Profiling with `perf` revealed the fprintf() inside term_esc_rgb()
to dominate our performance. Replace it with a u8->str lookup table
and a hand-assembled string fed into a single fputs(). On a 80x70
xfce4-terminal geometry, this wins 20%+ FPS on the demo, w00t!
Huzzah for profiling!
* sync up some docs #244
* sync README and man page
* notcurses_output() man page work
* pull attr/channels from output functions #244
* witherworm: clean up explicit moves #244
* still more man page work
* notcurses_lines, last of the man pages i think
* panelreel man page #244
* debian: ruby-ronn->pandoc
* debian: full multiarch compliance
* debian: symbols file
* pandoc: fix syntax for lexgrog
* fm6.mkv: strip audio
* pandoc: fix up apropos man syntax #249
* ncurses_lines man page
* higher planes stomp wide glyphs
* broken unit test
* develop out widestomp PoC
* fix notcurses_at_yx()
* fix up dig_visible_cell() return value
* refuse wide glyph on last column #242
* set adjacent cell wide when rendering #158
* xray: eliminate weird color flicker
* witherworm: don't eat wide glyphs
* unit test for boxed glyph
* uniblock: no need to emit so many U+200Es
* witherworm: remove wide glyph hack
When we emit a glyph that has no background pixels (i.e.
the U+2588 FULL BLOCK glyph), there's no need to emit a
background color change.
Eagle demo currently has hand-coded elision. Results from
80x70 runs using the `-c` parameter:
No optimization: 12.63MiB
Hand-optimized: 12.48MiB
New scheme, no hand-coded optimization: 12.45MiB
w00t!
* introduce doctest over googletest #202
* call dtester in in targets
* doctest conversion #202
* channel.cpp -> doctest #202
* egcpool tests -> doctest #202
* input tests to doctester
* zaxis -> doctest
* drone: always define LANG
* libav to doctest #202
* panelreel tests to doctest #202
* spec that a C++17 compiler is now required for doctest #202
* enmetric tests -> doctest #202
* fade tests -> doctest #202
* notcurses test case -> doctest #202
* last conversion to doctest #202
* finish move to doctest #202
* drone: set up make test
enable mouse events in the demo. spawn a thread from main() which
brokers all input events. it will consume mouse events, and
publish keyboard events to a queue. rewrite all demo notcurses_getc()
calls to demo_getc() equivalents.
Allow -1 in move specification to remain where we are on that
axis (#210), necessary for context-sensitive aligned output.
Add _aligned forms to printf and vprintf. Invert various output
functions so that simpler form is static inline wrapper around
more complicated form, rather than complicated form being a
static inline composition, facilitating atomic move+output. All
output forms now have a simple form (no alignment, placement at
cursor), an _aligned() form, and a _yx() form.
Request and parse up mouse messages. We handle up to 11 mouse
buttons, 3 modifiers (currently thrown away), motion while
holding down a button, and loss/gain of focus. I've added twelve
new NCKEYs: one for each button, and one for release. In addition,
I've introduced the 'ncinput' struct, which encodes the nckey plus
extra data. The only extra data thus far is coordinates for mouse
events. It is not necessary to provide a ncinput to all input
functions; NULL can be provided if the caller doesn't care about
details. All demos are updated. notcurses-input has been updated
to decode full information of returned ncinputs.
The primary resource for this work was Dickey at al's "XTerm Control
Sequences", https://invisible-island.net/xterm/ctlseqs/ctlseqs.html.
* planereels tester #180
* suppress_banner in all tests
* tabletcb: start passing back tablet
* properly initialize fbbytes stat
* panelreel: logic fixes#178
* install all testing data
While ncurses 6.1 adds the `RGB` capability to terminfo, the flag is not
commonly used in terminfo entries as of now. The `COLORTERM` environment
variable is, otoh, commonly present and set to either `truecolor` or
`24bit` value by terminal emulators which support direct color.
Add a check for this if `RGB` is false (or absent)
With this check in I was able to properly appreciate the awesomeness of
notcurses demos :)
* boxdemo: clean up colors
* Fix crash on certain resizes due to corruption of damage map #152
* Rewrite ncplane_move_yx(), throwing out ~25 line of code
* implement notcurses_refresh() #150
Implement a fairly conservative, line-granularity, two-level damage map. One on the overall notcurses object is dirtied by planar moves, creations, deletions, and resizes. One on each ncplane is dirtied by glyph output, media rendering, fades, and erasure. #83 This has some definite false positives: a hidden plane which moves will damage a bunch of lines unnecessarily. For now, don't do things like that :D.
