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 2022-07-17 00:00:00Z and 2022-07-18 00:00:00Z

How can i expand a NetMessage and still be able to connect to older server versions that dont have it? (edited)

[ddnet-maps:master] 1 new commit

e977bbc A Campotle 1, A Campotle 2, A Campotle 3 - ddnet-maps

cauldron

servelist is refreshing, when you switch tabs, but not when you come from ingome to server browser. is it worth to add a refresh there?

Can't hurt i guess

network.py (edited)

Image attachment

06:28

^This compiles (edited)

06:30

but this does not:

06:30

Image attachment

06:31

Image attachment

06:31

why

i assume u didnt initialize the variable

Jupstar ✪

i assume u didnt initialize the variable

its automated

bluesky

network.py (edited)

this compiles and its the same just a different message

not really

06:59

u pass a pointer over it

07:00

if u dont pass a pointer over it, its uninitialized

yes

https://github.com/ddnet/ddnet/blob/3c6402830d0792d285852f5cd1159c7e4d558017/src/game/client/gameclient.cpp#L2078-L2110

ddnet/gameclient.cpp at 3c6402830d0792d285852f5cd1159c7e4d558017 · ...

DDraceNetwork, a cooperative racing mod of Teeworlds - ddnet/gameclient.cpp at 3c6402830d0792d285852f5cd1159c7e4d558017 · ddnet/ddnet

07:01

show me this code in ur version

Image attachment

see

07:03

skinbody missing

i know

bluesky

network.py (edited)

this still compiles

07:04

well let me initialize it

but compile != runtime

yeah i mean the upper one doesnt give me an error

ignore the python stuff, its just to generate structs

07:04

imo also useless for network structs

i cant judge that i dont know enough about it

bluesky

yeah i mean the upper one doesnt give me an error

doesnt or does?

I want to pass an aditional bool and 4 strings to every player as soon as one changes his skin

07:05

how would you do that

Jupstar ✪

doesnt or does?

doesnt

bluesky

how would you do that

u'd use a different packet anyway, bcs this rn destroys vanilla compability and the ddnet admins will roast u

jaouis

1

07:06

but generally, i'd do it similar to your version, u just need to initialize the variables

07:06

everything else was correct

okay i will try it

Jupstar ✪

doesnt or does?

its weird because i only get the error when i add a string to ChangeInfo, not StartInfo

what error does it display?

07:08

maybe someone added a sizecheck or smth

Jupstar ✪

u'd use a different packet anyway, bcs this rn destroys vanilla compability and the ddnet admins will roast u

i will just use updateskin packet like in 0.7

07:08

for skinchange

bluesky

Click to see attachment 🖼️

this error on adding a string in ChangeInfo

bluesky

this error on adding a string in ChangeInfo

then its just luck

<:(

uninitilaized values can be anything

Jupstar ✪

uninitilaized values can be anything

it works now obviously after initializing it

07:13

but i still wonder why the StartInfo works without initializing it

set a breakpoint, see what value m_pSkinBody has

07:14

then u know it

i searched references but didnt find any so i might do the brakpoint

07:15

yes

07:15

thank you

07:15

i hope it get it to work

bluesky

i searched references but didnt find any so i might do the brakpoint

u wont find anything. Imagine it like this: RAM consists of random bytes at the start, bcs always setting all bytes of your whole RAM to zero at the computer start would be useless, u dont use all RAM most of the time now c++ allows you to use uninitialized memory. This is basically the RAM as is. This can in some situations be useful if you dont care about initializing it to zero before using it

yeah memzero before using big data

07:22

I found out why the behaviour seemed weird

07:23

ChangeInfo gets used before StartInfo "OnRender" calls SendInfo(false) (edited)

i will make a pr soon and i hope its gets good constructive critic because i really wanna optimise and learn how to do things better (edited)

[ddnet/ddnet] Issue opened: #5635 Using /practice before crossing start results in broken team

After doing this, you can't change teams or start the race. The only way to get out is to unlock the team, kill your tee, and rejoin the team.

