5 replies to this topic

[Telios]

Hiya

I've just switched to XNA, and I had a couple of questions I was hoping someone can help with

Could anyone explain the relationship between eDRAM and render targets on the Xbox 360? Each render target must have some VRAM assigned for the texture data. When you bind a render target with DiscardContents, this texture isn't copied into eDRAM (since this would be slow?) but the texture data is cleared to purple to show this. I guess PreserveContents would cause the texture to be copied, although this would probably be very slow.

Is this correct? At what point is the render target in eDRAM copied back to the texture store in VRAM?

Also, could anyone tell me how depth/stencil buffers are handled on the 360? If I create a render target with a depth buffer attached, where does the buffer live? Does it have a backing store in VRAM?

Thanks for any help!

Cheers
Jim

Edited by Telios, 21 January 2013 - 07:06 AM.

[phil_t]

The render target data in EDRAM is copied back to regular memory when you resolve the render target. That is, when you switch to another render target or back to the back buffer.

 

The depth/stencil buffer lives in EDRAM with the render target*, so this takes up part of your 10MB limit (4 bytes per pixel). It never gets copied back to regular memory.

 

*I believe there is also HiZ/HiStencil that has a small but of special memory on the GPU, separate from EDRAM.

[Telios]

Thanks for the quick reply!

 

Just to clarify - are depth/stencil buffers shared in any way in EDRAM? For example, if I create a second render target with a depth/stencil attached, is this depth buffer a separate area of memory or is it shared with the backbuffer's depth?

 

I'm trying to avoid tiling so it would be useful to know how much EDRAM I'm using.

 

Thanks again.

[NightCreature83]

You have to remember that a 360 uses tiled rendering as a full 720p target(1280*720*24 = 21MiB) will not fit in 10MiB EDRAM so it will always tile.

[phil_t]

The backbuffer can be thought of as just another render target. In that, I mean that it doesn't take up any space in EDRAM if you've currently set another render target to be active. EDRAM is just used for the render target that is currently active. So all you need to do to determine if tiling is triggered is to calculate how much memory your render target + depth buffer requires.

 

So, for a 1280x720 Color render target with a depth/stencil buffer, you have: (1280*720 * 4) + (1280 * 720 * 4) = 7372800 bytes = 7.03MB. So that fits comfortably in EDRAM (NightCreature83's calculation is incorrect - a Color render target is 4 bytes per pixel, not 24).  If you're using multiple render targets (say for deferred rendering), then 2 of those plus a depth buffer gives you 10.54MB, which means you'll end up tiling.

 

FWIW, I've rarely noticed any performance degradation from tiling. I'm sure it could happen if your performance bottleneck is vertex processing, but I think in most cases it's not a big deal.

 

Note that you can also visually see the tiling if you write a shader that uses the VPOS semantic. VPOS, which gives the current pixel position, "resets" for each tile. So you can write a test shader that can show exactly where the tile boundaries are.

Edited by phil_t, 21 January 2013 - 01:21 PM.

[Telios]

Exactly what I was looking for, thanks very much!

Edited by Telios, 21 January 2013 - 02:02 PM.