#perf_prof_branch

1 messages · Page 6 of 1

heavy galleon
#

I suspect the freezes might be because I start a mission immediately. As I have not encountered it when "playing" normally.

whole cloud
#

Freeze or stuck in loading screen?

heavy galleon
#

Ah, it freezes for a second when I start singleplayer/multiplayer from editor. I can record it later.
Always freezes for a second or so 8 or so times and after that everything is good. But not always haha

whole cloud
#

I count that as a stutter

heavy galleon
#

Yeah, that would be more appropriate name

#

But it's so freaking random

echo vessel
#

@whole cloud Just letting you know i'm having crashing in SP with the build everytime I get a "saving game" prompt and gives me 0xC0000409 STATUS_STACK_BUFFER_OVERRUN

#

Doesn't do it with the default build

humble haven
#

Game crashes when saving in SP, no error message, simply closes Arma

#

This is on profiling branch

jagged minnow
quaint flame
#

None of the changes related to upgrading VS are present in dev branch right?

feral harness
#

we are not that lucky

patent dome
whole cloud
#

Fix on monday 🫂 notlikemeowcry

vale shoal
#

I don't know if this error is already known or fixed, but I had the error once that the server couldn't start because of "PhysX_64.dll wasn't found" even when the DLL is in the folder. I switched only the arma3server_x64 to the latest perf. On the second startup the server started without errors. I send the dump file if needed.

feral harness
#

PhysX or RotorLib?

vale shoal
#

PhysX

whole cloud
#

"in the folder" you mean Dll/x64 folder?

vale shoal
vale shoal
#

But also it never happened again. The server restarted 3 times by now and everything was ok.

restive pilot
vale shoal
woven loom
#

Personally played MP yesterday with a bunch of mods and didn't have any issues that I noticed

whole cloud
#

Cannot produce the saving crash in editor scenario, or in the art of war gallery 🤔
Only specific missions crashing on save?

patent dome
whole cloud
#

I shall try the "STARRFLÜGLER" mission 😄

quaint flame
#

Crashing on save? I had that as well. Let me see if I can get the dump file later.

whole cloud
#

Won't need the dump anymore if I can repro, I'll find that out in a few minutes

quaint flame
#

Ok

whole cloud
#

Reproduced, its rotorlib. So it only crashes when you have helicopters in the world

I have a weird bug in jets showcase, something is constantly turning off my engine, while taxing and even in the air ? wuht
Could be that I have some weird controllers or keybinds maybe 🤔 But I got no controllers attached that i know about

hollow canyon
whole cloud
#

Yep that was it. I can repro now.

echo vessel
#

Dynamic recon ops missions in vanilla was guaranteed to do it

whole cloud
#

I already fixed it

echo vessel
#

Awesome

whole cloud
#

https://github.com/acemod/ACE3/blob/master/addons/common/functions/fnc_changeProjectileDirection.sqf#L24
Bullet velocity -3.61852956, -58.0264931, -858.032715
normalized -0.00420759153, -0.997712255, -0.0674726516
https://github.com/acemod/ACE3/blob/master/addons/common/functions/fnc_changeProjectileDirection.sqf#L38 Here _vdir is -nan(ind), -nan(ind), -nan(ind)

If I remove the *_adjustDir and +_adjustUp then
_dir = -179.758
_up = -3.86884
_vdir = [-0.00420768,-0.997712,-0.0674727]

So if we ignore the input parameters to that function, then everything would be fine.

whole cloud
#

_adjustDir = 0
_adjustUp = -nan(ind)

https://github.com/acemod/ACE3/blob/master/addons/scopes/functions/fnc_firedEH.sqf#L57C25-L57C43
_adjustUp = _elevation + _zero

_zero = -nan(ind)
_elevation = 0.00000000
_zeroing = [0.00000000, 0.00000000, -nan(ind)]
_adjustment is all zeroes

so this https://github.com/acemod/ACE3/blob/master/addons/scopes/functions/fnc_firedEH.sqf#L45
_zerocorrection = -nan(ind)
_baseangle = 0.00000000

So https://github.com/acemod/ACE3/blob/master/addons/scopes/functions/fnc_calculateZeroAngleCorrection.sqf
_zeroangle = "0.0720524" at Line 36

_zeroangle = "-nan(ind)" at line 70

"ace_advanced_ballistics" callExtension format ["calcZeroAB:%1:%2:%3:%4:%5:%6:%7:%8:%9", _newZeroRange, _initSpeed, _boreHeight, GVAR(zeroReferenceTemperature), GVAR(zeroReferenceBarometricPressure), GVAR(zeroReferenceHumidity), _ballisticCoefficients select 0, _dragModel, _atmosphereModel];

_AmmoCacheEntry = [-0.00120000006, 0.00000000, 0.00000000, 0.00000000, 0.500000000, 1.00000000, [], [], "ICAO", [], [], [], 0.00300000003]

_newzerorange = 200.000000
_initspeed = 860.000000
_boreheight = 5.00000000
_dragmodel = 1.00000000
_atmospheremodel ="ICAO"
_ballisticCoefficients = [];

select 0 on an empty array?
I hope your extension handles nil.

"ace_advanced_ballistics" callExtension format ["calcZeroAB:%1:%2:%3:%4:%5:%6:%7:%8:%9", 200, 869, 5, 15, 1013.25, 0, [] select 0, 1, "ICAO"];
it does not. @stark falcon

Arma3Retail_DX11_x64_2024-03-11_13-08-28.png
#

Didn't see that that was meant to be the solution to the problem. Thought it was just a question notlikemeowcry

hybrid cradle
#

shortest conversation about advanced ballistics

whole cloud
#

2.14.151553 new PROFILING branch with PERFORMANCE binaries, v20, server and client, windows 32/64-bit, linux server 32/64-bit

  • Fixed: Crash when saving game while a helicopter is in the world

Note: CAUTION! Starting with Arma 3 version 2.16, we will not offer a 32-bit version of profiling branch anymore. The 32-bit binaries on profiling branch will be non-profiling/performance builds.

If you don't want to use the Steam branch, the files are also available for alternative download here:
https://drive.google.com/drive/folders/15p9j7C2nHUt6NoVfChX4YFuqzFXzblJh
Note: There are separate Dll files that also need to be placed into Game folder.

camoman

whole cloud
quaint flame
#

we will not offer a 32-bit version of profiling branch anymore
Out of curiosity. What the benefit of dropping 32 bit? Building time?

silk pewter
#

Windows 95 is old

hybrid cradle
#

think main reason for asking is servers

autumn timber
#

Btw I wonder how many servers rely on 32bit extensions that have no 64bit counterparts (or the users have somehow lost their source code and can't rebuild them in 64bit) as I see this as the only reason for not running 64bit Arma servers 🤔

#

If there's another reason for that, I'd be interested in knowing

quaint flame
#

I see. Missed that message from dedmen.

whole cloud
#

Well all the servers that rely on 32 bit, gotta use stable then and cannot have profiling branch benefits

fast hornet
#

i mean.. just checked and the x64 binaries got released at the start of 2017. So 7 years ago. Should've been plenty of time to move away from 32bit 😄

vivid rune
#

I would love to see a 64bit non-AVX build instead.

whole cloud
#

We don't use AVX

restive pilot
#

I was gonna say the same but I do see some avx stuff in the exe 😅
(there were some ymm registers and stuff)

inland dew
#

mainly sse

whole cloud
#

We have some code that checks CPU support and only runs if supported

#

Maybe thats AVX. But I know we don't require AVX

vivid rune
#

OK. Whatever it is the 64bit performance crashes my Phenom II

heavy vortex
#

Phenom II is missing SSE 4.1 and 4.2. Not sure if related though.

autumn timber
vivid rune
#
  1. First was the choice for the main board because it has many PCIe slots for building a storage server from that. Then I nearly maxed out the processor.
whole cloud
#

windows or linux?
windows crashdump?

silk pewter
#

ah no, it's a list start

#
  1. yes
vivid rune
whole cloud
#

Ah the pain

random isle
whole cloud
#

2.14.151568 new PROFILING branch with PERFORMANCE binaries, v21, server and client, windows 32/64-bit, linux server 32/64-bit
- Added: CallExtension feature flags, context stacktrace, and game functions: https://community.bistudio.com/wiki/Extensions#Feature_Flags
- Fixed: The random script command would sometimes return the maximum value; it is now exclusive again (broken since 2008)

Note: CAUTION! Starting with Arma 3 version 2.16, we will not offer a 32-bit version of profiling branch anymore. The 32-bit binaries on profiling branch will be non-profiling/performance builds.

If you don't want to use the Steam branch, the files are also available for alternative download here:
https://drive.google.com/drive/folders/15p9j7C2nHUt6NoVfChX4YFuqzFXzblJh
Note: There are separate Dll files that also need to be placed into Game folder.

blobdoggoshruggoogly

whole cloud
coarse plaza
#

i get this error when launching SOG PF without mods, with perf branch.
Just before main menu opens up.
No errors when using stable.

18:44:16 Error in expression <Name _x + "_intialised",true];
_display displayAddEventHandler [configName _x se>
18:44:16   Error position: <displayAddEventHandler [configName _x se>
18:44:16   Error Foreign error: Unknown enum value: "EMMMenuShow"
18:44:16 File vn\emm\main_menu\fnc\initMenu.sqf..., line 46
18:44:16  ➥ Context:     [] L1 ()
    [] L1 ()
    [] L172 (vn\emm\main_menu\fnc\init.sqf)
    [] L172 (vn\emm\main_menu\fnc\init.sqf)
    [] L34 (vn\emm\main_menu\fnc\initMenu.sqf)
    [] L54 (vn\emm\main_menu\fnc\initMenu.sqf)
    [] L53 (vn\emm\main_menu\fnc\initMenu.sqf)
    [] L44 (vn\emm\main_menu\fnc\initMenu.sqf)
    [] L46 (vn\emm\main_menu\fnc\initMenu.sqf)

18:44:16 Error in expression <Name _x + "_intialised",true];
_display displayAddEventHandler [configName _x se>
18:44:16   Error position: <displayAddEventHandler [configName _x se>
18:44:16   Error Foreign error: Unknown enum value: "EMMMenuHide"
18:44:16 File vn\emm\main_menu\fnc\initMenu.sqf..., line 46
18:44:16  ➥ Context:     [] L1 ()
    [] L1 ()
    [] L172 (vn\emm\main_menu\fnc\init.sqf)
    [] L172 (vn\emm\main_menu\fnc\init.sqf)
    [] L34 (vn\emm\main_menu\fnc\initMenu.sqf)
    [] L54 (vn\emm\main_menu\fnc\initMenu.sqf)
    [] L53 (vn\emm\main_menu\fnc\initMenu.sqf)
    [] L44 (vn\emm\main_menu\fnc\initMenu.sqf)
    [] L46 (vn\emm\main_menu\fnc\initMenu.sqf)
whole cloud
coarse plaza
#

ok 👍 , i have not used SOG while, so i just thought that it was weird that i get it only when using perf branch and not stable.

fickle geyser
#

Does not break anything. Will be fixed in next SOGPF patch

violet elbow
tribal pasture
#

Got LOD flickering bug after having game running for hours doing nothing, anything useful I can get out of it? Stable though, not perf.

#

Seems to be going for 30 minutes in the background already, lots of Fresnel k must be >0, given n=2.51,k=0 spam

restive pilot
#

It happens when the game consumes too much memory and it can't free it up anymore. It's easily reproducible if you simply make a giant hashmap

tribal pasture
restive pilot
tribal pasture
#

Seeing how everything cycles through all LODs, perhaps one of the LODs has a material that triggers this RPT line

#

What would be the best channel to post about this? I just remembered that it was talked about here.

