#161202 - Ruben - Wed Jul 30, 2008 1:33 pm
| Code: |
@ assuming that...
@ r3: left volume | (right volume << 16)
@ r4: ROM offset
@ r5: 22.10 offset from "ROM offset" (r4)
@ r9: mixed samples (16 bits left, 16 bits right)
add r12, r4, r5, lsr #shift
ldrsb r12, [r12, #0x10]
mul r12, r3, r12
add r9, r9, r12, asr #0x07 |
That's the mixing step. After doing a small calculation (16 channels * 127 max volume * (+127 highest sample) > 65535), I *assume* that it would overflow if all 16 channels were active and mixed 127. So I used that division/shift before adding thing and this is in my mix-down/clipping loop:
| Code: |
.LClipLoop:
ldrsh ip, [r0], #0x02
bic ip, ip, #0xFF00
mov ip, ip, asr #0x02
strb ip, [r1], #0x01
ldrsh ip, [r0], #0x02
mov ip, ip, asr #0x02
strb ip, [r3], #0x01
subs r2, r2, #0x01
bcs .LClipLoop |
So... I thought that all I would have to do is clear some bits of the left channel (which would be an overflow, I think) but I just get static. Any ideas why? Am I missing something TOTALLY obvious?
#161204 - Ruben - Wed Jul 30, 2008 1:47 pm
FOLLOWUP:
On further inspection, the right channel is perfectly fine but it's the left that has the static.
#161207 - eKid - Wed Jul 30, 2008 3:49 pm
Your problem is probably here:
| Code: |
| add r9, r9, r12, asr #0x07 |
The top value (right output) gets shifted fine (with sign extension), but the bottom value (left output) gets the bottom bits shifted out of the top value.
example:
| Code: |
samp = -20 [0FFFFFFECh]
vol = 127,127 [07F007F]
samp * vol = 0F613F614h (note also the 1 sample error you'll get with negative numbers)
value >> 7 = 0FFEC27ECh = 0FFEC, 27ECh = -20, 10220 :\
|
#161230 - Ruben - Thu Jul 31, 2008 2:11 am
Hmm... anyone have any idea what to do then? My mixer runs at 1.7% CPU per channel @ 18157Hz (any panning) which is... slow IMO.
#161232 - eKid - Thu Jul 31, 2008 4:12 am
Use unsigned samples :P
#161234 - Ruben - Thu Jul 31, 2008 4:19 am
LOL!!! Unsigned samples get rid of this problem?
#161235 - eKid - Thu Jul 31, 2008 4:28 am
Something like this:
| Code: |
signed sample data [-128..127]
-20, -10, 0, 10, 20, etc
unsigned sample data [+128, or ^0x80]
108, 118, 128, 138, 148
during calculations:
add up volume of each channel (used to convert unsigned to signed during the mixdown)
mixing loop:
read unsigned sample [can use ldrb which can shift the position register]
multiply by volume [left and right]
clear bits that will be shifted into low word [clear 0x1F0000 for 32 channels]
add shifted value to mix [lsr#5 for 32 channels]
mixdown:
read sample
subtract volume total * x (x equals amount to match the mixing bit depth)
shift sample to 8-bit
clamp sample to -128 to 127
write sample to output
|
#161678 - Ruben - Sun Aug 10, 2008 4:53 am
Hmm... well, OK... but that still doesn't answer my question! :P Is my hypothesis (WOOT! Learned a new word at school!) correct and causing an overflow, or am I just going nuts, lol?