01:14 < PerfeCtion_> Hey 01:14 <+PerfeCtion_> Can someone help me changing small things in the teeworlds source? I already compiled it I just need some help please 01:17 <@heinrich5991> hello also from over here ;) ask your question and you might get an answer within a day or so ;) 01:23 <@heinrich5991> good night 03:03 <+PerfeCtion_> Okayy so in which cpp file can I do unlimited ammo? 03:14 <+PerfeCtion_> i found it out myself 03:14 <+PerfeCtion_> but how can i make the gun autoshoot? 03:28 <+PerfeCtion_> Heheee i got ittttt 03:38 <+PerfeCtion_> Oh mann Wie kann ich die shotgun mehr spread machen? 03:46 <+PerfeCtion_> i need the shotgun to spread more but it is acting weird i dont get it :/ 05:43 <+bridge> [ddnet] i guess you didn't add new direction for your bullets 05:43 <+bridge> [ddnet] i guess you didn't add new directions for your bullets 06:00 <+bridge> [ddnet] ChillerDragon: can we also get a script to remove paths automatically xd 06:39 <+bridge> [ddnet] bruh 06:39 <+bridge> [ddnet] https://cdn.discordapp.com/attachments/293493549758939136/810747146755899432/unknown.png 06:40 <+bridge> [ddnet] intel driver 06:42 <+bridge> [ddnet] using nvidia tho lol 06:42 <+bridge> [ddnet] it is probably the site that is broken 08:49 <+bridge> [ddnet] man these botters that ask for source code help are increasing exponentially lately 09:48 <+fokkonaut> Hey fstd 10:03 <+fstd> hi fokkonaut 10:09 <+fokkonaut> heinrich told me you have a 512 player server? Could you show me how you did the collision between tees? My server sometimes need up to 100% of the CPU, I did some local tests and think it is related to the collision code, e.g. CCharacterCore::Move() 10:19 <+fstd> oh well he's probably referring to http://penenen.de/twparticle.png 10:20 <+fstd> in no way shape or form i run such a server, that screenshot is at least 6-8 years old 10:20 <+bridge> [ddnet] oh true, i remember this 10:20 <+bridge> [ddnet] ohhh 10:20 <+fstd> and i didn't do anything special collision wise, it performed horribly at 512 10:20 <+fstd> in fact the lesson learned was 10:20 <+bridge> [ddnet] exactly 10:20 <+fstd> many tees needs ac ollision overhaul 10:21 <+fstd> quadtree maybe 10:21 <+bridge> [ddnet] well, i probably dont know how :D 10:22 <+fstd> the general idea is to group tees together that are physically close 10:23 <+fstd> that way you only have to check collisions against your group rather than everything 10:23 <+fstd> of course maintaining the groups has some overhead on its own 10:24 <+bridge> [ddnet] uhm.. 10:24 <+bridge> [ddnet] why is it even that expensive? 10:26 <+fstd> what collisions? 10:26 <+fokkonaut> yea 10:27 <+fstd> well because the vanilla approach checks every tee vs every other tee 10:27 <+fstd> so N tees -> N^2 checks 10:27 <+fstd> that gets costly as N grows 10:28 <+bridge> [ddnet] hm 10:30 <+fstd> (technically it only takes half the amount of checks as after checking X vs Y you don't have to check Y vs X 10:30 <+fstd> so there's a /2 in there somewhere 10:30 <+fstd> but that becomes insignificant as N grows as well 10:30 <+bridge> [ddnet] but for 128 players it should be possible to improve it, right? 10:31 <+fstd> IIRC 128 was right at the edge of becoming noticably nasty but still kinda playable (i only tested with dummies though, not with 128 real clients) 10:32 <+fstd> but yes for 128 players the naive approach really gets to its limits 10:33 <+bridge> [ddnet] well for my server i can say it usually works, i even had 128 clients connected 2-3 times, but in the past few days im facing problems more and more, where CPU usage goes up to 100% 10:33 <+bridge> [ddnet] i dont really know why it suddenly started to do that so often 10:33 <+bridge> [ddnet] the easiest you could do without changing source too much is replacing distance checks with squared distance 10:33 <+bridge> [ddnet] isnt distance already checked by squared distance? 