Here are some changes that gave me a significant frame rate increase of about 10 fps with random objects disabled. The interesting thing is that these changes aren't in the main loop but are in tile loader. My guess is that I've reduced the memory footprint just enough to reduce CPU cache misses, though I have no hard evidence of this.
Initially I modified all SGBinObject member functions to pass/return by reference instead of by-value. This gives little or no speed up but allows for some optimizations in fgBinObjLoad(). It is these changes that reduce the number of memory allocations. Needless copying of vectors, and vectors of vectors, can be very memory intensive, especially if they are large.
Anyway I would be interested to see if you get similar results. I would emphasize that the frame rate increase only occurs with random objects disabled. I lose about 10-15 fps on my GF2MX 32MB with random objects, probably a fill-rate limitation or texture memory thing.
* Finally I think I have the partial ssg tree deletion routine working correctly
after I managed to break it (and other confusion in the code cause it to
never be called so I didn't notice the problem.)
* Converted several SG_INFO statements to SG_DEBUG to clean up some
extraneous console output.
* This *should* conclude my investigation into a massive memory leak. :-)
Animations are now contained within the scene graph itself and are
updated whenever the graph is traversed -- that saves time by not
updating animations not currently in sight, and it allows animations
to be used for static objects and random objects as well.
Added new FGModelLoader and FGTextureLoader classes. These are intern
tables for models, to guarantee (mostly) that no model is loaded more
than once. FGTextureLoader is not yet used anywhere, but
FGModelLoader is now in place everywhere that ssgLoad* used to be
used (thus adding the ability to use animations).
In the future, FGModelLoader will add some interesting functionality,
including the ability to reload 3D models on the fly.
I've merged FGProps and FGTelnet so there is just a single property server.
I've left in the --telnet=port# command line option but it could be removed
if we wanted to. The command line accepts two forms of the --props option.
The original (--props=medium,dir,hz,host,port#,style) and the shorter
--props=port#. If you accept this change then src/Network/telnet.[ch]xx
can be removed from the cvs repository.
Your 3D models cause a stackdump when the base is separate from the
root. I've attached a patched newmat.cxx - you may want to test it
before committing!
I changed the code such that textures terrain also takes ambient,
diffuse, specular and emissive from the materials.xml file.
[note from dpm: doesn't seem to work yet]
When the scroll bar slider is dragged, the list scrolls only far enough to see all items; only the arrow buttons can scroll it so far that the last item goes to the top of the view.
Fix scroll bar proportional size: was wrong when the list was only a little longer than the visible area.
Minor fixes such as "delete files[i];" -> "delete[] files[i];" (where the item being deleted is an array of characters) and removal of global variables.
Smooth scrolling when dragging the slider: one item at a time, rather than one tenth of the list at a time.
Fix a bug that would have occurred if instantiated with arrows=2.
Sort properties primarily by name and then by numerical index order, rather than a simple ASCII string order. E.g. "js[1]", "js[2]", "js[10]" rather than "js[1]", "js[10]", "js[2]".
Avoid crashing if the selected property path does not exist; display an empty list instead. This cannot happen when the property picker is working properly, but did happen due to missing indices prior to this patch, and could happen if the user is allowed to type a pathname, as in the http and telnet interfaces.
Fix truncation of strings to PUSTRING_MAX: was wrong when string length was exactly 80.
Fix: move the scroll bar to the top each time a new list is displayed. It was left at its previous position, while the top of the new list was displayed, not corresponding to the slider.
Use getDisplayName instead of duplicated code: gives a better decision on whether to display the index, and avoids invalid property paths being generated which would previously crash find_props().
Replace unnecessary node lookups by name with direct access: tidier and more efficient. E.g. "getValueTypeString (node->getNode(name.c_str()))" -> "getValueTypeString (child)".
name "random-objects".
Put the random objects for each tile inside a top-level
ssgRangeSelector with a maximum range of 20km. This saves a lot of
range tests for distant tiles, and gives about a 10% framerate boost
on my card at 1000ft AGL (possibly more on faster cards).
instead, use a separate dummy bounding sphere for each triangle and
each tile, with the actual bounds, to make sure that objects are
always added when they should be in sight.
the tile cache it's ssg elements are disconnected from the main ssg scene
graph, and then the tile is thrown on the end of a delete queue. The
tilemgr->update() routine runs every frame. It looks at this queue and if
it is non-empty, it incrementally frees the compents of the first tile
on the queue. When the tile is completely free it is removed from the queue.
The amount of time to free the memory for even a single tile can be quite
substantial, especially with the increased overhead of dynamic/random
ground objects. This change allows the system to spread the work of freeing
tile memory out over many frames so you don't get a noticable single frame
hit or stutter.
the triangles containing the objects were out of the view Frustum,
they were never traversed by ssg (and thus, never culled). Now, every
200 frames, do a pass through the whole scene graph with cull-testing
disabled and without drawing anything; that will allow random-objects
to be collected incrementally.
failed on /sim/time/gmt-string because the raw value contains an embedded
space. Now I use a regex instead of a plain split(). I also removed the
view_next and view_prev functions. These will have to be replaced with
command properties.
I've also added a FGFSDemo.py script which is a python version of David's java program.
