ValueError when tuning dense operators according to x86 tutorial

Troubleshooting
#1

Issue description

When auto-tuning “dense” operators by replacing “conv2d” in the tutorial “Auto-tuning a convolutional network for x86 CPU” (https://docs.tvm.ai/tutorials/autotvm/tune_relay_x86.html), the tuning fails with ValueError “Cannot find infer layout for task”.

Steps to reproduce the issue

  1. Prepare hardware and environment that meet the requirements for TVM auto-tuning on an x86 CPU
  2. Replace “conv2d” with “dense” in the code in the auto-tuning tutorial for x86 (https://docs.tvm.ai/tutorials/autotvm/tune_relay_x86.html)
  3. Execute the code as modified in step 2, in the environment prepared in step 1

What’s the expected result?

  • Auto-tuning of “dense” operators according to the x86 tutorial succeeds without errors

What’s the actual result?

  • Auto-tuning of “dense” operators according to the x86 tutorial fails with ValueError “Cannot find infer layout for task”

Additional details

  • Replacing “conv2d” with “dense” in the tutorial for NVIDIA GPU produces no errors and the tuning succeeds
  • The tutorial for x86 uses graph tuner (https://docs.tvm.ai/tutorials/autotvm/tune_relay_x86.html#configure-tensor-tuning-settings-and-create-tasks)
  • The tutorials for ARM CPU and NVIDIA GPU use autotvm.record.pick_best() instead of graph tuner (ARM CPU: https://tvm.apache.org/docs/tutorials/autotvm/tune_relay_arm.html#begin-tuning, NVIDIA GPU: https://docs.tvm.ai/tutorials/autotvm/tune_relay_cuda.html#begin-tuning)
  • Comments in the source code of graph tuner appear to imply that it supports dense operators (https://github.com/apache/incubator-tvm/blob/afc239aeb870d5c0a25a3e3e8e8c838f7122d9cf/python/tvm/autotvm/graph_tuner/base_graph_tuner.py#L161)
  • Yet get_infer_layout() throws the aforementioned ValueError for everything except conv2d and depthwise_conv2d

Suggested solutions

Any or all of the following:

  • Amend x86 auto-tuning tutorial to include pick_best() option and comments about tuning operators other than conv2d
  • Amend source code comment in graph tuner to make clear what operators are actually supported
  • Add dense operator support to graph tuner