[HdSt] Enhance the current implementation of batching in HdSt #528
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
With the current aggregation strategy, there is a chance to end up with one batch per draw item, which could affect rendering performance. For example, if you have a scene with many draw items, the draw items could map onto the same batchMap key which alternate into multiple different aggregated buffers. This PR provides managing multiple aggregated buffers for each batchMap key to reduce the number of draw batches. The test scene has - rprims: 8029 - instancer: 7429 (1 prototype for each instancer) - instances: 28132 - textures: 231 and everything is static with an animated camera. The current implementation generated - 156 draw batches, - 8029 draw items and it attempted to append a batch 3 times. This PR generated - 7 draw batches, - still 8029 draw items and it attempted to append a batch 11,319 times. The batchMap had 3 keys and the depth of each key's vector was 1, 5, and 1. The test scene was able to be rendered upto 10fps performance improvement.
|
GitHub #161889. |
pixar-oss
pushed a commit
that referenced
this pull request
Jul 17, 2019
Improve batching performance in Hydra GL, adding to PR #528. Thanks to Yoojin Jang for the initial idea. The draw item "key" used in the map accounts for the geometric shader, buffer array versions and material params. Two draw items sharing the same "key" could use resources that don't aggregate. This checks happens in HdSt_DrawBatch::Append. So, a "key" effectively maps to several batches. - Use a vector of batches as the value type in the map, instead of a single batch. - Keep track of the previous draw item's key and batch. If the current draw item's key matches the last one, attempt to append it to the batch. If that fails, create a new batch and append it to the key's batch. If the key doesn't match, create a new entry into the map. This should reduce the number of HdSt_DrawBatch's created to the necessary minimum, and improve GPU performance since more draw item's are bucketed into a batch than was the case earlier. (Internal change: 1987160)
|
Thanks so much @yjang for the awesome PR! This is now available in dev, we couldn't do a direct merge so we adapted it a bit. Please let us know if you run into any issues! |
AdamFelt
pushed a commit
to autodesk-forks/USD
that referenced
this pull request
Apr 16, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
With the current aggregation strategy, there is a chance to end up with one batch per draw item, which could affect rendering performance.
For example, if you have a scene with many draw items, the draw items could map onto the same batchMap key which alternate into multiple different aggregated buffers.
This PR provides managing multiple aggregated buffers for each batchMap key to reduce the number of draw batches.
The test scene has
and everything is static with an animated camera.
The current implementation generated
and it attempted to append a batch 3 times.
This PR generated
and it attempted to append a batch 11,319 times.
The batchMap had 3 keys and the depth of each key's vector was 1, 5, and 1.
The test scene was able to be rendered upto 10fps performance improvement.