Rosy

[vagrant musk]

yeah
driver crashes can cause other programs to have problems, including crashes

[astral hinge]

cascading program failure

[cloud rivet]

oh

I forgot to actually send the index

so I was sampling with UB

[wraith urchin]

They are actually

[cloud rivet]

no I am setting it

[vagrant musk]

I’m surprised driver crashes don’t happen when training a pytorch model that takes like 100x the amount of memory your GPU has
… instead the program just crashes before the driver

pt probably has a check ig

[cloud rivet]

I will not sample but look at the index in render doc and match to where the csm is the descriptor

if I don't sample it won't crash

so I should be able to see all this

just going to look through how I set up the descriptor first though

[vagrant musk]

inb4 renderdoc crashes

[cloud rivet]

oh

I was using the sampler I created for the debug image

no I wasn't

fuck

maybe it's the 8K image

[astral hinge]

that's only ~250MB

[true moon]

It's quite interesting that memory profiling tools for GPUs seem to be not super common

You'd think every vendor's profiler would just tell you exactly how much memory is committed

[astral hinge]

AMD has a really good one called RMV

nsight graphics debuggeroni has a view that shows the memory usage of every resource but it's not as detailed as RMV

like RMV lets you take snapshots and compare them to see what got allocated/freed during a period. I don't think nsight has that

I think pix also has a thing similar to nsight

but yeah memory profiling tools aren't too important imo because you have most of the info you need in your app already

[cloud rivet]

ok

so it was ub descriptor indexes

I fixed it

I was allocating indexes incorrectly

VVL no use in this case, this is what matty was talking about how you lose saftey with bindless

well I figured it out anyway

8K shadows

looks pretty good actually lol

needs some PCF

let me try 4k now

looks like I got some shadow acne too

holy fuck though shadows make such a huge difference

looks great

damn this was so worth all the pain

honestly that was the least painful shadows have ever been I am getting better

I'm gonna fix that acne in a second but I want to see 4k now

well definitely worse with 4k but I can live with it

I will fix the shadow acne then

idk my existing PCF is not very good so maybe I can look into that idk

maybe it doesn't matter at my game's camera distance

[wraith urchin]

Turn on hardware pcf

Also you are gonna wanna look into the different types of bias

To fix the acne

[cloud rivet]

hrm hardware pcf

I'm pretty happy with these shadows tbh

they could be better, but this is good for now I am going to work on either blending or normal maps next

I'm gonna post a #wip with a little shorter video and then go to bed

lol

doesn't count slang

🇬🇳

[true moon]

Dang

What is all that

[brisk chasm]

music please 🙂

and i am jealous

my shadows never looked like they work :>

[astral hinge]

yes they did

I remember clearly that debug session we had and the big rock model

[brisk chasm]

ja that one time in the holiday camp

good times 🙂

rock4 something

https://sketchfab.com/3d-models/rock-terrain-3-8d16b9d28b1841ada0c3c3595feccf52 : )

[steady venture]

I just use "phi" "theta" "alpha". Not that this answers your problem.

[steady venture]

Awesome!

[steady venture]

Wait what, that's a thing?

(intense googling)

[cloud rivet]

A lot of vk create structs

It’s all public code https://github.com/btipling/Rosy/blob/master/src/Engine/Graphics.cpp

I am really happy with this absolutely literal, explicit approach with no helper functions. I am never getting annoyed with my own abstractions. Making changes, finding code, it’s trivial

It’s all in a single file

My build is fast, it’s just great

There’s a lot of explicit repetitive blocks of code for creating buffers, samplers etc for example

[cloud rivet]

I am gonna add all the (probably AI generated) music I listen to while writing code on this playlist https://youtube.com/playlist?list=PLba6ETlcAXxCAbOovAsgxVSaID_hqVK3D&si=UvICstTu0ADEmLHl

[cloud rivet]

It is in OpenGL with GLSL, just the type of sampler. I don’t know if there’s a slang version of it

I think I looked into it last month when working on shadows and couldn’t find anything

I am going to not work on shadows for a bit

Once I have an actual game and understand how shadows should look and work in the game I will revisit the topic

[steady venture]

Yeah I just looked it up shadow2DSampler or something

[cloud rivet]

https://github.com/shader-slang/slang/pull/5789/files looks like slang does have it

the SPIR-V analogue is OpSampledImage maybe

[cloud rivet]

I need to read about SPIRV sampler filters

[cloud rivet]

   // If a variable ID or parameter ID is found in this set, a sampler is actually a shadow/comparison sampler.
   // SPIR-V does not support this distinction, so we must keep track of this information outside the type system.

https://github.com/KhronosGroup/SPIRV-Cross/blob/main/spirv_cross.hpp#L938C1-L939C113

I just think it's not a thing in spirv

I am not sure that shadow2DSampler is "hardware PCF"

