| commit | b20ba96f7a83711b873afdfc11675daac6c2b5ab | [log] [tgz] |
|---|---|---|
| author | MaheshRavishankar <1663364+MaheshRavishankar@users.noreply.github.com> | Fri Nov 11 11:00:29 2022 -0800 |
| committer | GitHub <noreply@github.com> | Fri Nov 11 19:00:29 2022 +0000 |
| tree | db82f76924f1ffc1e1b31cf581cb36e2c19201cc | |
| parent | 91bb1ba4607b91a51750ae7c4f9ddf506b600cf1 [diff] |
Add a pass to materialize the encoding during executable translation. (#11056) Under the data tiling approach, at the Flow level an iree_linalg_ext.encoding attribute is added to tensors to capture the need for a data layout transformation. During executable translation this encoding + information about backend can be used to decide how the encoding is to be materialized (using pack and unpack operations). This PR adds a pass to materialize the encoding for the CPU backends. It uses patterns from a similar pass in IREE::LinalgExt::LinalgExtDialect, in addition to adding legalizations for hal.interface.binding, flow.dispatch.tensor.load/store and type conversion for !flow.dispatch.tensor type. To facilitate this - A new class called MaterializeEncodingTypeConverter, that derives from TypeConverter, is added that is meant to be used with patterns that implement the materialization. - A new class called MaterializeConversionTarget, that derives from ConversionTarget, is added that is meant to be used with the Dialect Conversion framework to drive the materialization. - Both the above classes rely on a callback being implemented by the users of these class that given the encoding return the MaterializeEncodingInfo object that has everything needed to materialize the pack/unpack operations. - This patch also adds the lowering for the flow.dispatch.workgroup_count_from_set_encoding_op needed to compute the number of workgroups correctly, based on the workload that was captured at the Flow level. With these changes a linalg.matmul can be compiled with IREE using the --iree-flow-enable-data-tiling flag.
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.