Nyaaori/rockbox
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity
rockbox/lib/rbcodec/dsp/dsp_cf.S
Michael Sevakis c9bcbe202d Fundamentally rewrite much of the audio DSP. 
Creates a standard buffer passing, local data passing and messaging
system for processing stages. Stages can be moved to their own source
files to reduce clutter and ease assimilation of new ones. dsp.c
becomes dsp_core.c which supports an engine and framework for effects.

Formats and change notifications are passed along with the buffer so
that they arrive at the correct time at each stage in the chain
regardless of the internal delays of a particular one.

Removes restrictions on the number of samples that can be processed at
a time and it pays attention to destination buffer size restrictions
without having to limit input count, which also allows pcmbuf to
remain fuller and safely set its own buffer limits as it sees fit.
There is no longer a need to query input/output counts given a certain
number of input samples; just give it the sizes of the source and
destination buffers.

Works in harmony with stages that are not deterministic in terms of
sample input/output ratio (like both resamplers but most notably
the timestretch). As a result it fixes quirks with timestretch hanging
up with certain settings and it now operates properly throughout its
full settings range.
Change-Id: Ib206ec78f6f6c79259c5af9009fe021d68be9734
Reviewed-on: http://gerrit.rockbox.org/200
Reviewed-by: Michael Sevakis <jethead71@rockbox.org>
Tested-by: Michael Sevakis <jethead71@rockbox.org>
2012-04-29 10:00:56 +02:00

719 lines
36 KiB
ArmAsm
Raw Blame History

/***************************************************************************
* __________ __ ___.
* Open \______ \ ____ ____ | | _\_ |__ _______ ___
* Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ /
* Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < <
* Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \
* \/ \/ \/ \/ \/
* $Id$
*
* Copyright (C) 2006 Thom Johansen
* Portions Copyright (C) 2007 Michael Sevakis
*
* 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 pga_process(struct dsp_proc_entry *this, struct dsp_buffer **buf_p)
*/
.section .text
.align 2
.global pga_process
pga_process:
| input: 4(sp) = this, 8(sp) = buf_p
movem.l 4(%sp), %a0-%a1 | %a0 = this, %a1 = buf_p
move.l (%a0), %a0 | %a0 = this->data = &pga_data
move.l (%a0), %a0 | %a0 = data->gain
move.l (%a1), %a1 | %a1 = buf = *buf_p
lea.l -20(%sp), %sp | save registers
movem.l %d2-%d4/%a2-%a3, (%sp) |
clr.l %d1 | %d1 = buf->format.num_channels
move.b 17(%a1), %d1 |
10: | channel loop |
move.l (%a1), %d0 | %d0 = buf->remcount
move.l (%a1, %d1.l*4), %a2 | %a2 = s = buf->p32[ch-1]
move.l %a2, %a3 | %a3 = d = s
move.l (%a2)+, %d2 | %d2 = *s++,
mac.l %a0, %d2, (%a2)+, %d2, %acc0 | %acc0 = S(n)*gain, load S(n+1)
subq.l #1, %d0 | --count > 0 ? : effectively n++
ble.b 30f | loop done | no? finish up
20: | loop |
move.l %accext01, %d4 | fetch S(n-1)[7:0]
movclr.l %acc0, %d3 | fetch S(n-1)[40:8] in %d5[31:0]
asl.l #8, %d3 | *s++ = (S(n-1)[40:8] << 8) | S(n-1)[7:0]
mac.l %a0, %d2, (%a2)+, %d2, %acc0 | %acc0 = S(n)*gain, load S(n+1)
move.b %d4, %d3 |
move.l %d3, (%a3)+ |
subq.l #1, %d0 | --count > 0 ? : effectively n++
bgt.b 20b | loop | yes? do more samples
30: | loop done |
move.l %accext01, %d4 | fetch S(n-1)[7:0]
movclr.l %acc0, %d3 | fetch S(n-1)[40:8] in %d5[31:0]
asl.l #8, %d3 | *s = (S(n-1)[40:8] << 8) | S(n-1)[7:0]
move.b %d4, %d3 |
move.l %d3, (%a3) |
subq.l #1, %d1 | next channel
