this is a fuckton faster lol
the old one did a gazillion draw calls now we do two, one for the fuzzy and the 2nd for the bars as both are handled within a single shader
thanks to RRs heat effect being super similar to the vhs bars lul
basically ported over from RR
water was massively toned down and adjusted, it visually almost matches software at default res (note that the software effects are not properly scaled lol)
the heat effect is just the water effect adjusted to look like the fixed function version from before but slightly slower to accomidate the slower water effect here
this should speed things up a little since we now only need ONE draw call instead of 20 per effect
small note: this does not zoom into the image like the shaderless version, which was quite jarring tbh but there can be some very slight "stretch" marks at screen border, but theyre extremely unnoticable
man idk why i even did the whole desktop res shit
somehow kept forgetting we should just check against the drawable/window size as it determines the main framebuffer size
so we just can use SDL_GetWindowSize and get rid of all the other crap
well "back" we still had it there, but the subsector marking logic can just be there
was kinda too lazy before to do this lul
saves some time as we dont need to go thru all subsectors again at mapload, which there can be many of
not super cool but seems to work just fine
the new subsector flag may even improve performance in some cases since the horizonspecial check seems to mess up the branch predictor often lol
quite a mess but "works"
the reason this didnt work before was
i mostly tried doing this per multisample textures which you cant really copy from
now we do it per color rbo which you also cant really copy from, but is the recommended way
so now each time we try to copy from the screen (screentextures, screenshots, etc.) we blit the msaa fbo into the normal fbo so we can copy from that
its probably a little slow, but cant really do much about it besides getting rid of screen textures and replacing them with more fbo logic.....
while this can look nice in some cases with msaa, it has the effect that everything transclucent is twice as translucent as it should be besides just being a kinda confusing option for almost everyone
so just delet it and simplify code
nearest-neighbour filtering is good from upscaling from a lower resolution, but will produce super ugly mixels and shimmering when downsampling from a higher resolution
while not compeletely needed, we use the stencil buffer for some stuff, so dont really feel like having to deal with potential issues here
besides, probably any gpu in almost the last 2 decades should support this and even gl3.0 anyways, so this is just bloat we do not need
dont do locale stuff from thread, use localtime_r && localtime_r as localtime is not threadsafe either
G_GetDemoDate returns malloc´d buffer again, tho not strictly necessary
also, fix dates not always being initialized in all cases and type issue with G_GetCreationTime
this whole sky handling is just incredibly stupid and wouldve been such an easy fix, but here we go
this is just more stupid hacks but i really dont feel bad here since the code already hacked around a bunch of shit
just hoping this wont cause any issues on some maps and needs testing
sigh
not sure why there werent any, the other map lumps have those, and a corrupted file may have those missing also which would just crash instead of telling you wtf is wrong
It was probably caused by menupath being modified when replay load
thread is in the middle of checking it, now the check for "valid"
replay load thread is simpler. Also as comment says, there still are
few reported data races for replaynamesloaded but i THINK it should be
fine, since the only "data race" that could be harmful here is
replaynamesloaded switching from true to false, which would mean
demolist_all is suddenly invalidated, and that shouldn't be possible
because we only switch replaynamesloaded from true to false inside main
thread while also locking the "search mutex"
(this is so wordy lol)
matches software now
even if most custom maps were made with opengl in mind, stuff did not work anyways for years so this shouldnt hopefully cause any issues
it should check its own "master" sector for remap behavior NOT the sector it resides in
note: i did not change the ceiling/floor behavior here as i noticed there might be more issues than this with fofs
W_CheckFileContains returns -1 if file was not found which still triggered those conditions
so just handle if file was not found to skip doing other stuff in those cases
read the sectornode list stuff
this means we reuse old stale data in some cases, which also causes strange bugs with orbiting items in some cases, but thats what vanilla does and we HAVE to keep vanilla compat after all
i really hope this does not cause any fun impossible to debug crashes, but what can you do
I was thinking to myself about how 2.1 and before did thinkers and how 2.2 made it into thinker lists and thought "wait isn't memory allocated similar to that?" and well here we are.
this appears to mess up remapping behavior in certain cases (ex. sneaker/spring panels)
since i cba looking into properly fixing this, ill just disable this for now
have a distinct name for that build "Phobos", as a nod to Doom
also invert the flag, as it should make adding stuff less headachey
will make this a compiler flag later but too lazy atm
can just be enabled per MAXPLAYERS define in doomdef
this will also auto undefine COMPAT_VANILLA
no support for tab rankings or intermission rankings at all yet
will port those over eventually from saturn 32p or blankart and clean stuff up
please note this may have bugs since not everything is ported over yet
prob did that while sleep deprived, P_WeaponOrPanel includes a few more things than intended to be there
doubt we even use any of those, but i dont want to cause any potential issues agaion
dont think this would cause any issues at all, but i dont feel like maybe causing some dumbass edge cases anywhere really
can test this further after v10 release since its already taking ages
this omits alot of hooks in the profiler, tho not sure if anyone should even care about those back in the list
maybe just have it have multiple pages like perfstats, or better just integrate it into perfstats......
i dont want another stinky library
this would just be an arsepain to include in windoze builds
TODO: make this drawer less of a mess, maybe integrate into perfstats instead?
