Repository navigation
Conversation
…t tensor. Allocate the final tensor directly and copy contiguous rows through the executor.
This branch has not been deployed
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Constant Pad builds an ndarray and then copies the whole result into a tract tensor. Allocate the final tensor directly and copy contiguous input runs through the current executor, preserving signed zero and quantized metadata.
Related to #2739 and #2899. The broader OCR/ORT performance gap remains open.
M2, Criterion estimates for f32
[1,3,640,640]with one cell of H/W padding:Reproduce with
cargo bench -p tract-core --profile opt-no-lto --bench pad, saving/comparing Criterion baselines. These timings include allocation and output construction.Validation:
cargo fmt --all,cargo clippy --workspace,cargo test -p test-unit-core pad --profile opt-no-lto, and exact-output NNEF checks in serial and with four workers on M2 and x86. Tests include empty dimensions, asymmetric/all-axis padding, multiple numeric types, signed zero, and quantized metadata.🍍