[Brian00mc]

I've programmed the Neon on the Cortex A8 (using a mix of ARM, VFP and SIMD instructions). Is there any reason this code would not work on an A9 with Neon? I've heard the Neon's are different but transparent to the programmer.

Thanks!

[shervin]

Your ARM & VFP & NEON code from A8 should still work fine on A9. It will perform with very different speed, but it will do the same calculations atleast. The main complication is that not all A9 devices have NEON. If you know that your A9 has NEON then it will run your code from A8 without problem. Just don't expect the same speedup.

[Brian00mc]

Thankyou, Shervin.

It's been quoted that A9 runs at 2.5 Dry and A8 2.0 Drystone. Seems to imply a 25% improvement for A9. Is this in the ballpark?

Thanks Again
Brian

shervin, on 15 June 2012 - 12:53 AM, said:

Your ARM & VFP & NEON code from A8 should still work fine on A9. It will perform with very different speed, but it will do the same calculations atleast. The main complication is that not all A9 devices have NEON. If you know that your A9 has NEON then it will run your code from A8 without problem. Just don't expect the same speedup.

[shervin]

Actually no that is not typical at all! Cortex-A9 is a very different CPU than Cortex-A8, especially in the cache system and the regular CPU processing (ie: non-NEON). So from what I have seen, regular CPU code seems to run much better on A9 compared to A8 (probably mostly due to out-of-order speculative execution in A9 but not A8), but NEON does not necessarily run any better on A9 than A8 (in fact in some ways it is better in A8 than A9!). So I often see 10-30x boost by using NEON on A8 but only 2-5x boost by using NEON on A9. This doesn't mean that NEON is crippled on A9, it mostly means than regular CPU code runs badly on A8 but does much better on Cortex-A9, so NEON does not have as much to improve on in an A9 whereas it can improve A8 CPU code a lot!

-Shervin.