I think it's driver implemented PCF

https://github.com/SaschaWillems/Vulkan/blob/6c7f72f0bdb066a95f2187f2455bfff9c23414a6/shaders/hlsl/deferredshadows/deferred.frag#L10-L12

https://github.com/SaschaWillems/Vulkan/blob/6c7f72f0bdb066a95f2187f2455bfff9c23414a6/shaders/hlsl/deferredshadows/deferred.frag#L54-L76

that's just what I am sort of already doing

I think the future of shadows is just RT

and I will probably add hybrid RT for better shadows rather than going down this rabbithole and running geometry through a render pass multiple times for worse looking shadows

and if you don't have RT hardware you don't have shadows idk

or you have shadowmap shadows that aren't as nice tbh, I can at least do that

directional light rays without any bounces are probably cheap? idk how hardware RT works

I'll probably do that for shadows as a next step, just turn on hybrid RT

conditionally turn it on

[wraith urchin]

It's not, but you need it in order to use hardware PCF. In GL you also need to set the compare mode and filter in the sampler object

It's something like sampler.compareEnable = VK_TRUE; sampler.compareOp = VK_COMPARE_OP_LESS; in bulkan

And probably VK_FILTER_LINEAR too

[cloud rivet]

PCF requires sampling multiple points yes

my sampler already uses those settings that are correct for a reverse depth

it doesn't automatically perform any PCF at the moment

without compare enable and compare op it just doesn't work at all

and results in a VVL I think

but with it turned on doesn't like multi sample

I think https://vulkan-tutorial.com/Multisampling is what I need to look at

yeah I wasn't claiming that

the op (i.e. VK_COMPARE_OP_LESS )is different depending on what direction your depth goes (0 to 1 vs 1 to 0) based on your application's needs

is why I even said that

[brisk chasm]

yeah

[cloud rivet]

so you need an op, but it might not be VK_COMPARE_OP_LESS

[brisk chasm]

infintinitnie 1 to 0 you need GREATER, generally speaking

[cloud rivet]

but that doesn't give you any PCF like output

I just get the value for one fragment with that sampling

[brisk chasm]

you are talking about the depth you pass into the shadowShrimpler

[cloud rivet]

shadow2DSampler works totally different iirc

I used it shadow2DSampler in foundations, but I forget the details now

I think I need to do that multisampling thing as per that vulkan tutorial to get the same behavior

so yes, it does exist in VK but it's not shadow specific

and I don't know anything about it yet

it's not as simple as just using shadow2DSampler

that's probably something I'll look into soon

clearly it makes a big difference

[wraith urchin]

hardware PCF is more like bilinear filtering than multisampling

[wraith urchin]

Did you also set the magFilter & minFilter fields of VkSamplerCreateInfo to VK_FILTER_LINEAR?

[cloud rivet]

not to linear no, to nearest, maybe that was a mistake https://github.com/btipling/Rosy/blob/master/src/Engine/Graphics.cpp#L2372-L2373

[wraith urchin]

Yeah in GL you need linear, I assume Bulkan is the same

[cloud rivet]

Linear filtering (VK_FILTER_LINEAR) computes a set of neighboring coordinates which bound the unnormalized coordinates. The integer texel coordinates are combinations of i0 or i1, j0 or j1, k0 or k1, as well as weights α, β, and γ.

yeah

oops maybe easy win

I just change that and things will look better

thank you

damn

easy to check right now tbh

[wraith urchin]

This VkSamplerCreateInfo stuff is actually really similar to the GL version

Maybe Bulkan isn't so much extra after all

Are you gonna do MSAA?

[cloud rivet]

ok so this is how it is now, let me try linear

[cloud rivet]

I will definitely do something, I don't know if MSAA is the best option

aw shit

[wraith urchin]

Depends what you mean by best

[cloud rivet]

                    sampler_create_info.magFilter = VK_FILTER_LINEAR;
                    sampler_create_info.minFilter = VK_FILTER_LINEAR;

                    sampler_create_info.borderColor = VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK;
                    sampler_create_info.unnormalizedCoordinates = VK_FALSE;
                    sampler_create_info.flags = 0;
                    if (VkResult res = vkCreateSampler(device, &sampler_create_info, nullptr, &shadow_map_image.shadow_sampler); res != VK_SUCCESS) return res;

I am using linear already

nm

I had previously linked NEAREST where I was referencing the sampler for the imgui image

oh well

I can do better PCF though in the shader

but I am not going to work on it now, I want to understand what my game actually needs and it may just be hybrid RT shadows and not a shadowmap at all

[wraith urchin]

The best thing to do is both hardware PCF and shader PCF

[cloud rivet]

yeah I have a shader PCF it is garbage though

you can kind of see what it's doing in that image

[wraith urchin]

can I see your shader currently?

I have a hunch on why this isn't working

[cloud rivet]

ok, but it's embarrasingly bad https://github.com/btipling/Rosy/blob/master/src/shaders/pcf.slang 😰

[wraith urchin]

The one that is supposed to be hardware PCF?

[cloud rivet]

I don't have any hardware PCF

[wraith urchin]

then what was sampler_create_info.minFilter = VK_FILTER_LINEAR; all about?

