iiinova — where your digital self keeps evolving

Reposted by Falvyu

Tom Forsyth @tomforsyth.bsky.social · 21/11/2025

Recent discussion about the perils of doors in gamedev reminded me of a bug caused by a door in a game you may have heard of called "Half Life 2". I wrote it up over on Mastodon (I find it's better at long threads): mastodon.gamedev.place/@TomF/115589...

mastodon.gamedev.place

Tom Forsyth (@TomF@mastodon.gamedev.place)

Attached: 1 image Recent discussion about the perils of doors in gamedev reminded me of a bug caused by a door in a game you may heard of called "Half Life 2". Are you sitting comfortably? Then I sha...

18361115

Reposted by Falvyu

camel-cdr.bsky.social @camel-cdr.bsky.social · 25/09/2025

Tenstorrent decided to publish the first benchmark data for Ascalon's RVV implementation using the instruction throughput benchmark of my rvv-bench benchmark suite. <3 camel-cdr.github.io/rvv-bench-re... Overall, the results look really good so far:

133

Falvyu @falvyu.bsky.social · 02/07/2025

You can probably find this on consumer-grade products too (e.g. smartphones). But the issue is finding how to 'access' these settings.

020

Falvyu @falvyu.bsky.social · 02/07/2025

Yes absolutely. Some astronomy sensors do support it internally: the Sony IMX571, IMX533, and IMX294 (among others) do support binning. But these models are mainly for astrophotography and industrial-grade products.

110

Falvyu @falvyu.bsky.social · 14/06/2025

If a pixel is saturated, then you can't recover the data. That's a big issue in bright areas (e.g stars, Andromeda/Orion nebula cores, ...), and that's why people stack pictures.

010

Falvyu @falvyu.bsky.social · 14/06/2025

It's also the reason why temperature made such a difference (and why people use cooled cameras for astrophotography): it doubles every 3 degrees C. But what really limits sub-exposure time is astrophotography is the histogram (+ mount/atmosphere stability + satellite/plane trails + sky glow).

110

Falvyu @falvyu.bsky.social · 14/06/2025

Lower shutter speed contains more 'dark current' noise, but that's just because dark current is proportional to exposure time, but so is the main signal: The signal-to-noise ratio remains the same. Taking multiple exposures does not help with it (other than making dark flat more convenient).

110

Reposted by Falvyu

InstLatX64 @instlatx64.bsky.social · 17/05/2025

The most underutilized #AVX512 instruction, VPSHUFBITQMB, is actually very useful for compactly generating a mask with arbitrary criteria based on b[5:0], e.g. to select special characters for parsing: if the second operand is a broadcasted 64-bit vector, where the ones represent the searched values

3176

Reposted by Falvyu

Underfox @underfox3.bsky.social · 15/04/2025

Intel researchers have proposed Earth, an efficient architecture for RISC-V vector memory access patterns, enabling coalesced strided instruction memory access and buffer-free segment instruction processing. arxiv.org/pdf/2504.08334

2136

Reposted by Falvyu

Tom Forsyth @tomforsyth.bsky.social · 29/03/2025

Oh shit. I just realised no - it really is. I thought I was joking. This is going to require a long and complex backstory. Sorry absolutely not sorry it'll be funny maybe but also you'll learn something.

15013

Reposted by Falvyu

FCLC (being a silly goose) @fclc.bsky.social · 27/03/2025

Hey Folks! cool little FOSS community tool released for simulating the way @tenstorrent.bsky.social Hardware works! Namely the way the tensix instruction stream operates: github.com/tenstorrent/...

github.com

GitHub - tenstorrent/tensix-isa-simulator

Contribute to tenstorrent/tensix-isa-simulator development by creating an account on GitHub.

091

Reposted by Falvyu

Raph Levien @raphlinus.bsky.social · 21/03/2025

I've now published my blog post, "I want a good parallel computer." raphlinus.github.io/gpu/2025/03/... . Thanks much to all the feedback on the draft, I'd like to think I've clarified some things that might have been confusing.

raphlinus.github.io

I want a good parallel computer

The GPU in your computer is about 10 to 100 times more powerful than the CPU, depending on workload. For real-time graphics rendering and machine learning, you are enjoying that power, and doing those...

26218

Reposted by Falvyu

InstLatX64 @instlatx64.bsky.social · 19/03/2025

Another round of #Intel ISA defragmentation: Finally, #AVX10_2 is a common ISA for all CPU segments!