#

Opened map and it finally crashed

#

The only thing I had out of ordinary was a single AI unit 1 km away from me doing nothing

whole cloud
#

Well it happens on stable. Not really anything to get out of it.
Use profiling and get higher memory limits and probably have it just not happen

tribal pasture
# 
Type: Public
Build: Stable
Version: 2.14.150957

Allocator: C:\Program Files (x86)\Steam\steamapps\common\Arma 3\Dll\tbb4malloc_bi_x64.dll [2017.0.0.0] [2017.0.0.0]
PhysMem: 32 GiB, VirtMem : 131072 GiB, AvailPhys : 18 GiB, AvailVirt : 131068 GiB, AvailPage : 14 GiB, PageSize : 4.0 KiB/2.0 MiB/HasLockMemory, CPUCount : 16
restive turtle
tribal pasture
restive turtle
tribal pasture
restive turtle
#

strange

#

i had it a while ago, Zulu Headless Client was causing it for my unit.

spiral pond
#

Fresnel k must be >0, given n=2.51,k=0

#

could be from a3 animals

#

14:13:48 Fresnel k must be >0, given n=2.51,k=0
14:13:48 ➥ Context: #(ai,64,64,1)fresnel(2.51,0) <- ShapeLoad: N:'a3\animals_f\snakes\snake_f.p3d' P:'a3\animals_f\snakes\snake_f.p3d'

restive turtle
silk pewter
analog acorn
feral harness
spiral pond
#

@tribal pasture ^ probably meant for you

silk summit
whole cloud
whole cloud
#

Huh that worked better than expected... aviator

Arma3Retail_DX11_x64_2024-03-19_11-26-59.png
silk pewter
#

gib moar fps!!1!

whole cloud
#

Calculating AI visibility through particle effects takes alot of time.
But its a quite small area to optimize, without many connections to other systems

Arma3Retail_DX11_x64_2024-03-19_11-35-48.png
#

From 1.7% of main thread time spent on particle visibility checks, down to 0.3%

#

2% of time (total mainthread time during a YAAB run) is spent by a windows API function.
That in windows 8 received a much lighter weight alternative to use instead. But we are windows 7 :/

hybrid cradle
#

i assume its more tricky than if windowsversion > 7 then {} else {}

whole cloud
#

No, it could work like that. But that if check would be somewhat expensive

hybrid cradle
#

ah

#

more expensive than the gain?

whole cloud
#

maybe, shall see

inland dew
#

few FPS is margin of error
like 3 and up and that after like 5-7 runs, that can clearly be considered as real improvement

only avg FPS is +- "reliable" in YAAB

whole cloud
#

I am not measuring on fps. I'm profiling the whole run in profiler.
YAAB results are easier to show though

inland dew
#

ah, ok

#

53.7 vs 49.5 is clearly not margin of error

autumn timber
whole cloud
#

SetEvent

#

The bool would likely not be in cache, so it would end up a ram fetch

restive pilot
#

drop win7 support 😠🪧 🥺🪧 hmmyes 🪧

silk pewter
whole cloud
#

Just saw I already implemented teh windows 8 method two years ago.
Just disabled with // Requires windows 8 :( 😄

whole cloud
restive pilot
#

a fnc pointer is equally slow right (assuming they both have the same signature)?

#

I guess in some way it could be even slower think_turtle

autumn timber
#

How do you even get access to both APIs on runtime? Isn't it by getting pointers to them? (since it would fail loading on win7 if it were the loader's job)
Maybe there's some kind of error that is being thrown out when using win8 api that doesn't exist? So that you can optimistically call that new API and then fall back to the old one, if the first one fails?

#

(would make it slower for the 0.1% of users that are still on win7, though hmmyes )

whole cloud
#

seems better enough 🤔

vtune-gui_2024-03-19_13-15-40.png vtune-gui_2024-03-19_13-16-03.png
whole cloud
restive turtle
whole cloud
#

Profiling says 2.6% down to 1.5%. Of total yaab runtime.
But in yaab fps its basically not noticable. And a 1% improvement but having to require windows 8 🤔

We already have about 12% fps improvement relative to 5 hours ago. Good enough for today 🤣

silk pewter
autumn timber
#

We already have about 12% fps improvement relative to 5 hours ago. Good enough for today 🤣
❤️

Btw, honestly, I don't really see a reason to care that much about the tiny minority of players who have not moved to more recent OSes, especially if this will benefit the majority (and that these people will have to move eventually, anyway). Also, ever since Steam dropped support for Win7 on the 1st of January, these players should be happy you're keeping backward compatibility, at all.

whole cloud
#

Maybe we can drop win7 for profiling branch, like we're dropping 32bit.
But not now

autumn timber
#

To be clear, I'm not suggesting dropping support completely (but I'm not against that either, though). I'm just saying that if there's an optimization that will make it faster for players with current OSes (better performance minus cost of the if check) but a bit slower for really old OSes (same performance minus cost of the if check), I'd suggest applying it anyway.

vivid rune
whole cloud
#

no

vivid rune
#

^^ Why?!? (the Hug)

plain trout
#

I think he wrote earlier that it take a lot of time to compile/build the exes

vivid rune
cloud nacelle
vivid rune
cloud nacelle
analog acorn
#

You'll still get them when they come to stable branch

vivid rune
inland dew
#

I might be wrong...

vivid rune
inland dew
#

anyway, I don't see why majority should suffer or miss improvements/progress because of minority...
sorry!

hybrid cradle
#

you see, you can resolve this by changing the time on your computer to 2024

#

and not using 32 bit any more

inland dew
#

and no win 7 and even 8

vivid rune
inland dew
#

xp - 7 - 10 - 12

cloud nacelle
#

The "64bit kids" don't suffer, compile times do

inland dew
vivid rune
inland dew
#

or instead, does much more useful stuff, for the majority

vivid rune
#

Who is the majority? The majority are the stable branch people. And there is still the 32bit. So the more "useful things" can only be performance improvements, because when it comes to features, it has to run on the 32bit.

cloud nacelle
woven loom
cloud nacelle
#

Ofc there are exceptions and sometimes the changes are actually noticeable in FPS (like what was shown today) but even then it's like a 5%-10% increase in FPS (being optimistic here) so if the Athlon gets 20FPS in YAAB on stable, that's a whopping 22FPS on perf 👍

vivid rune
woven loom
#

Yes what 😅 server or client?

vivid rune
#

As server of course.

woven loom
#

oh sorry, you said, i missed it

#

so i think this is primarily relevant to clients

cloud nacelle
#

But, perf doesn't even exist for server (correct me if I am wrong about this)

woven loom
#

and i assume you're running linux server?

vivid rune
vivid rune
woven loom
#

OK

woven loom
hybrid cradle
#

i just looked

vivid rune
#

When it only depends on compile time, what if dedmen kills only the 32bit clients and lkeeps the server?

woven loom
#

they can get more builds per day

#

also i don't understand, they're only only doing this on perf branch, if you use stable then you're not really affected

vivid rune
#

yeah but I use perf

fickle geyser
#

too bad

woven loom
#

Alternatively, switch to 64 bit windows server

#

But i think we're so far past 32-bit machines being used that it's only a matter of time that devs find it unreasonable to support it

vivid rune
kindred radish
#

Just use Windows problem solved blobdoggoshruggoogly

vivid rune
cloud nacelle
#

Or just get a "new" server that has a CPU made during the last decade. Plenty of CPU's newer than your 2008 Athlon are already e-waste for most people, you can get 4th gen intel quad cores for 10€ in the used market and probably for free if you just ask around

inland dew
hybrid cradle
#

i will buy one and ship it to him if thats so much of a problem

vivid rune
woven loom
vivid rune
woven loom
#

Ah

#

yeah buddy time to upgrade then 😅

#

if not now then soon

vivid rune
woven loom
#

probably just switch to stable going forward, i doubt perf branch updates will improve things significantly for the old phenom II

vivid rune
#

Maybe with the new versions the difference is not so big anymore.

woven loom
whole cloud
#

You're the one person affected by this. Sorry. 🤷
This decision was made knowing that one or two people would still like to use it, but decision's been made.

woven loom
#

i guess technically they're using 32 bit because 64 bit leads to a seg fault 😛

whole cloud
#

run gdb, get me offset inside file where the bad instruction is at, I might take a look
But I cannot repro having an old cpu and linux crashdumps are very annoying to work with. And if the instruction is somehow of benefit to all other users, there is no incentive to get rid of it just because there is one user that's affected who has a viable alternative to use anyway.

woven loom
#

@vivid rune see ☝️

#

Phenom II is otherwise x86-64 so the hardware is not too old to run 64-bit as such

#

Wiki says instuctions supported are

MMX, extended 3DNow!, SSE, SSE2, SSE3, SSE4a, AMD64, Cool'n'Quiet, NX bit, AMD-V

vivid rune
woven loom
fast hornet
#

i mean.. after reading all of the latest 200 lines or whatever... just drop 32 bit AND win7/8 support. Noone needs that ancient shit.

whole cloud
whole cloud
#

2.14.151599 new PROFILING branch with PERFORMANCE binaries, v22, server and client, windows 32/64-bit, linux server 32/64-bit
- Tweaked: Rendering performance improvement
- Tweaked: Performance improvement for calculating visibility through particle effects
- Fixed: Some UserActionEventHandler's did not trigger when a UI was capturing the keypress (This hopefully doesn't break any existing UA-EHs) - https://feedback.bistudio.com/T178835
- Fixed: Mortar gunner anim desync fix was breaking other get in turret animations - https://feedback.bistudio.com/T124694

Note: CAUTION! Starting with Arma 3 version 2.16, we will not offer a 32-bit version of profiling branch anymore. The 32-bit binaries on profiling branch will be non-profiling/performance builds.

If you don't want to use the Steam branch, the files are also available for alternative download here:
https://drive.google.com/drive/folders/15p9j7C2nHUt6NoVfChX4YFuqzFXzblJh
Note: There are separate Dll files that also need to be placed into Game folder.

meowthis

whole cloud
empty goblet
whole cloud
#

What you're saying is irrelevant to what we were talking about

#

we are not talking about loading a cache file from disk like at all

#

I can make some self-modifying code so it dispatches to the correct functions, but antivirus doesn't tend to like that and thats stupid

silk pewter
#

not denying anything you said, but I wonder how a "simple" if bool can be expensive - is it called 1e9× / s ?

empty goblet
#

especially when it's done on early runtime init, once

autumn timber
#

What about using the [[likely]] (or likely(expr) for GCC) keywords in the if to make use of speculative execution and minimize the impact?

whole cloud
whole cloud
autumn timber
#

I'm not an expert in the field, but from my understanding, "if condition is true the execution is more efficient" is precisely done because of speculative execution done by the CPU (because of instruction pipelining, to be exact)

whole cloud
#

The usual optimization that does is keep the hot-path code close together. So that it is in instruction cache.
In likely case just execute in a straight line without any jumps. When unlikely happens, allow to jump away and potentially fail instruction cache

vivid rune
silk summit
woven loom
#

@silk summit profiling still has 32 bit

#

Also I'm assuming wetzer is running YAAB on a client and not the phenom II server 😛

whole cloud
silk summit
vivid rune
heavy vortex
#

32-bit is still very common with paid servers. So I hope it dies.

vivid rune
# heavy vortex 32-bit is still very common with paid servers. So I hope it dies.

https://en.wikipedia.org/wiki/Association_fallacy

Association fallacy

The association fallacy is a formal logical fallacy that asserts that properties of one thing must also be properties of another thing, if both things belong to the same group. For example, a fallacious arguer may claim that "bears are animals, and bears are dangerous; therefore your dog, which is also an animal, must be dangerous."
When it is a...

whole cloud
autumn timber
#

Too bad. That would be less work for extension makers 😦

vivid rune
#

No chips and a free day for Wetzer

autumn timber
#

And then 32bit Arma executables would stop being needed 😄

whole cloud
heavy vortex
#

cackles

whole cloud
#

I should just make new extension features only work on 64bit.
So that you as extension dev can say "We need that feature and bohemia doesn't support that on 32bit so we have no choice"

heavy vortex
#

ACE removing 32-bit support might even be enough.

autumn timber
vivid rune
heavy vortex
#

Server providers would probably balk at adding the "You can't run ACE" condition.

autumn timber
#

Seriously, if you report it and provide dumps and whatnot, then it's probably going to be fixed, and then you won't need 32bit anymore (and will be able to make use of 64bit address space and other 64bit goodness)

heavy vortex
#

Is this still the Phenom illegal instruction issue?

hybrid cradle
vivid rune
whole cloud
#

no

#

segfault means crashed.
Searching for "game crash" won't lead to usable results

vivid rune
#

"when you start the arma server without the complete chain of 32 bit PAM libraries present, it segfaults."
Is not related?!?

autumn timber
#

This doesn't sound related to any kind of "illegal instruction", at first glance, tbh

vivid rune
autumn timber
#

Unless you're told otherwise, you don't have to dump anything yourself. Just collect the crash dump that's created when the application crashes

vivid rune
autumn timber
#

Make sure that the host is set to create crashdumps on crash, as said in the article

vivid rune
woven loom
whole cloud
whole cloud
autumn timber
#

run gdb, get me offset inside file where the bad instruction is at, I might take a look
Yeah, but that's after you get the crashdump and to get the crashdump, which is tricky (because you're running inside docker in this case), you don't need gdb, that's what I meant.

