Phobos - Vulkan abstraction library in Rust | Graphics Programming | Page 2

iron shadow Jul 4, 2023, 11:32 AM

#

LMAOO

#

IT FIXED ITSELF

#

IDK WHAT I DID

#

BUT I DID SOMETHING

#

🥳

#

WOOOOH

#

SPONZA NEVER HAVE I BEEN EVER HAPPY TO SEE YOU

#

i am crying rn i figured it out. i set the # of instances for my build_info's range but didn't do it for querying the build size 😭

night charm Jul 4, 2023, 3:14 PM

#

Ahhh

#

Lmao

#

That’ll do it

languid fern Jul 4, 2023, 3:57 PM

#

My laptop 1050 starts slowing down at around 3m triangles which is pretty amazing of itself

#

So if a desktop 2080 ti is getting bottlenecked by a single 1m tris then something must be very wrong

iron shadow Jul 4, 2023, 3:59 PM

#

oh no it loads fast now

#

idk if i should multithread it and make it concurrent

#

eh premature optimization is bad

limber rapids Jul 4, 2023, 3:59 PM

#

I can reach nanite level of performance with a single thread, you're good

iron shadow Jul 4, 2023, 3:59 PM

#

KEKW UE5

limber rapids Jul 4, 2023, 4:00 PM

#

Strictly in terms of triangles/s KEKW

iron shadow Jul 5, 2023, 2:05 AM

#

fellas

#

im proud to announce

#

after years, decades, centuries

#

#

monke

#

i think i did the rotation wrong however ;kek

iron shadow Jul 5, 2023, 2:29 AM

#

so how would define

#

the .transform

#

I'm so confused

#

.transform seems to take in a TransformMatrix type

#

but I cannot find on it's documentation how I would input my floats + vecs

#

let transform = vk::TransformMatrixKHR {
  matrix: [...}
}
phobos::TransformMatrix(transform);

Is not possible as it is a private constructor

#

bleakekw

#

bro what is there not a way to create a transform matrix with your own transforms. there is literally only identity

#

desperate times call for desperate measures im gonna make a pr

iron shadow Jul 5, 2023, 4:51 AM

#

uh penguin is texture loading not supported?

#

let image = images.get(gltf_image.index()).unwrap();
            let image_data = &image.pixels;
            let image_phobos = phobos::image::Image::new(
                ctx.device.clone(),
                &mut ctx.allocator,
                image.width,
                image.height,
                vk::ImageUsageFlags::SAMPLED | vk::ImageUsageFlags::TRANSFER_DST,
                get_image_type(image.format),
                vk::SampleCountFlags::TYPE_1,
            )
            .unwrap();

the format is R8G8B8A8_SRGB

iron shadow Jul 5, 2023, 6:12 AM

#

It keeps saying I need to specify a depth?

night charm Jul 5, 2023, 6:28 AM

#

iron shadow .transform seems to take in a TransformMatrix type

I forgor

#

You can’t actually construct it with things yet

night charm Jul 5, 2023, 6:29 AM

#

iron shadow It keeps saying I need to specify a depth?

Where depth

#

Image loading should work

iron shadow Jul 5, 2023, 6:39 AM

#

night charm You can’t actually construct it with things yet

LMAOOO pls fix

iron shadow Jul 5, 2023, 6:39 AM

#

night charm Image loading should work

I might be messing it up? I’m not entirely sure. It’s late rn so i’ll send you the code later

#

I simply want to create images for my gltf asset. I also create a buffer to store the image itself, but iirc you don’t need to do anything prior to get images to work, yeah?

night charm Jul 5, 2023, 6:43 AM

#

iron shadow LMAOOO pls fix

Yeah I’ll do it later today, I’m a bit busy rn

iron shadow Jul 5, 2023, 6:44 AM

#

Eh i can make a pr tbh if you can’t

night charm Jul 5, 2023, 6:44 AM

#

iron shadow I simply want to create images for my gltf asset. I also create a buffer to stor...

You don’t have to create a buffer at all

#

The image manages its memory

iron shadow Jul 5, 2023, 6:44 AM

#

Wait I think I’m screwing up then 💀

#

Where would you specify that actual data of the image

night charm Jul 5, 2023, 6:45 AM

#

You create a command buffer and upload the data

#

Probably through a buffer-image copy yeah

iron shadow Jul 5, 2023, 6:45 AM

#

Ohh ok cuz I’m trying to get textures

#

Ah that might be why then KEKW
I was just making an image without doing the copy

night charm Jul 5, 2023, 6:47 AM

#

Yeah it’s gonna be empty then lol

iron shadow Jul 5, 2023, 6:47 AM

#

But that shouldn’t be causing any errors thonk

night charm Jul 5, 2023, 6:48 AM

#

Idk what your full code is

iron shadow Jul 5, 2023, 6:49 AM

#

Ill share it when i wake up tomorrow

night charm Jul 5, 2023, 6:49 AM

#

Alright

iron shadow Jul 5, 2023, 9:00 AM

#

Huh I realized. I should be moving my buffers onto my gpu

#

All of my buffers have been CpuToGpu bleakekw

#

Iirc i make a staging buffer which is cputogpu then add to the command buffers cmds to copy that gpu information over

night charm Jul 5, 2023, 9:10 AM

#

Yep that’s pretty much it

iron shadow Jul 5, 2023, 9:24 AM

#

I guess I should also merge all my vertices together into one monolithic one while I’m at it. Wasn;t this called bindless or something?

#

Ah wait nvm it means not making binds (referring to buffers by their bda) im stupid lmaooo

night charm Jul 5, 2023, 9:44 AM

#

iron shadow I guess I should also merge all my vertices together into one monolithic one whi...

You can do that yeah

iron shadow Jul 5, 2023, 9:51 AM

#

Wait there’s no cpu only memory option?

limber rapids Jul 5, 2023, 9:53 AM

#

HOST_VISIBLE | HOST_COHERENT | HOST_CACHED is CPU_ONLY

iron shadow Jul 5, 2023, 9:53 AM

#

Ohh it’s been moved into flags I gotcha

iron shadow Jul 5, 2023, 11:23 AM

#

Penguin

#

here is the exact error:

 ERROR phobos::core::debug             > [VALIDATION]: VUID-VkImageCreateInfo-imageCreateMaxMipLevels-02251 (-1094930823): Validation Error: [ VUID-VkImageCreateInfo-imageCreateMaxMipLevels-02251 ] Object 0: handle = 0x1b9bf6ba320, type = VK_OBJECT_TYPE_DEVICE; | MessageID = 0xbebcae79 | vkCreateImage(): Format VK_FORMAT_R8G8B8_SRGB is not supported for this combination of parameters and VkGetPhysicalDeviceImageFormatProperties returned back VK_ERROR_FORMAT_NOT_SUPPORTED. The Vulkan spec states: Each of the following values (as described in Image Creation Limits) must not be undefined : imageCreateMaxMipLevels, imageCreateMaxArrayLayers, imageCreateMaxExtent, and imageCreateSampleCounts (https://vulkan.lunarg.com/doc/view/1.3.243.0/windows/1.3-extensions/vkspec.html#VUID-VkImageCreateInfo-imageCreateMaxMipLevels-02251)
error: process didn't exit successfully: `target\debug\DARE.exe`

hybrid pilot Jul 5, 2023, 11:23 AM

#

How DARE.exe you

iron shadow Jul 5, 2023, 11:23 AM

#

KEKW Danny's Awesome Rendering Engine

hybrid pilot Jul 5, 2023, 11:24 AM

#

Lol

iron shadow Jul 5, 2023, 11:24 AM

#

oh shoot

#

mayube srgb is it

hybrid pilot Jul 5, 2023, 11:24 AM

#

iron shadow mayube srgb is it

Try rgba

iron shadow Jul 5, 2023, 11:26 AM

#

Oh im loading from a file

#

i dont think i can

#

wait nvm

#

i can im a dumbass

#

R32G32B32A32_SFLOAT works idfk why everything else doesn't

#

RGBA seems to only be supported

#

wtf

#

@hybrid pilot yeah that fixed it...? But how tf am I supposed to conform my textures which im loading from gltf to it

limber rapids Jul 5, 2023, 11:28 AM

#

Are you trying to create an image on HOST_VISIBLE memory

iron shadow Jul 5, 2023, 11:29 AM

#

My only flags rn are SAMPLED and TRANSFER_DST so I doubt so

limber rapids Jul 5, 2023, 11:29 AM

#

Show me the create info

#

VmaCreateInfo for the image

iron shadow Jul 5, 2023, 11:30 AM

#

this is phobos

#

let image_phobos = phobos::image::Image::new(
                ctx.device.clone(),
                &mut ctx.allocator,
                image.width,
                image.height,
                vk::ImageUsageFlags::SAMPLED | vk::ImageUsageFlags::TRANSFER_DST,
                vk::Format::R8G8B8_SRGB,
                //get_image_type(image.format),
                vk::SampleCountFlags::TYPE_1,
            )
            .unwrap();

#

i don't really have access to the vmaCreateInfo cuz:

this is rust
phobo'ing time

limber rapids Jul 5, 2023, 11:32 AM

#

Huh

#

it's basically always device local

#

https://github.com/NotAPenguin0/phobos-rs/blob/master/src/resource/image.rs#L170

#

You can just look at the source you know

#

Anyways looks like some values are undefined for some reason

iron shadow Jul 5, 2023, 11:33 AM

#

oh it is but that really shouldn't be causing any issues?

limber rapids Jul 5, 2023, 11:33 AM

#

Yeah

iron shadow Jul 5, 2023, 11:34 AM

#

🫡 ethereal bug

#

ill run it through the debugger and see what is causing it

#

thats odd, everything it is complaining about is valid

hybrid pilot Jul 5, 2023, 11:41 AM

#

iron shadow <@137076228875091968> yeah that fixed it...? But how tf am I supposed to conform...

You might have to do some conversions

iron shadow Jul 5, 2023, 11:41 AM

#

seems like a massive waste of gpu memory though

#

ahh ill dig more into it

#

well seems like my gpu only supports r8g8b8a8

#

sadge

limber rapids Jul 5, 2023, 11:50 AM

#

Wait you were trying to create a r8g8b8 image?

#

bruh

#

Quite literally no GPU in the world supports that format for color KEKW

iron shadow Jul 5, 2023, 11:51 AM

#

LMAOOO i didn't realize it's not supported whatsoever i was so fixated on the vulkan specs 💀 saying stuff was null

limber rapids Jul 5, 2023, 11:52 AM

#

Check on GPU info for support

iron shadow Jul 5, 2023, 11:52 AM

#

yeah thats what i did i was coping maybe gpu didnt support it and my code wasnt broken as my copium came true. the real question is how to convert it all to rgba bleakekw

limber rapids Jul 5, 2023, 11:55 AM

#

If you are using stb image you can ask it to load all 4 channels regardless

iron shadow Jul 5, 2023, 11:55 AM

#

i wish froge_sad i'm using the in-built crate loader. doesn't seem like they offer a way to force it all rgba

limber rapids Jul 5, 2023, 11:55 AM

#

Otherwise you can just do a shrimple conversion

#

I can only speak C++ though so my example will be C++

#

// assume it is a 3-channel texture
auto texture = load_texture();
const unsigned char* data = texture.data;
auto rgba = std::vector<uint32>(texture.width * texture.height);
for (auto i = 0; i < rgba.size(): ++i) {
    texel =
        (data[i * 3 + 0] << 24) | 
        (data[i * 3 + 1] << 16) | 
        (data[i * 3 + 2] << 8) | 
        255;
}

iron shadow Jul 5, 2023, 11:59 AM

#

its been so long c++ KEKW

#

oh wtf iterating through the array by ptr

#

KEKW https://github.com/gltf-rs/gltf/issues/376 seems like im not alone in this issue

GitHub

Add an option to load R, RG, RGB images as RGBA. · Issue #376 · glt...

