> fully-asm coded polyfill (flat shaded) and it uses NO variables, only registers

As someone who writes small hobby OpenGL 4 games with zig for fun and because I'm not using a game engine I think I'm doing low level stuff, but this really appropriately puts what I'm doing into perspective. This quote is pretty awesome. I love how we got to where we are these days. Early game developers were true pioneers and we owe you so much!

LOL. I was still in secondary school back then. It'd be another 3 or 4 years before I became a game dev for real. The same 3D engine I'm talking about there, I ported it to DirectX and used its poly-fill for this game:

https://www.youtube.com/watch?v=t2kdKB18c7I&t=332s

Were they really getting 16000 texture-mapped polys on a DX2?

We will never know. I think I was probably 16 at the time there. With a lot of hindsight and watching some videos lately about hacking the Mario 64 3D routines I wonder if the techniques we thought were fast back then are outdated now?

There are better profiling tools available now. I might be able to do it much better. Plus, not all registers are equal on the 486. Using FS and GS to hold data might be slower than pulling it from RAM. How much on chip cache does a 486 have? I don't know.

It sounds reasonable to me.

Around 1994/5 I was doing perspective-correct texture mapping on a Sun SPARC (with real-time dithering too), an Atari Jaguar and a 486 DX4/100. 16k/sec polys on the DX4 (which is only ~500 at 30fps) sounds modest to me. (We had technically more texture-mapped polys than that, but they had geometric constraints with each other so they weren't all independent polys and shouldn't be counted the same way.)

I was surprised, coming from earlier processors including the 386, to find floating-point was sometimes faster for calculations for things like texture UV interpolation, instead of integers (fixed-point), x86 sub-register tricks and hyperbolic-Bresenham or line-at-infinity techniques to avoid division for perspective.

These are the sorts of questions I would like to solve. Were we mistaken in using fixed point for everything? Are there actually faster ways of doing most of these tasks.

That recent Mario 64 rewrite has me questioning everything. At least Nintendo's first party developers made the same mistake, and they had the hardware guys right there.

https://www.youtube.com/watch?v=xFKFoGiGlXQ

Brian Hook was in there as well. Wasn't he at id at some point?

Also, lots of people on Netcom. That was my ISP as well around that time and I still have their DNS server IPs memorized.

Yeah, he was at iD. I think Michael Abrash would pop up too back then.