This repository has been archived by the owner on Nov 17, 2023. It is now read-only.
Performance improvement in ToTensor GPU Kernel #14099
Merged
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Outdated suggestions cannot be applied.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.
Description
Earlier, we were using Kernel Launch/Map way of launching kernel to write common CPU and GPU code for ToTensor operator. However, I observed there are too many threads and blocks being launched with kernel causing significant performance implication.
To overcome, I wrote a separate CUDA kernel for GPU and moved out of Kernel launch/map.
Benchmarks below.
Benchmarks
Ran 1000 ToTensor operation on (512, 512, 3)
GPU
Before
('Average time per ToTensor 512,512,3 - ', 39.17948246002197)
After
('Average time per ToTensor 512,512,3 - ', 0.44632863998413086)
CPU
Before
('Average time per ToTensor 512,512,3 - ', 3.7258052825927734)
After
('Averagetime per ToTensor 512,512,3 - ', 1.8473007678985596)
Checklist
Essentials
Please feel free to remove inapplicable items for your PR.
Changes
@zhreshold @stu1130