| Age | Commit message (Collapse) | Author |
|---|
|
Since blank is a builtin, at least for now it requires explicit
override since the builtins aren't added to the modules values
list and filtering is going to be always applied shortly.
|
|
The "tiller" base (base being a Rocket concept) was always a bit
spurious. Maybe "til" would make more sense, but "rkt" is more
contextually specific.
I think when I originally picked "tiller" I was prioritizing
picking something unlikely to collide with another directory
name. But the way Rocket is naming the directory in the
filesystem it gets suffixed with an _ anyways.
|
|
The "compose,compose,compose,compose" default was never intended
to be permanent, but gave a set of scenes to test the Rocket
integrations like scene selection and scene-specific tracks
without any additional effort.
Now that there's scener for easily adding/editing scenes, and
things are just generally more mature, I think it makes sense to
just go back to something minimal here.
I'd really rather just have it be "", but that's not handled well
presently. There isn't really a way to start with an empty
scenes set for rkt. Which is awkward, but "blank" is close.
It'd just be nice to start with an altogether blank slate rather
than having to always edit the default first scene when starting
anew...
|
|
Now that the "scener" interface seems to be semi usable and
capable of editing... things are looking more complete in the
sense that there's no huge gaping holes and a lot of the dust has
settled.
It's also looking pretty good for this sticking around long-term,
so I'm removing the experimental flag making this more
discoverable and visible in general.
There's still work to be done surrounding GNU Rocket the library,
like getting it using non-blocking connects, and there's a need
for forgetting tracks which the protocol doesn't support
currently. But it'd be silly to wait on getting those things
upstream before making rkt more visible.
|
|
Playing with libs/sig in 2D, this isn't really an interesting
module by itself in terms of visual output. But it might have
utility as a diagnostic thing if libs/sig becomes a more used
thing.
At the very least, for now, it's useful for observing affects of
and iterating on libs/sig development. So I'm merging this, just
gated behind TIL_MODULE_EXPERIMENTAL so it's not in rtv rotation
or presented as something in the usual modules list.
|
|
This augments the NEWSCENE_SETUP state to also handle editing
existing settings, which is slightly different but actually
overlaps with the already implemented invalid input stuff.
There's still work to do, and the UX is kind of awkward at best.
But this takes us from having no ability to edit existing scenes,
to being able to actually make edits interactively while it's all
live... with a modicum of interactive guidance via the setup
machinery.
It basically behaves just like creating a new scene, except
instead of the <enter>-accepted "preferred" values, you accept
the existing relevant setting. So as-is, when editing, you have
no shortcut for getting back the "preferred" value for a given
setting. That's been replaced with the existing value for that
setting.
You also get seemingly spurious redundant queries for module
names in things like compose::layers, but they're really not the
same since the first time you get asked it's actually the full
settings string you're getting an opportunity to specify
wholesale, but can accept to seed the layer's settings as-is,
which you will then be given an opportunity to edit piecemeal.
It's that subsequent piecemeal editing of the individual settings
within the nested instance that can feel like a spurious
duplication, especialy when a given layer has just a bare-value
module name and no subsequent settings.. like "plasma". You'd be
asked if you want "plasma" for the layers/[N], then asked if you
want "plasma" _again_ for the layers/[N]/[0] since the module
name is an unnamed setting at position 0 within the layers/[N]
instance.
It was tempting to try streamline that a bit, but there's
actually utility in having an opportunity to paste in a full
settings string for the layers/[N] if you have a serialized scene
onhand you want to dump in there. Then after that, you can juts
smash enter as much as necessary to accept what you pasted in
without editing those in the piecemeal phase. Or, if there was
actually something in what you pasted you did want to change,
change it during that piecemeal phase.
