Graphics Update, v1.1

Back in December, I posted a long, waffling post about some changes I made to the low-level graphics engine. Some people seemed to like it, so I thought I'd do another update.

My primary focus has still been on performance, with the occasional bug fix thrown in here and there for good measure. One of our big performance pain-points has been the effects system - that is, the part of the code that renders effects like explosions, rocket trails, the planet smash, weapon fire, and all that good stuff. This is also alternately known as "the particle system", or "argh argh argh" when I was feeling particularly annoyed.

Anyway, this system was originally written to be entirely simulated on the CPU, which allows us to do a lot of cool stuff without requiring the end user having the latest and greatest graphics hardware, and is also a lot quicker to implement. Unfortunately, as the game scaled up, the effects system did not, and it became an issue. For the terminally curious out there, it used to work by building up all the geometry for particles on the CPU, and then render them all out as if they were just a mesh. The upshot was the CPU had to do a TON of work when there was a lot going on, and there was also a lot of data being thrown around.

I spent the best part of the last month revisiting the system, and optimizing it. We still do the effect simulation on the CPU - but now particles are actually rendered as instanced quads (except the planet smash debris, which is instanced meshes), and all the of calculations required to set up their appearance are performed on the GPU. We also send a lot less data across the bus, as individual particle effects now understand the minimum set of data required by the GPU for rendering. Because of this we reduce the memory transfer per frame by a factor of 4 or more (actually, more like a full order of magnitude, so 10x or more).

A lot of this work was built on top of some of the low-level engine re-architecture changes I made back in December. It made putting all the bits together a lot easier than it would have been.

The net result is quite a decent speedup in late-game scenarios. This is not the definitive optimization by any means - we still have more to do, but it's made a difference. To give an example, typically in late-game scenarios I would see the CPU particle system take up 10ms or more (even upwards of 30ms when I sent 1000 bots on a rampage). This time is now in the order of between 1 and 3ms. It also speeds up other parts of the rendering (at least on NVidia GPUs), since the driver is better able to optimize around what we're doing (yes, I've been talking to the NVidia guys a lot lately - and for what it's worth, they've been amazing! ).

You should see this going live shortly (it wasn't quite done for the last build release). We're still in the process of testing it. But once it's out there - stuff should start getting a bit quicker.

Oh, in case anyone's curious about how I tested the performance as I was iterating... I set up the "ping" command to perform the planet-smash effect, after Metabolical did the same thing to test the original ping functionality (and posted a video about it, too). Then I'd do a bunch of them. So with around 10 planet-smash explosions going on, I was still seeing a solid 60fps. Which I'm quite happy with ...

It's not in the latest update (we did a bunch of *other* optimizations on that update). So expect to see even more speed in the next one (at least, in later game).

Of course I love AMD! And the improvements will apply on everything (since it's just more efficient all around). I wanted to call out NVidia especially because their developer relations guys were nothing short of amazing.

Let's not be premature... wait until you've played with these changes first...