10:34 <+bridge> [ddnet] oh in 0.7? i dunno ^^ 10:34 <+bridge> [ddnet] let me check it 10:34 <+fstd> fokkonaut: it's a little involved but a profiler would tell you what part is eating the CPU 10:35 <+bridge> [ddnet] @Jupstar ✪ sqrtf(a.x*a.x + a.y*a.y) 10:35 <+bridge> [ddnet] thats not squared 10:35 <+bridge> [ddnet] @Jupstar ✪ `sqrtf(a.x*a.x + a.y*a.y)` 10:35 <+bridge> [ddnet] its rooted 10:35 <+bridge> [ddnet] root function is expensive 10:35 <+bridge> [ddnet] ah lol 10:36 <+fstd> see this is the kind of thing you'd probably want to profile before changing 10:36 <+bridge> [ddnet] ChillerDragon tried to do that, but couldnt get it to work 10:36 <+bridge> [ddnet] i did some local tests with visual studio, and it looked like the distance function in CCharacterCore::Move (inside of the player <-> player loop) is very expensive 10:36 <+bridge> [ddnet] you could also compile with Ofast, then it optimizes floating calculations 10:37 <+fstd> okay 10:38 <+bridge> [ddnet] fstd: Jupstar: do you think you can help? 10:39 <+bridge> [ddnet] sqrt(9) == 3 10:39 <+bridge> [ddnet] becomes 10:39 <+bridge> [ddnet] 9 = 3*3 10:39 <+bridge> [ddnet] 10:39 <+bridge> [ddnet] so basically remove sqrt and instead increase the checks by ^ 2 10:40 <+bridge> [ddnet] but this is really only about optmizing the math a bit, fstd's approach of grouped physics will still be faster 10:44 <+bridge> [ddnet] last time i checked with Ofast vs O2 was about 10% performance increase, tho i didnt dig much deeper into what exactly is better 10:44 <+bridge> [ddnet] probs the static sized loops like MAX_CLIENTS 10:44 <+bridge> [ddnet] and maybe the floating point arithmetics 10:44 <+bridge> [ddnet] :D 10:44 <+bridge> [ddnet] creating a mod gets more complicated every time 11:20 <+bridge> [ddnet] I would really appreciate any help 11:20 <+bridge> [ddnet] with this specific lag problem 11:22 <+bridge> [ddnet] @Jupstar ✪ i will try the squared way< 11:33 <+bridge> [ddnet] @Jupstar ✪ `sqrtf(a.x*a.x + a.y*a.y)` 11:33 <+bridge> [ddnet] What do you I have to replace this with? 11:33 <+bridge> [ddnet] What do I have to replace this with? 11:57 <+bridge> [ddnet] We can all suggest optimizations but doubt anyone will know what to do without a profile 11:58 <+bridge> [ddnet] Locally with some server dummies I can see that its the distance function 11:58 <+bridge> [ddnet] What jupstar meant was to compare distances without sqrt which would work because sqrt is monotone increasing 11:58 <+bridge> [ddnet] I understood what he means, but not how to do it 11:58 <+bridge> [ddnet] (As long as we don't have negative coordinates, that will crash and burn) 11:59 <+bridge> [ddnet] oh wait, thats very important 11:59 <+bridge> [ddnet] I want to try it anyways 12:00 <+bridge> [ddnet] I don't think it would help significantly, check perf see how much of the time is spent on the sqrt 12:02 <+bridge> [ddnet] fstd was considering a quadtree which would help with much larger N but it's not trivial to implement. 12:05 <+bridge> [ddnet] I considered a precalculated BSP or spatial hashing (chunking) both are much easier to implement 12:08 <+bridge> [ddnet] I guess most engines now use a Bounding Volume Hierarchy though so that's also an option you can look into 12:09 <+bridge> [ddnet] Actually this will work just fine, sorry, just woke up 12:11 <+bridge> [ddnet] :D 12:11 <+bridge> [ddnet] Could you help me implement any of those ideas 12:13 <+bridge> [ddnet] If I had time to code anything I'd finish my own PRs 12:24 <+bridge> [ddnet] @fokkonaut You can try the trivial optimization of checking squared distance but I have a feeling gcc/clang already figures this out on it's own 12:26 <+bridge> [ddnet] hmm 12:27 <+bridge> [ddnet] I doubt that will do much though. I'd go straight to spatial hashing. You basically split the map into chunks and a hash function tells you which chunk you are in. When checking for collision you only check against other tees in your chunk 12:29 <+bridge> [ddnet] makes sense, but i doubt i can implement such thing 12:29 <+bridge> [ddnet] make