Performance Profiling and Numeric Python

Performance Profiling and Numeric Python

• 1.
  Performance Profiling and Numeric Python BENWELLER
• 2.
  Brief Problem Statement Reading DIRECTLY off the slides  Never posts the slides  Video is available  Automatically grab all the slides  Quickly underwhelmed with speed  How to speed this up  Speed has never been an issue previously
• 3.
  Python isn’t theProblem*  The fault dear programmer isn’t in Python  But sometimes it is  Get around it  Async  C extension  Pypy (JIT everything)  Numba (just JIT certain methods)
• 4.
  Before Anything Else: Software Hierarchy of needs Fast enough? Tests Does the code work?
• 5.
  Initial Code/Data  Readingthe frames:  Data:  1920 x 1080 x 3  All pixels recorded in the video, then 3 values for each pixel (RGB)  Test for validation:
• 6.
  Version 0.0  Establisha baseline  Time: 788.85
• 7.
  Version 0.1  Removeall the redundant lines:  Time: 728.88
• 8.
  Version 0.15  ApplyNumba  Time: 200.35
• 9.
  Version 0.175  Realizeyour calculation is wrong!  Redo with new formula  Time: 582.29
• 10.
  Version 0.2  Applynumba  Very slow odd results  Time: 1020.90  Results: [4458465513457839.0, 2137656293717757.8, 3461590301822020.5, 2435933916096979.0, 1020423444981550.0 …]  Not bounded between 0 and 100
• 12.
  Version 0.75  Realizethere’s something wrong with broadcast operations?  Time: 549.63  Output was all in the valid range
• 13.
  Version 1.0  Applynumba in full force  Time: 186.37
• 14.
  What I learned Follow the hierarchy of needs!  If you don’t you will get burned  Tests/initial benchmarks are key  Slowest code first  Incrementally solve the next bit  Very rarely, break out specialized tools  Know your tools  Has anyone ran in to the problem I had with broadcasting?