On 2018-05-08 10:17 +0800, Feng Longda wrote: > This is just for broadwell, but if the processor is skylake or some > newer type, it can't full use of the processor new microcode. Jonathan means that if "-march=skylake" has only 0.22% performance increase than "-march=broadwell", you can just use "-march=broadwell" happily. > Does we have better solution for this kind of use case. The target_clones attribute should work. > It seems intel ICC provide this kind of solution. Where? I can't find that in ICC documentation. > 2018-05-08 3:26 GMT+08:00 Jonathan Wakely <jwakely.gcc@xxxxxxxxx>: > > On 7 May 2018 at 11:46, Xi Ruoyao wrote: > > > On 2018-05-07 18:30 +0800, Feng Longda wrote: > > > > If we set -march/tune to define processor, we can obtain some extra > > > > performance improvement. > > > > > > > > I want to run my application on different intel processor family, for > > > > example: broadwell, skylake, haswell, skylake-avx512. > > > > > > > > What should I set? Does It like the following ? > > > > > > > > -march=broadwell -mtune=intel -mmmx -msse -msse -msse2 -msse3 -mavx > > > > -mavx2 -mavx=512f -mavx512pf ... > > > > > > No. Then GCC would use AVX-512, and the program can't run on processors > > > without AVX-512. (You'll see "Illegal Instruction - Core Dumped"). > > > > > > You should use target_clone attribute. > > > c.f. https://gcc.gnu.org/onlinedocs/gcc-8.1.0/gcc/Common-Function-Attributes.html > > > > Or just use -march=broadwell and see if the performance is good enough. -- Xi Ruoyao <ryxi@xxxxxxxxxxxxxxxxx> School of Aerospace Science and Technology, Xidian University
