Time  Nick      Message
15:43 hecks     hi, what does m_ stand for in class fields?
15:43 sfan5     member
15:44 hecks     so this just means "not a local or outside variable"?
15:45 sfan5     yes
15:45 sfan5     it's not supposed to be used for new classes though, just for consistency with old code
15:46 hecks     hm, okay, I actually like the distinction, but I'll keep that in mind
15:47 hecks     while you're here, could you please roughly explain to me the path that a mapblock takes from a network message to the mesh displayed in clientmap.cpp, if you're any familiar with this code?
15:49 hecks     I'm trying to understand this stuff in order to do something about the performance, I know how they're being drawn so far
15:50 sfan5     not familiar with it but I imagine it's somewhat like this: "packet handler deserializes and puts it into local map; client requests a mesh and doesn't find it in cache; mesh generator thread takes it from local map and generates the mesh, putting it in cache; at next frame(?) the mesh is rendered too"
15:51 hecks     right, the cache, I forgot about that
15:51 hecks     is there anyone still active who knows more about this area?
15:53 sfan5     unfortunately not, your best bet is to look at the code until it makes sense
15:53 sfan5     the performance of what exactly did you want to look at?
15:53 hecks     I want to reduce drawcalls, possibly by merging block meshes or straight up decoupling drawing from the mapblock structure
15:54 hecks     the problem is that minetest mapblocks are tiny and the chatter between the game and the graphics driver is killing it
15:54 hecks     I've already eliminated geometry counts, fragment fill and other causes
15:55 hecks     meanwhile batches are an obvious hog, starting in the hundreds at low draw distances and getting into thousands with more than that
15:55 hecks     also getting worse the more diverse a mapblock is
15:56 hecks     materials are already batched but this is not enough, a block-layer can contain as little as one face if we're unlucky
15:57 sfan5     I have nothing useful to comment about that, good luck
15:57 hecks     binding and drawing buffers with these ancient GL methods is not cheap sadly
16:13 oil_boi   So you're finally gonna defragment the opengl implementation to speed up new graphics cards hecks?
16:13 hecks     no, just the meshes
16:14 oil_boi   That's still super noice, there's no reason my rx580 black edition should only be getting 350fps on view distance 20
16:14 hecks     there's something up with that actually
16:14 hecks     I get 240 fps just looking at the sky, 4ms are being eaten for *nothing*
16:15 oil_boi   Yes, especially when gta v runs 120fps on ultimate, what are those 4ms even been doing, I'm going to check right now
16:15 oil_boi   draw meshes is taking 0.05555 ms
16:15 hecks     check it; fly high up, look at the skybox
16:15 hecks     drawtime will still somehow be 4ms
16:16 oil_boi   for some reason only getting 220fps on superflat
16:16 hecks     maybe less for you but something noticeable for sure
16:16 hecks     but anyway, upgrading GL would take a new irrlicht video driver implementation, and dramatically different ways of doing things
16:16 hecks     instancing sure would solve this issue elegantly
16:16 hecks     but it wouldn't be compatible and batching is kinda easier to do
16:17 hecks     I'll just merge things into mega-blocks or something
16:17 oil_boi   sky rendering 0.00015 fps at y 350, I'm an fps snob so this is still irritating
16:17 hecks     sure is; "empty scene" is generally 2000+ fps stuff
16:17 Krock     will merge #10143 in 15 minutes
16:17 ShadowBot https://github.com/minetest/minetest/issues/10143 -- FormSpec: Don't stop style parsing on unknown property by v-rob
16:17 hecks     even irrlicht itself renders its demos this fast, with things being drawn and all
16:18 hecks     also who the hell draws the skybox before the world? should fix this if possible
16:18 oil_boi   That is an old technique
16:19 hecks     for... what exactly? don't you draw it last to cull as many fragments as you can?