I tried to use R8G8B8_SRGB format in Vulkan but got an error. Since most graphics card does not support RGB only image, it would be convenient and faster to load them as RGBA images (using the imag...

limber rapids Jul 5, 2023, 12:03 PM

#

Less raw version

#

Perhaps more readable to a Rust apostle

iron shadow Jul 5, 2023, 12:04 PM

#

oh no i havent gotten close to bit stuff yet thats why it's indecipherable

limber rapids Jul 5, 2023, 12:04 PM

#

It's easy, r8g8b8a8 means 8 bits per channel

iron shadow Jul 5, 2023, 12:04 PM

#

yeah that makes sense, im just confused about <<

limber rapids Jul 5, 2023, 12:05 PM

#

You want to convert r8g8b8 to r8g8b8a8, so what you do is shift everything 8 bits to the left, and add 255 or 0xff to the end

iron shadow Jul 5, 2023, 12:06 PM

#

ohh this makes sense

#

should i ask chatgpt KEKW

limber rapids Jul 5, 2023, 12:14 PM

#

no

limber rapids Jul 5, 2023, 12:14 PM

#

limber rapids ```cpp // assume it is a 3-channel texture auto texture = load_texture(); const ...

understand this

#

written by a human for a human (I assume)

iron shadow Jul 5, 2023, 12:14 PM

#

big assumptions we're making

#

rust has no way to bit offset

limber rapids Jul 5, 2023, 12:14 PM

#

wat

iron shadow Jul 5, 2023, 12:14 PM

#

i think ill have to do that manually then sadge

limber rapids Jul 5, 2023, 12:15 PM

#

You have no << in rust

iron shadow Jul 5, 2023, 12:15 PM

#

no

#

afaik

limber rapids Jul 5, 2023, 12:15 PM

#

https://doc.rust-lang.org/std/ops/trait.Shl.html you are lying

iron shadow Jul 5, 2023, 12:15 PM

#

wtf

#

i couldn't find it 💀

#

oh wait there is a better way so i can also support multiple component counts

hybrid pilot Jul 5, 2023, 12:19 PM

#

iron shadow i think ill have to do that manually then sadge

I do it with iterator

#

like you have your u8 vec and you do .chunks(3).flat_map(|p| [p[0], p[1], p[2], 255])

iron shadow Jul 5, 2023, 12:20 PM

#

im trying to figure out a way to have it work across all the possible types
thonk

limber rapids Jul 5, 2023, 12:21 PM

#

Do you see the i * 3

#

change 3 to component_count

#

and then switch on component count, with fallthrough bitwise shifts

#

multiplied by the component count

hybrid pilot Jul 5, 2023, 12:21 PM

#

iron shadow im trying to figure out a way to have it work across all the possible types <:th...

Same but T::MAX instead od 255

iron shadow Jul 5, 2023, 12:21 PM

#

nono like 8, 16, 32 bits

hybrid pilot Jul 5, 2023, 12:22 PM

#

Mmm

#

Not sure what you mean

limber rapids Jul 5, 2023, 12:22 PM

#

Don't think about that

iron shadow Jul 5, 2023, 12:22 PM

#

AHHHH i realize

#

yeah i shouldn't

limber rapids Jul 5, 2023, 12:22 PM

#

You won't have to deal with those textures

iron shadow Jul 5, 2023, 12:22 PM

#

KEKW

iron shadow Jul 5, 2023, 12:39 PM

#

what do yall think the chances of me running into a r8 texture gonna be

iron shadow Jul 5, 2023, 12:41 PM

#

limber rapids ```cpp // assume it is a 3-channel texture auto texture = load_texture(); const ...

sorry for asking but would this do an arithmetic overflow? At least that is what the rust compiler is complaining about however we are oparting on two different-ish types i think? I'm using a &[u8] which is basically an array of bytes

limber rapids Jul 5, 2023, 12:41 PM

#

cast to u32

#

C++ integer promotion rules are fucked and Rust decided to not deal with any of that and making you do everything

iron shadow Jul 5, 2023, 12:43 PM

#

oh it works

#

seems uhh

#

very hacky? im probably not used to it

limber rapids Jul 5, 2023, 12:45 PM

#

what do you think you've been programming up until now

#

programming itself is a bunch of hacks piled up on top of each other

#

we hacked sand and tricked it into thinking, what's one more hack

iron shadow Jul 5, 2023, 12:46 PM

#

that's the most real advice i have heard

#

thank you frogapprove

hybrid pilot Jul 5, 2023, 12:46 PM

#

iron shadow what do yall think the chances of me running into a r8 texture gonna be

My thought is to worry about it when you come across it

iron shadow Jul 5, 2023, 12:47 PM

#

heightmaps bleakekw

#

ah it doesn't work nvm

#

seems like ill have to do what pac85 does cuz rust is throwing a fit about how i shouldn't be going over my array

limber rapids Jul 5, 2023, 12:50 PM

#

I have no idea about idiomatic ways in rust

#

I have no idea about Rust in general KEKW

iron shadow Jul 5, 2023, 12:50 PM

#

me neither KEKW

#

i just get a happiness hit when i see
vec.iter().map().collect() it feels so nice pretty sure cpp has its own way

limber rapids Jul 5, 2023, 12:52 PM

#

Yeah it's called: no we don't have anything

#

you use for like god intended

iron shadow Jul 5, 2023, 12:52 PM

#

istg you guys have a map function

iron shadow Jul 5, 2023, 12:52 PM

#

limber rapids you use `for` like god intended

but me functional programming :(

limber rapids Jul 5, 2023, 12:53 PM

#

Probably std::ranges has something

iron shadow Jul 5, 2023, 12:53 PM

#

well for is functional but shhh

limber rapids Jul 5, 2023, 12:53 PM

#

But ranges are trash so nobody uses them

iron shadow Jul 5, 2023, 12:53 PM

#

LMAOO

#

ill go back to cpp when they finally get modules to work (they will never)

#

oh shoot i could technically make your way work, it's just that i need to allow arithmetic overflow

limber rapids Jul 5, 2023, 12:54 PM

#

Just do what pac did if it works

iron shadow Jul 5, 2023, 12:55 PM

#

yeah but i wanna use this learn more about bitwise operators

#

eh pac is probably better as it would be "safer" idfk know though i'm assuming as there is a possibility you read outside of the intended data

limber rapids Jul 5, 2023, 12:56 PM

#

I am only capable of writing unsafe code

iron shadow Jul 5, 2023, 12:57 PM

#

KEKW this is what c++ does to you

iron shadow Jul 5, 2023, 1:35 PM

#

data.chunks(component_size * component_count)
        .flat_map(|p| {
            vec![
                p.get(0..component_size).unwrap_or(&[0]).to_vec(),
                p.get((component_size)..(2 * component_size))
                    .unwrap_or(&[0])
                    .to_vec(),
                p.get((component_size * 2)..(3 * component_size))
                    .unwrap_or(&[0])
                    .to_vec(),
                match component_size {
                    2 => (65025u16).to_be_bytes().to_vec(),
                    4 => (4228250625f32).to_be_bytes().to_vec(),
                    _ => vec![255u8],
                },
            ]
            .into_iter()
            .flatten()
        })
        .collect::<Vec<u8>>()

bleakekw

night charm Jul 5, 2023, 2:37 PM

#

I have returned from the glacier

#

I can take a look in a few hours

night charm Jul 5, 2023, 5:31 PM

#

@iron shadow i added methods to the transform matrix

#

the plan is to finish up 0.10 tomorrow and release it

night charm Jul 5, 2023, 6:03 PM

#

iron shadow ```rust data.chunks(component_size * component_count) .flat_map(|p| { ...

why are you writing 65025u16 instead of u16::MAX

torn kettle Jul 5, 2023, 6:06 PM

#

night charm why are you writing `65025u16` instead of `u16::MAX`

KEKW

night charm Jul 5, 2023, 6:07 PM

#

wait thats not the max u16 value

#

is it supposed to be

torn kettle Jul 5, 2023, 6:07 PM

#

indeed

night charm Jul 5, 2023, 6:07 PM

#

KEKW

#

tbh i have no idea what the code above does

iron shadow Jul 5, 2023, 9:43 PM

#

iron shadow Jul 5, 2023, 9:44 PM

#

night charm tbh i have no idea what the code above does

me neither but it converts any given component size to rgba KEKW

night charm Jul 5, 2023, 9:44 PM

#

guh

#

magick

#

alright then

#

well the transform matrix stuff is added

#

a little different from your (now closed) pr but simpler I think

#

btw the old commits are added because i squashed yours when merging

iron shadow Jul 5, 2023, 9:46 PM

#

ohh i see

#

how would i get those commits ommited cuz if i did pr again, it would be annoying to have those prs still merged again

night charm Jul 5, 2023, 9:48 PM

#

overwrite the branch with origin/master

#

simplest non git fuckery way is probably just to delete the fork and make a new one, better is to make each pr on a "patch" branch and sync master periodically to the main repo

iron shadow Jul 5, 2023, 9:49 PM

#

ahh alright will do that later

#

hmm penguin

night charm Jul 5, 2023, 9:50 PM

#

aye

iron shadow Jul 5, 2023, 9:53 PM

#

how would you chain a bunch of commands together? I got something like this

let mut image_commands = ctx
            .execution_manager
            .on_domain::<phobos::domain::Compute>()
            .unwrap();

for image in images {
  image_commands = image_commands
                .copy_buffer_to_image(
                    &staging_buffer.view_full(),
                    &image_phobos.view(vk::ImageAspectFlags::NONE_KHR).unwrap(),
                )
                .unwrap();
}

But i have no clue if this would be considered best practice KEKW and also seems like it does f up everything else as if i wanted to use ctx, it would block it due to it to the top line

#

could i really just make a bunch of commands then do a batch submission? iirc seeing that on the execution manager

night charm Jul 5, 2023, 9:53 PM

#

what youre doing now is also what i do

#

ImageAspectFlags::NONE is suspicious

#

use COLOR

#

also avoid unwrapping as a general best practice, using the ? operator is much nicer

iron shadow Jul 5, 2023, 9:54 PM

#

ah i should probably return a result

night charm Jul 5, 2023, 9:54 PM

#

yeh

#

also seems like it does f up everything else as if i wanted to use ctx, it would block it due to it to the top line
If this is an actual issue you can clone the exec manager and use that, its refcounted internally

torn kettle Jul 5, 2023, 10:03 PM

#

? succinctly tells the reader that you have no idea what is happening in the given code and that the prayerbook should be kept close

iron shadow Jul 5, 2023, 10:04 PM

#

i see it more of a procasination tool : im too lazy to deal with this error, go deal with it later KEKW

#

oh god my image is a black screen bleakekw it isn't copying over correctly

night charm Jul 5, 2023, 10:12 PM

#

torn kettle `?` succinctly tells the reader that you have no idea what is happening in the g...

