JPEG XL

They also say that JXL doesn't provide enough of a difference from other formats, is slower, and doesn't have much interest from the community.

Regardless of if it's targeted or ad-related, they should not be putting that much influence on their ads over content. I remember a number of sites even back in 2019 would quickly load content if, for instance, their ad server went down, it would remove the ad and play the content. Now, most sites will just sit with an non-working ad. Some would continue running the countdown even if the ads stopped playing and went back to the content when the timer hit zero. Now it sits and does nothing - potentially indefinitely - when the ad fails. This is horrible user experience and goes completely against when their web development and the IAB have promoted over the years as focusing on experience and not putting ads in charge of it.

yeah that's the point

Requirements change and when they decide that the service requires making money then 🤷. People feel entitled just because it used to be "free." But really it never has been for normal users. It's already been a while since the current change and the same thing has already passed for twitch, which already has a more aggressive ad system.

you guys still have ads before your videos ? ublock origin is your savior 😄

i never got the ads

but i got the popup and decided i'd rather not participate in the battle

That's what drives people away from platforms (and sometimes to piracy). I already watch a fraction of YT I used to watch years back. Both because the content quality is declining and the ads are increasing. I remember just a few years ago they were promoting quality over quantity. Now they have the TikTok-like shorts and all that is out the window. It's a cycle that has nothing to do with entitlement and free-vs-paid. Even if I paid for ad-free YT I would be watching less and less of it. There's no entitlement. Public pressure is what pushes companies to stop doing anti-competitive and anti-consumer changes. It's public pressure that will get JXL added into browsers.

or you just grew out of it

I never though YT was going to get into the cat-and-mouse game. ublock will eventually be figured out as well.

Or you're just making excuses, w

well, you are also making excuses

youtube is just a platform

and they just happen to need to increase their tax

hmm that's a good analogy

you dont *need* to pay taxes

you don't want to pay taxes or you're now a criminal? just go to another country

That's likely what is going to happen in the long run - people will start using another service. Had they been a smaller service they likely would be losing a large % of viewers already. It's likely smallish at the moment but will likely continue to grow.

so why doesn't everyone just move to new zealand

But it doesn't mean people shouldn't try to push them in a better direction first.

or uhh

oregon

A lot of people already move to Texas and Florida for that very reason. Not Oregon because... it's Oregon.

New Zealand because it's difficult to get citizenship in other countries and New Zealand is fairly small and would get overrun pretty quickly.

one French presidential candidate wants France to tax its expats, the same way that the US and Eritrea do it

but Texas has rather high taxes...

But no income tax.

Yes. So if you're rich, you pay less overall tax in Texas than in say, California. But if you're a normal person, the state of Texas gets more of your money than California would.

A lot of people from NY and CA are moving to TX. The biggest complaints I've seen are incompetent gov employees. So either people are not aware that they are paying more and/or are willing to pay a bit more for gov that is not constantly getting things wrong and harassing them about it. Both NY and CA are also in the news alot for some pretty bad gov actions and just trying to gloss over a lot of issues. Granted, as more people move to TX and with their massive infrastructure issues, it's likely a number of people will be fleeing from there as well over the coming decade.

<https://lwn.net/SubscriberLink/952034/9d9742e4a537ff5b/> Linus Torvalds: > What is really needed, he said, is to find ways to get away from the email patch model, which is not really working anymore. He feels that way now, even though he is "an old-school email person". I can't believe it

FFmpeg has been going on and off for years about this exact issue

since it still uses conventional ML patches

but nobody wants to move to gitlab because of "a metric fuckton of javascript" or the like

Some parts of the kernel have already moved to GitLab, funnily enough

mostly display and GPU stuff

GitLab is absolutely terrible. It feels like they're changing UI at least once a year, and it's crap. Absolutely crap.

sourcehut is another option

Im not the biggest fan of sourcehut myself, I find it... clunky?

i like github personally

there is a lot I dont like about github, but the pros far outweigh the cons imo

What are you talking about?

It's great

I like GitHub's aesthetics

But GitLab has better functionality

at least in the ways that were relevant for me in the past

* change UI yearly making it nice and new each time * PRs and Issues and releases are are behind submenus, needs project to pin them * code search not worth half a damn * some repos need you to log in to search * issue search itself uses stupid tag thing which makes minor corrections a pain * zero discovery tools * pipelines are a pain to read into * issue/PR notifications behind a hamburger menu * can no longer collapse right side bar on PR overview (but you can on issue tickets)

and this is off the top of my head,

im also not sure what functionaility gitlab has that github doesn't myself, nothing ive noticed, well aside from self hosting

well, GitHub for example simply failed at importing an external project... don't remember the other reasons why I preferred GitLab, had something to do with the editor and stuff I think, but among other things... was long ago (didn't notice an UI change over two years though)

Merge Requests and Issues are pinned by default though

almost all projects i've seen have only started to do this like within a week or two now, so I don't think they were

mesa, qemu and fdroid being three projects I very often frequent

yes, might not have been default behaviour for more than a a few months...

the main thing for me I wish gitlab had was discoverability, but in terms of large project management like something like ffmpeg, I dunno. beyond that gerrit is an option, kinda sucks for community contrib tho

Guys, does anybody have an experience with xtensor lib?

