VapourSynth: baixar versions anteriors

- improved gpu up/download speed
- fixed memory model for gpu filters that export memory to cuda/vulkan, input frames they take are copied at most once per source frame and shared between requests
- fixed access violation in matroska muxer
- fixed missing includes in anticipation of libc++ 23 insanity
- logging is now much less likely to create deadlocks
- requests for negative frame numbers are now simply clamped unlike before where it was a fatal error to request negative frame numbers
- setmaxcpu no longer accepts random strings
- fixed avisynth compat having flipped rgb and added rgbs support
- many small bug fixes and improved error messages

GPU SUPPORT FOR EVERYONE!!!

Also some other minor fixes but there's a changelog you can read for that.

- implemented uncompressed mkv output support in vspipe, it's mainly intended to pass on audio, video and timestamps together into ffmpeg
- improved caching behavior for scripts using long cyclical request pattern such as looping 20+ frame clips, the cache heuritics can now over time adapt to much larger cycles
- fixed zimg api version check

- a compiler supporting c++20 is now required to build the project, compatibility has been verified with recent msvc, clang(-cl) and gcc
- relaxed the zen4 level instruction check to not include avx512bf16 since compilers never use these instructions on their own, this allows the faster binaries to be used on intel ice lake cpus and later as well
- added turn90 and turn270 functions
- fixed multiple rare threading issues, the most likely of these to be actually encountered in r76 and r77 was a deadlock that could happen when synchronous getFrame() is called from a filter getFrame function and a pipeline flush is in progress
- added basic avx512 as an additional cpu instruction level meaning that on x64 it's now baseline, avx2, avx512 and zen4 level
- fixed audiogain and audiomix destroying audio with a bit depth between 17 and 31 bits, most commonly 24 bit, where samples were clamped against the significant bit range instead of the range of the container they are stored left aligned in
- added the AudioResample filter for audio sample rate, channel layout and sample format conversion, the downmix coefficients follow itu-r bs.775 and the mix matrix is derived from the two layouts so any combination works
- added waveform, amplitude and frequency arguments to BlankAudio for generating slope, sine, square and triangle test signals
- added chromatic_adaptation argument to resize filters, it defaults to True which is identical to the previous behavior
- mergefulldiff now handles float appropriately
- maskedmerge and premultiply will no longer reject clips with different chroma location if only luma is being processed
- generic filters now checks the minimum plane dimension restrictions only the actually processed planes
- fixed the levels filter not actually reading the second and third plane gamma/max_in/min_in and so on values
- many fixes to the python bindings such as fixing potential deadlocks when freeing the core and using frames and other references after a core has been destroyed (ichunjo)
- added 7x7, 9x9 and 11x11 square convolution sizes
- added half-precision float support to all functions where it makes sense at all (basically only lut/lut2 excluded) in a portable way that generates fast code on both aarch64 and avx2 class x86
- all floating point masks passed to maskedmerge are now clamped to the 0-1 range before applying, this includes UV planes as well
- fixed possible memory leak audionode.output if an error is triggered early in the output process
- the core filters have now been split out and on x64 and are also compiled for avx2 and zen4 class cpus
- fixed proptoclip frame reference leak on error
- fixed potential crash in shufflechannels with multiple audio clip inputs
- lut optimization for all cpus and lut2 optimizations on avx512 cpus
- boxblur optimization
- improved planestats optimizations
- improved merge filters optimization and added avx512 path
- added avx512 optimizations to min, max, convolution and other filters internally known as genericfilters
- greatly improved the memory management and speed in very memory constrained scenarios, speedups of 2-3x are possible if you run scripts with far too low memory max cache size
- now the initial max cache size default is half of the physical available ram (HolyWu)
- remembered to actually remove avfs from the source tree
- several more small fixes for ai found bugs
- added aarch64 neon intrinsics for many for many functions
- added YUV410, 411 and 440 constants for 16 bit integer, half and single precision
- plugin function return type declarations are now checked and warnings will be printed on mismatch, this will be an error in future versions

- vsrepo is now no longer installed automatically for portable installs due to most plugins being available on pypi
- the register-install command can now be used on *nix as well (HolyWu)
- fixed ctrl-c handling in vspipe on non-windows systems
- fixed an issue where certain compilers would occasionally compile the avx2 code path for some internal filters like maximum in a way that could produce pixels garbage along the edges
- once again uses XDG_CONFIG_HOME environment variable for config location, was accidentally forgotten in r74
- fixed issue where caches could be forgotten with frames still in them and leak memory
- completely reworked cache limits so they are much stricter and the number of threads will now be dynamically lowered to avoid using too much memory

- slightly improved memory usage when many frames are queued for output
- improved things for distro packagers
- improved ccfEnableFrameRefDebug flag, it now outputs a list of all allocated frames as a log message after every external frame request has been completed
- fixed config issues on fedora and alpine linux and probably their closest friends too
- plugins can now ship with multiple copies targeting different optimization levels on x64 and have the best one automatically loaded
- simple plugin manifests have been implemented that stops further recursive plugin loading and enables loading of plugins compiled with optimized settings
- multiple improvements to the portable version experience where vsrepo is now installed by default again and forwarding bat files are provided to closer match pre R74 paths

