Natalie Vock who is one of several Valve contractors working on the Linux graphics driver stack worked out these newest RADV ray-tracing optimizations.
The patch series adds support for gfx11/gfx12 ds_bvh_stack* instructions along with other GFX12 (RDNA4) ray-tracing performance improvements. GFX11 as a reminder is for RDNA3 and RDNA3.5 GPUs. Natalie explained in the merge request :
"Adds support for both the gfx11 ds_bvh_stack_rtn instruction as well as the gfx12 ds_bvh_stack_push8_pop1_rtn (for BVH8 traversal) instruction.
ds_bvh_push8_pop1 needs to take in the 8x32 BVH intersection result as a vector-aligned operand in order to avoid copies. This is not as big of a deal on gfx11 because we only forward the intersection result as a whole - on gfx12 we need to manipulate individual components.
The first two commits are rather trivial changes which would've been quite annoying to rebase around if I would split them out into other MRs, but they too contribute to a considerable perf improvement on gfx12.
I also included "aco/isel: Use vector-aligned operands for image_bvh8_intersect_ray" in this MR because it has support for vector-aligned operands tied to defs as a shared dependency (ds_bvh_push8_pop1 also depends on it) and is rather trivial in itself."
The real kicker though is the performance improvement for end-users with today's merge to Mesa 25.2:
"All combined q2rtx for example gets a 14% fps increase on gfx12 in my testing."
Around 14% better performance on AMD Radeon RX 9000 series compared to Mesa 25.2-devel up to this point, which already featured a number of ray-tracing optimizations. It will be interesting to see the ray-tracing benefits as well outside of Quake II RTX.
Just last week I featured a number of benchmarks in AMD Radeon RX 9070 Ray-Tracing Performance Improving With Mesa 25.2 that showed already the much improved Vulkan ray-tracing performance coming for RADV in this quarter's Mesa3D feature release. With today's merge, those numbers are now out-of-date and I am eager to run some new tests on AMD Radeon RX 7000 series and RX 9000 series GPUs.