I don't know about linux crashdumps, but if it were a windows .dmp, I'd open it in [Redacted] Pro, and it would show me the place where the exception ocurred. Not sure if that would work with linux dumps, though 🤔

ashen helm
#

game interface does not resize properly anymore in recent prof branch. I'm on 21:9, playing on small interafce by default, when changing to normal then even after restart most of it is hidden outside of the screen view. Setting ratio to stretch/non-stretch does not do anything

#

main menu and editor looks okay but the SL's team view and scrollwheel menu are obscured, everything else is displayed way too close to the screen border

#

screen within the vignette is what I see in normal UI size

image.png
whole cloud
whole cloud
#

Do you have custom FOV in your profile?

ashen helm
#

yes, yes, no

#

however on a new profile it seems to be all good so not sure what's changed

whole cloud
#

Started game on ultrawide very small. Set to normal. Restarted.
All seems good.

But. The group menu bottom right, and the action menu, can be repositioned in the gameplay options. Maybe its related to that

#

I tried messing with them too, but can't reproduce a problem

whole cloud
#

I switched my window resolution backto 16:9 and restarted the game a couple times.
Now I noticed this

Maybe it only happens when resolution is changed?

Arma3Retail_DX11_x64_Hfvk529N1c.png
ashen helm
#

no idea, I didn't move the window around, unless it's caused by the auto-stretch

#

(which does not change when you select a different screen ratio)

tribal pasture
#

windowed borderless to windowed too

violet elbow
heavy galleon
#

Already have to compile my extensions for Linux and windows, screw 32bit.

stark falcon
#

Got this one time when launching profiling, next launch it worked fine. But I'm not sure if it has anything to do with recent changes.

ErrorMessage: Fatal error: F:\SteamLibrary\steamapps\common\Arma 3\arma3_x64.exeDll\x64\GFSDK_SSAO_D3D11.win64.dll could not be loaded
Application terminated intentionally - Exception code: 0000DEAD  at C741CF19
vivid rune
#

A Buddy had in YAAB 108.2 FPS with 7950X3D/4070TI 💯
Interesting: with standard exe it is 107.2 fps 👀
(Mh. Something seems go wrong with std exe. We retry.)

woven loom
#

I was just testing actually

#

12% improvement with a 7800X3D and 4070 Ti

#

But it's a single run

#

I'll do more in depth testing later

#

YAAB is rather heavy in particles so it makes sense

#

Smokescreens would very noticeably impact frame rates on my old computer

woven loom
woven loom
#

With settings turned up to ultra, 6k object view distance, 12k overall view distance; i get ~67.5 with perf and ~62 fps with stable. so around 8% improvement

plain trout
woven loom
#

Yes

plain trout
#

Oh sorry I thought I was on another discord. I’ve been asking around and people see improvements after this change, more so on the non-highend machines.

vivid rune
#

12% on 4700HQ (24.7 FPS -> 27.7 FPS)

whole cloud
vale shoal
whole cloud
#

Ah I see it should cut before the "arma3_x64.exe" and then append the Dll...
It's not cutting it. 🤔
I'll look at that next week but I didn't touch it 😄

#

Well I guess we kinda touched literally everything

empty goblet
inland dew
woven loom
woven loom
#

Along with 1% and 0.1% lows

#

I'd really love to see who gets 155 fps for 0.1% lows lol

#

Avg was 110-112

#

Max was way higher

woven loom
#

Didn't look at frame times etc though

vale shoal
#

,

whole cloud
vale shoal
random isle
#

so PREF now needs vcomp140.dll ?

heavy vortex
#

Whatever's in the vc 2022 redist, I imagine.

#

I think it's supposed to auto-install when you switch branches but I haven't tried it.

random isle
random isle
whole cloud
whole cloud
#

2.16.151633 new PROFILING branch with PERFORMANCE binaries, v0, server and client, windows 64-bit, linux server 64-bit
- Sync with 2.16 main branch update
- Previous Profiling branch improvements from 2.14 are still present
- Removed: 32-bit version of Profiling branch, 32-bit binaries now match the stable branch

If you don't want to use the Steam branch, the files are also available for alternative download here:
https://drive.google.com/drive/folders/15p9j7C2nHUt6NoVfChX4YFuqzFXzblJh
Note: There are separate Dll files that also need to be placed into Game folder.
rf

whole cloud
fast hornet
#

did the "death by helicopter tail rotor"-desync fix make it into 2.16 stable btw.?

whole cloud
#

its not listed in the changelog

patent dome
#

Was that an issue in Stable 2.14? I thought it was introduced with a prof update.

restive pilot
#

iirc the desync was always an issue, which was "fixed" with a prof update that caused weird deaths without actual collision with tail rotors, so it was reverted

#

I don't remember if it was actually fixed in the end

fast hornet
#

would've been such a great fix sigh

heavy galleon
#

My game just crashed while testing our mod. Classic ACCESS_VIOLATION. Do you want dump? Made the full one too.
(did not happen before when testing btw)

whole cloud
#

always dump

fast hornet
feral harness
#

anyway 2.18 will be a blast, worth waiting

restive turtle
#

just looking at new commands and stuff

feral harness
#

it will basically contain fixes and updates for 2 versions, 2.16(that didnt make it into 2.16)+2.18

whole cloud
#

2.16.151648 new PROFILING branch with PERFORMANCE binaries, v1, server and client, windows 64-bit, linux server 64-bit
- Tweaked: Minor performance improvements for particles and helicopter simulation
- Tweaked: Upgraded Linux compiler (no behavior changes expected)
- Fixed: Custom radio markers did not sync for JIP players - https://feedback.bistudio.com/T180034

If you don't want to use the Steam branch, the files are also available for alternative download here:
https://drive.google.com/drive/folders/15p9j7C2nHUt6NoVfChX4YFuqzFXzblJh
Note: There are separate Dll files that also need to be placed into Game folder.
🐧

whole cloud
night mulch
#

and the instant disconnect with no explanation when joining listen servers

ashen helm
#

did you make a ticket for this

silk pewter
#

is it Arma-related though, and not Steam-?

whole cloud
#

There is a ticket for the stutter stuff. But no repro has been achieved by me

#

I'd like to hear from some linux server people if everything is still fine.
There should be no change in required libraries and such

night mulch
#

#dev_rc_branch message
This is the instant disconnect issue, if you could take a look again i would be very happy to help repro and provide data🙏

whole cloud
gritty wasp
#

Do we need dll from v0? Should it be in game root or dll folder?

patent dome
#

If are using the Perf/Prof branch on Steam the DLL is already in the correct place.
The DLL has been the same for a while, so if you already got it with the later 2.14 version you don't need to get it. It goes in the DLL folder in the same way as it is in the GDrive v0 folder.
You shouldn't need it for the server EXE's

night mulch
#

#dev_rc_branch message
you said you can make a custom exe to diagnose it, we moved to DM and you said And I'll need to figure out where I can set traps for client disconnecting. I can maybe do that tomorrow

#

but i guess it kinda got forgotten

whole cloud
#

Yeah.
I still have over a dozen unread DM's open as reminder to do things, some reaching back months. I don't have the time
Maybe yours is somewhere on my todo list

tribal pasture
#

@whole cloud Big thank you for including HandleDamage's context parameter with 2.16!

#

So many headaches solved for me!

whole cloud
#

It's Easter now, please don't break stuff 🫂
And don't find broken stuff Ded

Eggs don't count

hybrid cradle
#

wait its easter?

heavy galleon
#

But breaking stuff is what I do the most sadge

feral harness
heavy galleon
#

I already found something broken PU_PepeWave

analog acorn
ashen helm
feral harness
heavy galleon
blazing wagon
#

Is there a way to run the equivalent of diag_captureSlowFrame ["total", 0.2]; on the server rather than your client?

#

(preferably still with the dialog)

night mulch
#

Really hard to play with friends because of that bug

whole cloud
hushed seal
whole cloud
fickle geyser
#

It does not

whole cloud
restive turtle
silk pewter
full nova
#

but what happen after Dedmen Fixed ?

woven loom
#

the fixing will continue until morale improves

hybrid cradle
# ashen helm

can we get KK to put this on dedmens desk in a frame

bright latch
#

I know we were told not to break anything...

#

Getting this when going from lobby to in-game, anyone else noticed it

quaint flame
#

Been profiling a mission and every now and then I get these huge lags.

wFram; 40.34702; 403.33545;""
3dSwp; 40.35285; 403.32861;""

Possible to find out what these are?

#

total -> wSimu -> rendr -> wFrame -> 3dSwp

feral harness
blazing wagon
# silk pewter

gonna use this meme until TFAR exists in Reforger, thanks ❤️

full nova
#

the ingame voice chat in Reforger is already TFAR inspired

restive turtle
blazing wagon
unreal sparrow
#

It seems that latest perf branch build crashes the game when saving in helicopter.

#

Dunno if AFM has anything to do with it, but it was also active.

#

Mission in which it happened was the Reaction Forces training day scenario.

whole cloud
#

There was a crash with AFM on game save, but we fixed it?
Are you using the steam branch?

unreal sparrow
#

Used latest branch from Google drive.

whole cloud
#

then you need to also use the latest dll

#

if you use an old one it'd crash

unreal sparrow
#

Ah, I seem to have missed the .dll in the V00 folder, I only looked in the V01 folder.

#

Thanks, I'll try in a bit.

unreal sparrow
#

Hmmm, tried it with latest DLL from the 2.16 V00 folder, still crashes on mission save with exit code 0xC06D007F.

#

Can't upload .RPT file, but last entries were multiple instances of HELI (H245 SAR): Unknown action ListLeftVehicleDisplay, target HELI and HELI (H245 SAR): Unknown action ListRightVehicleDisplay, target HELI followed by one final instance of Serialization of Display variables is not implemented.

#

Both latest .DLL and performance .EXE and running 64 bit.

# 
Mods: A3;curator;kart;heli;mark;expansion;jets;argo;orange;tacops;tank;enoch;aow;RF
Version 2.16.151648
Fault time: 2024/04/02 03:43:22
Fault address:  CDF2CF19 01:0002BF19ll C:\WINDOWS\System32\KERNELBASE.dll
file:     M01_TrainingDay
world:    Malden
Prev. code bytes: 6A 95 08 00 48 8D 4C 24 20 48 FF 15 1F 81 1B 00
Fault code bytes: 0F 1F 44 00 00 48 8B 8C 24 C0 00 00 00 48 33 CC
#

Ran the mission using the basic flight model, same result.

spiral pond
#

@unreal sparrow you need to upload the mdmp to the A3 FT/DM it to Dedmen

unreal sparrow
#

Should I just send it under the normal Arma project or under Reaction Forces?

#

Or is there a perf branch project?

full nova
#

Only BI employee can open Mdmp files

tribal pasture
spiral pond
full nova
whole cloud
#

you can open (almost) everything with notepad

shell lion
#

Hi, i have a problem with the newest perf branch.

Error:
Fatal error ....\a3master\arma3server_x64Dll\x64\PhysX_64.dll could not be loaded

  • server instance a3master
  • start instance with cmd.file

On my client version everything works fine (i'm using it since 5-6 years).

Normaly i'm using it on my servers to, but since version 2.16.151633 and 2.16.151648 i can't used it on my first server instance.
I install it with steam cmd or manually with the google drive files (also the dll file), but every time the same issue is coming.
Also I have 3 server instance on my server since 5-6 years with no problems.
Only 1 (main) instance is runnig 24/7, the others are only for testing or events.
But on the others instance the perf branch is active with no problem.
Only on the main instance its not working. I'm looking into the problem since 1 week now and i don't know what to do.

#

If you need rpt, mdmp and bidmp file i can send it to you.

whole cloud
#

I had that issue once, couldn't see how/why it happened and also cannot reproduce it
Can you reproduce it?

