* fix(python): a screen uploaded by the worker goes into a texture of its own, put in place after a fence
With python.threadedupload (the default) the worker wrote the image into the texture the drawing thread
draws from, with glTexImage2D and glGenerateMipmap from its own context, and only a glFlush after. Nothing
ordered the two contexts: the drawing one could sample the texture while its storage was being made anew,
its level 0 half written or its mipmaps not made yet, and on a change of size the worker deleted the texture
and wrote a new name into `id` with no synchronisation.
The worker now uploads each image into a new texture, sets a fence after it and flushes. The drawing thread,
once per frame in python_taskqueue::update(), waits for that fence on the gpu (glWaitSync, the cpu goes on),
deletes the old texture and puts the new one in place. The worker no longer touches the texture drawn from,
nor `id`. An image the drawing thread didn't take yet is dropped by the next one. Without sync objects the
worker waits for the upload with glFinish.
The windows of python.displaywindows with python.sharectx bound the texture once; they now bind the name of
the last image put in place, which python_rt keeps for them.
Uploads on the drawing thread (python.threadedupload no, the Lua and WebAssembly screens, the editor) are
as they were.
* fix(python): the image of the worker goes in place once its fence is signalled, the viewer binds it each frame
From the review of 8811ee5a:
- glWaitSync ordered only the context of the main window. The extra viewports and the windows of
python.displaywindows draw in contexts of their own and could bind the new texture before the upload
was done. commit_thread_upload() now asks the fence without waiting (glClientWaitSync, timeout 0) and
puts the image in place only once it is signalled, so every context binding it afterwards has it whole.
Until then the task stays in the upload queue and is asked again the next frame.
- The drawing thread chose between upload and commit by python.threadedupload, the worker by whether it
got a context. With python.threadedupload on and no context made for the worker, the screens were no
longer updated. The task now records that a worker uploaded it.
- The viewer of python.displaywindows told a new image by its texture name, but a deleted name comes back
for a later texture. It binds the name of the last image each frame, read under the lock of the render
target, and the name it starts with is read on the drawing thread when it is made.
* fix(python): the old texture deleted under the lock the viewers bind under, a fence not made waited for with glFinish
From the review of #411:
- finish_upload() deleted the old texture before it took the lock of the render target, and the
windows of python.displaywindows read the name under that lock but bound it after letting go of it.
A window could bind a name already deleted, or already given to the next upload. In a compatibility
context (the legacy renderer) that makes a new empty texture: a black frame and a texture never
deleted. The commit and the publishing of the new name now go under the lock, and the viewer binds
under it too. A name bound before the deletion stays alive in that window's context while bound.
- A glFenceSync that returns no fence was taken for a signalled one: the image went in place at once,
with nothing ordering the contexts. The worker now waits for its upload with glFinish then, as
without sync objects.
* fix(python): the worker waits for the lock of the screen without the gil, no flush after a finish
From the review of 939514ea:
- The drawing thread now holds the lock of the render target for a whole upload with
python.threadedupload off. The worker, copying the next image under that lock, waited for it holding
the gil. It now lets the gil go around the lock and the copy: the bytes it copies stay valid, the
object is immutable and held by the reference the worker has.
- glFlush after glFinish did nothing; the worker flushes only when it made a fence.
---------
Co-authored-by: maj00r <maj00r@host>
* fix(render): the python upload thread only marks the texture unit cache stale, the drawing thread clears it
update_from_memory() runs on the python upload thread with python.threadedupload (the default), and since
e230d7b9 it cleared opengl_texture::units there outright. That raced Bind_Material filling the same array
for its multi-bind: a unit reset between the fill and glBindTextures() made the bind of that unit fail, and
the object was drawn with the texture of the one drawn before it, for a frame, at the rate of the screens.
The thread now sets an atomic flag, and the drawing thread clears the cache when it next binds textures
(bind() and the multi-bind of Bind_Material), so a screen texture given a new gl name is still rebound.
* fix(app): fpslimit asks Windows for the 1 ms timer, the frames no longer come in turns long and short
The wait of fpslimit sleeps, and Windows wakes a process which didn't ask for a finer timer in steps of
up to ~15.6 ms. A 10 ms slot of fpslimit 100 ended ~15.6 ms later, the pacer made it up with no wait in the
next frame, and with vsync on the frames were shown for one or two refreshes in turn: visible stutter at a
steady FPS. timeBeginPeriod(1) is held from the settings to the exit, only when fpslimit is set.
