@gsuberland
Dance music we know, all genres. Pop, indie, rock and metal not so much.

@gsuberland Chopstick Dubplate feat. Top Cat and Mr Williams "Worldwide Traveler"

https://youtu.be/75ciLcHS6EQ

@headfirstonly Gawd. You remember that video the hypergolic mixture that I flimed - I had a horrible cold, I'd almost lost my voice and the mixture almost blew me out of my skin.
That was astonishing. Only later would I read John Clarke's "Ignition" where hypergolics were just the beginning on the insanity… https://archive.org/details/ignition_201612

The recording of the papers session on Simulation now available! https://youtu.be/MjmiCUhXyew
#I3D2025

Back on the search for IEEE papers. Looking to see if this one is any good, although the abstract has a lot of spurious factoids about "99.967% accuracy" alongside the phrase "major breakthrough" (some people seem to believe 108.56MHz is fast for an FPGA, let's see what this one says)

https://ieeexplore.ieee.org/document/11012211

@lritter you're right, can't skip the screenshot

And when I say "raise it to a power", I mean the range of integer, fractional, negative, positive, full floating point powers. A full menu of techniques, each one a polynomial, some bit flips and a single multiply.

The most bananas paper, how to raise to a power using a table, a multiply, and operand modification.

So, you want a reciprocal sqrt? Take your input operand X, split it into two parts X1 and X2, use the X1 to calc this polynomial C', flip some bits in X2, reassemble X1 and X2 into X', multiply C'*X' and voila, 6 or 7 decimals of precision!

It's like alien technology, how the hell is it so good?

https://sci-hub.se/10.1109/12.736432

In what way is this an open source "graphics" GPU when there's nothing in there but integer CUs?

https://arxiv.org/pdf/2505.08421

oh shit! I tried to run a diff but nothing happened!

Permalink: https://wizardzines.com/comics/oh-shit-diff-nothing-happened/

@GDCPresoReviews @breakin @TomF
Yes I was, but if you've ever working inside a large corporation nobody does anything "extra" without paying mind to how it will affect your ongoing career or compensation.

Hobby projects are to escape or to advance, and receiving a "low score talk" meant all that effort wasn't going receive any boost or resources. In fact continuing to ask for help would have been a negative "chasing stupid dreams, the market has spoken".

Maybe I should revisit it.

@breakin @TomF @GDCPresoReviews
When management measures the success of research by the GDC talk score, and allocates resources and time by that feedback, you need pretty pictures more than actual innovation.

@lisyarus @artificialmind @TomF @GDCPresoReviews
Negative values are a side effect of global-ness. Localized basis functions don't pollute far away bases with crud, but all frames and bases will be deviations from a constant or DC term. Avoiding bad values is about locality and speed of convergence.

@artificialmind @lisyarus @TomF @GDCPresoReviews Sampling a signal with one basis and reconstructing it with another (biorthogonal sampling) can guarantee perfect reconstruction if you have control over both bases - Modern sampling theory replaces Shannon's "band limit before sampling" rules.

https://basesandframes.wordpress.com/2016/05/12/50-years-after-shannon/

@artificialmind @lisyarus @TomF @GDCPresoReviews Orthonormal bases do not need to ring - see the Haar basis, no ringing because it's localized, it's just jaggy AF instead. Ringing comes from the global smoothness constraint in SH being followed above all else. SH will converge to the source signal given enough (i.e. infinite) bands of coefficients, it just has no reason to converge *quickly*. Converging "at the limit" is an evil thing to do to a basis you actually want to use.

@lritter You get to choose the number of needlets - 4, 5, 6, 7. You get to trade off storage/compute against fidelity with fine control.

@lritter @TomF https://www.youtube.com/watch?v=ij5mw_eqKuc

@TomF @lritter Rhyming slang, Scooby-Doo = Clue

@lritter Light Propagation Volumes? Not a scoobie, haven't been keeping up with engine tech. I was just trying to solve the "function in all directions as coefficients that don't ring like a bitch" problem.

@TomF @GDCPresoReviews Oh, and Needlets are expensive to sample without weighted importance sampling, you'll miss all the detail if you use regular unbiased spherical sampling.

@GDCPresoReviews Did you get the PDF of the slides?
https://basesandframes.wordpress.com/2016/05/22/implementing-needlets/