Alternative clipper

infuzion · 5 Dec 2009

Not sure how many cycles saved; the CPU Cycle Analyzer is a bit wonky on my Core2. I'm guessing 5-8 cycles, up to half. // Dynamic clipper by Mhz, re8 redux streamin in; streamin thresh; streamin angle; streamout out; float F1=1; float FN1=-1; //float F0P1=0.999985;float F0P2=-0.999985; //final min/max clip handled by angle GUI; set max angle to 0.999985 movaps xmm0,thresh; movaps xmm1,xmm0; //movaps xmm6,in; minps xmm0,in; mulps xmm1,FN1; movaps xmm2,F1; maxps xmm1,xmm0; movaps xmm3,angle; subps xmm2,xmm3; mulps xmm1,xmm2; mulps xmm3,in; addps xmm1,xmm3;//(clipping dac) //minps xmm1,F0P1; //maxps xmm1,F0P2; movaps out,xmm1;//(clipped)

//float maxn,minn,sub,clip; maxn=min(in,thresh); minn=max(maxn,(thresh*-1)); sub=(minn*(1-angle))+(angle*in); clip=min(sub,0.999984741); out=max(clip,-0.999984741); A bit of re-arranging; preloading all variables can be counter-productive sometimes. I also shrunk slightly the max range for the Angle knob to simulate a min/max instead of clipping it at the end of the ASM. I guess one can do that in the Threshhold knob also.

I noticed some DC offset? I guess it's a "tube" emulator. 😉

sunsynth · 5 Dec 2009

Thanks MegaHurtz!

I like ur Valve type 'Clipper' 🙂

MyCo · 5 Dec 2009

for this line:

mulps xmm1,FN1;

you could use this:

int mask=-2147483648; orps xmm1,mask;

because thresh is always positive.

MegaHurtz · 5 Dec 2009

Not exactly shure what this is supposed to be.. Risky naming these things 🙂 Im having trouble measuring this aswell, especcially with the added changes, here it runs at -3/2 cycles. So atleast it`s hardly noticed. Didn`t have a hard look at the code yet, but I fully agree with the clipper removed. Thanks for the update, much appreciated!

infuzion · 6 Dec 2009

MegaHurtz wrote: Not exactly sure what this is supposed to be.. Risky naming these things

My first impression: "Linear Waveshaper".

Thanks MyCo for the reminder!

MegaHurtz · 6 Dec 2009

would be nice to do a soft satturating version one of these days.

btw.. why or "-2147483648" ?

nix · 7 Dec 2009

ASM is so hard to understand, and where does that number come from MyCo 🙂 ?

stw · 7 Dec 2009

nix wrote: ASM is so hard to understand, and where does that number come from MyCo 🙂 ?

2147483648= 2^31 and one bit for the sign. Makes a 32 bit integer with all bits set. The same number without the sign bit is used for he "absmask" IIRC.

EDIT: oops, obviously it's the other way round (see MyCo post) 😊 . It actaually wouldn't make any sense the way i said. 😀

MyCo · 7 Dec 2009

The bitmask is a 1 followed by 31 zeros. So an orps with this always sets the most significand bit in a value which for float is the sign bit.

BTW: you can use a more readable version for the same mask effect:

float mask= -0;

yeah you see it right, it's a negative zero....

sunsynth · 7 Dec 2009

ASM brains must be... different 😉

...no offense - dudes

MegaHurtz · 8 Dec 2009

so.. the msb portion of a float is the sign bit.

infuzion · 8 Dec 2009

sunsynth wrote: ASM brains must be... different 😉

ASM is just like playing Legos; you have to figure out how they snap together & in what order.

I'm not sure if the "int mask=-2147483648;orps xmm1,mask;" works for this particular app though.

MegaHurtz · 12 Dec 2009

There is one thing bugging me though, To me, it seems that xmm0 tm xmm7 can be thought of as paralel. While processing these pairs serially. But howcome using it in it`s most paralel form, isn`t beneficial to cycles at all. Is this becouse it`s still a serial method of processing ??

Dell · 19 Dec 2009

Thank you for the "Alternative Clipper."

It sounds really nice. 🙂

infuzion · 19 Dec 2009

MegaHurtz wrote: To me, it seems that xmm0 t0 xmm7 can be thought of as parallel. While processing these pairs serially. But how come using it in it`s most parallel form, isn`t beneficial to cycles at all. Is this because it`s still a serial method of processing ??

xmm0-7 has nothing to do with parallel processing really, it is just several memory addresses that are easier to get to. For the most part, ASM is sequential, but most modern CPUs can be thought of having 2-3 pipelines that may be able to process at the same time if their destinations are different & there is no dependencies on the 2.

MegaHurtz · 20 Dec 2009

tx.. So a quadcore might do a little of that for you, but probbably not..?

infuzion · 21 Dec 2009

MegaHurtz wrote: tx.. So a quadcore might do a little of that for you, but probbably not..?

Intel Core2 self-optimize well, but I can always do better by hand. 😉