To see how the previous implementation was faulty, see either the added
tests, or
- In Aegisub, open a dummy video with a frame rate of 23.976
- Make a subtitle event with start time 04:44.41
- Double-click the line to (supposedly) seek to its first frame
- This will seek one frame earlier than it should, and the event will
not be displayed on the resulting frame.
The new formula is the mathematical inverse of the TimeAtFrame formula
(for the EXACT case). The derivation is the following:
FrameAtTime(ms, EXACT) should be the largest possible integer `frame`
such that:
ms >= floor(((frame - timecodes.size() + 1) * 1000 * denominator + last + floor(numerator / 2)) / numerator)
where `/` denotes division of real numbers (i.e. without rounding down).
The expression on the right is the formula used in TimeAtFrame.
This inequality can be transformed as follows:
ms >= floor(((frame - timecodes.size() + 1) * 1000 * denominator + last + floor(numerator / 2)) / numerator)
((frame - timecodes.size() + 1) * 1000 * denominator + last + floor(numerator / 2)) / numerator < ms + 1
(frame - timecodes.size() + 1) * 1000 * denominator + last + floor(numerator / 2) < (ms + 1) * numerator
(frame - timecodes.size() + 1) * 1000 * denominator < (ms + 1) * numerator - last - floor(numerator / 2)
frame - timecodes.size() + 1 < ((ms + 1) * numerator - last - floor(numerator / 2)) / (1000 * denominator)
frame - timecodes.size() + 1 < ceil(((ms + 1) * numerator - last - floor(numerator / 2)) / (1000 * denominator))
frame - timecodes.size() + 1 < floor(((ms + 1) * numerator - last - floor(numerator / 2) + (1000 * denominator - 1)) / (1000 * denominator))
frame < floor(((ms + 1) * numerator - last - floor(numerator / 2) + (1000 * denominator - 1)) / (1000 * denominator)) + timecodes.size() - 1
frame <= floor(((ms + 1) * numerator - last - floor(numerator / 2) + (1000 * denominator - 1)) / (1000 * denominator)) + timecodes.size() - 2
Since we are looking for the largest integer `frame` satisfying this
inequality, FrameAtTime(ms, EXACT) must be equal to the right-hand side
in the last inequality.
* Bump boost to 1.83
* Switch to building as C++20
Co-authored-by: Ryan Lucia <ryan@luciaonline.net>
* Fix many warnings
* Rip out boost::locale and just use ICU directly
* Fix
* Replace boost::filesystem with std::filesystem
* Fix
* Addditional c++ modernization
* Use string_view all over the place and rip out more boost stuff
* Simply BOM handling in charset conversions
* libaegisub: fix dispatch types
* Fix
* Bump subprojects
* Reorganize tests build file
* Fix remaining compilation errors on Linux
* Rip out the last bits of boost::filesystem
* Remove remaining uses of boost string joins
* Fix some errors introduced with refactors
* Revert "Simply BOM handling in charset conversions"
This reverts commit 2e6b26d5ffe69d7fb280bb1d18c6758112244b64.
Taking out the BOM handling broke tests, so it'll probably break more
stuff.
* Bring back IconvWrapper::RequiredBufferSize
This partially reverts 62befa996126659d139e9ad53e631780b6ed6122 .
This function is actually used in charset_conv_win.cpp
* Fix ifind after moving to ICU
The previous logic didn't check if the match was on parts of
decomposed characters, so it also failed the corresponding test.
* Remove incorrect karaoke_matcher test
This was clearly incorrect and probably just unfinished.
* Remove leftover boost::locale code
* Move iconv include to charset_conv.h
On newer mac sdks iconv_t is defined differently, so it's harder to
just have a typedef for it.
* Fix compilation on arm64 mac
wx uses a different string implementation here, and utf8_string()
doesn't exist there.
* Fix luajit dependency in luabins project
Since luajit always first tried using dependency(), further calls
of dependency() will also always return system luajit.
meson.override_dependency() won't work.
