Acceleration Structure setup using RayTracingAccelerationStructure.CullInstances #5988
Conversation
|
Hi! This comment will help you figure out which jobs to run before merging your PR. The suggestions are dynamic based on what files you have changed. HDRP Depending on the scope of your PR, you may need to run more jobs than what has been suggested. Please speak to your lead or a Graphics SDET (#devs-graphics-automation) if you are unsure. |
|
It appears that you made a non-draft PR! |
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
anisunity
requested review from
remi-chapelain,
anisunity and
sebastienlagarde
IonutNedelcuUnity
changed the title
|
did a pass on the PR, tests are passing locally, @remi-chapelain you can check it out |
|
@INedelcu any notes what to focus on for qa here? |
@TomasKiniulis if you can test a project that uses path tracing, make sure that path tracing convergence is reset when changing the Transform by moving the object around or changing Material properties. I tested this already locally but there are no graphics tests that do this exactly. You can also use one of the path tracing tests that we have in HDRP_DXR_Tests project, click on Play and play with the Material (change properties or just delete whole Materials) and move objects around. This should reset path tracing convergence. Here's a video - https://drive.google.com/file/d/1RSUtATyYybvRQpa15J9jNXAHEVdH0-WC/view |
There was a problem hiding this comment.
Here's what I tried:
- Adding pathtracer and testing the dirtiness of most change (transform, materials, volume...) to reset accumulation ✔️
- Checked performance in the HDRP template (not much difference, makes sense since building RTAS is probably not the bottleneck here)
- Tested performance (FPS) with huge number of objects ✔️ (details below)
- Let the project run in player and editor a bit to see if performance was staying the same ✔️
- Checked RaytracingBuildAccelerationStructure new markers timings. Going from 52.88ms to 4.02ms so a 13x decrease ! ✔️
Before
After
10k emissive sphere in a box.
Without PR: average frame time 225ms
With PR: average frame time 136ms
40% decrease ! ✔️
https://user-images.githubusercontent.com/57442369/137454155-6b7b8a30-434e-4650-9904-9e85028bf6f1.mp4
Complex static scene lighted only with emissive objects
Without PR: average frame time 196ms
With PR: average frame time: 99ms
50% decrease ! ✔️
…eleration structure in HDRaytraingManager.cs
…late the acceleration structure.
anisunity
force-pushed
the
HDRP/raytracing/rtas-cullinstances
branch
from
a8d3043
to
2066ada
Compare
|
Merged with master and did the formatting |
| public List<RayTracingInstanceCullingTest> instanceTestArray = new List<RayTracingInstanceCullingTest>(); | ||
|
|
||
| // Culling tests | ||
| RayTracingInstanceCullingTest ShT_CT = new RayTracingInstanceCullingTest(); |
There was a problem hiding this comment.
Sorry but from the comment I don't understand why this code is here?
But I guess "Test" here doesn't mean "Testing" right? i.e not part of our test framework
There was a problem hiding this comment.
yeah not part of the test framework, it's just the set of tests that are used to inject into the RTAS @sebastienlagarde
Use RayTracingAccelerationStructure.CullInstances function (latest trunk - 2022.a13) to filter Renderers and populate the acceleration structure used for ray tracing (DXR). This greatly increases CPU performance on the main thread when using ray tracing effects.
On my PC with i9 with 12 cores, the performance improvement can vary between 8X and 30X or more, depending on scene size.
On a quad core laptop the CPU speedup for acceleration structure setup on the HDRP template scene is around 16x.
More information about testing and CPU profiling is on this PR which was merged already to trunk : https://ono.unity3d.com/unity/unity/pull-request/132659/_/graphics/raytracing/rtas-cull-instances
HDRP DXR tests on trunk and this branch is green: https://unity-ci.cds.internal.unity3d.com/job/9294644