Re: neon acceleration via mmx/sse (was: Re: [Qemu-devel] CFP: 1st International QEMU Users Forum)

[Frédéric Pétrot]

Peter Maydell a écrit :
> 2010/11/28 Frédéric Pétrot <address@hidden>:
>> PS: We have indeed ourselves worked on the acceleration of the neon support
>>    (neon on mmx/sse instead of helpers)
> 
> Slight tangent, but: How well did you find that worked?
> Were you trying to retain bit-for-bit accuracy in the results?
> 
> -- PMM
> 
Ok, we worked on the integer neon only, as in integrated devices, we
prefer fixed point for energy efficiency reasons (no plug, no fan,
specialized applications).
It works quite well, on synthetic benchmarks (loops with a growing
number of simd instructions) we have a good speedup (4 to 6x)
compared to the helper approach.
Amdahl's law may well reduce this to almost nothing for most applications,
but for video decoding and the like, it may have a value.
We (do our best to) have an exact translation, and we checked against
the arm of a beagleboard for the behavior.
Incidentally, we have an interactive paper at DATE this year on this very
topic.
Don't hesitate to drop me a mail if you want to have a look at it.
Frédéric Pétrot
--
+-----------------------------------------------------------------+
| Frédéric Pétrot, Pr. ENSIMAG-TIMA/SLS   address@hidden |
| Phone : +33 4 76 57 48 70               Fluctuat  nec  mergitur |
| Mobile: +33 6 74 57 99 65               Ad augusta  per angusta |
| Fax   : +33 4 76 57 49 81               Eppur si muove          |
+-----------------------------------------------------------------+

-- 
+-----------------------------------------------------------------+
| Frédéric Pétrot, Pr. ENSIMAG-TIMA/SLS   address@hidden |
| Phone : +33 4 76 57 48 70               Fluctuat  nec  mergitur |
| Mobile: +33 6 74 57 99 65               Ad augusta  per angusta |
| Fax   : +33 4 76 57 49 81               Eppur si muove          |
+-----------------------------------------------------------------+