I have been refactoring GStreamer's spectrum analyzer. It is now working faster and also allows to get per-channel results. For testing I updated buzztard to check if one has the new version and if so show per channel graphs. While hacking on the analyzer graphs I have also refactored the code, so that now it is possible to probe the final output as well.
I spend the bigger part of the month on the tests again. I had an unref too much somewhere. Even after all the work in refdbg this is still difficult to track down. Eventually I found it and now all tests are green.
43 files changed, 796 insertions(+), 100 deletions(-)
gtk_window_set_transient_for(window, main_window); gtk_window_set_destroy_with_parent(window, TRUE);On my desktop it works fine on the netbook the problem still happens :/ No idea right now. As a good side effect I ensured that all dialogs have a properly set default response.
I was also running the performance tests and did some oprofiling. I knew already that my data-fixup code for buzzmachines caused quite some overhead. It is handling denormals and marking full zero buffers as empty. I now added code that sets the FPU on x86 machines to DAZ|FZ mode from the application and disable the fixup code. On my netbook my 11 min benchmark song renders in about 1 min instead of 1:20 min before. On my desktop most of the change are hard to measure as it already only takes 5 seconds :). Anyway the oprofile runs confirm the speedup. The code disappeared from the top 20.
The pattern editor widget becomes slow when getting large. Some text drawing operations look expensive. For a start I ported the whole widget to use cairo, but that does not take care of the slowness. After more measurements it becomes clear that I need to rework this more. I need to render it offscreen and reply to expose events by copying pixels.
I also worked a bit more on the interaction-controller library. udev/hal code is refactored into extra objects (saves me a lot of #ifdefs). It also has some tests now. I still have a mysterious issue with udev. Run this stand alone example as told in the header - for me valgrind complains about uninitialized data access. It only seems to happen when you have a /dev/input/mice which all my computers have. I have a workaround, but believe more testing from other would be great.
44 files changed, 1087 insertions(+), 317 deletions(-)
As there was no project news over x-mas, I have a slightly longer one this time. Before I dive into the usual code changes, something else - tadaa - we have a new webpage. The previous page was a mediawiki with a few customizations. Now we use a combo of wordpress-3 and mediawiki. Both use a similar theme (based on the arch linux one). Wordpres gives as news, forum, social-network features and much more. The first days I was suffering a lot from wordpress account creation spam. Luckily this seems to be under control now. The concept and a lot of the work was done by Joe Pea. We both have quite a few more things on the todo list. One of the next items is single sign in for wordpress and mediawiki. Stay tuned.
In the last two month I improved the song-recovery workflow. The recovery is now giving better feedback to the user and we're properly cleaning up afterwards. The machine machine and pattern view do now have almost full undo/redo.
Quite several times already I had to deal with a dialog containing a treeview showing a list of items. The latest example is the song-recovery dialog. Giving such a dialog a nice good initial size is tricky. The treeview usually goes into a scrolled window. By default one can see the headers and maybe one line. Of course one can resize the dialog a bit, but as the line height is not easy to figure it is still not easy. One approach I came up with is to grow the dialog to show all lines until it would exceed a certain limit. The limit would be e.g. half of screen or window height. The code for that is quite simple:
list=gtk_tree_view_new();
g_signal_connect(list,"size-request",G_CALLBACK(on_list_size_request),NULL);
...
static void
on_list_size_request(GtkWidget *widget,GtkRequisition *requisition,gpointer user_data)
{
gint max_height=gdk_screen_get_height(gdk_screen_get_default())/2;
gint height=2+MIN(requisition->height,max_height);
/* make sure the dialog resize without showing a scrollbar until it would
* reach half screen height */
gtk_widget_set_size_request(gtk_widget_get_parent(widget),-1,height);
}
Another thing in gtk that I wonder if it has a better
solution is the disposal of popup-menus. The problem is that
gtk_menu_popup is not taking ownership of the menu, but
there seems to be now event when the menu is done, so that
one can unref it. The only solution I can see is the pre-
create the menu and ref-sink it. When the entity that owns
it is disposed it can destroy and unref it.
As we're quite gtk related in this time anyway one more discovery. Listening for key-press-event instead of key- release-event and widgets makes key-repeat work. Have added that to the gtk api docs.
