commit 9aaae342bde3a469e5156214669866fd4ba57531
author Max191 <44243577+Max191@users.noreply.github.com> Tue Jul 30 08:37:35 2024 -0700
committer GitHub <noreply@github.com> Tue Jul 30 11:37:35 2024 -0400
tree e7110f104aeaa9479a21e60231d448fee62f6f25
parent 2e6dbfab0eb9786a9716f017ece893eb42c9ecc9

[GlobalOpt] Transition SetEncoding to use round_dims_to and stop creating tensor.pad (#17931)

This PR removes all padding introduced by SetEncoding, and instead
places the `round_dims_to` attribute on encodings. The consequences of
this PR are:

1. `iree_encoding.upper_bound_tile_size` is no longer necessary, since
the information needed for stream buffer allocation is already encoded
in `round_dims_to`, and there is no tensor.pad that uses the upper bound
size anymore.
2. The `original_type` field in encodings is no longer necessary, since
the types of the encoded tensors are no longer padded by the upper bound
size. This means the types of the encoded tensors have directly the
shape sizes of the `original_type`.
3. With the encoded types no longer having padded shapes, the
`matmul_narrow_M` and `matmul_narrow_N` fields can also be removed,
since the narrow sizes can be taken from directly from the tensor
shapes.
4. Since `iree_encoding.upper_bound_tile_size` is no longer generated,
the `CPUMaterializeUpperBoundTileSizePass` is no longer necessary, so we
can remove it.

This PR simply removes the padding from SetEncoding, and updates tests
to reflect the new expected form, but some follow up PRs to remove the
above mentioned fields can now be easily done.

This also updates the `PadFactor` value for `SetEncoding` to be a
command line flag with a default value of `32`. The previous value was
`16`, but some microkernels require a tile size of `32`, so the
`round_dims_to` needs to support padding up to `32` in those cases.

---------

Signed-off-by: Max Dawkins <max.dawkins@gmail.com>