less then 10 samples delay in code??

[Nubeat7]

hello, i have a question about small samples delays (less then10) and want to ask if this code works as i understand ?

Code: Select all

streamin i;
streamout o1;
streamout o2;
streamout o3;

float x1,x2,x3;

o3=x3; //3 samples delay
o2=x2; //2 samples delay
o1=x1; //1 samples delay
x3=x2;
x2=x1;
x1=i;

i also thought about to write some kind of recursive loop function in assembler to loop the input signal around to get one sample delay with each loop, like this you could set the delaytime in samples (like loop 5 times would be 5 samples delay) or do i think the wrong direction?

the goal is to use such a simple code for small latency compensation between dry and effected signals, or should i just use the one sample delay primitives for this?

[martinvicanek]

Yes, it works the expected way!

smallDelay.png (30.6 KiB) Viewed 30534 times

If you just want the delays you could even do without x1, x2 etc.

As you say this bucket brigade approach is only viable for short delays of a few samples. For long delays you want to use arrays instead. However, arrays are slow, so bucket brigade is preferable for short delays. Would be interesting to find out where the break even point is.

[nix]

it's a mem,
so writing to memory.
just to clarify for some maybe.

Cheers, that's interesting fellas

Would yous like the delay I built out of prims and snippets?
Seems to work OK

[KG_is_back]

As you say this bucket brigade approach is only viable for short delays of a few samples. For long delays you want to use arrays instead. However, arrays are slow, so bucket brigade is preferable for short delays. Would be interesting to find out where the break even point is.

Arrays are slow only in default Code, because each channel might need different index to read/write, each of the 4 channels need to be read individually. That means: 1sample delay bucked brigade takes 4times less CPU than using array delay (of any length). I've just checked that quick with analyzer primitive and it's true... That means bucket brigade is more efficient for delays upto 4 (maybe even a little more, depending on the index counter in array delay).

However, when the delay is fixed (by fixed I mean both - hard coded delay length and variable delay length that is set on first sample)and identical for all 4 channels, you can use assembly to greatly improve the delay algorithm because you can read and write to array all 4 channels at the time (to the same index). With this approach can be as effective as bucket brigade.

here's a schematic that shows it all. 3 code examples all with loop(1024) to boost CPU load to visible range.

first one is just o code that involves one read and one write from/to array. Default code form code component that every array based delay must contain ( + index counter which is actually not here). on my machine it takes 25-28cycles per sample.

second one is a bucket brigade type one sample delay. it takes 5-6cycles on my machine.

third one is assembly optimized fixed delay (defined on first sample, then stays fixed) and it takes 5-7cycles in my machine. And it supports virtually any delay (the given example is limited to 4096, but you can increase that easily by boosting mem size and both integers in "and eax,???;" operation to (memsize*16-1) )

note - inside the module the delay is also with 1024 loop - fully working code is the one outside the module.

[Nubeat7]

wow, thx KG thats very enlightening, your optimized delay already needs less cycles then the bucket brigade with 2 samples!

[martinvicanek]

Good work, KG! I think your array optimization is fair if you compare to bucket brigade where you also have the same delay for all SSE channels. However, sometimes this may be a limitation. Have you looked at Trogz Toolz optimized delays over at the SM forum?

[KG_is_back]

Good work, KG! I think your array optimization is fair if you compare to bucket brigade where you also have the same delay for all SSE channels. However, sometimes this may be a limitation. Have you looked at Trogz Toolz optimized delays over at the SM forum?

Where is it? I can't find it anywhere...

anyway, one of main things how to optimize a delay is to make use of SSE instructions as much as possible

here another comparison:

stock FS delay:
minimal delay is 1
uses cca. 36-40cycles (possibly -5cycles, because it has one extra 1024loop which uses cca 5 cycles);
My reworked delay:
works for 0 samples delay too
uses cca.15cycles

[martinvicanek]

Have you looked at Trogz Toolz optimized delays over at the SM forum?

Where is it? I can't find it anywhere...

It's visible to registered users only: Trogz Tools. You have to insist, the pages in the SM forum almost never load at the first intent, instead a misleading "Forbidden" message appears. Probably some timeout issue. It is a shame to see such a valuable source of material go down the drain.

[tester]

With the SM forum it is happening from time to time. Malc isn't sure why. I noticed, that it usually starts when spammers start to play.

[Tepeix]

Hi,

This old topic inspire me to make my first asm delay.
A mini mod delay of 7 samples.

What do you think of it ? Could it be more optimized ?

Here's the code :

Code: Select all

streaminin;streamoutout;
streamin mod;
float a=0; float a1=0;
float a2=0;float a3=0;
float a4=0;float a5=0;
float a6=0;float f0=0;
int abs=2147483647;   
movaps xmm0,in;
movaps xmm1,a;
movaps a,xmm0;
movaps xmm2,a1;
movaps a1,xmm1;
movaps xmm3,a2;
movaps a2,xmm2;
movaps xmm4,a3;
movaps a3,xmm3;
movaps xmm5,a4;
movaps a4,xmm4;
movaps xmm6,a5;
movaps a5,xmm5;
movaps xmm7,a6;
movaps a6,xmm6;
// mod //
subps xmm3,xmm7;
movaps xmm0,mod;
movaps xmm1,xmm0;
cmpps xmm1,f0,1;
andps xmm1,xmm3;
addps xmm7,xmm1;
// abs //
andps xmm0,abs;
// mix //
subps xmm7,xmm5;
mulps xmm7,xmm0;
addps xmm5,xmm7;
movaps out,xmm5;

Combining in series the 4 channels we get a little vibrato.)