Asynchronous Render List Generation and Frame-Independent Task Scheduling #2887

douira · 2024-11-22T20:00:25Z

Uses an octree to generate render lists independently of the, now asynchronous, slow graph search.

It runs the occlusion culler in a separate thread, which allows a large speedup in render list generation time. Correct culling results are ensured with a combination of different types of BFS, synchronous occlusion culling is used as a last resort if the camera teleports or moves extremely quickly.
Tasks are ordered based on a combination score of how long they've been pending, their distance from the camera, their type, and whether they're visible in the camera frustum. How many tasks can be scheduled is now independent of the frame rate and instead are limited based on their estimated duration and size.
There's an upload limit to ensure not too many tasks are submitted that will exceed the upload buffer's size. This isn't a hard limit and this PR doesn't implement a new way of handling task buffers, to avoid expanding the scope too far.

Testing has not shown regressions and generally frame rate has improved a little if a system was not limited by render list generation, and a lot if it was. (see testing thread)

Combination with #2886 will avoid even more of the asynchronous work since a graph search isn't necessary to schedule tasks with this PR. (will merge conflict, but should be easy to resolve)

Companion PR in Iris: IrisShaders/Iris#2539

Makes #2780 unnecessary, since the octree traversal is correctly ordered already.

…led yet, it just does the tree frustum test right after each bfs for testing purposes atm)

…g RenderSection objects, improve tree render list generation performance

…the tree, this greatly improves performance

…learer code

…ync bfs if there is very fast movement

…trees of varying accuracy to present as few sections as possible while not generating any errors when the camera is in motion.

… unloading of chunks

…ssues when the distance changes (for example under water)

- rebuild tasks are scheduled in a queue and pruned each frame - async culling tasks and results are classes in their own package - chunk rebuild tasks are prioritized based on their distance to the camera, their type, how long the task has been pending, and whether the section is currently visible (in the frustum) - generally cleaned up the update method in RSM

… its own "run" token

…plicity. also changed the frustum test to be before adding a section to the queue, and not after. from my measurements by looking at how long the OcclusionCuller call inside FrustumCullTask takes, this has no impact on performance.

…graph search culling

…ta all at once instead of for each section separately

… distance

…sk duration estimation

… douira:sharp-angle-traversal-occlusion

… variants of, add support for very tall worlds

… different tree implementation

…estimation, limit upload size based on previous mesh task result size or an estimate of it, the limit behavior changes depending on which type of upload buffer is used

… not automatically correctly ordered since they're not coming from a tree � Conflicts: � common/src/main/java/net/caffeinemc/mods/sodium/client/render/chunk/lists/VisibleChunkCollectorAsync.java

…her issues with (at least vertical) sorting though and even important zero-frame blocking tasks are not scheduled within the same frame

…tum task list update

This was referenced Nov 22, 2024

Integration with Asynchronous Render List Generation in Sodium IrisShaders/Iris#2539

Draft

Sort render lists for regions and sections after traversal (#2266) #2780

Merged

douira force-pushed the decoupled-frustum-test branch from 94885b5 to 158cd78 Compare November 25, 2024 19:12

douira added the T-enhancement Type: Enhancement label Nov 26, 2024

douira force-pushed the decoupled-frustum-test branch from 1ad3a3f to f969cbb Compare November 27, 2024 22:35

jellysquid3 added this to the Sodium 0.7 milestone Dec 2, 2024

douira added 24 commits December 7, 2024 21:23

documentation cleanup

7beb4f9

do frustum cull using a linear octree, WIP (BFS isn't actually decoup…

f98abcc

…led yet, it just does the tree frustum test right after each bfs for testing purposes atm)