TIL iOS contains a fairly up to date copy of Dear ImGui

why do they have fsr

i have no idea

I just saw "Copyright Advanced Micro Devices (FX Super Resolution)" in the open source licenses list.

I would have had no idea with that poorly formatted issue had they opened

doesn't matter who that person is, what you said is still true 😛

in particular, if it were up to me, I'd ignore anything reported by anybody saying that they compiled with -Ofast

for the most part, yes. code that breaks rules like the strict aliasing rule can get mangled by optimizations because the compiler starts making assumptions about the code that may not be true

-O3 should work fine.. back in the day it didn't actually help speed, but perhaps nowadays it does, who knows

if the code is fully correct though, -O3 promises that it will not break it (bugs notwithstanding)

but -Ofast makes no such promise

the problem with -O3 is that it aggressively unrolls loops which means that it increases code size

which can possibly have a negative impact on performance

it can also be undesirable for other reasons, of course 😄

another potential issue is that you gain a lot less from -O3's tree-vectorize when you already are using simd

since autovectorization doesn't really help if you're manually vectorizing anyway

because libjxl relies on highway to do a lot of the heavy lifting I don't think -O3 is actually necessary for it

yup, although apparently the first post mentions a 9 to 14% speedup?

perhaps that's more true in modular mode

for pure C code that does numbercrunching I'd believe that

but for code with simd or highway simd I'd imagine it's a lot less

hydrium is noticeably faster with O3 vs O0 for example

I think those numbers were on libjxl itself, at least the gcc bugzilla seems to say that

(also, O0 is super slow :P)

¯\_(ツ)_/¯

even -O0 can explode the universe if you've got UB in ya code

At that point isn't your code just broken?

The same is true of code that explodes at -O3 (assuming O3 is not so buggy these days)

rust uses O3 by default for release builds so its probably fine in llvm/clang at least

PCSX2 dumps Wayland support in https://github.com/PCSX2/pcsx2/pull/10179 (merged), one of the reasons being the lack of an absolute-coordinate window positioning protocol.

not a surprise, sure, X is broken by design, but wayland is designed to be broken, i've given up on wayland, target xwayland and call it good

Also see: the earlier debate in the forum on the PCSX2 thread.

im not dipping my toes in this one, I get enough notifs from the pcsx2 thread lmao

I don't think Wayland is ready to replace X for everyone yet, but the reality is that the Linux desktop has always been really deficient and Wayland just has a different set of deficiencies. And no one works on foundational X stuff for Linux anymore, Wayland will be the only functional choice.

'course, there's always the option to run an entire X11 DE under XWayland.

wayland is terrible, x11 is terrible linux users in shambles

X11 is bad, but most of the terribleness is in Xorg

shoulda just re-wrote X in rust xD

You know, someone in the comment section of that Brodie Robertson video talking about PCSX2 disabling Wayland by default said that someone he knew said that X11 would go down because it got designed by a committee (likely false), and Wayland's going down the exact same path (also not too sure about that).

I wonder if the chance wlroots becoming the only game in town would also mean the chance of it becoming the new X11.

I don't see GNOME and KDE's Wayland implementations going away.

And it'd be the new Xorg, not the new X11

X11 has a few server implementations, but Xorg is the only non-nested one anyone uses on Linux.

I mean being the new Xorg would mean return of the mess plaguing the current one.

what even is Xwayland?

Is it an X11 server and wayland compositor in one?

it's an X server and a wayland _client_

You can even run a full X session under Xwayland, e.g. a full KDE Plasma desktop or something.

that's in the "rootful" mode right?

yeah

It's a pretty decent option if you need X11 but are on hardware that Xorg has poor support for (e.g. Apple laptops)

so it's a server implementation of X11 that exists as a wayland application that other wayland applications sit next to?

yeah

look ma, I'm running an X application inside Xwayland inside a Wayland compositor running nested under Xorg

does Wayland mean the end of being able to run graphical Linux apps on Windows?

with X11, one could use an X11 server for Windows (such as https://sourceforge.net/projects/vcxsrv/) and point WSL apps to that

WSL has built in support for Wayland IIRC

yeah https://github.com/microsoft/wslg

Enabled automatically for new WSL2 installations

ah, I’m still on WSL1

last I checked, they still hadn’t fixed the interference between WSL2 and VirtualBox

still, that’s good to know, thanks

this is how termux works

nope, wayland actually made this marginally easier

virtualbox "fixed" this by disabling hardware acceleration, unlike every other VMM which can be accelerated using windows WHPX. I reccomend using vmware or qemu typically. Vmware workstation is a lot more preformance then vbox, but it is "payed software" (IE. they don't care when people share personal use licences)

there is a free version of VMware and the only real useful feature I can think of that it's missing compared to it's paid counterpart is snapshots

oh and also I think the free version is also limtied to only one VM running at the same time

networking stuff from workstation is significantly better also workstation has a marginally better graphics emulation

challenge.

