Untilize with unpadding only supports parallelization over the height #17537
Assignees
Labels
bug
Something isn't working
Comments
6 tasks
nardoTT
added a commit
that referenced
this issue
Feb 5, 2025
…adding (#17538) ### Ticket Link to Github Issue #17537 ### Problem description Currently, the untilize with unpadding implementation supports parallelization only along the height dimension. This affects perf for wide tensors, as they are mapped to a limited number of cores. ### What's changed In this PR, we introduce support for parallelizing the untiling operation along the width dimension, similar to tilize with padding. The operation executes the parallelization over the dimension with the larger number of tiles. In future versions: - we want the operation to support the parallelization along both dimensions simultaneously - we want the compute kernel to support the processing of an entire column block at once instead of one tile at a time For the tests added in test_to_layout.py, the kernel duration of the previous implementation is around 1.8 to 24.8 times larger than the current implementation ### Checklist - [x] Post commit CI passes https://github.com/tenstorrent/tt-metal/actions/runs/13121055787 - [ ] Blackhole Post commit (if applicable) - [ ] Model regression CI testing passes (if applicable) - [ ] Device performance regression CI testing passes (if applicable) - [ ] **(For models and ops writers)** Full [new models](https://github.com/tenstorrent/tt-metal/actions/workflows/full-new-models-suite.yaml) tests passes - [ ] New/Existing tests provide coverage for changes
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Describe the bug
The untilize with unpadding operation parallelizes the tensors along the height. Wide tensors are mapped to few cores only which affects performance
To Reproduce
Profile any wide tensor with untilize with unpadding operation and check the number of cores
Expected behavior
Using more cores for wide tensors
The text was updated successfully, but these errors were encountered: