Point sprites in XNA Game Studio 4.0

Originally posted to Shawn Hargreaves Blog on MSDN, Monday, March 22, 2010

As of version 4.0, XNA Game Studio no longer supports point sprites.

Why?

One big reason, plus two small ones.

The big reason is that DirectX 10 and 11 do not support point sprites, so if we kept them in our API, we would be unable to someday move that API to future DirectX versions. This would violate our "break it good" goal of taking all the pain now to avoid future compatibility breaks.

It is theoretically possible to emulate point sprites using geometry shaders, but after some investigation we concluded this would be:

• Complex. Not always a bad thing, but complexity can be an early warning of designs that turn out to be buggy or slow, so it makes me nervous if I see too much of it!

• Not fast enough to be useful on several popular GPUs.

A smaller reason is that point sprites behave differently on Windows and Xbox, and are not even consistent from one Windows GPU to another (the max size varies, as does the result of TEXCOORD interpolators).

The final reason is that point sprites are slower than triangles on some common graphics chipsets. Hardware manufacturers are sure to keep point sprites around for DirectX 9 compatibility, but they are unlikely to spend much time optimizing what is now a legacy feature, so this performance delta will only grow over time.

What can I do instead?

If you are worried about the idea of life without point sprites, the first thing you should know is that this worried me at first, too. But I'm not worried any more. By the end of this article, hopefully you won't be either.

When we looked at why people use point sprites, we found two main purposes:

• As a simple way to draw 2D dots, such as in our Primitives sample. These scenarios can easily (and with great performance) be changed to use SpriteBatch with a 1x1 white texture, so point sprites aren't necessary for them.

• As a high performance way to draw particles, such as in our Particle 3D sample. This scenario was more concerning. Anything you can do with point sprites, you can also do with triangles, so our worry was about performance rather than functionality. It would be a bad thing if the 4.0 particle sample could not run as fast as previous versions!

If you've been reading my blog for long, you probably already heard me mention how the first step in performance work is to measure. Using GS 3.1, I changed the particle sample to draw triangles instead of point sprites. In fact I made three versions:

• The original point sprite implementation. This requires one vertex per particle. There are two versions of the pixel shader: a cheap one for particles that do not rotate, plus a more expensive shader for rotating particles.

• Using triangle lists, with one triangle and three vertices per particle. To fit all of a square texture onto a single triangle, the triangle must be expanded somewhat larger than the final particle size. Because it can rotate in the vertex shader, this implementation is able to use the cheaper pixel shader for rotating as well as non rotating particles.

• Using indexed triangle lists, with two triangles and four vertices per particle. Although it requires one more vertex than the previous implementation, this can fit the entire texture without needing to expand the triangle, so fewer pixels need to be shaded.

All three versions used insignificantly tiny amounts of CPU time. The difference was in GPU performance:

• Triangle lists took 129% as long as indexed triangles (turns out that having to shade fewer pixels more than makes up the cost of that extra vertex).

• For particles that do not rotate, indexed triangles took 180% as long as point sprites (ouch! those extra three vertices per particle are really costing us. But it isn't 4x slower, like we might naively expect). 

• For particles that rotate, indexed triangles took 70% as long as point sprites (whoah! This got FASTER. Sure, we have to shade an extra three vertices per particle, but the cheaper pixel shader more than makes up for that extra vertex work).

A 30% gain from not using point sprites is pretty sweet, but an 80% penalty is painful indeed. I see more people using rotating particles than otherwise, but still... ouch!

One of the best things about my job is the chance to work with amazingly smart people. When I mailed around these performance figures, I was lucky enough that one of the graphics gurus who created the Xbox GPU driver happened to see them, and was sufficiently intrigued to spend some time looking at PIX captures of my test app. Jason discovered that my indexed triangle implementation was bottlenecked by rasterizer performance, specifically by an obscure hardware penalty which occurs if you use both center and centroid mode interpolators with a shader that is fast enough not to be bottlenecked by anything else (an extremely rare situation).

Once identified, it was trivial to remove this bottleneck, which left my indexed triangles bottlenecked by texture fetches, just like the point sprite original. Revised timing data:

• For particles that do not rotate, indexed triangles take 104% as long as point sprites.

• For particles that rotate, indexed triangles take just 40% as long as point sprites.

Armed with these figures, removing point sprites didn't seem so painful any more.

I can't promise exactly when, but we are working on an updated version of the Particle 3D sample which shows how to use indexed triangle lists, and includes Jason's magic optimization. We will be sure to get this out by the time 4.0 ships, if not before.