What is this

solve it

I don't even understand the question/challenge

maybe <@794205442175402004>

uh no sorry, too cryptic for me

😪

Well, I describe the task that, it's ready-encoded data. The binary data type is converted to decimal digits, so that less space is occupied. 8x8 indicates the size of the image. The picture with arrows indicates the rules of movement in the image. The Huffman tree contains drawn blocks that need to be written to the image using binary data that needs to be translated from decimal

although of course it was possible to think a little

👉🌳

Too difficult for my brain. Fab will solve it.

<@416586441058025472> go fun https://discord.com/channels/794206087879852103/806898911091753051/1179420337880965180

88.116 e7 epf 2 p gaborish 0

I have already dropped this task in a telegram to a group of 1,500 people, and no one understood a damn thing... in short, I just came up with a block algorithm for compressing very small 1-bit images (max 128 bytes.) it seems to me that no one has done this yet. perhaps when I write an encoding implementation at least in python (I hope i will do this in December) and equate it with current formats like png gif webp jxl

assuming the decimal numbers represent 8 bits each

processing-wise it's probably going to be annoying if you have variable shapes per symbol — it might be easier to use fixed 2x2 blocks or something, but then it basically becomes a 4-bit image at 1:2 resolution

here is a translated text that I wrote to my friend regarding what I want to do: Wowa DZɢαs: I need an input matrix of values ranging from 1 to 256 in two planes (1 bit image) Make a complete search of all the options for the location of the paralepipeds with coordinates from 1 to 5 (That is, 16 options) While doing the counting for the same - the total occupied space. Creating a binary huffmon tree based on the frequency of occurrence By a complete search to find the variant that after the transformation (from the structural arrangement of the paralepipeds, creating a tree and immediately encoding into it) would take up the least bit As a result, getting information about which paralepipeds were used, their number, the size of the tree, the number of links in the graph, and the encoded data — this would all make up the total amount of information for choosing the best option Aleksander: it will be a long time, it seems to me Wowa DZɢαs: It depends on the size of the image, but in general, you can use this tactic: fill in the image once, then remove the most unpopular paralepiped, for example, somewhere a single block of 5 by 5, and make the next filling

exactly what I did on a piece of paper. and by the way, yes, I didn't notice initially, but it was possible to compress not into 5 bits but into 4 bits, using a 2 by 3 block, it was possible to remove an entire tree branch

for images this small, signalling the huffman tree itself (including the interpretation of each code) is probably more expensive than the image data

that's it. What I'm thinking about. if use blocks of a maximum of 5 by 5, then can use 4 bits to transmit each, instead of building a tree, can transmit the blocks themselves in the order of the tree. The size of the image I'm still thinking. either 256 or 64 pixels maximum, but this is at least 2 bytes at the beginning, because need to transfer the number of blocks

I think I'll figure something out. 🥴👌

have you checked out the JBIG2 encoding schemes? If you're limiting yourself to 1-bit images, there are some nice tricks that can be used...

all I know is that all of them are relatively outdated, and that all, without exception, were created for much larger images than 8x8 (I remember how I tried to compile it here, and then I was very disappointed in the result)

https://discord.com/channels/794206087879852103/794206087879852106/1082432983832731678

I used a good image then.

optipng -- 286 bytes jxl -e 10 -- 206 bytes 1 bit bmp paq8px -- 158 bytes jbig2 -- 279 bytes

png - 209 bytes

👌

jxl -e 10, converted from <@1034873369314730065> png - 196 bytes

New <:JXL:805850130203934781> <:Stonks:806137886726553651>

damn, losless webp is very good on this one - 190 bytes file, the encoder reports: `Header size: 45 bytes, image data size: 119`

It still has windows on it too. The testing was done from an arch live CD image.

got access to the manga109 dataset and yum, crunchy jpeg artifacts

when you're good enough to make sure your screentone scans arent full of moire but cant be bothered to set the quality slider higher than 5

https://github.com/victorvde/jpeg2png https://github.com/ilyakurdyukov/jpeg-quantsmooth

how do these compare to djpegli?

gimp native, djpegli, jpeg2png (500 iter), jpegqs

looking at the whole image (wont post bc copyright), jpegqs reduces less artifacts than jpeg2png and leaves the lines a lot harsher / more bitonal. djpegli reduces artifacts a little but seems to thin the lines a lot, maybe its getting interpreted as the wrong transfer curve somewhere? 1s

ah yeah that was it. pfm output

jpegli. no artifact reduction

looks like what jpeg2png does when trying to find an output that encodes to the same dct as the input, is optimize for a weighted combination of total generalized variation and mse

according to the dev its specifically for non photographic jpegs which is what im dealing with here

sadly no float output :(

thanks, looks like ill go with jpeg2png for now

is something like this likely to be added to djpegli at some point?