Extra byte per line per plane, shouldn't be a problem.
Two new stats for cell elisions and emissions
Allow keypress to interrupt view-demo
ncvisual_stream() now allows a callback per frame
Allow ncvisual_open() to create its own, perfectly-sized, ncplane #128
Typical performance prior to this PR:
4655 renders, 18.3s total (0.000305s min, 0.196s max, 0.0039s avg 253.9 fps)
401046.505KB total (9.688KB min, 139.697KB max, 86.15KB avg)
Emits/elides: def 1082115/116196 fg 10547624/7236460 bg 10602717/6208644
Elide rates: 9.70% 40.69% 36.93%
4680 renders, 18.4s total (0.000285s min, 0.15s max, 0.0039s avg 255.0 fps)
403078.188KB total (9.688KB min, 139.697KB max, 86.13KB avg)
Emits/elides: def 1088994/116196 fg 10604983/7267750 bg 10655426/6237472
Elide rates: 9.64% 40.66% 36.92%
4699 renders, 17.8s total (0.000227s min, 0.192s max, 0.0038s avg 263.8 fps)
403266.907KB total (9.688KB min, 139.697KB max, 85.82KB avg)
Emits/elides: def 1086511/116196 fg 10601709/7359116 bg 10661910/6326744
Elide rates: 9.66% 40.97% 37.24%
After this PR:
5625 renders, 15s total (9.36e-05s min, 0.187s max, 0.0027s avg 375.2 fps)
168365.640KB total (0.930KB min, 139.600KB max, 29.93KB avg)
Emits/elides: def 310575/116196 fg 4486002/4473416 bg 4116835/4630666
Elide rates: 27.23% 49.93% 52.94%
Cells emitted; 9928000 elided: 12572000 (55.88%)
5642 renders, 14.2s total (9.17e-05s min, 0.154s max, 0.0025s avg 397.0 fps)
168669.009KB total (0.605KB min, 139.600KB max, 29.90KB avg)
Emits/elides: def 310819/116196 fg 4499833/4482134 bg 4118562/4652470
Elide rates: 27.21% 49.90% 53.04%
Cells emitted; 9962160 elided: 12605840 (55.86%)
5650 renders, 14.3s total (0.000118s min, 0.143s max, 0.0025s avg 395.7 fps)
169461.884KB total (0.860KB min, 139.600KB max, 29.99KB avg)
Emits/elides: def 305431/116196 fg 4515396/4456376 bg 4149967/4613668
Elide rates: 27.56% 49.67% 52.65%
Cells emitted; 9945200 elided: 12654800 (55.99%)
on netcurses-demo, we're eliding about half of the total cells via this damage map. that's pretty fucking sweet! FPS increase of about 50% -- I'll take that any day of the fuckin' week, boyo. w00t!
https://www.youtube.com/watch?v=XbGs_qK2PQA
We were doing an entire copy of the framebuffer every fucking
render for no good reason whatsoever. Found this while working
on #83, but let's get it in master immediately. JFC. I suck.
Before (delay factor of 0.5):
[schwarzgerat](0) $ time ./notcurses-demo -d.5
Term: xterm with direct-color indexing
1737 renders, 24.8s total (0.00131s min, 0.362s max, 0.014s avg 70.1 fps)
230217.182KB total (12.696KB min, 223.310KB max, 132.54KB avg)
Emits/elides: def 1074878/126257 fg 6615070/3023957 bg 5519815/3130192
Elide rates: 10.51% 31.37% 36.19%
real 0m47.653s
user 0m8.977s
sys 0m1.414s
[schwarzgerat](0) $
After (same delay factor of 0.5)
[schwarzgerat](0) $ time ./notcurses-demo -d.5
Term: xterm with direct-color indexing
1741 renders, 23.2s total (0.00125s min, 0.351s max, 0.013s avg 75.0 fps)
231551.089KB total (12.696KB min, 223.310KB max, 133.00KB avg)
Emits/elides: def 1065445/126257 fg 6655295/3006679 bg 5566155/3115184
Elide rates: 10.59% 31.12% 35.88%
real 0m46.140s
user 0m8.932s
sys 0m1.350s
[schwarzgerat](0) $
Now *there's* the loss we saw during #139 work, thinking we'd get it
back in #143. Good. Huzzah! But also fuck!
nick black
Marek Habersack