I think it at least kind of works.
|
|
There's an outstanding issue surrounding the need for context
clones, and I'd just like to write something down somewhere
before it falls off my radar. Presently it's just checkers that
exercises this need, so it makes sense to put it here for now,
until I get around to actually taking action on the issue.
|
|
Mechanical change switching til_fragmenter_slice_per_cpu() users
over to til_fragmenter_slice_per_cpu_x16(), except sparkler where
it's quite detrimental to performance.
|
|
Better spearate the generic error line from surrounding text with
an extra newline
|
|
This was a bit of an aspirational misnomer, editing scenes isn't
actually implemented yet. What the EDITSCENE state currently
implements is the per-scene dialog+prompt, which originally was
going to just be the scene editing flow but became more of a
"view a specified scene's details" with a prompt of its own.
Nothing functionally changes, just mechanical internal renames.
|
|
This is leftover from development when it used a fmt string in
combination with the key, before the desc path was getting
properly appended.
|
|
With the addition of the "radcache" in b6362c, the need for a
faster approximate atan2f() is largely eliminated.
And there seems to be a bug in the implementation as-is taken
from https://mazzo.li/posts/vectorized-atan2.html
You can see the bug as vertical line artifact around the center
where the X coordinate would be 0. Rather than debug what's
wrong with this approximation's implementation surrounding its
quadrant adjustments, let's just resume using atan2f() and let
the cache keep things quick.
|
|
On Linux I don't notice a significant affect on anything letting
rkt try connect every frame when offline but in creative mode.
On Windows however, Dan reported significant latencies in the
Scener prompt responsiveness and visible slowdowns in this
condition.
I suspect the WIN32 Rocket library's sync_tcp_connect() code is
the real problem here. But for now I can ameliorate things a bit
by just hammering on that code path less when unconnected.
|
|
This is only relevant to creative mode. Stops RocketEditor from
continuing playback endlessly until user intervention beyond the
current end of the demo.
|
|
Prepartory commit for pausing playback upon entering 99999 scene
It needs to trigger only on the edge of entering the scene to
permit RocketEditor to unpause playback even if still in scene
99999, if that's what the user is trying to do. It'd be annoying
to have it just keep asserting a paused state until the scene idx
leaves 99999...
But this also enables triggering anything on scene change edges,
for future stuff.
|
|
- strip off the leading /path/to/rkt/module prefix
- separate taps from their scene context path with ':' vs. '/'
RocketEditor doesn't currently support recursive grouping, so
this is as good as it gets.
Note this commit will break the existing tracks for alphazed, so
you'll have to use a newer .zip for track data if building your
rototiller from source. Or build from a prior commit.
|
|
Preparatory commit for rewriting track paths a bit to better
group things in RocketEditor. I'll need access to
rkt_context_t.til_module_context_t.setup->path for prefix
matching purposes..
|
|
This makes it possible to tiled+rotate the output of another
module in the same manner checkers::fill_module fills cells with
module output.
The default stays "none" for the classic roto with the
psychedelic color cycling. When !"none" the color cycling
doesn't get applied currently. It might be interesting to try
support that in the future though.
|
|
During rkt_scener development this append was at one time needed,
as there was no retained reference to the scenes_settings for
deriving paths from.
Now that the path is derived from the actual scenes setting
instance it's just resulting in a double trailing /scenes in the
"/module/rkt/scenes/scenes:" heading
Purely cosmetic fix
|
|
Pressing '=' at either prompts now makes scener's current scene
the current Rocket scene.
While you could already do this manually by just looking at the
scenes list for the one with the '*' in the Rocket column while
either watching a production and pressing <enter> repeatedly to
keep refreshing the scenes list... that's cumbersome and
annoying, now just use this shortcut.
Since this just copies Rocket's scene to the Scener scene index,
it needed to properly handle scene 99999... hence the previous
commits.
|
|
While there's no actual context for 99999, it's a state we need
to represent visibly somehow.. so just make it appear like an
epilogue scene off the end.
I've included the Rocket/Scener/Pinned status columns
consistently as well so you at least still get a visible
indication when you've done something like pinned 99999 somehow
(not that this is possible presently, but with future changes
there will be more ways to copy the Rocket idx into Scener's idx)
|
|
rkt_scener needs to know this value so define it in rkt.h and
switch over all the existing 99999 instances.
|
|
Until channel context paths are distinct it's buggy to let the
contexts linger while constructing the next channel's contexts.