Beside all the things above a couple of bug-fixes went in (make preferences in machines work (e.g. choosing fluidsynth instruments). I did some refactoring on midi controller handling; one can use the pitchwheel now and continuous controllers work better. Finally the dialogs warning about unsaved changes are more user-friendly; they now show the time passed since last saved/created in addition to the timestamp.
50 files changed, 2402 insertions(+), 1183 deletions(-)
Last month I made good progress on the undo/redo. I am now running the undo/redo in sync with the object lifecycle. The downside of this is that undo/redo fails when you have ref-count leaks. This now makes them more apparent. When trying to debug them I noticed that refdbg is not working anymore since quite a while. I was googling what other people use for such problems. Well, apparently everybody hates ref-count issues. Also most of the info you find is how to get traces for refcount activity. That is what refdbg does also. As this did not work for me anymore I've tried systemtap. Unfortunately right now its not useful for user-space work on most distros (except fedora) - bugs filed. Next step: gdb scripting. I managed to get a gdb script running. Now even for simple cases you quickly end up with a log of 150 backtraces. That is a lot to manually review. I went one step further and wrote a script to filter the traces. Over time I got the trace-log down to 20 entries. This is practical. Unfortunately the gdb script only work for tracing the lifecycle of one particular object. If I extend it to work for all refcounting activity I hit a gdb limitation to not allow nested commands :/. Back to refdbg :) I finally figured why it does not work and have started to fix it up. Lets see how far I get next month.
Buzztard
Naturally this used most of my hacking time. Still I managed to make some progress there as well. The buzztard irc channel was much more active this month. People reported small things here and there and most of it got fixed right away. The manual got some more love too to help people to get started and improve the description of several concepts. Finally all the refcount-debugging leads to a lot of code reviews. I found several areas with lots of needless refcount activity - fixes are in the repository.
81 files changed, 2682 insertions(+), 817 deletions(-)
In October I had not much time for buzztard. I went to the GStreamer conference and the CELF Europe. There I gave a talk about the "MeeGo multimedia stack". I also hacked on two GStreamer tool projects; gst-tracelib and gst-mediainfo. In tracelib I worked on a UI to monitor gstreamer pipelines on the fly. My approach to use libgvc (graphviz) for the layout seems to challenge it a bit though - I get very mysterious failures. At some point realloc is failing to e.g. allocate 96 bytes and the lib calls abort(). mallinfo() from libc is telling that the process uses 200 mb and valgind has nothing serious to complain about. That's weired as I was suspecting a corrupted memory list. I have to put that aside for a while until I get some new ideas for tackling it.
gst-mediainfo is the UI for the new discoverer API in GStreamer. I'll make a first release when the current gstreamer core and base is out as it depends on them.
In buzztard I mostly worked a bit on the crash recovery and did some bug fixing. Hopefully will have more time in November.
14 files changed, 325 insertions(+), 166 deletions(-)
After all the initial work for journaling of edit events in the editor, I wanted to get a first proof of concept that everything fits together. And that I just got. The pattern editor implements a change-logger interfaces. It grabs the changelog singleton instance and logs all changes together with the undo action. The redo-action gets serialized and written to the changelog. The ui can undo/redo activities by using the change-logger interface. Upon startup the ui checks for leftover change logs, auto-cleans a few of them and offers a dialog for deleting/replaying them. I can now load a song, make changes in the pattern editor, ctrl-c/kill the app and replay the log when starting it again. There are a couple of details that need to be cleanup and some things need a bit more thinking: Next I will implement this in larger scale also.
As a continuation from last month I changed the whole editor UI to use G_DEFINE_TYPE and co. Boring work, but good to have it done now.
I had some strange crashes in the test suite - all pointing to gconf and only happening with the commandline client. After adding lots of extra logging there I still could see nothing wrong. Then I remember the hack I added to get a eventloop into the app - I run it as a new thread. By doing that I was doing gconf bits from different threads too and it does not seem to like it. I have now done the mainloop stuff more properly and get no more crashers. While looking into this problem, I cleaned the settings classes and the gconf backend a bit. It now does more stuff only as needed, improving the startup time of buzztard app a little further.
In GStreamer next core release will have extra metadata on element factories. This mechanism allows me to kill the one-property help interface. I rewrote all the code to use that conditionally and will remove it in a year or so.
