Confused about Comparing the Different Schedules

#1

I’m goning thtough the tutorial (Working with Operators Using Tensor Expression), example 1.
The results on my machine are here:

    Operator                  Timing             Performance
       numpy    4.4620100379688665e-06                   1.0
       naive            0.0122167051      2737.9376101899393
    parallel    0.015120362299999999       3388.688544251433
      vector    0.015216546000000001      3410.2446813245324

My questions are:
What dose the timing mean, execution time？
Take the case of vector as an example, why more Timing means better Performance?

#2

Timing should be kernel execution time in seconds. Performance is Timing of current row divided by timing of numpy, so lower values are better. The question is why, is tvm performing ~3000x worse than numpy. There most likely is something wrong with your target, are you by any chance using a cuda or opencl target.

1 Like