HIPRT (MIT) behind the CMake option DLU_HIPRT (off): its headers come from its
SDK (HIPRT_ROOT, else ROCm's /opt/rocm) and are copied next to the servers; its
library is loaded when first used (hiprtew), as HIP or CUDA are by Orochi
(MIT, fetched pinned by hash; CUDA when its toolkit is found). The trace
kernels (nearest hit skipping the triangle a ray leaves, any hit) are compiled
the first time and kept in cache/hiprt. One GPU context for the process
(hiprt_device picks the GPU); the workers take turns on it. When HIPRT, the
GPU or a scene's upload fails, Embree is used instead, and the UGC server logs
why at start.
For a GPU the rays go in batches (UgcRays::Scene gets batch queries; the CPU
backends answer them a ray at a time):
- hidden faces: with a batch backend the paths are traced side by side, a
bounce at a time (the path code split into Start, Scatter and Bounce, the one
by one tracing unchanged); the same paths with the same random numbers, so
the same triangles are decided (tested with builtin side by side)
- the occlusion bake and the denoised icons' traced occlusion always ask in
batches (the same rays, the same results)
Its symbols are hidden: the servers export theirs (-rdynamic), and HIPRT's
library, which has an Orochi of its own, would otherwise call ours.
Check: configure with -DDLU_HIPRT=ON on a machine with ROCm (or HIPRT's SDK)
and an AMD RDNA or NVIDIA GPU; UgcServer --make-model x.lxfml out hiprt; the
UGC tests (hits, hidden faces and occlusion against builtin); a build without
it leaves everything as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