As a last bigger task I migrated from deprecated HAL to GUdev. I have still conditional support for both, but will kill the HAL code once everything works nicely. I was using hal to discover input devices (joysticks, wiimotes,..) and midi-devices that can be used to control effects. Most of the conversion was straight forward. I spent a bit more time on getting nice names of the found devices for the UI and to classify midi-devices. For the midi-device classification I copied the matching rules from HAL - after all they seemed to do the trick there. For the ui names HAL was actually grepping through various files (e.g. /proc/asound/cards) for soundcard names. That I did not really wanted to copy. I got some great help from the udev people and learned that udev devices have properties like ID_VENDOR. One just needs to walk the device tree to get to the root device and grab the names there.
119 files changed, 3377 insertions(+), 1953 deletions(-)
I still face some problems on my new computer system. One of
them iswas that udev was hanging in 8/10
cases at bootup.
Thanks to great support from the udev folks I learned how to
track it down. I previously tried to add options to
/etc/udev/udev.conf (e.g. to raise the log level), but one
can actually just add them to the kernel options.
This has the great advantage that they don't persist. Things
you want to add to debug are:
udev.log-priority=info udev.children-max=1 udev.exec-delay=1
This logs a lot of things (the boot can go from 30 second to minutes, so be patient). The option children-max is turning parallel event execution off and exec-delay is delaying run actions. With those I saw that the system hangs when loading the cdrom.ko kernel module. As I had some problems with the drive as well I swapped it and now bootup is fine.
If you like to work on free software and like snow - we're having a few open positions:
Ping me on irc (ensonic @ freenode/gimpnet) if you have questions.After some more work bt-dec starts to works better. I can now play buzztard songs in totem, banshee and co. I still see some issues with seeking though.
I have be going through my pending gtk bugs and patches. I can see a lot of activity in gtk+ these days. People have been helpful reviewing and commenting the changes. I could close two of my bugs with patches to code and docs committed. At the same time I also understand a few things better and reworked the focus handling on the tabs. Instead of listening to focus notifies and then grabbing the primary widget on the page, I am now overriding the focus vmethod and grab there. The disadvantage of the previous method was that the widget has to be realized and mapped, but when you receive the notify for the tab switch, it is not necessarily the case yet.
There were some changes on the logging in gstreamer. I had to adjust some parts of my tests that where intercepting log messages.
I've started to migrate all modules to use G_DEFINE_TYPE, G_DEFINE_INTERFACE and GST_BOILERPLATE macros to shorten the code. I hope I can finish it next month. At the same time I have cleaned-up my gobject use:
I improved the error handling in the commandline client and also in the ui (when rendering to disk). It catches and handles missing element messages now. Before the processing could hang if that happened :/
While doing more work on the changelog I noticed that I could also use that to show last save date when warning about unsaved changes. The change log uses UTC timestamps, so I had to convert them. I haven't found any nice API for this (like give a date and tell what tz it is and ask to convert to local or any other tz). The startup check for old change logs does now some auto cleanup and sanity checks. Next I'll add some ui for picking one for recovery.
I recently upgraded on machine to suse 11.3. After that all machine icons where in the wrong size. After some digging I noticed that the gnome/index.theme misses all the 'scalable' entries and that seems to break my svg icons. I have added a workaround, but with that I get 48x48 png scaled up. Readding the missing entries works. I filed a bug, but nobody feels responsible so far :/
Bored from all the refactoring and bug fixing I hacked a bit on the spectrum analyser. Its now using cairo (no more gdk_draw_*) and thus is ready for gtk3 and looks much nicer. In addition is can do log plots and one can increase the resolution.
While playing with orc for gstreamer and googling for simd information, I can across some info about the gcc autovectorizer. One can ask for quite verbose reports of gcc attempting to use simd for loops. I could improve the performance in a few place based on the reports. Unfortunately the only places where the vectorization finally succeeded where those where I just should have used memset() :/
78 files changed, 4300 insertions(+), 1832 deletions(-)
New HTML Parser: The long-awaited libxml2 based HTML parser code is live. It needs further work but already handles most markup better than the original parser.
Keep up with the latest Advogato features by reading the Advogato status blog.
If you're a C programmer with some spare time, take a look at the mod_virgule project page and help us with one of the tasks on the ToDo list!