a 2nd function called distance_squared 12:29 <+bridge> [ddnet] and change all affected calculations to the squared versions 12:30 <+bridge> [ddnet] i still dont know what i should calculate in there 12:31 <+bridge> [ddnet] e.g. 12:31 <+bridge> [ddnet] 12:31 <+bridge> [ddnet] ``` 12:31 <+bridge> [ddnet] float D = distance(Pos, pCharCore->m_Pos); 12:31 <+bridge> [ddnet] if(D < 28.0f && D >= 0.0f) 12:31 <+bridge> [ddnet] { 12:31 <+bridge> [ddnet] if(a > 0.0f) 12:31 <+bridge> [ddnet] m_Pos = LastPos; 12:31 <+bridge> [ddnet] else if(distance(NewPos, pCharCore->m_Pos) > D) 12:31 <+bridge> [ddnet] m_Pos = NewPos; 12:31 <+bridge> [ddnet] return; 12:31 <+bridge> [ddnet] } 12:31 <+bridge> [ddnet] ``` 12:31 <+bridge> [ddnet] gets 12:31 <+bridge> [ddnet] 12:31 <+bridge> [ddnet] ``` 12:31 <+bridge> [ddnet] float D = distance_squared(Pos, pCharCore->m_Pos); 12:31 <+bridge> [ddnet] if(D < 28.0f * 28.0f && D >= 0.0f) 12:31 <+bridge> [ddnet] { 12:31 <+bridge> [ddnet] if(a > 0.0f) 12:31 <+bridge> [ddnet] m_Pos = LastPos; 12:31 <+bridge> [ddnet] else if(distance_squared(NewPos, pCharCore->m_Pos) > D) 12:31 <+bridge> [ddnet] m_Pos = NewPos; 12:31 <+bridge> [ddnet] return; 12:31 <+bridge> [ddnet] } 12:31 <+bridge> [ddnet] ``` 12:31 <+bridge> [ddnet] i mean inside of that function... 12:31 <+bridge> [ddnet] remove sqrt 12:31 <+bridge> [ddnet] Well doesn't the name tell you what it is? 😄 12:31 <+bridge> [ddnet] but better create a 2nd function 12:32 <+bridge> [ddnet] Definitely create a 2nd function, changing it everywhere at once won't be fun 12:32 <+bridge> [ddnet] just retuirn `a.x*a.x + a.y*a.y` 12:32 <+bridge> [ddnet] lol 12:32 <+bridge> [ddnet] and that works without problems? 12:32 <+bridge> [ddnet] up to a certain point 12:32 <+bridge> [ddnet] Shouldn't be too hard btw, in CCharacterCore::Move instead of going thru all the tees, you only go thru tees in your and neighbouring chunks 12:35 <+bridge> [ddnet] how would i define where a chunk is and ends? 12:36 <+bridge> [ddnet] what are the boundaries? 12:36 <+bridge> [ddnet] When the squares get way too large 12:37 <+bridge> [ddnet] At some point floats will become too inaccurate 12:37 <+bridge> [ddnet] u shouldnt care too much i think, in theory u could also use doubles, bcs x87 uses 80bit precision internally 12:38 <+bridge> [ddnet] Also I think there is a conversion to an integer there so floor(D) + 1 can't be larger than INT_MAX 12:38 <+bridge> [ddnet] So that's another limit 12:38 <+bridge> [ddnet] Ooooh now that I think about it you shouldn't replace that first distance with a distance squared e.g. 12:39 <+bridge> [ddnet] It will square the amount of iterations you do along the line 😄 12:39 <+bridge> [ddnet] yea 12:39 <+bridge> [ddnet] yeah only inside the MAX_CLIENTS loop 12:40 <+bridge> [ddnet] I think you can skip thru tees aswell 12:40 <+bridge> [ddnet] didn't know this 12:44 <+bridge> [ddnet] it performs better i think, but not yet perfect of course 12:44 <+bridge> [ddnet] is ur server on linux? 12:46 <+bridge> [ddnet] for the release build u could deffs try out Ofast and lto 12:46 <+bridge> [ddnet] the good thing here is, if they dont work as expected, you can just remove them again without changing any code 12:52 <+bridge> [ddnet] ChillerDragon: 12:52 <+bridge> [ddnet] you here for this? :D 12:52 <+bridge> [ddnet] @Jupstar ✪ https://github.com/fokkonaut/F-DDrace/commit/c2017bcd2e4016c4171bbd2deb1a2016d9753b6e 12:54 <+bridge> [ddnet] i think the reference looks wrong 12:54 <+bridge> [ddnet] why? 12:54 <+bridge> [ddnet] &v => v 12:54 <+bridge> [ddnet] bcs u take a reference of a tmp object 12:54 <+bridge> [ddnet] just as the length function would do 12:54 <+bridge> [ddnet] no, it accepts a const reference 12:55 <+bridge> [ddnet] i never understoof this xD 12:55 <+bridge> [ddnet] just remove the reference 12:55 <+bridge> [ddnet] the compiler is smart enough to not produce slower code 12:55 <+bridge> [ddnet] i dont even know what that is 12:55 <+bridge> [ddnet] its like a pointer just that it cant be null xd 12:55 <+bridge> [ddnet] what is it used for 12:55 <+bridge> [ddnet] like 12:55 <+bridge> [ddnet] idk 12:56 <+bridge> [ddnet] e.g. if you have an object of class CPlayer 12:56 <+bridge> [ddnet] and u take it by reference, you modify the object itself 12:56 <+bridge> [ddnet] not a copy of it 12:57 <+bridge> [ddnet] but isnt that the same with a pointer? 