This makes luabins link system luajit where it's available while aegisub
itself uses the subproject's luajit, which causes all kinds of fun
issues and definitely didn't baffle me four hours...
The added solution for this is horribly ugly (and also has problems when
reconfiguring...) but it's the only one I found that works. Maybe it's
better to always require interal luajit after all, or make the user
choose with a meson option?
* Fix locale initialization
Previously this would fail on startup because the automation menu
uses boost::locale::comparator.
... Or maybe the locale init change should just be reverted entirely?
Or it should be something different? I don't really know.
* Revert "Fix luajit dependency in luabins project"
This reverts commit 340fb9c6125412d18f23c08c00f2bc8f46774b6d.
* Fix luajit dependency in luabins project, take 2
Thinking about it some more, just copying the detection logic is
probably the lesser evil here.
* Fix agi::split_iterator after refactor
is_end being removed caused it to not output an empty segment at the
end if the input ends with a delimiter, but existing usages relied
on it doing that.
* Fix style parsing after refactor
* Fix tons of implicit this captures
* Enable CI to test
* Update deprecated hunspell usage
* Fix tests compilation on mac
* Make sure wx subproject builds with c++14
* Fix compilation on Windows
* Revert "Bring back IconvWrapper::RequiredBufferSize"
This reverts commit 04f4b260a08de6497c583f025090ca4f9fe1ef6d.
* Pin libass wrap for now
Apparently dependency('iconv') breaks when iconv is overridden??
* Fix compilation with wx 3.0
* Fix startup crash on Windows
windows.h was defining the ERROR macro, which shadowed the
DialogueTokenType enum variant, which broke the lexer construction.
* Fix SplitText ICU logic
Include UBRK_WORD_IDEO and check the entire rules vec. This now matches
the logic of boost::locale.
* Add test for character_count with \N and friends
* Fix ass_dialogue parsing after refactor
* Revert "Pin libass wrap for now"
This reverts commit 3802bb7272a8bf4861c09b020c9eee9e643804fc.
* Remove iconv's stdbool.h
This was breaking things (libass) and doesn't seem to be
needed any more.
* Revert changes to to_wx
These broke some things, in particular FromUTF8Unchecked seems to not
like empty strings. Probably safer to just revert.
* Fix kara replacer after refactor
* Fix karaoke timing mode after refactor
* Revert "Enable CI to test"
This reverts commit 256cbebbfc9cf4eb5c497898670ac28ce24ce2fd.
---------
Co-authored-by: Ryan Lucia <ryan@luciaonline.net>
Co-authored-by: Thomas Goyne <plorkyeran@gmail.com>
* [git] ignore IDE setting, Meson subprojects
* [win_installer] rename *.mo -> *.gmo
* [win_installer] only add aeg translations
* [win_installer] split aeg/wx translations
* [win_installer] gen & pack translations
* [git] ignore hunspell source dir
* [git] ignore uchardet source dir
* [test] Generate the test executable
run with `meson test`
* [test] add tools to remove test data files
`unset.bat FULL_PATH`
* [test] set to correct Error type
* [test] Set Source character sets to UTF-8 for MSVC
* [test] move test data files to build_root
* [git] remove IDE dir
* [git] ignore gtest subprojects
* [ci] run test in ci
* [meson/test] use more meaningful name
* [test] Add more comments and help msg
The thes file was using \r\n line ends, and \r was not stripped, causing the encoding string has a \r at its end.
Use binary mode and always use \n as line ends.
$(subst .c,.o,...) replaces '.c' with '.o' everywhere in pathnames. For
example, renaming the "Aegisub" folder to "Aegisub.cool" will make the
build system generate "Aegisub.oool/.../.o" objects.
https://www.gnu.org/software/make/manual/make.html#Text-Functions
boost::split_iterator type-erases the predicate, which makes it require
a virtual call per character (!) along with a heap allocation. As it
turns out we only ever need one predicate (comparing to a single
character), so replace it with a split_iterator that just does that.