?? tells the reader this code is a big blunder and should not be taken seriously

night charm Jul 5, 2023, 10:12 PM

#

iron shadow oh god my image is a black screen <:bleakekw:1082598350303539240> it isn't copyi...

yeah you have NONE as your aspect flags

iron shadow Jul 5, 2023, 10:12 PM

#

nono i fixed that

night charm Jul 5, 2023, 10:12 PM

#

oh alright

iron shadow Jul 5, 2023, 10:12 PM

#

i think i fixed it? now it realized i should probably keep my buffers alive as they're dropping before the command buffer has a chance to submit it

night charm Jul 5, 2023, 10:13 PM

#

that is not good yeah

iron shadow Jul 5, 2023, 10:18 PM

#

yooo images finally work

night charm Jul 5, 2023, 10:18 PM

#

https://cdn.discordapp.com/emojis/576487221113258057.webp?size=48&name=Pog&quality=lossless

iron shadow Jul 5, 2023, 10:21 PM

#

so when i submit all those commands, is it doing in parallel or 1 by 1? or is that depends on how the gpu is feeling on that day

night charm Jul 5, 2023, 10:22 PM

#

depends on the weather

iron shadow Jul 5, 2023, 10:22 PM

#

i see i see ill tell god to make it snowy for the gpu

night charm Jul 5, 2023, 10:22 PM

#

if the gpu likes snow sure

#

but you can count on a certain bandwidth of copy being parallel id expect

hybrid pilot Jul 5, 2023, 10:32 PM

#

iron shadow so when i submit all those commands, is it doing in parallel or 1 by 1? or is th...

Comands start in the order you start them, but if commands need the same hw unit they can't always run in parallel

#

Buffer copies are compute shaders on some GPUs (amd)

iron shadow Jul 5, 2023, 10:33 PM

#

that's weirdd

#

oh well seems like nvidia is fine with me putting it on the compute as well

hybrid pilot Jul 5, 2023, 10:34 PM

#

What is weird?

iron shadow Jul 5, 2023, 10:34 PM

#

how it is compute instead of transfer unless im tripping

hybrid pilot Jul 5, 2023, 10:35 PM

#

What I mean is that copies are done by a compute shader by the driver lol

#

But you can run different compute workloads in parallel

night charm Jul 5, 2023, 10:36 PM

#

Most big gpus do have a separate async copy unit

hybrid pilot Jul 5, 2023, 10:36 PM

#

So you can expect transfers to happen in parallel with other stuff

hybrid pilot Jul 5, 2023, 10:36 PM

#

night charm Most big gpus do have a separate async copy unit

Amd uses to have it

night charm Jul 5, 2023, 10:36 PM

#

Not anymore?

hybrid pilot Jul 5, 2023, 10:36 PM

#

I think not

night charm Jul 5, 2023, 10:36 PM

#

Hmm

limber rapids Jul 5, 2023, 10:36 PM

#

Wat

#

AMD does have a DMA queue

#

since probably ATI

night charm Jul 5, 2023, 10:37 PM

#

Well anyway, compute implies transfer support

#

Thats why the compute domain also exposes transfer commands in phobos

hybrid pilot Jul 5, 2023, 10:42 PM

#

limber rapids since probably ATI

~~Well I can't tell for sure what the hw has but I can tell for sure that radv does copies with a compute shader for anything newer than southern islands~~~ no that was bullshit, I misunderstood stuff. The actually code to chose between cp_dma and cs is more involved

#

Some legacy hw lingers on

#

Could still be there dunno

torn kettle Jul 5, 2023, 10:42 PM

#

it does some copies with a compute shader, but amd has dedicated dma hw

#

called sdma

hybrid pilot Jul 5, 2023, 10:43 PM

#

Mmm

torn kettle Jul 5, 2023, 10:44 PM

#

there is also dma hw in the CP

#

and then finally you can use shaders

#

cornucopia of options

hybrid pilot Jul 5, 2023, 10:44 PM

#

Yeah that's what is used in the si path

#

In any case it was just an example

#

To say that just because something looks like dedicated hw in vulkan it might not be

night charm Jul 5, 2023, 10:47 PM

#

Yeah even different queue families don’t necessarily map to different hw

torn kettle Jul 5, 2023, 10:48 PM

#

do you mean queues

iron shadow Jul 5, 2023, 10:50 PM

#

uh dumb question, i'm a bit confused as to how the gltf sampler maps to vulkan?

#

OHH NVM im being brain dead

#

i thought the values represented like pixels but they represent enums

night charm Jul 5, 2023, 10:55 PM

#

Most of them probably map to vk sampler settings

iron shadow Jul 5, 2023, 10:59 PM

#

let phobos_sampler = phobos::Sampler::new(
                    ctx.device.clone(),
                    vk::SamplerCreateInfo {
                        mag_filter: translate_filter(
                            gltf_sampler
                                .mag_filter()
                                .unwrap_or(gltf::texture::MagFilter::Linear),
                        ),
                        min_filter: translate_filter(
                            gltf_sampler
                                .mag_filter()
                                .unwrap_or(gltf::texture::MagFilter::Linear),
                        ),
                        mipmap_mode: vk::SamplerMipmapMode::LINEAR,
                        address_mode_u: translate_wrap_mode(gltf_sampler.wrap_s()),
                        address_mode_v: translate_wrap_mode(gltf_sampler.wrap_t()),
                        address_mode_w: vk::SamplerAddressMode::REPEAT,
                        mip_lod_bias: 0.0,
                        anisotropy_enable: vk::FALSE,
                        max_anisotropy: 0.0,
                        compare_enable: vk::FALSE,
                        compare_op: vk::CompareOp::ALWAYS,
                        min_lod: 0.0,
                        max_lod: 0.0,
                        border_color: vk::BorderColor::INT_OPAQUE_BLACK,
                        unnormalized_coordinates: vk::FALSE,
                        ..Default::default()
                    },
                );

I geniunely have no clue if these are good defaults

night charm Jul 5, 2023, 11:03 PM

#

Make max_lod VK_LOD_CLAMP_NONE I think is what it’s called

#

Enabling anisotropic filtering is probably also a good idea

hybrid pilot Jul 5, 2023, 11:04 PM

#

hybrid pilot ~~Well I can't tell for sure what the hw has but I can tell for sure that radv d...

Ok strike that, I assumed that because the function to encue the copy command started with si it was only used for si 🤦‍♂️
The acwual condition is more complicated

iron shadow Jul 5, 2023, 11:08 PM

#

oh wtf what is the gltf default sampler

hybrid pilot Jul 5, 2023, 11:12 PM

#

I was gonna send you my code to handle the gltf sampler but it looks like I just ignore it

#

Guess you can take inspiration from that :p

iron shadow Jul 5, 2023, 11:13 PM

#

KEKW ill rummage around in the gltf spec to see if they defined it

limber rapids Jul 5, 2023, 11:34 PM

#

Just define your own

#

You most likely want linear filtering and linear mip transitions with anisotropy enabled and repeat wrap mode

languid fern Jul 5, 2023, 11:47 PM

#

Yep that's what I do if not defined

iron shadow Jul 6, 2023, 12:38 AM

#

ah how do you convert a gltf transform to a vulkan one again

#

also how tf do you do textures 💀

#

i want it so i can just refer to textures by their bda

limber rapids Jul 6, 2023, 12:49 AM

#

descriptor indexing

limber rapids Jul 6, 2023, 12:49 AM

#

iron shadow ah how do you convert a gltf transform to a vulkan one again

there is no vulkan transform

iron shadow Jul 6, 2023, 12:58 AM

#

limber rapids descriptor indexing

okok

#

wil look into it

#

I;m a bit confused,..

#

What is the purpose…?

#

So is it a way to bind multiple resources and refer to them as an index in an array?

iron shadow Jul 6, 2023, 2:34 AM

#

hmm how would i do this in phobos

iron shadow Jul 6, 2023, 3:08 AM

#

phobose seems like it can support it, but like i'm not 100% sure if:

im meant to touch those functions
allocation callback (???)

iron shadow Jul 6, 2023, 5:11 AM

#

ahh im so confused on descriptor indexing

hybrid pilot Jul 6, 2023, 7:54 AM

#

iron shadow phobose seems like it *can* support it, but like i'm not 100% sure if: 1. im mea...

Allocations callbacks are for using custom allocators for the host side

iron shadow Jul 6, 2023, 8:02 AM

#

hmm pretty sure phobos doesn't want me to accesss that

#

i wanna see if penguin added an impl

night charm Jul 6, 2023, 8:38 AM

#

What do you need allocation callbacks for thonk

iron shadow Jul 6, 2023, 8:38 AM

#

im trying to do descriptor indexing and its leading me down a path of making a descriptor pool + sets

night charm Jul 6, 2023, 8:38 AM

#

You can do descriptor indexing I just need to add some internal code to set the proper flags

iron shadow Jul 6, 2023, 8:38 AM

#

oh

night charm Jul 6, 2023, 8:39 AM

#

Ill try to add it today

iron shadow Jul 6, 2023, 8:39 AM

#

ill wait then. hmmmm what other things i should do. oh a camera

iron shadow Jul 6, 2023, 8:39 AM

#

night charm Ill try to add it today

take your time bro i feel so bad

night charm Jul 6, 2023, 8:39 AM

#

Nah it’s fine lol

iron shadow Jul 6, 2023, 8:40 AM

#

i should probably also relearn vulkan KEKW
ill do both vkguide using vulkan-hpp then vulkan-tutorial on rust copium that prevents me from going brain dead and ctrl+c + ctrl+v my way through

#

oh my god i spent over 60+ hours in the ide in the past week bleakekw . the addiction cannot be stopped

hybrid pilot Jul 6, 2023, 8:48 AM

#

Wtf that's like 9h a day

iron shadow Jul 6, 2023, 8:49 AM

#

granted a lot of is me afk

#

but also bleakekw

night charm Jul 6, 2023, 8:49 AM

#

bleakekw

iron shadow Jul 6, 2023, 8:49 AM

#

my friends all normal talking about elden ring and im frothing at the renderdoc of it KEKW

night charm Jul 6, 2023, 9:40 AM

#

time to remove buffer usage flags

iron shadow Jul 6, 2023, 10:02 AM

#

breakage 3000 incoming

night charm Jul 6, 2023, 10:03 AM

#

it does break but easy to fix

iron shadow Jul 6, 2023, 10:03 AM

#

How would do you plan on supporting descriptor indexing? I’m so comfused by the process 💀

night charm Jul 6, 2023, 10:03 AM

#

Pretty easy

#

An unbounded array in the shader is seen as a descriptor indexed array using reflection

#

Then there’s just a bind_sampled_image_array function to bind all your images

#

And done

limber rapids Jul 6, 2023, 10:04 AM

#

As long as you non-uniformly index it KEKW

iron shadow Jul 6, 2023, 10:04 AM

#

Ohh so that’s what you plan on adding

#

Oh yeah csn you do the same to buffers or not?

