SIGGRAPH 09: Days 4 and 5

Friday, after the whirlwind of the show, I couldn’t muster the energy to write up the fullness of days 4 and 5. Luckily, a 3 hour layover in St. Louis accomodated me a bit. Since this will be a long post, here’s the executive summary.

  • More from Beyond Programmable Shading – This talk was very good, and I think it has some pieces that would be really nice to share even with co-workers who aren’t focused on graphics and GPU computing just so they understand the problem space better.
  • New Exporters for PhysX – NVIDIA had the new exporters showing in their booth. They look much slicker than the current stuff.
  • Papers from Friday – Some of these were really sweet. I have some summaries below. I think I’ll break out the more exciting ones in a later post.
  • Teapot FAIL. Chris wanted a teapot, but the line was too long. See image below.

To Shading and Beyond

The course notes are here, and they are worth a read. I’ll probably mention topics from this talk for some time. Here’s the relatively quick bullet points.

  • There was a great discussion of how GPUs are SIMD devices. We’re all familiar with the explicit 4X SIMD that we get in SSE instructions, but a lot of times we lose sight of the fact that GPUs are really SIMD in the same way. They are just 16X or 32X or more. This is the portion I mentioned above that I want to share with others in my company. When you have a very wide SIMD product, it behooves you to recast many if not all problems as compute intensive even if that’s more work.
  • The information presented by Johan at DICE about their expieriments with Frostbite was top notch. He showed a prototype running 1000 lights using a compute shader to sort, tile, and render the lights. That portion was the most valuable. The cool thing was that the compute shader used a few synchronization barriers. At first, it was parallel on pixels, then on lights, then on pixels again. It’s a very cool trick, and I never thought about parallelizing on multiple dimensions in a single shader to amortize the cost of dispatch.
  • In the afternoon, Paul Lalonde from Intel talked about extensible graphics pipelines. In short, he advocated a graphics pipeline that’s in place and usable but that has fully extensible pieces in software. On the one hand, everything that he said was straight Larrabee marketing. On the other hand, I really agree with his points. I don’t want to have to write a pipeline from scratch every time I grab a GPU. However, it’s nice to be able to violate any “restriction” that is there.

Let’s Get Physical

PhysX is a great physical simulation system, but the art pipeline has always left something to be desired. Luckily, there are new exporters for Max and Maya coming from NVIDIA. I got a demo at the show from Dan Horowitz. (I hope that’s his name. It was really loud the night before when I met him.) The new plugins look really slick and offer a much better interface for editing. However, they may be a bit of work for Emergent to update the Gamebryo exporters. I’ve definitely got to touch base with Adam and Stephen tomorrow to make sure we’re on top of this.

Papers and Talks.

Multi-Layer, Dual-Resolution Screen-Space Ambient Occlusion. The multilayer stuff was a technique similar to depth peeling to deal with single-sample artifacts in SSAO. I don’t think it’s as valuable as the dual resolution stuff which has some good performance implications.

RACBVH. That expands to Random Access Compressed Bounding Volume Hierarchies. This was a great presentation on compressed bounding volume hierarchies for large scenes. I don’t think it has a lot of value for client side graphics in games in the short term, but it could be useful on the server side or perhaps as part of a streaming system in a large world game.

Bucket Depth Peeling. This talk had very impressive results, but I think the cost is too much. 16 full resolution buffers for storing 8 color and 8 depth samples. That’s a pile of memory.

An Efficient GPU-based Approach for Interactive Global Illumination. This was a very impressiveĀ  implementation that ran a pretty robust global illumination on the GPU at 3-4 Hz. It’s way too costly for now, but the results were very impressive. Here’s a link to the paper.

Beyond Triangles: GigaVoxels Effects in Games. This talk was mistitled. I don’t think we’ll see this in games in the near term. However, the medical imaging applications in the short term could be enormous. This technique could render a 2,048*3 grid of voxels in real time. A full CT scan was one of the examples. Here’s a site from the author.

Teapots! We don’t need no stinking teapots!

Ok, I lie. The teapot is cool. Apparently, this year’s theme was 3D so it had some 3D glasses on. However, I couldn’t bring myself to burn what little time I had on Thursday waiting in line. Sorry, Chris.

Line for Teapots

Line for Teapots

For context, Pixar’s booth was a 20′ by 40′ affair. That makes the line a good 300+ feet long when compared with the booths on the map.

I hit publish because I’m tired of typing. Comment about mistakes. dba


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: