@apeforest Sorry to pick a random PR to comment:
How to run the benchmark test for a single operator? I tried the code snippet in https://github.com/apache/incubator-mxnet/blob/master/benchmark/opperf/README.md.

import mxnet as mx
from mxnet import nd

from benchmark.opperf.utils.benchmark_utils import run_performance_test

add_res = run_performance_test(nd.add, run_backward=True, dtype='float32', ctx=mx.cpu(),
                               inputs=[{"lhs": (1024, 1024),
                                        "rhs": (1024, 1024)}],
                               warmup=10, runs=25)

And just get below printings, no performance report:

$ python op_benchmark.py
INFO:root:Begin Benchmark - add
INFO:root:Complete Benchmark - add

Another question is, the script can only be executed under the root folder of MXNet, right? So users always need to clone MXNet source code to do the benchmarking?

@TaoLv thanks for giving opperf a try. Jumping in quick to answer your 2 doubts

The result is stored in add_res
So if you print add_res you will see the details.

Yes, as it currently stands opperf utility can only be used after cloning the repo and setting the PYTHONPATH to the repo.

Thank you @ChaiBapchya

So if you print add_res you will see the details.

It would be better to add print(add_res) explicitly to the code snippet in the README.md.

Yes, as it currently stands opperf utility can only be used after cloning the repo and setting the PYTHONPATH to the repo.

Either explicitly document this requirement in the README or remove this requirement so the users can run benchmark directly after mxnet is pip installed.

Either explicitly document this requirement in the README or remove this requirement so the users can run benchmark directly after mxnet is pip installed.

https://github.com/apache/incubator-mxnet/tree/master/benchmark/opperf#prerequisites

Doesn't this say "add path to your cloned MXNet repository to the PYTHONPATH"

I'll add a comment that "explicitly" asks to "clone" the repo.
Also I'll try to see if this requirement can be removed and user can run benchmark directly with pip. But that would involve broader (time-consuming possibly) discussion so will do it later.

It would be better to add print(add_res) explicitly to the code snippet in the README.md.

I will add that print() statement explicitly to README too.

Thanks for pointing out.

Thank you @ChaiBapchya

So if you print add_res you will see the details.

It would be better to add print(add_res) explicitly to the code snippet in the README.md.

Yes, as it currently stands opperf utility can only be used after cloning the repo and setting the PYTHONPATH to the repo.

Either explicitly document this requirement in the README or remove this requirement so the users can run benchmark directly after mxnet is pip installed.

@TaoLv As requested, add the 2 changes to my existing PR -
specifically this commit - 4228baa
Does this help?

@ChaiBapchya Thank you for the quick turn around. It looks good to me.