Hi!
I think we should add an -mtune option to the global CFLAGS since the vast
majority of AArch64 CPUs are either Cortex-A53, Cortex-A57 or a big.LITTLE
combination of them. And I believe that a proper option to be added is
-mtune=cortex-a57.cortex-a53.
My opinion is based on the performance results I got building Spec2006 with
Tizen toolchain and running it on Juno board (which is equipped with a
big.LITTLE CPU). I tried options -mtune=cortex-a53, -mtune=cortex-a57 and
-mtune=cortex-a57.cortex-a53. Then, I executed the benchmarks separately
on A53
and A57 cores. Here are the results:
| config | INT/a53 | FP/a53 | INT/a57 | FP/a57 |
|---------------+---------+--------+---------+--------|
| O2-a53 | -0.0% | -0.0% | -0.0% | -0.0% |
| O2-a57 | -0.0% | +0.5% | -0.5% | -2.1% |
| O2-a57.a53 | +0.5% | +0.3% | +0.8% | +1.5% |
As you can see, the code compiled with -mtune=cortex-a57.cortex-a53 is
faster
than with any other tuning options. Surprisingly, it is even faster on
an A53
core than the code compiled with -mtune=cortex-a53, and faster on a A57 core
than the code compiled with -mtune=cortex-a57.
Thanks,
Nikolai Bozhenov
_______________________________________________
Dev mailing list
[email protected]
https://lists.tizen.org/listinfo/dev
