I have not been able to use AutoTVM with Scan, for instance with a recurrent neural network. Is that possible? Are there any examples, for instance for something like https://github.com/apache/incubator-tvm/blob/master/topi/recipe/rnn/matexp.py?
Particularly is there any docs for this error:
RuntimeError: FLOP estimator fails for this operator. Error msg: Only support tvm.compute currently. Other ops like tvm.scan/tvm.extern is not supported. Please use cfg.add_flop to manually set FLOP for this operator
There hasn’t been a lot of experiences in Autotvm for scan. But the particular error message should be resolved by adding cfg.add_flops to the template. This is mainly used for flop reference and you can put in a rough number
Thanks. I was able to get it working, but still getting crushed by CuDNN even starting with the TOPI code. https://github.com/apache/incubator-tvm/blob/master/topi/recipe/rnn/matexp.py . RNNs seem very difficult to optimize.