* fix: the 1 ms timer also for an fpslimit set by the scenery, the stale unit cache checked with a plain load first
A config block of the scenery is parsed after the settings and can set fpslimit, so the timer period is
asked for at the first wait of the pacer too, once. bind() checks the stale flag of the unit cache with a
relaxed load before the exchange, which it does for every texture.
* fix(textures): the stale flag of the unit cache uses the default memory order (Sonar S8417)
On x86 a sequentially consistent load and exchange cost the same as the
relaxed and acq_rel ones; only the store in invalidate_unit_cache() becomes
an xchg, and it runs rarely.
---------
Co-authored-by: maj00r <maj00r@host>
An image of a material read the material with all its textures, the normal and specular maps and the variants, and only the first one was let go of: those stayed in the memory for every material scrolled past. They are let go of together now (Unload_Material). texture_manager::unload also takes the textures read and not made yet, waiting for a worker still reading one, and deletes the gl texture itself: release() would read it back from the card with gfx.resource.move only for it to be dropped.
The Refresh of the texture browser let go of the images in the middle of the frame, while panels drawn before it could still refer to them; it now marks them, and they go at the start of the next frame.
Checked on td.scn by a scripted run: 8040 images of the picker read a batch every few frames end at 2.0 GB, against 2.3-2.6 GB before this and 1.1 GB at the start; a trackbed material picked from the folder is drawn on the track after its image was let go.
The picker of the textures of a track offers every texture of the textures folder, and each tile that came into sight read its texture full size, into the graphics card and the memory, and kept it there for good. Scrolling through the list for a trackbed texture took about 2.5 MB a tile: some 900 tiles made 3.5 GB, the whole folder 28 GB, and the game went down once the memory ran out.
An image out of sight for 3 seconds is now taken out of both again (texture_manager::unload, Unload_Texture of the renderers) and read anew when it's shown once more, a few a frame as before. Only a texture the scene has never drawn is let go; those of the scene stay to the resource sweep. The texture browser lets go of its images on Refresh too. The picker no longer lists textures/previews: the 4100 previews of the node bank are images of models, not textures of anything.
Checked on td.scn by a scripted run: 8000 images of the picker read one batch a few frames after another held the game at 2-3 GB, where it went to 28 GB and fell before; images shown again after they were let go come back; a trackbed texture picked from the folder is drawn on the track after its image was let go.
The textures, materials and sounds are kept by the names they're asked for.
Given back as named on the disk, a file in another case missed the bank and
was loaded again on every request.
- FileExists with names and extensions finds the file in any case but gives
back the name asked for; FileOnDisk gives the file to open, used where
textures, sounds and scripts are opened
- FileExists of a single name is a plain check again, the places checking
it and opening the same name stay as they were
- the texture format is told by the extension asked for again
- a folder changed shortly before it was read is read again for a name
missing from it, the time of a change being as coarse as the clock of the
file system; a listing cut short isn't kept
The parser and the loaders lowercase every name they read, while the data
made on Windows has files and folders named in any case (scenery/Braniewo,
textures/siec_trakcyjna_ph/.../STS_*.tga, sounds/sip/Gorki.wav). FileCased
matched the file name only, so a folder in another case wasn't found.
- FileCased walks every part of the path on Linux, the folder listings kept
by their lowercase and read again once the folder changed; Windows as it was
- FileExists finds the file in any case on Linux, and the variant with names
and extensions gives back the name as it's on the disk, so textures,
materials, sounds, station announcements and scripts open that name
- models, sound captions, timetables and nodebank previews open the file
as it's named on the disk; the texture format is told by the extension in
any case
Deleting a model in the editor left the renderer holding it as the picked node, the next pick used freed memory: the renderer forgets it now (Forget_Node). A texture filled straight from memory (internalsrc:) can't be read again, released by the garbage sweep it came back grey; it stays now, and the units bound with the old name are rebound. The arrays of the ED brake force sharing are vectors, they leaked once per update. The scene hierarchy and the editor menu items go through the translation. Python caches ignored.