i don't think it's needed for decoder, more accurate rounding and light smoothing is ok (like in djpegli), but strong filtering is dangerous, it's not always needed and not good for all images and can smooth also important details, so such separate tools are better to use for that

there’s a previous project that did something somewhat along those lines: https://github.com/google/knusperli

but AFAIK, it doesn’t go as far as jpeg2png and jpegqs seem to

yeah

reading That One Upscaling article and. Caught this detail here

the dataset doesn't have just 109 manga covers, it has 109 manga volumes unless, of course, you're using the only easily findable torrent of the dataset which contains low quality pngified versions of just the covers

perhaps meaning that only manga covers are used from this dataset

"[t]he entire manga109 dataset will be used"

i think he just grabbed the torrent and thought that was the whole dataset

then maybe, although I don't think non-color images would be needed at all for such comparisons

when comparing traditional upsampling its pretty common to work on the greyscale version of a image afaik, since you're applying the same operation to all channels

seen many papers with a line basically saying "yeah this'll probably generalize to multi-channel images fine, we didn't check"

yeah, but manga is mostly b&w line art and textured backgrounds

yup

for more real-life stuff i'm currently using RealSRv3

its the real world super resolution dataset with the least alignment issues ive found

idk I read that as "this datasaet includes these manga covers" without a statement that it is limited to those

Not sure how many others noticed, but it seems Youtube keeps changing scaling methods depending on your scroll position, being blurry when still

that's a chrome thing

might be to do with compositing and gpu overlay

https://blog.beeper.com/p/introducing-beeper-mini-get-blue

Real client-side, end-to-end encrypted iMessage for Android 🥳

hopefully this doesn't turn out to be a privacy nightmare like Sunbird/Nothing Chats

it really does sound too good to be true

It might be short lived with Apple RCS in 2024

who cares about blue bubbles

the US, apparently

here we just settle for the privacy nightmare that is Facebook Messenger

No, I mean

why doew it matter if your bubbles are blue

Bubble color is a proxy for iMessage features

And a proxy for whether you can afford an iPhone, if you hang out what those kinds of people :)

but like, nobody cares if they text u and it's green

Ive never seen someone scoff at that

I've honestly never heard of the bubble color or seen it up until all the fuss about it being blue for android now

it apparently breaks group chats (reverts to mass SMS text), is less secure (no E2EE), less features (read receipts, typing indicators, reactions) and there's apparently some social stigma associated with that can't comment on the last one though since i personally haven't experienced anything like that

>some social stigma

nobody actually cares

anyone has nightly libjxl build links? I need jpegli for windows.

Thanks

https://www.ddinstagram.com/reel/C0Mhy9SRTcC/?igshid=ODhhZWM5NmIwOQ==

https://youtu.be/fg4t_XxJ8Bw entrancing

a phone charger has more power than the apollo guidance computer

<@1028567873007927297> `f{read,write,open}` are what is in standard C. They give you FILE pointers, and the standard says that reading and writing FILE pointers that have text streams does stuff like newline conversion.

Windows has HANDLEs, which are similar to File Descriptors, but of course, those are only used in Windows APIs, not in standard C ones.

(just like how FDs are not used for standard C functions, only POSIX functions)

near context for the above messages: https://discord.com/channels/794206087879852103/804324493420920833/1183741909773062164

https://www.phoronix.com/news/Linux-6.6.6-Released

<:monkaMega:809252622900789269>

the seal is broken..... it begins.... the year of linux on desktop....

turns out putting 20k image pairs through butteraugli takes. a while

TIL: `yt-dlp` can download "1080p Premium" videos, even without YouTube Premium. This has probably been here for a while; I've just never had to download YouTube videos in a while.

that's a much higher bitrate wow

well, it's just an estimate, the actual file size was only around 300 MB

do wonder why it's literally twice the actual filesize tho

it's probably estimated from only the first chunk or something

patched by Apple

it is back again

This sounds like a horrible game of cat & mouse that I would not want to pay to play

yeah. In my country no one uses apple messages so I don't care. I installed beeper to see what it looks like.

A funny observation. I started making an archive, which was 5 gigabytes, on a fat32 flash drive. after 10 minutes, 7 zip returned an error....... but WinRar, after reaching the size limit, just started making second part.rar Small thing.

yeah, WinRAR tends to handle most situations wayy better then 7-Zip

*than 😄

Seems like Thorium is under fire currently... https://www.reddit.com/r/browsers/comments/18ihls8/dont_use_thorium_browser_if_you_have_it_installed/

🙄

It is just a photo of circumcision and people call it porn. 😒

thats a remarkably stupid point lmao

age is irrelevant nowadays, even a barely tech-savy toddler can handle build scripts and copypastas

the issue isnt the inclusion of that stuff, it's the owner's attitude bleeding into the project.

