[cache] sometimes there is a cache miss #783
Description
When benchmarking, I like to trigger tritonbenchmark once with the full autotuning pipeline, wait a bit, and run the tritonbenchmark again to get the results. For the second run, I expect results to be read from cache so there won't be any expensive autotunings.
Recently, I noticed that this is not the case anymore -- meaning that sometimes the full autotuning pipeline would be triggered even in the second rerun. This starts happening after the autotuner update from Differential Evolution to Pattern Search, so that could be a possible reason. Also, it is hit or miss -- sometimes it works and sometimes it does not.