Nyaaori/rockbox
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity
rockbox/apps/dsp_cf.S
Michael Sevakis 7c120fb0ce SWCODEC-Coldfire: Small upsampling tweek improves speed and size a tiny bit and prevents unneeded reloading of the previous sample. I imagine it would help most with non-iram source buffers. 
git-svn-id: svn://svn.rockbox.org/rockbox/trunk@12425 a1c6a512-1295-4272-9138-f99709370657
2007-02-21 07:06:58 +00:00

223 lines
11 KiB
ArmAsm
Raw Blame History

/***************************************************************************
* __________ __ ___.
* Open \______ \ ____ ____ | | _\_ |__ _______ ___
* Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ /
* Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < <
* Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \
* \/ \/ \/ \/ \/
* $Id$
*
* Copyright (C) 2006 Thom Johansen
*
* 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.
*
****************************************************************************/
/****************************************************************************
* apply_crossfeed(int32_t* src[], int count)
*/
.section .text
.global apply_crossfeed
apply_crossfeed:
lea.l (-44, %sp), %sp
movem.l %d2-%d7/%a2-%a6, (%sp) | save all regs
move.l (44+4, %sp), %a4
movem.l (%a4), %a4-%a5 | a4 = src[0], a5 = src[1]
move.l (44+8, %sp), %d7 | d7 = count
lea.l crossfeed_data, %a1
lea.l (8*4, %a1), %a0 | a0 = &delay[0][0]
move.l (%a1)+, %a6 | a6 = direct gain
movem.l (3*4, %a1), %d0-%d3 | fetch filter history samples
move.l (33*4, %a1), %d4 | fetch delay line index
movem.l (%a1), %a1-%a3 | load filter coefs
move.l %d4, %d5
lsl.l #3, %d5
add.l %d5, %a0 | point a0 to current delay position
| lea.l (%d4*4, %a0), %a0
| lea.l (%d4*4, %a0), %a0 | point a0 to current delay position
/* Register usage in loop:
* a0 = &delay[index][0], a1..a3 = b0, b1, a1 (filter coefs),
* a4 = src[0], a5 = src[1], a6 = direct gain,
* d0..d3 = history
* d4 = delay line index,
* d5,d6 = temp.
* d7 = count
*/
.cfloop:
mac.l %a2, %d0, (4, %a0), %d0, %acc0 | acc = b1*dr[n - 1] d0 = dr[n]
mac.l %a1, %d0, %acc0 | acc += b0*dr[n]
mac.l %a3, %d1, (%a4), %d5, %acc0 | acc += a1*y_l[n - 1], load left input
move.l %acc0, %d1 | get filtered delayed sample
mac.l %a6, %d5, %acc0 | acc += gain*x_l[n]
movclr.l %acc0, %d6
move.l %d6, (%a4)+ | write result
mac.l %a2, %d2, (%a0), %d2, %acc0 | acc = b1*dl[n - 1], d2 = dl[n]
move.l %d5, (%a0)+ | save left input to delay line
mac.l %a1, %d2, %acc0 | acc += b0*dl[n]
mac.l %a3, %d3, (%a5), %d5, %acc0 | acc += a1*y_r[n - 1], load right input
move.l %acc0, %d3 | get filtered delayed sample
mac.l %a6, %d5, %acc0 | acc += gain*x_r[n]
move.l %d5, (%a0)+ | save right input to delay line
movclr.l %acc0, %d6
move.l %d6, (%a5)+ | write result
addq.l #1, %d4 | index++
moveq.l #13, %d6
cmp.l %d6, %d4 | wrap index to 0 if it overflows
jlt .cfnowrap
moveq.l #13*8, %d4
sub.l %d4, %a0 | wrap back delay line ptr as well
clr.l %d4
.cfnowrap:
subq.l #1, %d7
jne .cfloop
| save data back to struct
lea.l crossfeed_data + 4*4, %a1
movem.l %d0-%d3, (%a1)
move.l %d4, (30*4, %a1)
movem.l (%sp), %d2-%d7/%a2-%a6
lea.l (44, %sp), %sp
rts
.cfend:
.size apply_crossfeed,.cfend-apply_crossfeed
/****************************************************************************
* dsp_downsample(int channels, int count, struct resample_data *r,
* in32_t **src, int32_t **dst)
*/
.section .text
.global dsp_downsample
dsp_downsample:
lea.l -40(%sp), %sp | save non-clobberables
movem.l %d2-%d7/%a2-%a5, (%sp) |
movem.l 44(%sp), %d2-%d3/%a0-%a2| %d2 = ch = channels
| %d3 = count
| %a0 = r
| %a1 = src
| %a2 = dst
move.l 4(%a0), %d4 | %d4 = delta = r->delta
move.l #16, %d7 | %d7 = shift
.dschannel_loop:
move.l (%a0), %d5 | %d5 = phase = r->phase
move.l -4(%a1, %d2.l*4), %a3 | %a3 = s = src[ch-1]
move.l -4(%a2, %d2.l*4), %a4 | %a4 = d = dst[ch-1]
lea.l 4(%a0, %d2.l*4), %a5 | %a5 = &r->last_sample[ch-1]
move.l (%a5), %d0 | %d0 = last = r->last_sample[ch-1]
move.l -4(%a3, %d3.l*4), %d1 | r->last_sample[ch-1] = s[count-1]
move.l %d1, (%a5) |
move.l %d5, %d6 | %d6 = pos = phase >> 16
lsr.l %d7, %d6 |
cmp.l %d3, %d6 | past end of samples?
