Closed WeiqunZhang closed 4 months ago
I tested this with Castro and GCC 13.2 and I didn't see any measureable performance difference.
The hotspot functions in Castro are perhaps already inlined. Or they are so expensive that the cost of not inlining does not matter.
By default, set GCC inline limit to a huge number (43210). GCC's default value is version dependent. It appears to be smaller than 2000 in recent versions. The motivation for setting this to a huge number is that GCC especially the recent versions (e.g., 12) no longer inline big lambda functions passed to ParallelFor, resulting in serious performance issues.