claude @mathr@post.lurk.org

Updated the Gem example with a spinning [cube] with desktop OpenGL 1.x Fixed Function Pipeline lighting emulated in WebGL by Regal.

Other primitives like [sphere] and [torus] fail due to a bug with GL_QUAD_STRIP, have reported it to the Emscripten port.

[teapot] fails for other reasons (incomplete GLU? something else?), seems to be an infinite loop rather than a simple failure...

added meta tags to my blog generator:

https://code.mathr.co.uk/cmcms/commitdiff/a0d4e304679ae51f7e5b18e8355755957bb07b67

https://mathr.co.uk/empd/#build-cleanroom added some scripts for building in a chroot, so you don't risk potential weird interactions with your main dev system

http://mathr.co.uk/empd/gemtest.html victory! #puredata #gem #web #browser #emscripten #regal #opengl

working on cleaning up / committing my changes to the Gem sources now, will push as soon as I've got a reproducible build script

fighting may have got easier with this
https://github.com/emscripten-ports/regal

currently requires emsdk "incoming" tag, afaict

fighting with Emscripten, Regal, GL, GLU, SDL, Gem

made a mini-site for my emscripten pd experiments:
https://mathr.co.uk/empd/

managed to get a slightly modified version of #puredata working in the #web #browser using #emscripten

quick start guide:
https://github.com/claudeha/libpd/blob/emscripten/samples/emscripten/pdtest/README.md

demo:
https://cafe.mathr.co.uk/eiskaffee/empd/pdtest.html

don't look:
https://github.com/claudeha/pure-data/commit/56ec8f1e763b307dc5757380f885282c58adb18e

I uploaded a short video of nnirror in action here (5mins, 60MB, no sound):

https://mathr.co.uk/misc/2019-01-07_nnirror.mp4

not sure about the border colour changing on the training view (it is redundant since I added the graphs), and the multicoloured lines on the other view for showing the learning parameters that led to success or failure are rather unilluminating...

Today I'm mostly working on 'nnirror', my art project about training neural networks to recognize themselves.

The ego network is trained using a generative adversarial network against the id network. Ego aims to recognize its own weights (output 1) vs everything else (output 0 for id's attempts to fool ego, output 0 for random input too).

The network weights are visualized at the top left of the first image, below is the normalized change since the previous epoch.

The second image plots the parameters (learning rates, momementa, etc), on the left if the ego network failed to achieve enlightenment after 1000 epochs, on the right if it managed to score above 4.5 in that time. The total score is twice the top graph minus the two lower graphs.

implemented common subexpression elimination using the uniplate package, not yet integrated into the formula compiler proper, but seems promising in ghci exploration...

I think I'll punt these compiler changes to a future `kf-2.15` branch, the generated code for reference and perturbation iterations is just not good enough yet.

Would be nice though, adding a formula would become 2 lines of code, and the huge formula.cpp generated from the semi-automatically crafted formula.xml via formua.xsl takes a long time to compile, at least the generated code is split into many files so it can be compiled in parallel...

Hopefully will manage to release `kf-2.14.4` today with some less invasive changes.

the stripped exe is 16MB for gcc and 27MB for clang, so I guess gcc wins here too...

it linked and even ran, but no faster than the code compiled by gcc-7

gcc took 20mins to compile it all on my dual core laptop, clang over 10x as long, and gcc output is marginally faster in one simple test.

clang-7 is taking forever to compile my formulas, it better link ok when its done or I'll be well annoyed

gcc-7 did take less time to compile it, though

trying clang-7, took some time to find the right triple: -target x86_64-pc-mingw32

(though I haven't checked yet that it links correctly)

improved my fractal formula compiler to reduce one Burning Ship benchmark from 21.5s to 20.5s \o/

then I upgraded gcc-mingw to version 7 (was 6.3 before) and it slowed down again to 22.5s /o\

quite literally, 1 step forward 2 steps backward

now throwing more -Ofast at it to see if it helps...

I enjoyed this release:

Bebhionn - Interplanetary
https://archive.org/details/Zimmer155
minimal, techno, deep
CC BY-NC-ND

While hacking on a reference implementation to go with my work-in-progress paper, I found several bugs in the paper. Mostly embarrassing sign errors, but some off-by-one too.

I also accidentally fixed one algorithm: arbitrarily setting `sgn(0) = 1` instead of `sgn(0) = 0` makes it succeed in more cases (less risk of a singular matrix).

But now I'm considering deleting 3/4 of a page of vagueness about some advanced techniques and replacing with a fuller treatment of the basics: as it is now, you need to know already how to render escape time fractals for my first page to make any sense. Running very tight to the 8-page hard limit...

I might be able to shrink some of the images to 2/3 page instead of 1 whole page, but whether the page breaks will still be good is another issue.