commit 537b07c4640d5c2e263aea849c60f60a64eca3ea
author Nicolas Vasilache <nicolasvasilache@users.noreply.github.com> Tue Feb 07 19:28:52 2023 +0100
committer GitHub <noreply@github.com> Tue Feb 07 10:28:52 2023 -0800
tree 71d9626f1764ee72a422d65a29b8d0d4c0be4a6f
parent 21d08080d5d18dda411a972bc0c66f464d8ebf3b

Add an example of generalized packing (#12076)

This revision demonstrates how the generalized packing transformation is
a one-size fits all implementation
for tiling linalg ops of rank N to linalg ops of higher rank.

The tensor.pack / tensor.unpack representation allows us to complete our
panoply of composable linalg transformations.
Now, one can either:
1. lower a linalg op to loops and use classical loop-based tiling
techniques (e.g. Allen&Kennedy, polyhdral/affine etc)
2. tile an N-D linalg op to N-D loops surrounding an N-D linalg op. This
often preserves the name of the linalg op and is at the basis of the
TilingInterface.
3. (this PR) tile an N-D linalg op to a 2*N-D linalg.generic. This step
requires that the tile dimensions divide the problem dimension.
tensor.pack / tensor.unpack provide this guarantee.

Step 3. can easily be adapted to produce a new named op (e.g. mmt4d)
when relevant, the point of this PR is to demonstrate generality.

This is related to discussion #12075.

An additional pattern is added to convert tensor.pack/unpack to
linalg_ext.pack/unpack until IREE adopts the upstream variants.
With this, it is possible to form dispatch regions without failing to
lower.

One thing to note in the IREE pass pipeline is that the
InterchangeGenericOps breaks the normalization property of the packing.
This can be recovered after the fact but it would be better to disable
in such cases if possible.

At this time, `iree-compile` fails on the `iree_linalg_ext` ops as it
wants statically allocated buffers to be hoisted to the top of the
function.

Both these issues can be left for a followup investigation.