Closed debadeepta closed 1 year ago
there is no plan to provide generic performance models, and matmul_perf_model
is more intended as an example than anything else :) But the autotuner should be generic enough that users should be able to plug-in their own model if they want.
"‘perf_model’: performance model used to predicate running time with different configs, returns running time."
I saw in the code that there is a matrix multiplication performance estimator. Are there any others? For e.g. it is increasingly common in Neural Architecture Search papers to use some function approximator (e.g. RandomForest, Neural Network) etc as a performance estimator. These can also be configured to output uncertainty with multiple independent instances.
Say for a particular GPU one could imagine having a library of such estimators that can be periodically (or online) updated as more configurations are tried out.