As discussed with Tim on irc: Here is a quick fix for the memory-hungry tile
manager. Due to bugs in FGNewCache, old tiles were never deleted.
I left the timestamp-updates in the cull-traversal. but imho things work
just as well when timestamps are updated in FGNewCache::insert_tile() and
FGNewCache::get_tile()
Make an OSG file reader for .stg files.
New class flightgear::SceneryPager, which is a subclass
osg::DatabasePager to handle explicit delete requests.
Modify FGNewCache, FGTileEntry, and FGTileManager to use
SceneryPager. Mostly this involved removing the queues that talked to
FGTileLoader.
Calculate accurate tile timestamps from the time they are traversed in
the cull stage (which means that they are visible) instead of updating
them periodically.
Replace tile entry transform and range node with one LOD node
I tried to make sure accessor functions which return by reference act
on const objects. also replaced some iterators with const_iterator
and a few return/pass by reference that were missed the first time
around.
This is a new improved patch for the previous tile manager fixes.
Rather than building dependencies between FGlocation or the viewer or fdm with
tilemgr what I ended up doing was linking the pieces together in the Mainloop
in main.cxx. You'll see what I mean...it's been commented fairly well. More
than likely we should move that chunk somewhere...just not sure where yet.
The changes seem clean now. As I get more ideas there could be some further
improvement in organizing the update in tilemgr. You'll note that I left an
override in there for the tilemgr::update() function to preserve earlier
functionality if someone needs it (e.g. usage independent of an fdm or
viewer), not to mention there are a few places in flightgear that call it
directly that have not been changed to the new interface (and may not need to be).
The code has been optimized to avoid duplicate traversals and seems to run
generally quite well. Note that there can be a short delay reloading tiles
that have been dropped from static views. We could call the tile scheduler on
a view switch, but it's not a big deal and at the moment I'd like to get this
in so people can try it and comment on it as it is.
Everything has been resycned with CVS tonight and I've included the
description submitted earlier (below).
Best,
Jim
Changes synced with CVS approx 20:30EDT 2002-05-09 (after this evenings updates).
Files:
http://www.spiderbark.com/fgfs/viewer-update-20020516.tar.gz
or
http://www.spiderbark.com/fgfs/viewer-update-20020516.diffs.gz
Description:
In a nutshell, these patches begin to take what was one value for ground
elevation and calculate ground elevation values seperately for the FDM and the
viewer (eye position). Several outstanding view related bugs have been fixed.
With the introduction of the new viewer code a lot of that Flight Gear code
broke related to use of a global variable called "scenery.cur_elev".
Therefore the ground_elevation and other associated items (like the current
tile bucket) is maintained per FDM instance and per View. Each of these has a
"point" or location that can be identified. See changes to FGLocation class
and main.cxx.
Most of the problems related to the new viewer in terms of sky, ground and
runway lights, and tower views are fixed.
There are four minor problems remaining. 1) The sun/moon spins when you pan
the "lookat" tower view only (view #3). 2) Under stress (esp. magic carpet
full speed with max visibility), there is a memory leak in the tile caching
that was not introduced with these changes. 3) I have not tested these
changes or made corrections to the ADA or External FDM interfaces. 4) The
change view function doesn't call the time/light update (not a problem unless
a tower is very far away).
Details:
FDM/flight.cxx, flight.hxx - FGInterface ties to FGAircraftModel so that it's
location data can be accessed for runway (ground elevation under aircraft)
elevation.
FDM/larsim.cxx, larcsim.hxx - gets runway elevation from FGInterface now.
Commented out function that is causing a namespace conflict, hasn't been
called with recent code anyway.
FDM/JSBSim/JSBSim.cxx, YASim/YASim.cxx - gets runway elevation from
FGInterface now.
Scenery/newcache.cxx, newcache.hxx - changed caching scheme to time based
(oldest tiles discard).
Scenery/tileentry.cxx, tileentry.hxx - added place to record time, changed
rendering to reference viewer altitude in order to fix a problem with ground
and runway lights.
Scenery/tilemgr.cxx, tilemgr.hxx - Modified update() to accept values for
multiple locations. Refresh function added in order to periodically make
the tiles current for a non-moving view (like a tower).
Main/fg_init.cxx - register event for making tiles current in a non-moving
view (like a tower).
Main/location.hxx - added support for current ground elevation data.
Main/main.cxx - added second tilemgr call for fdm, fixed places where viewer
position data was required for correct sky rendering.
Main/options.cxx - fixed segfault reported by Curtis when using --view-offset
command line parameter.
Main/viewer.cxx, viewer.hxx - removed fudging of view position. Fixed numerous
bugs that were causing eye and target values to get mixed up.
but no entries qualify for removal. It will keep trying to schedule the
tile(s) until an entry frees up. Entries in the cache do not qualify for
removal if they are in the process of being loaded.
a crash when relocating to a new airport. Pending work from the old
area is now just completed as normal, rather than trying to empty the various
queues in their various stages when can lead to many problems in a threaded
environment.
He writes:
Here are the final changes to add threads to the tile loading. All the
thread related code is in the new FGTileLoader class.
./configure.in
./acconfig.h
Added --with-threads option and corresponding ENABLE_THREADS
definition. The default is no threads.
./src/Scenery/tilemgr
Removed load_queue and associated references. This has been replaced by
a new class FGTileLoader in FGNewCache.
Made the global variable global_tile_cache a member.
schedule_needed(): removed global_tile_cache.exists() tests since
sched_tile() effectively repeats the test.
initialize_queue(): removed code that loads tiles since this is now
performed by FGTileLoader.
update(): ditto
./src/Scenery/newcache
Added new class FGTileLoader to manage tile queuing and loading.
tile_map typedefs are private.
exists() is a const member function.
fill_in(): deleted
load_tile(): added.
./src/Scenery/FGTileLoader
The new threaded tile loader. Maintains a queue of tiles waiting to be
loaded and an array of one or more threads to load the tiles. Currently
only a single thread is created. The queue is guarded by a mutex to
synchronize access. A condition variable signals the thread when the
queue is non-empty.
CLO: I made a few tweaks to address a couple issues, hopefully what we
have is solid, but now we kick it out to the general public to see. :-)
structure. The new approach is simpler, more flexible, and more dynamics.
We can now dynamically size the tile cache up and down. Also, the range
of tiles to load is now dependent on visibility and is calculated to always
bring in enough tiles.
