search

arporter / habakkuk

Fortran code analysis for performance prediction
11 stars 0 forks source link

Support MAX/MIN intrinsics #13

Closed arporter closed 7 years ago

arporter commented 7 years ago

We don't currently recognise MAX and MIN as Fortran intrinsics. We need to add them (including an estimation of their cost in FLOPs and cycles).

arporter commented 7 years ago

I've extended dl_microbench to include benchmarks for MIN and MAX. Running these on an E5-1620v.2 with the Intel compiler shows that they do not count as a FLOP. Assuming that they pretty much consist of the CMP instruction, I've assumed they get executed on port 5 (Agner says it can be done on 0, 1, or 5 but 0 and 1 will generally be busy for the floating-point heavy codes we're looking at) of the Intel Ivy Bridge and cost 1 cycle.

arporter commented 7 years ago

Created branch support_min_max for work on this issue.

arporter commented 7 years ago

Both MIN and MAX can take an arbitrary no. of arguments (>= 2).

arporter commented 7 years ago

Only modified source files are dag.py and config_ivy_bridge.py. Coverage of latter is 100%. Coverage of former is 97%. Missed lines are due to FMA, latency and division operands. None due to this work.

arporter commented 7 years ago

All changed files are pep8 and pylint clean.

arporter commented 7 years ago

Branch merged to master. All tests pass. Closing issue.