#

Or is this an image exclusive thing?

night charm Jul 6, 2023, 10:05 AM

#

For buffers just use the device address

iron shadow Jul 6, 2023, 10:05 AM

#

Ohh ok wait what?

night charm Jul 6, 2023, 10:05 AM

#

Just send device address and use it in the shader

hybrid pilot Jul 6, 2023, 10:05 AM

#

night charm And done

How do you change individual images?

night charm Jul 6, 2023, 10:06 AM

#

That would involve updating descriptor sets which isnt being done atm

#

Might be worth looking into though

hybrid pilot Jul 6, 2023, 10:07 AM

#

Mmm

iron shadow Jul 6, 2023, 10:07 AM

#

night charm Just send device address and use it in the shader

Ohh what bind function would that be?
Oh wait is it ubo

hybrid pilot Jul 6, 2023, 10:07 AM

#

Wonder how other rendegraphs do it

limber rapids Jul 6, 2023, 10:07 AM

#

iron shadow Ohh what bind function would that be? Oh wait is it ubo

There is no bind

hybrid pilot Jul 6, 2023, 10:07 AM

#

Maybe by making bindless a "first class citizen"

limber rapids Jul 6, 2023, 10:07 AM

#

Bindless in my thing is the only citizen bleakekw

hybrid pilot Jul 6, 2023, 10:08 AM

#

Lol

night charm Jul 6, 2023, 10:08 AM

#

iron shadow Ohh what bind function would that be? Oh wait is it ubo

A new one

iron shadow Jul 6, 2023, 10:08 AM

#

Wait only bindless is really bad?

limber rapids Jul 6, 2023, 10:08 AM

#

The exact opposite, it's ridiculously good

hybrid pilot Jul 6, 2023, 10:08 AM

#

iron shadow Wait only bindless is really bad?

I don't think so

iron shadow Jul 6, 2023, 10:08 AM

#

That’s what I thought, but it also means graphics programmers can get a bit diabolical with vram

hybrid pilot Jul 6, 2023, 10:09 AM

#

It means you don't have to pretend the hw can't di stuff it can do

iron shadow Jul 6, 2023, 10:09 AM

#

OHHHH

hybrid pilot Jul 6, 2023, 10:09 AM

#

Which means you can make better use of it

#

Imho

night charm Jul 6, 2023, 10:09 AM

#

Yeah basically you just get to throw pointers at the gpu and dereference them

limber rapids Jul 6, 2023, 10:10 AM

#

I wish I had access to the whole GPU memory subsystem

#

That way I wouldn't have to readback marked pages and make them resident/evict them

#

I could manage memory directly on the GPU 🥲

iron shadow Jul 6, 2023, 10:10 AM

#

I wish I could write rust on the gpu

hybrid pilot Jul 6, 2023, 10:10 AM

#

limber rapids That way I wouldn't have to readback marked pages and make them resident/evict t...

Is that exposed to userspace at all?

limber rapids Jul 6, 2023, 10:10 AM

#

Through the GAPI yes

#

It's called sparse

night charm Jul 6, 2023, 10:10 AM

#

iron shadow I wish I could write rust on the gpu

Well

#

There is a thing for rust shaders

#

But meh

hybrid pilot Jul 6, 2023, 10:11 AM

#

limber rapids It's called sparse

Sparse binding?

limber rapids Jul 6, 2023, 10:11 AM

#

yeah

hybrid pilot Jul 6, 2023, 10:11 AM

#

Yeah I'm aware of that

#

I was talking of managing memory from the GPU

#

Lol

limber rapids Jul 6, 2023, 10:11 AM

#

Ah

#

no way KEKW

hybrid pilot Jul 6, 2023, 10:11 AM

#

Yeah

#

It would be a security hole I guess

limber rapids Jul 6, 2023, 10:12 AM

#

I do not care

#

I will throw away all security for 2% more convenience

hybrid pilot Jul 6, 2023, 10:12 AM

#

But maybe in principle the GPU could write it'sn own page table?

limber rapids Jul 6, 2023, 10:12 AM

#

yes

hybrid pilot Jul 6, 2023, 10:12 AM

#

As long as you map it

#

But you would have to hack the kernel driver to do that I guess

#

Like maybe map it to an hardcoded address then access it from a shader

#

So you don't have to touch the rest of the stack

night charm Jul 6, 2023, 10:13 AM

#

Reminds me of gpu exceptions

#

from martty’s blog

iron shadow Jul 6, 2023, 10:16 AM

#

night charm There is a thing for rust shaders

Oh wtf rustgpu

#

That looks cool, is it any good?

hybrid pilot Jul 6, 2023, 10:16 AM

#

We where discussing it

#

Outside of here

#

Seems like it lacks essential featured for compute shaders

#

(Barriers)

#

But it is promising at least

limber rapids Jul 6, 2023, 10:16 AM

#

hybrid pilot (Barriers)

wtf

night charm Jul 6, 2023, 10:18 AM

#

Yeah I personally wouldn’t use it

hybrid pilot Jul 6, 2023, 10:18 AM

#

limber rapids wtf

What's not clear?

limber rapids Jul 6, 2023, 10:18 AM

#

me when
me when no GL_KHR_shader_subgroup_ballot

hybrid pilot Jul 6, 2023, 10:18 AM

#

Lol

limber rapids Jul 6, 2023, 10:19 AM

#

hybrid pilot What's not clear?

it's not clear how rustgpu doesn't have barrier()

#

third class language

#

smh

hybrid pilot Jul 6, 2023, 10:19 AM

#

limber rapids it's not clear how rustgpu doesn't have `barrier()`

It doesn't have it yet

#

Should have it eventually

#

No idea why it doesn't have then yet

night charm Jul 6, 2023, 10:20 AM

#

Seems pretty essential yeah

#

I assume BDA is also not implemented

limber rapids Jul 6, 2023, 10:20 AM

#

I don't understand why features like these aren't given absolute maximum priority

#

BDA is literally necessary to move on from the garbage "bind and use" model

iron shadow Jul 6, 2023, 10:21 AM

#

Petition to deprecate bin-

night charm Jul 6, 2023, 10:22 AM

#

One day

torn kettle Jul 6, 2023, 10:22 AM

#

iron shadow Petition to deprecate bin-

(he got shot by nvidia snipers)

iron shadow Jul 6, 2023, 10:22 AM

#

LMAOOO is nvidia that adamant on bind and use

hybrid pilot Jul 6, 2023, 10:23 AM

#

torn kettle (he got shot by nvidia snipers)

Does nvidia hw like binds?

#

I wanted to look into how descriptor sets work on nvk and never got around to it

torn kettle Jul 6, 2023, 10:25 AM

#

i was just doing a meme, but nv descriptors are a bit more convoluted than amd ones

iron shadow Jul 6, 2023, 10:26 AM

#

nvk bleakekw

#

All I can remember was jerry rigging that to work so I could submit my raytracing cs project

hybrid pilot Jul 6, 2023, 10:28 AM

#

I don't think nvk does rt

iron shadow Jul 6, 2023, 10:34 AM

#

It doesnt i thik that was nvh

#

But iirc nvk is also used in their tutorial

#

Which was what i built off

torn kettle Jul 6, 2023, 10:35 AM

#

nvk doesn't really work yet

#

you are talking about something else

iron shadow Jul 6, 2023, 10:37 AM

#

Ohh i think it was nvvk then

#

Ah it was nvh

night charm Jul 6, 2023, 11:04 AM

#

https://github.com/NotAPenguin0/phobos-rs/releases/tag/v0.10.0 new release

GitHub

Release v0.10.0 · NotAPenguin0/phobos-rs

Changes

Phobos no longer requires Rust nightly to compile, except when using FSR2.
change: don't poll in fence implementation by @pac85 in #56
Implement VK_EXT_debug_utils naming by @DannyLeWa...

iron shadow Jul 6, 2023, 11:07 AM

#

Poggies

night charm Jul 6, 2023, 11:07 AM

#

ill upgrade the egui integration later today

iron shadow Jul 6, 2023, 11:07 AM

#

So can I do

#

Descriptor indexing?

night charm Jul 6, 2023, 11:08 AM

#

thats next on the list

#

but i wanted to get this release out already

iron shadow Jul 6, 2023, 11:08 AM

#

Fair

iron shadow Jul 6, 2023, 7:42 PM

#

Hey Pengiun, you got an idea why ash and phobos' ash includes it exposes just are never detected by my IDE?

#

Do I need to do static linking?

hybrid pilot Jul 6, 2023, 8:01 PM

#

? It is like the default in rust.
What do you mean they aren't detected?

night charm Jul 6, 2023, 8:03 PM

#

huh

#

I re-export ash as phobos::vk

#

you can just use phobos::vk and use that

iron shadow Jul 6, 2023, 8:03 PM

#

Yeah, but it's really weird as there is like no autocomplete whatsoever

#

like it won't recognize classes exist

#

i tried importing regular ash too

#

and the issue persisit which is weird

night charm Jul 6, 2023, 8:04 PM

#

thats weird

#

it works for me

#

takes some time to index sometimes but it works eventually

iron shadow Jul 6, 2023, 8:08 PM

#

bleakekw time to do a hard ide reset

iron shadow Jul 6, 2023, 8:27 PM

#

like it's not detecting types like vk::Bool32

#

is there a way to force clion to index it?

hybrid pilot Jul 6, 2023, 9:04 PM

#

Use vim

night charm Jul 6, 2023, 9:14 PM

#

KEKW

iron shadow Jul 6, 2023, 9:47 PM

#

bleakekw

#

ping me when support for descriptor indexing and giving shaders buffers via bda is added to phobos

limber rapids Jul 6, 2023, 9:50 PM

#

All you need for bda is GL_EXT_buffer_reference

night charm Jul 6, 2023, 9:50 PM

#

you can do bda already

limber rapids Jul 6, 2023, 9:50 PM

#

And vkGetBufferDeviceAddress

night charm Jul 6, 2023, 9:50 PM

#

a bda is just a uint

#

just send it through a push constant or something

#

or in a ubo if you have lots of them

#

but you can literally just write the return value of address() to some memory visible in the shader and use it like you would in glsl

iron shadow Jul 6, 2023, 9:51 PM

#

oh 💀 so im just realistically waiting for the descriptor indexing for textures

night charm Jul 6, 2023, 10:02 PM

#

Pretty much

#

Hopefully I get around to it tomorrow

iron shadow Jul 6, 2023, 10:04 PM

#

wait hm how would I synchronize all those buffers?

#

like "No one else is allowed to write to this until X is finished reading" like a RwLock

night charm Jul 6, 2023, 10:08 PM

#

Inside the shader with a barrier(), outside you have to add barriers yourself

iron shadow Jul 6, 2023, 10:10 PM

#

OHH memory_barrier that makes sense

limber rapids Jul 6, 2023, 10:11 PM

#

Disclaimer: barrier only synchronizes a workgroup

iron shadow Jul 6, 2023, 10:13 PM

#

workgroup?

