Repository navigation
ENH: Investigate NPY_GCC_OPT_3 macro for other compilers #21164
Description
Activity
Hi Sebastian,
I've started digging into it. Can we hop in a meeting someday? I have a few questions related to this issue.
Reacted by Sebastian Berg@seberg
A year later, but not forgotten. Now, with more time on my hands!Is something like this that we'd be looking for?
#ifdef _MSC_VER #define NPY_OPT_3 __pragma(optimize("gt", on)) #elif defined(__GNUC__) || defined(__clang__) #define NPY_OPT_3 __attribute__((optimize("O3"))) #else #define NPY_OPT_3 #endif
I can set up a few environments (Docker, possibly, because of personal laziness...even though there's the overhead associated with it that would have to be considered) to benchmark MSVC, Clang, GCC, and ICC?
Let me know if that's similar to what you're looking for.
Yes, probably that is all there is to it. I think we now compile more parts with high optimization anyway, but there are places that should benefit.
I'll run some benchmarking mumbo jumbo here and post the results to see if the gain is worthwhile.
Since a very long time (predating clang, I believe), NumPy uses
NPY_GCC_OPT_3in a few places. This macro is useful to locally enable a high optimization level for functions we know should be optimized: tight, simple (usually 1-D) loops.However, due to its age, the macro only applies to
GCC(unless clang picks it up?). It would be nice to generalize the macro a bit to other compilers, probably using#pragmadepending on the compilers.This may need some care, since different compilers have different ideas of what
O3means, IIRC. So it may be that e.g. the Intel compiler enables unsafe fast-math when GCC does not.The task are:
#define(maybe renaming it)In some cases functions that currently use this, may end up as universal-intrinsics eventually. But I somewhat expect that this macro will stay useful for things where maximum performance is less important or just as a stop gap, because it adds no complexity.
EDIT: I expect this is a fairly nice sprintable thing to investigate, although it is best if an MSVC setup is available. (Clang likely supports the gcc attributes, but I am not sure. This is a useful reference probably: https://stackoverflow.com/questions/31373885/how-to-change-optimization-level-of-one-function.)