-split emeraldhunt handling into its own function
-merge hoop spawn loop into the main mapthing spawnloop
-continue early for things like rings and coins
- reduce the amount of R_PointInSubsector on maps with nights hoops
hopefully reduces levelload time slightly
this "works" with a skinlimit of 255
but would cause hard to debug crashes with raised skinlimits and was generally really unsafe and could cause soundnames to not be null terminated aswell as buffer overflows from blind strcpy
the sprintf fix is from Indev originally for neptune
NOTE: if the skinlimit is ever raised above 1024 then freeslotnames NEEDS to be atleast one byte bigger to not truncuate stuff
NOTE2: pretty sure localskins are fucked if any are loaded in servers with 255 skins loaded, but i dont care atm
when using localskins and going over the skinlimit
this should use numskins not numallskins there
no i dont feel like supporting localskins here the other skin commands also dont care
this sucks
but the final screen texture with the palette render postprocess step will not be in the downsample fbo
so we have to redraw it there with the postprocess step......
this is still better than drawing another screen texture when using paletterender which costs perf
still horrid, but atleast this wont give me braindamage each time i have to look at it
this also read stale data from the previous frames and was just insanely messy
this still sucks ass but oh well
this basically brings it to the latest state from Doom-Legacy before the big refactor (which i had previously ported and tested as well) + some extra stuff from it
sadly this does not seem to fix plane holes, but its a bit cleaner and more accurate in some places
while the big refactor did fix issues with plane holes, it also introduced a few new issues, like z-fighting sneaker panels and significantly longer loading times
may as well just wait at this point for what (or rather if) RR will be doing
most of those were added cause they´re cool to look at, but without much thought about potentially giving players that enable such features an "unfair" advantage over others.
- eggboxes wont be rotated while stationary and on the ground anymore (they are meant to deceive players, being rotated makes them stand out alot!)
- players that are about to explode due to eggboxes wont spawn extra red speedlines anymore
- option to show driftlevel per speedlines has been removed
dont need all those yaw vars, since they effectively map to gl_viewsin and -gl_viewcos respectively
while those are less accurate due to being fixed point, there appears to be no visual issues from this and seems to even slightly improve performance in some cases
move stuff into a helper func, which just returns if mo != NULL in vanilla compat mode, otherwise returns P_MobjWasRemoved
reasoning for this crap is in the comment above said helper
very scuffed up since we dont have the srb2::Draw lib
so it may display stuff out of screen bounds
i also rather not use fmt here but va since i dont want another dependency
needs to be fixed up sometime
- Add lua_profile cheat: calculate average time spent in
hooks over a period
- Show hook times in a sorted table, along with lump name,
script line number and hook type
- Show cumulative time spent in game logic hooks and
percentage of overhead to game logic
stuff added per -file or whatever
this incorrectly rechecked the iwads again instead of the pwads themselves, which let to the game flagging vanilla maps as replaced, setting the game to be modified
pretty sure none of those cause issues currently, but rather not risk stuff as that condition might change from smth else in those cases
its not an expensive call anyways really
before this flagged every file with lumps that contained the MAP as string somewhere, like palettes and such
this is kinda messy but i really do not want to dupe this huge block of code, neither do i want to convert this to ceepeepee again to mess about with templates
oh well
never seen srb2, saturn, RR or blan or anything else have this return null
there is a possibility but whatever
just continuing in such cases was wrong eitherways as it wouldve just corrupted the vres thing
per https://git.do.srb2.org/STJr/SRB2/-/merge_requests/2784
chat offsets and character count
offsets are added to the base offset like we do for hud stuff, the og pr will directly use the cvar values for this, but wanna keep consistent with our rest
needs to be added in the menu somewhere but idk where
theres guards just above this which should just skip shit but for some reason the clamps still do stuff, idk man
fixes the weird grass fof in green hill
akin to RR, just toned down for smaller kart chars
also toggable ofc
theres not that many stairs in kart but this looks cute with smol bumps in the road
wish this would work on slopes but theyre kinda messed up lol
noticed sloperoll does not properly work from certain angles
so just reset rollangle if needed in K_RollMobjBySlopes since couldnt find a better place tbh
decided to have it i_error for now, not sure what exactly is the best approach here
should we just "correct" some of those instead?