bgt.b 10b | channel loop |
movem.l (%sp), %d2-%d4/%a2-%a3 | restore registers
lea.l 20(%sp), %sp | cleanup stack
rts |
.size pga_process, .-pga_process
/****************************************************************************
* void crossfeed_process(struct dsp_proc_entry *this,
* struct dsp_buffer **buf_p)
*/
.section .text
.align 2
.global crossfeed_process
crossfeed_process:
| input: 4(sp) = this, 8(sp) = buf_p
lea.l -44(%sp), %sp |
movem.l %d2-%d7/%a2-%a6, (%sp) | save all regs
movem.l 48(%sp), %a1/%a4 | %a1 = this, %a4 = buf_p
move.l (%a4), %a4 | %a4 = buf = *buf_p
movem.l (%a4), %d7/%a4-%a5 | %d7 = buf->remcount, %a4 = buf->p32[0],
| %a5 = buf->p32[1]
move.l (%a1), %a1 | %a1 = &crossfeed_state
move.l (%a1)+, %d6 | %d6 = direct gain
movem.l 12(%a1), %d0-%d3 | fetch filter history samples
lea.l 132(%a1), %a6 | %a6 = delay line wrap limit
move.l (%a6), %a0 | fetch delay line address
movem.l (%a1), %a1-%a3 | load filter coefs
bra.b 20f | loop start | go to loop start point
/* Register usage in loop:
* %a0 = delay_p, %a1..%a3 = b0, b1, a1 (filter coefs),
* %a4 = buf[0], %a5 = buf[1],
* %a6 = delay line pointer wrap limit,
* %d0..%d3 = history
* %d4..%d5 = temp.
* %d6 = direct gain,
* %d7 = count
*/
10: | loop |
movclr.l %acc0, %d4 | write outputs
move.l %d4, (%a4)+ | .
movclr.l %acc1, %d5 | .
move.l %d5, (%a5)+ | .
20: | loop start |
mac.l %a2, %d0, (%a0)+, %d0, %acc0 | %acc0 = b1*dl[n - 1], %d0 = dl[n]
mac.l %a1, %d0 , %acc0 | %acc0 += b0*dl[n]
mac.l %a3, %d1, (%a5), %d5, %acc0 | %acc0 += a1*y_r[n - 1], load R
mac.l %a2, %d2, (%a0)+, %d2, %acc1 | %acc1 = b1*dr[n - 1], %d2 = dr[n]
mac.l %a1, %d2 , %acc1 | %acc1 += b0*dr[n]
mac.l %a3, %d3, (%a4), %d4, %acc1 | %acc1 += a1*y_l[n - 1], load L
movem.l %d4-%d5, -8(%a0) | save left & right inputs to delay line
move.l %acc0, %d3 | get filtered delayed left sample (y_l[n])
move.l %acc1, %d1 | get filtered delayed right sample (y_r[n])
mac.l %d6, %d4, %acc0 | %acc0 += gain*x_l[n]
mac.l %d6, %d5, %acc1 | %acc1 += gain*x_r[n]
cmp.l %a6, %a0 | wrap %a0 if passed end
bhs.b 30f | wrap buffer |
tpf.l | trap the buffer wrap
30: | wrap buffer | ...fwd taken branches more costly
lea.l -104(%a6), %a0 | wrap it up
subq.l #1, %d7 | --count > 0 ?
bgt.b 10b | loop | yes? do more
movclr.l %acc0, %d4 | write last outputs
move.l %d4, (%a4) | .
movclr.l %acc1, %d5 | .
move.l %d5, (%a5) | .
movem.l %d0-%d3, -120(%a6) | ...history
move.l %a0, (%a6) | ...delay_p
movem.l (%sp), %d2-%d7/%a2-%a6 | restore all regs
lea.l 44(%sp), %sp |
rts |
.size crossfeed_process,.-crossfeed_process
/****************************************************************************
* int lin_resample_resample(struct resample_data *data,
* struct dsp_buffer *src,
* struct dsp_buffer *dst)
*/
.section .text
.align 2
.global lin_resample_resample
lin_resample_resample:
| input: 4(sp) = data, 8(sp) = src, 12(sp) = dst
lea.l -44(%sp), %sp | save non-volatiles
movem.l %d2-%d7/%a2-%a6, (%sp) |
movem.l 48(%sp), %a0-%a2 | %a0 = data
| %a1 = src
| %a2 = dst
clr.l %d1 | %d1 = ch = src->format.num_channels
move.b 17(%a1), %d1 |
moveq.l #16, %d7 | %d7 = shift
.lrs_channel_loop: |
movem.l (%a0), %d2-%d3 | %d2 = delta = data->delta,
| %d3 = phase = data->phase
move.l (%a1), %d4 | %d4 = srcrem = src->remcount
move.l 12(%a2), %d5 | %d5 = dstrem = dst->bufcount
cmp.l #0x8000, %d4 | %d4 = MIN(srcrem, 0x8000)
ble.b 10f |
move.l #0x8000, %d4 |
10: |
move.l (%a1, %d1.l*4), %a3 | %a3 = s = src->p32[ch]
move.l (%a2, %d1.l*4), %a4 | %a4 = d = dst->p32[ch]
move.l %d3, %d0 | %d0 = pos
lsr.l %d7, %d0 | ...