12:57 <+bridge> [ddnet] https://github.com/fokkonaut/F-DDrace/commit/1ac880a085ecd7376d3570fd60da52c1444ed72f 12:57 <+bridge> [ddnet] yes 12:57 <+bridge> [ddnet] references are basically pointers 12:58 <+bridge> [ddnet] its not directly true, but they are implemented like that 12:58 <+bridge> [ddnet] the rest looks fine to me 12:58 <+bridge> [ddnet] okay, thanks 12:58 <+bridge> [ddnet] if its broken you'd notice anyway, bcs the collision wont work 😄 12:58 <+bridge> [ddnet] it all works fine, as it seems 12:59 <+bridge> [ddnet] I wonder if my mod has more of these "bugs" 13:00 <+bridge> [ddnet] performance issues 13:00 <+bridge> [ddnet] I think I should replace the distance by distance_squared in gameworld too, right? 13:00 <+bridge> [ddnet] all these FindEntities etc 13:00 <+bridge> [ddnet] ClosestCharacter 13:04 <+bridge> [ddnet] u can replace it basically everywhere, where the distance is not used directly, but only compared against others 13:04 <+bridge> [ddnet] but it wont improve it everywhere as much 13:04 <+bridge> [ddnet] makes sense 13:05 <+bridge> [ddnet] why? 13:05 <+bridge> [ddnet] i guess a single call is not that bad? :d 13:05 <+bridge> [ddnet] :D* 13:06 <+bridge> [ddnet] well the FPU might not even profit from a signle core, if it waits on another call anyway 13:06 <+bridge> [ddnet] its really hard to predict what happens inside the CPU 13:06 <+bridge> [ddnet] ask an intel or amd dev xd 13:06 <+bridge> [ddnet] xd 13:06 <+bridge> [ddnet] so basically you lol 13:06 <+bridge> [ddnet] well the FPU might not even profit from a signle call, if it waits on another call anyway 13:06 <+bridge> [ddnet] what you've done to ddnet is immaculate xD 13:07 <+bridge> [ddnet] but i havent done anything FPU related xD 13:07 <+bridge> [ddnet] i also dont know how GPU drivers are implemented in detail, its more guessing 13:07 <+bridge> [ddnet] also dont know what that is tbh 13:07 <+bridge> [ddnet] floating point unit 13:07 <+bridge> [ddnet] xD thats why its experimental? 13:07 <+bridge> [ddnet] its like a CPU just for floating points 13:07 <+bridge> [ddnet] lol 13:07 <+bridge> [ddnet] ah no, opengl is well defined 13:08 <+bridge> [ddnet] its experimental, bcs we had quite a few bugs on Mac and older hardware on windows 13:08 <+bridge> [ddnet] we even have an unknown bug in 6+7th gen intel 13:08 <+bridge> [ddnet] tho it also affects older GL versions 13:09 <+bridge> [ddnet] but i'd say GL 3.3 is stable, i didnt really encounter an issue lately 13:13 <+bridge> [ddnet] anyway fokko, if u have the motivation to change every distance check do it. it wont hurt^^ 13:20 <+bridge> [ddnet] :D 13:26 <+bridge> [ddnet] something something premature optimization something something 13:27 <+bridge> [ddnet] Profile your code before making small insignificant optimizations a modern CPU has so many tricks up it's sleeve, small changes usually mean nothing 13:31 <+bridge> [ddnet] i ran my fng server with 256 slots, 255 dbg dummies which randomly moved and jumped etc. and it worked fine without code modications, but maybe i just used a stronger CPU 13:32 <+bridge> [ddnet] I had trouble with cpu usage so I'd guess that's what helps 13:33 <+bridge> [ddnet] Maybe just throw money at the problem @fokkonaut 13:33 <+bridge> [ddnet] ChillerDragon said that we have the best CPU already 13:33 <+bridge> [ddnet] i wondered too 13:33 <+bridge> [ddnet] he said anything better costs > 100€ 13:34 <+bridge> [ddnet] p/m 13:41 <+bridge> [ddnet] https://github.com/fokkonaut/F-DDrace/runs/1903237972 13:41 <+bridge> [ddnet] doesnt work 13:43 <+bridge> [ddnet] ah 13:43 <+bridge> [ddnet] i think i know how this works 13:43 <+bridge> [ddnet] yea i fucked it up lol 13:48 <+bridge> [ddnet] still doesnt work 14:00 <+bridge> [ddnet] https://cdn.discordapp.com/attachments/293493549758939136/810858123291131944/unknown.png 14:02 <+bridge> [ddnet] cant u use smth like push "-Wno-..." 