[Libav-user] gcc auto-vectorisation

[Claudio Freire]

On Mon, Feb 25, 2013 at 4:04 PM, "René J.V. Bertin" <rjvbertin at gmail.com> wrote:
>
>> René J.V. Bertin <rjvbertin at ...> writes:
>>
>>> The SIMD version ran twice as fast as the scalar version
>>> until I used gcc 4.7, which has auto-vectorisation
>>
>> Unfortunately, turning auto-vectorisation on triggers
>> bugs in gcc and is therefore no option;-(
>
> I've had some time to check up on this. On my system (a Macbook Pro 13" with a 2.6Ghz i7 running OS X 10.6.8), I only had 4 failing tests after building for i386 with MacPort's gcc-4.7 with -ftree-vectorize . However, if the duration of make check is any indication, auto-vectorisation does not give any performance benefit (or hit); for both, the tests (2nd make check call) take about 249 seconds.


That's because __builtin_assume_aligned isn't being called (most
likely, didn't check). That results in **far** sub-optimal
vectorization. I don't know about the failing tests though.