beq.b 11f | pos == 0?
cmp.l %d4, %d0 | pos = MIN(pos, srcrem)
blt.b 12f |
move.l %d4, %d0 | pos = srcrem
move.l -4(%a3, %d0.l*4), %d6 | %d6 = last = s[pos - 1]
bra.w .lrs_channel_complete | at limit; nothing to do but next
11: |
move.l 4(%a0, %d1.l*4), %d6 | %d6 = last = last_sample[ch]
tpf.l | trap next move.l (last = s[pos - 1])
12: |
move.l -4(%a3, %d0.l*4), %d6 | %d6 = last = s[pos - 1]
cmp.l #0x10000, %d2 | delta >= 1.0?
bhs.b .lrs_downsample | yes? downsampling
|
/** Upsampling **/ |
lea.l (%a3, %d0.l*4), %a3 | %a3 = &s[pos]
sub.l %d4, %d0 | %d0 = pos - srcrem = -dte
lsl.l %d7, %d2 | move delta to bits 30..15
lsr.l #1, %d2 |
lsl.l %d7, %d3 | move phase to bits 30..15
lsr.l #1, %d3 |
move.l (%a3)+, %a5 | %a5 = s[pos]
move.l %a5, %a6 | %a6 = diff = s[pos] - last
sub.l %d6, %a6 |
bra.b 22f |
/* Funky loop structure is to avoid emac latency stalls */
20: |
move.l (%a3)+, %a5 | %a5 = s[pos]
move.l %a5, %a6 | %a6 = diff = s[pos] - last
sub.l %d6, %a6 |
21: |
movclr.l %acc0, %d7 | *d++ = %d7 = result
move.l %d7, (%a4)+ |
22: |
move.l %d6, %acc0 | %acc0 = last
mac.l %d3, %a6, %acc0 | %acc0 += frac * diff
subq.l #1, %d5 | dstrem <= 0?
ble.b 23f | yes? stop
add.l %d2, %d3 | phase += delta
bpl.b 21b | load next values?
move.l %a5, %d6 |
bclr.l #31, %d3 | clear sign bit
addq.l #1, %d0 | dte > 0?
bmi.b 20b | yes? continue resampling
tpf.w | trap next add.l (phase += delta)
23: |
add.l %d2, %d3 | phase += delta
lsl.l #1, %d3 | frac -> phase
bcs.b 24f | was sign bit set?
tpf.l |
24: |
move.l %a5, %d6 | yes? was going to move to new s[pos]
addq.l #1, %d0 |
movclr.l %acc0, %d7 | *d = %d7 = result
move.l %d7, (%a4) |
add.l %d4, %d0 | %d0 = -dte + srcrem = pos
or.l %d0, %d3 | restore phase
swap.w %d3 |
moveq.l #16, %d7 | %d7 = shift
bra.b .lrs_channel_complete |
|
/** Downsampling **/ |
.lrs_downsample: |
move.l (%a3, %d0.l*4), %a5 | %a5 = s[pos]
bra.b 31f |
30: |
lea.l -4(%a3, %d0.l*4), %a5 | %d6 = s[pos - 1], %a5 = s[pos]
movem.l (%a5), %d6/%a5 |
31: |
move.l %d6, %acc0 | %acc0 = last
sub.l %d6, %a5 | %a5 = diff = s[pos] - s[pos - 1]
move.l %d3, %d0 | frac = (phase << 16) >> 1
lsl.l %d7, %d0 |
lsr.l #1, %d0 |
mac.l %d0, %a5, %acc0 | %acc0 += frac * diff
add.l %d2, %d3 | phase += delta
move.l %d3, %d0 | pos = phase >> 16
lsr.l %d7, %d0 |
movclr.l %acc0, %a5 |
move.l %a5, (%a4)+ | *d++ = %d0
subq.l #1, %d5 | dst full?
