stringstreams were shown to be slow when reading/writing single characters
and there is lots of potential by having functions perform on existing
buffers whenever possible.
* Rewrite shadow filtering for the new distortion
* Calculate penumbra radius using a single sample
* Avoid peter-panning effect due to filtering of short shadows
* Add adaptive filter quality for soft shadows
* Avoid sharp shadows on surfaces without normals (e.g. plants)
* Increase default and maximum soft shadow radius
* Make line numbers in shader errors match the code
This reduces the
(absolute) time spent in Server::SendBlocks() from 700ms to 300ms
(relative) share of MapBlock::serialize() from 80% to 60%
in a test setup with 10 players and many block changes
The logging streams now do almost no work when there is no output target for them.
For example, if LL_VERBOSE has no output targets, then `verbosestream << x` will return a StreamProxy with a null target. Any further `<<` operations applied to it will do nothing.
If a newly started thread immediately exits then m_running would
immediately be set to false again and the caller would be stuck
waiting for m_running to become true forever.
Since a mutex for synchronizing startup already exists we can
simply move the while loop into it.
see also: #5134 which introduced m_start_finished_mutex
This commit introduces mt_localtime() in src/gettime.h, a wrapper
around the OS-specific thread-safe versions of localtime()
(resp. localtime_s on Windows and localtime_r in other systems).
Per the Open Group recommendation,
«portable applications should call tzset() explicitly before using
ctime_r() or localtime_r() because setting timezone information is
optional for those functions», so we also do a one-shot
call of tzset() (_tzset() on Windows to avoid warning C4996).
The function is used to replace the localtime() calls in
getTimestamp() and makeScreenshot().
(The only reminaing call to localtime() in the tree now is the one in
the local copy of the Lua source code.)
As reported in #12197, b0b9732359
introduces a regression in worldalign textures.
The specific change that seems to be responsible for this issue is the
change in order between the computation of the cuboid texture
coordinates and the box edge correction.
Fix #12197 by moving the box edge correction back to before the cuboid
texture coordinates, as it used to be.
Since b2eb44afc5, a texture defined as
`[combine:16x512:0,0=some_file.png;etc`
will not be sent correctly from a 5.5 server to a 5.4 client due to the
overeager detection of unsupported base modifier `[` introducing a
spurious `blank.png^` before the modifier.
Fix this by whitelisting which base modifiers can be passed through
unchanged to the client, and prefix `blank.png` for the others
(which at the moment is just [png:, but the list may grow larger
as new base modifiers are added.)
Stop scaling images to POT immediately when loaded. The 'combine'
modifier hardcodes X and Y coordinates, and so behaves incorrectly
if applied to a scaled image. Images emitted by generateImage()
are already scaled to POT before being used as a texture, so
nothing should break.
* Move shadow position calculation to vertex shaders
* Animate entire scene before rendering shadows to prevent lagging of shadows
* Remove unnecessary use of PolygonOffsetFactor
* Apply normal offset to both nodes and objects
* Rename getPerspectiveFactor -> applyPerspectiveDistortion
* Remove perspective distortion from fragment shaders