16:19 oil_boi   it is easier to draw geometry on top of an existing background elements, it's the lazy way to do it, that's why the skybox is the "background"
16:19 hecks     only in a really bad software renderer
16:20 oil_boi   Welcome to irrlicht m8
16:20 hecks     :]
16:20 hecks     ah yes, that wonderful soft driver
16:20 hecks     try switching that on sometime for a laugh
16:20 oil_boi   One thing I'm having an interesting time with is dynamic map updates with pistons, causing some interesting fps mesh rebuilding problems
16:21 hecks     every time a mapblock is touched, it's being remeshed
16:21 oil_boi   Oi, you're going to give me a hernia talking like that one, I'm gonna pretend I'm just a bad programmer
16:21 oil_boi   Oh wait
16:21 hecks     I hope that doesn't ruin my idea of merging blocks
16:22 oil_boi   Do it and create something with mesecons or redstone with pistons as a test and see how it goes
16:22 hecks     a frameship
16:22 hecks     no better way to kill a server than a frameship
16:23 hecks     and it's also a full remesh
16:23 oil_boi   Is that's what is happening now?
16:23 hecks     hm? no, I'm getting thousands of drawcalls on static scenes
16:23 hecks     that's my problem
16:24 hecks     there's also one funny thing about irrlicht here
16:24 hecks     it has... a minimum vertex count for VBOs (?????)
16:24 hecks     500 by default
16:25 oil_boi   https://i.imgur.com/XkJyaEL.png there are no animated meshes in this scene
16:25 hecks     there is in fact a line somewhere in minetest that disables that madness, but someone commented it out a very long time ago
16:25 hecks     so you're getting 3ms with nothing; but keep in mind that the hud and text also eat a lot
16:26 hecks     text rendering is pretty crap
16:27 oil_boi   Ah there's my fps https://i.imgur.com/LhfU16R.png
16:27 oil_boi   You're right, 800 with no text at all
16:27 oil_boi   When I build the game on windows I get like 50 errors about the font rendering engine
16:27 hecks     still crap; also even the debug display counts as text
16:28 hecks     see how it very briefly lerps down from a higher fps when you enable it
16:28 oil_boi   Warning! Unsafe mix of BOOL and INT!
16:28 oil_boi   Oh of course
16:28 hecks     text rendering might also be having batching issues
16:29 oil_boi   I bet it's also a cause of the memory leak
16:29 hecks     crap, what memory leak... do you too get an "out of memory" after leaving and entering the game a few times?
16:30 oil_boi   No, I have 16 gb of memory, but I can make it reach 12GB at the main menu
16:30 hecks     ahahahahaha what
16:30 hecks     just by generating strings?
16:30 Krock     old news. it was already like this in 0.4.7-dev
16:30 oil_boi   Ah I have some free time, let me show you, I'll be back in a little while
16:30 Krock     media not fully unloading after game leave
16:31 hecks     all media?
16:32 Krock     cannot tell anything more precise - it's the only thing in Minetest that consumes this much RAM and varies per server
16:32 hecks     and is vorbis decoded to a wave in ram instead of being streamed, by any chance?
16:32 Krock     not all media - since the RAM usage decreases after leaving
16:32 Krock     merging trivial PR....
16:32 hecks     cause I've been getting this a lot since adding audio
16:44 sfan5     pretty sure it's decoded entirely in ram yes
16:45 oil_boi   I finally caught this bug during my endeavor https://youtu.be/7HNBXRKnkas
16:47 hecks     have you ever been so angry that you just vibrated your feet in place at 60hz
16:50 sfan5     literally never seen that bug
16:51 sfan5     did you compile the engine yourself? using msvc?
16:52 hecks     closest I've ever seen to that was when setting jump force to 0
17:01 oil_boi   Yes and visual studio everything's stock, only happens in MTG
17:04 hecks     wait, is that snow full nodes or those tiny slabs on top?
17:06 oil_boi   I have no idea, it happens on anything randomly
17:06 hecks     oh okay
17:06 oil_boi   Sometimes I don't even have to look down and it blows my ears up
17:06 hecks     but... only mtg? curious
17:07 oil_boi   Yep
17:07 oil_boi   Also the game hit 7 gb and reset somehow, it went to 0MB and froze
17:07 hecks     what if you transplanted some of the sounds to devtest and tried to repro it
17:07 oil_boi   But here, I did a little more and https://i.imgur.com/IZILm7i.png oof oof oof
17:08 oil_boi   I just realized the main menu is using 50% of my gpu there, WAT
17:13 oil_boi   I forgot I had the game open oh my god
17:14 oil_boi   My cpu went from 70C to 28C
17:15 oil_boi   It's water cooled
17:17 hecks     oh yeah, this is the primary reason I'm looking into map drawing perf
17:18 hecks     I just want to have minetest open without being cooked alive
17:21 oil_boi   This was left open at the main menu man
17:22 hecks     but I thought the main menu was capped at 30 fps?