[cloud rivet]

https://github.com/btipling/Rosy/blob/master/src/shaders/basic.slang#L71

[cloud rivet]

yes that sampler with that configuration is what I am using in that case

so on line 71 in basic.slang

that sampler has the VK_FILTER_LINEAR

I sample float4 mapValue = shadowMap.Sample(shadowSampler, float2(vertex.x, vertex.y));

[wraith urchin]

VK_FILTER_LINEAR is for hardware PCF

[cloud rivet]

I don't understand I guess what you are trying to ask me

[wraith urchin]

I'm trying to get you to use hardware PCF to make your shadows look better

[cloud rivet]

yes, thank you

I have a shadow map texture that I wrote to in a render pass where each texel represents a projection of a coordinate in 3D space into that image, I sample by taking a vertex in my vertex shader and applying the same perspective projection to the dimensions of the image with xy coordinates

the depth value is recorded in the first color channel of the image

[wraith urchin]

How do you do a sample with a comparison in slang?

[cloud rivet]

it's just the color value in the red channel vs the z value of the projected vertex z

so the red channel's value in the shadow map is compared to my z

for that fragment

[wraith urchin]

like the equivalent of float textureLod( sampler2DShadow sampler, vec3 P, float lod); in glsl

[cloud rivet]

and if the value is greater than I show a shadow because it is occluded

yeah

that's different

there's an HLSL equivalent that I would use

[wraith urchin]

with hardware PCF you don't do the comparison yourself, you just provide the z value in P.z and the hardware returns a value between 0.0-1.0

[cloud rivet]

ok I think opengl drivers are doing some magic there you don't get in vulkan but that could be

[wraith urchin]

I can find the HLSL version 1 sec

[cloud rivet]

because I looked at sascha's implementation

and he's just doing it himself

[wraith urchin]

its used in MJP's thing

[cloud rivet]

and I looked at slang

yes I linked it previously

[wraith urchin]

in HLSL it looks the the func is SampleCmpLevelZero

maybe slang is the same?

[cloud rivet]

https://github.com/shader-slang/slang/pull/5789/files

yep

[astral hinge]

it do be

[cloud rivet]

yeah I am using linear

I am not using SampleCmpLevelZero

I can try it out

[wraith urchin]

It should be something like float pcf(float4 fragCoord, float4 vertex, texture2D shadowMap, SamplerState shadowSampler) { float result = 0.0f; float shadowOffset = 0.00025; // Sample some more locations; result += shadowMap.SampleCmpLevelZero(shadowSampler, float2(vertex.x - shadowOffset, vertex.y - shadowOffset), vertex.z); result += shadowMap.SampleCmpLevelZero(shadowSampler, float2(vertex.x + shadowOffset * 2.f, vertex.y), vertex.z); result += shadowMap.SampleCmpLevelZero(shadowSampler, float2(vertex.x, vertex.y + shadowOffset * 2.f), vertex.z); result += shadowMap.SampleCmpLevelZero(shadowSampler, float2(vertex.x + shadowOffset * 2.f, vertex.y + shadowOffset * 2.f), vertex.z); result += shadowMap.SampleCmpLevelZero(shadowSampler, float2(vertex.x - shadowOffset * 2.f, vertex.y - shadowOffset * 2.f), vertex.z); return result * (1.0 / 5.0f); }

[cloud rivet]

I will try that thank you

ah

it requires a different binding index

fuck

[wraith urchin]

Bulkan thing?

[cloud rivet]

it needs a separate descriptor binding index

because slange is like, no you're already using a sampled image sampler at that binding

and this is like a different type of sampler for slang

I think it would just be a new binding with the same type of sampler? this is init code stuff

hrm

I don't know if that's even possible

or I don't know how to do it

I have to read about it

[wraith urchin]

I'm not sure, I think you might need to ask #vulkan about that one

[cloud rivet]

maybe I can have a second SAMPLED_IMAGE on this

            const auto bindings = std::vector<VkDescriptorSetLayoutBinding>({
                {desc_storage_images->binding, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, descriptor_max_storage_image_descriptors, VK_SHADER_STAGE_ALL},
                {desc_sampled_images->binding, VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE, descriptor_max_sampled_image_descriptors, VK_SHADER_STAGE_ALL},
                {desc_samples->binding, VK_DESCRIPTOR_TYPE_SAMPLER, descriptor_max_sample_descriptors, VK_SHADER_STAGE_ALL},
                });

If flags contains VK_DESCRIPTOR_SET_LAYOUT_CREATE_PER_STAGE_BIT_NV, then there must not be any two elements of the pBindings array with the same VkDescriptorSetLayoutBinding::binding value and their VkDescriptorSetLayoutBinding::stageFlags containing the same bit

hrm

[astral hinge]

why do you need a second sampled image binding 🤔

[cloud rivet]

Compiled debug.slang successfully
C:\Users\Bjorn\source\repos\btipling\Rosy\src\shaders/basic.slang(6): warning 39001: explicit binding for parameter 'shadowSampler' overlaps with parameter 'imageSampler'
[[vk::binding(1, 0)]] SamplerComparisonState shadowSampler[];
                                             ^~~~~~~~~~~~~
C:\Users\Bjorn\source\repos\btipling\Rosy\src\shaders/basic.slang(5): note: see declaration of 'imageSampler'
[[vk::binding(1, 0)]] SamplerState imageSampler[];
                                   ^~~~~~~~~~~~
C:\Users\Bjorn\source\repos\btipling\Rosy\src\shaders/basic.slang(6): warning 39001: explicit binding for parameter 'shadowSampler' overlaps with parameter 'imageSampler'
[[vk::binding(1, 0)]] SamplerComparisonState shadowSampler[];
                                             ^~~~~~~~~~~~~
C:\Users\Bjorn\source\repos\btipling\Rosy\src\shaders/basic.slang(5): note: see declaration of 'imageSampler'
[[vk::binding(1, 0)]] SamplerState imageSampler[];
                                   ^~~~~~~~~~~~

