DDraceNetwork - developer

DDraceNetwork

Development / developer

Development discussion. Logged to https://ddnet.tw/irclogs/ Connected with DDNet's IRC channel, Matrix room and GitHub repositories — IRC: #ddnet on Quakenet | Matrix: #ddnet-developer:matrix.org GitHub: https://github.com/ddnet

Between 2020-07-13 00:00:00Z and 2020-07-14 00:00:00Z

@nexusrf can you try the build in #2502?

https://github.com/ddnet/ddnet/pull/2502

Improved dummy switching by Fireball-Teeworlds · Pull Request #2502...

Fix a use-after-free when there are no new snapshots for the cl_dummy tee after the switch: #2499 . When one of the tees is hooked, cycling through them will no longer show phantom hook for the oth...

@heinrich5991 can we add to the tail of netobjects without breakage or do we need to add a new object and backwards compatibility code?

@Learath2 i never used this channel, where is #2502

https://github.com/ddnet/ddnet/pull/2502

Improved dummy switching by Fireball-Teeworlds · Pull Request #2502...

Fix a use-after-free when there are no new snapshots for the cl_dummy tee after the switch: #2499 . When one of the tees is hooked, cycling through them will no longer show phantom hook for the oth...

The bot even links it for you

09:15

Go to the checks tab, there is artifacts towards top right, you can download from there

09:16

Image attachment

@Learath2 this is only for authorized users

sure? I thought they changed that

i dont see a difference either

oh

09:30

wel

neither one takes effect in my game

Image attachment

09:31

i did download the file, do i have to do some sort of preparation first? maybe download a further build or something

09:32

i mean an earlier build

Nope, if that artifact doesn't change your dummy gameplay that's fine

[ddnet:master] 2 new commits

85a2268 Show "Free View" for Sixup - Learath2 37a906d Merge #2501 - bors[bot]

09:58

[ddnet/ddnet] Pull request closed: #2501 Show "Free View" for Sixup

10:01

[ddnet/ddnet] Issue opened: #2507 does /save not keep hook...

As reported by Clavata on Discord. @Zwelf ?

neither one takes effect in my game

@nexusrf here's what I meant by phantom hook: https://youtu.be/mxVT4pdyGnU.

Phantom Hook Bug

xd, never did see it

@Learath2 not really, not without extra checks

10:32

the trouble is that the unpacking fails if the object is too small

@heinrich5991 so if I were to introduce a new field to player@netobj.ddnet.tw I'd need to instead add a new netobj called player2 AND keep the code for the old netobj too?

short of improving the general problem, yes

10:36

the only problematic combination is new client <-> old server

10:36

i.e. we can fix it

Uf, I hadn't noticed this rather problematic limitation of the protocol yet

10:40

@heinrich5991 could we even improve the general problem now?

yes

10:41

since the only problem is new client <-> old server: the new client expects the netobj to be longer

10:41

so we can make the new client add default values for short netobjs

something similar to your get or default thing on vanilla?

don't remember what you're referring to, possibly

You made it to fix the race record thing iirc

ah yea

10:43

something similar to that, just for netobjs

Problem is that netobjs aren't "unpacked"

hm yes

10:46

can add that ofc

Yeah, there is no good way to handle this problem

10:47

We only have a Size, no way to figure out how many fields are missing

10:47

honestly, just that cast even if it has the correct size is unsafe

yes @ unsafe

10:47

why not? we know the size, we should be good

Ah but everything is ints

10:48

and we already assume sizeof(int) == 4

10:49

maybe a "default" instance of each netobj that we can chop up and append to the end of the small one?

if you are talking about anything with 0.7: I will just break compability since there arent any ddnet7 servers anymore and i will just replace my player object with the current one

I'm more concerned about the new field we want to add causing some ugly compatibility code

you could make the snap code support not only invalidating objs but also replacing them

10:51

in a different buffer maybe

can you elaborate further?

in addition to the other functions like SnapInvalidateItem(int SnapID, int Index), the client interface gets a new function SnapReplaceItem(int SnapID, int Index, void *pItem, int ItemSize);. in the snap sanity check loop, this function is called at the appropriate places when a old netobj is detected

