rockbox/lib/rbcodec/dsp/dsp_arm.S

/***************************************************************************
 *             __________               __   ___.
 *   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"

/****************************************************************************
 *  void channels_process_sound_chan_mono(int count, int32_t *buf[])
 */

#include "config.h"

    .section .icode, "ax", %progbits
    .align  2
    .global channels_process_sound_chan_mono
    .type   channels_process_sound_chan_mono, %function
channels_process_sound_chan_mono:
    @ input: r0 = count, r1 = buf
    stmfd   sp!, { r4, lr }            @
                                       @
    ldmia   r1, { r1, r2 }             @ r1 = buf[0], r2 = buf[1]
    subs    r0, r0, #1                 @ odd: end at 0; even: end at -1
    beq     .mono_singlesample         @ Zero? Only one sample!
                                       @
.monoloop:                             @
    ldmia   r1, { r3, r4 }             @ r3, r4 = Li0, Li1
    ldmia   r2, { r12, r14 }           @ r12, r14 = Ri0, Ri1
    mov     r3, r3, asr #1             @ Mo0 = Li0 / 2 + Ri0 / 2
    mov     r4, r4, asr #1             @ Mo1 = Li1 / 2 + Ri1 / 2
    add     r12, r3, r12, asr #1       @
    add     r14, r4, r14, asr #1       @
    subs    r0, r0, #2                 @
    stmia   r1!, { r12, r14 }          @ store Mo0, Mo1
    stmia   r2!, { r12, r14 }          @ store Mo0, Mo1
    bgt     .monoloop                  @
                                       @
    ldmpc   cond=lt, regs=r4           @ if count was even, we're done
                                       @
.mono_singlesample:                    @
    ldr     r3, [r1]                   @ r3 = Ls
    ldr     r12, [r2]                  @ r12 = Rs
    mov     r3, r3, asr #1             @ Mo = Ls / 2 + Rs / 2
    add     r12, r3, r12, asr #1       @
    str     r12, [r1]                  @ store Mo
    str     r12, [r2]                  @ store Mo
                                       @
    ldmpc   regs=r4                    @
    .size   channels_process_sound_chan_mono, \
                .-channels_process_sound_chan_mono

/****************************************************************************
 * void channels_process_sound_chan_custom(int count, int32_t *buf[])
 */
    .section .icode, "ax", %progbits
    .align  2
    .global channels_process_sound_chan_custom
    .type   channels_process_sound_chan_custom, %function
channels_process_sound_chan_custom:
    stmfd   sp!, { r4-r10, lr }
    
    ldr     r3, =dsp_sw_gain
    ldr     r4, =dsp_sw_cross

    ldmia   r1, { r1, r2 }             @ r1 = buf[0], r2 = buf[1]
    ldr     r3, [r3]                   @ r3 = dsp_sw_gain
    ldr     r4, [r4]                   @ r4 = dsp_sw_cross
    
    subs    r0, r0, #1
    beq     .custom_single_sample      @ Zero? Only one sample!
    
.custom_loop:
    ldmia   r1, { r5, r6 }             @ r5 = Li0, r6 = Li1
    ldmia   r2, { r7, r8 }             @ r7 = Ri0, r8 = Ri1

    subs    r0, r0, #2

    smull   r9, r10, r5, r3            @ Lc0 = Li0*gain
    smull   r12, r14, r7, r3           @ Rc0 = Ri0*gain
    smlal   r9, r10, r7, r4            @ Lc0 += Ri0*cross
    smlal   r12, r14, r5, r4           @ Rc0 += Li0*cross
    
    mov     r9, r9, lsr #31            @ Convert to s0.31
    mov     r12, r12, lsr #31
    orr     r5, r9, r10, asl #1
    orr     r7, r12, r14, asl #1
    
    smull   r9, r10, r6, r3            @ Lc1 = Li1*gain
    smull   r12, r14, r8, r3           @ Rc1 = Ri1*gain
    smlal   r9, r10, r8, r4            @ Lc1 += Ri1*cross
    smlal   r12, r14, r6, r4           @ Rc1 += Li1*cross
    
    mov     r9, r9, lsr #31            @ Convert to s0.31
    mov     r12, r12, lsr #31
    orr     r6, r9, r10, asl #1
    orr     r8, r12, r14, asl #1
    
    stmia   r1!, { r5, r6 }            @ Store Lc0, Lc1
    stmia   r2!, { r7, r8 }            @ Store Rc0, Rc1