Compiled basic.slang successfully
Compiled all shaders successfully

the samplers are bound at 1

and I can't have two of them

[[vk::binding(1, 0)]] SamplerState imageSampler[];
[[vk::binding(1, 0)]] SamplerComparisonState shadowSampler[];

that's not a thing you can do

so how else would I do it but create a separate binding

[astral hinge]

it's just a warning

[cloud rivet]

hrmmm

ok let me run it then

[astral hinge]

it's 100% legal in vulkan to do this

[cloud rivet]

wow way better

Jake confirmed Vulkan MVP

[wraith urchin]

Jaker I think, not me

[cloud rivet]

I have the nicest shadows 🤗 thanks ya'll

[astral hinge]

por que no los dos

[cloud rivet]

it's not AAA shadows (aka Jaker shadows) but I am very very happy with this outcome

[brisk chasm]

is that CSM already?

[cloud rivet]

I set it up to be a CSM but I am only rendering one cascade because of how expensive it is

and it seems to be good enough

for sponza anyway

[wraith urchin]

how many ms?

[cloud rivet]

it added 2 ms to the frame

so I am up to 3-4 ms per frame now

so if each cascade costs 2ms

that seems a lot ya

and this is just for tiny sponza

I need culling to do cascade

but I don't think I need it for my game

[wraith urchin]

yeah it shouldn't take 2ms per cascade on a 3070

[cloud rivet]

I mean I need culling

well it's just rendering the same geometry as the render

[wraith urchin]

this is crytek sponza?

[cloud rivet]

yeah

[wraith urchin]

you should be getting better than 2ms for one cascade on a 3070 on crytek sponza

[cloud rivet]

without culling?

maybe it's the layers

let me disable layer messages

[wraith urchin]

yes without culling, unless maybe you have a pathologically large cascade

[cloud rivet]

the layers add overhead

[wraith urchin]

what is the cascade resolution

[cloud rivet]

4096^2

[wraith urchin]

yeah try disabling layers maybe

[cloud rivet]

doesn't seem any better

I can profile later

[wraith urchin]

GPU trace time

[cloud rivet]

yeah

it's just a bunch of draw calls

[echo crystal]

are u rendering only to depth

[cloud rivet]

yeah

looks like PCIe throughput

so uh

[echo crystal]

hmm

i have seperate position buffer don't know how much that affects

[cloud rivet]

I have that too

hrm

I think I need to put all my meshes into a single buffer

so the draws took 1ms

those pipeline barriers heh

[wraith urchin]

Something is seriously wrong then, during my shadow pass PCIe throughput is almost zero

[cloud rivet]

yeah I don't know I have to manage all that myself, and I have never looked at it, it's using VMA maybe my VMA setup is wrong

because I have prefer device set up for those buffers

also

I need to just use one buffer

every mesh is its own buffer which is bad

[wraith urchin]

You sure you didn't allocate the vertex buffer in CPU memory by mistake?

[cloud rivet]

I don't dicate it

you tell VMA your usage

I tell it prefer device

[wraith urchin]

Yeah I dunno, ask in #vulkan maybe

but PCIe throughput being anywhere close to that looks very wrong

[cloud rivet]

actually I have this set

vma_alloc_info.usage = VMA_MEMORY_USAGE_AUTO;
                    vma_alloc_info.flags = VMA_ALLOCATION_CREATE_MAPPED_BIT | VMA_ALLOCATION_CREATE_HOST_ACCESS_RANDOM_BIT;

that's wrong

oops

wait that's my scene data

I don't need to read any data from host so I don't need access random bit

frame time went up +3x

@wraith urchin figuratively saving my life today, teaching me programming 101

[wraith urchin]

Np

[cloud rivet]

1.4ms -> 0.14ms for shadow pass

[echo crystal]

hehe nice

[cloud rivet]

all I did was not flag usage for random access why did I do that, literally had brain turned off

[echo crystal]

it's ok

everyone makes mistakes

[cloud rivet]

imgui theme update

[astral hinge]

AMD theme

[cloud rivet]

I got it from the graphics programming github org

I had to change it a little since it was outdated

something seems to have gone wrong

[true moon]

Apologies for the shitty take I'm a fingerstyle noob

[cloud rivet]

that sounds nice

[true moon]

Hopefully it sounds familiar

[cloud rivet]

I don't know music, so I do not recognize it

[true moon]

It's the dwarf fortress theme that plays in fortress mode

[cloud rivet]

ohhhh

yes

I recognize it!