#

Dumb fix would be, copy your Dll folder and name the copy arma3server_x64Dll

spiral pond
shell lion
unreal sparrow
#

I just tested it again to make sure it's not related to the RF helicopters and tested without DLC, I am uploading the logs to the feedback tracker right now.

whole cloud
#

huh. Profiling 2.14 v22 on google drive also has the wrong dll

whole cloud
unreal sparrow
#

Hmmmm, I downloaded that one from the V00 in the 2.16 folder.

#

As I said.

#

So I don't know why it would crash like that.

whole cloud
#

Because thats the wrong one, outdated

unreal sparrow
#

lmao

#

Ah, I missed the part that wasn't highlighted by the ping, my bad.

#

Thanks for solving it.

marble mason
#

When remote controlling (from Zeus or UAV from terminal) while in helicopter pilot seat, heli slowly rolls backwards onto tail until it explodes. Perf only and AFM only.

#

Tested with AH-9

random isle
marble mason
#

Video of what im talking about.
https://youtu.be/G3waDEnBui8

YouTube
Fire-39_1 ‍
Remote control bug

My Specs

Win 10 64bit
MOBO: ASUS TUF X570-PLUS
CPU: AMD 5950X
RAM: 32GB 3000 MHz DDR4
GPU: MSI GeForce GTX 1080 GAMING X 8G
Storage: 2X 1TB M.2 NVME Samsung 970 EVO
Headset: Corsair H2100
Keyboard: Corsair K95 RGB
Joystick: Logitech Extreme 3D Pro
Mouse: Alienware TactX Gaming Mouse
Settings: Ultra, 10000m draw distance
Gameplay recording soft...

▶ Play video
celest sparrow
#

never tried it from that side, interesting bug

random isle
ashen helm
#

with AFM enabled?

random isle
ashen helm
#

on arcade model I haven't seen any such issue

marble mason
#

yes only AFM

whole cloud
#

There is also age old issue where when using joystick input, going to menu would drop throttle to zero
maybe same thing?

marble mason
feral harness
marble mason
feral harness
#

probably always been like that no one touched this in years until recently

ashen helm
#

I firmly believe this wasn't the case before but I cannot change the audio output anymore, once game is started with specific device selected in Windows you cannot change it mid game. Same for the screen ratio settings, also get defaulted no matter what option you choose

image.png
whole cloud
#

I can atleast tell that if you play the game with -noSound, then changing audio output doesn't work

#

Seems to be fine for me.
I can atleast confirm that it goes away, and comes back switching with a channel that I cannot hear

ashen helm
#

I have nopauseaudio enabled and game running with reaction forces only. I'll see if getting rid of any of those makes difference

#

right now I can select whatever device I want and sound will be played through the one chosen in windows anyway

#

if I select my speakers from the dropdown and click OK, upon reopening the audio settings the output device selected is Default. If I select the headphones (and are default in windows), on reopening the menu output is as headphones, not default

ashen helm
heavy galleon
#

Oh yeah, engine power going to 0 when in pause menu has been happening for a long time.
I kinda assumed it was intended at this point

inland dew
#

has always been like this

woven loom
#

Yup, would be great if it didn't happen

#

Great way to crash if you want to quickly change a keybind or video option

marble mason
#

Or any GUI really, god forbid you want to type something in chat, or look at ctab or change a radio channel...

woven loom
#

ah, i don't fly in MP so didn't realise it happened while accessing chat as well

light cargo
autumn timber
random isle
light cargo
patent dome
#

Technically anyone can open the MDMP files in VS, but it's useless without the source code or debug symbols.

whole cloud
light cargo
#

i cant seem to limit the ram used by the game

woven loom
#

i feel like you've provided far too little info to even guess at what's happening

#

vanilla or mods? over how much time are you seeing the behaviour? do you have graphs? what is the outcome of this supposed leak? does your game crash?

azure jewel
#

a little bug:
BWMod -> MELLS launcher: missile will explode in air, it should explode after hit something
addons:
cba_a3
ace(required for manual missile control)
BWMod

silk pewter
azure jewel
# silk pewter is it perf-prof related or just a mod issue?

maybe its a prof branch bug:
this bug happens in prof branch(2.16.151648) while working normally in steam default branch(2.16.151618).

moreover, in the 2.14, theres no bug in v16(2.14.151459), the bug appeared since v17(2.14.151512).
by the way, v17 use a RTDynamics_64.dll file

random isle
#

Hmm just had server crash on startup said something about physics didn’t get the crash log 😦

light cargo
whole cloud
random isle
random isle
#

off topic a bit, is there something like YAAB that can run in headless mode say on server ? Building new server for our community and trying to fine tune settings.

light cargo
woven loom
#

I've seen page file usage with 16GB and chrome open in the background with dozens of tabs

#

not seen arma noticeably use page file with 32GB of RAM

light cargo
#

also it doesnt help that my motherboard is cursed

#

it works for everyday operations but it has memory errors and when the cpu usage is too high it may become unresponsive

woven loom
#

well that seems like a hardware issue 😅

light cargo
#

motherboard issue

woven loom
#

memory errors = replace RAM

light cargo
#

only in memtest i had memory errors

#

there doesnt seem to be any errors in general use

#

but i guess i might as well be using dirty ram without knowing it

woven loom
#

yes that means the RAM has issues and can lead to issues like freezing and becoming unresponsive

light cargo
#

it been almost 10 years

woven loom
#

definitely would suggest replacing the faulty RAM sticks

light cargo
#

i suppose ddr3 cl9 isnt very compatible with this motherboard or something

#

anyways, i want to limit memory usage for arma 3 and i seem unable to do so in this branch or im not seeing the effect

woven loom
#

motherboards usually try and run the most stable configuration. if it's running JDEC specs and still throwing errors then you have bad sticks

light cargo
#

xmp makes it unstable

#

anyways sorry for the offtopic

woven loom
light cargo
#

i still would like to limit arma 3 memory usage to somewhat alliveate part of the issue

cloud nacelle
light cargo
heavy vortex
#

That's working as intended since 2.16, I think. You can set -maxmem=4096 to approximate the previous behaviour. Or 5-6GB for a bit of margin.

light cargo
#

what was the issue?

woven loom
#

Issue was that arma wasn't using enough RAM lol

cunning acorn
#

does this code for profile brenching still work? CautionSpecialProfilingAndTestingBranchArma3

worldly badge
#

No. Profiling doesn't require pass anymore

whole cloud
#

2.16.151697 new PROFILING branch with PERFORMANCE binaries, v2, server and client, windows 64-bit, linux server 64-bit
- Tweaked: Added a cache of recent Regex patterns for regex* script commands
- Fixed: Heal actions did not appear when touching water (now they only need to have the upper body out of water) - https://feedback.bistudio.com/T178578
- Fixed: "soundFly" of shots was not simulating the speed of sound - https://feedback.bistudio.com/T178099
- Fixed: Potential error at game start complaining about a missing DLL
- Fixed: Feature objects, that are under the mouse cursor while an Eden Editor world is cleaned up, could get stuck as ghost objects - https://feedback.bistudio.com/T174009
- Fixed: Script errors in params would not print context to RPT (thanks @boreal wigeon)
- Fixed: Script errors in params would not trigger the ScriptError Event Handler
- Fixed: Game freeze and later crash when pressing Alt+F4 with an inventory display open
- Fixed: Possible crash on game exit (introduced with 2.14)
- Fixed: Editor mission could get stuck as "Multiplayer" even with only one playable unit (thanks @patent wraith)
- Fixed: setUnitLoadout did not trigger the SlotItemChanged Event Handler as expected - https://feedback.bistudio.com/T180338
- Fixed: Mission Event Handlers for markers were not triggered locally with Zeus curator changes
- Fixed: Zeus edited marker position and direction did not broadcast to remote clients

If you don't want to use the Steam branch, the files are also available for alternative download here:
https://drive.google.com/drive/folders/15p9j7C2nHUt6NoVfChX4YFuqzFXzblJh
Note: There are separate Dll files that also need to be placed into Game folder.
morphine

whole cloud
heavy vortex
# light cargo what was the issue?

Well, 4GB isn't always enough. Depends on detail settings, map, mod usage. Trouble is that Arma doesn't have much understanding of how much RAM it needs. It keeps allocating memory until it hits the defined limit, and only then does it look for things to free.

whole cloud
#

You mean that? I aimed the missile into the ground, but the explosion particle effect is high up in the air?

arma3_x64_2024-04-11_11-47-18.png
azure jewel
whole cloud
#

Mh weird. I just got it to fly over 15 seconds in VR.
On Stratis its indeed about 10.

But also on Stratis, if I aim the missile into the ground immediately after it launches, it "hits" about 20 meters early and explodes in the air above the ground. interesting

#

Huh.. On my internal build everything is fine though 😕

whole cloud
#

bwa3_mells\functions\handleCameraAiming.sqf
Line 32 calls setDamage to trigger the missile, even while its still in the air

private _dirCam = vectorDir bwa3_mells_camera;
private _dirMissile = vectorDir bwa3_mells_currentMissile;
private _difference = acos ((_dirCam vectorDotProduct _dirMissile) / ((vectorMagnitude _dircam) * (vectorMagnitude _dirMissile)));
if (_difference > 130) then {private _charge = "DemoCharge_Remote_Ammo_Scripted" createVehicle (getPos bwa3_mells_currentMissile); deleteVehicle bwa3_mells_currentMissile; _charge setDamage 1;};

If the difference between camera aim direction and missile fly direction is too high?

# 
_dirMissile = [0.255782813, 0.949058771, -0.184018314]
_dirCam = [0.255783439, 0.949061632, -0.184002817]
_difference = -nan(ind);

mh.

#

exact same code in my local internal build.

What the frick

Arma3NoOptInt_DX11_x64_2024-04-11_13-11-22.png
silk pewter
#

what did you brokeded

whole cloud
#

(_dirCam vectorDotProduct _dirMissile) is 1
acos (_dirCam vectorDotProduct _dirMissile) is -nan(ind)
Thus
acos 1 is... 0 ? ??? ???? ? Wat

#

Aha, rounding! toFixed to the rescue

#

acos is only valid with inputs between -1 and 1
acos 1.00000011920928955078 is greater than 1, and returns nan

# 
_dirMissile = [0.255782813, 0.949058771, -0.184018314]; 
_dirCam = [0.255783439, 0.949061632, -0.184002817];

((_dirCam vectorDotProduct _dirMissile) / ((vectorMagnitude _dircam) * (vectorMagnitude _dirMissile))) toFixed 20

Internal build: "0.99999988079071044922"
Profiling branch: "1.00000011920928955078"

How is that possible, its the same code

silk pewter
whole cloud
analog acorn
silk pewter
#

fixed

#

(tell me if aTan, aSin etc too)

whole cloud
silk pewter
#

(thanks)

whole cloud
#

(_dirCam vectorDotProduct _dirMissile) toFixed 20
Internal: "1.00000011920928955078"
Prof: "1.00000011920928955078"
2.12 profiling: "1.00000011920928960000"

(((vectorMagnitude _dircam) * (vectorMagnitude _dirMissile))) toFixed 20
Internal: "1.00000023841857910156"
Prof: "1.00000000000000000000"
2.12 profiling: "1.00000023841857910000"

(vectorMagnitude _dircam) toFixed 20
Internal: "1.00000011920928955078"
Prof: "1.00000000000000000000"
2.12 profiling: "1.00000011920928960000"

((_dirCam vectorDotProduct _dirMissile) / ((vectorMagnitude _dircam) * (vectorMagnitude _dirMissile))) toFixed 20
Internal: "0.99999988079071044922"
Profiling: "1.00000011920928955078"
2.12 profiling: "0.99999988079071045000"

acos ((_dirCam vectorDotProduct _dirMissile) / ((vectorMagnitude _dircam) * (vectorMagnitude _dirMissile)));
Internal: 0.0279765
Profiling: -nan(ind)
2.12 profiling: 0.0279765

🤔
vectorMagnitude supposed to be 1, vectorDir of a entity should always be normalized.

silk pewter
#

x32 vs x64? ^^'

whole cloud
#

both 64

#

In 2.12 acos already behaved like that

arma3_perfv7_x64_2024-04-11_13-26-56.png
whole cloud
#

(vectorMagnitude [0.255783439, 0.949061632, -0.184002817]) toFixed 20
2.16 Profiling from steam: "1.00000000000000000000"
2.12 Profiling from steam: "1.00000011920928960000"
My local profiling build: "1.00000000000000000000"
My local internal debug build: "1.00000011920928955078"
My local internal release build: "1.00000000000000000000" 😠
Visual Studio you 🤡

restive pilot
#

for scaled objects it's not

whole cloud
#

They are not scaled

#

And when this code was written, scaling objects wasn't possible

restive pilot
#

the game itself could scale them tho

#

anyway I know what you mean 😅

whole cloud
#

It really reallys shouldn't

restive pilot
#

maybe it does the math differently due to ordering

#

(iirc Arma uses fast)

#

yeah, fast

heavy vortex
#

I don't think it's even considered incorrect to reorder a+b+c

#

well, maybe

restive pilot
#

with floats some changes do make a difference

heavy vortex
#

I guess every time you use acos you just have to fence the damned thing.

