This code follows TVM tutorial: Optimizing Operators with Auto-scheduling.
First, set the environment variable PYTHONPATH:
export TVM_HOME=/path/to/tvm
export PYTHONPATH=$TVM_HOME/python:${PYTHONPATH}
Then, run tuner.py
to search for optimal schedules for each matmul input size. Specify the matrix sizes and target by setting -m
and -target
flags. For example, if you want to run search on Bert sizes with GPU, run the command:
python3 tuner.py -m ../benchmark_sizes/bert_large_matmul.txt -target cuda
After search is done, it automatically saves json and so files associate with each matrix size.