[true moon]

That's why I'm posting it in your thread haha I wasn't just spamming

[cloud rivet]

yeah that sounds just like it

it's been a minute since I played

[cloud rivet]

nice

I just found out I had UB all this time because I was setting the material buffer size to vector.size() * sizeof(vector) instead of type in vector which is and I didn't break until I added alpha mode

also the faster frame rate led to me discovering I had a time step bug

because I was suddenly zipping around too fast, I wasn't tracking SDL ticks correctly

that's also fixed

lots of fixes today

maybe my game is just flying around sponza I am having fun just doing that tbh

[cloud rivet]

support for alpha mode

surface details and normal mapping tomorrow then I guess

then do that fancy lighting jake was telling me about

[dry apex]

nice. it looks way better

[cloud rivet]

thanks!

after fancy lighting I will add frustum culling

then I'll try intel sponza

[wraith urchin]

by "fancy lighting" do you basically mean PBR materials?

[cloud rivet]

what you were sharing the other day

I just have the n dot l atm

#1323084490997895198 message

[wraith urchin]

Ah yeah that was my GGX lighting code

should work

[cloud rivet]

I gotta add you all to the credits on this

[wraith urchin]

Did you do any normal mapping stuff back when you were doing GL?

[cloud rivet]

yes

I stole your grass

I was normal mapping it

[wraith urchin]

Haha lol, I think we just stole used it with permission from the same place

[cloud rivet]

the lengyel math is really good for tangent space stuff

shader objects made that so easy

I didn't have to make a separate pipeline

it was just like how I would do it in opengl

wow sponza is a real place https://en.wikipedia.org/wiki/Sponza_Palace#/media/File:Pati_del_palau_Sponza_de_Dubrovnik.JPG

[dry apex]

imaging marrying there

[cloud rivet]

why do I not live in Crotia?

[cloud rivet]

I got married in Hawai'i that was nice :D

[wraith urchin]

Sibenik is in Croatia too lol

[cloud rivet]

yes

[dry apex]

Wow nice. you have children too?

[cloud rivet]

yes I have a 17 year old daughter

almost college age

[dry apex]

im about the same age as your child lmao

[cloud rivet]

:D

I'm young at heart

man croatia is pretty

[cloud rivet]

20 years ago this December btw

[cloud rivet]

I am going to at some point switch to premake

Not anytime soon but no reason to be able to not build cross platform with sdl and vulkan

[brisk dagger]

That's my goal 🫡

I need someone to volunteer however

[dry apex]

real

[cloud rivet]

this hobby doesn't help unfortunately

[cloud rivet]

MSAA looks trivial to add so I am going to do that next actually https://vulkan-tutorial.com/Multisampling

[wraith urchin]

Yessss do MSAA!!!!

You need those crisp clean edges

[silver slate]

Unless you rely on a lot of shader magic that could really use sub-pixel levels of detail. :')

[wraith urchin]

There are ways of dealing with that

MSAA is the default in ASO

[cloud rivet]

yes

gonna do it in a bit

[cloud rivet]

it's gotta be painful to follow me on github as I am constantly spamming public work repos

[wraith urchin]

Lol it's OK, I don't really look at that part of github anyways

[cloud rivet]

lol

[cloud rivet]

me neither

[wraith urchin]

Why is your work stuff public?

[cloud rivet]

it's open source

[wraith urchin]

Ah

[cloud rivet]

I just found a bunch that were missing license files

[wraith urchin]

Good thing my work code is hidden

[cloud rivet]

I'm mostly a manager these days so all my work stuff is just small bits, not a lot of interesting code

[cloud rivet]

ok I guess this is the most obvious example I could find of aliasing in sponza as a before MSAA picture

idk

[wraith urchin]

Can't wait to see the after

[cloud rivet]

is that better?

seems like it?

I seem to have lost shadows

what happened to the shadows

seems less aliased but no shadows

hrm

maybe I have to add a resolve mode on the depth attachment too?

oh

I deleted my shadow pass while debugging

I forgot I did that

[true moon]

Lmfao

[cloud rivet]

ok ok what do we think

seems better?

[echo crystal]

yep

[cloud rivet]

that was hard

the vulkan tutorial was no use

well a little bit useful

I now have two draw images

hrm

maybe I don't have to blit into the swapchain now, because I can just resolve into it?

my frametime doubled

I'm going try removing the blit

and resolve into swapchain

I also went with absolute maximum sample mask size

                            vkCmdSetRasterizationSamplesEXT(cf.command_buffer, msaa_samples);
                            VkSampleMask sample_mask{ ~0U };
                            vkCmdSetSampleMaskEXT(cf.command_buffer, msaa_samples, &sample_mask);
                            VkExtent2D fragment_size = { 1, 1 };
                            VkFragmentShadingRateCombinerOpKHR combiner_ops[2];
                            combiner_ops[0] = VK_FRAGMENT_SHADING_RATE_COMBINER_OP_KEEP_KHR;
                            combiner_ops[1] = VK_FRAGMENT_SHADING_RATE_COMBINER_OP_REPLACE_KHR;
                            vkCmdSetFragmentShadingRateKHR(cf.command_buffer, &fragment_size, combiner_ops);

maybe I can uh

sample less and still get good results

I'm going to test getting rid of the blit first

that actually made everything worse

time to profile

[wraith urchin]

So sharp so clean

How many samples is that?