#

I'm not sure that even precise would guarantee vectorMagnitude vectorNormalized _vec <= 1

whole cloud
#

Release build:
Our code: return sqSize * (1/sqrt(sqSize))
The generated assembly code: return sqrt(sqSize)
sqSize = 1.00000012
sqrt(sqSize) = 1.00000000 (thats wrong. Should be 1.00000006)
Enf result "1.00000000000000000000"

Debug build:
Our code: return sqSize * (1/sqrt(sqSize))
The generated assembly code: return sqSize * (1/sqrt(sqSize))
sqSize = 1.00000012
sqrt(sqSize) = 1.00000000 (thats wrong. Should be 1.00000006)
1/sqrt(sqSize) = 1.00000000
sqSize * (1/sqrt(sqSize)) = 1.00000012
End result "1.00000011920928955078"

#

Optimizer is trying to be smart

#

I don't understand why our code is so weird.
magnitude is supposed to be sqrt(sqSize). For some reason we do sqSize*(1/sqrt(sqSize)).
Result is generally the same, if the sqrt is calculated correctly.

If the sqrt is calculated wrong for close to 1.0, then that saves our butt. Maybe the code is there specifically for that messup

But MSVC decided to optimize out our butt saving. What journey...

heavy vortex
#

I guess if sqrt is consistently off when close to 1 then it makes some sense. Never seen that done before though.

gritty wasp
whole cloud
#

its not undefined, yeah we updated compiler

#

I followed that code back till july 1999 where
Coord Size() const {return sqrt(X()*X()+Y()*Y()+Z()*Z());}
was changed to

Coord Size() const
{
    //return sqrt(X()*X()+Y()*Y()+Z()*Z());}
    // optimization: |v|=|v|*|v|/|v|
    float size2=SquareSizeInline();
    Coord invSize=InvSqrt(size2);
    return size2*invSize;
}

The comment was "Improved attacks."
I guess I'm not getting any info out of this for why
But clearly this unexplained change from 1999, is now essential

heavy vortex
#

what exactly does InvSqrt() do there?

whole cloud
#

1/sqrt(x)

heavy vortex
#

you have a function for that? :P

whole cloud
#

Probably because it was once in the past fastInvSqrt

heavy vortex
#

right

#

And now compiler is probably doing rsqrtss + NR anyway

whole cloud
#

it optimizes
x * (1/sqrt(x))
->
sqrt(x)
and uses sqrtss

heavy vortex
#

Curious. sqrtss is supposed to be bit-accurate as far as I can tell.

#

ah wait, that's bit-accurate.

whole cloud
#

Ah yes
1.00000000 (thats wrong. Should be 1.00000006)
Correct.
But.
3f800001 is 1.00000012

1.00000006 is between 3f800000 and 3f800001
it cannot return the correct value, it would be half a bit

silk pewter
#

then fix it!

whole cloud
#

Should I even bother reporting this to microsoft 😓
It ends up with one bit being rounded down instead of up

heavy vortex
#

Pretty sure that's an acceptable result for fast

whole cloud
#

the sqrt yes

#

But we know that sqrt fails in fpfast. And thats why we use that workaround (I assume)
and that workaround worked fine till VS2017
That workaround would still work, if it wouldn't wrongly optimize it away

in fpprecise it can optimize it away, because sqrt is correct there and it doesn't matter

heavy vortex
#

hmm. Surely sqrt(1.00000012) is still 1 in fp:precise?

#

It's just not allowed to do the x/sqrt(x) optimization.

silk pewter
#

I believe you can use #pragma, or unsafe code, to prevent such optimisation?

heavy vortex
#

Not sure how deep this goes though, or whether it covers all the cases. Feels like the fundamental flaw is that acos is single-bit reliant and you shouldn't use it.

whole cloud
#

Ah yeah it turns the
x*1/sqrt into a x/sqrt

restive turtle
#

So smort meowheart

whole cloud
# heavy vortex Might work on MSVC at least: https://learn.microsoft.com/en-us/cpp/preprocessor/... 
#pragma float_control(precise, on)
float D(float x)
{
    return x * 1/std::sqrt(x);
}
#pragma float_control(precise, off)

That works but it turns the sqrt from sqrtss instruction, into a sqrt function call... ugh.
Its weird that it does that, with /fp:precise compiler flag it still uses sqrtss.

float D(float x)
{
    return _mm_div_ss(_mm_set_ps1(x), _mm_sqrt_ss(_mm_set_ps1(x))).m128_f32[0];
    //return x * 1/std::sqrt(x);
}

This works as desired.. I hate it 🤣

heavy vortex
#

Does it work in GCC? :P

#

(turns out five others things were compiler-broken in GCC already)

analog acorn
#

Is there not like a "I know. This is like this on purpose. Don't optimise it." flag to stop it touching that bit of code?

whole cloud
#

GCC we don't even set fast math
It goes the whole way

D(float):
        movaps  xmm2, xmm0
        rsqrtss xmm2, xmm2
        movaps  xmm1, xmm2
        mulss   xmm1, xmm0
        mulss   xmm1, xmm2
        mulss   xmm2, DWORD PTR .LC1[rip]
        addss   xmm1, DWORD PTR .LC0[rip]
        mulss   xmm1, xmm2
        mulss   xmm0, xmm1
        ret

I'm just not gonna ask how
sqrt, div, mul

turns into
rsqrt, mul, mul, mul, add, mul, mul

heavy vortex
#

That's the NR step

whole cloud
#

Ah Newton-Raphson

#

ye

#

Well GCC no issue so don't need to care ✅

heavy vortex
#

lol

whole cloud
#

I love these bugs.
A compiler optimization caused one least significant bit of a float to be flipped. Causing missiles to explode mid-air.

All fixed ✅ Next week prof

Arma3Retail_DX11_x64_2024-04-11_16-25-23.png
light cargo
#

'99 is what, ofp?

random isle
#

ofp

light cargo
#

yeah

random isle
#

same code still yep

pallid pebble
fickle geyser
#

but solving stuff like this is fun

vivid rune
#

Does AR also benefit from that finding?

restive turtle
woven loom
vivid rune
woven loom
#

yeah but the workaround that broke might not be relevant

heavy vortex
#

You really shouldn't use acos for that in Reforger because it'll be much slower than comparing the dot product result directly.

woven loom
#

(graphs from intel presentmon 2.0)

night mulch
#

Check if you have the steam infinite download loop bug

ashen helm
#

Alternatively, unsubscribe from all compositions and scenarios

woven loom
#

no steam isn't downloading anything, only had WS and RF loaded.

#

this isn't noticeable to the naked eye normally, i'm just curious what's happening

quaint flame
#

Use diag exe and captureSlowFrame

plain trout
#

If I look at steam when that happens I see how arma is trying to redownload every mod I have installed only to „realize“ it’s already properly installed and doesn’t need downloading. I’m subbed to 400ish mods and it’s always a joy. 4090 and 7950x3D made the situation better, but it is still recognizable.

light cargo
#

i got a crash on exit

light cargo
#

sent dump to dedmen already fyi

woven loom
#

Looks like wSimu is the culprit?

#

(Completely empty map, capture triggered when frame time falls below 0.016s, i.e. 60fps)

heavy vortex
#

wSimu has pretty much everything in it.

woven loom
#

Well i don't have the ability to diagnose this further...

heavy vortex
#

wsSet is interesting maybe

#

Compare against a normal frame.

woven loom
#

normal frame (~110 fps)

heavy vortex
#

Expand wSimu like in the other shot.

woven loom
heavy vortex
#

Yeah, most of the difference there is wsSet, although rendr time is a bit extended too.

woven loom
#

Yeah...

heavy vortex
#

If you hit "copy all" on a slow frame then it'll give you the whole snapshot as portable text that other people can look at.

woven loom
#

Awesome, will just do that and post! Thanks for the help!

heavy vortex
#

I need to sleep but I'll have a poke at it tomorrow.

#

I can only guess like 1/5 of the categories, but hey :P

woven loom
#

I appreciate it nonetheless 😄

#

This time, dlcSim and steamCbk also seemed to spike

heavy vortex
#

That one might be a relatively rare spike. Or the other way around.

#

Gotta figure which is the more common one.

woven loom
spiral pond
#

imo the main problem is

  1. data analysis - one off events can happen; unless very extreme impact (ie to cause mini freeze or worse), you rather want to know what has frequently/continuously a relevant impact. for that you need statistics and visualization

  2. lack of data export/pipe to even get that data - the diag command to log that is broken in A3

heavy vortex
#

So which was common? The dlcSim spike or the wsSet spike?

woven loom
#

Not sure how to even get that info, much less tell you the result

#

The wsSet spike happens even when the dlcSim and steamCbk spikes happen, so I'd assume wsSet is the regular one. But I don't even fully understand what the numbers mean or how they add or interact with each other.

#

But beyond that not sure what to do. I could capture 1200 frames and send it to dedmen/KK I guess, or make a feedback ticket with the captures attached (repro is easy, just sit on a blank map, I've seen it with both the WS map and Altis)

#

But it would be good to know if this is even something worth doing or if I'm wasting time

#

And that's if the capture to file works (kju suggests it's broken?)

#

But like I said on my system in this controlled test the effect is mostly not noticeable. However I'd assume on slower systems it will contribute to a stuttery experience, or even on fast systems it could stack with other sources of microstutters to become noticeable. Arma does a reasonable job of frame pacing so it manages these drops better than other games imo.

spiral pond
#

@woven loom do you execute this from console while simulation is halted?

#

if so, you need to use spawn with some sleep and close the console, so it catches the full simulation

#

otherwise do in MP (local host) where console doesnt stop simulation

#

it seems to me you also misread the info

#

the first value is a "timestamp" - the second is the "duration"

#

you can use right arrow key to automatically open the longest frame level by level

whole cloud
#

The captures are not detailed enough for precise analysis. It tells roughly where the issue is, but it's still alot of code.

You could do precise profiling with Intel v-tune. I can then analyze the files it recorded.
I'll need a game crash dump from same PC together with that though

woven loom
#

Hmm well there's no crash

woven loom
woven loom
woven loom
#

Okay no it looks beyond me lol. I'll just try and get better frame captures through Arma and report back.

spiral pond
whole cloud
#

Maybe I should make a tutorial on vtune

pallid pebble
woven loom
#

That would be super helpful!

woven loom
#

Ugh i did more testing but forgot that splendid camera pauses simulation
it actually doesn't if you start in splendid cam (i.e. if there is no player unit) in SP.

silk pewter
sour ridge
woven loom
#

Well there's a feedback ticket for this now
Periodic microstuttering observed with empty maps
https://feedback.bistudio.com/T180665

Documented stuff as best as I could with relevant screenshots and frame capture output in text form as well

woven loom
#

I have another weird performance issue, but this is a proper slowdown. While in 3DEN, Looking at the horizon drops the frame rate to ~15 fps.
I haven't tested extensively, but so far it happens:

  • only in eden
  • only when there are a several dozen units and objects (i don't know exact thresholds)
  • with the WS map but seems not to happen on Altis (tested with other scenarios)
  • on both stable and profiling branch
  • only when looking at the horizon in a particular angle
  • scales non-linearly with overall view distance: with 2km vd i get ~300 fps, with 3km it's already 140-ish, at 5km it's around 60, 12km it's 15.
#

I have a repro mission in case it's of interest

#

frame capture:

#

culprit seems to be wSimu -> rendr -> hudDr -> 3DEN_UI_onDraw -> 3DEN_PrepareDraw

quaint flame
#

3D icons maybe?

woven loom
#

Possibly... i see the regression even when not looking towards any of the units/groups... and then doesn't quite explain why the performance keeps dropping with increasing view distance, because i don't have things beyond maybe 5km from the camera (might be even less really but i didn't measure).

