Open powderluv opened 1 year ago
This check is newly enforced. I'd assume we can vectorize depthwise_conv_2d_nchw_chw
no problem given the vectorization works on convolution op interface. Can we have the problematic dispatch so we can take a look?
@powderluv Can you provide the dispatch? Any additional info so far?
I have the .mlir for dispatch 28 (the failing dispatch) here: gs://shark-public/ean/reproducers/nvidia_mit-b0/module_forward_dispatch_28_vulkan_spirv_fb.mlir
The error can be reproduced with the following CLI input:
iree-compile --iree-hal-target-backends=vulkan module_forward_dispatch_28_vulkan_spirv_fb.mlir -o
mit-b0_dispatch_28.vmfb
We should probably add one of these models to IREE's CI so we have coverage there and doesn't escape to the SHARK tests.
@antiagainst could you assigned a priority here so that we can considered that this has been screened?
The reason is that we are missing a pattern to convert 2-D conv to 1-D for vectorization for this specific format. See https://github.com/llvm/llvm-project/blob/2b7ded2/mlir/lib/Dialect/Linalg/Transforms/Transforms.cpp#L1746-L1750. We just need to add a similar pattern to the existing one for nhwc-hwc format.
I won't be able to come to this soon. @powderluv is there somebody in your team that can add it? Should be fairly straightforward given we have an example there.
@powderluv Have to set this as a P2 on our side - if you have anyone to pick it up please feel free to assign!
What happened?
Recently we have been seeing an error like:
Will triage more but just wanted to file the issue in case it is known.
CI failures: https://github.com/nod-ai/SHARK/actions/runs/5171513432/jobs/9315382588
Steps to reproduce your issue
What component(s) does this issue relate to?
No response
Version information
No response
Additional context
No response