|
 |
Hello I noticed that SIMD optimization for for x86 only so I introduced
it for NEON also:
1. NEON-accelerated noise (AArch64): a hand-written ARM NEON kernel for
the core Perlin Noise/DNoise functions, plugged into the runtime noise
dispatch. ~6–13% faster on noise-heavy scenes, output pixel-identical.
2. SIMD BVH traversal (POV_BVH=bvh4): a contiguous SoA "BVH4" node
layout with a branchless 4-wide NEON ray-box slab test, applied to both
the scene-level bounding hierarchy and the per-mesh BVHs — scene trees
keep the best-first priority queue, mesh trees use a front-to-back
stack. ~1.9× on traversal-bound rendering (1.1–1.7× on typical GI/mesh
scenes, no change on non-BVH-bound ones), output pixel-identical.
Anybody cares for these? Main use case is Apple Silicon computers
Post a reply to this message
|
 |
|
 |
Alessio Sangalli <man### [at] gmail com> wrote:
> Hello I noticed that SIMD optimization for for x86 only so I introduced
> it for NEON also:
>
> 1. NEON-accelerated noise (AArch64): a hand-written ARM NEON kernel for
> the core Perlin Noise/DNoise functions, plugged into the runtime noise
> dispatch. ~6–13% faster on noise-heavy scenes, output pixel-identical.
>
> 2. SIMD BVH traversal (POV_BVH=bvh4): a contiguous SoA "BVH4" node
> layout with a branchless 4-wide NEON ray-box slab test, applied to both
> the scene-level bounding hierarchy and the per-mesh BVHs — scene trees
> keep the best-first priority queue, mesh trees use a front-to-back
> stack. ~1.9× on traversal-bound rendering (1.1–1.7× on typical GI/mesh
> scenes, no change on non-BVH-bound ones), output pixel-identical.
>
> Anybody cares for these? Main use case is Apple Silicon computers
Did you create a Git pull request? that's where official development usually
happens.
Post a reply to this message
|
 |