@sima It's condemned to linear for most scenarios. That said, that's not as bad as it sounds for dGPU+iGPU because you need to shove it across PCIe anyway so you may as well de-tile.
In theory, you should also be able to share tiled between two different NVIDIA GPUs if they're close enough in generation but there appear to still be kernel bugs there. Something to do with fencing.
@demofox There is not "gist". That's why it needs a blog post. 😂
The very short version is that the original implementaiton for nouveau that was donated by NVIDIA was deeply broken in several different ways. Also, the HW is tricky because of the way it stores tiling information in page tables.
With NVK, we're trying to sort all this out and do it properly but keep NVK and old GL playing nicely together. It's turning out to be a challenge.
Add comment