←
Jeremie Pelletier42,393 posts
Jeremie Pelletier avatar

Jeremie Pelletier

@HostOfMeta

Self-dual

◷ Joined May 2012▧ 2,515 Photos and videos
Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·6 Apr 2024

Need more tools to bind all these feature together:

What happens when you combine Emacs with gaming? You get an environment in which any panel can display any visualization for any data type. And if it's not there, you just develop it right there, live. Coded support for multiple viewports as it draws 60fps :) #gamedev

0:45
Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

Teams that continuously rewrite their systems from scratch will always end up outperforming teams that keep adding to the codebase. Less is exponentially more.

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

Could go with a full suite of UI tools, but that decouples the UI subsystem from the rest of the art pipeline. Instead can tweak the model format a bit to also fit UIs in it, then reuse its animation system, using bones for placement. A few powerful abstractions beat many teams.

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

Writing skeletal animations could start with a blend graph, or it can skip that and go with priority overrides. Skip half the work, get anims working directly on the end pose, and have other anims take priority on a subset of bones. No blending, no morphing and controls better.

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

Had I done it another way I probably would've started with a fat gbuffer, a forward+ pass, a post-process stack, and then realized bindless and BDA throw half of it away by changing what gets possible in the first place. Had I gone PBR I'd have irradiance baking and miss on art.

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

I thought I was doing lots of work generating and drawing immediate mode UI each frame, which gets rather big now, still takes only 0.1ms of frame time. Made a ubershader to iterate on lighting and material support and it draws straight to the window surface, skips many steps :D

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

Write a profiler built in the engine that renders its own stats, so each change sees not just its results but also its costs. That adds another dimension to learning. Write throwaway debugging views, add prog art gizmos, use AI to generate starting data and assets, it all works.

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

Iterating using compiled languages too slow? Use hashmaps for objects in a way that gets you to see changes as you make them. You'll learn a lot more. Is it performant? Hell no. But it hits 60fps anyways for a test scene. Is it fun to work with? Hell yes. Go AAAA scale later :D

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

One error we make comes from optimizing the design once it work, instead of making a new architecture on the side built from the lessons learned. Both improve perf, but the latter also improves everything else, and that pays off infinitely more in the long run. Keeps it simple.

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

That amounts to 0.5% of CPU time on a decade old PC. Use the naive GPU approach to learn how a subsystem works; even if it takes 1ms of frame time. Go for "we call json.serialize and have an exporter" just so data starts flowing between processes. Redesign when things can move.

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

More than this, seeing the better way often comes from seeing the previous way in motion, rather than staring at notes or plans that fail to get implemented when meeting reality. Burn through 100Mb of GC memory a second in the first prototype, so long as it gets things moving :D

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

I'm learning a ton going through different engines and their shipped content, even from those whose data layouts I could improve. Improvements come iteratively; make it work, then make it good. Nobody gets it perfect first try, and many ship the first version. Find a middle way.

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

Another studio I worked at had its architects openly criticizing Unity3D, while that engine could do hundreds of things their internal one couldn't. That prevents the team from learning about other ways to do things and developing perspective, then understanding tradeoffs better.

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

One AAA I worked at went "we don't study other engines" because they believed their flagship one led the industry, until artists started prototyping using UE5. Then it becomes a divide, because both engines do things the other cannot do without heavy investments. Adds friction.

Jeremie Pelletier avatar
Jeremie Pelletier@HostOfMeta·19h

And while it seems risky, this makes the fun paradox. It gets risky upfront while lowering the long term risk. Because understanding gets developed in those jumps. It keeps the system flexible to changes, it teaches edge cases, how to design it better each time and it compounds.