It also happens regardless of how far i am from units, like even if i'm super close to the ground and to some units. As long as i look towards the horizon it happens. If the (roughly) centre of the camera looks away from the horizon, performance goes back up.

#

Forgot to mention, i see this behaviour even in the loadout editor (virtual arsenal)

#

(i.e. click on unit on the map -> edit loadout)

whole cloud
#

Oki but who cares about performance in Eden as long as it's above 20fps meowtrash

worldly badge
#

3D icons are very expensive as far as I can see

gritty wasp
woven loom
#

☝️

whole cloud
woven loom
#

It's also that the wild swings in FPS makes it feel very stuttery. If it was dipping down to like 40 or 50 i wouldn't even notice

whole cloud
#

20fps.

Arma3Retail_DX11_x64_4qRgAW7ub2.png
#

20° to the left, 150fps.

I have hundreds of objects placed, but they are all behind me

Arma3Retail_DX11_x64_up16AJ9TsV.png
silk pewter
#

I suppose a failed culling?

whole cloud
#

It is in UI_onDraw, but not inside PrepareDraw for me.

Its drawing lines for objects, outside of terrain bounds, so it triggers terrain synth (Which vanilla terrains don't have)

#

Weird thing, there isn't any object outside of terrain bounds, and I also don't see any lines on screen.
Even when directly looking at hundreds of objects infront of me, I get 90fps, not barely 20

#

3000 signs are fine.
1000 building ruins are baaaad.

Dangit.
Pressing backspace to hide the UI makes the issue go away, and it doesn't come back. What is this 🤣

#

Having the 1000 buildings selected, behind you out of view. even goes down to 1.5fps.
Something is wrong with the check if something is in view indeed

#

No! Its again drawing lines.

The line that goes from object down onto terrain.
For some reason its drawing it in wrong place, when its behind camera.

My signs were fine because they were on the ground.
There is a minimum distance of 0.00001 meters for the lines. (lol why isn't it like.. 10cm, that should be enough)
So objects that are on ground, don't draw

#

omfg :/

For every line that is drawn. The camera calculates screenToWorld [0.5, 0.5]
So its does a lineIntersect from camera, in camera view direction, till it hits terrain (based on view distance).

When your camera aims far outside of terrain, boom.
In my 1.5fps case, that is a 9km lineIntersect, outside of terrain so it has to dynamically generate terrain data. For every line drawn under the 1000 buildings BEHIND my camera that I cannot even see.

#

After it has done this magnificent calculation.
It decides "Hey, I cannot see the beginning or end or middle of the line anyway. Lets return zero"

silk pewter
#

sooo… yeah 😄

#

"pls fix"? :D

whole cloud
#

This code is weird.

Its trying to check if the line passes through the camera.
To do that it does the very expensive lineIntersect to find a point on terrain that it hits.

It uses that point behind the line (if it passes infront of camera), to project it back onto the line.
That should (if the line is infront of camera) find a point on the line, that is visible.
Then it does a clipping check on that point.

But as far as I can see, where the actual ground is, is completely irrelevant. It just needs to be far enough infront of the camera.
It could be a ultra expensive check to find at which distance terrain is, or I can just hardcode 16km infront (turn 9+km lineIntersect, into a simple multiplication), to get the same clipping result.

None of this has comments for why it works the way that it works 😠

The code also looks completely messed up.
At first it calculates distanceSqr for the line length for fast comparison.
But when it decides to use that distance and calculate it back into actual meters, it uses InvSqrt, instead of Sqrt, and then uses that as a distance in meters. That's bullsh. You don't get the square root by using InvSqrt, this doesn't even work.

You can even see how that doesn't work. You should be able to see the line passing infront of your camera

quaint flame
#

Now I wonder where else in the game stuff like this happensmeowsweats

#

Fix it and sell it as FPS DLC

whole cloud
#

This specific function is only used for 3den line drawing

quaint flame
#

Yeah I guessed so. Was just generally wondering.

whole cloud
#

🙄
Lets do a line/plane intersection to find the proper midpoint.
Oh we have two functions with the same name and same arguments, in classes where the name only differs by one letter.
Oh its a copy paste with just one letter changed, one is in Eden code, the other in Zeus code.
The copy paste even includes all the same comments.

violet elbow
#

so two-for-one fix? 😄

vivid rune
#

Maybe it was separated because Zeus was/is a separate DLC?

whole cloud
#

no

heavy vortex
whole cloud
#

Now I can see the original code tried to do a plane/line intersection.
But their plane was rotated wrongly by 90°.
And if it failed, it just tried to use the line length to calculate the endpoint (which is nonsense as the endpoint was already stored in a variable), and then also failed doing that by using InvSqrt instead of Sqrt.

The terrain intersection point isn't needed at all. Nor is the camera view direction actually.

This code is so messed up.
Lines render correctly when you don't have the unit selected, because then it is a 3D line, which doesn't run any of this code.
But if you select the line, it becomes a 2D line, which runs this code..

The difference is 3D line is projected onto screen by the renderer.
2D line is projected onto screen by our bad code above, and given to the renderer as on-screen coordinates.

End result, the same line but the 2D line is more pixelated with darker looking edges.

With our bad code that means line is gone, you un-select the unit, and line appears 🤯

#

lol in the video you can't even see the difference

#

This is supposed to be roughly centered on the line infront of the camera. So much fun again
I'm spending more time researching how this even works and why its so broken, than I spend fixing it..

silk pewter
#

classic "how is it this much broken?" moment

pallid pebble
random isle
#

i just hope it will uncover something bigger 😄

whole cloud
random isle
whole cloud
#

Lots of reading and experimenting. There's like 7 stackoverflow tabs open

whole cloud
#

Old broken stuff

Vector3 screenCenterOnGround = IntersectWithTerrain(viewOrigin, viewDirection);
float onLineDistance = (screenCenterOnGround - lineBegin).DotProduct(lineDirection);
float lineLength2 = lineBegin.Distance2(lineEnd);
if (Square(onLineDistance) > lineLength2)
  onLineDistance = InvSqrt(lineLength2);
Vector3 pointOnLine = lineBegin + lineDirection * onLineDistance;

New stuff

Vector3 GetLineRayMidpoint(Vector3 lineBegin, Vector3 lineDirection, Vector3 viewOrigin, Vector3 viewDirection)
{
  // Check if the line passes through camera viewport
  // We create a plane on the line, facing the camera
  // We trace from camera, along camera direction, onto the plane. We now have a point that is perpendicular to the line, in the line plane
  // The project that point down onto the line
  // We have a midpoint on the line centered infront of camera

  // Point on line, to which camera would be perpendicular
  auto pointOnInfiniteLine = lineBegin + (viewOrigin - lineBegin).DotProduct(lineDirection) * lineDirection;
  auto directionFromLineToView = (viewOrigin - pointOnInfiniteLine).Normalized();
  // We could check if this conflicts with camera direction, thus is behind camera
  // But I'll save that step and directly do the plane projection

  // Find a point perpendicular to the line (on the plane) in our camera direction
  auto cameraOnLine = MathPlane::FindIntersectionWithPlane(directionFromLineToView, lineBegin, viewOrigin, viewDirection);

  // We now have a point on the plane of the line, project it down onto the line itself
  auto distanceToPointOnLine = (cameraOnLine - lineBegin).DotProduct(lineDirection);
  auto pointOnLine = lineBegin + distanceToPointOnLine * lineDirection;

  return pointOnLine;
}

Maybe some maths people can find a way to optimize that. Probably is some way.

random isle
#

just for shits and gigles here is chatgpt 4.0 optimization ask

Vector3 GetLineRayMidpoint(const Vector3& lineBegin, const Vector3& lineDirection, const Vector3& viewOrigin, const Vector3& viewDirection)
{
    // Calculate the projection of viewOrigin onto the line defined by lineBegin and lineDirection
    Vector3 lineToView = viewOrigin - lineBegin;
    float projectionScale = lineToView.DotProduct(lineDirection);

    // Calculate the point on the line corresponding to this projection
    Vector3 pointOnLine = lineBegin + projectionScale * lineDirection;

    // Normalize lineDirection for accurate projection calculations
    Vector3 lineDirectionNorm = lineDirection.Normalized();

    // Calculate intersection of view ray with the plane containing the line and orthogonal to viewDirection
    // This requires the plane normal, which we assume to be lineDirectionNorm (line must be normalized for this context)
    Vector3 planeNormal = lineDirectionNorm;  // Assuming the plane's normal is along the line direction, change if the context differs
    float denom = viewDirection.DotProduct(planeNormal);

    if (abs(denom) > 1e-6) {  // Avoid division by zero and ensure the view direction is not parallel to the plane
        float num = (lineBegin - viewOrigin).DotProduct(planeNormal);
        float t = num / denom;
        Vector3 intersectionPoint = viewOrigin + t * viewDirection;

        // Project intersectionPoint onto the line
        float distanceToPointOnLine = (intersectionPoint - lineBegin).DotProduct(lineDirectionNorm);
        pointOnLine = lineBegin + distanceToPointOnLine * lineDirectionNorm;
    }

    return pointOnLine;
}
whole cloud
#

Boi I imagine this would trigger some peoples OCD.
But I already fixed one bug in this function, I'm not gonna fix the whatever this offset is too.
The red block is my code, and its correct, goodbye 😄

whole cloud
#

prof with that tomorrow, please don't break it Ded

#

That code should actually be "find closest point between two rays"
Or
"Find closest point on ray A, to ray B going past it"
Or
"Find point on ray A, that's closest to ray B"
I'm sure such a thing must exist somewhere, in a better way. But that's a bit hard to Google for

#

Guess I'm redoing that tomorrow

light cargo
#

damn, what a hell of issues you're finding lately

#

maybe we can finally fix arma 3 bad optimization??

inland dew
#

too bad Dedmen wasn't born much earlier

light cargo
#

what

#

lol

whole cloud
#

If I were I wouldn't be doing this now

light cargo
#

hmm

#

ah i see, the optimization problem plagued the whole series

silk pewter
#

he spotted a bool _runWithoutBugs = false; too, let's see what comes out of it

light cargo
#

wat

woven loom
#

lmao what have i unleashed meowsweats

#

Thank you dedmen! And i'm glad you had some fun too 😄

#

and not just pain looking at the old code lol

#

i didn't quite get why it needs to do whatever it's doing though -- is it for projecting UI elements? (like what's the final goal of doing line intersect etc)

light cargo
#

sorry for the offtopic but this issue reminds me of a bug Control has where you can get tooltips if the object that produces the tooltip is behind the camera 180 degrees

kindred radish
#

Now you could please fix the low FPS problem in "high" density cities and we don't need Arma 4 blobcloseenjoy

whole cloud
woven loom
#

I see 🤔 thanks!

plain trout
#

(Is there a chance that the shadows settings not disabling shadow calculations will be fixed?)

random isle
#

ya i noticed as when turning off shadows gives you like extra 10 fames

empty goblet
plain trout
#

You also need to reduce down drawing distance as only that disables shadow calculations. It’s in the bug reports and I use this opportunity to poke that a change of the behavior gets added with the next patch

pallid pebble
#

Looks like you've got your work cut out for you dedmen blobcloseenjoy hmmyes

whole cloud
#

2.16.151758 new PROFILING branch with PERFORMANCE binaries, v3, server and client, windows 64-bit, linux server 64-bit
- Fixed: Advanced Flight Model helicopter simulation did not consider water buoyancy - https://feedback.bistudio.com/T180426
- Fixed: Float rounding error that could cause missile to explode mid-air (thanks @azure jewel)
- Fixed: Scripted min/max behaved incorrectly when provided with a NaN value (NaN is allowed on the left side)
- Fixed: createMarker would ignore the drawPriority of other markers and incorrectly draw on top of higher priority markers - https://feedback.bistudio.com/T180604
- Fixed: Performance issue in Eden Editor when looking outside of the terrain
- Fixed: Eden Editor lines that go through the camera (but don't start/end in the camera view) would randomly not show
- Changed: Scripted division by zero now returns infinity (previously 0, still triggers a script error)

If you don't want to use the Steam branch, the files are also available for alternative download here:
https://drive.google.com/drive/folders/15p9j7C2nHUt6NoVfChX4YFuqzFXzblJh
Note: There are separate Dll files that also need to be placed into Game folder.
nootnoot

whole cloud
azure jewel
light cargo
#

gg!