4x is the sweetspot for me

[cloud rivet]

this is 8

[wraith urchin]

Cool

[cloud rivet]

ID,,Origin,Source,Message
12,,Target,GPU Trace,GPU Trace failed to initialize periodic sampling library. You might not have permission to access NVIDIA GPU performance counters on the target device.

cannot profile

hrmm

[true moon]

Don't you need to run nsight as admin

[cloud rivet]

oh I didn't before

let me try that

[true moon]

Maybe you crashed the driver and it's in a weird state

[cloud rivet]

ok running as admin worked

yeah it's the sampling

lemme try 4x

also I don't know what the sample mask should be

the tutorial doesn't use a sample mask

because it's old

[wraith urchin]

Wait are you using sample mask?

[cloud rivet]

yes

VkSampleMask sample_mask{ ~0U };

[wraith urchin]

Be careful with that, you can accidentally force the shader to be executed per sample

[cloud rivet]

hrm

ok I will 4x and then I will try adjusting the sampler mask

[wraith urchin]

Msaa will slow things down, but it should be manageable

[cloud rivet]

wow 4 made a huge difference

[wraith urchin]

Aso runs 4x by default

[cloud rivet]

I see a little bit of aliasing

frame time is much better now

still much better than it was

maybe msaa is more important at a big camera distance

this looks fine

I'm sticking with 4

[astral hinge]

pls make a magnifier tool in your engine

[cloud rivet]

oh

[astral hinge]

you can do it with imgui

[cloud rivet]

I actually have a thing

hrm

is that using the draw image as an image in imgui

I can do that

but

hrm

[wraith urchin]

Are you doing HDR tonemapping @cloud rivet ?

[cloud rivet]

it's being rendered into

[cloud rivet]

I don't know what that is so no

[wraith urchin]

OK if you end up doing tonemapping you will need to tonemap per sample

[cloud rivet]

I am going to make a github issue for a magnfier

how do you do anything per sample

it's just something I configured

I will figure that out I guess

[wraith urchin]

Texture 2d multisample

Not sure what they call that in bulkan

[cloud rivet]

I see

[astral hinge]

Read mjp's article about MSAA

[cloud rivet]

https://therealmjp.github.io/posts/msaa-overview/ 👀

ohh

I see

the samples are stored

that explains why I can't blit the sampled image, it has to be "resolved"

into a 1 sample count

and so I am defining an image with 4 samples per pixel and the tone mapping would have to be applied to all the samples makes sense

I think I understand

[wraith urchin]

Yeah well you can do tonemapping post msaa resolve but it looks much worse

MJP's article explains why iirc

[cloud rivet]

yes it does

When resolving before tone mapping we’re oversampling the signal representing physical light being reflected towards the camera, and when resolving after tone mapping we’re oversampling the signal representing colors displayed on the screen.

1. During MSAA resolve of an HDR render target, apply tone mapping and exposure to each sub-sample
2. Apply a reconstruction filter to each sub-sample to compute the resolved, tone mapped value
3. Apply the inverse of tone mapping and exposure (or an approximation) to go back to linear HDR space

hrm

I don't know what those last two points mean and they aren't explained I think

[wraith urchin]

Yeah don't worry about it unless you are doing tonemapping though

Eventually you will likely end up doing it though im sure lol

[cloud rivet]

I found this video where MJP is talking about sampling in 2015 https://www.youtube.com/watch?v=nj4puag4hwc i

[astral hinge]

the reconstruction kernel is how we approximate values at points we haven't sampled

[cloud rivet]

I see

hrm

this is a nice slide

[astral hinge]

I guess in this case we are probably trying to calculate the average value of the function (radiance at a point) within the pixel (which we can assume is a square)

so one naive reconstruction kernel would be simply averaging the msaa samples

[cloud rivet]

and that has to be done manually?

[wraith urchin]

Basically a box filter, is what happens with the default blit

[astral hinge]

yeah

idk how other kernels would look. I never tried it

[wraith urchin]

MJP has another sample on GH that uses other kernels

[cloud rivet]

so you have like a fragment shader that reconstructs the previous draw in a special pass for that? I don't understand

[astral hinge]

"reconstruction" means trying to figure out what the function that we are discretely sampling looks like

[wraith urchin]

I wasn't super impressed with the alternative kernels but

[cloud rivet]

is that just configured?

[astral hinge]

wdym

[cloud rivet]

or do you have to write shader code

[wraith urchin]

If you want anything other than a box filter you need to use a shader

[cloud rivet]

what code would you specifically write for "Apply a reconstruction filter to each sub-sample to compute the resolved, tone mapped value"

ok

I see

so it is done in a shader

[wraith urchin]

But also tonemapping per sub-sample needs a shader

[cloud rivet]

but the samples are part of the image you rendered into right?

[wraith urchin]

mine looks like cpp for(int i = 0; i < MSAA_SAMPLES; i++) { pixel += tonemap(texelFetch(color_texture, coord, i).rgb); } pixel *= (1.0 / MSAA_SAMPLES);

[astral hinge]

btw this is all signal processing theory

[cloud rivet]

is that fragment shader doing anything else?

[astral hinge]

I think bart wronski has a bunch of articles on this topic

[wraith urchin]

theres alot of stuff I cut out, this shader has a bunch of stuff merged into it, but this is the tonemapping part

[cloud rivet]

so you have a shader that's sort of working with the surfaces and fragment data you got from the graphics pipeline, and then a later stage subsequent pass that now has the sampled data does all this work yeah?