ble.b 32f | yes? stop
cmp.l %d4, %d0 | pos < srcrem?
blt.b 30b | yes? continue resampling
tpf.l | trap cmp.l and ble.b
32: |
cmp.l %d4, %d0 | pos = MIN(pos, srcrem)
ble.b 33f |
move.l %d4, %d0 |
33: |
move.l -4(%a3, %d0.l*4), %d6 | %d6 = s[pos - 1]
|
.lrs_channel_complete: |
move.l %d6, 4(%a0, %d1.l*4) | last_sample[ch] = last
subq.l #1, %d1 | ch > 0?
bgt.w .lrs_channel_loop | yes? process next channel
|
move.l 12(%a2), %d1 | %d1 = dst->bufcount
sub.l %d5, %d1 | written = dst->bufcount - dstrem
move.l %d1, (%a2) | dst->remcount = written
move.l %d0, %d1 | wrap phase to position in next frame
lsl.l %d7, %d1 | data->phase = phase - (pos << 16)
sub.l %d1, %d3 | ...
move.l %d3, 4(%a0) | ...
movem.l (%sp), %d2-%d7/%a2-%a6 | restore non-volatiles
lea.l 44(%sp), %sp | cleanup stack
rts | buh-bye
.size lin_resample_resample, .-lin_resample_resample
/****************************************************************************
* void channel_mode_proc_mono(struct dsp_proc_entry *this,
* struct dsp_buffer **buf_p)
*
* Mix left and right channels 50/50 into a center channel.
*/
.section .text
.align 2
.global channel_mode_proc_mono
channel_mode_proc_mono:
| input: 4(sp) = this, 8(sp) = buf_p
move.l 8(%sp), %a0 | %a0 = buf_p
move.l (%a0), %a0 | %a0 = buf = *buf_p
lea.l -20(%sp), %sp | save registers
movem.l %d2-%d4/%a2-%a3, (%sp) |
movem.l (%a0), %d0/%a0-%a1 | %d0 = buf->remcount, %a0 = buf->p32[0],
| %a1 = buf->p32[1]
move.l %a0, %a2 | use separate dst pointers since read
move.l %a1, %a3 | pointers run one ahead of write
move.l #0x40000000, %d3 | %d3 = 0.5
move.l (%a0)+, %d1 | prime the input registers
move.l (%a1)+, %d2 |
mac.l %d1, %d3, (%a0)+, %d1, %acc0 |
mac.l %d2, %d3, (%a1)+, %d2, %acc0 |
subq.l #1, %d0 |
ble.s 20f | loop done |
10: | loop |
movclr.l %acc0, %d4 | L = R = l/2 + r/2
mac.l %d1, %d3, (%a0)+, %d1, %acc0 |
mac.l %d2, %d3, (%a1)+, %d2, %acc0 |
move.l %d4, (%a2)+ | output to original buffer
move.l %d4, (%a3)+ |
subq.l #1, %d0 |
bgt.s 10b | loop |
20: | loop done |
movclr.l %acc0, %d4 | output last sample
move.l %d4, (%a2) |
move.l %d4, (%a3) |
movem.l (%sp), %d2-%d4/%a2-%a3 | restore registers
lea.l 20(%sp), %sp | cleanup
rts |
.size channel_mode_proc_mono, .-channel_mode_proc_mono
/****************************************************************************
* void channel_mode_proc_custom(struct dsp_proc_entry *this,
* struct dsp_buffer **buf_p)
*
* Apply stereo width (narrowing/expanding) effect.