17:22 hecks     it sure feels that way
17:22 oil_boi   It's linked to the in game pause menu apparently, 2000fps at the main menu
17:23 hecks     oh so you have it unlimited in the config?
17:23 oil_boi   Yep, the pause menu in game basically stutters if I have it below 250
17:24 oil_boi   It's really, really annoying
17:24 hecks     well, it's either vsync or weld a refrigerator to your GPU
17:25 oil_boi   The real question is, why is the pause menu fps linked to the main menu fps?
17:28 oil_boi   vsync with 60fps limit at main menu is using 5% gpu and only rendering at 37 fps
17:28 oil_boi   What even is life
17:28 Krock     oil_boi: sleep_ms(25); in the code. this has already been reported
17:29 hecks     oh wow
17:29 oil_boi   What else is sleeping during the main client loop?
17:30 Krock     in-game it sleeps when vsync does not do the frame capping
17:30 Krock     game.cpp IIRC
17:31 Krock     oil_boi: regarding your "empty" world rendering: I suppose the GPU is not fully under load, but rather the CPU trying to render the text
17:33 Krock     TTF rendering can be quite slow, especially because it's rendering the text twice per frame (shadow and white surface text)
17:34 hecks     could probably do that with a shader instead
17:34 Krock     that would be lovely, yes. I have zero clue about GPU stuff
17:34 oil_boi   meshupdatethread is sleeping
17:34 Krock     oil_boi: separate thread. does not matter.
17:35 hecks     I wonder if it's something as bad as a draw per character, or even glBegin
17:35 Krock     hecks: it "plots" all same colored characters at the same time
17:35 Krock     FYI. dunno how important that is
17:35 hecks     but does it cache the text mesh?
17:36 Krock     mesh? All I can see are 2D image batch draws; https://github.com/minetest/minetest/blob/master/src/irrlicht_changes/CGUITTFont.cpp#L554-L566
17:37 hecks     uh oh
17:37 hecks     that's glBegin
17:37 hecks     per character
17:37 hecks     this is what kills statbars too
17:37 Krock     so glBegin is slow?
17:37 hecks     definitely more than a cached VBO
17:38 hecks     maybe the "batch" part here does something to make it faster but I know that text in minetest is atrociously slow so I wouldn't be surprised
17:38 oil_boi   client\fontengine.h(51,1): warning C4805: '|': unsafe mix of type
17:38 oil_boi   55 times
17:38 hecks     that ain't it, the font engine only makes the sprites out of TTFs, right?
17:38 Krock     unsafe but works just fine
17:38 hecks     bools *are* ints :]
17:39 oil_boi   Also, with the mesh generation thread disabled, I am now getting 350-430 fps vs 200
17:39 oil_boi   mesh generation sleep*
17:39 Krock     hecks: found this in my history: https://github.com/minetest/irrlicht/blob/6fe86df4b37db76bf9d17f515b704de3cc5e1a68/source/Irrlicht/COpenGLDriver.cpp#L2140-L2174
17:39 hecks     hmmmmmm
17:40 Krock     oil_boi: please point out the line you changed
17:40 Krock     this could mean there's an unopened lock
17:41 oil_boi   I went crazy so I will pastebin
17:41 hecks     yup, literally draws a quad by feeding it to the driver vert by vert, this is slow
17:41 hecks     the proper way to do text is to make an atlas out of the font, then make and cache meshes for strings since strings do not change often
17:42 hecks     but this is a can of worms and it's hard to do right
17:42 hecks     when you add unicode and especially CJK chars into the mix, it's a world of pain
17:43 Krock     if it were at least possible to cache it
17:43 Krock     render once to a texture, and then draw that
17:43 hecks     that too is possible
17:43 hecks     but the atlas approach lets you change it faster if the character set doesn't change
17:43 Krock     freetype should already do the rendering for us
17:44 hecks     but freetype only gives you bitmaps from ttf, doesn't it
17:44 hecks     it's a game engine's job to make the atlases for GPU friendly rendering
17:44 Krock     looks like that... yes.