- avfs has been moved to a separate repository/download
- configuration and and handling of autoloading directories has been completely reworked, see documentation for a full explanation
- the whole application can now be installed with "pip install vapoursynth" on windows, osx and linux (HolyWu)
- greatly improved support for python virtual environments
- added ccfEnableFrameRefDebug core creation flag that will log all frame allocations to help find memory leaks
- added getCoreInfo2 function which also returns the core creation flags used
- added the new _Range property which follows the H.273 numbering (0 and 1 switched) instead of the now deprecated _ColorRange property, all filters compiled against api version 4.2 and later must only use the _Range property
- fixed crash in vspipe when using --info with multiple output clips set and discarding all output (--)
- removed remaining leftovers of 32 bit windows support from the source code
- windows binaries now require windows 10
- dropped support for python versions older than 3.12 in order to make vsscript python version independent as well
- now requires visual studio 2026, python 3.14 and inno setup 6.6+ on windows to build

**LAST RELEASE WITH WINDOWS 7 SUPPORT**
- many minor python improvements (ichunjo)
- the api r3 headers are no longer distributed with the windows binaries, general api r3 support for plugins remains for now
- vsscript r3 api support has been completely removed since nobody has used it for 3+ years
- fixed mirroring at edges in convolution filters to be more consistent
- fixed cache state not being updated properly for nodes with no consumers ever and no calls to setcache

- added named pipe output support to vspipe in windows by specifying \.\pipe<pipename>
- fixed a bug that could lead to wrong timecodes being output by vspipe when a clip contains many different frame durations
- the portable install script now allows the installed python version to be specified using the PythonVersionMajor and PythonVersionMinor arguments, it can also handle pre-releases properly now for easier testing
- by default a limited python api compatible module will be compiled when python 3.12+ is used on all platforms
- now supports python 3.12 and later in addition to 3.8 on windows

- fixed vertical mirroring at edges in convolution filters
- improved trim and audiotrim error messages
- updated the supported python versions to 3.13 and 3.8 on windows
- updated visual studio 2022 runtime version

Uses latest visual studio compiler again. See https://github.com/Jaded-Encoding-Thaumaturgy/vs-preview/issues/190 for a workaround for vspreview crashes.

Other changes:

- the inspectable core flag is now correctly detected by the vsscript python environment when an external core is supplied
- fixed a potential crash that could happen when more than 2 audio clips are spliced at the same time

added YUV420PS/H and YUV422PS/H constants for convenience
vspipe --info will now output information for all set output nodes
added a new function to to the vsscript api to get a list of all set output indices, since most users don't need vsscript api 4.2 you now have to define VSSCRIPT_USE_LATEST_API or VSSCRIPT_USE_API_42 to actually require it
the python module will now warn if it doesn't match the core library version when creating a core
updated zimg to fix a rare resizing bug that only happens around 0.5x scaling with avx512 optimizations
now always sets console output codepage to utf8 in windows

fixed portable base path detection, was broken in r67
fixed is_inspectable, was broken in r67
reverted widestring print changes in vspipe from r67

updated the supported python versions to 3.12 and 3.8
added an option to write a full filter graph with time information to file when done processing in vspipe
clarified where different information will be output in vspipe
fixed backslashes and linebreaks not being properly escaped in dot graph output
copyframeprops can now optionally copy only a subset of properties
removeframeprops function now accepts wildcards in property names
added prop_src argument to shuffleplanes
added an access pattern mode when only repeatedly requesting the last frame of a clip
added approximate_gamma option to resizers (saindriches)
added plugin_path property to plugin object in python
audio is now clamped instead of wrapped around in audiomix and audiogain, these functions can also now warn when clipping happens
the vfw wrapper now provides simple waveformatex information automatically if the audio is mono or flagged as plain stereo and samples contain no padding bits
increased the get_frame error message buffer size in python, fewer messages should be truncated now
added missing st428 transfer characteristics to resize
made most graph information functions part of the stable api and bumped the version to 4.1
added new frame request pattern where only the last frame is requested multiple times to further improve cache usage
added new api functions for cache clearing
fixed request pattern set for selectevery with multiple identical offsets
fixed request pattern set for premultiply
expose number of bytes used by cache in python and clean up version and str representations of things (setsugennoao)
fixed modifyframe filter not properly copying its own videoinfo which may cause crashes if nothing else references the format node (jsaowji)
fixed setcache function leaking a node reference (jsaowji)

fixed inspection api sometimes causing memory corruption when used from python
frame properties in python are now return as str type instead of bytes when hinted as utf8 printable
fixed how unprintable data is returned from plugin functions in python, previously it would leak a ctypes pointer with no length instead of returning a bytes object
fixed a bug in the avx2 maskedmerge float premultiplied code path that would switch the two input clips
reverted the from_chars code a bit more to make no locale affect float parsing
fixed the sar adjustment for real this time