#

ohh compute? I think the ray tracing shaders still are covered

#

unless I'm wrong

night charm Jul 6, 2023, 10:14 PM

#

Probably ub

limber rapids Jul 6, 2023, 10:14 PM

#

you can use subgroup vote to synchronize in shaders other than compute (or mesh/task)

#

if (subgroupElect()) {
    write();
}
subgroupBarrier();

iron shadow Jul 6, 2023, 10:16 PM

#

shader code I assume?

#

hmmm okok

limber rapids Jul 6, 2023, 10:17 PM

#

Yes

#

You can't effectively synchronize all threads

#

You can only synchronize within a subgroup, so be careful

iron shadow Jul 6, 2023, 10:18 PM

#

im so confused by these words 😭 I havent done any compute

limber rapids Jul 6, 2023, 10:18 PM

#

Don't worry about it until you need it

iron shadow Jul 6, 2023, 10:20 PM

#

wait so when doing bda for my buffers, does this bind my buffers?

#

or is the shader seriously just reading it from gpu memory using that address

limber rapids Jul 6, 2023, 10:20 PM

#

it is a pointer

#

to god knows where in the GPU

iron shadow Jul 6, 2023, 10:21 PM

#

KEKW

limber rapids Jul 6, 2023, 10:21 PM

#

There is no binding

iron shadow Jul 6, 2023, 10:21 PM

#

huh okay so all this descriptor binding stuff ive been doing was mainly to just make a pointer?

#

sorry for asking such basic questions 💀

limber rapids Jul 6, 2023, 10:22 PM

#

Let me give you a basic rundown on how BDA actually works

#

The classic binding model involves the following

vkCreateDescriptorSetLayout(&set_layout);
// you allocate a descriptor set to *bind* to the GPU, a descriptor set is a fancy pointer to resources which right now is nullptr
vkAllocateDescriptorSets(&set, &set_layout);
// Now you *write* to the descriptor set with your resources
vkUpdateDescriptorSets(&set, 1, &write);
// Then you finally *bind* the set
vkCmdBindDescriptorSets(1, &set);```

#

We call this bindful because we need to bind descriptor sets

#

BDA works like this:

// In the shader
layout (buffer_reference) buffer X {
    LotsOfStuff[] data;
};

layout (push_constant) {
    uint64_t address;
};

void main() {
    X ptr = X(address);
}```

#

And in your vulkan app you just do a shrimple

auto address = buffer.address();
vkCmdPushConstants(1, &address);```

#

There is no vkCmdBind anything

torn kettle Jul 6, 2023, 10:29 PM

#

limber rapids you can use subgroup vote to synchronize in shaders other than compute (or mesh/...

this seems suspect to me

limber rapids Jul 6, 2023, 10:30 PM

#

frog_sweat

iron shadow Jul 6, 2023, 10:46 PM

#

limber rapids There is no vkCmdBind anything

ohhh that's why so many developers live and die by bind less

#

what if you wanted to push an arbitrary array of them? You simply just change it to?

layout (push_constant) {
    uint64_t[] addresses;
};

?

limber rapids Jul 6, 2023, 10:47 PM

#

no

#

if you need a lot of addresses you should first consider your life choices and then you make a UBO

iron shadow Jul 6, 2023, 10:48 PM

#

KEKW but my rt shaders can't know which vertex buffer is gonna actually be used

limber rapids Jul 6, 2023, 10:49 PM

#

Make only one vertex buffer

#

If not, have an object list with each object having a uint64 address to its vertex buffer

limber rapids Jul 6, 2023, 10:51 PM

#

torn kettle this seems suspect to me

What is sus sir

hybrid pilot Jul 6, 2023, 11:39 PM

#

night charm but you can literally just write the return value of `address()` to some memory ...

So those can't be tracked by the graph I imagine

#

Or can you explicitly say "I'm gonna need this buffer but don't bind it cuz I'll just send the address"

iron shadow Jul 6, 2023, 11:44 PM

#

oh real

hybrid pilot Jul 6, 2023, 11:45 PM

#

Though you'd probably use buffer address for persistent stuff where sync is not necessarily a concern

iron shadow Jul 6, 2023, 11:53 PM

#

yeah but what if my scene changes and stuff

#

add, removing, idfk about culling as i have no clue how that works exactly

limber rapids Jul 6, 2023, 11:55 PM

#

If you make incoherent read/writes that could even change from subgroup to subgroup, you need atomics/locks

#

But this scenario is never going to happen tbh

#

If you have a specific example I might give you more details/realistic approach

iron shadow Jul 7, 2023, 1:17 AM

#

Rn it's mainly jsut to ray trace a scene that is gonna be static for a bit

#

but I wanna do physics simulations with gravity where many objects are gonna be transformed as a result of it

iron shadow Jul 7, 2023, 2:55 AM

#

but idfk like if I removed an object from my scene, I should remove it from the device buffer as well

#

but I'm just unsure if that wouldn't be optimal as I think I would need to more or less restream the data from the cpu back to the gpu but now slightly changed

#

Hold on the neurons are firing what if I did an "inverted slice" of the removed memory, copy all data to the left and right of it into one massive new buffer I have allocated for the new size? My worry then is mainly sync issue as what if I have an AS or some other object that is referencing my buffer, but now that information is old?

night charm Jul 7, 2023, 8:15 AM

#

hybrid pilot Or can you explicitly say "I'm gonna need this buffer but don't bind it cuz I'll...

You can do that

hybrid pilot Jul 7, 2023, 8:16 AM

#

night charm You can do that

Fantastic!

#

So if you use a buffer bound that way across passes you can get it synced for you

#

Really nice

night charm Jul 7, 2023, 8:16 AM

#

Yep

night charm Jul 7, 2023, 8:17 AM

#

iron shadow Hold on the neurons are firing what if I did an "inverted slice" of the removed ...

You’re just reinventing an allocator at this point bleakekw

#

Look up some algorithms for that

hybrid pilot Jul 7, 2023, 8:17 AM

#

Sparse bindings make allocation so much easier

night charm Jul 7, 2023, 8:17 AM

#

It’ll give some inspiration

#

Yeah that’s true

limber rapids Jul 7, 2023, 8:22 AM

#

There are complex ways of solving this

hybrid pilot Jul 7, 2023, 8:22 AM

#

Yeah not a trivial problem

limber rapids Jul 7, 2023, 8:22 AM

#

But really this is so much easier than you think

#

dispatch(); //physics shader, writes to many buffers
barrier();
build_tlas();
barrier();
trace_rays();```

#

That's it

#

This is all host side btw

#

If you write from the host to a buffer used by the gpu it's even easier

night charm Jul 7, 2023, 8:25 AM

#

Yeah rebuild your tlas every frame

#

Its pretty cheap

limber rapids Jul 7, 2023, 8:25 AM

#

all you have to do is allocate some HOST_COHERENT memory

#

Yeah, you could implement caching, virtual memory allocators, sparse resources, write to and from a page table from the GPU, request and mark pages asynchronously and all the ridiculously complicated stuff you want

#

But just, don't? KEKW

night charm Jul 7, 2023, 1:09 PM

#

Probably no updates for a week or so

#

My laptop decided to delete its own kernel

#

And I didn’t bring my installer usb

hybrid pilot Jul 7, 2023, 1:25 PM

#

night charm And I didn’t bring my installer usb

You are a penguin, so you are linux. Stick your finger in the USB port and boot yourself

night charm Jul 7, 2023, 1:25 PM

#

bleakekw right

#

ill try that

#

Maybe I can buy one here but shits expensive here and I already bought a new phone charger

hybrid pilot Jul 7, 2023, 1:26 PM

#

Phobs needs more founding

#

BTW how did it happen?

#

Update?

night charm Jul 7, 2023, 1:32 PM

#

hybrid pilot BTW how did it happen?

I tried to install a new graphics driver

#

And it failed and decided to take the kernel with it

hybrid pilot Jul 7, 2023, 1:33 PM

#

night charm I tried to install a new graphics driver

Uhm do you use nvidia?

night charm Jul 7, 2023, 1:34 PM

#

I have a nvidia card in there but it doesn’t work with my compositor so I generally use the igpu

#

The driver was also for the igpu

hybrid pilot Jul 7, 2023, 1:34 PM

#

Mmm

iron shadow Jul 7, 2023, 8:54 PM

#

limber rapids ```cpp dispatch(); //physics shader, writes to many buffers barrier(); build_tla...

Huh yeah i forgot that tlas building is fast blas is a bit slower, but enough that animations could be donw

iron shadow Jul 7, 2023, 10:47 PM

#

ah wait I realized

#

how would I even identfiy the BLAS in my shader so I would know which buffer it refers to

limber rapids Jul 7, 2023, 10:49 PM

#

The TLAS identifies the BLAS

iron shadow Jul 7, 2023, 10:57 PM

#

gl_InstanceCustomIndexEXT ?

limber rapids Jul 7, 2023, 10:57 PM

#

no

#

the TLAS takes a BLAS as an argument

iron shadow Jul 7, 2023, 11:04 PM

#

oh wait it is gl_InstanceCustomIndexEXT

#

I literally just needed a way to determine which BLAS i was hitting in my shader

limber rapids Jul 7, 2023, 11:04 PM

#

why

#

You could use gl_InstanceCustomIndexEXT but why do you need to do that

iron shadow Jul 7, 2023, 11:05 PM

#

Because if I'm passing my vertex buffers into my shader, I need to be able to

#

determine which vertex buffer corresponds to which blas

night charm Jul 7, 2023, 11:05 PM

#

Yes you can use that value to identify mesh/object properties

#

The nvidia rt sample does this

iron shadow Jul 7, 2023, 11:06 PM

#

idk im generally following what NVIDIA is doing for their shaders

#

yeah nvidia's rt sample

limber rapids Jul 7, 2023, 11:06 PM

#

I still don't get why you need the BLAS

#

why can't you use gl_InstanceCustomIndexEXT to identify an object, which has the address to your vertex and index buffers

night charm Jul 7, 2023, 11:07 PM

#

That’s exactly the idea

#

Just poor wording of it KEKW

iron shadow Jul 7, 2023, 11:07 PM

#

KEKW

limber rapids Jul 7, 2023, 11:07 PM

#

epic

iron shadow Jul 7, 2023, 11:07 PM

#

step 2) camera

night charm Jul 7, 2023, 11:08 PM

#

Are you using rt pipelines or ray queries btw

limber rapids Jul 7, 2023, 11:09 PM

#

I will personally uninstall ray queries from the driver if anyone is using them

iron shadow Jul 7, 2023, 11:09 PM

#

raytracing pipelines

night charm Jul 7, 2023, 11:09 PM

#

They’re great wdym

limber rapids Jul 7, 2023, 11:09 PM

#

no

iron shadow Jul 7, 2023, 11:09 PM

#

iirc whatever the example is

night charm Jul 7, 2023, 11:09 PM

#

Ray query is epic

limber rapids Jul 7, 2023, 11:09 PM

#

CPU readback is cringe

night charm Jul 7, 2023, 11:10 PM

#

iron shadow raytracing pipelines

Right, quick note about that. SBT is currently fully generated because I don’t know of any reason not to. If you come across any limitation with this lmk and I’ll rework a little

night charm Jul 7, 2023, 11:10 PM

#

limber rapids CPU readback is cringe

Why do you need cpu readback

#

Just ray query in the fragment shader for shadows and stuff

#

Its great

#

Works nicely

limber rapids Jul 7, 2023, 11:10 PM

#

You see I can't read

night charm Jul 7, 2023, 11:10 PM

#

bleakekw

iron shadow Jul 7, 2023, 11:11 PM

#

bleakekw

#

me neither dw bro

limber rapids Jul 7, 2023, 11:11 PM

#

For some reason I thought you meant occlusion queries

#

I have no idea why

night charm Jul 7, 2023, 11:11 PM

#

Ah

#

nono

#

Ray query good

limber rapids Jul 7, 2023, 11:11 PM

#

The compute rasterizer brain worm ate my brain

#

Only 5% remains

night charm Jul 7, 2023, 11:11 PM

#

It did crash my driver for some inexplicable reason but that aside

#

Nvidia shader compiler nonsense

iron shadow Jul 7, 2023, 11:12 PM

#

so like any articles on how to do flying camera. learn opengl is fien but i can't go upside down :(

limber rapids Jul 7, 2023, 11:12 PM

#

Ah yes, I have crashed NVIDIA's internal compiler countless times

limber rapids Jul 7, 2023, 11:12 PM

#

iron shadow so like any articles on how to do flying camera. learn opengl is fien but i can'...