17:44 hecks     anyway if you're using an atlas and someone is typing ansi only, you don't have to reupload any textures and it's fast
17:45 oil_boi   https://pastebin.com/raw/M49LY7hX
17:45 oil_boi   Oh by the way, you guys know that weird like pause you get when you load into singleplayer? It's instant now
17:45 oil_boi   There is no delay entering a world after hitting play
17:45 hecks     so you just commented out all the sleeps and the game is smoother?
17:45 hecks     madman
17:45 hecks     I'm gonna try this
17:45 Krock     oil_boi: you set the network throughput to may speed
17:46 Krock     *max
17:46 hecks     whoops, so which sleep should be left alone?
17:46 hecks     then again it's okay to do this in singleplayer
17:46 oil_boi   If you think that's bad, this is with unlimited viewing range https://i.imgur.com/1LjNKwP.png, 320 fps while taking a screenshot with no weird stutter glitch happening
17:47 hecks     aaaaaaaa
17:47 oil_boi   With unlimited viewing range
17:48 Krock     oh no
17:48 Krock     QueuedMeshUpdate *MeshUpdateQueue::pop()   holds a lock
17:48 Krock     though only during pop... not after
17:48 oil_boi   Time to comment that out too
17:49 Krock     dno't
17:49 Krock     it won't change anything and make the rendering instable
17:49 Krock     *unstable
17:49 oil_boi   *breaking news today, a local man makes the entire local area internet go down as the server room catches fire, we'll be back with more at 11*
17:50 Krock     you cannot simply remove locks and expect it to work better. they're there to avoid concurrency issues
17:51 Krock     I don't understand how the FPS is higher with sleep_ms(m_generation_interval); commented out. are you sure about that?
17:51 hecks     but you're not supposed to sleep while locking something either, in case that's happening
17:51 Krock     hecks: well, that's what I thought. It doesn't happen.
17:51 Krock     *But that doesn't happen
17:53 oil_boi   Hahahaha I changed while to an if statement and it's crazy
17:53 hecks     what happened
17:54 Krock     I'd expect Minetest to completely stall due to held locks
17:54 hecks     hey, the block stutter is gone
17:56 oil_boi   hecks https://youtu.be/anfC9SMvseM
17:57 sfan5     sounds like you've reinvented this https://github.com/minetest/minetest/blob/master/minetest.conf.example#L894
17:57 oil_boi   hecks, what did you do to make the block stutter go away?
17:58 hecks     just killed the sleeps where you did
17:58 oil_boi   Amazing
17:58 hecks     I mean the stutter that happens when loading blocks, I guess there's a config var for that
17:58 Krock     that's what F4 did until the "Disable camera update" key was removed
17:58 hecks     but that makes sleeping in the first place questionable
17:59 oil_boi   q->crack_pos = m_client->getCrackPos(); smuggling drugs through code are we?
17:59 hecks     for the cracko.png obviously
18:00 hecks     but I was meaning to ask this actually, what is the crack visual doing in meshing
18:00 hecks     does the whole block get remeshed when someone just starts scraping a node?
18:00 sfan5     the answer may surprise you and I think yes
18:00 Krock     hecks: yes, also animated nodes
18:00 hecks     and couldn't it just be... drawn separately?
18:00 hecks     well animations I can understand
18:00 Krock     perhaps
18:01 oil_boi   hecks which line is rebuilding on scraping a node?
18:01 sfan5     pretty sure animations are a thing not affected by this
18:01 sfan5     since it just swaps the texture
18:01 sfan5     (each frame is in fact stores a separate texture)
18:01 hecks     animations don't scroll texcoords?
18:01 hecks     oh well this works too
18:01 sfan5     it's terrible
18:01 oil_boi   std::time_t t_now = std::time(0); what the, 0 just works too??
18:01 sfan5     it should just use texcoords
18:01 hecks     no wait
18:01 hecks     if it uses texcoords, you'll have to touch the vbo
18:01 hecks     if you swap the texture, it's better
18:02 hecks     you just bind a different texture but the vbo is intact
18:02 sfan5     hm yes that just crossed my mind
18:02 hecks     am I understanding this right?
18:02 sfan5     changing the texcoords requires uploading the mesh again, doesn't it?
