The RADV Vulkan driver’s ray-tracing performance has improved a lot over time such as shown within yesterday’s RADV vs. AMDVLK performance comparison on Strix Point. Coincidentally, merged today is yet another ray-tracing optimization to benefit RDNA3 (GFX11) and newer AMD graphics processors.
A six month old merge request by Natalie Vock was merged today for Mesa 25.2. This merge for the RADV driver enables using pointer flags on RDNA3 graphics processors and newer. Pointer flags for the newer AMD GPUs allow the hardware to do some of the culling work plus early-cull box nodes with CullOpaque/CullNonOpaque ray masks when all children are (not) opaque.
Natalie Vock explained in the merge request:
“AMD hardware can early-cull box nodes if all leaves are either opaque or not and the ray flags are set to discard (non-)opaque geometries. This works even across TLAS/BLAS boundaries.
Propagate info on whether all child nodes are opaque or not through the BVH to allow RADV to set these flags per box node.
Then, enable pointer flags on RADV/gfx11+ in the BVH descriptor and pass the ray flags through to the intersection instruction.”
There aren’t any gaming benchmark numbers noted in the merge request but Arseny Kapoulkine did comment and provide some details on the performance benefits:
“Before this MR, trace took ~4.45ms with cullMask = 0x1, and ~5.39ms with cullMask = 0xff on 7900 GRE
After this MR, trace took ~4.40ms with cullMask = 0x1, and ~4.35ms with cullMask = 0xff on 7900 GRE”
This work and more will be found in the Mesa 25.2 release due out next quarter.