    bgt     .custom_loop
    
    ldmpc   cond=lt, regs=r4-r10       @ < 0? even count
    
.custom_single_sample:
    ldr     r5, [r1]                   @ handle odd sample
    ldr     r7, [r2]

    smull   r9, r10, r5, r3            @ Lc0 = Li0*gain
    smull   r12, r14, r7, r3           @ Rc0 = Ri0*gain
    smlal   r9, r10, r7, r4            @ Lc0 += Ri0*cross
    smlal   r12, r14, r5, r4           @ Rc0 += Li0*cross

    mov     r9, r9, lsr #31            @ Convert to s0.31
    mov     r12, r12, lsr #31
    orr     r5, r9, r10, asl #1
    orr     r7, r12, r14, asl #1

    str     r5, [r1]                   @ Store Lc0
    str     r7, [r2]                   @ Store Rc0

    ldmpc   regs=r4-r10
    .size   channels_process_sound_chan_custom, \
                .-channels_process_sound_chan_custom

/****************************************************************************
 *  void channels_process_sound_chan_karaoke(int count, int32_t *buf[])
 */
    .section .icode, "ax", %progbits
    .align  2
    .global channels_process_sound_chan_karaoke
    .type   channels_process_sound_chan_karaoke, %function
channels_process_sound_chan_karaoke:
    @ input: r0 = count, r1 = buf
    stmfd   sp!, { r4, lr }            @
                                       @
    ldmia   r1, { r1, r2 }             @ r1 = buf[0], r2 = buf[1]
    subs    r0, r0, #1                 @ odd: end at 0; even: end at -1
    beq     .karaoke_singlesample      @ Zero? Only one sample!
                                       @
.karaokeloop:                          @
    ldmia   r1, { r3, r4 }             @ r3, r4  = Li0, Li1
    ldmia   r2, { r12, r14 }           @ r12, r14 = Ri0, Ri1
    mov     r3, r3, asr #1             @ Lo0 = Li0 / 2 - Ri0 / 2
    mov     r4, r4, asr #1             @ Lo1 = Li1 / 2 - Ri1 / 2
    sub     r3, r3, r12, asr #1        @
    sub     r4, r4, r14, asr #1        @
    rsb     r12, r3, #0                @ Ro0 = -Lk0 = Rs0 / 2 - Ls0 / 2
    rsb     r14, r4, #0                @ Ro1 = -Lk1 = Ri1 / 2 - Li1 / 2
    subs    r0, r0, #2                 @
    stmia   r1!, { r3, r4 }            @ store Lo0, Lo1
    stmia   r2!, { r12, r14 }          @ store Ro0, Ro1
    bgt     .karaokeloop               @
                                       @
    ldmpc   cond=lt, regs=r4           @ if count was even, we're done
                                       @
.karaoke_singlesample:                 @
    ldr     r3, [r1]                   @ r3 = Li
    ldr     r12, [r2]                  @ r12 = Ri
    mov     r3, r3, asr #1             @ Lk = Li / 2 - Ri /2
    sub     r3, r3, r12, asr #1        @
    rsb     r12, r3, #0                @ Rk = -Lo = Ri / 2 - Li / 2
    str     r3, [r1]                   @ store Lo
    str     r12, [r2]                  @ store Ro
                                       @
    ldmpc   regs=r4                    @
    .size   channels_process_sound_chan_karaoke, \
                .-channels_process_sound_chan_karaoke

#if ARM_ARCH < 6
/****************************************************************************
 *  void sample_output_mono(int count, struct dsp_data *data,
 *                          const int32_t *src[], int16_t *dst)
 */
    .section .icode, "ax", %progbits
    .align  2
    .global sample_output_mono
    .type   sample_output_mono, %function
sample_output_mono:
    @ input: r0 = count, r1 = data, r2 = src, r3 = dst
    stmfd   sp!, { r4-r6, lr }

    ldr     r1, [r1]                   @ lr = data->output_scale
    ldr     r2, [r2]                   @ r2 = src[0]

    mov     r4, #1
    mov     r4, r4, lsl r1             @ r4 = 1 << (scale-1)
    mov     r4, r4, lsr #1
    mvn     r14, #0x8000               @ r14 = 0xffff7fff, needed for
                                       @ clipping and masking
    subs    r0, r0, #1                 @
    beq     .som_singlesample          @ Zero? Only one sample!