Originally when the gc was added here the intention was to
support stuff like the "ref" module and get the channels settings
wired up immediately with more focus on rtv's details in this
area. Supporting stuff like contexts backing some layers
persisting across channels, while the others were swapped out,
seemed potentially interesting (and it still is).
But the rkt stuff became prioritized as rtv is more like a fuzzer
than anything despite being the default module. And rkt related
activities will continue for now, so let's just get rtv less
likely to crash.
A reliable repro for triggering an ASAN UAF bug without this
commit is:
--seed=0x64af3b05
'--module=rtv,duration=1,context_duration=1,channels=compose,caption_duration=2,snow_duration=0,snow_module=none,log_channels=on'
'--video=sdl,fullscreen=off,size=640x480'
A few channels in blinds will UAF while updating taps stored in a
freed context, because the previous channel has a blinds in the
same layer as the newly setup channel, putting the contexts at
exactly the same paths on-stream. There's probably another bug
in here that I need to dig into, but coexisting contexts at the
same path on-stream was never the intention. The syncronous
immediate gc ensures nothing remains of the previous channel
before constructing the new one at the same path.
|
|
The channel module name was being used as the settings instance
name, which is redundant to the entries[0] "value as label"
behavior with the module name also @ entries[0].
The resulting paths resembled:
/module/rtv/compose/compose/layers/[N]/...
/module/rtv/compose/compose/texture/...
Now they are:
/module/rtv/channel/compose/layers/[N]/...
/module/rtv/channel/compose/texture/...
There's still work to be done in this area of rtv. It's unclear
if even a static "/module/rtv/channel" is correct, or if it
should show the subscript of the channels[] entry currently being
used when that becomes a proper nested settings situation, i.e.
/module/rtv/channels/[0]/compose/layers/[N]/...
...
/module/rtv/channels/[1]/spiro
...
/module/rtv/channels/[2]/roto
...
for a "channels=compose,spiro,roto" kind of hypothetical
another option which might make sense is to have the channel path
more like an auto-increment identifier so the channel contexts
have unique paths and don't collide on the stream. This could be
important for scenarios utilizing the "ref" module and trying to
reuse parts of the context from one channel to the next. When
the paths collide things tend to break in weird ways presently.
this area needs more thought
|
|
Silly and unnecessarily racy to do this in render_fragment now
that blinds is threaded
|
|
This is a first stab at threading blinds, while here I got rid of
the reliance on _checked() put_pixel for clipping.
This could be better, things like using a block put/copy instead
of put_pixel would be a huge advantage for blinds due to its
naturally contiguous blocks per-blind.
While this module when non-threaded wasn't especially slow, any
module leaving cores idle is depriving other potentially threaded
modules of utilizing them for the duration of the non-threaded
render. So now that rototiller is being used for compositions
and increasingly becoming something of a meta-demo, it's become
important to make everything as threaded as possible.
|
|
Especially since taps are established on first use, it's
desirable to get them online immediately so they show in things
like --print-pipes and modules::rkt/RocketEditor and not
spuriously once the context becomes actively rendered...
This kind of thing will be done for all modules eventually
|
|
computing the angle for every pixel coordinate from the origin is
costly, even with the approximate method added by c690303.
An easy speedup is to only do this once for a given frame
dimensions, and cache those results. In the form of a 32-bit
float, it's equivalent to caching a full page of pixel data.
This is slightly complicated by needing to be an effectively
global cache and the potential for multiple shapes contexts
rendering concurrently when part of a composition.
I think this particular situation highlights a need for something
equivalent generalized on-stream where modules can register
discoverable caches of costly to compute information, having a
high probability of being useful to others.
In this particular case it was alphazed's use of shapes in two
layers that made it an obvious win, even without any other
modules needing atan2() per-pixel with a centered origin.
