View Full Version : 8.11: GL_EXT_transform_feedback on ATI card!

11-13-2008, 07:28 AM

Seth Hoffert
11-13-2008, 08:26 AM
It is good to see that AMD is being true to their word. :D

11-13-2008, 02:38 PM
There are also entrypoints for these extensions: GL_ARB_map_buffer_range, GL_NV_conditional_render
Those might come next time.

11-13-2008, 04:24 PM
What in the Sam Hill is GL_AMD_texture_texture4?

11-13-2008, 06:43 PM
This could be the OpenGL implementation of D3D10.1 Gather4. It allows a texture fetch of 4 texels in one call. Quite a cool feature for soft shadows!

11-13-2008, 07:01 PM
Good guess ;-)

11-14-2008, 03:03 AM

Amazing! Oo
ATI start to make me very happy!
I would just enjoy the release of specification for their extensions. :)

11-14-2008, 04:05 AM
Would such a 4-texel fetch by chance require DXTn texture compression to not get a performance penalty?

11-14-2008, 04:55 AM
I think it could provide more efficiently in any case. Even if we speak about "scalar" processor every think is processor by batch of data accordingly to cache and register size. This is just a way to make just that speficique fetch are processed together. Moreover, ATI had quite time to optimize this feature, its actually an old one integrated in X1600 chips that should work with DX9. That was missing feature in X1800 but then including X1900, big buzz at the time but I don't really think it diserved it ... I mean that much even if it is great!

11-14-2008, 05:37 AM
This Fetch4 feature is present in DX9 on ATI. There is a paper about it in ATI SDK March 2006. They even planned GL_ATI_texture_nearest_four in OpenGL but as we know it has never been shipped.

11-14-2008, 10:22 AM
Would such a 4-texel fetch by chance require DXTn texture compression to not get a performance penalty?

I seriously doubt it. Most texture hardware fetches 4 pixels from the texture regardless, since it is designed to assume bilinear filtering will take place. Which is why most hardware swizzles the texture internally in such a way as to make 2x2 pixel fetches more efficient.

11-14-2008, 06:53 PM
Texture4 could actually be the more general gather that's part of sm5, which returns the full vector value (instead of just the red component as in sm4.1) and supports a slew of new sampling functions to boot.

11-15-2008, 08:25 AM
Thumbs up for AMD.