you have a forward vector and an up vector, cross them and you get your right vector, slam that shit into glm lookAt

iron shadow Jul 7, 2023, 11:12 PM

#

KEKW

limber rapids Jul 7, 2023, 11:12 PM

#

you now have a fps camera

night charm Jul 7, 2023, 11:12 PM

#

This is just speculation but it seems like it starts optimizing the shader after a few frames on a background thread, and just crashing inside that somewhere

limber rapids Jul 7, 2023, 11:13 PM

#

night charm This is just speculation but it seems like it starts optimizing the shader after...

Incredible

night charm Jul 7, 2023, 11:13 PM

#

Truly

limber rapids Jul 7, 2023, 11:13 PM

#

deferred optimization

#

truly a sentence I had hoped to never utter

night charm Jul 7, 2023, 11:13 PM

#

Its not unheard of apparently

#

You profile it for some frames and optimize based on that data

limber rapids Jul 7, 2023, 11:13 PM

#

Perhaps I'm too small brained

#

I don't know what kind of heuristics you can get from on the fly profiling KEKW

#

Perhaps instruction or branch reordering idk

night charm Jul 7, 2023, 11:15 PM

#

Not sure either

#

But I can imagine the big brain compiler engineers know

limber rapids Jul 7, 2023, 11:15 PM

#

yeah

night charm Jul 7, 2023, 11:16 PM

#

I got some pointers from pixelduck on what could cause this but haven’t found a fix

iron shadow Jul 7, 2023, 11:18 PM

#

fellas we have a problem

#

which direction should i make up

limber rapids Jul 7, 2023, 11:18 PM

#

0,1,0

#

no discussions are allowed

iron shadow Jul 7, 2023, 11:19 PM

#

KEKW the greatest debate of time

limber rapids Jul 7, 2023, 11:19 PM

#

Maybe if you do physics you can do 0,0,1

#

Anything else is forbidden by international law

iron shadow Jul 7, 2023, 11:20 PM

#

nah bro wtdym

#

z-buffering 💀 why tf would you also make that up

#

hmm

limber rapids Jul 7, 2023, 11:20 PM

#

z-up is quite common

iron shadow Jul 7, 2023, 11:20 PM

#

0,1,0 we do

limber rapids Jul 7, 2023, 11:20 PM

#

Blender does z-up, Unreal too

iron shadow Jul 7, 2023, 11:21 PM

#

iirc there was a nice image explaining all that stuff

#

where tf was it

#

KEKW cant find it

#

which system does godot use? lhs? whatever ill pick opengl and do rhs

limber rapids Jul 7, 2023, 11:24 PM

#

Do whatever is convention for the API you are using

#

which is right handed, y up, -z forward and x right

iron shadow Jul 7, 2023, 11:25 PM

#

i still think the potential of (1,0,0) to mean up is still there KEKW

night charm Jul 7, 2023, 11:30 PM

#

y up or I make more breaking changes in Phobos

iron shadow Jul 7, 2023, 11:32 PM

#

KEKW

#

oh shoot i need to update phobos

iron shadow Jul 8, 2023, 1:02 AM

#

oh no

languid fern Jul 8, 2023, 1:04 AM

#

ohno.png

iron shadow Jul 8, 2023, 2:15 AM

#

ohno+panic.png

#

i uh

#

oh no

night charm Jul 8, 2023, 6:30 AM

#

I even wrote some instructions in the release notes

iron shadow Jul 8, 2023, 7:37 AM

#

i fixed the issue 😎

#

camera working now

#

i can now view suzanne

#

https://tenor.com/view/yippee-happy-celebration-joy-confetti-gif-25557730

Tenor

night charm Jul 8, 2023, 8:00 AM

#

Epic

night charm Jul 8, 2023, 10:32 AM

#

usb acquired

#

I can fix my laptop later today

iron shadow Jul 8, 2023, 11:09 AM

#

Yoo pog

night charm Jul 8, 2023, 2:04 PM

#

ladies and gentlemen i once again have a kernel

iron shadow Jul 8, 2023, 2:37 PM

#

This is big

#

All you have to do is rewrite the rest of it

night charm Jul 8, 2023, 2:56 PM

#

bleakekw

iron shadow Jul 9, 2023, 7:18 AM

#

i have not been able to go cg in the past all day bleakekw i'm not even in my uni yet and a club got me to code a discord bot for them KEKW

#

how am i busy with uni already and i havent even started my first year its actually joever how do i phobo now

night charm Jul 9, 2023, 7:40 AM

#

My laptop died again

#

New chapter in the penguin laptop saga begins

#

laptop refuses to start at all for a whole night

iron shadow Jul 9, 2023, 7:51 AM

#

oh shoot i need one too for university

night charm Jul 9, 2023, 7:51 AM

#

The next morning grub has disappeared

#

okay

iron shadow Jul 9, 2023, 7:51 AM

#

night charm laptop refuses to start at all for a whole night

How old is it?

night charm Jul 9, 2023, 7:51 AM

#

4 years or so

iron shadow Jul 9, 2023, 7:52 AM

#

🫡 4 years

#

which one do you plan on getting?

night charm Jul 9, 2023, 7:52 AM

#

I can fix this

iron shadow Jul 9, 2023, 7:52 AM

#

oh?

night charm Jul 9, 2023, 7:52 AM

#

I can get into boot options now

#

So I can boot my arch usb again and reinstall grub

#

I should just give up on getting factorio to run

iron shadow Jul 9, 2023, 7:54 AM

#

so is this hardware or software

#

cuz if it is failing this much

#

i doubt it's software then

night charm Jul 9, 2023, 7:54 AM

#

No idea

iron shadow Jul 9, 2023, 7:56 AM

#

Good luck then 🫡 idk if it keeps happening im pretty its hardware at that point

#

or some really f'd up driver

night charm Jul 9, 2023, 8:00 AM

#

probably the latter

#

Im stuck using my igpu because wayland doesn’t like nvidia

iron shadow Jul 9, 2023, 8:06 AM

#

OHH

#

Linux + NVIDIA KEKW

#

Yeah good luck

#

Literally the only reason why I probably can't go over to linux rn:

i game a lot
nvidia

#

Wait so if I am doing BDA, how would indicate to phobos that I tend to use these buffers so it can generate the necessary fences/barriers

night charm Jul 9, 2023, 8:27 AM

#

Same as any buffer

#

you just add it to the pass

#

But you don’t need to do any sync if they’re readonly

#

Since there are no buffer layouts or whatever

iron shadow Jul 9, 2023, 8:28 AM

#

Ah alright

#

so I would be fine by just not even binding it?

#

so anything that would modify that buffer, would sync'd

night charm Jul 9, 2023, 8:49 AM

#

Yeah the resource sync and declaration is completely separate from binding

iron shadow Jul 9, 2023, 8:49 AM

#

wait so what is the point of declaring your bindings then?

night charm Jul 9, 2023, 8:51 AM

#

Its declaring resource usages

iron shadow Jul 9, 2023, 8:52 AM

#

oh like from vulkan prespective

night charm Jul 9, 2023, 8:52 AM

#

Not bindings specifically

#

But declaring what and how you will use a resource allows the system to sync it

#

Its up to you how you “bind”/access the resource, as long as it’s within the declared usage

#

Yeah

iron shadow Jul 9, 2023, 8:52 AM

#

ohh that makes sense

iron shadow Jul 9, 2023, 12:13 PM

#

hey penguin how much effort would it take for me to implement descriptor indexing

iron shadow Jul 9, 2023, 12:37 PM

#

oh i realized this was so much easier.I was gonna make a system where i had an object struct which described the index of the bda ubos, but then i realized

#

i could just place my bdas in place of those indices bleakekw

night charm Jul 9, 2023, 1:18 PM

#

Might try to implement descriptor indexing today

#

Or tomorrow

#

It’s not a lot of work but needs some annoying things

limber rapids Jul 9, 2023, 1:22 PM

#

You need to VkDescriptorBindingFlags~~ VARIABLE_DESCRIPTOR_COUNT~~ and PARTIALLY_BOUND when creating the descriptor layout and VkDescriptorSetVariableDescriptorCountAllocateInfo when allocating the set

torn kettle Jul 9, 2023, 1:24 PM

#

you don't need variable count

limber rapids Jul 9, 2023, 1:25 PM

#

my bad

iron shadow Jul 9, 2023, 1:25 PM

#

KEKW

torn kettle Jul 9, 2023, 1:28 PM

#

you in fact don't even need partially bound either KEKW , but that does make things a bit nicer

iron shadow Jul 9, 2023, 1:58 PM

#

oh no

night charm Jul 9, 2023, 3:19 PM

#

I was pinged

iron shadow Jul 9, 2023, 3:19 PM

#

false information

#

i was asking a question then i realized it was stupid and deleted it

night charm Jul 9, 2023, 3:22 PM

#

I see

hybrid pilot Jul 9, 2023, 3:26 PM

#

What will descriptor indexing look like in phobos?