vanilla probably just crashes anyways
completely forgot about this (sry!)
the original implementation was mostly based on hughsk´s 8x8 bayer shader while the current is mostly adapted from the libretro 4x4 bayer shader
set using the front side upper texture field like the play sound special per "DSXXXX"
the parsing part already was there but not the actual code to make it use the specified sound for some reason lul
shouldnt break anything and is vanilla compatible!
wadfile->filename has not a fixed size, so it theoretically might be bigger than wadfilename buffer
this probably wont ever happen but better be safe than sorry honestly
the "already loaded" check will not first compare the md5sum, which is a very simple memcmp, and only then it will do string copies, nameonly mutation and name compare should the md5sum match
this should save quite a few unnecessary string operations since the md5 match case will rarely ever hit
stuff like that should always be checked per equals/greater than just equals
in practice this shouldnt have caused any issues but its just better practice to check for greater
check for subsector validcount first, then if its a sky plane
so we can skip checking sector heights for sky planes, which may save a bit of maths on maps with many slopes
* Don't overwrite the first player's name when joining with multiple local splitscreen players. (resolves#151)
* The defaults for unprovided names are now consistently "Player A/B/C/D" as opposed to some being letter-based and some being number-based.
* Send local player info once per local splitscreen player, rather than n^2 times (where n is the number of local splitscreen players).
* Since the packet *requires* four names to be sent every time you join a server, send dummy names for local splitscreen players who aren't connecting, instead of your actual cvars (no data leakage you didn't ask for!)
fixes a 4byte hole so we can fit everything important into cacheline 1
even when compiled without HWRENDER this should be a better layout
checked per pahole
again ive never seen this hit, but the mobj gets checked right below its radius get dereferenced within the floorz/ceilingz functions which kinda annoyed me lol
as a fallback i just the floor/ceiling sectorheights, which were used before the addition of slopes
this has an insanely rare chance of happening in certain cases (train rush map has a few spots where this is somewhat likely)
abs does not really handle -2147483648 and its result is pretty much implementation defined/ub, since we compile with fwrapv flag it would just return -2147483648 which will cause this to try and access tantoangle at a negative index
we can just return INT32_MAX here since treating either as 2147483647 would return 2147483647 eitherways
decided to finally make smol stickers for the beta hud
this was never a really great solution since it made things look pretty bad at certain resolutions/scales
still keep the loading code for convenience but honestly not needed aswell at this point
also makes this alot more readable
like this most of the time we dont have to ever check the players floorz nor ceilingz at all since we check if the sector the players mobj is in is even offroad at all
nothing super big but i did see an improvement of around 1fps in my usual benchmark
on my linux system and usual benchmark setup this improves performance quite significantly (+~12fps)
this relies on jemalloc to be installed as a separate library for now and is more a convenience thing
does not fix whitescreen when switching resolutions while in intermission screen, but oh well
lets hope there will be no issues with calling D_Renderview in intermission state, but it seems to work from quick testing lul
otherwise this may still get truncated and lots of text editor also complain about a super long single line of text
also fix null ptr dereferencing into strtok should strdup return null
now works more like i intended too by more or less scaling the pattern to be roughly the same size on each resolution and trying to reduce precision issues by using floor to convert float values to integer range
not the most efficient but idk a better solution to this, and clipping code may cause some other undesired effects, but well were not supposed to be drawing textures with holes on single sided walls eitherways, but here we are
basically this copies columns into a buffer filled with cyan pixels, so the holes become cyan instead of nothing, which we can cut again using the twomultipatch drawers
fixes the "melt" effect on some things like spelunky or music plant (and well reading/writing oob)
like normal planes, polyobject planes will just reallocate their vertex buffer eitherways according to size, so that check does not make much sense, srb2 also did remove it a while ago
apparently reusing a va list like that is ub, so we should just copy it if we want to reuse it
also fix the resize check, since vsnprintf return does not include 0 terminator which we need to account for
also make this part of dehacked.c a bit more readable lol
-handle more memory allocation failures
-fix null pointer dereferences
-prevent writing out of bounds while writing filenames for D_ParseFileneeded
-prevent reading out of bounds while checking for new target in director
-fix Update_MS checking against an unitialized value in some cases
-fix assert for AddLumpToCache since strlen on non null terminated strings is UB
-cleanup and rearrange R_DrawColumn
-cleanup musicdef parser a little
all in one commit cause idk would be ugly if those were seperate ig
oftentimes lua scripts call CV_FindVar a fuckton of times during runtime, with many cvars this can actually tank performance more than one might think
ideally lua scripts should cache the result, but with so much stuff and servers we cant expect everyone to go through all scripts to do that
using a hashtable for CV_FindNetVar might improve join times aswell, but needs further testing to confirm
The constructor attribute causes the function to be called automatically before execution enters main (). Similarly, the destructor attribute causes the function to be called automatically after main () has completed or exit () has been called. Functions with these attributes are useful for initializing data that will be used implicitly during the execution of the program.
