07:49 <cyrozap> Hey, this is a little off-topic, but it involves bit-hacks so I figured graphics people might be able to help me. I'm reverse-engineering a thing, and I'm trying to understand the purpose of a function. This function takes a single 6-bit input value, returns a 6-bit output value, and when its inputs and outputs are plotted as unsigned integers, it looks like this: https://i.imgur.com/pV2AhCg.png

07:50 <cyrozap> So, my question is, does anyone recognize this? And if so, does this algorithm have a name? And what is it usally used for?

07:52 <cyrozap> The algorithm is implemented by masking the lower 6 bits of the input, reversing the order of those 6 bits (keeping everything in the lower 6 bits of the byte), adding 1, and then doing the bit reverse again.

07:53 <cyrozap> And then the final value is stored in a hardware register, the purpose of which is currently unknown to me.

07:58 <HdkR> Almost looks like some Nvidia shader bit twiddling

08:05 <cyrozap> To describe it by parts, it's: f(x) = { (x + 32) for x in [0,32); (x - 16) for x in [32,48); (x - 40) for x in [48,56); (x - 52) for x in [56,60); (x - 58) for x in [60,62); 1 if x is 62; 0 if x is 63 }

08:09 <cyrozap> For some more context, the hardware register the output value is getting written into may be part of some sort of data-moving hardware (though that's just a hunch at this point).

08:11 <cyrozap> So maybe this is some kind of block-size thing?

08:16 <cyrozap> The scatter plot vaguely reminds me of a wavelet transform, though I'm pretty sure that's not what this is since it wouldn't make sense for the hardware to have any compression capabilities (its firmware is stored uncompressed on flash, and it's just a USB host controller).

08:25 <cyrozap> This is specifically what it reminded me of: https://commons.wikimedia.org/wiki/File:Jpeg2000_2-level_wavelet_transform-lichtenstein.png

08:27 <cyrozap> Though that's probably because the differences between the starting points of each linear range is a geometric progression (with r = 1/2), and a geometric series can be illustrated like this: https://en.wikipedia.org/wiki/File:GeometricSquares.svg

11:14 <macc24> cyrozap: it sort-of kinda maybe looks familiar for mw

12:54 <alyssa> eyeballs

12:56 <HdkR> They're weird right? How they see things, ever watching.

13:00 <alyssa> cyrozap: It might be easier to think of "reverse, add 1, reverse" as instead "add 32, but carry in the wrong direction"

13:01 <alyssa> It has one very interesting proprtiety

13:02 <alyssa> Can't phrase it precisely but a bit being set in the output means the input is strictly less than something

13:04 <alyssa> so you can count leading zeroes and estimate log2(x)

13:04 <alyssa> although you could do that anyway...

13:04 <alyssa> sorry, estimate 6-log2(x0

13:05 <alyssa> Also, cruially, this map is bijective.

13:06 <alyssa> (invert in the obvious way.. reverse, subtract 1, reverse)

13:08 <alyssa> so tbh it really doesn't matter

13:09 <alyssa> You can calculate it, you can invert it, both are trivial for you... postpone the *why* until you have info on neighboring fields...

13:14 <adjtm_> cyrozap, it reminds me posits with 0 exponent bits

13:15 <adjtm_> the number of consecutive 1s at the beginning codifies a range, so you have more precision at the range it matters most for your application and the precision decreases where your are far from it

13:16 <adjtm_> the output could be a linear representation of it

13:17 <adjtm_> for example for comparing the numbers with an integer ALU

13:17 <adjtm_> I meant posits without sign and 0 exponent bits

13:23 <adjtm_> what alyssa said about using the codification in your favor for estimating functions make sense

13:24 <adjtm_> there are some of that kind, like the famous 1/sqrt(x) of quake (that can be generalize to any fractional exponential, positive or negative(

13:25 <adjtm_> with posit(8,0) there is a fast algorithm for estimating the sigmoid function

17:14 <alyssa> ok.. why is this test still failing..

19:18 <alyssa> bbrezillon: I think we still need to wire in indirect compute dispatch.

19:18 <alyssa> You up for that?

20:21 <alyssa> bbrezillon: es2_info: ../src/gallium/drivers/panfrost/pan_resource.c:120: panfrost_resource_from_handle: Assertion `!drm_is_afbc(whandle->modifier) || !panfrost_afbc_format_needs_fixup(dev, rsc->internal_format)' failed.

20:21 <alyssa> Uh oh ...