bge.b .dsloop_skip | yes? skip loop
tst.l %d6 | need last sample of prev. frame?
bne.b .dsloop | no? start main loop
move.l (%a3, %d6.l*4), %d1 | %d1 = s[pos]
bra.b .dsuse_last_start | start with last (last in %d0)
.dsloop:
lea.l -4(%a3, %d6.l*4), %a5 | load s[pos-1] and s[pos]
movem.l (%a5), %d0-%d1 |
.dsuse_last_start:
sub.l %d0, %d1 | %d1 = diff = s[pos] - s[pos-1]
move.l %d0, %acc0 | %acc0 = previous sample
move.l %d5, %d0 | frac = (phase << 16) >> 1
lsl.l %d7, %d0 |
lsr.l #1, %d0 |
mac.l %d0, %d1, %acc0 | %acc0 += frac * diff
move.l %acc0, %d0 |
add.l %d4, %d5 | phase += delta
move.l %d5, %d6 | pos = phase >> 16
lsr.l %d7, %d6 |
move.l %d0, (%a4)+ | *d++ = %d0
cmp.l %d3, %d6 | pos < count?
blt.b .dsloop | yes? continue resampling
.dsloop_skip:
subq.l #1, %d2 | ch > 0?
bgt.b .dschannel_loop | yes? process next channel
asl.l %d7, %d3 | wrap phase to start of next frame
sub.l %d3, %d5 | r->phase = phase - (count << 16)
move.l %d5, (%a0) |
move.l %a4, %d0 | return d - d[0]
sub.l (%a2), %d0 |
asr.l #2, %d0 | convert bytes->samples
movem.l (%sp), %d2-%d7/%a2-%a5 | restore non-clobberables
move.l %acc1, %acc0 | clear %acc0
lea.l 40(%sp), %sp | cleanup stack
rts | buh-bye
.dsend:
.size dsp_downsample,.dsend-dsp_downsample
/****************************************************************************
* dsp_upsample(int channels, int count, struct resample_data *r,
* in32_t **src, int32_t **dst)
*/
.section .text
.global dsp_upsample
dsp_upsample:
lea.l -40(%sp), %sp | save non-clobberables
movem.l %d2-%d7/%a2-%a5, (%sp) |
movem.l 44(%sp), %d2-%d3/%a0-%a2| %d2 = ch = channels
| %d3 = count
| %a0 = r
| %a1 = src
| %a2 = dst
move.l 4(%a0), %d4 | %d4 = delta = r->delta
swap %d4 | swap delta to high word to use
| carries to increment position
.uschannel_loop:
move.l (%a0), %d5 | %d5 = phase = r->phase
move.l -4(%a1, %d2.l*4), %a3 | %a3 = s = src[ch-1]
lea.l 4(%a0, %d2.l*4), %a4 | %a4 = &r->last_sample[ch-1]
lea.l (%a3, %d3.l*4), %a5 | %a5 = src_end = &src[count]
move.l (%a4), %d0 | %d0 = last = r->last_sample[ch-1]
move.l -4(%a5), %d1 | r->last_sample[ch-1] = s[count-1]
move.l %d1, (%a4) |
move.l -4(%a2, %d2.l*4), %a4 | %a4 = d = dst[ch-1]
moveq.l #16, %d1 | %d0 = shift
move.l %d5, %d6 | %d6 = pos = phase >> 16
lsl.l %d1, %d5 | swap phase to high word to use
| carries to increment position
lsr.l %d1, %d6 | pos == 0?
beq.b .usstart_0 | no? transistion from down
cmp.l %d3, %d6 | past end of samples?
bge.b .usloop_skip | yes? skip loop
lea.l 4(%a3, %d6.l*4), %a3 | %a3 = s + pos + 1
movem.l -8(%a3), %d0-%d1 | %d0 = *(s - 2), %d1 = *(s - 1)
.word 0x51fa | tpf.w - trap next instruction
.usstart_0:
move.l (%a3)+, %d1 | %d1 = *s++
.word 0x51fb | tpf.l - trap next two instructions
.usloop_1:
move.l %d6, %d0 | move previous sample to %d0
move.l (%a3)+, %d1 | fetch next sample
move.l %d1, %d6 | save sample value
sub.l %d0, %d1 | %d1 = diff = s[pos] - s[pos-1]
.usloop_0:
lsr.l #1, %d5 | make phase into frac
mac.l %d1, %d5, %acc0 | %acc0 = diff * frac
movclr.l %acc0, %d7 | %d7 = product
lsl.l #1, %d5 | restore frac to phase
add.l %d0, %d7 | %d7 = last + product
move.l %d7, (%a4)+ | *d++ = %d7
add.l %d4, %d5 | phase += delta
bcc.b .usloop_0 | load next values?
cmp.l %a5, %a3 | src < src_end?
blt.b .usloop_1 | yes? continue resampling
.usloop_skip:
subq.l #1, %d2 | ch > 0?
bgt.b .uschannel_loop | yes? process next channel
swap %d5 | wrap phase to start of next frame
move.l %d5, (%a0) | ...and save in r->phase
move.l %a4, %d0 | return d - d[0]
sub.l (%a2), %d0 |
movem.l (%sp), %d2-%d7/%a2-%a5 | restore non-clobberables
asr.l #2, %d0 | convert bytes->samples
lea.l 40(%sp), %sp | cleanup stack
rts | buh-bye
.usend:
.size dsp_upsample,.usend-dsp_upsample
Reference in a new issue View git blame Copy permalink