^ really annoying imo

09:17

sometimes i need to show how to do something in a speedrun when none of us have started racing, but as soon as i do it we have to rejoin team again cus i forget this bug

09:18

tbqh didnt know how it worked exactly until now xd

09:18

okay why cant i reproduce it suddenly

09:19

there is SOMETHING broken about it, i thought for sure this was the way to produce it

09:19

grrr

[ddnet/ddnet] Issue closed: #5579 Show Practice Indicator in Scoreboard for teams that are in practice

09:21

[ddnet/ddnet] Issue closed: #5635 Using /practice before crossing start results in broken team

oh i see what codedev said

09:51

makes sense then

Jupstar ✪

Can't hurt i guess

well i was just wondering.

https://scratch.mit.edu/projects/322341152/

10:59

terraria in scratch

bluesky

How can i expand a NetMessage and still be able to connect to older server versions that dont have it? (edited)

Not possible

cauldron

well i was just wondering.

then transform your wondering into a pr xdd

@Ryozuki I added documentation to my rust stuff: https://github.com/heinrich5991/ddnet/commit/655c812992f6d0f46fcffaa28d59ddd9da12ace8

nice

@Jupstar ✪ @Tater I'm thinking about implementing a tilemap optimization, because the performance in browsers isn't optimal yet (apart from tilemap optimizations I could 1. add mipmaps or 2. add more stuff to the renderer like entities or smth) I currently can think of 2 things that could be big optimizations: 1. render the tilemap to a texture first and for each next frame only render the difference (this would make tilemaps basically a no-op if the camera doesn't move), as explained in the factorio blog here https://steamcommunity.com/app/387290/discussions/0/1694924244565179392/ 2. reduce how much of the screen we render for each tilemap each frame. maybe split the screen into multiple quads and calculate on the cpu which of the quads actually have tiles on them, or just calculate how much clipping we can do by kinda calculating a bounding box of the visible tiles on the screen for 2. there might be more options but I don't see them yet which of the two do you think will be the better improvement?

14:45

oop, wrong link

14:46

https://www.factorio.com/blog/post/fff-333

Patiga

@Jupstar ✪ @Tater I'm thinking about implementing a tilemap optimization, because the performance in browsers isn't optimal yet (apart from tilemap optimizations I could 1. add mipmaps or 2. add more stuff to the renderer like entities or smth) I currently can think of 2 things that could be big optimizations: 1. render the tilemap to a texture first and for each next frame only render the difference (this would make tilemaps basically a no-op if the camera doesn't move), as explained in the factorio blog here https://steamcommunity.com/app/387290/discussions/0/1694924244565179392/ 2. reduce how much of the screen we render for each tilemap each frame. maybe split the screen into multiple quads and calculate on the cpu which of the quads actually have tiles on them, or just calculate how much clipping we can do by kinda calculating a bounding box of the visible tiles on the screen for 2. there might be more options but I don't see them yet which of the two do you think will be the better improvement?

im interested in 1., but i think you shouldnt underestimate that waiting for the previous frame can cost lot of idle time, so if you not runnning into a bigger GPU bottleneck it might work out even worse, how often does the camera actually stand still? and has no animations etc.

I'd do that for each tilemap individually, there are no movement animations for tilemaps and I'd render it in the normal color and add the envelope color in the fast simple render step

that sounds like a lot of memory. it seems you could batch tilemaps with same parallax together

Patiga

I'd do that for each tilemap individually, there are no movement animations for tilemaps and I'd render it in the normal color and add the envelope color in the fast simple render step

ok but the whole layer?

same parallax and same color envelopes etc.

or your tilemap as you have them now?

@Jupstar ✪ I'm not sure what you mean, each layer would be rendered individually, yes

14:55

@heinrich5991 yea that would be a good optimizations, but I'd rather do that at the end, because I think thats quite some added complexity

Patiga

@Jupstar ✪ I'm not sure what you mean, each layer would be rendered individually, yes

yeah but u want to render the full layer?