added "used" attribute aswell so compiler does not omit non referenced functions (i hope)
gotta test this a bit but seems cool for stuff like verstable maps
apparently demos used to send a mapcmd before instead of directly calling G_InitNew, so this function was never called which meant that timedemos always returned a wrong result first time
so both CONS_Alert and CONS_Debug dont have to handle and copy the varargs twice
probably makes no actual difference in practice, but i prefer it to not do pointless work
removes the "force_mobj" field, gives a suprising perf boost from my timedemo testing +6-7fps (164 - 169/170fps)
this does however will prevent the generic hook from being called if the mobj being passed ever should be NULL
This shouldnt ever happen, vanilla assumes this aswell and way worse things would happen there, but just wanted to note
as per comment, this is merely a shitty workaround for a chain reaction of bad fixed point maths
since im honestly not feeling like rewriting most of the software renderer to either operate on 64bit or float maths, we just skip drawing anything that seems completely unreasonable
taken from raptorjit
we can check the hash first to skip quite a bunch of memcmps which should be faster in most cases (hopefully but quick timedemo tests confirmed that)
the string hashing was the most hit and most spent in part of the whole game in alot of cases
replacing it with luajits algo absolutely massively improves lua performance (10-15fps in my bench timedemos!)
however idk how all this stuff exactly works so this needs alot of testing to make sure nothing breaks
ive rewritten the unaligned ptr access slightly to use memcpy for safety with some platforms that require stricter alignment
luaV_execute should compute 'ra' only when the instruction uses it.
Computing an illegal address is undefined behavior even if the address
is never dereferenced.
seen this a bunch of times now that occasionally something tries to access uninit values from a surface info
since those are really small we just always init them now
not sure if this is more sane than the huge ass array from before
but atleast this is more accurate and can basically generate flashpals from any palette while also still working with all the colour and gamma adjustments
prob needs some cleanup idk this is all rather messy lol
iirc, if the provided palette currently in use did not provide subpalettes, this will read past the bounds of allocated memory, which aint good lol
i just dumped the white flashpal and threw it into there since idk if -file or autoload gets called before the palette gets initialised, which would carry a risk of copying more garbage, hence this big ass array
lump name may not be null terminated here, so strlen may read out of bounds garbo
using strnlen ensures we read max 8 characters which is the name buffer size
been rethinking this and been doing some digging online
the unordered_map was just a quick and simple fix for performance issues in blan, but came with caveats:
- unordered_map is simple to use but not particulary great in terms of performance during lookups especially
- we needed to convert w_wad to c++ just for the lumpnumcache
- due to needing to preserve the key since they gotta outlive the scope of the function that inserts things into the map, we effectively did a string copy on each map insert and lookup
after some research came across verstable, which is basically almost the exact same as unordered map but in C while being extremely fast (even more with its builtin string hash and comparisor)
only caveat with verstable is we need to rely on a new lib, though its just a single header file, and we gotta manage ensuring the input keys outlive functionscope ourselves, which we do with strdup in this case
good thing, verstable allows to call a "destructor" specified by us whenever it removes or replaces keys in the map
this also means we only need to copy strings on insert and not on lookup anymore
all in all this is much faster while giving us the simple usage of c++ unordered_map while staying in C
due to the strdup quirk i had to slightly modify verstable to allow to take const as inputs for the getter functions even if the keys themselves are not const marked with "KEY_GET_CONST"
Verstable can be found here: https://github.com/JacksonAllan/Verstable
and is licensed under MIT which is perfect!
this version changed so that many standard functions now propagate the constness of its input pointers
to not cause warnings for everyone on older versions this is now a really dumb macro lol
we *rely* on a bunch of gnu extensions so its better to explicitly set this
yes in theory we fully comply with c23/gnu23 and saturn compiles fine under that, but since its still fairly new theres alot of maschines that still dont support it :chaosleep:
every "flag" set by signal handlers must be of type "volatile sig_atomic_t"
everything else may fail to be correctly set, read or get optimized away which rendered this basically useless
the earlier approach still raised the default signal within this, which basically just skipped everything and ended up NOT calling I_Quit at all
Basically this is to ensure we dont call I_Quit at all from a signal handler as it is NOT async safe and can cause serious issues like deadlocks and such, instead SIGINT and SIGTERM will just signal the main loop/thread to safely call I_Quit instead
lumpnumcache can be case sensitive W_CheckNumForName should also look for uppercase stuff and W_LumpExists is case sensitive anyways
this avoids having to always account for case (toupper/tolower) during hash and string comparisons and even allows us to even directly compare our strings instead of using fastcmp (std::strings allow this!)