*/
.section .text
.align 2
.global channel_mode_proc_custom
channel_mode_proc_custom:
| input: 4(sp) = this, 8(sp) = buf_p
lea.l -28(%sp), %sp | save registers
movem.l %d2-%d6/%a2-%a3, (%sp) |
movem.l 32(%sp), %a0-%a1 | %a0 = this, %a1 = buf_p
move.l (%a1), %a1 | %a1 = buf = *buf_p
move.l (%a0), %a2 | %a2 = this->data = &channel_mode_data
movem.l (%a1), %d0/%a0-%a1 | %d0 = buf->remcount, %a0 = buf->p32[0],
| %a1 = buf->p32[1]
movem.l (%a2), %d3-%d4 | %d3 = sw_gain, %d4 = sw_cross
move.l %a0, %a2 | use separate dst pointers since read
move.l %a1, %a3 | pointers run one ahead of write
move.l (%a0)+, %d1 | prime the input registers
move.l (%a1)+, %d2 |
mac.l %d1, %d3 , %acc0 | L = l*gain + r*cross
mac.l %d1, %d4, (%a0)+, %d1, %acc1 | R = r*gain + l*cross
mac.l %d2, %d4 , %acc0 |
mac.l %d2, %d3, (%a1)+, %d2, %acc1 |
subq.l #1, %d0 |
ble.b 20f | loop done |
10: | loop |
movclr.l %acc0, %d5 |
movclr.l %acc1, %d6 |
mac.l %d1, %d3 , %acc0 | L = l*gain + r*cross
mac.l %d1, %d4, (%a0)+, %d1, %acc1 | R = r*gain + l*cross
mac.l %d2, %d4 , %acc0 |
mac.l %d2, %d3, (%a1)+, %d2, %acc1 |
move.l %d5, (%a2)+ |
move.l %d6, (%a3)+ |
subq.l #1, %d0 |
bgt.s 10b | loop |
20: | loop done |
movclr.l %acc0, %d5 | output last sample
movclr.l %acc1, %d6 |
move.l %d5, (%a2) |
move.l %d6, (%a3) |
movem.l (%sp), %d2-%d6/%a2-%a3 | restore registers
lea.l 28(%sp), %sp | cleanup
rts |
.size channel_mode_proc_custom, .-channel_mode_proc_custom
/****************************************************************************
* void channel_mode_proc_karaoke(struct dsp_proc_entry *this,
* struct dsp_buffer **buf_p)
*
* Separate channels into side channels.
*/
.section .text
.align 2
.global channel_mode_proc_karaoke
channel_mode_proc_karaoke:
| input: 4(sp) = this, 8(sp) = buf_p
move.l 8(%sp), %a0 | %a0 = buf_p
move.l (%a0), %a0 | %a0 = buf = *buf_p
lea.l -20(%sp), %sp | save registers
movem.l %d2-%d4/%a2-%a3, (%sp) |
movem.l (%a0), %d0/%a0-%a1 | %d0 = buf->remcount, %a0 = buf->p32[0],
| %a1 = buf->p32[1]
move.l %a0, %a2 | use separate dst pointers since read
move.l %a1, %a3 | pointers run one ahead of write
move.l #0x40000000, %d3 | %d3 = 0.5
move.l (%a0)+, %d1 | prime the input registers
move.l (%a1)+, %d2 |
mac.l %d1, %d3, (%a0)+, %d1, %acc0 | L = l/2 - r/2
msac.l %d2, %d3, (%a1)+, %d2, %acc0 |
subq.l #1, %d0 |
ble.b 20f | loop done |
10: | loop |
movclr.l %acc0, %d4 |
mac.l %d1, %d3, (%a0)+, %d1, %acc0 | L = l/2 - r/2
msac.l %d2, %d3, (%a1)+, %d2, %acc0 |
move.l %d4, (%a2)+ |
neg.l %d4 | R = -L = -(l/2 - r/2) = r/2 - l/2
move.l %d4, (%a3)+ |
subq.l #1, %d0 |
bgt.s 10b | loop |
20: | loop done |
movclr.l %acc0, %d4 | output last sample
move.l %d4, (%a2) |
neg.l %d4 | R = -L = -(l/2 - r/2) = r/2 - l/2
move.l %d4, (%a3) |
movem.l (%sp), %d2-%d4/%a2-%a3 | restore registers
lea.l 20(%sp), %sp | cleanup
rts |
.size channel_mode_proc_karaoke, .-channel_mode_proc_karaoke
/****************************************************************************
* void filter_process(struct dsp_filter *f, int32_t *buf[], int count,
* unsigned int channels)
*
* define HIGH_PRECISION as '1' to make filtering calculate lower bits after
* shifting. without this, "shift" - 1 of the lower bits will be lost here.
*/
#define HIGH_PRECISION 0
.text
.global filter_process
filter_process:
| input: 4(sp) = f, 8(sp) = buf, 12(sp) = count, 16(sp) = channels
lea.l -44(%sp), %sp | save clobbered regs
#if HIGH_PRECISION
movem.l %d2-%d7/%a2-%a6, (%sp) | .