use faster deinterleave with one fewer steps

79163a9

implement front-to-back ordering during tree traversal

8282957

refactor plumbing between RSM and the chunk renderer to avoid touchin…

b6a1237

…g RenderSection objects, improve tree render list generation performance

filter out sections that don't render anything before adding them to …

c183714

…the tree, this greatly improves performance

change tree read timing to average 100 samples

cf29615

doesn't need to be public anymore

bdf87a4

more reasonable flag mask ordering

f50e55d

take 2000 samples for measurement

82bb032

fix entity culling

beab645

split node iteration into separate loops, no change in performance, c…

8620f68

…learer code

use section coordinates for tree compatibility test

07fa1a0

remove unnecessary frustum test from entity culling

d594dd0

refactor update state management in RSM, add option to fall back to s…

6b80f22

…ync bfs if there is very fast movement

add better async bfs and tree management. It now calculates multiple …

d7fa6c8

…trees of varying accuracy to present as few sections as possible while not generating any errors when the camera is in motion.

prevent some NPEs that happen because of concurrency with loading and…

4d3a1e0

… unloading of chunks

comment

422c24a

increase tasks per thread per frame to 3, this seems to usually be ok

72ec071

do render distance and fog culling in the tree traversal to prevent i…

bbfc02e

…ssues when the distance changes (for example under water)

don't do frustum bfs if the frame was dirty to prevent unnecessary work

8a5defe

remove debug changes

de2828e

fix issues with occlusion culling by making the graph search generate…

9cfd218

… its own "run" token

douira added 29 commits December 7, 2024 21:24

added efficient approximative ray cast backed by a bitmap to improve …

473a550

…graph search culling

further improve speed of tree building by writing the reduced tree da…

5b0963e

…ta all at once instead of for each section separately

restore some info on the debug screen

91c43f6

fix section ordering

4ae5b82

improve how coordinate parameters are passed into section tree traversal

6726440

fix issues when camera is outside of world

000ae28

notes

5a489c5

fix crash when loading under water

f59ebf4

fix missing world when fog distance changes, and flickering of render…

329ec66

… distance

remove unused best type variable

d4b9698

fix broken cancellation behavior on sync bfs

3c5b01a

implement frame rate independent task scheduling with effort-based ta…

dd072e0

…sk duration estimation

add angle-based section visibility path occlusion, cherry picked from…

97d9adf

… douira:sharp-angle-traversal-occlusion

make inner tree in section tree static

d807d7e

refactor tree structures to be more self-contained and easier to make…

b327952

… variants of, add support for very tall worlds

fix rebase changes

317fa45

make traversable trees abstract so that they can be subclassed with a…

e68d426

… different tree implementation

ordering consistency

aacb6f3

fix incorrect multi forest size calculation

3099307

cleanup forest build distance handling

146c965

fix section sorting by correcting the camera offset

32b9527

update todos

80000c7

update debug info to be relevant and useful again

36978a8

refactor job effort estimation, add category-based meshing task size …

08191f7

…estimation, limit upload size based on previous mesh task result size or an estimate of it, the limit behavior changes depending on which type of upload buffer is used

sort render lists separately when doing sync rendering, since they're…

882e65c

… not automatically correctly ordered since they're not coming from a tree � Conflicts: � common/src/main/java/net/caffeinemc/mods/sodium/client/render/chunk/lists/VisibleChunkCollectorAsync.java

remove completed todos

e7229d3

fix some issues with very delayed section sorting. There are still ot…

b3d4b7c

…her issues with (at least vertical) sorting though and even important zero-frame blocking tasks are not scheduled within the same frame

fix sorting by using correct section pos calculation and improve frus…

df32670

…tum task list update

douira force-pushed the decoupled-frustum-test branch from ecdd040 to df32670 Compare December 7, 2024 22:31

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Asynchronous Render List Generation and Frame-Independent Task Scheduling #2887

Asynchronous Render List Generation and Frame-Independent Task Scheduling #2887

douira commented Nov 22, 2024 •

edited

Loading

Asynchronous Render List Generation and Frame-Independent Task Scheduling #2887

Are you sure you want to change the base?

Asynchronous Render List Generation and Frame-Independent Task Scheduling #2887

Conversation

douira commented Nov 22, 2024 • edited Loading

douira commented Nov 22, 2024 •

edited

Loading