#

cheers for the release

whole cloud
#

Today I found out.
diag_codePerformance was doing a thing badly.
{objNull} in diag_codePerformance, would run 10% slower than the same code would run inside real script.

So next prof update, all scripts will look faster when profiling them.
I'm also making changes to make scripts run faster though, so I guess these will be hidden/made look more extreme by this fix.

#

_x = []; _x resize 2000; _x apply {true}; 0.0547725 ms // 2.14 (But I think that is already with SimpleVM)
_x = []; _x resize 2000; _x apply {true}; 0.0205002 ms // 2.18
_x = []; _x resize 2000; _x apply {true}; 0.0171497 ms // 2.18 internal

We're making good progress

autumn timber
#

So next prof update, all scripts will look faster when profiling them.
Now that's the "Performance DLC" that I have been waiting for!

silk pewter
#

"just show '60 FPS' top right of the screen, they won't notice"

whole cloud
#

Previously when running a nular command, only about 40% of the execution time would be the command itself, the rest is overhead of other things done around it.
Next prof update will have that at 80%. So way less overhead on the side of commands.
That should make everything a tiny bit faster.

#

More benchmarks to see how far we've come

I'd say the efforts of improving script performance are paying off.
_x = []; _x resize 2000; _x apply {true};
2.16 stable: 0.374646 ms
2.16 perf: 0.0174524 ms
2.18 trunk: 0.0172388 ms

_x = []; _x resize 2000; _x apply {objNull};
2.16 stable: 0.413851 ms
2.16 perf: 0.0595997 ms
2.18 trunk: 0.0577047 ms

toFixed 8; diag_codePerformance [{isAutotest}, nil, 1e+6, true]; (just returns false)
2.16 stable: 0.00015126 ms
2.16 perf: 0.00012084 ms
2.18 trunk: 0.00005475 ms

toFixed 8; diag_codePerformance [{objNull}, nil, 1e+6, true];
2.16 stable: 0.00016533 ms
2.16 perf: 0.00015065 ms
2.18 trunk: 0.00006980 ms

quaint flame
#

Wow..

empty goblet
fickle geyser
heavy vortex
#

Do you have current results for perf vs stable there?

heavy vortex
#

hmm:

perf, no mods: 280fps
perf, ace: 274fps
stable, no mods: 281fps
stable, ace: 268fps
plain trout
#

Try with rhs enabled

fickle geyser
#

RHS does not run much unscheduled code

#

CBA does with ACE due to player event handlers etc.

heavy vortex
#

Surprised it's that much for one unit tbh.

fickle geyser
#

It does not scale with amount of units that much

#

A lot of player "event handlers" are a loop due to engine lacking hooking points for them.

heavy vortex
#

fair enough. I generally put the ACE penalty at about 5% for real missions.

fickle geyser
empty goblet
#

well, now someone needs to find most common scripting inhibitor 😁

coarse plaza
#

When using setMarkerDrawPriority it puts markers with other than 0 priority to wrong positions tunres_respawn_RespawnPosLocal and tunres_respawn_MainBaseLocal, those should be same position as tunres_respawn_west.
Looks like they take the position from the original array position, so in this case from 1 and 2 array elements.

[marker name - pos - markerDrawPriority]

Stable - with setMarkerDrawPriority
[["tunres_respawn_west",[1683.92,5485.9,0],0],
["tunres_respawn_RespawnPosLocal",[1683.92,5485.9,0],100],
["tunres_respawn_MainBaseLocal",[1683.92,5485.9,0],100],
["respawn_west",[1855.93,5979.63,0],0],
["tunres_respawn_Name 1",[1672.92,5481.86,0],0],
["tunres_respawn_Name 2",[1531.44,4962.02,0],0]]


Stable - no setMarkerDrawPriority
[["tunres_respawn_west",[1683.92,5485.9,0],0],
["tunres_respawn_RespawnPosLocal",[1683.92,5485.9,0],0],
["tunres_respawn_MainBaseLocal",[1683.92,5485.9,0],0],
["respawn_west",[1855.93,5979.63,0],0],
["tunres_respawn_Name 2",[1531.44,4962.02,0],0],
["tunres_respawn_Name 1",[1672.92,5481.86,0],0]]



Perf - with setMarkerDrawPriority
[["tunres_respawn_west",[1683.92,5485.9,0],0],
["respawn_west",[1855.93,5979.63,0],0],
["tunres_respawn_Name 2",[1531.44,4962.02,0],0],
["tunres_respawn_Name 1",[1672.92,5481.86,0],0],
["tunres_respawn_MainBaseLocal",[1855.93,5979.63,0],0],
["tunres_respawn_RespawnPosLocal",[1531.44,4962.02,0],0]]


Perf - no setMarkerDrawPriority
[["tunres_respawn_west",[1683.92,5485.9,0],0],
["tunres_respawn_RespawnPosLocal",[1683.92,5485.9,0],0],
["tunres_respawn_MainBaseLocal",[1683.92,5485.9,0],0],
["respawn_west",[1855.93,5979.63,0],0],
["tunres_respawn_Name 1",[1672.92,5481.86,0],0],
["tunres_respawn_Name 2",[1531.44,4962.02,0],0]]
spiral pond
#

something seems bugged with marker reference now

#

i am seeing marker types shown for other markers. like the indexing/mapping seems to mix up

#

(we use setMarkerDrawPriority)

quaint flame
#

See above

whole cloud
#

I don't see why markers with 100 prior should be at same pos as one with 0

#

Your perf shows wrong draw priority. All zero.
The Local markers should have a 100 prio, going by your stable example.
That means they should be at the end. That means the perf with draw priority is correctly sorted.

#

Ah I see your setPos position.
Not the position (order) in the array

#

Ohhhhh I see

#

Well that's annoying

#

The elements after the insert get bumped, but their index is referenced in hashmap that doesn't get updated.
That is a annoying fix but rather trivial.
Though means on every insert (that is not at end) I'll need to iterate all markers. Should be fast enough

whole cloud
#

I can't release the fix today, I'll rever that build and we'll repush it tomorrow 🤞

whole cloud
#

This only affects using setMarkerDrawPriority, so official servers aren't affected

#

😄 @woven loom looking at your stutter issue.
Fun.

The CPU utilization is not even one core, its zero cores XD
Seems its mostly steam's fault, I'll move some stuff around then blobdoggoshruggoogly

vtune-gui_Fjx7braUcS.png
#

This causes a 8ms lagspike, every 5 seconds.
And all this code does is make that "Yay you have purchased a DLC" popup that you'll get ingame. In case some player purchases a DLC while currently playing the game (In A3's lifetime there were probably dozens of people that have seen that popup)

Fun fact, yeah that actually works.
When you get a ingame "Cannot get into this vehicle seat because you don't own the DLC". You can go buy the DLC in steam, without exiting the game, and then you can get into that vehicle.
So much attention to detail

woven loom
#

Lol so it's the check for that popup??? Lmao

#

It's weird that it doesn't disable it then if all loaded DLCs are owned

#

(I own all the first party DLC and was only running with WS and RF for these tests, both of which I own)

#

I will say, I've never used that pop up interface to buy DLC, but it has annoyed me enough to buy DLCs later 😛

silk pewter
#

"mission accomplished"

woven loom
#

Oh wait, i misunderstood

#

It just says you bought a DLC IF you buy during gameplay

#

Nevermind 🤣

#

Definitely not worth the lag spike 😛

whole cloud
#

It also has an event if the "dlc is installed" state changes. Which might happen even if you own all.
But.. That event is never used

#

I don't want to ask questions about that. I'll just make it async and the lag will disappear and I can't break anything

#

Previously running 10k iterations would freeze the game for a minute.
Now I run 10k iterations and it only drops my fps from 750 to 690
✅

woven loom
#

Thank you!

feral harness
silk pewter
#

unrealistic beauty standards

feral harness
#

1975 unrealistic

plain trout
#

that stuttering issue is so annoying. Tysm

woven loom
#

Not sure how that'll map to our systems but better than the 50% drop for sure

whole cloud
#

I don't understand how anyone could even notice a 8ms stutter every 5 seconds
But I guess some people don't play Arma on 30fps

woven loom
#

Well I at least didn't notice it by myself lol. I was looking at presentmon, noticed the suspicious frame times and thought it was worth reporting

#

The 8ms spike isn't noticeable to me since I play at 70-120 fps, but if you're playing at say 30-60 fps (so >16ms) and you're getting +8ms spikes it might become noticeable

#

Especially when other sources of frame time spikes get added in

#

There was also another ~8ms spike because of bgWait, will see if that persists once this is gone

plain trout
#

The stutters froze the game a lot longer

whole cloud
whole cloud
plain trout
#

Steam/arma checks every mod that’s installed when I start up arma

#

At least that’s what it looks like when I look into stream while the game is running

cobalt forge
#

Not every mod, but it checks all scenarios and composition you have to make sure they are updated I assume (Though it is a bit annoying, especially when updating a scenario, as it re-checks every scenario and comp. again)

plain trout
#

Other people had that issue too. One needs to be subbed to a lot of mods though

#

It’s checking after the game is started though

obsidian relic
whole cloud
night mulch
#

Yeah thats the steam infinite download loop bug

#

Which also causes stutters

plain trout
#

That’s unfortunate then

whole cloud
harsh grotto
# 
20:32:45 Error in expression <Max == _fMin}) exitwith {
_fMin = _fMin / (1000^_logMin);
_fMin = _fMin tofixed >
20:32:45   Error position: </ (1000^_logMin);
_fMin = _fMin tofixed >
20:32:45   Error #: делитель нуля (zero divisor in English)
20:32:45 File a3\Functions_F_Enoch\EM_Core\fn_frequencyToString.sqf..., line 50
20:32:45  ➥ Context:     [] L1 ()
    [] L17 (a3\UI_F_Enoch\UI\RscIngameUI\RscWeaponSpectrumAnalyzerGeneric.sqf)
    [] L143 (a3\UI_F_Enoch\UI\RscIngameUI\RscWeaponSpectrumAnalyzerGeneric.sqf)
    [] L100 (a3\Functions_F_Enoch\EM_Core\fn_showSpectrumAnalyzer.sqf)
    [] L105 (a3\Functions_F_Enoch\EM_Core\fn_showSpectrumAnalyzer.sqf)
    [] L104 (a3\Functions_F_Enoch\EM_Core\fn_showSpectrumAnalyzer.sqf)
    [] L49 (a3\Functions_F_Enoch\EM_Core\fn_frequencyToString.sqf)
    [] L52 (a3\Functions_F_Enoch\EM_Core\fn_frequencyToString.sqf)

😦 fix pls pls pls

#

every second this error with Spectrum Analyzer in my hand and i can't check for other errors while playing 😦

heavy vortex
#

you could make a tiny mod that overrides the function :P

harsh grotto
heavy vortex
#

You use the spectrum analyzer in warlords?

harsh grotto
#

yep

heavy vortex
#

I guess that game mode isn't what I thought it was.

harsh grotto
#

yes, not work, only for more faster movement, cuz spectrum analyzer weight superlow

heavy vortex
autumn timber
#

So kinda like running with knives in CS? 😆

harsh grotto
#

yep 🙂

whole cloud
#

You can also just ignore it

harsh grotto
#

meowheart love yu

light cargo
tawdry gazelle
fast hornet
#

oh nice.. is 2.1.4 battleye whitelisted already?

#

hm nope, doesn't look like it

tribal pasture
#

Not really a perf issue exactly but also happens in perf build. We get weird situation where people have their game freezing after next round in KotH. Basically they get back into lobby, join and it freezes on loading screen forever.

#

Apparently this started since 2.16 but I'm not sure

#

I was having similar issue recently with my local server but it was happening under difference circumstances. I launch local dedicated server, join it with game instance, after 1 minute I get kicked with Player Sa-Matra kicked off by BattlEye: Client not responding, join the server again, select role in lobby, start and it freezes during loading.

#

I tried both stable and perf, same issue. I also tried my mission build from 2 years ago, also same issue.

#

I also tried empty basic mission and it doesn't happen there.

#

So it is something that happens in my mission that triggers freeze on second join/round but wasn't doing that until recently.