was that inclusion well-known to the project's users? otherwise it would be indeed problematic.

for example, many FOSS projects (and others) do advertise many things besides from political stuffs, and while many don't care others do care about them, either positively or negatively

but those advertisements are almost always in public, because otherwise why would you advertise them

in comparison a sneaky inclusion is all about the trust, especially for binary releases where users have virtually no way to verify them

regardless of your stance on these political issues, binary browser releases really are not the place for this

it's not an issue of whether you think it's "actually porn" or not, that "yiff" image should never have been shipped out

I wouldn't care if they did it, if it was an actual easter egg thing, not accsessible from a standard page

then it shouldn't be in binary releases

if it's not accessible it's just bad form

if it is accessible it's worse form

it is accessible, i'm pretty sure this was mentioned in this server before

I know it's accessible, I'm pointing out that making it inaccessible is still bad form

ah, okay

yeah it's bloat

It was basically a backup of http://sexuallymutilatedchild.org/ (not available in its old state anymore), including all assets, inside a Thorium-related repository.

Actually, still there.

The "yiff" furry porn pic was actually distributed with the browser.

terribly bloated PNG

should have used JXL

ugh

… I stand corrected, it seems you really are supposed to answer logically to the question

“Yes” leads to ^

where indeed both options are available

could have just shown that window to begin with

https://zamundaaa.github.io/wayland/2023/12/18/update-on-hdr-and-colormanagement-in-plasma.html

color management and HDR coming to Plamsa soon

> With that, Steam starts in a nested gamescope instance and games started in it have HDR working, as long as they use Proton 8 or newer. When this was initially implemented as a bunch of hacks at XDC this year there were issues with a few games, but right now, almost all the HDR capable games in my own Steam library work fine and look good in HDR. dang, soon not just the Steam Deck OLED, but any Linux desktop will be able to totally leapfrog Windows in terms of HDR support

Ive had two users tell me that MPV will trigger it too

which is massive since before that you need to use amdvlk to get HDR on linux

nice you have the choice between yes and no, if I had to implement this dialog I would only put the [ Yes ] button

https://youtu.be/sm-h0kVtGxE?t=1m29s

a master at work

(later: https://youtu.be/1vqHA5DbXBQ?t=5m30s )

anecdote

Sony patent https://patents.google.com/patent/US8246454?oq=8246454

https://tenor.com/view/wheeze-wheeze-laugh-nikkal-kundhal-nikkalkundhal-gif-23223780

**J**oyeu**X** noë**L**!

(why couldn’t I have thought of this 12-24h ago)

Jolly XLmas!

when will a rapper name themself after JXL? It took like 25 years for JPEG to achieve that feat.

I expect that JPEG 2000 probably never will

it’s such a ’90s name

1990s, that is

(won’t fare better in the 2090s)

not counting https://en.wikipedia.org/wiki/Tom_Holkenborg right?

metal subgenres, where you find very normal names like "progressive deathgrind"

For Windows, this is good Custom Resolution Utility (CRU). If overkill, system Display Preferences will show a toggle next to HDR if supported (and a message if not).

Happy, lean and artifact-free new year to all! https://www.theguardian.com/artanddesign/gallery/2023/dec/30/acrobatic-cats-inspired-by-salvador-dali-in-pictures

Happy new year

Right -> Jpegxl viewer app. It is very dark.

Is it https://github.com/oupson/jxlviewer on the right ?

If it is the case, make sure that you are using the latest release (0.4.0).

Thanks. I updated it and now it works well.

For izzyondroid the process is automatic, it will get the new version in the next few days

https://youtu.be/2GRnrFfURYo

looks like a peak detection issue

https://vxtwitter.com/senokay/status/1744220752805765452

uuh

https://flic.kr/p/8Wrghy

<@416586441058025472> benchmark jxl with this

Alert loud audio

I reached 800000hz

Now I reached 1800000hz

https://tenor.com/view/discord-discord-devs-discord-developers-discord-api-discord-employees-gif-25780576

Media unable to load INCIDENT REPORT FOR DISCORD Investigating We are currently investigating an issue that is preventing embedded media from loading. Posted 17 minutes ago. Jan 13, 2024 - 06:18 PST This incident affects: Media Proxy.

🤔

Always gotta love when 'compress' options end up bigger than the original

Probably they used av1/avif lossless.

😆

wowza SVE has a lot of fancy instructions

AVX has like, 4 instructions total that the docs state do math on complex numbers

meanwhile SVE:

even MVE (Helium), the vector extensions for ARM microcontrollers, have a bunch of complex number instructions, lol.

guess I should switch hydrium to float

only downside is that storing a full Group in RAM costs 256x256x6 bytes which is about 400k

and changing that from 16-bit ints to 32-bit floats changes it to about 800k

which is a lot more when the library only uses in the order of 1500k total

It would be nice if CPUs and GPUs supported something like an 3x 21-bit float type natively, i.e. fitting a full pixel in one 64-bit register and with a little more precision than 16-bit int or half-float. Say 1 sign bit, 5 exp bits and 15 mantissa bits, so it can represent int16, uint16 and float16 exactly but it has that little wiggle room of extra precision that could be useful for intermediate results. Plus one bonus bit per pixel that can be used for something.