I think I understand

[wraith urchin]

Yeah after everything is rendered to a multisampled target, this shader runs to sample all that and writes to a non-multisampled output

[astral hinge]

input: msaa texture(s)
output: single sample texture

[cloud rivet]

so this is done instead of just the "resolveMode" mode on an attachment

[wraith urchin]

If you need something more complex than that yeah

[cloud rivet]

typedef enum VkResolveModeFlagBits {
    VK_RESOLVE_MODE_NONE = 0,
    VK_RESOLVE_MODE_SAMPLE_ZERO_BIT = 0x00000001,
    VK_RESOLVE_MODE_AVERAGE_BIT = 0x00000002,
    VK_RESOLVE_MODE_MIN_BIT = 0x00000004,
    VK_RESOLVE_MODE_MAX_BIT = 0x00000008,
    VK_RESOLVE_MODE_EXTERNAL_FORMAT_DOWNSAMPLE_ANDROID = 0x00000010,
    VK_RESOLVE_MODE_NONE_KHR = VK_RESOLVE_MODE_NONE,
    VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT,
    VK_RESOLVE_MODE_AVERAGE_BIT_KHR = VK_RESOLVE_MODE_AVERAGE_BIT,
    VK_RESOLVE_MODE_MIN_BIT_KHR = VK_RESOLVE_MODE_MIN_BIT,
    VK_RESOLVE_MODE_MAX_BIT_KHR = VK_RESOLVE_MODE_MAX_BIT,
    VK_RESOLVE_MODE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
} VkResolveModeFlagBits;

is all the modes you can use with resolving

depending on the image format

I am using VK_RESOLVE_MODE_AVERAGE_BIT

because VVL yelled at me to use it

because of my format

that sounds like what jaker was saying

so one naive reconstruction kernel would be simply averaging the msaa samples

[astral hinge]

ye that's the box kernel the other Jake mentioned

mathematicians have weird names for things

[wraith urchin]

One of these days I might mess with a Lanczos resolve kernel

[cloud rivet]

the video goes into detail on dealing with flickers and temporal stability

which looks like signal processing

alrighty thanks for all the information and resources I know a lot about msaa now

I didn't know anything this morning

I'm going to look at normal mapping

oh even my debug lines are smooth now

those were super aliased

[wraith urchin]

so smooth

[cloud rivet]

even imgui looks nicer

[astral hinge]

Hmm

[cloud rivet]

hrm it really makes a difference

video capture stutters

from a previous video I shared

clearly better

let me try to get exact same picture as that shot

even the roof texture looks better?

idk

could be because it's a screenshot of a paused video

anyway

It's all novel for me since I am new to this

but normal mapping now

[astral hinge]

The textures shouldn't look any different tbh

[cloud rivet]

well

[astral hinge]

MSAA only affects triangle edges

[cloud rivet]

I also enabled another thing

[astral hinge]

anisotropic filtering?

idk

[cloud rivet]

shading rate

it was also in that tutorial https://vulkan-tutorial.com/Multisampling

e. MSAA only smoothens out the edges of geometry but not the interior filling. This may lead to a situation when you get a smooth polygon rendered on screen but the applied texture will still look aliased if it contains high contrasting colors. One way to approach this problem is to enable Sample Shading which will improve the image quality even further, though at an additional performance cost:

So I added this:

VkExtent2D fragment_size = { 1, 1 };
VkFragmentShadingRateCombinerOpKHR combiner_ops[2];
combiner_ops[0] = VK_FRAGMENT_SHADING_RATE_COMBINER_OP_KEEP_KHR;
combiner_ops[1] = VK_FRAGMENT_SHADING_RATE_COMBINER_OP_REPLACE_KHR;
vkCmdSetFragmentShadingRateKHR(cf.command_buffer, &fragment_size, combiner_ops)

from the spec:

Each pixel in the framebuffer is assigned an attachment fragment shading rate by the corresponding texel in the fragment shading rate attachment, according to:

x' = floor(x / regionx)

y' = floor(y / regiony)

where x' and y' are the coordinates of a texel in the fragment shading rate attachment, x and y are the coordinates of the pixel in the framebuffer, and regionx and regiony are the size of the region each texel corresponds to, as defined by the shadingRateAttachmentTexelSize member of VkFragmentShadingRateAttachmentInfoKHR.

VK_FRAGMENT_SHADING_RATE_COMBINER_OP_REPLACE_KHR specifies a combiner operation of combine(Axy,Bxy) = Bxy.

oh there's a glossary

Shading Rate
The ratio of the number of fragment shader invocations generated in a fully covered framebuffer region to the size (in pixels) of that region.

idk

was just trying to follow the blog post

[astral hinge]

Sample shading is kinda bad tbh (in terms of perf)

Regular MSAA guarantees exactly one fs invocation per texel where the triangle touches at least one sample

Sample shading just turns it into SSAA with a better sample pattern

[cloud rivet]

idk

[astral hinge]

Do you have mipmaps?

