Multisample of non-uniform resolution Patent

[OpenGL guy]

Can the R300 switch between sampling patterns in real time?

Centroid sampling can be changed just like any other state. The application/driver can specify the sampling mode on a per primitive basis (i.e. every draw command could have a different setting). This is all documented in the PS 3.0 specs.

 

[K.I.L.E.R]

Can the R300 switch between sampling patterns in real time?

Centroid sampling can be changed just like any other state. The application/driver can specify the sampling mode on a per primitive basis (i.e. every draw command could have a different setting). This is all documented in the PS 3.0 specs.

I guess the answer is no.
R300 = PS 2.0.

 

[Xmas]

Can the R300 switch between sampling patterns in real time?

Centroid sampling can be changed just like any other state. The application/driver can specify the sampling mode on a per primitive basis (i.e. every draw command could have a different setting). This is all documented in the PS 3.0 specs.

Centroid is an instruction modifier, so changing it means using another shader. That should be at least as expensive as changing a global render state (if the shader is cached). But I really see no reason to toggle centroid sampling on and off if you can't do it per edge.

On the "blurry" seam - wouldn't that be more like supersampling, i.e. the shared edge pixels' color should be more precise than that of interior pixels? Or is the LOD too low to give an improvement?

 

[Dave H]

A minor one: the corresponding memory bandwidth hit.

Isn't the bandwidth hit the same for MSAA and SSAA the only difference is that with SSAA you get half the fillrate while with MSAA there is 0 hit on fillrate?

2x SSAA will cut your fillrate in half; 4x will cut it to one quarter, and so on.

For a naive implementation, the bandwidth hit for MSAA and SSAA is the same. But because with MSAA all sub-samples of a given non-edge pixel will have the same color, there are huge opportunities for lossless framebuffer compression, which both NV3x and R3x0 implement. (This is why NV3x and R3x0 take such a smaller performance hit from MSAA than NV2x, which didn't do framebuffer compression.)

 

[OpenGL guy]

On the "blurry" seam - wouldn't that be more like supersampling, i.e. the shared edge pixels' color should be more precise than that of interior pixels? Or is the LOD too low to give an improvement?

If the edge is an interior edge (like my quad rendered as two triangles), this extra filtering is often undesirable.