#else
movem.l %d2-%d6/%a2-%a6, (%sp) |
#endif
move.l 48(%sp), %a5 | fetch filter structure address
clr.l %d6 | load shift count
move.b 52(%a5), %d6 | .
subq.l #1, %d6 | EMAC gives us one free shift
#if HIGH_PRECISION
moveq.l #8, %d7
sub.l %d6, %d7 | shift for lower part of accumulator
#endif
movem.l (%a5), %a0-%a4 | load coefs
lea.l 20(%a5), %a5 | point to filter history
10: | channel loop
move.l 52(%sp), %a6 | load input channel pointer
addq.l #4, 52(%sp) | point x to next channel
move.l (%a6), %a6 |
move.l 56(%sp), %d5 | number of samples
movem.l (%a5), %d0-%d3 | load filter history
| d0-d3 = history, d4 = temp, d5 = sample count, d6 = upper shift amount,
| d7 = lower shift amount,a0-a4 = coefs, a5 = history pointer, a6 = buf[ch]
20: | loop
| Direct form 1 filtering code. We assume DSP has put EMAC in frac mode.
| 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 do parallel load of input value.
mac.l %a2, %d1, %acc0 | acc = b2*x[i - 2]
move.l %d0, %d1 | fix input history
mac.l %a1, %d0, (%a6), %d0, %acc0 | acc += b1*x[i - 1], x[i] -> d0
mac.l %a0, %d0, %acc0 | acc += b0*x[i]
mac.l %a3, %d2, %acc0 | acc += a1*y[i - 1]
mac.l %a4, %d3, %acc0 | acc += a2*y[i - 2]
move.l %d2, %d3 | fix output history
#if HIGH_PRECISION
move.l %accext01, %d2 | fetch lower part of accumulator
move.b %d2, %d4 | clear upper three bytes
lsr.l %d7, %d4 | shift lower bits
#endif
movclr.l %acc0, %d2 | fetch upper part of result
asl.l %d6, %d2 | restore fixed point format
#if HIGH_PRECISION
or.l %d2, %d4 | combine lower and upper parts
#endif
move.l %d2, (%a6)+ | save result
subq.l #1, %d5 | are we done with this channel?
bgt 20b | loop
movem.l %d0-%d3, (%a5) | save history back to struct
lea.l 16(%a5), %a5 | point to next channel's history
subq.l #1, 60(%sp) | have we processed both channels?
bhi 10b | channel loop
#if HIGH_PRECISION
movem.l (%sp), %d2-%d7/%a2-%a6
#else
movem.l (%sp), %d2-%d6/%a2-%a6
#endif
lea.l 44(%sp), %sp
rts
.size filter_process, .-filter_process
/****************************************************************************
* void sample_output_stereo(struct sample_io_data *this,
* struct dsp_buffer *src,
* struct dsp_buffer *dst)
*
* Framework based on the ubiquitous Rockbox line transfer logic for
* Coldfire CPUs.
*
* Does emac clamping and scaling (which proved faster than the usual
* checks and branches - even single test clamping) and writes using
* line burst transfers. Also better than writing a single L-R pair per
* loop but a good deal more code.
*
* Attemping bursting during reads is rather futile since the source and
* destination alignments rarely agree and too much complication will
* slow us up. The parallel loads seem to do a bit better at least until
* a pcm buffer can always give line aligned chunk and then aligning the
* dest can then imply the source is aligned if the source buffers are.
* For now longword alignment is assumed of both the source and dest.
*
*/
.section .text
.align 2
.global sample_output_stereo
sample_output_stereo:
| input: 4(sp) = count, 8(sp) = src, 12(sp) = dst
lea.l -48(%sp), %sp | save registers
move.l %macsr, %d1 | do it now as at many lines will
movem.l %d1-%d7/%a2-%a6, (%sp) | be the far more common condition
move.l #0x80, %macsr | put emac unit in signed int mode
movem.l 52(%sp), %a0-%a2 | %a0 = this, %a1 = src, %a2 = dst
move.l (%a0), %a0 | %a0 = this->outcount
move.l 4(%a2), %a4 | %a4 = dst->p16out
lea.l (%a4, %a0.l*4), %a0 | %a0 = count -> end address
movem.l 4(%a1), %a2-%a3 | %a2 = src->p32[0], %a3 = src->p32[1]
clr.l %d1 | %a1 = multiplier: (1 << (16 - scale))
move.b 19(%a1), %d1 | %d1 = src->format.output_scale
sub.l #16, %d1 |
neg.l %d1 |
moveq.l #1, %d0 |
asl.l %d1, %d0 |
move.l %d0, %a1 |
move.l #0x8000, %a6 | %a6 = rounding term
moveq.l #28, %d0 | %d0 = second line bound
add.l %a4, %d0 |
and.l #0xfffffff0, %d0 |
cmp.l %a0, %d0 | at least a full line?