14:03 <+bridge> [ddnet] ah nvm that looks right indeed 14:03 <+bridge> [ddnet] maybe outdated compiler? 14:04 <+bridge> [ddnet] not sure, its github actions test 14:04 <+bridge> [ddnet] the first error just occured out of nowhere 14:05 <+bridge> [ddnet] and the fix doesnt work 14:06 <+bridge> [ddnet] its also missing a pop isnt it? 14:06 <+bridge> [ddnet] but probs not causing this issue 14:07 <+bridge> [ddnet] but ubuntu latest doest fail? 14:07 <+bridge> [ddnet] then its probs the compiler version, i'd say 14:07 <+bridge> [ddnet] gcc 5.5 is pretty old xd 14:08 <+bridge> [ddnet] added the pop later on 14:08 <+bridge> [ddnet] how do i update it? 14:08 <+bridge> [ddnet] i think you can just check the gcc version somehow 14:09 <+bridge> [ddnet] i am not familiar with all of this xD 14:09 <+bridge> [ddnet] at all 14:09 <+bridge> [ddnet] #if defined(__GNUC__) && .... 14:09 <+bridge> [ddnet] `#if defined(__GNUC__) && ....` 14:10 <+bridge> [ddnet] ``` 14:10 <+bridge> [ddnet] /* Test for GCC > 3.2.0 */ 14:10 <+bridge> [ddnet] #if __GNUC__ > 3 || \ 14:10 <+bridge> [ddnet] (__GNUC__ == 3 && (__GNUC_MINOR__ > 2 || \ 14:10 <+bridge> [ddnet] (__GNUC_MINOR__ == 2 && \ 14:10 <+bridge> [ddnet] __GNUC_PATCHLEVEL__ > 0)) 14:10 <+bridge> [ddnet] ``` 14:10 <+bridge> [ddnet] https://gcc.gnu.org/onlinedocs/cpp/Common-Predefined-Macros.html 14:10 <+bridge> [ddnet] so a `&& __GNUC__ > 7` might do the job 14:10 <+bridge> [ddnet] i dunno when they added that pragma support 14:11 <+bridge> [ddnet] well, i do have a fix already in f-ddrace 14:11 <+bridge> [ddnet] for other false positives 14:11 <+bridge> [ddnet] maybe it just fails bcs the actual warning is unsupported 14:11 <+bridge> [ddnet] no idea, but how can it suddenly pop up? 14:11 <+bridge> [ddnet] like the whole warning why i want to add this ignore thing 14:12 <+bridge> [ddnet] yes, but like this: 14:12 <+bridge> [ddnet] warning was added in gcc 8 14:12 <+bridge> [ddnet] warning causes werror in gcc 8 14:12 <+bridge> [ddnet] 14:12 <+bridge> [ddnet] so u only want to remove it for gcc 8 14:12 <+bridge> [ddnet] i mean 14:13 <+bridge> [ddnet] https://cdn.discordapp.com/attachments/293493549758939136/810861277910925312/unknown.png 14:13 <+bridge> [ddnet] from one day to another 14:13 <+bridge> [ddnet] this error appeared 14:13 <+bridge> [ddnet] https://cdn.discordapp.com/attachments/293493549758939136/810861345544994837/unknown.png 14:13 <+bridge> [ddnet] yes, maybe ubuntu latest was updated 14:13 <+bridge> [ddnet] oh 14:13 <+bridge> [ddnet] true 14:15 <+bridge> [ddnet] that warning is considered not very useful anyway, i already read it in the gcc mailing lists, and they were devided, if its useful or not, in the end, not warning might cause more issues 😄 14:15 <+ChillerDragon> @fokkonaut haha we do not have the best single core cpu on the market but i think the best we can afford at the current hoster 14:15 <+bridge> [ddnet] meh 14:15 <+ChillerDragon> anything i missed i did not read all? 14:15 <+bridge> [ddnet] yes 14:16 <+bridge> [ddnet] Can you add a script to remove the paths for svgs? XD 14:16 <+ChillerDragon> sure 14:16 <+bridge> [ddnet] nice thanks 14:16 <+ChillerDragon> make an issue please 14:16 <+ChillerDragon> wait i can make one :D 14:16 <+bridge> [ddnet] ok 14:16 <+ChillerDragon> oh no then i do not get a mail 14:16 <+ChillerDragon> can you do it ? :D 14:16 <+bridge> [ddnet] ok xD 14:16 <+ChillerDragon> nice thanks 14:20 <+ChillerDragon> ofast and lto is just to boost performance right? not to solve the actual issue but more compating the sympthoms? 