| commit | 5c0434a50aacc93e19fbff5dc7e71007285cb8ba | [log] [tgz] |
|---|---|---|
| author | Han-Chung Wang <hanhan0912@gmail.com> | Thu Apr 04 13:59:35 2024 -0700 |
| committer | GitHub <noreply@github.com> | Thu Apr 04 13:59:35 2024 -0700 |
| tree | e3ebabf0ea411c1af9656c6aaab2a38c60464434 | |
| parent | 7f9062614b7da58d87db79fc665c749424635167 [diff] |
[CPU] Do not use pre-configured tile sizes as input vector sizes for x86 (#16692)
It is not reliable for some cases. E.g., the below snippet is extracted
from mobilebert_int8 model and the pre-configured tile sizes of unpack
op is `[8, 4]`. The tile sizes are driven by consumers. We are able to
tile tensor.unpack op even the tile sizes are not aligned with
inner_tiles. The tradeoff is to generate a stack buffer to hold all data
required by tensor.unpack op. It is TilingInterface implementation
details. The op will always result in `*x8x8` shape, and the vector
input sizes are `*x8x8`. In this context, the configured tile sizes are
incorrect for vectorization. It is only correct for tiling.
```
%unpack = tensor.unpack %extracted_slice
outer_dims_perm = [0, 1]
inner_dims_pos = [0, 1]
inner_tiles = [8, 8]
into %17 : tensor<1x1x8x8xi32> -> tensor<8x8xi32>
%19 = linalg.generic ... { ... } -> tensor<8x4xi8>
%pack = tensor.pack %19
outer_dims_perm = [0, 1]
inner_dims_pos = [0, 1]
inner_tiles = [8, 4]
into %extracted_slice_2 : tensor<8x4xi8> -> tensor<1x1x8x4xi8>
```
The revision triggers an issue in quant_matmul_fusion pipeline_tests. It
is added a long while ago, and it's being outdated. All the matmuls on
x86 path are data-tiled. Thus, we retire the test in the revision.IREE (Intermediate Representation Execution Environment, pronounced as “eerie”) is an MLIR-based end-to-end compiler and runtime that lowers Machine Learning (ML) models to a unified IR that scales up to meet the needs of the datacenter and down to satisfy the constraints and special considerations of mobile and edge deployments.
See our website for project details, user guides, and instructions on building from source.
IREE is still in its early phase. We have settled down on the overarching infrastructure and are actively improving various software components as well as project logistics. It is still quite far from ready for everyday use and is made available without any support at the moment. With that said, we welcome any kind of feedback on any communication channels!
See our website for more information.
IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.