Sidney

My hot take here is that this is fundamentally how LLMs work and also exactly why I wouldn’t touch them for development work. I have tried a few times with Vulkan questions that I knew the answer to and my heart goes out to anyone who listens to Claude instead of reading the spec.

Amazing, thank you so much for the tip! This is now a must have extension together with Sponsor Block to make Youtube usable.

Flight sims are super fascinating projects (citation needed, author is heavily biased). One extremely large level, the main character is crazy fast and there are tons of mods/plugins/scenery add-ons. Should be fun, I hope 😄

Looks like I’ll be giving a talk about working on X-Plane at GPC 🥳. Covering how we went from a GL 2.1(!) codebase to a modern Vulkan/Metal one and the challenges and advantages of working on a single product for 30 years, with an eye on the unique challenges of a small studio working on flight sims

I’ve been enjoying doing CAD work on the Surface with a pen, but the keyboard has been incredibly awkward. Figured this was a good opportunity to finally use one of the Pico Pis collecting dust and make a macro pad thingy. Quite pleased with how it turned out.

Made the mistake of looking at the Hacker News comments on the preliminary report of the Air India crash. People who regularly rant about managers demanding features with no understanding of the problem space are now seriously saying “planes should just not allow actions that can crash it” 🙄

Be still, my beating heart <3

Before that they were actually floppy too!

They are growing up so fast 🥹

Still no way to query hardware features and limits programmatically though. Just that one PDF that we all get to hardcode into our games and apps. It’s my biggest gripe and annoyance with Metal and I can’t believe it’s still like this.

Yeah, the 900 series is just at the edge of our support matrix. It’s the one odd one out in a few other ways, like no async compute, no Aftermath support (just don’t crash, duh). It’s a step up though, our previous major version supported down to the 6xx gen and those are rough these days.

Also when I wrote first gen I really meant second gen. First gen might do this too though, but thankfully it’s out of the support matrix and thus not my problem anymore.

It’s one of those insidious bugs that keeps popping up ever so often since no sane developer wants to use a 980 as their daily dev GPU. We colour NaNs/INFs magenta in the final render to make debugging easier and if I see a bug report with 9xx and magenta in the title, I know we screwed up again.

Fun quirk of first gen Maxwell GPUs: If you have a f16 render target, output to it won’t be clamped to the highest f16 value but instead overflows into INF. This gets particular fun when you turn on blending and no amount of shader side clamping can save you.

Also if you publish samples, please name all the resources in it so when I inevitably double check my implementation against the sample in RenderDoc, I don’t have to rename all resources by hand. Especially if they are ping ponged a lot. Looking at you, FFX SSSR and CACAO.

Pure evil! What did the syntax highlighting look like for the call?

My (now) wife and I met on Omegle and then talked a lot on Skype. That was in 2012. We also talked quite a bit on Facebook at the time, my only hope is that this trend continues and we are going 3 for 3.

Anyways, it doesn’t really matter because the end result looks totally amazing. I wish I had even an ounce of this artistic talent!

As far as I know the HDR and LDR astc formats are both 16 bytes per block. Even the uncompressed HDR formats just take bits from the alpha channel, eg rgb10a2, so still 32 bits per texel just like rgba8 would be. Even rgba16f should be “just” 67mb, so a compressed format should beat this handsomely.

Those 3 formats all have a 4x4 block size with 16 bytes per block. (2048 / 4) * (4096 / 4) * 16 comes out to 8.3mb. Although that’s without any mips, not that they weigh much in the grand scheme of things.

That math checks out, but where is the 25 coming from? This might be the part where I’m making a huge fool of myself online, I’m a desktop guy so basis u and astc are exotic to me. If I understand basis u correctly you get either a 4x4 astc hdr/ldr or BC6H texture out of it.

This looks really awesome, but the 200mb number is throwing me for a loop right now. With a 2k by 4k image that’d be 25 bytes per pixel. Even with mips and HDR that feels astonishingly high. I feel very stupid, what am I missing?

Honestly, forget about AI tools. The trifecta of power tools for game engine development is RenderDoc, Superluminal and Live++. Hands down the most bang for your buck you can ever get and RenderDoc in particular is criminally underpriced.

I had a similar shower thought this morning and ended up thinking that it’s impossible to change because it would break basically every little endian encoded file.

Edit and Continue has been fully replaced by Live++ for me. That one just works and it’s nice to be able trigger it even outside of debug sessions with a global hot key. But yes, this looks amazing! I really hope it works well, it would make my life in the low level trenches so much easier.

C++ Dynamic Debugging: Full Debuggability for Optimized Builds - C++ Team Blog Over the past 5 years, we’ve had many incredible opportunities to engage with game developers. From AAA studios to indie developers, the passion for slashing iteration times is significant. Amidst all...

Tired of debugging optimized code with missing variables and inlined functions?

@visualstudio.com preview now has Dynamic Debugging which automatically unoptimizes functions you step into while debugging 👀

Available in UE 5.6, or with a simple cherry pick:
devblogs.microsoft.com/cppblog/cpp-...

Crazy to think that this message could be prompted by your work on Superluminal or just because your wifi and/or audio stopped working. Btw, if you guys ever need testers for the Linux version of Superluminal, I would love to get my hands on it.

But descriptor sets are a gold mine for Vulkan blog content. Also job security.

Reading this and the replies is like walking into an alternate reality. I haven’t had a single Windows update brick my PC. Is this really a common experience?

I am so used to doing this in seconds by pure default, I hd to actually look at this twice to make sense of the result. And now I feel a little stupid for never even considering this to get milliseconds out.