14:22 <+bridge> [ddnet] yes, but it heavily relies on defined behavior 14:22 <+bridge> [ddnet] so better write good code with it xD, there are edge cases you never heard of, before you googled them xD 14:39 <+ChillerDragon> i wrote "compating" sometimes i wonder how people understand me 14:40 <+bridge> [ddnet] i just ignore words i dont understand xd 14:49 <+bridge> [ddnet] https://9to5linux.com/linux-kernel-5-11-officially-released-this-is-whats-new 14:49 <+bridge> [ddnet] :poggers: 15:39 <+bridge> [ddnet] @Learath2 splitting the map into different chunks doesnt make mush sense if all players are in the same chunk, or in chunks close to it 15:39 <+bridge> [ddnet] You make the chunks rather small 15:40 <+bridge> [ddnet] There is a version of this scheme that makes the density of the grid variable at a given point (quadtrees), quite a bit harder to implement though 16:29 <+bridge> [ddnet] https://www.drdobbs.com/jvm/an-algorithm-for-compressing-space-and-t/184406478 16:29 <+bridge> [ddnet] 👀 16:34 <+bridge> [ddnet] https://cdn.discordapp.com/attachments/293493549758939136/810896764474621972/unknown.png 16:34 <+bridge> [ddnet] u have to scroll all the way down 16:34 <+bridge> [ddnet] to reject all their shit 16:34 <+bridge> [ddnet] but allow all is up there 16:34 <+bridge> [ddnet] https://cdn.discordapp.com/attachments/293493549758939136/810896845773209600/unknown.png 16:35 <+bridge> [ddnet] i wonder if u can law suit them 16:35 <+bridge> [ddnet] doesnt it have to be clear? 16:36 <+bridge> [ddnet] gdrp should make this a opt in feature, where it is disabled by default, and the user has to explicitly go and activate it by his own will 18:34 <+bridge> [ddnet] @Jupstar ✪ @fstd: it probably is not even related to that part of the code, i did some tests with real clients, and it seems the DoSnapshot function is just taking so much 18:35 <+bridge> [ddnet] and all the entities in general 18:36 <+bridge> [ddnet] i did some intense debugging with chillerdragon but idk, i cant really find anything, but i also never had anything like this before, sucks really hard 18:38 <+bridge> [ddnet] I dont know why it ran so good for months now, I dont want to remove 128p support again 18:52 <+bridge> [ddnet] Can i for example thread some specific stuff? 18:52 <+bridge> [ddnet] for example the snapshotting which will go up to 60% iirc, or some other stuff? 19:04 <+bridge> [ddnet] Eh the gun is broken xD 19:05 <+bridge> [ddnet] You can see if you shoot with grenade or smth 19:05 <+bridge> [ddnet] You cant see if you shoot with grenade or smth 19:07 <+bridge> [ddnet] @Marius you made a bad update. I think the version you got is already removed 19:07 <+bridge> [ddnet] Check out if you can download the latest version 19:09 <+bridge> [ddnet] So its not the problem from the srv? 19:10 <+bridge> [ddnet] Ah I see 19:10 <+bridge> [ddnet] @fokkonaut would it kill you to learn how to profile the code? Speculating about where the issue might be and randomly moving things to threads won’t help 19:11 <+bridge> [ddnet] I profiled it 19:11 <+bridge> [ddnet] DoSnapshot is using 50-60% of the CPU on my local server 19:12 <+bridge> [ddnet] CGameWorld::Tick and Snap takes a lot too 19:12 <+bridge> [ddnet] iterating over all entities 19:12 <+bridge> [ddnet] CSnapshotBuilder::Finish is also very expensive due to the tl_swap 19:12 <+bridge> [ddnet] and all the loops in CCharacterCore::Move and ::Tick 19:13 <+bridge> [ddnet] Snap on what? 19:14 <+bridge> [ddnet] CDragger takes 10%, character just about 2-5 19:14 <+bridge> [ddnet] but creating the snapshots takes a lot 19:16 <+bridge> [ddnet] There is a tl_swap in ::Finish? 19:16 <+bridge> [ddnet] 3 19:16 <+bridge> [ddnet] well not in finish i think 19:16 <+bridge> [ddnet] let me see 19:17 <+bridge> [ddnet] yes 19:17 <+bridge> [ddnet] 3 times per index 19:17 <+bridge> [ddnet] in the loop 19:17 <+bridge> [ddnet] Also is this profile from a live server? People moving around and chatting might change the profile drastically due to the way collision is handled 19:18 <+bridge> [ddnet] i tested it with clients locally because chillerdragon said its not possible to get a profiler on the live server 19:18 <+bridge> [ddnet] There is no loop in CSnapshotBuilder::Finish? 