from some benchmarking externally this makes this faster quite a bit
prefer to use standard lib functions if possible
lrint does the same thing (well ofc as long as no one touches the rounding direction löl) and is even quite a lot faster (though this is negligible in this case)
Older versions break and have unstable ids on Windows, cant check for linux but ill just add a version check at this point
This was part of updating sdl with v9, which got reverted due to other issues but this was not reverted oopsie
-add LTO option, enabled by default
-set optimization level to O3 and O0 for debug builds
-add ASAN and UBSAN options
-always compile with gnu debug symbols
dont call rand within the loop, use xorshift instead
while not massive, we dont need to call rand thousands of times, the xorshift is visually indistinguishable
this is still rather slow due to the crapton of draw calls and should best be replaced by a shader or something
whoops, not sure why that was even done like that
string.c will include strcasestr.c and SRB2_HAVE_STRLCPY will ensure eitherways we dont use the included strlcpy stuff
this shouldnt account for highreshudscale at all
on very high resolutions the bars are still only half size, but atleast they dont become minuscule or just dont draw at all
also ensures both bars are clamped, on some resolutions sometimes they just go out of visible screen range
this is incredibly ugly and hacked together
theres some weird culling issues and shit doesent really "scale" with different resolutions which needs to be fixed
but it kinda works
toggles certain effects, which may be troublesome to some players
-disabled hyudoro flashing, instead makes player translucent
- Makes rainbow colour effects, such as invul, change colours very slowly
- Disables underwater and heatwave screen effects
- Disables palette flashing
- Disables vhs effect
not the most pretty implementation, but im lazy and this works
feel free to suggest more things
toggles certain effects, which may be troublesome to some players
-disabled hyudoro flashing, instead makes player translucent
- Makes rainbow colour effects, such as invul, change colours very slowly
- Disables underwater and heatwave screen effects
- Disables palette flashing
- Disables vhs effect
not the most pretty implementation, but im lazy and this works
feel free to suggest more things
properly match the old linear lookup by returning the FIRST match for doomednum
also shrinks the array, doomednum max is 4095, no need for such massive array, also fixed a off by
should fix some join desynchs and desynchs on certain maps
bunch of smaller issues before, no real checks for strltol and potential free on a null pointer lul
also skips replays with shorter than 10 char names and handles errors better
This seems to improve performance on maps with moderate to lots of fofs quite a bit.
On maps with only a few fofs, this makes no difference or has very very little regression
Need to test and check if there are any visual errors, but so far everything seems good
no need to constantly reallocate new memory
we dont have renderer switch, and the next level will render drawnodes anyways
those wont grow super big also, so theres no big worry about wasting tons of memory
yes i like ceepeepee vectors, how can you tell
While i assume this to be fine currently, i quite frankly dont trust the srb2 codebase enough aswell as preventing any future issues if there should ever be need to switch locales, since toupper WILL change behaviour in such cases which does not qualify as pure
this might let the compiler optimize aways subsequent calls in some cases
strcmp etc from glibc also does it
prayin nothin changes the locale at runtime lmao
fixes an issue for me which led to having the game threads getting stuck when sending an interrupt signal to the game due to some thread issues with the sdl subsystems
shoutout to dsda doom for this
well this still aint working all to great, but atleast this allows me to quit the game if it gets stuck without having to always send a kill signal to it
copy pasted from what the dedicated build does
we dont use it for anything, we also cannot use it for anything.
Due to massive amounts of unterminated string being passed to everything
So we dont need to waste memory and resources
in two parts:
- Use our lumpcache lookup first (our unordered map), BUT since it is case insensitive we extract our lump out of it to do another case sensitive name compare
- In the manual lump seach falltthrough case, we can also check against the string lenghts and hash (its case insensitive but will still reject differing names, same with stringlenght)
this improves the performance of the leaderboard map browser quite a bit and everything else using v.patchExists
needs testing if actually better
but the old lumpnumcache was basically useless due to being incredibly small
even just booting up the game we have around 1000 entries already
this hopefully should still be faster than potentially going through ALL lumps in ALL wads we have loaded to find the lumpnums
not sure if longnames should get any special handling like before, but i dont think it really matters if theyre long honestly
thanks ceepeepeereference for the basis of this
Apparently in some cases the server will connect to itself, in this case the server tries to send itself the savegame, leading to a crash in some cases. oops
- expose SAVEGAMESIZE
- pass node number to SV_AddWaitingPlayers (saves iterating through all 128 net nodes everytime a player joins) lifted from RR
- move all packet handling to the packet handlers
- Reorder packethandling switches a little, so common/important packets get checked first
- fix wrong clamp for realfirsttic that was too early
- add nowtime division by NEWTICRATERATIO (no change unless ticrate changes)
shouldnt change the behavior much as we wont see anything without a camera, but this is explicit about what exactly is wanted (i.e. thiscam may be null, we only check due to null deref, we only care about if freecam is active)
phew so every line now has two names
otherwise we cannot fit everything and anyone into this anymore
this even allowed to put back the tooltip at the bottom
i still want to redo the credits completely at some point with some cool stuff though
- Fix serverlist results not being properly reset, which lead to reopening/refreshing the serverlist to show the old state, selecting one of those "old" servers then lead to just connecting to nothing
- Fix the resfresh button making servers disappear if hit too often too fast (from srb2)
https://codeberg.org/srb2classic/srb2classic/commit/3acc312923da0fe5e07c0517d9bff9e8b657b1eb
due to ghosts occasionally ending to early and game interpreting them as still valid uhhhhh
lets pray this wont explode or cause any weird issues......
