Hacker News Reader

Comment by psyc | Hacker News Reader

I also wrote software renderers in the 90s. I would say it was mixed. Some people used matrices, and some didn't. You really can't say that everybody used the matrix formalism across the board. For that matter, I worked on an engine at a well-known game company as recently as 2013, where there were definitely hand-optimized paths for several different cases of transforms where the constraints were known. Generally speaking, game programmers will do whatever it takes, and painstakingly optimizing algos down to minimum arithmetic operations has always been pretty common in the field.

View on HN

I too wrote software rasterizers in 90s and even shipped a game with one. If you used a single sine for each vertix, least computing full rotation, you were fucked. There was no hardware, which could handle this at real time in 90s.

If you did division for each scanline, you were running at 1/4th of the speed at most - division was that expensive on 386-Pentium, not to mention other platforms, where CPU's could not even have hardware div.

> Generally speaking, game programmers will do whatever it takes, and painstakingly optimizing algos down to minimum processor operations has always been pretty common in the field.

I am not even talking about crazy optimizations (like using intel's x86 half register to do ghetto fixed point), I am talking about common sense stuff.

Ah. I think you're talking about something completely different from what jblow and I were talking about. If you mean that the code in the article is not optimized, and is doing crippling amounts of redundant computation, then of course I agree with you.

The guy asked what kind of magic is calling trig functions in inner loops? I replied that nobody really did it and I am confused as to what the author was trying to show.

I, frankly, do not understand what you and jblow read into this other than what I said.

Hahaha, I guess it's a big misunderstanding. You wrote:

"Nobody really wrote software rendering like that beyond CG classes".

I read this as a claim that nobody in general wrote software renderers. When by "like that" you just meant using the specific techniques he used.

That said, I still have to disagree, in the sense that, to get to a fast software renderer, you start with a slow software renderer. Nobody does all the crazy optimizations a priori ... so stuff like a divide per pixel was common, say. Calling trig functions in inner loops is of course goofy, but my presumption is that in the next step of refinement those would be lifted out of the loops, because that is the way things are always done.

As I said, it's bad as a tutorial:

It enforces the stupid idea, that you rasterize 2D triangles for a 3D renderer. People, who think this way are then completely shocked and mystified as to why there is perspective distortion to the surface attributes.

And then, on 2D triangles, it enforces a bizarre notion that a triangle is not flat, so the surface attributes gradients can change from scanline to scanline.

> Nobody does all the crazy optimizations a priori ... so stuff like a divide per pixel was common, say. Calling trig functions in inner loops is of course goofy, but my presumption is that in the next step of refinement those would be lifted out of the loops, because that is the way things are always done.

Yeah, .. but no. Depends on what era you're talking about I guess.

When I wrote my first low-level rasterizer + basic 3D poly engine-ish-thing in 1998, I honestly wouldn't have considered for a second to do a division per scanline. An integer add was one tick, a mul was 3-10 (iirc), but a division was 10-40 ticks. Depending on what point you start considering the optimizatons "crazy", yes, had I known (and cared) about perspective correct shading back then, I would have started designing the algorithm, a priori (which for me usually meant on grid-paper) hunting for some way of faking a sufficiently accurate reciprocal using adds, shifts and at most 2 muls (per scanline, cause per pixel even a single mul was madness, obviously). With what I knew back then, probably go for a 2nd degree polynomial that might be sufficient to at least give the impression it was doing better than naive bilinear? :) Had I been aware of Carmack's (objectively crazy) inverse sqrt hack, I would probably have started looking in that direction (abusing the IEEE float spec on the bit level woohooooo).

Sure you could write a perspective correct triangle rasteriser with a div per scanline, and it would be too slow, it would be a nice theoretical proof of concept, but it would also be kinda useless if it turned out you couldn't make the above crazy hacks run fast enough. They were a real hurdle that had to be crossed or you might as well not bother. Also, why save the fun stuff for last? ;-)

> like using intel's x86 half register to do ghetto fixed point

Just curious, what's this and how does it work?

x86 has a bizarre register architecture. 4 "general purpose" registers are like this : EAX & 0xffff == AX == (AH<<8) | AL i.e. the lower 16 bit portion of a 32bit register is a separate 16bit register, which, in turn, is made of two 8bit registers. If you stored a 8.8 value in a AX, for example, you could immediately access the integer part through AH. Of course, 8.8 numbers are not good for all values but you could write fast texturing loops with them (256x256 used to be a solid size for a texture so 8.8 was okay for texcoords).

whoa, my compliments how you managed to explain in a single paragraph both the bizarre x86 registers and fixed point math, succinctly :)

Reply on news.ycombinator.com