Why?

gotta go fast

👀

it's of course rather specific to "stuff that works with vectors of 3 numbers", for which I know no other examples than images/video, so you could say it doesn't really belong in a general-purpose instruction set

still, if there's a wishlist somewhere of "new instructions that would be cool", something like this might be a good candidate. When 16-bit is just a little too tight but 32-bit is a bit wasteful, having something in between is nice — and not having to deal with alignment of multiple-of-3 stuff is also nice.

maybe 3x int21 instead of 3x float21 would also be useful

I think libopus has both floating point and fixed point, so you can Go Fast on typical hardware without suffering soft FP on embedded stuff.

it should be faster if I'm not using hand-written assembly, simd, or intrinsics

ah

since I'm not using manual instructions, a lot of my arithmetic is actually 32-bit

which means 32-bit float is likely faster

ic

400k is probably worth the memory overhead though

I store an entire Group in RAM

which is ~800k as 32-bit and ~400k as 16-bit

Maybe I should write a PR contributing some super not portable fixed point SIMD stuff.

(actually, I wonder how hard it would be to find appropriate portable code patterns that reliably get autovectorized.)

I won't accept it

I'm intentionally not including that stuff not because I don't know how to write it (which I don't) but because I don't want it in there

source-level portability is an important design goal

basically I want it to do things that libjxl doesn't, which is be very portable and have a low-memory-footprint

things that libjxl does well, such as coding efficiency and use of simd, my goal is not to try to beat it at those

mm

https://www.macrumors.com/2024/01/15/app-store-to-be-split-in-two/

maybe Kakao etc. are actually threatening both but one is abiding while one is ignoring and the legal action is too slow to observe

if it's the case, it is really sad IMO, maybe the publisher should let abiding ones to survive but require ads for affected comic sources (which would then benefit publishers)

https://fxtwitter.com/kakaoent_pcok/status/1744242383376781461

they did the funny

wtf

I mean, it is clear that such apps do violate the copyright

that doesn't still make such allegation true

should just block SK

the correct statement would be that advertising on their app directly benefits creators while those free apps don't and can't

which is technically and ethically correct and can even pave a way to coexist without such issues

(as I've said above...)

technically speaking it can contain malware if you can't build one yourself though :p

but otherwise it's a false claim...

I'm more amused by its strange use of hashtags: `#PCOK #Protecting #the #Content #of #Kakao...`

`#the`

anyway, I'm aware of reasons that publishers say such things and I generally *want* to support them... but their apps are crappy at best

not even mentioning the DRM

so is it a self-closing emphasis tag? XD

I have too many ebooks in my ridibooks account, which is one of the largest SK ebook services and in my opinion better than all other SK alternatives, but its desktop app is still crappy

I once analyzed their app and found that it was doing all gymnastics to render a single page of JPEG in whooping 2 seconds...

seriously, JPEG rendering at that size should never take more than 100ms

it's not even like the file has a dimension of 3000 by 4000 pixels (so to say), I think it was about ~1200 px vertically

yeah it's good to have APIs for such cases

I can't see any way for ebook service or book publisher to release some sort of API without risking a copyright violation though

(which always happens, by the way, but in much easier ways I meant)

more amusingly, the content decryption was not even the biggest factor (thought it was significant, thanks to the fact that everything was written in JS!)

for some reason it was keeping its own canvas to render decoded images, wtf?

my guess is that it's an Electron app, as everything today is, and it didn't want any decoded image to stay in the world-readable cache

but I refuse to believe that this was the optimal solution even if that was the intent

(I believe the incognito/private mode can be enabled in such webviews which encrypts any disk cache with an in-memory-only key)

and all this attempt was futile because it never blocked the remote desktop service

it does use a built-in windows mechanism to hide itself from screenshots, but that's all

at least it didn't come with a much stupider kernel-level "security" solution, it's why I keep using that service

not really I think

there are security solutions that do try to do more, often used in games or financial apps, but I'm not sure if it really helps

I've seen dedicated attackers were always able to get around them anyway, so the question is probably the relative effectiveness for simple-minded attackers

and I have no evidence that it generally works, but I don't have any counterevidence as well

(cause I can do a substantial amount of reverse engineering, I don't know much about how "typical" attackers would behave)

https://youtu.be/W4PHhurAhwc?si=7aU28TyzJ4VdWloB

That file manager looks cool. I'm going to try it.