credits to RR
REVERT THIS ONCE I GO AND FIXUP THE REPLAY MESS
clips sprites that are out of our view, especially with precip this may lead to significant speedup in certain cases
huge thanks to indev for helping with this <3
huge thanks to dsda-doom for the maths needed!
in both cases "R_Subsector" already got the frontsectors ceiling z, so no need to get it again within R_Prep3DFloors
should save some P_GetSectorCeilingZAt per frame
since we mostly use it in fixed point maths, this just makes alot more sense
moving it into inline functions ensures that in C code this wont evaluate its arguments multiple times
range from 0.25 to 4
to account for all types of controllers and such
base 1 was set around what felt good on my dualsense, not entirely sure if this is the best baseline to go with
otherwise it could still be adjusted
big oopsie from my side
this caused textures that werent translucent to be marked with translucent in a bunch of cases, which led to them being rendered in a slower path
I_LocateWad returns mallocd memory that needs to be freed
and our home search does not need a tmp ptr
just do as we did before and sprintf it into returnwadpath
also check home directory before the hardcoded lists like before (might speed up startup a little on nix)
also cleanup iwad search in dedi builds and apply everything there aswell
- skincolors are uint8 type not uint16
- ensure desired colour is within skincolor range
- rename G_GetSkinColor to G_GetSkinColorForGamepad
- replace pointless array with single variable for current color in D_DeviceLEDTick
- remove pointless call to G_GetSkinColorForGamepad when hotplugging gamepads (G_SetPlayerGamepadIndicatorColor will call it itself when 0/skincolor_none is passed)
honestly doesent make that much sense in most cases to grow capacity * 2 (slightly misunderstood some stuff on implementation)
and in other cases dont need to manually check for resize
we do culling per our viewmatrix since a while, which wasnt properly set up for portals beforehand and used whatever was last which was then incorrectly applied to portal culling :chaobaba:
players first, then kart items, shadows and then some srb2 things that are commonly used
makes more sense to check for all of those first everywhere, so we may skip a crapton of pointless checks in many cases
avoid mixing integer maths AFTER casting to float types, even though its all double this may lead to wonky precision issues
I_GetPrecisePrecision now also doesent always call SDL_GetPerformanceFrequency everytime, since its return value will never change since its solely dependant on the platform
and also removing some dead code
- use sqrtf for fade/tint alpha for shaderless light calc
- save recalcing the cosine/sine from angle on angled flats 3 times in HWR_RenderPlane
- precalc float "int32 min/max "
continue the loop, not return, it might just be an empty line, so returning might break some luas
0 lenght would make realloc allocate nothing at all, so just skip it
there were a bunch of cases where inf and nan float results were casted and used with integer things, which is undefined in c
yes i did use ceepeepee for that crap, cause this wouldve been really annoying without templating
if someone really wants that crap to use outside ceepeepee, feel free to make a bunch of wrappers cause im too lazy for that stuff
i.e. since we allow texture filtering on translucent textures now, the min filter may specifcy one of the mipmap filters, this means we NEED to call GL_GENERATE_MIPMAP and set the LODs for them, otherwise no texture will be generated
alternatively we could do a bunch of special cases for the mipmap filters, but this is simpler
actually replace everything correctly
my dumbass forgor about the fast both strcmp and stricmp matter in sort stuff so we cannot use the fastcmp stuff there, which means we should not redifine stricmp as fasticmp in case something relies on that behaviour ugh
apparently some gl drivers may use a slower path for streamed textures that only get partly updated, this tricks them into thinking this did a "full update" and resulting in using a faster renderpath
This used to be used for `P_RelinkPointers()`, but was superseded by the far saner mobjnum. HOWEVER, its lingering effect was to leave `mobj->info` in an invalid state until a later `P_FinishMobjs()` was called. This is memory unsafety :D
After removing the last remnants of this ancient hack, it is now once again possible to connect to a server without crashing immediately. However, I did get a crash after a few seconds, so there is definitely still *something* nasty going on under the hood.
