[Pulento]

First of all, many thanks for this very detailed and valuable feedback...

You're welcome, your work is truly inspiring after trying XBMC on lots of ARM small boards and failing.

1. The Kernel reports 512MB of RAM
You are perfectly right !
Short story : My mistake with kernel config.
Long story : I changed a lot of settings regarding memory while I was trying to understand why I did not have the theorical 184MiB available to allocate DMA buffers.
At the begining I tweaked the config to have a 2G/2G split and no highmem at all to avoid highmem overhead. During my tests I reverted to a standard 3G/1G split and forgot to reenable highmem

Anyway, it is already solved : If you want I can even repost the kernel which gives all your memory back right now....

No problem. I compiled a Kernel from your repo; will test it later though and report back.

2. Once again you are right : I have not properly installed CP1252 locales in my RFS
Yet I am unsure about the required package to solve this issue
Be aware that the zypper package manager utility is installed and that I host a repository with additional packages....
you can do :
zypper refresh
zypper search xxxxx
zypper install xxxxx

I'm a complete newbie on Yocto didn't know about zipper and that you tied the image to a functional repository indeed installing eglic-gconv-cp1252 fixed my problem with ISO-8859-1 subs files. Why XBMC decode them as CP1252 instead of ISO-8859-1 I don't know but it's solved now

3. For the dropped frames : I think I know the issue : Could you check the XBMC log when it happens please ?
You will find it in /home/root/.xbmc/temp/xbmc.log
If it is the issue I am aware of, you should have some explicit messages which state that I am voluntary dropping frames...
Here again you are right : Not a resource issue (just a buggy behavior I have to investigate regarding the way V4L buffers are sometimes given back after rendering)

Indeed, here is a log of a file MKV (H.264) @23.976 fps it has a periodic hiccup when playing (most people will not note them)

Code:

23:11:53 T:1219159136  NOTICE: Decode - No frame buffer available (min Free 6 - Ready picture 1)
23:11:53 T:1219159136  NOTICE: Decode - Dropping one frame at idx 2 to free frame buffer
23:11:53 T:1219159136  NOTICE: Decode - No frame buffer available (min Free 6 - Ready picture 1)
23:11:53 T:1219159136  NOTICE: Decode - Dropping one frame at idx 7 to free frame buffer
23:11:54 T:1467466848   DEBUG: CDVDPlayerAudio:: Discontinuity2 - was:173735482.519434, should be:173724812.641069, error:-10669.878365
23:11:55 T:1219159136   DEBUG: CPullupCorrection: detected pattern of length 47: 41500.00 41500.00 42000.00 41500.00 41500.00 42000.00 42000.00 41500.00 41500.00 42000.00 41500.00 41500.00 42000.0
0 41500.00 42000.00 41500.00 42000.00 41500.00 41500.00 42000.00 41500.00 41500.00 41500.00 42000.00 42000.00 41500.00 41500.00 42000.00 41500.00 41500.00 42000.00 41500.00 42000.00 41500.00 42000
.00 41500.00 41500.00 42000.00 83500.00 83500.00 41500.00 41500.00 42000.00 41500.00 42000.00 41500.00 42000.00, frameduration: 43478.723404
23:11:55 T:1219159136   DEBUG: CPullupCorrection: pattern lost on diff 42000.000000
23:11:56 T:1219159136  NOTICE: Decode - No frame buffer available (min Free 6 - Ready picture 1)
23:11:56 T:1219159136  NOTICE: Decode - Dropping one frame at idx 7 to free frame buffer

I don't know H.264 internals but I thought that PULLUP/PULLDOWN was something that was only signaled on interlaced MPEG-2 streams. And this particular stream seems progressive to me. I think (correct me if I wrong) the video signal is keep at 1080p@60 so this XBMC CPullupCorrection is to correct the 23.976 HZ stream played on a 60HZ signal.

Reproducing the build is not so trivial at this time : kernel and xbmc sources are already public but I made a rfs build 2 months ago and I patched some components by hand...
I plan to clean up everything and enable to perform a full build with yocto in the near future : If you are not in a hurry it is the best way.
Otherwise I can tell you and to rebuild specific components : kernel would be easy for instance

I'll wait since I do think it is better to rebuild based on last Yocto, any help testing build or Yocto layers I can help Building last master following Otavio's blog was fairly easy; modifying it and adding bb recipes should be a different matter.

Regarding your remark about ramfs (you mean tmpfs in fact right ?), be aware that my image already takes care of storing logs in tmpfs var/tmp var/log and a few others (including /home/root/.xbmc/tmp/) end up in tmpfs in order to save erase/write cycles on sdcard... Not perfect but already much better than many images which just don't care at all ;-)
Yet on my side, I don't aim at having all the rfs in tmpfs as I wish to keep my xmbc database (what the hell if scrapers had to run at each boot lol) and I wish to be able to easily change a configuration file and so on....

I use to install XBMC on top of a minimal Debian, there is a package on it called flashybrid. You tell it which parts go on volatile tmpfs and what goes to non-volatile tmpfs; it basically rsync changes on reboot and when you want it (using crontab) and it mounts these tmps on top of the original fs so is somewhat transparent. I want to have something similar if it speeds things and keep my cards safe. I use it on work on about 700 units and is working like a champ.

But first I'll like to enable my PS3 remote

At last, I was aware of the udev perm rule and was too lazy to change it before release ;-) Sorry for that...

Finally, sorry for the public key : My secret plan to take over the world thanks to a wandboard botnet has just failed ! ;-)
No, really sorry : I will add a comment in my blog regarding this issue and it will be solved in future images for sure...

No problem. There are lot of ways to take over the world, one of them is making XBMC really works on a cheap SBC

I minor thing I just noted. On some H.264 files (using iMX-H264 decoder) the last (bottom) line of macroblocks looks shifted somehow it is very subtle and maybe 16 or 32 pixels wide I cannot tell but it seems to be one or two bottom line of macroblocks. It actually seems like that the first line of the last macroblock line was copied on the next lines of the macroblock line.

I tried to take screenshots but I just got 0 byte PNGs so I just taked a crappy picture with my phone and I cannot find how to attach it here so I'll send it by email on the Wandboard forum. O my god this thing is complaining about the number of emoticons I use !!! I had to erase them. I miss when mailing lists ruled the world.