Re: 16bit vs 32bit render targets
From: Eyal Teler (et_at_nospam-et3d.com)
Date: 05/16/04
- Next message: Eyal Teler: "Re: Problem with CreateOffscreenPlainSurface"
- Previous message: Joris Slob: "Problem with CreateOffscreenPlainSurface"
- In reply to: fry: "16bit vs 32bit render targets"
- Next in thread: Eyal Teler: "Re: 16bit vs 32bit render targets"
- Messages sorted by: [ date ] [ thread ]
Date: Mon, 17 May 2004 01:06:11 +0200
fry wrote:
> Hi,
>
> I did some tests with D3DFMT_A32B32G32R32F and D3DFMT_A16B16G16R16F render
> targets, I use them to render 128 by 128 textures in multiple passes ( I
> also use 3 render targets as output from pixel shader) after rendering I am
> copying RT to systemmemory to read its contents. I expected 16 bit format to
> be around twice as fast but it appears to be at the most 10-15 % faster. Is
> it possible that GPU uses 16 bit as 32 bit inside? - I use ATI 9800 Pro,
> thanks for any help.
Why do you expect 16 bit to be twice as fast? If you're purely
bandwidth bound then that could be the case, assuming you're accessing
only those FP textures. But otherwise there's no more reason to expect
that than there's reason to expect RGB565 to be twice as fast as ARGB8888.
THe Radeon 9800 Pro's pixel shader use 24 bit per component. But even
on the GeForceFX chips, which do differenciate between 16 bit and 32
bit processing, you can't expect to double performance automatically
by moving to 16 bit. And remember that to make the shader use less
accuracy, you should use the 'half' type (or _pp hint in shader assembly).
Eyal
- Next message: Eyal Teler: "Re: Problem with CreateOffscreenPlainSurface"
- Previous message: Joris Slob: "Problem with CreateOffscreenPlainSurface"
- In reply to: fry: "16bit vs 32bit render targets"
- Next in thread: Eyal Teler: "Re: 16bit vs 32bit render targets"
- Messages sorted by: [ date ] [ thread ]
Relevant Pages
|