.somloop:
    ldmia   r2!, { r5, r6 }
    add     r5, r5, r4                 @ r6 = (r6 + 1<<(scale-1)) >> scale
    mov     r5, r5, asr r1
    mov     r12, r5, asr #15
    teq     r12, r12, asr #31
    eorne   r5, r14, r5, asr #31       @ Clip (-32768...+32767)
    add     r6, r6, r4
    mov     r6, r6, asr r1             @ r7 = (r7 + 1<<(scale-1)) >> scale
    mov     r12, r6, asr #15
    teq     r12, r12, asr #31
    eorne   r6, r14, r6, asr #31       @ Clip (-32768...+32767)
    
    and     r5, r5, r14, lsr #16
    and     r6, r6, r14, lsr #16
    orr     r5, r5, r5, lsl #16        @ pack first 2 halfwords into 1 word
    orr     r6, r6, r6, lsl #16        @ pack last 2 halfwords into 1 word
    stmia   r3!, { r5, r6 }
    
    subs    r0, r0, #2
    bgt     .somloop     
       
    ldmpc   cond=lt, regs=r4-r6        @ even 'count'? return

.som_singlesample:
    ldr     r5, [r2]                   @ do odd sample
    add     r5, r5, r4
    mov     r5, r5, asr r1
    mov     r12, r5, asr #15
    teq     r12, r12, asr #31
    eorne   r5, r14, r5, asr #31

    and     r5, r5, r14, lsr #16       @ pack 2 halfwords into 1 word
    orr     r5, r5, r5, lsl #16
    str     r5, [r3]

    ldmpc   regs=r4-r6
    .size   sample_output_mono, .-sample_output_mono
    
/****************************************************************************
 * void sample_output_stereo(int count, struct dsp_data *data,
 *                           const int32_t *src[], int16_t *dst)
 */
    .section .icode, "ax", %progbits
    .align  2
    .global sample_output_stereo
    .type   sample_output_stereo, %function
sample_output_stereo:
    @ input: r0 = count, r1 = data, r2 = src, r3 = dst
    stmfd   sp!, { r4-r9, lr }

    ldr     r1, [r1]                   @ r1 = data->output_scale
    ldmia   r2, { r2, r5 }             @ r2 = src[0], r5 = src[1]

    mov     r4, #1
    mov     r4, r4, lsl r1             @ r4 = 1 << (scale-1)
    mov     r4, r4, lsr #1             @
    
    mvn     r14, #0x8000               @ r14 = 0xffff7fff, needed for
                                       @ clipping and masking
    subs    r0, r0, #1                 @
    beq     .sos_singlesample          @ Zero? Only one sample!

.sosloop:
    ldmia   r2!, { r6, r7 }            @ 2 left
    ldmia   r5!, { r8, r9 }            @ 2 right

    add     r6, r6, r4                 @ r6 = (r6 + 1<<(scale-1)) >> scale
    mov     r6, r6, asr r1
    mov     r12, r6, asr #15
    teq     r12, r12, asr #31
    eorne   r6, r14, r6, asr #31       @ Clip (-32768...+32767)
    add     r7, r7, r4
    mov     r7, r7, asr r1             @ r7 = (r7 + 1<<(scale-1)) >> scale
    mov     r12, r7, asr #15
    teq     r12, r12, asr #31
    eorne   r7, r14, r7, asr #31       @ Clip (-32768...+32767)
    
    add     r8, r8, r4                 @ r8 = (r8 + 1<<(scale-1)) >> scale
    mov     r8, r8, asr r1
    mov     r12, r8, asr #15
    teq     r12, r12, asr #31
    eorne   r8, r14, r8, asr #31       @ Clip (-32768...+32767)
    add     r9, r9, r4                 @ r9 = (r9 + 1<<(scale-1)) >> scale
    mov     r9, r9, asr r1
    mov     r12, r9, asr #15
    teq     r12, r12, asr #31
    eorne   r9, r14, r9, asr #31       @ Clip (-32768...+32767)
    
    and     r6, r6, r14, lsr #16       @ pack first 2 halfwords into 1 word
    orr     r8, r6, r8, asl #16
    and     r7, r7, r14, lsr #16       @ pack last 2 halfwords into 1 word
    orr     r9, r7, r9, asl #16

    stmia   r3!, { r8, r9 }

    subs    r0, r0, #2
    bgt     .sosloop

    ldmpc   cond=lt, regs=r4-r9        @ even 'count'? return

.sos_singlesample:    
    ldr     r6, [r2]                   @ left odd sample
    ldr     r8, [r5]                   @ right odd sample

