Next Ops to work on

[AlexandreEichenberger]

Idea: put here a quick comment to claim the operations that you care currently working on, so that we do not replicate work.
Can also add a request for new op.

[AlexandreEichenberger]

Working on compress

[tungld]

Working on NonMaxSuppression

[etiotto]

I am working on SpaceToDepth (#926) and DepthToSpace (#927).

[AlexandreEichenberger]

FYI, here are some of the benchmarks we are focusing on and that have ops that are not working yet.

high priority: (from model zoo)

• Roberta,
• Bertsquad (onehot),
• Bidaf (‘compress’, [edit: now worked on] ‘hardmax’, ‘categorymapper’ [edit: now worked on] ),
• yolo3 nonmaxsuppression [edit: now worked on]
• tigny-yolo3: ‘round’[edit: now supported], ‘nonmaxsuppression’ [edit: now worked on]

high priority: support compile models compiled to their lowest component (like RNNs not exported as high level ONNX ops). No crash.

medium prio: hugging face GBERTQnA

A list of ops currently not supported and present in Model Zoo are listed at the end of this issue #128

[etiotto]

I am going to look at categorymapper (#941).

[AlexandreEichenberger]

working on one hot to work with multiple types

[tungld]

working on Hardmax to support Bidaf. PR #950 (merged).

[chentong319]

working on Resize.

[mmoldawsky]

Working on IsNaN op

[etiotto]

Working on ScatterElements (needed by fasterrcnn-10.onnx, maskrcnn-10.onnx).
PR is #1352

Scatter is deprecated but we map it to ScatterElements. PR is #1337

[etiotto]

Working on ScatterND. PR is #1370

[etiotto]

Implemented GatherElements. PR is #1375.

[etiotto]

Working on GahterND. PR is #1382.

[AlexandreEichenberger]

Status of implemented ops are listed here now: https://github.com/onnx/onnx-mlir/blob/main/docs/SupportedONNXOps-cpu.md

[airMeng]

Hi, thank you for your excellent work!
I am quite new to MLIR so the questions may be stupid, please never mind. I see ArgMax is supported in onnx-mlir but ArgMin not, is there any special issue for ArgMin? If not, can I open a PR about ArgMin just based on ArgMax with little modification?

[tungld]

@airMeng please go ahead with a PR for ArgMin. Thank you!

[Ris-Bali]

Hi, can I work on celu op ?

[muzafferkal]

Somebody please support QuantizeLinear/DequantizeLinear ops for quantized networks.

[srcarroll]

is anyone working on extending the decompose-onnx pass (or a similar pass) to support more onnx.Custom ops? In particular i am trying to compile a basic backwards graphs and get InPlaceAccumulatorV2 custom op after converting onnx to mlir with onnx-mlir --EmitONNXIR.

[AlexandreEichenberger]

@srcarroll CustomOps is a non-standard op that @chentong319 added to easily convert a custom op into a function call. While Tong knows best, my recollection (and I may be wrong here) is that custom ops are mainly generated within onnx-mlir, and not parsed in from an ONNX protobuf.

Tong is away for a bit, if you wanted to add support for more custom ops, we would certainly be interested in taking in the changes. There are also ONNX functions, maybe that may help too.

[srcarroll]

@AlexandreEichenberger thanks for the response. i'd be happy add support, but I can't find any info on the definition of InPlaceAccumulatorV2. do you know where i can find that?

could you also point me to the ONNX functions you are referring to and how to emit them? thanks

[AlexandreEichenberger]

I could not find info about your "InPlaceAccumulatorV2", it does not appears in the ONNX specs or what I could find about the OnnxRuntimeExtensions, though I may have overlooked something in the ORT as I am not very familiar with it. How did you create the ONNX graph?

As far as creating custom ONNX functions, you can see a reference here in the ORT literature as a preferred way to make new custom ops: https://onnxruntime.ai/docs/reference/operators/add-custom-op.html

The ONNX specs also have a section on ONNX functions: https://onnx.ai/onnx/intro/concepts.html#functions