I've since blocked him from all of my repos, since I got twenty six emails a couple of days ago as he was repeatedly commenting on random commits and issues. I recommend others do the same. He also DM'd me asking me not to take anything down for some reason

I wrote my own script to create recovery data in case of accidental byte corruption. I decided to do this because when I moved the 451 GB gl folder from disk to disk, I broke 3 bytes, which I was able to restore, 10 hours before creating an ICE ECC file. The recxorf recovery files contain 1 byte of recovery per 64 kilobytes of data. There is no hierarchy of files, so the recovery information is recorded in their alphabetical order. This means that there should be no other folders in the folder. Renaming files or adding new files after the recovery is created will break everything. The algorithm is based on XOR using a hash to detect corruption and then iterating through all 65536 bytes in the block until the hash function matches. How to use: 0. Install numba: pip install numba 1. Place the RECXORF.py to the file folder 2. Use the terminal to go to the same folder (cd) 3. python RECXORF.py Files "RECXORF.py " and "recxorf" are ignored during processing.

A variant of the script for working with a single file, the path is set at the bottom of the code.

It uses Reed–Solomon error correction. This is a much more complex and powerful method, which is much better in terms of the effective repair of information. But it's very, very, very slow. That's why I wrote mine. My algorithm, when working on 1 thread, creates recovery blocks with disk write speed.

I guess you rely on cryptographic hashes to detect corruptions and can only correct a single byte per each 64K block, okay? the former is okay, but I'm not sure about the latter.

for example, faulty network transmission may result in a burst error that your scheme would be never able to correct...

incidentally, I'm going to develop a new parchive tool soon

If there is a byte shift, then yes. But I did my own thing precisely because I have to move terabytes between HDD disks, and 1-2 bytes of errors appear there, which are shit

in that case you don't even need a per-block checksum, you can just verify and if needed recover opportunistically.

That would be extremely long. If anything, I checked if also Bruteforce all the byte options until it matches the hash amount, this also works when it's only about 1 byte. but then you have to make a hash larger, 64 or 80 bits, and all this loses its meaning.

what’s your motivation for doing so? will it work with the same format or would you be changing that too?

hm, erasure coding in zfs works at disk write speed unless you have a very fast disk. I wonder what the difference is there?

maybe you should try Merkle trees if that's your concern.

the ones that exist seem very focused on the usenet usecase, while I'm more interested in durable distributed storage. someone pointed out that while that may be true, the format itself may still be good for my usecases, so I probably don't need to make a new format. though making a format with stronger recovery guarantees for common storage failure modes is something I might look at later.

it's not about the speed of the disk, but reed solomon is just slow on any cpu - much slower than the speed of reading data

you still keep hashing 64K blocks, but they are increasingly merged into a single hash covering larger "blocks", and eventually a single hash for the entirety. if that doesn't match, you can traverse down to the actual block(s) that had an error in log time.

I was thinking of using this when I thought that there might be a problem when 2 bytes are broken in 1 block, but at the same time there are no broken bytes(or only one broken per block) in the largest 128 blocks. Then it would be possible to make 2 bytes of XOR for this block chain - but I just gave up on it and didn't complicate anything.

I mean, you would keep all the intermediate hashes anyway, wouldn't you?

I think I don't care anymore and I don't want to come up with something more complicated.

I suppose reed-solomon in ZFS is quite fast because it only does N+{1,2,3} parity. (where N is the number of data pieces)

while schemes like parchive usually uses hundreds to tens of thousands of pieces total between data and parity.

> In the early 20th century, an attempt was made to introduce an artificial distinction between *gray* and *grey*, with the former being used for a "mixture of white and blue", but the latter being used for a "mixture made by white and black"[1]; this has not been generally adopted.

I’m glad it hasn’t

in french we should have used gris and grys ?

not clear how to make the distinction orally either

https://arstechnica.com/gadgets/2024/01/inventor-of-ntp-protocol-that-keeps-time-on-billions-of-devices-dies-at-age-85/

https://www.reddit.com/r/TeenagersITA/s/jGqV5nAkXv

I helped re a bit

Where is the gui for this?

https://pypi.org/project/img2pdf/

if you had to guess, how does this user interface work?

Mousing over it highlights the box

what do you think this UI widget is

it's a toggle knob switch. clicking it toggles it to yes

UI design is my passion

please tell me that it shows a half of both "No" and "Yes" in the middle

no animation, it just changes instantaneously when you click it

https://www.forbes.com/sites/barrycollins/2024/01/22/meet-avif-the-sharp-new-image-format-thats-set-to-replace-jpeg/

> When images in both formats were shrunk down to around 80,000 bytes (0.08MB) in size, the AVIF image still looked very close to the high-quality original

“very close”, sure

(and why write it 0.08MB instead of 80kB?)