    add     r6, r6, r4                 @ r6 = (r7 + 1<<(scale-1)) >> scale
    mov     r6, r6, asr r1
    mov     r12, r6, asr #15
    teq     r12, r12, asr #31
    eorne   r6, r14, r6, asr #31       @ Clip (-32768...+32767)
    add     r8, r8, r4                 @ r8 = (r8 + 1<<(scale-1)) >> scale
    mov     r8, r8, asr r1
    mov     r12, r8, asr #15
    teq     r12, r12, asr #31
    eorne   r8, r14, r8, asr #31       @ Clip (-32768...+32767)
    
    and     r6, r6, r14, lsr #16       @ pack 2 halfwords into 1 word
    orr     r8, r6, r8, asl #16

    str     r8, [r3]

    ldmpc   regs=r4-r9
    .size   sample_output_stereo, .-sample_output_stereo
#endif /* ARM_ARCH < 6 */    

/****************************************************************************
 * void apply_crossfeed(int count, int32_t* src[])
 */
    .section .text
    .global apply_crossfeed 
apply_crossfeed:
    @ unfortunately, we ended up in a bit of a register squeeze here, and need
    @ to keep the count on the stack :/
    stmdb   sp!, { r4-r11, lr }        @ stack modified regs
    ldmia   r1, { r2-r3 }              @ r2 = src[0], r3 = src[1]
    
    ldr     r1, =crossfeed_data
    ldmia   r1!, { r4-r11 }            @ load direct gain and filter data
    mov     r12, r0                    @ better to ldm delay + count later
    add     r0, r1, #13*4*2            @ calculate end of delay
    stmdb   sp!, { r0, r12 }           @ stack end of delay adr and count
    ldr     r0, [r1, #13*4*2]          @ fetch current delay line address

    /* Register usage in loop:
     * r0 = &delay[index][0], r1 = accumulator high, r2 = src[0], r3 = src[1],
     * r4 = direct gain, r5-r7 = b0, b1, a1 (filter coefs),
     * r8-r11 = filter history, r12 = temp, r14 = accumulator low
     */
.cfloop:
    smull   r14, r1, r6, r8            @ acc = b1*dr[n - 1]
    smlal   r14, r1, r7, r9            @ acc += a1*y_l[n - 1]
    ldr     r8, [r0, #4]               @ r8 = dr[n]
    smlal   r14, r1, r5, r8            @ acc += b0*dr[n]
    mov     r9, r1, lsl #1             @ fix format for filter history
    ldr     r12, [r2]                  @ load left input
    smlal   r14, r1, r4, r12           @ acc += gain*x_l[n] 
    mov     r1, r1, lsl #1             @ fix format
    str     r1, [r2], #4               @ save result

    smull   r14, r1, r6, r10           @ acc = b1*dl[n - 1]
    smlal   r14, r1, r7, r11           @ acc += a1*y_r[n - 1]
    ldr     r10, [r0]                  @ r10 = dl[n]
    str     r12, [r0], #4              @ save left input to delay line
    smlal   r14, r1, r5, r10           @ acc += b0*dl[n]
    mov     r11, r1, lsl #1            @ fix format for filter history
    ldr     r12, [r3]                  @ load right input
    smlal   r14, r1, r4, r12           @ acc += gain*x_r[n]
    str     r12, [r0], #4              @ save right input to delay line
    mov     r1, r1, lsl #1             @ fix format
    ldmia   sp, { r12, r14 }           @ fetch delay line end addr and count from stack
    str     r1, [r3], #4               @ save result

    cmp     r0, r12                    @ need to wrap to start of delay?
    subeq   r0, r0, #13*4*2            @ wrap back delay line ptr to start
 
    subs    r14, r14, #1               @ are we finished?
    strne   r14, [sp, #4]              @ nope, save count back to stack
    bne     .cfloop
    
    @ save data back to struct
    ldr     r12, =crossfeed_data + 4*4
    stmia   r12, { r8-r11 }            @ save filter history
    str     r0, [r12, #30*4]           @ save delay line index
    add     sp, sp, #8                 @ remove temp variables from stack
    ldmpc   regs=r4-r11
    .size   apply_crossfeed, .-apply_crossfeed

