My (Dominic) experiments with profiling for the code generator were really interesting and we should keep up doing them once in a while to remove some bottlenecks. The setup is as follows:
- Make a build directory with
- Build a target
- Look for the .prof file with the same name as a given generated operator file
- Open it with the python tool
Dynamic exploration of the profile needs a bit of experienced but is very insightful. Often, the "cutoff" needs to be reduced to "1/100" (whatever the measure of that is).