Skip to content

Conversation

@kvark
Copy link
Owner

@kvark kvark commented Aug 22, 2024

Implements an experimental code path where the spatial reuse is only done within the workgroup.

TODO

  • workgroup-local spatial re-use
    • doesn't provide perf benefit on its own, surprisingly
    • facing a driver bug / GPU crash on AMD if the group is larger than 8x4
    • save temporal intermediate results
  • jittered workgroup mapping
    • looks to be helping greatly
    • debug visualization
  • redo the gbuffer changes, since the experiment was slower
  • remove extra buffers
    • gbuffer targets
    • reservoirs
  • find best temporal pixel in 2x2 grid
  • results
    • confirm faster convergence
    • test performance on NVidia, AMD, Intel
  • additional experiments

Debug visualization

pattern

Convergence

TODO

Performance

Nvidia GeForce 3050

Old (11ms, including the gbuf-fill, ray-trace, and temporal):
Local0

TODO

@kvark kvark force-pushed the local-restir branch 2 times, most recently from e2b9c63 to 410969b Compare August 23, 2024 06:40
@kvark kvark force-pushed the local-restir branch 2 times, most recently from c1051a3 to 7cc23f2 Compare September 3, 2024 05:00
@kvark kvark mentioned this pull request Sep 17, 2024
@kvark kvark added the type: experiment Experimental code label Sep 17, 2024
@kvark kvark mentioned this pull request Sep 19, 2024
6 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

type: experiment Experimental code

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants