[gst-devel] RE : Re: aligned(16) GstBuffer Payload ?

[Stefan Kost]

Hi,

SP GLE schrieb:
> Hi,
> thanks for all answers.
> 
> - I had seen the problem with posix_memalign/memalign requesting free
> to be used and probably incompatible with g_free(), so some parts of my
> code rely on posix_memalign + free(), but most of the processing is
> done outside of this code ::
> 
> I'm writing a server handling a lot of buffers and adding,
> substracting, multiplicating all samples in each buffer (160 samples
> each 20ms, for each member (it's a conference server, with a total of
>> 100 members)). Even if liboil is able to handle unaligned buffers it
> seems to use raw loops to compute (remaining) unaligned data. So any
> improvement is interesting .....
> 
> But switching GstBuffer to posix_memalign is not a good thing (IMHO), :
> - GstBuffers are often allocated for a limited time, there are groups
> of GstBuffers with the same size (eg before decoding, before encoding,
> after...etc), so relying on g_slice dramatticaly improves performances
> over malloc and posix_memalign. 
>
We don't use g_slice() and we should not, because if we allocate in one thread
and free in another, g_slice needs to use locking for the thread local pools.
Ideally we use freelists in all mini-object bases subclasses for the instances
and maybe extra one in GstBuffer for the buffer-memory.

Stefan

> 
> - Aligned buffers are interesting if the compute time gain is over
> posix_memalign cost (so if GstBuffer payload is often used with
> SSE/liboil).
> 
> Regards.
> 
> --- Stefan Kost <ensonic at hora-obscura.de> a écrit :
> 
>> hi,
>>
>> David Schleef wrote:
>>> On Fri, Oct 12, 2007 at 05:05:58PM +0200, SP GLE wrote:
>>>> is there a way to get a GstBuffer with a configurable memaligned
>>>> payload (aligned(16)), in order to use payload with SSE2 and other
>>>> liboil methods (even if liboil can work on not aligned data, using
>>>> aligned could speed up some processing).
>>> There are very few situations where allocating aligned buffers to
>>> use with liboil will result in faster performance.  Liboil was
>> designed
>>> to allow you to use it without thinking of such things.
>>>
>>> In any case, you can create aligned buffers by allocating a normal
>>> buffer with extra padding, and creating an appropriate subbuffer.
>>> Alternately, you could create a subclass of GstBuffer similar to
>>> the efence element in -good.  Or you could patch gstbuffer.c and
>>> add a function similar to gst_buffer_new_and_alloc() that allocates
>>> memory using posix_memalign() instead of mallc().
>>>
>> I was recently looking at it. Unfortunately gst_buffer_new_and_malloc
>> uses
>> g_malloc and g_free (we need to fix the docs regarding this). data
>> allocated
>> using posix_memalign() should be free'ed with free(). It could
>> confuse glib if
>> its free'ed with g_free(). So a subclass shoudl override the
>> GstMiniObjectFinalizeFunction and properly release the memory. On the
>> other hand
>> I wonder if we should perhaps switch GstBuffer to use
>> posix_memalign() and free().
>>
>> Stefan
>>
>>>
>>> dave...
>>>
> 
> 
> 
>       ____________________________________________________________________________________________
> Découvrez le blog Yahoo! Mail : dernières nouveautés, astuces, conseils.. et vos réactions !
> http://blog.mail.yahoo.fr
> 
> -------------------------------------------------------------------------
> This SF.net email is sponsored by: Splunk Inc.
> Still grepping through log files to find problems?  Stop.
> Now Search log events and configuration files using AJAX and a browser.
> Download your FREE copy of Splunk now >> http://get.splunk.com/
> _______________________________________________
> gstreamer-devel mailing list
> gstreamer-devel at lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/gstreamer-devel