autumn

autumn

@froggi.es

i work on drivers, hdr/colour, linux kernel, gamescope, dxvk and vkd3d-proton, old games. source engine masochist. opinions are my own. frogs are cool. github: https://github.com/misyltoad fedi: autumn@autumn.moe email: hello@froggi.es

1.3K followers335 following152 posts

autumn
@froggi.es

Can memcpy hang and never return? Yes. Today, I hit an issue where memcpy was not returning, making no forward progress at all, not even a single instruction. 1/3

autumn
@froggi.es

__memcpy_sve () at ../sysdeps/aarch64/multiarch/memcpy_sve.S:126 126 stp C_q, D_q, [dst, 48] (gdb) bt #0 __memcpy_sve () at ../sysdeps/aarch64/multiarch/memcpy_sve.S:126

autumn
@froggi.es

Based on the dmesg spew, I can only assume stp was faulting, the kernel was handling it, but failing to page in the needed stuff for the process, but ended up telling it to retry anyway... So it was faulting, and retrying stp over and over again...

autumn reposted

Tholp
@tholps.site

rude neighbors using leafblowers and lawnmowers while im trying to sleep at 2pm in the morning

autumn
@froggi.es

To whoever sent me a direct message on this godforsaken platform, I cannot see it as I live in the UK. You should contact me by another means.

autumn reposted

autumn
@froggi.es

After about 3 days of work, I got DLSS to work on NVK (the open source NVIDIA driver in Mesa)! 🎉 DLSS is essentially just some launched Cuda kernels, so it was just a matter of implementing the VK_NVX_binary_import and VK_NVX_image_view_handle extensions. 1/5 (Control, 960x540 -> 1920x1080)

autumn
@froggi.es

After about 3 days of work, I got DLSS to work on NVK (the open source NVIDIA driver in Mesa)! 🎉 DLSS is essentially just some launched Cuda kernels, so it was just a matter of implementing the VK_NVX_binary_import and VK_NVX_image_view_handle extensions. 1/5 (Control, 960x540 -> 1920x1080)

View full thread
autumn
@froggi.es

The code is not ready to land yet (lots of clean-ups needed!), but it's working now in Control! For those who read, here is a picture of the initial froggy sample app and kernel I wrote to test the extensions before delving into the extra bits that DLSS needs to work. ;D 5/5

autumn
@froggi.es

(also special thanks to Mary for helping me find the smem in the qmd thing despite the kernel launch at first glance not appearing to specify any)

autumn
@froggi.es

After about 3 days of work, I got DLSS to work on NVK (the open source NVIDIA driver in Mesa)! 🎉 DLSS is essentially just some launched Cuda kernels, so it was just a matter of implementing the VK_NVX_binary_import and VK_NVX_image_view_handle extensions. 1/5 (Control, 960x540 -> 1920x1080)

View full thread
autumn
@froggi.es

multiple sections an ELF where some of the metadata is ordinal-based and some of it is name-based and oh, the ELF is also inside another container of ELFs (elves?). But once you parse all that out, it's actually not *that* bad. 3/5

autumn
@froggi.es

The surface area is mostly in the parsing of the cubins, the actual dispatch code is tiny! ATM, it only supports DLSS versions that have code compiled for the GPU you are using. If you try and run old DLSS on a new GPU, that won't work currently. PTX->NIR is much harder. 4/5

autumn
@froggi.es

After about 3 days of work, I got DLSS to work on NVK (the open source NVIDIA driver in Mesa)! 🎉 DLSS is essentially just some launched Cuda kernels, so it was just a matter of implementing the VK_NVX_binary_import and VK_NVX_image_view_handle extensions. 1/5 (Control, 960x540 -> 1920x1080)

autumn
@froggi.es

These extensions are used by DXVK and VKD3D-Proton via. DXVK-NVAPI to implement DLSS. Although, that alone was not an easy feat because the surface area of Cuda is pretty huge and parsing the cubins is difficult as there's these weirdly packed attributes spread across- 2/5

autumn
@froggi.es

Not once have I ever wanted clangd to insert headers into my code, or touch it in ANY way. Stop doing that! WHY IS THAT DEFAULT?!

autumn
@froggi.es

my hot take is that the majority of desktop applications (not games) shouldn't need to be and should not be "rendering" anything. simply forwarding their scene graphs to the higher level compositor which can do a much better job.

autumn
@froggi.es

Google is a bad keeper of Android. The new "you need to give Google your ID for sideloadable apps" thing is a huge and disgusting overstep. I wish there was a serious, properly funded, open-source competitor in the phone space.

autumn reposted

Bruno Dies
@brunodias.dev

ah yeah the familiar whiplash of seeing a legacy media org cover something you're an actual expert in and realizing "oh if this is how badly they're fucking up this, how badly are they fucking up every other topic?"

Bryant Francis@bryantfrancis.me

This is an insane thing to write about game development. www.nytimes.com/2025/07/28/a...

autumn
@froggi.es

the lidl candy creme mini bars are soooo much better than real milky ways. just tried a real milky way again recently and it was just awful, the nougat is wayyy too sweet and grainy with shit chocolate. just get the off brand lidl ones. they're just 200x better.

autumn
@froggi.es

went to the Newcastle pride event today only to find out they want £13 to enter. wtf?! i refuse to be extorted for being gay. that is insane.

autumn
@froggi.es

hopefully i am not stuck in Newcastle with the current train situation... not hopeful right now though after an hour delay that keeps ticking up :x