/****************************************************************************
 * int dsp_downsample(int count, struct dsp_data *data,
 *                    in32_t *src[], int32_t *dst[])
 */
    .section    .text
    .global     dsp_downsample
dsp_downsample:
    stmdb   sp!, { r4-r11, lr }     @ stack modified regs
    ldmib   r1, { r5-r6 }           @ r5 = num_channels,r6 = resample_data.delta
    sub     r5, r5, #1              @ pre-decrement num_channels for use
    add     r4, r1, #12             @ r4 = &resample_data.phase
    mov     r12, #0xff
    orr     r12, r12, #0xff00       @ r12 = 0xffff
.dschannel_loop:
    ldr     r1, [r4]                @ r1 = resample_data.phase
    ldr     r7, [r2, r5, lsl #2]    @ r7 = s = src[ch - 1]
    ldr     r8, [r3, r5, lsl #2]    @ r8 = d = dst[ch - 1]
    add     r9, r4, #4              @ r9 = &last_sample[0]
    ldr     r10, [r9, r5, lsl #2]   @ r10 = last_sample[ch - 1]
    sub     r11, r0, #1             
    ldr     r14, [r7, r11, lsl #2]  @ load last sample in s[] ...
    str     r14, [r9, r5, lsl #2]   @ and write as next frame's last_sample
    movs    r9, r1, lsr #16         @ r9 = pos = phase >> 16
    ldreq   r11, [r7]               @ if pos = 0, load src[0] and jump into loop
    beq     .dsuse_last_start
    cmp     r9, r0                  @ if pos >= count, we're already done
    bge     .dsloop_skip

    @ Register usage in loop:
    @ r0 = count, r1 = phase, r4 = &resample_data.phase, r5 = cur_channel,
    @ r6 = delta, r7 = s, r8 = d, r9 = pos, r10 = s[pos - 1], r11 = s[pos]
.dsloop:
    add     r9, r7, r9, lsl #2      @ r9 = &s[pos]
    ldmda   r9, { r10, r11 }        @ r10 = s[pos - 1], r11 = s[pos]
.dsuse_last_start:
    sub     r11, r11, r10           @ r11 = diff = s[pos] - s[pos - 1]
    @ keep frac in lower bits to take advantage of multiplier early termination
    and     r9, r1, r12             @ frac = phase & 0xffff
    smull   r9, r14, r11, r9
    add     r1, r1, r6              @ phase += delta
    add     r10, r10, r9, lsr #16   @ r10 = out = s[pos - 1] + frac*diff
    add     r10, r10, r14, lsl #16
    str     r10, [r8], #4           @ *d++ = out
    mov     r9, r1, lsr #16         @ pos = phase >> 16
    cmp     r9, r0                  @ pos < count?
    blt     .dsloop                 @ yup, do more samples
.dsloop_skip:
    subs    r5, r5, #1
    bpl     .dschannel_loop         @ if (--ch) >= 0, do another channel
    sub     r1, r1, r0, lsl #16     @ wrap phase back to start
    str     r1, [r4]                @ store back
    ldr     r1, [r3]                @ r1 = &dst[0]
    sub     r8, r8, r1              @ dst - &dst[0]
    mov     r0, r8, lsr #2          @ convert bytes->samples
    ldmpc   regs=r4-r11             @ ... and we're out
    .size   dsp_downsample, .-dsp_downsample

/****************************************************************************
 * int dsp_upsample(int count, struct dsp_data *dsp,
 *                  in32_t *src[], int32_t *dst[])
 */
    .section    .text
    .global     dsp_upsample
dsp_upsample:
    stmfd   sp!, { r4-r11, lr }     @ stack modified regs
    ldmib   r1, { r5-r6 }           @ r5 = num_channels,r6 = resample_data.delta
    sub     r5, r5, #1              @ pre-decrement num_channels for use
    add     r4, r1, #12             @ r4 = &resample_data.phase
    mov     r6, r6, lsl #16         @ we'll use carry to detect pos increments
    stmfd   sp!, { r0, r4 }         @ stack count and &resample_data.phase
.uschannel_loop:
    ldr     r12, [r4]               @ r12 = resample_data.phase
    ldr     r7, [r2, r5, lsl #2]    @ r7 = s = src[ch - 1]
    ldr     r8, [r3, r5, lsl #2]    @ r8 = d = dst[ch - 1]
    add     r9, r4, #4              @ r9 = &last_sample[0]
    mov     r1, r12, lsl #16        @ we'll use carry to detect pos increments
    sub     r11, r0, #1             
    ldr     r14, [r7, r11, lsl #2]  @ load last sample in s[] ...
    ldr     r10, [r9, r5, lsl #2]   @ r10 = last_sample[ch - 1]
    str     r14, [r9, r5, lsl #2]   @ and write as next frame's last_sample
    movs    r14, r12, lsr #16       @ pos = resample_data.phase >> 16
    beq     .usstart_0              @ pos = 0
    cmp     r14, r0                 @ if pos >= count, we're already done
    bge     .usloop_skip
    add     r7, r7, r14, lsl #2     @ r7 = &s[pos]
    ldr     r10, [r7, #-4]          @ r11 = s[pos - 1]
    b       .usstart_0

