Which parts of the GPU exactly process 2x2 pixels at once?

[Kaotik]

Is it the ROPs or which?

Trying to get 100% clear on the "under 4 pixel polygons get wasted" (Huddys statement about 75% of work going to waste at max settings tesselation in StoneGiant)

 

[bmcnett]

a quad is 2x2 pixels

The pixel shader always processes 2 X 2 groups of pixels (called "quads") but there is no other part of a GPU that necessarily does. Quads exist for the purpose of mip mapped texture filtering; if you compute texture coordinates for each pixel in a 2 X 2 quad, you can do a little bit of math on the four coordinates to determine which mip maps should be sampled.

This is one of many reasons why a GPU is poor at rendering tiny triangles. A one-pixel triangle is shaded as a 2 X 2 quad, which wastes 75% of pixel shader resources.

 

[ccanan]

nice reply

 

[Kaotik]

Cheers, that cleared it up perfectly

 

[Grall]

If extensive tesselation support starts catching hold, what's the chance of future GPUs dropping the pixel quad concept entirely and shading only individual pixels?

 

[AlexV]

If extensive tesselation support starts catching hold, what's the chance of future GPUs dropping the pixel quad concept entirely and shading only individual pixels?

It's more likely that they'll do something along the lines of quad-fragment-merging.

 

[sebbbi]

If extensive tesselation support starts catching hold, what's the chance of future GPUs dropping the pixel quad concept entirely and shading only individual pixels?

If most rendered triangles are one pixel (or sub-pixel) sized, it's better just to render the points as a point cloud, and drop the triangles (polygons) entirely.

 

[ebola]

If most rendered triangles are one pixel (or sub-pixel) sized, it's better just to render the points as a point cloud, and drop the triangles (polygons) entirely.

i heard they looked into this sort of thing during the ps3's development - and you still need a triangle concept for filtering "or it looks like a mosaic.."

 

[fellix]

Other factors in the fragment pipeline would make shading of tiny primitives inefficient long before hitting the quad boundary, like rasterization granularity and batch-size limitations.