Turbine Camp

[pdhirajkumarprasad]

Introduction

This page provides information on implementing complete support of ONNX operators in Shark/IREE front-end. This effort is a part of overall ONNX quality improvement tracked by #812

Pre-requisites

• Must read full ONNX
• Read torch-MLIR architecture
• Read add new torch op

Development Workspace

Create a fork of https://github.com/llvm/torch-mlir and follow build instructions to build a clone of your fork and push changes using Pull Requests

Guidelines

• Assign an ONNX Ops from the list below to your GitHub ID by hovering your mouse on it and clicking on the bullseye (that will show to the right) to create a new issue, assign that issue to yourself and use that to track your work
• Study (thoroughly) the specification of the operator you picked on onnx-ops
• Use alt_e2eshark test to create multiple tests to comprehensively test all features and attributes of the operator -- this also help you understand how the operator is expected to behave
• Test torch-mlir/iree using the tests created and make code changes in torch-mlir to have correct inference numeric for llvm-cpu target working, if an iree codgen issue, file an issue on iree and note that in issue you created and assigned to self for the op

Contact:

• @vivekkhandelwal1 (Indian Time Zone)
• @AmosLewis (US Pacific Time Zone)
• @zjgarvey (US Central Time Zone)
• @marbre (Central EU Time Zone)

Onnx Ops( Known missing features)

• AveragePool (auto_pad ) #835
• CastLike (saturate ) #840
• ConvInteger (auto_pad) #836
• ConvTranspose (auto_pad) #839
• GridSample (mode - Cubic, paddding_mode only zero is supported)
• MaxPool (auto_pad, storage_order, indices output) #843
• LpPool (auto_pad) #838
• MaxUnpool (output_shape) #858
• NonMaxSuppression (center_point_box ) #844
• ScatterElements (reduction : max, min, add) onnx.ScatterElements #823
• Resize (antialias, axes, keep_aspect_ratio_policy , extrapolation_value, tf_crop_and_resize, coord tf mode, bicubic mode) #842
• RoiAlign (roi max pooling without default, coordTfMode and sampling_ratio)
• Upsample (unsupported interpolation mode other than nearest, linear) #834
• SplitToSequence (keepdims) #837
• TfIdfVectorizer (IDF and TFIDF mode) #833
• DFT (onesided, axis input tensor) #841
• Col2Im (5D input tensor)

Onnx Ops( Known Quality Issues)

• AffineGrid
• Bernoulli
• Cast
• DequantizeLinear
• GRU Quality Issues #856
• MelWeightMatrix
• QuantizeLinear
• RandomUniformLike

Other tasks:

llvm/torch-mlir#3796

[knwng]

Seems LSTM also has quality issue. But it seems the testcases themselves are buggy(wrong # of inputs & outputs). Besides, RNN only has 1 testcase rn. I'm afraid it can't cover all the corner cases.

[vivekkhandelwal1]

Seems LSTM also has quality issue. But it seems the testcases themselves are buggy(wrong # of inputs & outputs). Besides, RNN only has 1 testcase rn. I'm afraid it can't cover all the corner cases.

Yeah, the issue for LSTM is already opened here #315.