People used to see only "plastic" compressed to full texture loss videos from phones 🤦‍♂️

forbes isn't exactly a great source of good technology info

for example, they give you "four articles a month" for free by their paywall

if you scroll to the bottom of the page it pops up this obnoxious "this article is paywalled" ad, and you can't scroll back up to actually read the supposed free article of four

reloading the page uses another one of your "four free articles"

Also isn't that section of Forbes basically random blogs

And after all that, they don't post a single image for it

They're referring to images in that Netflix blog post, no?

yes, and I had a look again, to see if I remembered it properly

I guess it depends on your definition of “very close”

for me, it was only true of the images that were low-quality to begin with

Well that's neat, guess I got water in my phone line from the storms in England lately. When I ring the home phone I get data coming out and eventual disconnect of the broadband

https://www.apple.com/newsroom/2024/01/apple-announces-changes-to-ios-safari-and-the-app-store-in-the-european-union/

> For users, the changes include new controls and disclosures, and expanded protections to reduce privacy and security risks the DMA creates

what security risks does the DMA create?

> This change is a result of the DMA’s requirements, and means that EU users will be confronted with a list of default browsers before they have the opportunity to understand the options available to them. The screen also interrupts EU users’ experience the first time they open Safari intending to navigate to a webpage.

oh man, their PR team is really salty

I can't wait for them to go with malicious compliance and have the EU courts tell them they can't do that

the entire article goes on about security but they won't actually say what the "security risks" are

https://9to5mac.com/2024/01/24/whatsapp-third-party-chats-eu-imessage/

Now this is an EU fail

how is it an EU fail?

oh the interop problem

yea that is a fail

interop is a problem when the other chats don't have good spam filters and the like

it's why we get garbage from the matrix bridge

interop doesn't mean any random spammer can just connect and send messages to anyone

it's just so pointless

i dont want everything app!!

go away elon

Looks like we might get operator overloading in C2y

Also, according to a TikTok uploaded by the C project editor, defer has been approved by the committee for C2y

https://twitter.com/uwukko/status/1750815285509083630?t=8cYhl3G5S1RHNinn_vZfCQ&s=19

I personally think C++-style operator overloading is a mistake, though I do feel needs for operator overloading in general

the best design in my opinion comes from OCaml: https://v2.ocaml.org/manual/expr.html#sss:local-opens

for example, `Foo.(3 + 4)` is equivalent to `Foo.(+) 3 4` (the infix `a + b` is internally `(+) a b` where `(+)` is a function with a weird name)

so you can locally overload operators without too much complication, but the overloading is clearly indicated in the syntax. best of both worlds.

"the inventor of the .webp file extension" *welp*

`.png.jxl`

at some point we will also see `.pam.jxl`

A friend talked about it yesterday, I guess now I know why

people can't really distinguish lossy webp from lossless webp, sadly

in fact, the general sentiment seems that most even don't realize that webp has a lossless codec at all

so the fact that lossless webp *does* perform as expected doesn't help

and is outweighted by the fact that lossy webp didn't perform as expected

btw what did cause the introduction of lossless webp in the first place?

https://discord.com/channels/794206087879852103/804324493420920833/1082651956436291654

oh, thanks!

Yeah, my friend joked "You're into JXL right? We'll I've lost all hope, the guy who made WebP is on it" So I then had to explain about the lossless mode which still beats JXL occasionally too Followed up with me sending them a WebP on Telegram

I've found an old mozilla bug that discusses those requirements: https://bugzilla.mozilla.org/show_bug.cgi?id=600919#c83

> From Jeff Muizlaar's blog (http://muizelaar.blogspot.com/2011/04/webp.html) we could see that the four features missing in WebP as far as Mozilla is concerned are: > * Color spaces > * Alpha channel > * EXIF (or something similar) > * ICC color profiles

> There appears to be four simple things that Mozilla require to adopt this format: EXIF data, Alpha transparency, lossless support and i8x8 intra compression (AFAICT). That's not a lot to ask for what Google are proposing to become the worlds new standardised image format. Thus it's simply a matter of Google putting their money and resources where there mouth is and in a much more proactive manner than they've done with WebM.

hmm, so there was no fixed set of lists arose from that bug I guess? but the general concensus was made and google devs took a notice...

> This is even more confusing. PSNR has, for a while now, been accepted as a poor measure of visual quality and I can't understand why Google continues to use it > still true in 2024...

great foreshadowing!

I briefly considered a possibility that the decision to add a lossless codec was motivated by comparably poor performance of lossy codec before username posted that link

The fact twitter push people, who would never do this in real life, to openly insult and threaten others, is appalling

> only one person (that i know of) in the world has so far been able to use .webp without the help of format converters. oh no, _format converters_

(as opposed to…?)

Shape-shifters formats ?

Schrodinger's File

probably they mean "as opposed to their software just supporting webp out of the box"

not sure which piece of relevant software doesn't

https://pbs.twimg.com/media/GE1BP6ab0AAc0Wx?format=jpg&name=large

Interesting wallpaper

https://youtu.be/Umkoxvi0QC4

synth pop cover of the classic

https://cdn.discordapp.com/attachments/834553178086440960/1201974010997645382/2024-01-30_21-25-08_edit_VHS.mp4

I recorded on the real VHS: clip -【original anime MV】SHINKIRO【hololive/宝鐘マリン・Gawr Gura】 https://youtu.be/9ehwhQJ50gs