OpenCL grouped convolutions "did you forget to bind"

[Wheest]

It's possible that this is a usage question rather than a bug.

However, I have been trying to run grouped convolutions on OpenCL devices, using models imported from ONNX.

OpenCL models work when the number of groups==1, and when groups==input_channels, but get errors like Direct host side access to device memory is detected in fused_nn_conv2d_9 for other values of groups.

Interestingly, varying the optimisation level changes the error message slightly, but does not fix results.

I have an MWE gist notebook that shows the problem here.

Prior issues with grouped convolutions were dealt with in #3070.

Perhaps grouped convolutions in OpenCL require a specific data layout? However I'm unsure how I would specify this for imported ONNX models.

[vinx13]

group conv2d is not fully supported, we don't have schedule on CUDA/OpenCL. Currently only int8 group conv2d on CUDA is supported.

[tqchen]

@vinx13 can we add a more specific exception for other cases for now?