10:55

this requires support code in datasrc/{network,datatypes}.py

That'd imply we need to keep all the older versions in the code still, no?

we only need to add a default value to new fields in datasrc/network.py

I'm not sure ItemSize > ObjSize is a good enough way to detect an old message though

why not? we theoretically have unique IDs (edited)

I guess it is as long as we only append at the tail

10:58

This sounds like a good plan

11:00

@heinrich5991 or maybe keep all this transparent to the client?

cannot do that

11:01

the old server is the problem

I meant as in CClient, I was thinking of fixing it while validating, but noticed that'd be quite ugly now that I checked how we store the snaps

11:02

so yeah, SnapReplaceItem sounds about right

11:04

@heim

11:04

@heinrich5991 doesn't look trivial to replace it either actually

yep, needs a side buffer

I guess the cache is efficient enough at shifting data over unless the snap is larger than the cache line

ehm what?

11:05

I wouldn't try to modify a snap in place

Rebuild it?

no, keep a side buffer with all the replaced objects

And when the client tries to reference them, redirect it?

yes

11:07

if we had actual data structures, this wouldn't pose such a problem.. :/

which data structure would you use to solve this problem?

I would have the snap as a key-value mapping, probably a hashmap

11:08

they support modifying items

11:08

(in fact that's how I implemented it in libtw2)

Hashmaps have good theoretical performance, but I wonder how they'd compare to the array of ints approach right now?

11:09

Arrays have great cache behaviour and they play really well with prefetching

our array of ints approach has a linear search for ddnet and a binary search for teeworlds 0.7

Do we lookup enough for the O(1) lookup to be a good optimization?

dunno, typical snap size?

Don't we traverse snaps more often?

let's say 128 items

11:11

so finding the entry in the hash map must not be 64 times slower than the comparison we do for our linear search

11:12

in order for the hash map to be more effective

and we'd have to be doing the lookup often enough for it to be worth it, just traversing through an array has much better performance than traversing through a hashmap

11:17

Traversing a hashmap has the overhead of going through the empty buckets and worse cache performance

11:18

I think the unordered_map in c++ even uses linked lists, which are even worse

yes. rust hashmaps are fine

11:20

unordered_map has to have bad performance according to the standard

@heinrich5991 do they use open addressing?

it changed a couple of times

11:21

the algorithm used

Well even with open addressing, you have a layer of indirection and a prefetch unfriendly data structure

The hash table implementation is a Rust port of Google's SwissTable. The original C++ version of SwissTable can be found here, and this CppCon talk gives an overview of how the algorithm works.

11:23

https://abseil.io/blog/20180927-swisstables

abseil / Swiss Tables and absl::Hash

An open-source collection of core C++ library code

I've been watching a couple different talks on data structures and performance, I've been very surprised at some of the benchmarks I've seen

so no indirection if I see this blog post correctly

Well with a catch, if you don't have indirection, you can only store one type of thing (or a union) and the buckets then get rather large

11:26

We generally recommend that you use absl::flat_hash_map<K, std::unique_ptr<V>> instead of absl::node_hash_map<K, V>.

11:26

Ah, that's not saying what I thought it's saying

11:27

That's rather odd, I'd have thought the library would be smarter than the programmer at handling the indirection

we want fixed size buckets here, though

Do we? Items have varying size

pointer, len

11:27

(like we currently have with the array btw)

whats this hash map talk about?

11:28

related to ddnet

yes, kinda

Considering whether a hash map would be a good choice to store snaps as

I was lamenting the fact that we don't have data structures

instead of a large rather flat array of ints

do we iterate snaps a lot?

11:29

well i dont know much :)

Once every time we receive a snapshot

whats the advantage of the hash map then?

Then once more for the events

and in the game code as well

@Ryozuki we do linear lookups for specific times

11:30

s/times/types/

ah

11:31

maybe we can index by type or smth

11:31

if i got this right

We traverse it as many times as we do a linear lookup

11:33

Ah I see what you mean @heinrich5991 you want the snaps themselves to stay as is, but we reference into it using a hashmap?

