0!, {r4, r5, r6, r7, r8, lr}
stmia r1!, {r4, r5, r6, r7, r8, lr}
ldmia sp!, {r4, r5, r6, r7, r8, pc}
.section .ibss,"aw",%nobits
synth_full_sp:
.space 4