Add overview of ONNX operator (sets) supported

pixelspark commented 2 years ago

Is your feature request related to a problem? Please describe.

For users it would be very helpful to know which operators are supported.

Describe the solution you'd like

A table listing all ONNX operators, indicating the level of support in WONNX (complete, partial/incorrect or no implementation) and the shader file the implementation is in (for developers).

The full list can be found here. It seems WONNX currently implements:

Abs, Acos, Asin, Atan, Ceil, Cos, Cosh, Exp, Floor, Log, Round, Sign, Sin, Sinh, Sqrt, Tan, Tanh (endomorphism/map.wgsl)
Reshape, Dropout, Flatten, Squeeze, Softmax (endomorphism/copy.wgsl)
Add, And, Div, Equal, Greater, GreaterOrEqual, Less, LessOrEqual, Mod, Mul, Or, Sub (endomorphism/arithmetic.wgsl)
BatchNormalization (endomorphism/batchnormalization.wgsl)
Celu, Elu (endomorphism/activation.wgsl)
Concat (matrix/concat.wgsl)
MaxPool, AveragePool (Conv only support NxCxHxW for the moment.) (pool/aggregate.wgsl)
Conv, ConvRelu (Conv only support NxCxHxW for the moment.) (pool/conv_kernel_1.wgsl, pool/conv_kernel_3.wgsl, pool/conv.wgsl).
SqueezenetConvGroup (containers/SqueezenetConvGroup.wgsl) (Not sure if this is actually an ONNX operator?)
Gemm, MatMul (matrix/gemm_1.wgsl, matrix/gemm.wgsl)
Relu, Sigmoid, Softsign, Softplus, Clip (endomorphism/activation.wgsl)
Transpose (matrix/transpose.wgsl)

Describe alternatives you've considered

Additional context

pixelspark commented 2 years ago

Hm, seems you are still missing quite a bunch of ONNX ops? (e.g. Relu is not on the list).

haixuanTao commented 2 years ago

Yes, well, the thing is that:

I'm still trying to find the best way to do the templates WGSL for each OP. And so I tried to build Op that I really use so that I can do test and avoid tech debt and legacy.
Many Ops are actually not used in Inference such as batchnormalization or can be replaced using onnxsimplifier by more general Ops.
There is a lot of stuff required to build really efficient WGSL shader and my first iteration of Ops was 20x slower than the one I'm building currently. And I want to make sure to have a good enough grasp of every optimisation before rolling out a lot of Ops.

But I am going to had a couple of Op before the end of the week to enable yolo.

If you want to run a specific model with unsupported Op, let me know and I can prioritise its ops. :smile:

pixelspark commented 2 years ago

@haixuanTao I fully understand! I'm not trying to run any specific models (from the commit log I see you are trying to make YOLO work - already quite an impressive list, more so if YOLO works!), just suggesting things that could bring the project forward to a more generally usable state.

I think it would be very helpful to have a list somewhere of the ops that onnxsimplifier can output with the current implementation status. If we manage to (at some point) implement all of them you can basically say that there is support for any ONNX model (given you run through the simplifier first). A list like this allows contributors to easily see where they can make the most impact towards that goal (I have no experience with WGSL but might have a go at implementing one or two ops if I can find the time. Currently I have no idea however about what would be the most sought-after op to implement?).

haixuanTao commented 2 years ago

Yes, full Support for all onnx model is definitely an objective! :rocket:

I have searched but haven't found any list of onnx simplifier node... My general experience is that you can sort of guess what node can be simplified and what node can't.

I guess you can try to find a model you like within: https://github.com/onnx/models and see what node are missing when you try to run it.

I have made a guide on building new op within the Readme. You can check it out here: https://github.com/haixuanTao/wonnx#contribution-on-implementing-a-new-operator

pixelspark commented 2 years ago

I have made a guide on building new op within the Readme. You can check it out here: https://github.com/haixuanTao/wonnx#contribution-on-implementing-a-new-operator

This is brilliant, more projects should have this!

Really interested in trying this, hopefully after Christmas there will be some more free time to dive in to this :-)

webonnx / wonnx

Add overview of ONNX operator (sets) supported #17