I think the full list of things that have to be the same is parallax, envelope, envelope offset

like 4000 x 4000 tiles? XD

no, only what the camera sees

or just apply it per layer

14:56

ok

14:56

but still, how often does ur camera stand still

14:56

especially when in an "editor" mode

well also if it moves, the amount of tilemap calculations is much much less I think

but anyway, please try it out, i am interested in it

1

for each frame you probably move less than a pixel I'd guess

14:57

maybe thats a little exaggerated

well for zooming out far it might give some nice performance extras ^^

Jupstar ✪

then transform your wondering into a pr xdd

sure, i guess thats an easy one

[ddnet/ddnet] Issue closed: #5569 Global static constants in `SCREAMING_SNAKE_CASE` or `gs_CamelCaseWithPrefix`?

Jupstar ✪

well for zooming out far it might give some nice performance extras ^^

but actually maybe not

Jupstar ✪

but actually maybe not

why not?

heinrich5991

why not?

his method already kinda is linear

15:01

and doest require to wait for the last frame

15:02

zooming out or not doesnt make a big difference in his method

15:02

bcs it doesnt rely on the vertex shader

Embedded image

15:10

Many people report this, not only on my server but als others

15:11

They say they need to connect multiple times to fix it

15:11

This doesnt seem to have arrived perfectly ^^

15:11

Whatever it is made for

@Jupstar ✪ ah, another thing I'm wondering about is if there is a difference between doing many render passes in a command buffer or simply doing a new command buffer for each render pass and submitting them together

Patiga

@Jupstar ✪ ah, another thing I'm wondering about is if there is a difference between doing many render passes in a command buffer or simply doing a new command buffer for each render pass and submitting them together

i dunno, but many render passes might be a problem ofc, at least if you load the previous framebuffer

15:14

OP_LOAD vs OP_DONT_CARE

Patiga

@Jupstar ✪ ah, another thing I'm wondering about is if there is a difference between doing many render passes in a command buffer or simply doing a new command buffer for each render pass and submitting them together

the idea behind multiple command buffers is really more for multi threading, or if you are able to have completly static command buffers

15:15

for ddnet i use them only for multi threading

oh, to be able to record them faster?

yeah kinda, you can build a signle frame in many threads

15:16

https://user-images.githubusercontent.com/6654924/153448356-941222a3-8bd3-424d-8685-a43389a4f691.png vk_st vs vk_mt is mostly about building the command buffers

Embedded image

I don't think I need to load the previous framebuffer, though I also don't have DONT_CARE, only clear with a color

15:18

ah right, I do need load for the tilemap render textures

mhh, i assume the opengl backend doesnt support it

load? eh :c

but yeah u need to load anyway

15:20

ah

15:20

that also doesnt exists?

load exists

ok

ah right, my message there is confusing

anyway, my prediction is, fast GPUs will not gain anything might even loose FPS, very slow GPUs might gain from it. Let's see how it turns out

15:23

and i assume web performs bad bcs of other reasons than your renderer, it simply sucks xdd

15:23

maybe too many fallbacks for GLES3 compability

heh ^^

Patiga

heh ^^

btw it ran quite ok for me, do u use firefox?

yes, but not at 144 fps as on chromium :d

15:33

mostly at 80 iirc

15:35

ah before I start trying that out, can one of your tools figure out how much time is spent in the tilemap fragment shader opposed to the rest?

on the browser?

that would be best

15:36

but shouldn't the native time distribution be about the same?

if webgl doesnt fallback to software rendering or smth

sooo, do you happen to have such a tool? ^^

Patiga

sooo, do you happen to have such a tool? ^^

i looked a bit, i'd say the wgpu backend does lots of useless state changes

Patiga

sooo, do you happen to have such a tool? ^^

lots of repeated webgl state changes

Image attachment

15:47

2303 commands for map "tutorial" on default zoom

15:47

sounds insane

that is the amount of commands inside of the command buffer per frame?