With this commit:
Configured settings as flags:
--seed=0x64adabae '--module=compose,layers=blank\,shapes\\\,type\\\=pinwheel\\\,scale\\\=.9\\\,pinch\\\=.25\\\,pinch_spin\\\=.1\\\,pinches\\\=7\\\,points\\\=19\\\,spin\\\=.01\,shapes\\\,type\\\=pinwheel\\\,scale\\\=.9\\\,pinch\\\=1\\\,pinch_spin\\\=-.25\\\,pinches\\\=8\\\,points\\\=5\\\,spin\\\=0,texture=moire\,centers\=3' '--video=mem,size=1366x768'
FPS: 73
FPS: 74
FPS: 73
Without:
Configured settings as flags:
--seed=0x64adb857 '--module=compose,layers=blank\,shapes\\\,type\\\=pinwheel\\\,scale\\\=.9\\\,pinch\\\=.25\\\,pinch_spin\\\=.1\\\,pinches\\\=7\\\,points\\\=19\\\,spin\\\=.01\,shapes\\\,type\\\=pinwheel\\\,scale\\\=.9\\\,pinch\\\=1\\\,pinch_spin\\\=-.25\\\,pinches\\\=8\\\,points\\\=5\\\,spin\\\=0,texture=moire\,centers\=3' '--video=mem,size=1366x768'
FPS: 55
FPS: 54
FPS: 54
So it's significant, and in alphazed there's also a transition
from one scene with two full-screen shapes layers into a
checkered scene with shapes as the fill_module. Further
amplifying the payoff.. infact whenever shapes is used for a
fill_module in checkers, there's n_cpus shapes contexts created
because checkers is threaded. All of those would be benefitting
from the cache.
|
|
The color stuff as-is isn't cheap and doesn't even get used if
there's a texture present, so don't bother with it at all.
This is especially significant since this module isn't threaded,
so it ties up all the cores leaving most of them idle when part
of a composition.
Also since spiro doesn't clamp its coordinates to the fragment
dimensions, only considering the frame dimensions, it must
continue using the "checked" put_pixel variant...
|
|
This seems to make things work well enough for mingw+wine
Will probably revisit in the future. Adding an ewouldblock
helper rather than duplicating the ifdeffery seems likely
Let's just leave it like this for now and find out if a real
windows test succeeds
|
|
This is available in win32, unlike inet_aton()
|
|
In the early days of checkers when I introduced fill_module= with
the per-cpu contexts to still allow threaded rendering, the whole
seed passing to contexts thing wasn't as well sorted out. This
meant the contexts often produced vastly different outputs
despite being the same module, same seed, and same settings.
The consequence of that was that w/fill_module checkers would
produce crazy randomized output when you expected the same output
in the filled cells. But by using .cpu_affinity (which I had to
implement just for this use case actually) at least the different
outputs would become stable. It was a band-aid over a different
problem that still needed sorting out.
Nowadays, it seems like this is improved enough at least for
alphazed to look correct without the affinity hack, so I'm
removing it because it really kills checkers threaded
performance.
Whatever modules remain uncooperative WRT seed reproducibility,
they'll just need to be fixed up.
|
|
While here also did more minor optimizations moving things out of
the inner loops that were only there out of laziness...
|
|
Taken from this excellent post:
https://mazzo.li/posts/vectorized-atan2.html
While I haven't gone full vectorized, just getting rid of the
regular atan2f() call will be a big improvement.
This just adds the functionality, nothing is calling it yet.
|
|
This is just a first stab at threading shapes... whenever shapes
finds itself in a scene it easily becomes a significant
bottleneck, threading is a trivial path to improving that
somewhat.
While at it I also got rid of the need for _checked() variants
which also helps a bit.
But nothing here is proper optimization of the routines.. there's
too much math happening per pixel in a naive fashion.
|
|
|
|
trivial simplification
|
|
Silly typo, one of those fun C instances where it's surprising
how silently mostly-working such a blatant mistake can be.
For posterity:
The way this was even observed as having an affect is while
verifying graceful handling of connections broken while in the
listen backlog.
With an active scener session idle at the prompt, start another
telnet, connecting without receiving any banner (queued via
backlog), ^]cl that backlogged telnet. Then start another
telnet in the same way. Now go to the idle scener session and
quit. The latest telnet would just sit there, seemingly blocked
behind the broken-while-backlogged connection.