* Make sure the consvar's properties are all initialised to zeros as defaults
* Error if the consvar is not given a name
* Error if the consvar has CV_CALL but no call function
* Error if the consvar has CV_NOINIT but not CV_CALL
See c92378fa and a7e20f70. I didn't realize that light
userdata's metatable is shared--like numbers or strings.
So it cannot be paired with a metatable.
I also made a few minor tweaks to Lua cvars, other than
accounting for the double pointer in the userdata.
while this is more accurate to how the player experienced stuff in the actual game, this also can be somewhat jittery and stuttery depending on the net conditions the replay was recorded in, so make this toggable if someone wants it
fixes an issue where in some rare cases, the game may get stuck in a weird limbo state due to the text command buffer being full and never be cleared, so net commands can never be executed
struct with two sets of maptextures, one for normal another for encore textures
functionally identical, hopefully also performancewise
more explicit than doing this with just one set and indexing stuff :bloblcatpats:
- SOCK_Get return true when receiving good packet from a lan player
- SOCK_GetAddr fix handling of ipv6 addresses, so ipv6 servers can be found in the serverlist
taken from srb2/classic
we only need R_ExecuteSetViewSize from splitscreen_onchange, the rest doesent do anything due to SV_StopServer call earlier setting cl_mode to CL_ABORTED
SV_StopServer also calls Y_EndIntermission and Y_EndVote if necessary
by reusing the special sector calculated earlier, this didnt change for when its checked another time
also one fixedmul in P_SceneryTryMove in some cases
allow sending a fake player to the ms and serverbrowser when your dedicated server is empty
effectively functioning as a "seed" player without manual intervention and occupieng a server slot
lots of confusion
but it does more brightness adjustment than gamma adjustments, yes this does some gamma correction but not user changable global gamma adjustments like globalgamma
uhhh
:shook:
use the CLEANUP macro everywhere
init some things that may not be initialized
and make this more readable
lets hope didnt break anything Zzz...
TODO: if there are still issues with demo records and stuff crashing might consider porting some of the fixes from RR
just treat them as "normal" buttons instead of manually polling for them
alot cleaner and gets rid of yet another polling loop
also clean up some other stuff
needs to be tested with multiple controllers in splitscreen but cba to do that atm
Fixes keyboard input issues for certain languages when using the "native" keyboard input setting
probably fixes some other issues aswell
needs regression testing
in theory this should improve levelload times and reduce memory usage?
gotta see if it really does, visually so far there seems to be no difference at all
message(FATAL_ERROR"In-source builds will bring you a world of pain. Please make a separate directory to invoke CMake from.")
endif()
# DO NOT CHANGE THIS SRB2 STRING! Some variable names depend on this string.
# Version change is fine.
project(SRB2
VERSION1.6
LANGUAGESCCXX)
set(CMAKE_CXX_STANDARD17)
set(CMAKE_CXX_STANDARD_REQUIREDON)
#set(CMAKE_CXX_EXTENSIONS OFF) # optional but recommended: disables GNU extensions, forces pure ISO C++
set(CMAKE_CXX_EXTENSIONSON)
set(CMAKE_C_EXTENSIONSON)
##### PACKAGE CONFIGURATION #####
set(SRB2_CPACK_GENERATOR""CACHESTRING"Generator to use for making a package. E.g., ZIP, TGZ, DragNDrop (OSX only). Leave blank for default generator.")
if("${SRB2_CPACK_GENERATOR}"STREQUAL"")
if("${CMAKE_SYSTEM_NAME}"MATCHES"Windows")
set(SRB2_CPACK_GENERATOR"ZIP")
elseif("${CMAKE_SYSTEM_NAME}"MATCHES"Linux")
set(SRB2_CPACK_GENERATOR"TGZ")
elseif("${CMAKE_SYSTEM_NAME}"MATCHES"Darwin")
set(SRB2_CPACK_GENERATOR"TGZ")
endif()
endif()
set(CPACK_GENERATOR${SRB2_CPACK_GENERATOR})
set(CPACK_PACKAGE_DESCRIPTION_SUMMARY"Sonic Robo Blast 2 Kart Saturn"CACHESTRING"Program name for display purposes")
set(CPACK_PACKAGE_VENDOR"Kart Krew"CACHESTRING"Vendor name for display purposes")
@@ -22,13 +22,14 @@ We also have a Flatpak available on [Flathub](https://flathub.org/en/apps/org.sr
See [SRB2 Wiki/Source code compiling](http://wiki.srb2.org/wiki/Source_code_compiling). The compiling process for SRB2Kart is largely identical to SRB2.