@Patiga ddnet with GLES 3 does 603 for everything(HUD, text, etc.)

Patiga

that is the amount of commands inside of the command buffer per frame?

yes

15:49

at least i assume it is?

15:49

or is that a record of multiple frames, let me see

how is that tool called?

spectate.js or smth i alraedy gave it to you a while ago

15:50

but it also crashes my firefox often rn xD

heh ^^

but yeah i'd say single frame

15:51

https://addons.mozilla.org/de/firefox/addon/spector-js/

Spector.js – Holen Sie sich diese Erweiterung für

Laden Sie Spector.js für Firefox herunter. Easily explore and debug any WebGL context.

when I go to the Information tab up there it says

Commands Summary total: 2301 draw: 61 clear: 0

15:58

so maybe its actually the total amount of commands?

i assume so

15:58

but ddnet uses 603 commands for everything

15:58

hud, text and ingame rendering

but then I'd only use

15:59

~37 each frame, right? but that sounds too little

what?

is each draw call a frame?

no

16:00

1 frame 603 commands

ah, each draw is a draw call on the pipeline

ddnet has more drawcalls anyway, you cannot directly compare it, but it doesnt call random opengl commands for no reason xD

fokkonaut

Click to see attachment 🖼️

these messages are just troubleshooting tips when you don't connect to the game server instantly

16:01

they're not stopping you from connecting

@Patiga i get > 600 fps btw

Image attachment

16:02

on firefox

16:03

chrome sadly forces vsync, dunno how i can disable it

heinrich5991

they're not stopping you from connecting

Ahhhh

Jupstar ✪

@Patiga i get > 600 fps btw

oh that would be more than enough, what graphics card do you have?

16:03

I have nvidia 1060 3GB

rx 6900 xt xd

Jupstar ✪

anyway, my prediction is, fast GPUs will not gain anything might even loose FPS, very slow GPUs might gain from it. Let's see how it turns out

my GPU probably isn't fast, but zooming out very far goes down to just a couple of FPS

so maybe little OP

that might be a tiny bit OP

that's probably going to be fixed by not re-rendering everything

heinrich5991

my GPU probably isn't fast, but zooming out very far goes down to just a couple of FPS

on patigas renderer?

no, in ddnet

yes ddnet has vertex shader overhead

ah sorry, didn't read the whole conversation

I gtg off for now, will look into it more later

ok

Patiga

I gtg off for now, will look into it more later

540 commands on ddnet web edition

[ddnet-scripts:master] 1 new commit

171c114 Keep builds directory but delete content - def-

@Jupstar ✪ can I also see how long the rendering took / how long each fragment shader took?

17:03

or can you see that information on native maybe?

renderdoc has that but not per stage, but its defs ur fragment shader anyway

why is it definitely the fragment shader?

bcs ur vertex shader makes basically nothing

I guess I could test it by seeing how performance scales with the amount of tilemap layers

17:05

well it could maybe also just be slow in general, no? I'm not sure how big the wgpu abstraction overhead is

yes could be, i have to compile without vsync for my version, then i can see how much fps i get

[ddnet/ddnet-web] Pull request opened: #205 σℓí♡ as FNG mod

@Jupstar ✪ my measurements on native: empty map - 40,000 fps 16 tiles layers - 6750 fps 32 tiles layers - 3700 fps 64 tiles layers - 2050 fps 16 quads layers - 36550 fps 32 quads layers - 35950 fps 64 quads layers - 33300 fps

17:43

so yea, the tiles layers are quite heavy

lmao, but ur broken fps

ah right

17:46

did I ask already how to properly count fps? ^^

take the average xd

17:49

or just use external tools

average of what exactly?

the frame times

and can I measure the frame time on the cpu?

i assume ur whole code is single threaded?