But what was really happening was the banner send got the error
on the broken connection after accepting, as you'd expect. This
bug in the errno tests prevented detecting the genuine error
though, leaving the broken session connected indefinitely.
Fun!
|
|
After putting the recv() in a for(;;) to not have to render a
frame per byte received, I completely dropped the ball on moving
the return and adding the continue to actually finish the change.
This makes creating new scenes via pasting long settings strings
far less laggy. A future improvement would be to not recv() a
byte at a time, but this really isn't a perf-sensitive thing.
|
|
This changes things so rkt won't exit with an error @ startup if
RocketEditor isn't already listening. It also tolerates
RocketEditor going away, and will show a "OFFLINE" overlay status
text should that happen w/connect=on.
Some status text has also been added to the "EXIT SCENE" 99999
scene for both the RocketEditor connection and the scener
enabled/disabled status. No indicator yet for if scener has a
connection though, only if it's listening or not via listen=on.
|
|
Remove spurious space
|
|
|
|
Trivial related indentation adjustment too
|
|
This adds a BBS-style interface for creating new scenes in a live
rkt session.
It listens on tcp port 54321 on localhost by default, just use
telnet to connect, the rest is fairly self-explanatory.
This is still early days, but it's a whole lot more than nothing.
|
|
With the "ref" builtin module established and seeming to work
well enough, it looks unlikely we'll need access to unresolvable
module names in the contexts.
The thinking originally was that these names might have special
syntax making them more generic. E.g something like
"@/module/rkt/scenes/[1]/drizzle" for a module_name would have
been supported, which would get resolved either at context create
or even later (as in the ref builtin) at render time.
But the ref builtin is using a path setting, so module names just
stay module names.
Maybe in the future a special syntax will be added for brevity
reasons, but it does make the code more complicated vs. module
names just being names and resolving them entirely at setup time.
Anyhow, this commit does away with the module_name in the
context's scenes. You can still access it via
til_module_context_t.module.name anyways... it's basically just a
resolution-of-name-to-context time constraint that's being
codified now.
|
|
In the interests of improving error handling of interactive
setups, return the setting that was invalid when setup returns
-EINVAL.
For now this is only supported for non-finalized/non-baking setup
calls, i.e. (!res_setup).
In the future I want this also supported for finalizing when
res_setup is set. A lot of the -EINVAL returns are actually
during that stage (or need to be added) due to that being when
the more sensitive parsing occurs going from strings to native
numeric types and such.
The main reason it's not already being done in this commit is
it'll churn quite a bit to get there, since the setup funcs don't
generally have the setting pointers onhand at that phase. It'll
require changing the string-value-centric local variables to
instead all be the til_setting_t* holding those values.
And if setup funcs aren't going to be so value-centric, the
settings API will likely change to not even return the values
directly anymore and only return the full-blown til_settings_t as
the main return pointer, perhaps rid of their res_setting use
even.
Anyway, while here I did some random little cleanups too.
|
|
Make this a distinct heap allocation so it can be enlarged when
editing the scenes... (preparatory commit for scenes editing)
|
|
Preparatory commit for adding an interactive scene editing server
of sorts. It'll go in a separate listing, but needs these types
as it'll operate on rkt_context_t->scenes[].
|
|
I had benchmarked this change and it showed no difference at all
on my 2c/4c i7 X230.
But having just tried it on an RPi4B where it moved the test case
from 54FPS to 60FPS, a +10% improvement, it's worth the
readability loss.
It's interesting how Intel's cleverness discourages optimizing in
ways that benefit probably *all* the competition... even when the
optimization is such a minor change in terms of effort.
|
|
Major gain comes from eliminating the cosf() from the inner loop...
There's still a bunch left on the table for moire but even just
these changes turn 19FPS into 81FPS over here for:
'--module=compose,layers=moire\\\,centers\\\=2\,moire\\\,centers\\\=2\,moire\\\,centers\\\=2\,moire\\\,centers\\\=2,texture=none' '--video=mem,size=1366x768'
|