19:18 <+bridge> [ddnet] i am on 0.7 19:18 <+bridge> [ddnet] my mod 19:18 <+bridge> [ddnet] ChillerDragon learn to use perf nobo 19:19 <+bridge> [ddnet] his argument was that the one he used didnt give any information 19:19 <+bridge> [ddnet] Oh 0.7 now sorts the snapshots? 19:19 <+bridge> [ddnet] and valgrind or some others are way to expensive 19:19 <+bridge> [ddnet] i have no idea 19:19 <+bridge> [ddnet] maybe? it seems like 19:19 <+bridge> [ddnet] why tho? 19:20 <+bridge> [ddnet] Well bubble sort has awful performance characteristics as n grows. Maybe try replacing that with std::sort? 19:20 <+bridge> [ddnet] I see, I should have based on 0.6 xD 19:20 <+bridge> [ddnet] how would that work? 19:21 <+bridge> [ddnet] `std::sort(pSnap->SortedKeys()[0], pSnap->SortedKeys()[NumItems]);`? 19:22 <+bridge> [ddnet] ah no 19:22 <+bridge> [ddnet] You need a custom swap function to swap the offsets too 19:22 <+bridge> [ddnet] why is that even needed 19:24 <+bridge> [ddnet] Because of the way split 1 item into 3, (data, size, offset) all go into separate arrays 19:25 <+bridge> [ddnet] So when sorting the data by key, you also need to swap around the size and the offset so they stay correct 19:25 <+bridge> [ddnet] Oh, yea of course! i mean why does 0.7 sort them? 19:25 <+bridge> [ddnet] Oh, so they can do binary search on the client 19:26 <+bridge> [ddnet] Avoids the costly linear search 0.6 does 19:26 <+bridge> [ddnet] ohh 19:27 <+bridge> [ddnet] But I don’t like the idea honestly. The client sorting it would distribute the cost. The server pays a huge cost per client all on it’s own like this 19:27 <+bridge> [ddnet] And why dont you need to do that for 0.6? 19:28 <+bridge> [ddnet] i mean for the bridge servers 19:28 <+bridge> [ddnet] I think I did, didn’t I? 19:28 <+bridge> [ddnet] not sure 19:28 <+bridge> [ddnet] Hm I actually don’t remember 19:28 <+bridge> [ddnet] you sounded surprised 19:29 <+bridge> [ddnet] I did but then I do remember discussing this before 19:29 <+bridge> [ddnet] Well let me check 19:29 <+bridge> [ddnet] i dont think you sort them, tbh 19:30 <+bridge> [ddnet] Hm, okay than idk what the bubble sort on the server is for. Maybe to reduce the cost for the client? 19:30 <+fstd> fokkonaut: if nothing about your build changed then i guess something changed on the host (duh) 19:30 <+fstd> this would be a good time to have been running munin for a few months :) 19:30 <+bridge> [ddnet] Well I push changes every day almost 19:30 <+bridge> [ddnet] and Mr. Anderson says nothing has changed on his side 19:31 <+fstd> so you can revert to a commit that has okay performance? 19:31 <+fstd> in that case use git bisect 19:31 <+bridge> [ddnet] https://github.com/teeworlds/teeworlds/commit/0c4b3ab62bb3ef798c77f8d591f921bffe6b30b0 19:31 <+bridge> [ddnet] thats the point, i dont know when it started 19:31 <+bridge> [ddnet] This is indeed why they bubble sort, I wonder how our bridge even works at all now 😄 19:32 <+fstd> finding out at what point it started is the whole point of using git bisect 19:32 <+bridge> [ddnet] x 19:32 <+bridge> [ddnet] D 19:32 <+fstd> ? 19:32 <+bridge> [ddnet] how does it work? 19:32 <+fstd> you tell it a good and a bad commit 19:32 <+bridge> [ddnet] i reacted to learaths message fstd 19:32 <+fstd> and it does a binary search 19:33 <+bridge> [ddnet] no idea what any of those means 19:33 <+bridge> [ddnet] binary search is rly simple 19:33 <+bridge> [ddnet] just look it up 19:33 <+fstd> it makes find teh bug fast 19:33 <+fstd> is what it means 19:33 <+bridge> [ddnet] But I would like to understand xD 19:33 <+fstd> or let's say the commit that introduced the bug 19:34 <+fstd> well 19:34 <+bridge> [ddnet] binary search discards half 19:34 <+bridge> [ddnet] every iteration 19:34 <+bridge> [ddnet] https://blog.penjee.com/wp-content/uploads/2015/04/binary-and-linear-search-animations.gif 