Currently only the Makefile build system with GCC or Clang compiler is supported.
For Windows only the Makefile build system with GCC or Clang compiler is supported.
Visual C++ (MSVC) is unsupported.
64-bit Windows builds are supported and recommended over 32-bit builds, they will build automatically when using the MSYS2 MinGW 64-bit executable (requires the mingw-w64-x86_64-gcc package), no extra flags needed!
The specific DLLs can be found under /libs/dll-binaries/(i686 for 32bit/x86_64 for 64-bit)/ or in the Release packages
MacOS and Linux can be build with CMake, but it is recommended to use the Makefile for Linux over CMake
Alternatively the [srb2bld](https://github.com/Bijman/srb2bld) script by Bijman can be used aswell!
The latest version of this library is available from GitHub:
https://github.com/libsdl-org/SDL_mixer/releases
The latest version of this library is available from:
http://www.libsdl.org/projects/SDL_mixer/
Due to popular demand, here is a simple multi-channel audio mixer.
It supports 8 channels of 16 bit stereo audio, plus a single channel of music. It can load FLAC, MP3, Ogg, VOC, and WAV format audio. It can also load MIDI, MOD, and Opus audio, depending on build options (see the note below for details.)
It supports 8 channels of 16 bit stereo audio, plus a single channel
of music.
See the header file SDL_mixer.h and the examples playwave.c and playmus.c for documentation on this mixer library. This documentation is also available online at https://wiki.libsdl.org/SDL2_mixer
See the header file SDL_mixer.h and the examples playwave.c and playmus.c
for documentation on this mixer library.
The process of mixing MIDI files to wave output is very CPU intensive, so if playing regular WAVE files sound great, but playing MIDI files sound choppy, try using 8-bit audio, mono audio, or lower frequencies.
The mixer can currently load Microsoft WAVE files and Creative Labs VOC
files as audio samples, it can load FLAC files with libFLAC, it can load
Ogg Vorbis files with Ogg Vorbis or Tremor libraries, it can load MP3 files
using mpg123 or libmad, and it can load MIDI files with Timidity,
FluidSynth, and natively on Windows, Mac OSX, and Linux, and finally it can
load the following file formats via ModPlug or MikMod: .MOD .S3M .IT .XM.
If you have built with FluidSynth support, you'll need to set the SDL_SOUNDFONTS environment variable to a Sound Font 2 (.sf2) file containing the musical instruments you want to use for MIDI playback.
Tremor decoding is disabled by default; you can enable it by passing
--enable-music-ogg-tremor
to configure, or by defining MUSIC_OGG and OGG_USE_TREMOR.
libmad decoding is disabled by default; you can enable it by passing
and unpack them in /usr/local/lib under UNIX, and C:\ under Win32.
This library is under the zlib license, see the file "LICENSE.txt" for details.
iOS:
In order to use this library on iOS, you should include the SDL.xcodeproj
and Xcode-iOS/SDL_mixer.xcodeproj in your application, add the SDL/include
and SDL_mixer directories to your "Header Search Paths" setting, then add the
libSDL2.a and libSDL2_mixer.a to your "Link Binary with Libraries" setting.
Note:
Support for software MIDI, MOD, and Opus are not included by default because of the size of the decode libraries, but you can get them by running external/download.sh
- When building with CMake, you can enable the appropriate SDL2MIXER_* options defined in CMakeLists.txt. SDL2MIXER_VENDORED allows switching between system and vendored libraries.
- When building with configure/make, you can build and install them normally and the configure script will detect and use them.
- When building with Visual Studio, you will need to build the libraries and then add the appropriate LOAD_* preprocessor define to the Visual Studio project.
- When building with Xcode, you can edit the config at the top of the project to enable them, and you will need to include the appropriate framework in your application.
- For Android, you can edit the config at the top of Android.mk to enable them.
This library is under the zlib license, see the file "COPYING.txt" for details.
The default MP3 support is provided using minimp3. SDL_mixer also supports using libmpg123: you can enable it by passing --enable-music-mp3-mpg123 to configure.
Loaded 100 of 324 files, more files were not shown because too many files have changed in this diff.
Show more
Reference in new issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
alufolie91
indev
SteelT
Alug
orbitalviolet
Roberto Ierusalimschy
NepDisk
GenericHeroGuy
Gustaf Alhäll
James R
Monster Iestyn
MIDIMan
romoney5
Eidolon
GLide KS
mazmazz
entryway
AJ Martinez
Zwip-Zwap Zapony
toaster
minenice55
X.organic
Alug
PAS
cobalt2727
LJ Sonic
Blue Staggo