#

Will appreciate ideas what I could do to figure it out, it is very server killing for us, people get their game freezing on second round on live servers.

#

Maybe an extended logging client build to get better idea at what point it freezes?

#

Apparently it doesn't even get to vehicle init scripts stage, as that's where all my scripts start and instantly log a line that I don't see in such freeze.

#

There is no crash dump in such freeze

whole cloud
#

I heard that before.
Something with the server and rejoining due to BE thinking you're not responding

Is it freeze? doesn't sound like it actually is a freeze. You mean they get stuck in loading screen and it never finishes? WITHOUT the game freezing?

stray prawn
#

noticed this on wasteland too, players either get a battleye not responding issue or the game crashes with a access violation. (it started with 2.14 (barely) but became worse since 2.16)

thin wyvern
#

I have this problem after server kick me because high ping. My game is freez and I must kill it in task manager arma3.exe not responding.

whole cloud
#

crash is not a freeze or stuck in loading screen

whole cloud
rare compass
tribal pasture
#

When you kill it it doesnt generate anything

whole cloud
#

then you can do a dump in task manager

tribal pasture
#

On it

thin wyvern
tribal pasture
#

Been frozen for good 15 minutes, 0 changes in RPT, no freeze dumps

#

Other processes dump fine thonk

#

Same with both stable and perf

#

Okay, managed to do it with procdump64

random isle
whole cloud
#

Thanks to the memory dump I can see where and why it freezes. But don't yet know how it got into that situation.
It looks like some code is not cleaning up properly on kick. It ends up in an infinite loop trying to resolve a script variable.
Because the scope's parent is, itself. If it tries to look up variable and fails, it looks in parent scope. And if that is a self reference thats infinite loop.

That should (obviously) not be possible to happen.
There is some nonsense code in there. Like vehicle init scripts, have a parent scope. The script should be at "root" level without a parent, that is definitely wrong. But its old code and has been like that for a long time. Easy to fix that, but there must be some issue elsewhere too.

thin wyvern
#

I don't have any dump file only see this in .rpt

#

When try to re-connect

#

20:37:49 EPE manager release (4|34|0)
20:37:49 Number of actors in scene after release: 4
20:37:49 EPE manager release (0|4|0)
20:37:50 Starting mission:
20:37:50 Mission file: introOrange
20:37:50 Mission world: Altis
20:37:50 Mission directory: a3\missions_f_orange\scenes\introOrange.Altis
20:37:51 [CBA] (xeh) INFO: [45149,410.823,0] PreInit started. v3.17.0.240401
20:37:51 [CBA] (xeh) WARNING: One or more children of class CAManBase do not support Extended Event Handlers. Fall back to loop.
20:37:51 [CBA] (settings) INFO: Reading settings from settings file.
20:37:51 [CBA] (settings) INFO: Finished reading settings from settings file.
20:37:51 [CBA] (xeh) INFO: [45149,411.009,0] PreInit finished.

#

and after this nothing...

#

And what I see in all .rpt files after I try to re-connect to the server I get this message

#

19:26:48 EPE manager release (x|x|x)
19:26:48 Starting mission:
19:26:48 Mission file: introOrange
19:26:48 Mission world: Altis
19:26:48 Mission directory: *a3\missions_f_orange\scenes\introOrange.Altis*

#

but missions is not on Altis 🤔

whole cloud
#

Like vehicle init scripts, have a parent scope.
That code that does that, has been copy pasted about 130 times in tons of places.
Its likely there are moer places that get parent scope, when they really shouldn't.
That init script from cba probably too

Most scripts that run at start can probably run into that issue

silk pewter
#

helper method incoming

kindred radish
#

I can trigger that by closing down my test server exe while im still on it.
If i logout/dc before i don't get that problem.

inland dew
empty goblet
#

combined with the not released socket problem (you can't rejoin from the same IP for 10 minutes (or longer) or unless server restarts)

indigo anvil
tribal pasture
#

What I also hate about leaving the server is that you can randomly have mission file locked

#

And if server updated the mission on next start, it redownloads it but can't overwrite the old one, ditching you back into server browser

#

Been an issue since Apex

light cargo
#

wouldnt using -loadmissiontomemory help at all?

worn obsidian
#

its my birthday ! a gift? perf1 ? :o

kindred radish
#

I had the missionfile still under lock for the server after closing the server exe, it stopped when i moved all my .paa's to the mod instead of missonfile (been a while tho (1,5-2 years)) blobdoggoshruggoogly

whole cloud
#

there is generally a bug where server doesn't release file locks

woven loom
fast hornet
# woven loom how did you check?

ms process explorer can tell you about dll handles.. but the most obvious one in this case would be the BE message on launch 😄

woven loom
#

ah okay there's a message

#

i thought you were checking some sort of list somewhere

#

thanks

novel pier
#

Question, I have heard that there is improved exe

#

files

#

where can I found them

worldly badge
novel pier
#

And what do I download from there

#

this one

worldly badge
#

It literally is written there

#
  1. Use Steam Branch (search for steam game beta branch)
  2. Download it from GDrive and put it into Arma 3 folder
    Pick one
novel pier
#

is there autodownloading for those files?

whole cloud
#

Use the steam branch.

novel pier
whole cloud
#

yes... The branch. indeed.

worldly badge
#

It literally says the option

novel pier
#

profiling?

novel pier
worldly badge
#

You are stupid, I get it

novel pier
#

so choose profiling - performance

#

?

#

and that is it?

whole cloud
#

You are in the profiling branch channel.
You linked the profiling branch forum thread above.

Yes. You want the branch that says profiling.
Yes that is it.

novel pier
#

ty dedmen ❤️

#

and what now?

worldly badge
#

Play the game?

novel pier
#

that is it?

#

no more steps?

worldly badge
#

Yes?

novel pier
#

ty

indigo anvil
fast hornet
woven loom
#

awesome! thank you

whole cloud
#

2.16.151782 new PROFILING branch with PERFORMANCE binaries, v4, server and client, windows 64-bit, linux server 64-bit
- Tweaked: Improved performance of getModelInfo/getEntityInfo/getShotInfo
- Tweaked: Slightly improved script performance
- Fixed: diag_codePerformance had unexpected overhead, the measurement result could be 10% off from the real execution time
- Fixed: HashMap circular reference check when saving the game had a memory leak and false positive - https://feedback.bistudio.com/T180508
- Fixed: Map markers were mixed up when setMarkerDrawPriority was used (since v3)
- Fixed: ~8 millisecond micro stutter every 5 seconds - https://feedback.bistudio.com/T180665
- Fixed: Possible game freeze when rejoining a MP server after being kicked

If you don't want to use the Steam branch, the files are also available for alternative download here:
https://drive.google.com/drive/folders/15p9j7C2nHUt6NoVfChX4YFuqzFXzblJh
Note: There are separate Dll files that also need to be placed into Game folder.
🪖

whole cloud
charred holly
#
  • Tweaked: Improved performance of getModelInfo/getEntityInfo/getShotInfo
    Those are already in the profiling branch?
shell lion
opal hound
#

I have some dumps if they're needed

shell lion
#

me to from a lot of players

orchid mesa
#

king of the hill with rhs mods also crashes your game when entering a vehicle on profiling build

midnight oyster
#

Most of the game seems busted in profiling branch after the latest update

light cargo
#

downgrading to previous build via google drive then

woven loom
#

Seems fine with vanilla though

whole cloud
#

huh I broke getModelInfo. interesting. How did I manage that

#

Ah that'd do it. Why did I now see that on my tests...
Creating a 4 element array and then setting the 5th element

"Improved performance of getModelInfo/getEntityInfo/getShotInfo"
Ah that's wrong, cuz only getModelInfo is in prof.

And there is also the bug with that 5th element, that exists in dev-branch but not prof so I merged that badly

full nova
whole cloud
#

Crash fix is up, we replaced v4 with it

heavy galleon
#

Does that include the one I encountered?

harsh grotto
empty goblet
#

forgot about to pin the updated allocator

feral harness
harsh grotto
#

ahahah

random isle
# 
_vehicle deleteVehicleCrew _unit

Doesnt seem to work in latest pref

heavy vortex
#

Can we just get rid of deleteVehicleCrew and make deleteVehicle work everywhere :P

silk pewter
light cargo
#

which custom mallocs are recommended for use with arma 3? especially on windows 7

random isle
light cargo
#

just that one?

inland dew
#

yes

midnight oyster
#

Can confirm that yesterday's hotfix cleared up my crashing problem 👌🏾

tribal pasture
#

(deleteVehicleCrew still doesn't work in all cases btw)

heavy vortex
#

deleteVehicleCrew doesn't even specify what it's for. It's just some vague threat.

#

I haven't found anything that actually breaks by just using deleteVehicle on crew. At least since ACE fixed that bug in their last version.

tribal pasture
#

I had issues with deleting "expired" player body recently

#

if player keeps control of the dead body for a long time (minute+), it breaks and no longer properly deletes with either command

heavy vortex
#

makes sense

tribal pasture
#

Hmm, did some tests and issues happen when vehicle is upside down 🤔

#

So, nevermind my comments, the issue was with upside down vehicle, deleteVehicleCrew worked with server-client setup, deleteVehicle on the other hand left ghost body.

#

To my surprise vehicle you can't move into upside down vehicles with getInDriver thonk

#

I thought deleteVehicleCrew breaks the seat

#

Still, deleteVehicle should just work with any dead bodies in the first place

random isle
#

i also noticed locally when i killed the driver of vehicle and do car deleteVehicleCrew dude it auto moves another ai from cargo into driver seat, maybe it was a hacky way of getting dead dude eject 😄

analog acorn
#

I'm pretty sure that's just AI wanting to have a driver. If you put an AI alone in a vehicle passenger seat to begin with, they will also switch to driver if possible

feral harness
fast hornet
#

AI really do move in vehicles on their own.. but i've mostly seen them moving to the gunner position

spiral pond
#

AI crew gets out to take the driver seat when the driver gets killed

quaint flame
#

I found something weird, not sure if it's related to perf yet, though.

Executing ["GetAllDisplayControls"] call (uiNamespace getVariable ["RscDisplayEGSpectator_script", {}]) will return
[00007FF613E23910,Control #42609,Control #43109,Control #43609,Control #52909,Control #48909,Control #43910,Control #73189,Control #72812,Control #83909]

#

The first one should be a control as well, but ctrlType returns "scalar", I have never seen that before

#

Ah, seems like a bad define in the script causing that

[
_display displayCtrl IDC_RSCDISPLAYEGSPECTATOR_GAMETIME,
_display displayCtrl         42609,
_display displayCtrl             43109,
_display displayCtrl             43609,
_display displayCtrl     52909,
_display displayCtrl     48909,
_display displayCtrl     43910,
_display displayCtrl             73189,
_display displayCtrl     72812,
_display displayCtrl         83909
];
#

But shouldn't that return controllNull then?

#

The related script is a3\ui_f\scripts\gui\rscdisplayegspectator.sqf btw

tribal pasture
#

That address is result of nil format

#

Some nil types aren't formatted into readable string but to some address

#

[format ["%1", agent nil], str agent nil, format ["%1", getPos nil], getPos nil]
=>
["00007FF6E0173650",string,"array",array]

quaint flame
#

Question is, why does displayCtrl return nil?

#

Ah well, because it's never executed.

#

Because of that missing define

tribal pasture
#

yeah, it treats it as missing global variable

quaint flame
trim ether
#

Strange issue with the perf build. Our (3CB) airburst rounds do not detonate in the air above the target. I've narrowed it down to a change in createVehicle

[] spawn {
sleep 1;
_pos = getPos player;
_shell = "UK3CB_BAF_Sh_60mm_AMOS_AB_helper"createVehicle [_pos select 0, _pos select 1, 30];
}```
This code causes the ammo to explode above the player in vanilla.  In perf build it does nothing.
feral harness
trim ether
#

The alt syntax makes no different. Still no effect in Perf build

[] spawn {
sleep 1;
_pos = getPos player;
_shell = createVehicle ["UK3CB_BAF_Sh_60mm_AMOS_AB_helper", [_pos select 0, _pos select 1, 30],[],0,"CAN_COLLIDE"];
}```
feral harness
#

ok weird. and if you _shell setpos [_pos select 0, _pos select 1, 30]; additionally?

trim ether
#

no effect

#

i even tried _shell setDamage 1.0