| commit | c087bfb61d958e022f6229973b2c144310616053 | [log] [tgz] |
|---|---|---|
| author | MaheshRavishankar <1663364+MaheshRavishankar@users.noreply.github.com> | Thu Feb 22 23:18:21 2024 -0800 |
| committer | GitHub <noreply@github.com> | Thu Feb 22 23:18:21 2024 -0800 |
| tree | 083f76dc1614fbbe9d9254cc699c9ed15db2ab42 | |
| parent | 11089e899da87054637404a0484074ca593c22da [diff] |
Add support for using PDL to replicate the functionality in MLP sample that uses Transform dialect. (#16453)
This PR adds a sample that uses PDL to match a subgraph corresponding to
MLP and replaces with a `flow.dispatch`, that invokes an external
function which is provided by a system plugin.
To enable this an new pass `--iree-preprocessing-apply-pdl-patterns` is
added that has an option to read in the PDL pattern file and applies it
to the input program.
To support this a custom rewrite function `rewriteAsFlowDispatch` is
added that takes as arguments
- the root of the matched DAG (this is replaced by the matcher)
- A list of values that represent the dynamic dimensions of the results
of the root
- The name of the external function provided by the plugin
- The operands to the external function.
What is missing is the support to specify the workload and number of
workgroups to use while invoking the external function. This could be
solved by having a custom PDL operation (if possible) that accepts the
workload and a region that computes the number of workgroups based on
the workload. For now that is not handled, and the nubmer of workgroups
is set to `{1, 1, 1}`. This is still a useful thing to
prototype/checkpoint, but for any reasonable deployment this needs to be
fixed.
This PR adds a sample that matches the input in TOSA dialect. Due to the
TOSA dialect definition, the matmul now has a batch dimension as well.
To be possible to use the same plugin implementation, the `llvm.bareptr`
calling convention is used for the external function so that the inputs
(outputs) are passed (passed by reference) using pointer, offset only,
and `memref.extract_strided_metadata` is used to extract this
information from the multi-dimensional memrefs within the dispatch.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.