Silk.NET•5y ago

isn't glGetTextureSubImage one of the DSA methods

PPerksey isn't glGetTextureSubImage one of the DSA methods

TechPizza•9/9/21, 4:28 PM

it is but glReadPixels didn't work either

TechPizza•9/9/21, 4:30 PM

i forgot to tell that i do a image copy operation that may or may not be broken
i will debug it further but not today

TechPizza•9/10/21, 7:46 AM

yeah, glGetTextureSubImage immediately after glTexSubImage2D doesn't work, how strange

TechPizza•9/10/21, 8:44 AM

glGetTexImage didn't work either, fantastic

queenvicky•9/10/21, 8:55 AM

I'm not absolutely sure but I think you have to glTexImage2D before you can glTexSubimage2D or something along those lines

TechPizza•9/10/21, 9:02 AM

well i'm not getting any gl error or debug messages so that's both good and bad

novaleaf•9/10/21, 3:56 PM

Coding / Perf question. First the setup:
I'm trying to code review some rendering code (I'm not a rendering person)

We are trying to make a game engine, so we are trying to learn how to make a pipeline using opengl via silk.net .

We have a simple e2e stress test: create 90k objects and move them around every frame. Right now the bottleneck seems to be cpu bound, in this code:

with 90k objects, all moving every frame, about 50% of the cpu is spent in this function, with the times split about equally between the various functions called, which is Silk.NET

I'm not a rendering person, but I am guessing that this is cpu blocking while sending data to the gpu.

So finally my question:
Any ideas on how to make this faster? Specifically, is there a way to do this multithreaded? like calling this function in parallel? or is that not valid for silk.

novaleaf•9/10/21, 3:58 PM

fyi, the funky is a hacky workaround to let us send a directly to the method. otherwise we need to do a byte[] copy

Deleted User•9/10/21, 4:06 PM

first of all delegates can be a problem. why can't you use the overload? (or )

PerkseyOP•9/10/21, 4:06 PM

they are

novaleaf•9/10/21, 4:07 PM

yeah that's something I did as part of the code review, just to stop doing byte[] coppies. improved framerate by like 15%

novaleaf•9/10/21, 4:08 PM

but now hitting the cpu bound "issue". maybe nothing to do (moving 90k items/frame is overkill) but I'm just trying to learn how to improve rendering systems

PerkseyOP•9/10/21, 4:09 PM

it doesn't look like you're capturing so you can just make the lambda a static function

PerkseyOP•9/10/21, 4:09 PM

and i recommend using function pointers instead of delegates if you're microoptimising this hard

Deleted User•9/10/21, 4:09 PM

just don't use either

novaleaf•9/10/21, 4:09 PM

Yeah the lambda is there just so the implementation is "close" to the usage. it's cached so okay perf / no allocations

PerkseyOP•9/10/21, 4:09 PM

you can have static local functions

Deleted User•9/10/21, 4:10 PM

just don't do this, probably a bad idea to have that in there

novaleaf•9/10/21, 4:10 PM

all the perf hits are in the 3 functions. so looking at how to improve that

PerkseyOP•9/10/21, 4:10 PM

oh, well then there's no way

PerkseyOP•9/10/21, 4:11 PM

just minimise your gl calls as much as possible

PerkseyOP•9/10/21, 4:11 PM

you could also try out direct state access which may have a marginal buff

PerkseyOP•9/10/21, 4:11 PM

but that limits your platform targets to only the modern stuff

novaleaf•9/10/21, 4:12 PM

PerkseyOP•9/10/21, 4:12 PM

yeah binding buffers is expensive at a conceptual level

PerkseyOP•9/10/21, 4:13 PM

you can try using NamedBufferData

PerkseyOP•9/10/21, 4:14 PM

available on OpenGL 4.5 or in the extension

novaleaf•9/10/21, 4:14 PM

is there any way to do this multithreaded?

PerkseyOP•9/10/21, 4:14 PM

lol no

PerkseyOP•9/10/21, 4:14 PM

this is opengl

Deleted User•9/10/21, 4:14 PM

also, somewhat unrelated to your issue, is this code available somewhere, or could you share it just with me/the maintainers?
I have tried making some perf test apps, but it seems like you have found a use case that has a lot of impact from GL calls, would like to investigate whether we can do anything in terms of overhead here.
Also are you sure that you need to bind all those buffers?

PerkseyOP•9/10/21, 4:15 PM

Also are you sure that you need to bind all those buffers?

you have to bind the buffer before you write to it

novaleaf•9/10/21, 4:15 PM

I didn't write the code, I need to ask if I can share it. Sadly I don't know a thing about render pipelines so I can't critique how he's doing instancing

novaleaf•9/10/21, 4:18 PM

Okay thanks... I am going to try porting this code to my ecs engine, so I will ask for review then. (probably a month though, still working on the ecs....)

novaleaf•9/10/21, 4:20 PM

In the samples, are there examples on "best practice" for doing static and dynamic instancing?

PerkseyOP•9/10/21, 4:21 PM

not currently

novaleaf•9/10/21, 4:21 PM

okie. no biggie. too noob for it to really matter to me

novaleaf•9/10/21, 4:22 PM

I think it's probably more important for me to abstract the rendering away from the rest of the pipeline/engine. Anything in the samples showing that?

PerkseyOP•9/10/21, 4:24 PM

we haven't made a game engine atop silk.net, no.

PerkseyOP•9/10/21, 4:24 PM

others have, but we don't have a sample showcasing what this would look like in the repo.

novaleaf•9/10/21, 4:25 PM

okay thanks. I'll fiddle with it! hope to open source my engine if it survives the egg laying process

Deleted User•9/10/21, 4:31 PM

any real world profiler data or potentially code is worth gold to us at this point

novaleaf•9/10/21, 4:53 PM

okay I will ask the person who did the current render experiment if he's willing to share the code. I don't think he wants to though.

PerkseyOP•9/10/21, 5:00 PM

fair enough

PerkseyOP•9/10/21, 5:00 PM

but for things like this, assume other people not being able to see the code == other people not being able to help optimise the code

novaleaf•9/10/21, 5:11 PM

oh I agree, totally. I'm trying to get him to agree to do things under agpl at least.

novaleaf•9/10/21, 5:12 PM

I guess he got burnt in the past by people making money off his hobby projects

Deleted User•9/10/21, 5:13 PM

it'd be great if he could share it with just us, no publication needed. I'd also be happy to sign an NDA or something if he wants that. Just looking to gain some insight where most time is spent in Silk.NET

TechPizza•9/10/21, 5:16 PM

are you calling that buffer update 90000 times?

novaleaf•9/10/21, 5:18 PM

cool thanks! We are honestly just starting out on this... I have some indie game experience in the xna days and he has done some open source games. Ultimately we want to do vulkan, but he's doing opengl first to get used to silk.

I want to take this "reference" and get it working in the ecs system I'm making. (I'm a low level systems guy)