[WIP] TensorTiler2D #1870
Draft
[WIP] TensorTiler2D #1870
+4,230
−37
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
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Coverage ReportCreated: 2024-10-25 14:52Click here for information about interpreting this report.
Generated by llvm-cov -- llvm version 14.0.0 |
|
How hard is it to generalize from 2d? |
|
@fifield Great question but I'm not sure... This needs some work before it's ready for merge due to some complexities with translating between the sizes/strides generated by the TensorTiler2D and the limitations in sizes/sizes allowed by the hardware. My plan right now is to get a bit more confident about the problem space in 2D (which is also easier to visualize) before moving on to something more complex. |
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
|
|
||
| # Define tilers | ||
| c_tiler = TensorTiler2D(M, N, m * n_aie_rows, n) | ||
| c_iter = c_tiler.tile_iter(tile_repeat_step_horizontal=n_aie_cols, iter_step=tb_n_rows) |
There was a problem hiding this comment.
[black] reported by reviewdog 🐶
Suggested change
| c_iter = c_tiler.tile_iter(tile_repeat_step_horizontal=n_aie_cols, iter_step=tb_n_rows) | |
| c_iter = c_tiler.tile_iter( | |
| tile_repeat_step_horizontal=n_aie_cols, iter_step=tb_n_rows | |
| ) |
| print(f"Expected: {expected_c_tile}") | ||
| print(f"Actual: {c_tile}") | ||
|
|
||
| def matrix_whole_array_tiling_sweep(): |
There was a problem hiding this comment.
[black] reported by reviewdog 🐶
Suggested change
| def matrix_whole_array_tiling_sweep(): | |
| def matrix_whole_array_tiling_sweep(): |
| ) | ||
| return | ||
|
|
||
| if __name__ == "__main__": |
There was a problem hiding this comment.
[black] reported by reviewdog 🐶
Suggested change
| if __name__ == "__main__": | |
| if __name__ == "__main__": |
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.
Description
This PR introduces some classes to help with data transformations/tiling and data visualization. The main classes introduced are:
TensorTile: A class representing information for specifying a 'chunk' of data, suitable for use in describing anpu_dma_memcpy_ndoperation; also includes code for visualizing theTensorTile.TensorTile2DIter: An iterator class that producesTensorTiles based on anoffset_fnTensorTiler2D: A class for representing data tiling. Provides atile_iter()method to aid in generating iterative patterns of tiling over the tensor; includes code for visualizing the the transformation.I also include:
test/python/tensortilerdma_transpose,matrix_scalar_add,row_wise_bias_add) to show how to use theTensorTiler2Dprogramming_examples/basic/tiling_exploration) that use the device to create an access order map that can be compared to the map provided by theTensorTiler2D. These examples are meant to verify theTensorTiler2Dcan produce valid access patterns but also illustrate how to use the tensor tiling code.npu_dma_memcpy_ndto support taking aTensorTiledirectly instead of sizes/strides/offsetSome incidental things that got included in this PR:
TODO List
tiling_explorationexamplestile_repeatintile_iter()tile_stride(skips, for height and width) intile_iter()TensorTile2DIterand/ortiler_iter()where some iterations may be a 'partial' transformation (e.g., in some cases where factors do not divide evenly)Wish List (probably not included in this PR)
Notes
Credit to Andre for his really excellent tool here: https://andreroesti.com/data-layout-viz/data_layout.html
This is my attempt at making something similar that can be integrated into a development tool directly, with the additional step of generating sizes/strides for you.