I want the snap data to be stored in an array

11:33

but I want an index

11:34

so yes

That would probably have better performance

11:35

With a flat hashmap, the next item would probably fit into the cache rather nicely

by snap types u mean CNetObjs?

not very prefetch friendly but it happens

11:35

Well types of netobjs

anymore

11:39

they made discord not gamer

Image attachment

11:39

justatest

What does "invite-only place" mean? I always see invites on public websites, like on DDNet.tw

that you get to decide who joins I guess

There are also communities that don't create permanent invites posted publicly - instead they create expiring ones or even one-use-only

14:20

let's say community of specific paid course

pls fix this no sound hook

15:22

some players have it, i will add an example in a bit

15:25

sound is on, everyone's hook is normal except for his. his hook has no sound

was the double hook sound in 0.7 just fixed

and if he hits an unhookable tile you still hear the usual sound

@noby so you did make a 0.6 - 0.7 compatible server right? Can we make a window compile out of it?

i didnt make it, i applied the fixes timakro made to https://github.com/unique-clan/unique-race this server

but can i compile it to windows? since according to the instruction, not even git bash can recognize "cmake" or "make" commands (edited)

15:35

is it linux related?

that server should work on windows but i dont know / dont have windows to test it on

15:36

u probably have to install cmake

right

actually no

15:36

https://github.com/ddnet/ddnet

15:36

use the instructions here

the fuck?

Image attachment

16:10

im not sure if this is intended xd

16:14

maybe programming isn't for me lol

lol

im always using cmake gui before start compiling huh

cmake gui generation cannot detect visual studio on my computer

16:15

so i tried visual studio itself kek

kek

(edited)

kek

16:18

reinstall/check your VS Installer

@Learath2 sorry but i dont see any difference at all with that build and my current ddnet version

That's great

17:55

I was worried it'd break something like the dummy triple fly u do

u mean this thing?

17:56

i didnt try playing with it yet

17:56

i will do some 3flies with dummy and see

18:02

i dont see any major difference @Learath2 my triplefly has not been cucked

1

Perfect

Cucked

does hook have a hitbox or is it just a straight line

Is it possible I can add the AntiBot in my own mod @Learath2 ?

18:46

Its a library, right?

18:46

not open source but free to use?

It's not my library to share

18:47

You should ask @noby he wrote it

DDNet has it, is it possible to have it for local testing?

18:47

Or isnt it public?

It isnt public

Oh, ok

If it was really public people can reverse engineer it

True

18:48

thanks

Fireball-Teeworlds

[ddnet/ddnet] Issue opened: #2508 Hook sounds disappearing for players hooking other players

Example: https://youtu.be/ZKXxNgFj-HA.

Some players will not make hook sounds when hooking other players
Wall hooking sounds are fine
It seems that everyone will be missing hook sounds coming from that player (confirmed by asking another person who was there if they hear anything)
If you record a demo, hook sounds will be missing from that demo too
Rejoining the server doesn't fix the problem (if you're not the player that doesn't make sounds)
It's possible that changing t...

sound is on, everyone's hook is normal except for his. his hook has no sound

@nexusrf I've encountered this too! I've created an issue (see above)

does ddnet not use cl_isddnet anymore?

20:33

instead NETMSG_CLIENTVER now?

20:40

oh, it is sending both

yes

[ddnet/ddnet] Pull request opened: #2509 Keep hook after load

Provides a fix for #2507 on the server side. The client shows no hook, as it sends the server to release the hook (Input.m_Hook = 0). Made a demo how it currently looks: !client_demo Looks a bit ugly in the client :/ (sorry for the gif, experimented a bit with video recording and editing)

[ddnet/ddnet] Issue closed: #2497 Crash in zoom out

21:42

[ddnet/ddnet] Pull request closed: #2498 Limit zooming out further (fixes #2497)

21:42

[ddnet:master] 2 new commits

4d34dd3 Limit zooming out further (fixes #2497) - def- 3ff0a7f Merge #2498 - bors[bot]

f10 (screenshot) doesn't work when holding alt, ctrl etc.

22:09

it did work before

22:10

because alt is "show chat" for me and I wanted to screenshot it

2

Exported 204 message(s)