Repository navigation
Fixes the non-NVIDIA GPU issues reported in #2.
Bug fixes for timing accuracy, and some performance optimizations to the initial scene load.
Cluster LOD
- Read streamed cluster groups on worker threads, so page faults on the cluster cache no
longer stall the render thread. On Zorah with a cold file cache, this removes the
stalls of up to 100 ms per frame that lasted about 9 seconds while the cache paged in
after a load - Share BLASes across more instances, and skip LOD traversal for geometries with a single
LOD level - Make
--lpe-sweepavailable in default builds
UI and Profiling
- Average the profiler readouts over the last second, and skip the CPU frame time of the
first frame after a load or after the window regains focus - Add per-frame CPU, GPU and streaming times to the
--dump-statsoutput
Bug Fixes
- When DLSS was unavailable, the samplers were recreated every frame until the D3D12
sampler heap ran out and the device was removed. The sample now falls back to no
denoiser and grays out the DLSS control - CLAS and template builds reused the scratch size from the size query instead of the
size for the build mode - Cluster tiling ran group barriers in a lane-0 branch, and could drop surfaces when the
first wave of a thread group had no surface - Vulkan CLAS byte counters wrapped past 4 GB, returning wrong CLAS addresses from the
streaming CLAS compactor. It is off by default and enabled only by--linearallocin
RTXMG_DEV_FEATURESbuilds --dump-statstiming statistics included frames from before a shot-list reset, used a
different window for mean/min/max than for median/p95, and left the GPU frame and pass
timers empty on headless runs- A cluster group load that did not fit the CLAS build scratch asserted instead of
waiting for a later frame