yes

so every new render call is a new frame basically

17:51

it doesnt buffer for ever

but how is that different from my approach? counting the frames and dividing by the amount of seconds

i dunno u did that for like 5 seconds

17:53

the average for 5 seconds xd

buts thats what I do

17:53

for 5 seconds I increment a counter

17:53

and divide that counter by 5 to display the fps

but 40k fps never xD

hmmm

17:57

I forgot to divide by 5 xd

17:58

it seems I neglected that tool a bit too much

18:03

adjusted fps counts: empty map - 8,000 fps 16 tiles layers - 1350 fps 32 tiles layers - 740 fps 64 tiles layers - 410 fps 16 quads layers - 7310 fps 32 quads layers - 7190 fps 64 quads layers - 6660 fps

Patiga

adjusted fps counts: empty map - 8,000 fps 16 tiles layers - 1350 fps 32 tiles layers - 740 fps 64 tiles layers - 410 fps 16 quads layers - 7310 fps 32 quads layers - 7190 fps 64 quads layers - 6660 fps

mhh, what do u think is the bottleneck, GPU cores?

18:07

VRAM usage is also relativly high with this rendering technique right?

hm I'm not sure, we'd need to measure that

normally textures and vertex buffers are highly optimized, but ofc u access the textures for every single fragment

18:10

u basically have a texture lookup more than ddnets renderer

well the vram usage of each tilemap is easy to calculate with this rendering technique, how many does the ddnet version use for each tile on this map?

18:11

width * height * 4 * 4

18:11

I could reduce it by a factor of 2 by using f16 instead of f32 for the tilemap

do u need more than u8?

18:13

uint 8

I don't know what the bottleneck is

its just tile index right?

18:13

or other stuff in it?

yes, I used floats so that it is less work for the fragment shader

18:14

its index, flip_x, flip_y, rotate

mh ok

index has to be (index + 0.5) / 256 in the fragment shader

18:15

and flip_x, flip_y, rotate would need to be cast to float

18:16

the index calculation coz it has to index into the mapres 3d texture

yeaj

18:17

yeah

18:17

well anyway, vram isnt biggest issue, its really just about how much overhead comes from sampling

18:17

f16 might be lighter too ofc

ah

[ddnet/ddnet] Pull request opened: #5636 Minor refactoring of `CNetBase`

Checklist

[X] Tested the change ingame
[ ] Provided screenshots if it is a visual change
[ ] Tested in combination with possibly related configuration options
[ ] Written a unit test (especially base/) or added coverage to integration test
[ ] Considered possible null pointers and out of bounds array indexing
[ ] Changed no physics that affect existing maps
[ ] Tested the change with [ASan+UBSan or valgrind's memcheck](https://github.com/ddnet/ddnet/#using-addres...

[ddnet/ddnet-web] Pull request closed: #205 σℓí♡ as FNG mod

20:25

[ddnet-web:master] 2 new commits

e38df14 σℓí♡ as FNG mod - murpii 4d10d36 Merge pull request #205 from murpii/patch-11 - def-

[ddnet:master] 4 new commits

4970a61 Remove dead code - Robyt3 5f7d19d Use sizeof instead of hardcoded value - Robyt3 b594a32 Change return type of CNetBase::IsSeqInBackroom from int to bool - Robyt3 989dda7 Merge #5636 - bors[bot]

20:44

[ddnet/ddnet] Pull request closed: #5636 Minor refactoring of `CNetBase`

if(DoButton_CheckBox(&m_Dummy, Localize("Dummy settings"), m_Dummy, &Dummy))
{
    m_Dummy ^= 1;
}

21:15

menu_settings.cpp why is it ^= instead of =

^ is bitwise XOR

21:19

so this means m_Dummy = m_Dummy ^ 1

21:20

it's equivalent to m_Dummy = !m_Dummy

oh okay now i get it

21:21

damn everything is so extra in cpp when there is a more common solution to most stuff

21:21

xD

^ is not extra its a common lang feature

22:27

what other symbol would you use?

XOR is an operation lost to users of newer languages

Exported 268 message(s)