With some work currently pending to be pulled into Mesa, the open source AMD RADV driver could be about to get much faster with Ray Tracing which is exciting to see.
As noticed by Phoronix with the "Batch acceleration structure builds" merge request:
This converts radv_CmdBuildAccelerationStructuresKHR to a simple shim that pushes the actual build commands to a queue, where they are accumulated and dispatched as late as possible.
This helps especially with games that don't do any build command batching of their own. For example, it triples the performance of Hitman 3.
Then in one of the comments from testing it:
the perf uplift from this PR is huge in a few games, out of the ones I own lego builders journey gets a 2x improvement in performance. Control gets an additional ~5 fps at 1080p, Minecraft RTX (education edition) gets ~10 fps more. When paired with the monolithic pipeline MR it completely blows amdvlk out of water.
Tested on my rx6800.
So if these two merge requests are pulled in, AMD GPU owners could expect to see a rather nice performance uplift in quite a number of games in later release of Mesa. So we could potentially see this in Mesa 23.3 due later on this year.
As a recent convert over to team AMD GPU with a Radeon 6800 XT, I'm going to be keeping a closer eye on stuff like this.
Update: turns out this might not get merged into Mesa, as there's also an already merged upgrade for VKD3D-Proton that does the Direct3D 12 to Vulkan translation that does some of the same as the MR noted:
Significantly improves Raytracing performance on RADV in a number of games, especially Hitman 3.
Handling WriteBufferImmediate is a bit janky as always, since we want to maintain correct command ordering with
BOTTOM
markers.
Either way, things are about to get speedy. Exciting!
It's kind of funny that the author of the MR suggests that it probably won't get merged because it only benefits D3D12 and vkd3d already has an MR that can do the same thing. Could be two separate projects solving the same problem at the same time.
They have both communicated together while it was going on so both projects weren't working on it independently without knowledge of the other. The mesa work started first and a MR was opened but after some talk it was decided it made more sense for it to be implemented in vkd3d-proton for now.
As a recent convert over to team AMD GPU with a Radeon 6800 XT
Congrats! You decided not to get the latest RDNA 3 cards?
Last edited by Shmerl on 14 August 2023 at 4:27 pm UTC
See more from me