Profile Operators' running time

Questions
#1

Hi! I want to run a onnx model on tvm. According to the docs, it’s easy to use timeit to get the inference time of the model. However, I didn’t find a method to profile the operator’s runing time in a model. Such as, if the resNet has runed for about 2 seconds, I hope to know how long the time is used for calculating convolution during the 2 seconds; I found that pytorch provide a Class named Profiler, which is workful for me. So, Is there a tool like that in TVM?

#2

using debug_executor

#3

Yes see https://github.com/apache/tvm/blob/961a7c70d75c81503c8c1d7c2e0db66bac4a1859/tests/python/unittest/test_runtime_profiling.py#L100

#4

Thank you very much for the answer. I have found that debug_executor is working. But I was trying to run yolov5 modle which has some switches. I have to build with VM instead of graph runtime. So is there a way to calculating the operators running time while it’s build with VM?

#5

Thank you for the answer. I have found that debug_executor is working. But I was trying to run yolov5 modle which has some switches. I have to build with VM instead of graph runtime. So is there a way to calculating the operators running time while it’s build with VM?

#6

Yes https://github.com/apache/tvm/blob/961a7c70d75c81503c8c1d7c2e0db66bac4a1859/tests/python/unittest/test_runtime_profiling.py#L68

#7

Thank you very much！！！