bhi.w 40f | long loop 1 start | no? do as trailing longwords
sub.l #16, %d0 | %d1 = first line bound
cmp.l %a4, %d0 | any leading longwords?
bls.b 20f | line loop start | no? start line loop
10: | long loop 0 |
move.l (%a2)+, %d1 | read longword from L and R
move.l %a6, %acc0 |
move.l %acc0, %acc1 |
mac.l %d1, %a1, (%a3)+, %d2, %acc0 | shift L to high word
mac.l %d2, %a1, %acc1 | shift R to high word
movclr.l %acc0, %d1 | get possibly saturated results
movclr.l %acc1, %d2 |
swap.w %d2 | move R to low word
move.w %d2, %d1 | interleave MS 16 bits of each
move.l %d1, (%a4)+ | ...and write both
cmp.l %a4, %d0 |
bhi.b 10b | long loop 0 |
20: | line loop start |
lea.l -12(%a0), %a5 | %a5 = at or just before last line bound
30: | line loop |
move.l (%a3)+, %d4 | get next 4 R samples and scale
move.l %a6, %acc0 |
move.l %acc0, %acc1 |
move.l %acc1, %acc2 |
move.l %acc2, %acc3 |
mac.l %d4, %a1, (%a3)+, %d5, %acc0 | with saturation
mac.l %d5, %a1, (%a3)+, %d6, %acc1 |
mac.l %d6, %a1, (%a3)+, %d7, %acc2 |
mac.l %d7, %a1, (%a2)+, %d0, %acc3 |
lea.l 16(%a4), %a4 | increment dest here, mitigate stalls
movclr.l %acc0, %d4 | obtain R results
movclr.l %acc1, %d5 |
movclr.l %acc2, %d6 |
movclr.l %acc3, %d7 |
move.l %a6, %acc0 |
move.l %acc0, %acc1 |
move.l %acc1, %acc2 |
move.l %acc2, %acc3 |
mac.l %d0, %a1, (%a2)+, %d1, %acc0 | get next 4 L samples and scale
mac.l %d1, %a1, (%a2)+, %d2, %acc1 | with saturation
mac.l %d2, %a1, (%a2)+, %d3, %acc2 |
mac.l %d3, %a1 , %acc3 |
swap.w %d4 | a) interleave most significant...
swap.w %d5 |
swap.w %d6 |
swap.w %d7 |
movclr.l %acc0, %d0 | obtain L results
movclr.l %acc1, %d1 |
movclr.l %acc2, %d2 |
movclr.l %acc3, %d3 |
move.w %d4, %d0 | a) ... 16 bits of L and R
move.w %d5, %d1 |
move.w %d6, %d2 |
move.w %d7, %d3 |
movem.l %d0-%d3, -16(%a4) | write four stereo samples
cmp.l %a4, %a5 |
bhi.b 30b | line loop |
40: | long loop 1 start |
cmp.l %a4, %a0 | any longwords left?
bls.b 60f | output end | no? stop
50: | long loop 1 |
move.l (%a2)+, %d1 | handle trailing longwords
move.l %a6, %acc0 |
move.l %acc0, %acc1 |
mac.l %d1, %a1, (%a3)+, %d2, %acc0 | the same way as leading ones
mac.l %d2, %a1, %acc1 |
movclr.l %acc0, %d1 |
movclr.l %acc1, %d2 |
swap.w %d2 |
move.w %d2, %d1 |
move.l %d1, (%a4)+ |
cmp.l %a4, %a0 |
bhi.b 50b | long loop 1
60: | output end |
movem.l (%sp), %d1-%d7/%a2-%a6 | restore registers
move.l %d1, %macsr |
lea.l 48(%sp), %sp | cleanup
rts |
.size sample_output_stereo, .-sample_output_stereo
/****************************************************************************
* void sample_output_mono(struct sample_io_data *this,
* struct dsp_buffer *src,
* struct dsp_buffer *dst)
*
* Same treatment as sample_output_stereo but for one channel.