#

Like what kind of api will it expose for it?

iron shadow Jul 9, 2023, 3:38 PM

#

Yeah that ^

#

also I'm gonna evangelize my university to use phobos KEKW

#

should i convince the guy who is making a vercel physics engine to join phobos

iron shadow Jul 9, 2023, 4:01 PM

#

Hmm how would I enable an extension in phobos

#

I need specifically: GL_EXT_scalar_block_layout

limber rapids Jul 9, 2023, 4:03 PM

#

You enable them in the shader

#

#extension NAME : enable/require```

iron shadow Jul 9, 2023, 4:17 PM

#

huh yeah

#

i don't need

#

OHH

#

IT WAS FOR SHADER INT 64

#

#extension GL_EXT_shader_explicit_arithmetic_types_int64 : require

#

I'm so confused

#

let mut settings = phobos::AppBuilder::new()
            .version((1, 0, 0))
            .name(name)
            .validation(true)
            .present_mode(phobos::vk::PresentModeKHR::MAILBOX)
            .scratch_size(1 * 1024u64)
            .gpu(phobos::GPURequirements {
                dedicated: false,
                min_video_memory: 1 * 1024 * 1024 * 1024, // 1 GiB
                min_dedicated_video_memory: 1 * 1024 * 1024 * 1024,
                queues: vec![
                    phobos::QueueRequest {
                        dedicated: false,
                        queue_type: phobos::QueueType::Graphics,
                    },
                    phobos::QueueRequest {
                        dedicated: true,
                        queue_type: phobos::QueueType::Transfer,
                    },
                    phobos::QueueRequest {
                        dedicated: true,
                        queue_type: phobos::QueueType::Compute,
                    },
                ],
                features: vk::PhysicalDeviceFeatures::builder().shader_int64,
                ..Default::default()
            });

Doesn't work for obvious reasons

#

shader_int64 returns bool32

limber rapids Jul 9, 2023, 4:23 PM

#

You need to set that to true

iron shadow Jul 9, 2023, 4:23 PM

#

yeah for soem reason

limber rapids Jul 9, 2023, 4:23 PM

#

Sorry, wrong question mark lol

iron shadow Jul 9, 2023, 4:23 PM

#

i solved it lmaoo

#

now i have a descriptor type error bleakekw

#

also can soemone explain to me

#

why the fuck does it take like 90 good seconds

#

to the vulkan docs to load

torn kettle Jul 9, 2023, 4:27 PM

#

gnomes

limber rapids Jul 9, 2023, 4:38 PM

#

bookmark this https://registry.khronos.org/vulkan/specs/1.3-extensions/html/

iron shadow Jul 9, 2023, 4:49 PM

#

Hey uh penguin

#

so if I were to set the extensions variable in the gpu settings

#

would it like mess up everything i.e. rt extension loading, validation, etc.

#

since i would be overriding the default values?

night charm Jul 9, 2023, 5:06 PM

#

I think you just want to .push() values into it

night charm Jul 9, 2023, 5:07 PM

#

hybrid pilot Like what kind of api will it expose for it?

.bind_sampled_image_array() with a slice of images and a sampler will bind all those to 0..n in the shader

#

Whether a binding is using descriptor indexing or not is checked using reflection

hybrid pilot Jul 9, 2023, 5:08 PM

#

Nice

#

When I'll rewrite my renderer it will be bindless from the start

night charm Jul 9, 2023, 5:09 PM

#

Bindless is great

#

You do need to set a max value for the amount of descriptors in such an unbounded array but I’m thinking to just provide that on init

#

And default to like 4k

limber rapids Jul 9, 2023, 6:16 PM

#

mr HLSL enthusiast

#

What does this ternary mean? KEKW
https://github.com/EpicGames/UnrealEngine/blob/ue5-early-access/Engine/Shaders/Private/Nanite/Rasterizer.usf#L155C1-L155C1

iron shadow Jul 9, 2023, 6:17 PM

#

oh no

#

my condolenseces 🫡 epic games

night charm Jul 9, 2023, 6:39 PM

#

my laptop be back

night charm Jul 9, 2023, 6:40 PM

#

limber rapids What does this ternary mean? <:KEKW:666849321462792234> <https://github.com/Epi...

bleakekw

#

no idea

#

rcp might be 1/x?

limber rapids Jul 9, 2023, 6:41 PM

#

ye but I mean

#

What does vecN ? x : y mean

#

all(vecN) ? x : y?

#

Or any?

night charm Jul 9, 2023, 7:07 PM

#

ah

#

Probably all

#

I think the zero is casted to float3(0,0,0)

languid fern Jul 11, 2023, 12:05 AM

#

Ok so I might be using phobos in the near future...

#

Is it stable/usable for rasterization/compute yet?

iron shadow Jul 11, 2023, 12:48 AM

#

We're mainly waiting for descriptor indexing for images

#

other than that, Penguin is using phobos in his terrain project iirc

#

so it should be?

#

It's ready for ray tracing 100%

languid fern Jul 11, 2023, 12:51 AM

#

Oh nice

#

I dont even have an RT card lol so I'm mostly going to be using it for game engine dev

#

Are you one of the maintainers or just using it?

#

If the code isn't too complex (like wgpu's imo) I might be able to help out if I find bugs / issues with it
As long as I can rasterize / compute stuff like I can in wgpu I'll definitely use it

iron shadow Jul 11, 2023, 1:18 AM

#

I'm just using it rn

night charm Jul 11, 2023, 6:15 AM

#

Definitely stable-ish yeah

iron shadow Jul 11, 2023, 6:28 AM

#

Oh yeah penguin is very open and chill about you contributing to the project
it'd be cool if we made this project community developed than solely penguin developed KEKW i'm relearning vulkan so i can help out more meaningfully

night charm Jul 11, 2023, 6:28 AM

#

heh im always open to take prs

night charm Jul 11, 2023, 6:58 AM

#

languid fern If the code isn't too complex (like wgpu's imo) I might be able to help out if I...

I think apart from the render graph internals, the code isn’t too complex

#

I might rewrite that part at some point in time, it truly is a beast

iron shadow Jul 11, 2023, 6:59 AM

#

i wanna learn how a render graph works KEKW

night charm Jul 11, 2023, 6:59 AM

#

Me too

#

bleakekw

iron shadow Jul 11, 2023, 6:59 AM

#

bleakekw

#

all i understand is that it's a nice way to abstract resource bindings

night charm Jul 11, 2023, 6:59 AM

#

Mine only abstracts sync

#

It doesn’t do any binding or resource management/aliasing

iron shadow Jul 11, 2023, 7:00 AM

#

ohh i see

#

i'm mainly waiting for vulkan to add the task graph from dx12

#

as it looks insane

torn kettle Jul 11, 2023, 7:14 AM

#

.. to debug

languid fern Jul 11, 2023, 7:14 AM

#

iron shadow i'm mainly waiting for vulkan to add the task graph from dx12

There's going to be support for what now??

languid fern Jul 11, 2023, 7:14 AM

#

torn kettle .. to debug

Oh...

iron shadow Jul 11, 2023, 7:16 AM

#

bleakekw

languid fern Jul 11, 2023, 2:32 PM

#

night charm Definitely stable-ish yeah

Oh alright cool!

#

I already have a good front-end API rn that uses WGPU as a back end so if I could just swap out the backend for phobos it should be pretty easy

languid fern Jul 11, 2023, 3:06 PM

#

Ohh wait phobos allows you to do uploads using a transfer only queue? That's amazing

night charm Jul 11, 2023, 3:15 PM

#

yep

#

well, its a "transfer domain", which maps to a transfer only queue if possible

#

If thats not possible it shares a queue

languid fern Jul 11, 2023, 3:15 PM

#

Awesome

#

There are a few missing things that I'm gonna need like vkCmdDrawIndexedIndirectCount but they shouldn't be hard to implement

#

I could just make a PR

#

Also I love how you actually made use of the trait system for "Domains"

#

Really refreshing to see a vulkan API actually make use of traits properly

night charm Jul 11, 2023, 3:19 PM

#

froge

#

yeah some commands are missing still

#

easy enough to add

limber rapids Jul 11, 2023, 3:19 PM

#

Penguin is just encouraging you to use vkCmdDispatch for everything bleakekw

night charm Jul 11, 2023, 3:19 PM

#

bleakekw

#

when i add mesh shaders ill convert you too

limber rapids Jul 11, 2023, 3:20 PM

#

It's difficult enough without learning a new language thank you

night charm Jul 11, 2023, 3:20 PM

#

fair enough

#

one day

#

inshallah you will learn rust

languid fern Jul 11, 2023, 3:22 PM

#

night charm inshallah you will learn rust

We must convert them.

night charm Jul 11, 2023, 3:22 PM

#

🦀

languid fern Jul 11, 2023, 3:22 PM

#

night charm when i add mesh shaders ill convert you too

I wish I had an RT capable card

#

Mesh/Task/Accel Structures seem so fun

night charm Jul 11, 2023, 3:23 PM

#

what card do you have

limber rapids Jul 11, 2023, 3:23 PM

#

they are not

night charm Jul 11, 2023, 3:23 PM

#

bleakekw

limber rapids Jul 11, 2023, 3:23 PM

#

do not be fooled

languid fern Jul 11, 2023, 3:23 PM

#

Lol

night charm Jul 11, 2023, 3:23 PM

#

1060+ has compute emulation for rt

limber rapids Jul 11, 2023, 3:23 PM

#

I am currently struggling, meshoptimizer prefers good mesh topology to good spatial locality

#

Which is eating my brain

languid fern Jul 11, 2023, 3:23 PM

#

night charm what card do you have

Laptop 1050 bleakekw

night charm Jul 11, 2023, 3:24 PM

#

ah

#

just barely not bleakekw

languid fern Jul 11, 2023, 3:24 PM

#

Yep which makes it even more painful tbh

#

Anyways as long as I can do some compute shenanigans and rasterization I am ver happy

night charm Jul 11, 2023, 3:24 PM

#

definitely possible

limber rapids Jul 11, 2023, 3:25 PM

#

I wish we had access to the vertex pipeline in compute

#

mesh shaders kinda solve this but they are meh

#

NVIDIA also puts absurd limits on workgroup sizes

#

"it must be 32 or else"

night charm Jul 11, 2023, 3:26 PM

#

https://cdn.discordapp.com/emojis/754090061141639189.webp?size=48&name=kekw&quality=lossless

#

is it not a multiple of 32

#

instead of just 32

limber rapids Jul 11, 2023, 3:26 PM

#

With VK_EXT_mesh_shader it can be whatever

#

but if you use anything other than 32, occupancy dies (on NVIDIA)

night charm Jul 11, 2023, 3:27 PM

#

rip

limber rapids Jul 11, 2023, 3:27 PM

#

VK_NV_mesh_shader explicitly limits max invocation size to 32

night charm Jul 11, 2023, 3:27 PM

#

hmm i see

#

32 it is

limber rapids Jul 11, 2023, 3:27 PM

#

AMD is chad and likes big workgroup sizes

#

Up to 256

languid fern Jul 11, 2023, 3:28 PM

#

Oh also Penguin do you mind if I make some QoL improvements to context creation? (like the ability to select GPU stuff like that)

night charm Jul 11, 2023, 3:29 PM

#

oh yeah go ahead, it should already have some options to prefer a certain gpu but an explicit picker might not be bad either

languid fern Jul 11, 2023, 3:29 PM

#

Aight thanks yea

night charm Jul 11, 2023, 3:29 PM

#

should be mostly extending AppSettings/GpuRequirements

languid fern Jul 11, 2023, 3:29 PM

#

I have a scoring system already in my graphics API impl so I'd rather just pick manually

languid fern Jul 11, 2023, 3:29 PM

#

night charm should be mostly extending `AppSettings`/`GpuRequirements`

Gotcha

night charm Jul 11, 2023, 3:29 PM

#

scoring instead of discarding would be nice to have

#

currently it just discards gpus that dont fit the requirements given

languid fern Jul 11, 2023, 3:30 PM

#

I'll work on that too then lel

night charm Jul 11, 2023, 3:30 PM

#

frOK

iron shadow Jul 11, 2023, 5:59 PM

#

limber rapids they are not

i agree. do not let the pretty rays trick you. in serious words: the debugging experience is horrible. you will get bugs that are completely silent KEKW

languid fern Jul 11, 2023, 11:57 PM

#

https://github.com/NotAPenguin0/phobos-rs/issues/61

GitHub

Custom PhysicalDevice Selection Logic · Issue #61 · NotAPenguin0/ph...