182

Reposted by Falvyu

NSF–DOE Rubin Observatory @rubinobservatory.org · 12/03/2025

Lights, camera, action! The world's largest digital camera has been installed at NSF–DOE Rubin Observatory! 🤩 The LSST Camera was the final major component of the observatory. With it in place, Rubin officially enters its final phase of testing!🔭 🔗: rubinobservatory.org/news/lsst-camera-installed

521570

Reposted by Falvyu

FCLC (being a silly goose) @fclc.bsky.social · 24/02/2025

"All lessons in tech will be relearned every 25 years" -Me, looking at us relearning the same thing again

2102

Reposted by Falvyu

Björkus "No time_t to Die" Dorkus @thephd.dev · 07/02/2025

I gotta see so many posts saying "Rust is slower" and like I think these people just evaluate languages and designs based on vibes and not based on like the material output from the compilers.

61074

Reposted by Falvyu

Jonathan O'Callaghan @astrojonny.bsky.social · 04/02/2025

The chance of asteroid 2024 YR4 hitting out planet in 2032 is now 1.5%, or 1 in 67. The OVERWHELMING likelihood is that the asteroid will miss Earth. But for the first time ever, we might have to seriously consider a deflection mission soon. Let me explain. 🧵 (1/x)

media.tenor.com

a picture of the earth and an asteroid with the national geographic logo

Alt: a picture of the earth and an asteroid heading towards it

95805325

Falvyu @falvyu.bsky.social · 01/02/2025

I don't know anything about Raku, but these ads/posters are quite cool. #FOSDEM

000

Reposted by Falvyu

FCLC (being a silly goose) @fclc.bsky.social · 01/02/2025

Hey folks! Will be talking about the wonderful, wacky world of floating point Sunday afternoon in the HPC, Big data, & Data science dev room at 15:30! #fosdem

1279

Reposted by Falvyu

Kenneth Hoste (boegel) @boegel.bsky.social · 01/02/2025

“Welcome to the largest open source conference on earth!” #FOSDEM

411210

Reposted by Falvyu

Gaia Mission @gaiamission.bsky.social · 15/01/2025

So long and thanks for all the fish! I just observed my last star this morning... #GaiaDPAC and @esa.int Gaia teams will do their best to make some great data releases from the data I gathered! #GaiaDR4 in 2026, #GaiaDR5 around the end of the decade. Keep posted for our news today at 10:00 CET.

giphy.com

rotating european space agency GIF - Find & Share on GIPHY

Discover & share this rotating european space agency GIF with everyone you know. GIPHY is how you search, share, discover, and create GIFs.

0114

Reposted by Falvyu

NSF–DOE Rubin Observatory @rubinobservatory.org · 14/01/2025

When you turn it on and it works spectacularly! The Rubin team has successfully completed the first on-sky engineering tests using the 144-Mpx engineering camera. Up next: installing the largest digital camera in the world, LSST Camera! 🔭🧪 rubinobservatory.org/news/rubin-completes-comcam-tests

115538

Falvyu @falvyu.bsky.social · 10/01/2025

Has the author ever programmed in anything other than C ?

010

Reposted by Falvyu

八咫烏 @yatagarasu.bsky.social · 01/01/2025

today i learned x264 has a setting specifically for encoding touhou games

61173505

Falvyu @falvyu.bsky.social · 01/01/2025

Will the talks be recorded & available online ? In any case, I wish you good luck.

110

Falvyu @falvyu.bsky.social · 30/12/2024

Oh yep. Good catch. Thanks, it works perfectly.

000

Falvyu @falvyu.bsky.social · 30/12/2024

Also, do you happen to know if someone has compiled a list of bitwise scans/segmented-scans ? I searched for one earlier, but did not find anything. However, I'm likely not the first to come across this problem. And I wouldn't be surprised if they are listed on some personal blog.

300

Falvyu @falvyu.bsky.social · 30/12/2024

Thanks. These work like a charm and are indeed simpler than what I had found.

100

Reposted by Falvyu

efroach76.bsky.social @efroach76.bsky.social · 30/12/2024

As Harold mentions, Prefix-Or is just v | -v. Segment-scan-or can also be simplified, for example uint64_t s = (v | m) >> 1; return s ^ (s - v) (4 instructions instead of 7).

121

Falvyu @falvyu.bsky.social · 29/12/2024

Note #1: in the segmented scan version, `(v[i] == 1) & (mreset[i] == 1)` => `1` (i.e. input has priority over 'reset'). Note #2: Expression can be simplified if 'input' can not happen at the same time as a 'reset'.

