Add multisampled antialiasing #377
Conversation
This is ported from the multi branch. Configuration of antialiasing mode is currently set statically, but could become more dynamic. In addition, the mask LUT is computed and uploaded every frame, rather than being persistent.
shader/fine.wgsl
Outdated
| #ifdef have_uniform | ||
| let total = workgroupUniformLoad(&sh_count[slice_size - 1u]); | ||
| #else | ||
| workgroupBarrier(); | ||
| let total = sh_count[slice_size - 1u]; | ||
| #endif |
There was a problem hiding this comment.
I think we always use workgroupUniformLoad in the other shaders now, so you can do that here too. Then again I'm not sure if the second barrier that it injects is useful in this case.
There was a problem hiding this comment.
I think we need it to pass uniformity analysis. In any case, this was a holdover from the original multi branch being on an older wgpu that didn't have it, so good catch.
src/mask.rs
Outdated
There was a problem hiding this comment.
Would you mind moving this into crates/encoding? That way this utility can be available to any renderer that desires it. It doesn't pertain to the scene encoding per se but it is related to managing GPU buffers which is currently a feature of the encoding crate.
I'm also happy to move it over in a follow up if you don't have time. I'm going to start working on enabling this in the crates once the PR lands.
| let dy_dxdy = dy / (dx + dy); | ||
| let a = dx / (dx + dy); |
There was a problem hiding this comment.
You may want to define and reuse idxdy = 1.0 / (dx + dy) to avoid computing it twice. An optimizing shader compiler could reduce the divisor in both lines into one computation but I don't know if that's guaranteed.
There was a problem hiding this comment.
I've added a TODO here. When #378 is fixed, we'll want to apply it here as well.
An optimizing compiler could do it, but one that satisfied IEEE arithmetic could not.
|
I'm seeing artifacts on cartioid with both msaa8 and msaa16. I don't know if I want to necessarily block this PR on this as it can be investigated in a follow up. Unless you want to fix that in this PR, consider setting the default static AA config to
|
|
I think the artifacts you're seeing are because of only 4 bits for winding number accumulation, and that test case is a good one for slamming lots of winding number increments into a small area. I'm not 100% sure though. An obvious way to validate that would be to try an 8 bit accumulator. In any case, agree with keeping area as the default and considering this a followup. |
Minor cleanup of the multisampling logic, and switch default to Area.
This is ported from the multi branch.
Configuration of antialiasing mode is currently set statically, but could become more dynamic. In addition, the mask LUT is computed and uploaded every frame, rather than being persistent.