*/
.section .text
.align 2
.global sample_output_mono
sample_output_mono:
| input: 4(sp) = count, 8(sp) = src, 12(sp) = dst
lea.l -32(%sp), %sp | save registers
move.l %macsr, %d1 | do it now as at many lines will
movem.l %d1-%d5/%a2-%a4, (%sp) | be the far more common condition
move.l #0x80, %macsr | put emac unit in signed int mode
movem.l 36(%sp), %a0-%a2 | %a0 = this, %a1 = src, %a2 = dst
move.l (%a0), %a0 | %a0 = this->outcount
move.l 4(%a2), %a3 | %a3 = dst->p16out
movem.l 4(%a1), %a2 | %a2 = src->p32[0]
lea.l (%a3, %a0.l*4), %a0 | %a0 = count -> end address
clr.l %d1 | %d5 = multiplier: (1 << (16 - scale))
move.b 19(%a1), %d1 | %d1 = src->format.output_scale
sub.l #16, %d1 |
neg.l %d1 |
moveq.l #1, %d5 |
asl.l %d1, %d5 |
move.l #0x8000, %a4 | %a4 = rounding term
moveq.l #28, %d0 | %d0 = second line bound
add.l %a3, %d0 |
and.l #0xfffffff0, %d0 |
cmp.l %a0, %d0 | at least a full line?
bhi.w 40f | long loop 1 start | no? do as trailing longwords
sub.l #16, %d0 | %d1 = first line bound
cmp.l %a3, %d0 | any leading longwords?
bls.b 20f | line loop start | no? start line loop
10: | long loop 0 |
move.l (%a2)+, %d1 | read longword from L and R
move.l %a4, %acc0 |
mac.l %d1, %d5, %acc0 | shift L to high word
movclr.l %acc0, %d1 | get possibly saturated results
move.l %d1, %d2 |
swap.w %d2 | move R to low word
move.w %d2, %d1 | duplicate single channel into
move.l %d1, (%a3)+ | L and R
cmp.l %a3, %d0 |
bhi.b 10b | long loop 0 |
20: | line loop start |
lea.l -12(%a0), %a1 | %a1 = at or just before last line bound
30: | line loop |
move.l (%a2)+, %d0 | get next 4 L samples and scale
move.l %a4, %acc0 |
move.l %acc0, %acc1 |
move.l %acc1, %acc2 |
move.l %acc2, %acc3 |
mac.l %d0, %d5, (%a2)+, %d1, %acc0 | with saturation
mac.l %d1, %d5, (%a2)+, %d2, %acc1 |
mac.l %d2, %d5, (%a2)+, %d3, %acc2 |
mac.l %d3, %d5 , %acc3 |
lea.l 16(%a3), %a3 | increment dest here, mitigate stalls
movclr.l %acc0, %d0 | obtain results
movclr.l %acc1, %d1 |
movclr.l %acc2, %d2 |
movclr.l %acc3, %d3 |
move.l %d0, %d4 | duplicate single channel
swap.w %d4 | into L and R
move.w %d4, %d0 |
move.l %d1, %d4 |
swap.w %d4 |
move.w %d4, %d1 |
move.l %d2, %d4 |
swap.w %d4 |
move.w %d4, %d2 |
move.l %d3, %d4 |
swap.w %d4 |
move.w %d4, %d3 |
movem.l %d0-%d3, -16(%a3) | write four stereo samples
cmp.l %a3, %a1 |
bhi.b 30b | line loop |
40: | long loop 1 start |
cmp.l %a3, %a0 | any longwords left?
bls.b 60f | output end | no? stop
50: | loop loop 1 |
move.l (%a2)+, %d1 | handle trailing longwords
move.l %a4, %acc0 |
mac.l %d1, %d5, %acc0 | the same way as leading ones
movclr.l %acc0, %d1 |
move.l %d1, %d2 |
swap.w %d2 |
move.w %d2, %d1 |
move.l %d1, (%a3)+ |
cmp.l %a3, %a0 |
bhi.b 50b | long loop 1 |
60: | output end |
movem.l (%sp), %d1-%d5/%a2-%a4 | restore registers
move.l %d1, %macsr |
lea.l 32(%sp), %sp | cleanup
rts |
.size sample_output_mono, .-sample_output_mono
Reference in a new issue View git blame Copy permalink