'Tis my first feature request. Tell me if this is good or if I should add more detail. Implement a proper interface to allow users to select their PhysicalDevice (with potential fallback if non...

#

First time doing a feature request on dis so hopefully it goes well

iron shadow Jul 12, 2023, 12:10 AM

#

yo i should implement this so i can actually learn ash and vulkan

#

scoring? I feel like the user themselves could keep track of that

languid fern Jul 12, 2023, 12:12 AM

#

Right yea that's why I don't recomend that method of doing it

#

And just letting the user pick their own PhysicalDevice

languid fern Jul 12, 2023, 12:52 AM

#

iron shadow yo i should implement this so i can actually learn ash and vulkan

Also yea sure. Idk if it's gonna help with ash / vulkan too much though but should get you started

#

It's mostly just rust / iterator stuff

#

Now the question would be "how to retrieve all possible PhysicalDevices" before creating the instance itself

#

You could pass a callback that will be called to "select" a PhysicalDevice

#

So something like this

        .gpu(|physical_devices_iter| {
              // select one physical device from the iterator and return it
        })

#

Then you could pass what queues / features you want for said physical device / device

night charm Jul 12, 2023, 6:40 AM

#

That’s not bad yeah

languid fern Jul 12, 2023, 7:07 AM

#

I just realized that the reason you might be so offline is because of different time zones lol

#

It's 3 am in canada rn kek

night charm Jul 12, 2023, 7:08 AM

#

yeah im in eu https://cdn.discordapp.com/emojis/754090061141639189.webp?size=48&name=kekw&quality=lossless

#

its 9am now

languid fern Jul 12, 2023, 7:08 AM

#

Dang complete opposite eh

#

Anyways I'll get to work on phobos soon

night charm Jul 12, 2023, 7:08 AM

#

aye

languid fern Jul 12, 2023, 7:08 AM

#

I alr cloned it and started doing some testing just to get used to the API

night charm Jul 12, 2023, 7:09 AM

#

yeah thats perfect

#

I think for returning a default gpu in case the selector fails we could just pick the first dedicated gpu if there is one and log a warning that the selector failed

languid fern Jul 12, 2023, 7:13 AM

#

night charm I think for returning a default gpu in case the selector fails we could just pic...

That's assuming we use the one that uses the F: Fn(&PhysicalDevice) -> bool method

languid fern Jul 12, 2023, 7:13 AM

#

languid fern So something like this ```rs .gpu(|physical_devices_iter| { ...

If we force the user to always return a physical device like in here I don't think a fallback would be needed

night charm Jul 12, 2023, 7:15 AM

#

Hmm that’s possible too

#

But it requires tracking state between each iteration to see which is better

languid fern Jul 12, 2023, 7:18 AM

#

Up to the user KEKW

#

They could pick randomly if they wanted to as well

#

It's how wgpu does it too, just let's the user pick what adapter (phys device) they want

torn kettle Jul 12, 2023, 7:33 AM

#

nb that it is a bit questionable if the app should do any heuristic for picking

night charm Jul 12, 2023, 7:50 AM

#

Yeah I think just giving some requirements is the most you want to do

torn kettle Jul 12, 2023, 8:01 AM

#

even that

#

if the user wants to use the integrated gpu bc they want to conserve power

#

how can the app know?

night charm Jul 12, 2023, 8:56 AM

#

That is true tbh

#

Theres os settings for that usually right

iron shadow Jul 12, 2023, 10:26 AM

#

languid fern So something like this ```rs .gpu(|physical_devices_iter| { ...

This looks good, but how would you indicate what gpu you want?

#

OH

#

LMAOOOO it;s an iter, i thought you were gonna pass each gpu individually

#

Ok this makes sense now. I assume you iterate over that iter then return the desired gpu you want in same function

hybrid pilot Jul 12, 2023, 11:00 AM

#

torn kettle how can the app know?

So we need Ai and neuralink integration

night charm Jul 12, 2023, 11:18 AM

#

bleakekw

languid fern Jul 12, 2023, 2:36 PM

#

iron shadow Ok this makes sense now. I assume you iterate over that iter then return the des...

Yepp

#

Doesn't even need to be an iter, could just be a slice

languid fern Jul 12, 2023, 3:24 PM

#

night charm Theres os settings for that usually right

Yep

#

What I do in my rust app is check if the user is running on laptop battery (if they have one) and force the use of integrated graphics only then

#

If not it always uses the "best" dedicated card (one with the most vram and highest max buffer size)

torn kettle Jul 12, 2023, 3:44 PM

#

didn't understand my point :/

languid fern Jul 12, 2023, 3:47 PM

#

Oh no I was just talking about how my upstream API does it. Phobos IMO shouldn't handle that

#

GPU selection is completely up to the user is what I meant*

torn kettle Jul 12, 2023, 3:53 PM

#

thats not what i said

languid fern Jul 12, 2023, 3:56 PM

#

torn kettle nb that it is a bit questionable if the app should do any heuristic for picking

Oh sorry then I misunderstood. What'd you mean by this then?

torn kettle Jul 12, 2023, 4:02 PM

#

i explained a bit below

night charm Jul 12, 2023, 6:19 PM

#

i think the point is that the app shouldnt care what gpu is used, thats up to the user to configure in os settings

hybrid pilot Jul 12, 2023, 7:28 PM

#

night charm i think the point is that the app shouldnt care what gpu is used, thats up to th...

OS settings just reorder them right?

night charm Jul 12, 2023, 7:28 PM

#

depends

hybrid pilot Jul 12, 2023, 7:28 PM

#

So the first one is the preferred one?

night charm Jul 12, 2023, 7:28 PM

#

im pretty sure on linux at least you can disable them entirely

hybrid pilot Jul 12, 2023, 7:30 PM

#

night charm im pretty sure on linux at least you can disable them entirely

At what level?

#

Do you mean shutting down the device?

night charm Jul 12, 2023, 7:30 PM

#

i think so

#

or just disabling the driver

hybrid pilot Jul 12, 2023, 7:30 PM

#

You can't shut down the integrated gpu in some setups though

#

Because those drive the display

night charm Jul 12, 2023, 7:30 PM

#

oh yeah, igpu might be a special case

hybrid pilot Jul 12, 2023, 7:31 PM

#

But Xorg had a way of choosing devices

#

DRI prime something

night charm Jul 12, 2023, 7:31 PM

#

yeah

#

should work on wayland too I think thats a driver thing

hybrid pilot Jul 12, 2023, 7:31 PM

#

I don't have such a devices so I can't test what it does to vulkan

night charm Jul 12, 2023, 7:31 PM

#

either way xorg bad

hybrid pilot Jul 12, 2023, 7:31 PM

#

Mboth xorg and Wayland compositors have to manage multi gpu afaik

languid fern Jul 12, 2023, 9:44 PM

#

Descriptor indexing allows you to do bindless shenanigans right?

#

Is it supported in Phobos yet? Cause I saw an issue stating that you can't change descriptor sets after you call draw / dispatch, which might be problematic for my renderer (if not using bindless)

night charm Jul 12, 2023, 10:17 PM

#

that issue is about something else, descriptor indexing is coming soon

#

i just have to add the api for it really

#

I added a comment to the issue to explain the problem better

#

its more of a design annoyance really

#

ill try to get descriptor indexing out by this weekend or so

#

im back from vacation friday night

languid fern Jul 12, 2023, 10:20 PM

#

Oh deal deal

#

It's gonna be nice utilizing bindless for the first time

night charm Jul 12, 2023, 10:20 PM

#

it is really nice

#

though vk api wise bindful isnt really annoying anymore with phobos

#

not like with a lot of vk renderers that store descriptor sets per material 🤢

languid fern Jul 12, 2023, 10:22 PM

#

... that's what I do lel

#

very naive way but worked fine 4 me

#

scales horribly though kek

night charm Jul 12, 2023, 10:23 PM

#

it really does lol

#

here you just call bind with an image or buffer and it Just Works ™️

languid fern Jul 12, 2023, 10:24 PM

#

Yep that's how my current high level graphics API worked, but it does some hideous stuff under the hood to make it work.

#

Good to know I don't have to deal with that anymore lel

night charm Jul 12, 2023, 10:25 PM

#

but it does some hideous stuff under the hood to make it work.

#

ok its not that bad tbh

#

just a hashmap full of descriptor sets

languid fern Jul 12, 2023, 10:26 PM

#

Yea it just "feels" very ugly lol

night charm Jul 12, 2023, 10:26 PM

#

speaking of i should replace those with ahash before i forget

languid fern Jul 12, 2023, 10:26 PM

#

Yess ahash all the things

night charm Jul 12, 2023, 10:26 PM

#

especially pipeline lookup can get expensive

languid fern Jul 12, 2023, 10:26 PM

#

There's also LruCache that automatically drops old values

night charm Jul 12, 2023, 10:26 PM

#

but thats mostly because it needs to hash a lot of things

languid fern Jul 12, 2023, 10:26 PM

#

Right yea

night charm Jul 12, 2023, 10:27 PM

#

languid fern There's also LruCache that automatically drops old values

that seems similar to what I do except this "oldness" is determined by the last frame it was used on

languid fern Jul 12, 2023, 10:27 PM

#

Neat

#

Is there a hard-coded lifetime limit?

#

Like 2 - 3 frames?

night charm Jul 12, 2023, 10:28 PM

#

I think its 4 or 8

#

it is hardcoded though yeah

languid fern Jul 12, 2023, 10:28 PM

#

Awesome

night charm Jul 12, 2023, 10:28 PM

#

if its used sporadically then recreating it is no big deal

languid fern Jul 12, 2023, 10:32 PM

#

Okkk I was confused for a sec with how methods (bind storage texture, bind sampled texture) in incomplete command buffer automatically handled sync

#

They don't, I needed to look in the pass builder lol

#

I was scared for a sec thinking this lib doesn't do auto-sync and that I'll need to handle putting pipeline barriers myself

night charm Jul 12, 2023, 10:36 PM

#

no worries it does

#

just not when binding no

#

you declare them when you build the pass

languid fern Jul 12, 2023, 10:37 PM

#

Yep that's all I care about really

#

I'll just need to change my graphics lib to allow to build passes like that but it shouldn't be too hard

#Phobos - Vulkan abstraction library in Rust