    @ Register usage in loop:
    @ r0 = count, r1 = phase, r4 = &resample_data.phase, r5 = cur_channel,
    @ r6 = delta, r7 = s, r8 = d, r9 = diff, r10 = s[pos - 1], r11 = s[pos]
.usloop_1:
    mov     r10, r11                @ r10 = previous sample
.usstart_0:
    ldr     r11, [r7], #4           @ r11 = next sample
    mov     r4, r1, lsr #16         @ r4 = frac = phase >> 16
    sub     r9, r11, r10            @ r9 = diff = s[pos] - s[pos - 1]
.usloop_0:
    smull   r12, r14, r4, r9
    adds    r1, r1, r6              @ phase += delta << 16
    mov     r4, r1, lsr #16         @ r4 = frac = phase >> 16
    add     r14, r10, r14, lsl #16
    add     r14, r14, r12, lsr #16  @ r14 = out = s[pos - 1] + frac*diff
    str     r14, [r8], #4           @ *d++ = out
    bcc     .usloop_0               @ if carry is set, pos is incremented
    subs    r0, r0, #1              @ if count > 0, do another sample
    bgt     .usloop_1
.usloop_skip:
    subs    r5, r5, #1
    ldmfd   sp, { r0, r4 }          @ reload count and &resample_data.phase
    bpl     .uschannel_loop         @ if (--ch) >= 0, do another channel
    mov     r1, r1, lsr #16         @ wrap phase back to start of next frame
    ldr     r2, [r3]                @ r1 = &dst[0]
    str     r1, [r4]                @ store phase
    sub     r8, r8, r2              @ dst - &dst[0]
    mov     r0, r8, lsr #2          @ convert bytes->samples
    add     sp, sp, #8              @ adjust stack for temp variables
    ldmpc   regs=r4-r11             @ ... and we're out
    .size       dsp_upsample, .-dsp_upsample

/****************************************************************************
 *  void dsp_apply_gain(int count, struct dsp_data *data, int32_t *buf[])
 */
    .section .icode, "ax", %progbits
    .align  2
    .global dsp_apply_gain
    .type   dsp_apply_gain, %function
dsp_apply_gain:
    @ input: r0 = count, r1 = data, r2 = buf[]
    stmfd   sp!, { r4-r8, lr }

    ldr     r3, [r1,  #4]           @ r3 = data->num_channels
    ldr     r4, [r1, #32]           @ r5 = data->gain

.dag_outerloop:
    ldr     r1, [r2], #4            @ r1 = buf[0] and increment index of buf[]
    subs    r12, r0, #1             @ r12 = r0 = count - 1
    beq     .dag_singlesample       @ Zero? Only one sample!

.dag_innerloop:
    ldmia   r1, { r5, r6 }          @ load r5, r6 from r1
    smull   r7, r8, r5, r4          @ r7 = FRACMUL_SHL(r5, r4, 8)
    smull   r14, r5, r6, r4         @ r14 = FRACMUL_SHL(r6, r4, 8)
    subs    r12, r12, #2
    mov     r7, r7, lsr #23
    mov     r14, r14, lsr #23
    orr     r7, r7, r8, asl #9
    orr     r14, r14, r5, asl #9
    stmia   r1!, { r7, r14 }        @ save r7, r14 to [r1] and increment r1
    bgt     .dag_innerloop          @ end of inner loop

    blt     .dag_evencount          @ < 0? even count

.dag_singlesample:
    ldr     r5, [r1]                @ handle odd sample
    smull   r7, r8, r5, r4          @ r7 = FRACMUL_SHL(r5, r4, 8)
    mov     r7, r7, lsr #23
    orr     r7, r7, r8, asl #9
    str     r7, [r1]

.dag_evencount:
    subs    r3, r3, #1
    bgt     .dag_outerloop          @ end of outer loop
               
    ldmpc   regs=r4-r8
    .size   dsp_apply_gain, .-dsp_apply_gain