Could PlayStation 4 breathe new life into Software Based Rendering?

If I take it from what happened on PS3, one possibility could be that the CU would actually drive the engine. On PS3, initially the PPU would drive the SPU jobs by distributing and feeding the work to the SPUs. But the PPU itself was actually a bottleneck in this, and making the SPU the main job distributer turned out to be more efficient, leaving the PPU to other work.

I don't know if this is possible with CUs, but the AMD presentation suggests that it just might be. And in that case, that would indeed free up additional CPU resources, which are currently claimed to be a more likely bottleneck than GPU resources are.

 

Let's post this from Cerny again

So backface culling for one.

 

This IS the software based rendering thread, yes? One which you yourself started by the way. So tell me... Why would you render polygons...in software...when you have hardware rasterizers that could draw the same poly much faster?

Software based rendering - when used today, in realtime 3D context - typically entails non-polygonal rendering (voxels, splines, bezier patches, solid geometry modeling, whathaveyou), since we already have very competent hardware accelerators if polys is what you want to use. Offline software renderers as used by the movie visual effects studios and such also use polygons a lot of course, but those aren't exactly realtime rendered so is a different discussion.

Yes, but like I said, just because you're using GPU compute doesn't mean you're doing software rendering.

No, I'm more inclined to you seeing buzzwords on a slide and then jumping to conclusions that aren't there, or at least not in the form you think they are. (Or using commonly used terminology in novel ways, causing confusion.) GPU compute is just a fancy new name for stuff that has existed in more primitive versions since directx 9 came out ages ago now. SSAO, tone mapping, FXAA, bokeh DoF and things like that are all implemented as GPU compute, today, on graphics rendered using rasterized polygons. There's no indications any of that is going away, at all.

Not sure what you're going on about, TBH. In order for the industry to shift to a dramatically different method of rendering there has to be incentive to do so; drawbacks to be overcome and benefits to be gained. You don't just do it for the hell of it, or because of fking powerpoint marketing slides. Where are the drawbacks and the benefits?

 

You didn't. But you were talking about moving from fixed function pipelines, yet the current GPU is very programmable, so not 'fixed function'.

You didn't describe any graphical workloads that could be moved to compute, so Taisui asked for clarification over what you meant, and you're clarification talked of fixed-function hardware...

Please explain what you mean by 'programmable pipeline'. What are the workloads that can be moved to compute, and how are these workloads not currently programmable so that "compute == software based rendering" and "shaders != software based rendering".

There's certainly a discussion there, but you haven't expressed it such that anyone can get involved because everyone's still crashing heads over the definitions.

 

Yup.
Sounds a lot like what they had for the ps3 which could cull triangles not affecting the final image.
http://www.jonolick.com/uploads/7/9/2/1/7921194/gdc_edge_07_final.pdf

 

what makes something software anymore...it all runs on programmable hardware


in opengl4, for example, i write the T&L code (or use someone else's)...

 

FWIW my definition is software rendering = pure general purpose CPU code (inc. modern SIMD FPU code) with no use of specialist co-processors such as GPUs. Regardless of the programmable nature of a modern GPU it is still fundamentally geared towards large scale matrix ops and if your code is 'branchy' and serial it will be a very poor match even for GPGPU (to my understanding of it). I look forward to being corrected!

 

When they say "efficient branching", are they talking about something that was already part of the GCN architecture, or a Sony customization?

 

GCN has a separate scalar pipeline that is generally used for branching and control flow. This makes coherent branches free in many cases.

Fully software based GPU rasterizers already exist. Link: https://research.nvidia.com/publication/high-performance-software-rasterization-gpus

 

That guys "insider" info is so lame. SDKs is improving (like every other generation as time passes), Uncharted will have amazing graphics (no shit), Microsoft will show an exclusive (when have they not?), both studios will pad their presentations with 3rd party content (as usual). I mean, what a genius.

 

Great link, cheers! Man it's going to take quite a few passes to understand much beyond the abstract

 

No all the actual polygon rasterization / shading is still done with the hardware

 

He was wrong about several things one example is The Witcher 3 resolution.