010

Falvyu @falvyu.bsky.social · 29/12/2024

Some neat tricks for computing bit-wise prefix-or and segmented-prefix-or within scalar registers.

341

Falvyu @falvyu.bsky.social · 28/12/2024

It turns that the 'unrolled loop' for the 64-bits bitwise-or segmented scan can actually be reduced down to a few instructions. (not sure if that's a known trick) Note: compilers optimize `((~mreset) | v)` so that it is only computed once.

010

Falvyu @falvyu.bsky.social · 27/12/2024

(or just replicate whatever simdjson is doing)

010

Falvyu @falvyu.bsky.social · 27/12/2024

PS: It only handles 8-bits ASCII (no UTF-8) For quotes (master only handles double-quotes): complexity increases quickly if you want to support single-quote & cie. Instead of doing everything at once, a RLE of quoted areas and a scalar post-processing ay be a good option.

110

Reposted by Falvyu

camel-cdr.bsky.social @camel-cdr.bsky.social · 27/12/2024

Removing single line comments in multiple lines in parallel with RVV. If anyone has a good AVX512 solution, please share. Inspired by this reddit post: www.reddit.com/r/simd/comme... #RVV #RISC-V #SIMD

$/* Note: '\' is '\n' */ 0123456789ABCDEFGHIJKLMNOPQRSTUVWX | vid Line 1# # # ## #\Line2 # #\Line 3 | v0 = input 0000001001010110110000001011000000 | m = vmor(vmseq(v0, '#'), vmseq(v0, '\n')) 000000111223345567777777889AAAAAAA | iota = vslide1down(viota(m)) /* vl-=1 */ 0######\##\ | v1 = vslide1up(vcompress(v0, m), 0) 000000###########\\\\\\\###\\\\\\\ | seg-cp-scan = vrgather(v1, iota) 0000001111111111100000001110000000 | seg-cp-scan == '#' Line 1\Line2\Line 3 | vcompress(v0, seg-cp-scan)$

272

Falvyu @falvyu.bsky.social · 27/12/2024

BTW the code requires VBMI2: pre-Icelake implementations likely have to widen to 32-bits, compress, and narrow to 8-bits. I also have ideas to improve the 64-bits segmented scan: instead of doing in scalar, perform SWAR within a SIMD lane and replace 'bit-select' patterns (4 ops) with vpternlog.

110

Falvyu @falvyu.bsky.social · 27/12/2024

I've taken a look at the problem: github.com/natmaurice/s... (code on master branch also handle quotes) The segmented scan is done in scalar/SWP to avoid inter-lane overheads. I've tested it, but not benchmarked it yet.

github.com

GitHub - natmaurice/simd-comment at comment-simple

Contribute to natmaurice/simd-comment development by creating an account on GitHub.

321

Reposted by Falvyu

FCLC (being a silly goose) @fclc.bsky.social · 23/12/2024

Mini thread on developer trust for #AMD. Effectively: the default assumption for AMD is that the stack will be broken. The community tried to help, but as Dave Airlie said years ago “throwing source code over the fence doesn’t make you FOSS” #HPC #ROCm

2245

Falvyu @falvyu.bsky.social · 23/12/2024

Of course, revenues likely lower on desktop than on datacenter. But poor results on desktop aren't a good display for other markets. And it's a shame that AMD makes great hardware but shoots itself in the foot due to poor software support.

120

Falvyu @falvyu.bsky.social · 23/12/2024

Even on the desktop CPU side, things are far from perfect: znver4 was submitted to gcc nearly 6 months after launch (vs ~1 year before launch for Intel). Zen 5 also under-performed on launch on Windows due to drivers issues (gcc support is there, so that's at least an improvement).

120

Reposted by Falvyu

halvarflake.bsky.social @halvarflake.bsky.social · 23/12/2024

Ok, reading semianalysis.com/2024/12/22/m... it seems that *most* of AMDs problems could indeed be remedied by having a team of 4-10 very strong engineers focus on always keeping default PyTorch fast on AMD hardware, with nightly regression tests.

semianalysis.com

MI300X vs H100 vs H200 Benchmark Part 1: Training – CUDA Moat Still Alive

Intro SemiAnalysis has been on a five-month long quest to settle the reality of MI300X. In theory, the MI300X should be at a huge advantage over Nvidia’s H100 and H200 in terms of specifications an…

4175