Benchmarking Criteria

Questions
1

Dear @tqchen , I’m new to TVM compiler and also to DNN. Just exploring a bit of source and benchmarking criteria. I have couple of doubts regarding the same…

  1. Is there any documentation which stats what all passes are enabled for which opt_level? OR any API to dump all passes list, I could find “PrintIR()” for dumping IR.
  2. What are the benchmarking criteria/parameters at pass level or optimization level? I have explored the benchmarking from TVM docs, however I could see only the execution time as a overall parameter.

Regards,

2

Hi,

@srkreddy1238 and @FrozenGene can you please help me with these simple queries?

Regards,

3

re 1. you could refer this link to get all passes at different level https://docs.tvm.ai/api/python/relay/transform.html

re 2. current default opt level is 2, but we use 3 normally. You could see the directory of benchmark. (We should upgrade NNVM to Relay in this benchmark directory)

4

Thank you @FrozenGene. It was helpful.

Regards,