GCC 4.8 Compiler On AMD's Eight-Core Piledriver
Phoronix: GCC 4.8 Compiler On AMD's Eight-Core Piledriver
This month from CPUs based upon AMD's new Piledriver micro-architecture I have delivered results of compiler tuning on AMD's Open64 compiler as well as GCC bdver2 tuning. That initial testing from an AMD FX-8350 Eight-Core processor didn't show any big boost out of the "bdver2" target with the new BMI/TBM/F16C/FMA3 instruction set extensions. Testing in this article from the AMD FX-8350 are GCC compiler benchmarks of the 4.6.3, 4.7.2, and 4.8.0 development snapshots to look for performance improvements on this new high-end AMD processor when using the very latest GCC compiler code.
i dont see any software (and libraries) actually using FMA. Even AVX is pretty hard to find in libs. And XOP is dead-on-arrival. I have yet to see a software to use XOP to gain anything.
AVX is already 2 years old and not many softwares can benefit positively from it.
Does building linux kernel with march=native on Piledriver make use of all these fancy instructions ? IOW, can the linux kernel use AVX, FMA , XOP. Even SSE4.x ?
You shouldn't need to - if the compiler was aware enough, it would use those for you. FMA is extremely common in any code, but it seems gcc can't yet take advantage of it.
If you mean something very specific like x264's asm, then obviously they only use things that help in video encoding, etc.
Tags for this Thread