765 Commits over 638 Days - 0.05cph!
- Clean up texture streaming jobs
- Fix wrong camera position in Streaming_FallbackSquaredDistanceJob
- Add various debugging ConVars
- Add "Indirect Instancing" UI Indicator
Experimental frame synchronisation ConVars
Fix monster mesh rendering issue
Reset some CodeGen to their state on /main
- Create buffers with correct flags according to `buffer_upload_mode`
- Set default buffer upload mode to be "indirect" (1)
- Disallow changing `buffer_upload_mode` during gameplay
un-gate indirect instancing for testing purposes
Fix build by scrapping the function pointer idea.
- Fix rendering issues due to various jobs executing for command indices that are no longer in use.
- Fix broken debug draw when damage region bitscan job is active.
- Split `surviving_command_counts` into `command_counts_after_culling` and `command_counts_after_reordering` for clarity.
- Create a universal IsCallVisible() function instead of the mess we had before
- Rewrite the buffer update logic to ensure a direct update only happens once per frame to make DX11 happy.
- Always update the entire buffer to make greedy GPU drivers happy.
- Add IGraphicsBuffer wrapper because Unity didn't do it so I can finally stop copy/pasting everything.
- Use a function pointer (👻) in damage region processing for similar reasions.
- Unlock the arrays in Update in case OnPreCull wasn't called.
- New debugging ConVars: `reorder_instances_job`, `damage_bitscan_job` and `buffer_upload_mode`
A whole marathon of Indirect Instancing changes:
- Schedule all jobs at the beginning of LateUpdate() for better worker utilisation.
- Further improve scheduling behavior by modeling actual dependencies between jobs using writes_to_* and reads_from_* handles.
- Allow instance reordering and command aggregation to run in paralell to give even more flexibility to the scheduler.
- Run the debug overlay (where possible) in parallel with the jobs to improve performance
- Avoid garbage allocation when scheduling the frustum culling job. (GeometryUtility.CalculateFrustumPlanes)
- Schedule all jobs using ScheduleByRef.
- Correctly set `[ReadOnly]` and `readonly` for all jobs to help Burst as much as possible.
- Make batch size tunable: `indirect_instancing.batches_per_worker`
- Add somewhat automatic instrumentation to CrudeProfiler via `using` statements.
- Add ProfilerMarker to mimic what Unity 6 is doing with `using` statements. (Still need to verify it this compiles out properly.)
- Add more profiler instrumentation.
- Rename a ton of things for clarity.
- Remove CommandBuffer. We won't need it where we're going.
Fix burst compile issues by bringing back `DamageRegion` struct
Jobify the damage region scanning for instance data
First implementation of hybrid monster meshes
Merge from /indirect_instancing/layers
Slightly optimise command reordering
Jobified instance reordering
Merge bitmapped_damage_tracking → indirect_instancing
Implement bitmapped damage tracking
Rebase on current /indirect_instancing
Clean up some failed experiments
Fix nothing rendering for real this time.
Fix nothing actually rendering with new Graphics.DrawMeshInstancedIndirect based draw call submission
Fix NRE in Indirect Instancing debug overlay
Add agressive optimisation attributes to the hot path of Indirect Instancing.
Rebase on /main (save 271)
Regenerate ConVars and force bounds to be always huge (experiment)
Add a third submission mode to indirect instancing which works through `Graphics.DrawMeshInstancedIndirect` and avoids copying to a command buffer.
Enable `indirect_instancing.command_buffer` to switch back to the old way.
Finalise SIMD terrain culling implementation:
- Fix oscillation/flicker bug by tracking individual ray progress properly.
- Adjust step size dynamically with altitude to reduce iterations.
- Limit overall iterations to ConVar `indirect_instancing.tc_max_iterations`.
- Ignore rays to instances that have been culled in a previous step.
Add more debug output to the indirect_instancing.debug overlay
Add new SIMD terrain culling implementation (WIP) providing ~3x speedup (up to 30x in certain cases)
Only allow StartMotion() when IsDynamic is set
Speed up motion list execution 30x by caching matrices and checking them for changes.
Regenerate ConsoleSystem.cs
Regenerate ConsoleSystem.cs
Merge other fixes from /indirect_instancing
Merge build fix from /indirect_instancing_fixes
Restore constant that I removed because I was an idiot.
Merge from /indirect_instancing_fixes
Fix multiple issues with LOD bounds/distance calculations:
- Never return zero-bounds for anything, try multiple things and fall back to 1 m³ as a last resort.
- Always use bounds (and center) of LOD0 for consensus between LOD states.
- Fix fallback renderer using transform position instead of bounds center for distance calculations.