19:34 <+fstd> there's this magical network of information 19:34 <+fstd> i'd explain it to you but only if you genuinely can't figure out out by yourself using google or wikipedia etc 19:35 <+bridge> [ddnet] yeah 19:35 <+bridge> [ddnet] literally wikipedia can tell u all this 19:35 <+bridge> [ddnet] u just a lazy bones 19:36 <+bridge> [ddnet] so how do i use git bisect 19:36 <+bridge> [ddnet] google it 19:36 <+bridge> [ddnet] or run "git bisect" 19:37 <+bridge> [ddnet] git bisect --help 19:37 <+bridge> [ddnet] wtf 19:37 <+bridge> [ddnet] it runs git-bisect.hmtk 19:37 <+bridge> [ddnet] it runs git-bisect.hmtl 19:37 <+bridge> [ddnet] on linux it opens a man page 19:38 <+bridge> [ddnet] Ah the client when unpacking the delta also uses the snapshotbuilder so it sorts too 19:38 <+fstd> even the .html file will do. i think you can download .html viewers somewhere 19:39 <+bridge> [ddnet] okay, but git bisect wont help me much i think. since i need about 60-80 players for it to trigger 19:39 <+bridge> [ddnet] @fokkonaut you can probably abuse that to distribute the cost of sorting. Just send unsorted snaps 19:39 <+fstd> you obviously need some way to tell whether a given version is good or bad 19:39 <+bridge> [ddnet] i dont really have that 19:40 <+bridge> [ddnet] because it just starts to lag very hard, there is no value that pops over :D 19:40 <+bridge> [ddnet] so removing the bubble sort should work? 19:40 <+fstd> well but you notice whether it lags very hard or not...? 19:40 <+fstd> so you can tell when it's bad 19:41 <+fstd> now with the same test load if it doesn't lag very hard, it's good, no? 19:41 <+bridge> [ddnet] yea, true 19:41 <+bridge> [ddnet] but its not always happening 19:41 <+fstd> okay that sucks 19:41 <+bridge> [ddnet] its not just, oh okay 70 players i will start to lag now 19:41 <+bridge> [ddnet] yea 19:41 <+bridge> [ddnet] it sucks so bad 19:42 <+fstd> well bring 70 players then every time? 19:42 <+bridge> [ddnet] its at 72 right now and doesnt lag 19:42 <+fstd> sigh 19:42 <+bridge> [ddnet] There is actually a way to handle this. perf is perfectly capable of saving to a ring buffer and you can send a SIGUSR2 to it to get it to dump the buffer to a file 19:42 <+bridge> [ddnet] its pure randomness so the only thing i can do is improving everything 19:43 <+bridge> [ddnet] So to profile transient issues like this keep perf running and send it a SIGUSR2 when it gets laggy 19:43 <+fstd> Learath2: nice idea 19:43 <+bridge> [ddnet] ChillerDragon 19:43 <+bridge> [ddnet] u there :/ 19:43 <+fstd> Learath2: what's perf tho? not gprof? 19:44 <+bridge> [ddnet] @Learath2 ? 19:44 <+bridge> [ddnet] @fokkonaut I think so, try it it should break very obviously if my hypothesis is wrong 19:44 <+bridge> [ddnet] okay, thanks 19:44 <+bridge> [ddnet] fstd: perf is part of the kernel 19:45 <+bridge> [ddnet] it can do the kind of sampling profiling gprof does 19:46 <+bridge> [ddnet] works @Learath2 19:49 <+bridge> [ddnet] Again not sure how much it'll help but it's one less bubble sort per player per snap less 19:51 <+bridge> [ddnet] O(nk^2) where k is the number of snapitems which I guess is fair to assume is of the same order as n so an O(n^3) process gone 19:52 <+bridge> [ddnet] (atleast I think, I'm usually not very good with the theory part of CS) 19:57 <+bridge> [ddnet] well, just locally already it is a huge improvement 20:04 <+bridge> [ddnet] fstd: perf works without instrumenting the code so a little less information but it's still very useful imho, you should add it to your toolkit for profiling stuff 20:05 <+bridge> [ddnet] ulan ziyaaa 20:25 <+bridge> [ddnet] rider is better than vs but running trying to run 32 bits on it sucks 20:25 <+bridge> [ddnet] you gotta download alot of shit 20:25 <+bridge> [ddnet] sad 20:26 <+bridge> [ddnet] you gotta download some of shit 20:50 <+bridge> [ddnet] @Learath2 why is the CDragger::Snap so expensive? 20:50 <+bridge> [ddnet]