Hi TVM community,
I am trying to use autotvm/tvm to autotune a quantized gemm operation using an external library, FBGEMM. I have two operators, one which packs the weight matrix, and the second which does the gemm using the packed weights. For each configuration of the autotvm search space, I need to pack weights before calling gemm. Is there a way to have two ops as part of a single autotvm task, and have timing printed separately?