18:02 hecks     yup, at least the texcoords buffer on GL's side
18:02 hecks     but irrlicht might not give you this luxury
18:02 hecks     and just happily upload the whole mesh again, which would suck
18:03 sfan5     then it's not so bad I guess
18:03 hecks     if you had GL level control over this, it would be acceptable
18:03 sfan5     ..and I can cross looking into that off my todo list
18:03 oil_boi   So you're telling me that every time I start digging a node it's doing a full rebuild of the mesh
18:04 sfan5     perhaps
18:04 oil_boi   I thought I was a madman
18:04 hecks     what would it take to link to GL on windows with mingw?
18:04 sfan5     honestly I'm not sure, it also creates an extra texture for default_dirt.png^[crack:1:0
18:04 sfan5     so maybe it's just a swap and I was wrong
18:04 hecks     and yes, we need to take crack off the streets... I mean, the mapblocks
18:05 sfan5     hecks: what's special about linking to GL on mingw?
18:05 oil_boi   Also I can't believe the digging microstutter is gone with that sleep removed
18:05 hecks     someone said that windows builds fail if GL is used directly, in that uh alpha test discard PR
18:05 sfan5     oh that was me
18:06 hecks     I'd be happier if that was sorted out because talking to GL directly would let me bypass irrlicht's brain damage when needed
18:06 sfan5     ah yes, the engine does not directly link to OpenGL
18:06 hecks     also if cracko is applied to default_dirt then wouldn't it be also applied to every other texture in existence? and since it's a different texture, it still needs a remesh because of greedy meshing
18:06 sfan5     just transitively because Irrlicht sits in the middle
18:06 oil_boi   Yo why is the smooth lighting calculated in the mesh rebuild thread? https://i.imgur.com/GH2nDxX.png
18:07 sfan5     so you can't just include gl.h and use those functions (with the current windows setup)
18:07 sfan5     hecks: it creates a copy
18:07 oil_boi   That's why we have chunk border smooth lighting errors
18:07 hecks     this still poses a problem for my batching idea
18:08 hecks     suddenly temp copying one mapblock isn't viable
18:08 oil_boi   And that's why punching a node fixes this
18:08 hecks     punching a node also fixes physics errors, I wonder why
18:08 sfan5     oil_boi: the reason there are smooth lighting errors is https://github.com/minetest/minetest/pull/6765
18:08 oil_boi   Nah, you're just pulling my leg man
18:09 hecks     punching a node will fix the bug where entities cannot enter a mapblock and raycasts in that mapblock fail
18:10 oil_boi   If smooth lighting is disabled globally why is the update still attempted on cache fill?
18:40 hecks     so I'm experimenting with a bigger blocksize globally at first, it sure is slow to generate
18:40 hecks     but the framerate sure is promising, even though culling sucks
18:47 * oil_boi explores v3s64
19:15 oil_boi   Max mapblock size 512 gets me over 1000fps
19:21 oil_boi   Yep any higher that crashes the entire computer
19:21 hecks     but man, do higher block sizes render well!
19:21 hecks     meshblocks are officially on the table
19:40 pgimeno   [0703 20:02:25] <sfan5> changing the texcoords requires uploading the mesh again, doesn't it?
19:41 pgimeno   not if the meshcoord change happens in the shader
19:41 pgimeno   texcoord* not meshcoord
19:42 hecks     true, but then again it's not a good idea anyway because that would require you to tessellate geometry to tile the texture
19:46 pgimeno   what do you mean with tessellate geometry?
19:47 pgimeno   you would have the animation in a texture atlas
19:47 hecks     but how would you tile a single frame of the texture itself? you'd have to disable greedy meshing just for this
19:48 hecks     and it would not work on things like meshes that use UVs outside the 0-1 range
19:48 hecks     binding a different texture is more sane actually
20:28 hecks     finishGenOOM environment_OnGenerated(): not enough memory, what the hell
20:28 hecks     the program's working set was barely above 1 GB, VRAM barely above 100 MB
20:28 hecks     lua memory at 16 megs
20:28 sfan5     luajit memory limit?
20:30 hecks     but... how, all I did was bump the block size
20:30 hecks     also this wasn't happening at first
20:31 hecks     not that I want to keep a large block size, but it's a mystery
21:30 hecks     so guess I can confirm that cracko does not affect meshing, thank goodness
21:43 Zughy     hey there guys. Is there any way to remove the shadow bg of a text? (ie for a nametag)
21:43 Zughy     in C++ ofc
21:46 Calinou   it's applied on the client side with the `font_shadow` minetest.conf setting, I don't think you can control it
21:57 Zughy     ty