[cloud rivet]

yes

[astral hinge]

I guess the inner part of the roof consists of lots of thin triangles iirc, so MSAA should help with that

The other part is just big rects iirc so it should look the same in both pics

[astral hinge]

Btw on mobile it's extremely hard to inspect pixels

[cloud rivet]

ah yeah there's a lot of geometry there

[astral hinge]

Every app I have either downsamples the image or applies a bilinear filter when I magnify it

That's why you need a magnifier

[cloud rivet]

why are you on mobile

[astral hinge]

Because I just got back from the gym and I'm having a shake

[cloud rivet]

nice

I still have to exercise this evening

based on the msaa tutorial I thought I would do that in like 30 minutes

but nope

[astral hinge]

Even on desktop it's painful because I have to put the image into paint.net (or use sharex's magnifier I guess)

[cloud rivet]

I made a github issue for a magnifier

I will add it

[astral hinge]

MSAA touches a lot of stuff in even a basic renderer

[cloud rivet]

yeah

also every tutorial everywhere is still render pass

unless it's "modern" and then it just uses compute

but nobody talks about shader objects

[astral hinge]

It's kinda a viral AA effect, arguably more so than TAA

[cloud rivet]

or even dynamic rendering

although I didn't even think to look in any of my vulkan books about msaa

I bet they have that

fuck

[wraith urchin]

He's gotta do A2C now too

[astral hinge]

I wanted to use MSAA in my voxel renderer but I started rethinking it after I realized how annoying it'll be (it will likely not be a forward renderer)

[wraith urchin]

Yeah deferred rendering kinda killed msaa in the mainstream

But it just looks so good, I gotta have it

[cloud rivet]

looks cool

hrm

I feel like I should calculate the tangent in my asset packager

hrmmm

I'm gonna do some not very good tangents as a first step

[slender ember]

If you are using something like Assimp you can have it calculate them for you

[cloud rivet]

I am using fastgltf

oh maybe it has something

that's a good idea

[slender ember]

Yeah, not sure what it provides.

[cloud rivet]

btw loved your cargo loading asset building video you're so skilled at blender and uh everything

I learn a lot from your videos

[slender ember]

Thanks, yeah safe to say things didn't quite pan out how I imagined during the stream, but at least we got a cargo rack and a crate out of it 😄

[cloud rivet]

uv maps in blender are super weird for me too, so was happy that it wasn't just me

let me see about tangents in fastlgtf so I'm not doing unnecessary videos

[slender ember]

JUST uploaded this https://www.youtube.com/shorts/rfivu1wUWPg (from yesterday's stream) 😄

[cloud rivet]

already watching lol

I get those notifications

oh

I don't get notifications for shorts

why is that

[slender ember]

Oh, I deleted the notification actually from the channel. Most of my shorts are just short clips of stuff people have seen already, so I don't want to spam them there.

[cloud rivet]

oh ok

true

[slender ember]

I was trying to figure out if Sapphire can be configured to post different kinds of videos (normal, live, shorts) to different channels but haven't found a way yet.

That way people could just mute channel if they are, say, not interested in a certain kind of upload, but meh...

[cloud rivet]

I see other servers use roles for that

instead of channels

so if you want to be notified for one thing and not the other

you take the role

[slender ember]

Yep, I have seen that too but that kinda seems like a workaround.

[cloud rivet]

the music was good too

that's the kind of music I also listen to while working on my project

[slender ember]

Nice to hear, some people have asked me "Could you change the shit music?" 😄

[cloud rivet]

what, it's good

[slender ember]

I have a few "stream safe" pieces that I typically use. I should probably change them up a bit more.

[cloud rivet]

I didn't see anything for tangents so I'm gonna do a first attempt, and if it's ok for crytek sponza call it, but if it isn't I'll do something more complex I have in my math books

just going to yoink exactly what's in the book

wait

well the approach

not going to take copyrighted code 😅

[true moon]

I'll bet you can do it without seams if you just use "project from view"

Which is an easy way to do planar parts like that anyway

[slender ember]

Absolutely, it's not that I don't know how to fix it, but rather an interesting quirk with how the standard unwrapping in blender gets tripped up by this shape.

[true moon]

Honestly blender's UV editing leaves a lot to be desired

I remember when I switched to blender from 3ds max years ago the two things I missed were the nondestructive modifier stack workflow, and the more robust UV editor

Blender's is pretty rudimentary

[cloud rivet]

why did you switch?

[true moon]

3ds max is expensive and my student license expired

Blender is ubiquitous and still really good overall

It's just not better in every way

[slender ember]

I have only done rudimentary modeling in 3DS Max so I can't compare the two regarding their UV workflow.

But I will say that with the ZenUV and Packmaster3 extensions installed the UV workflow in Blender is much nicer.

[true moon]

It just had whole-island modification algorithms like being able to pin certain vertices and then relax the rest in various ways which gave more control over how organic shapes get unwrapped

Yeah I haven't messed with the extensions yet I think I have one UV editing one but it wasn't too helpful

[slender ember]

ZenUV + Packmaster3 =

[true moon]

I'll have to try them

[slender ember]

Keep in mind if the standard UV stuff starts to wear you down.

They are both paid though.

[true moon]

Ah alright

[slender ember]

But nothing compared to a 3DS Max license 🙂

[true moon]

I'll still see what features they have and if they solve problems I have

I don't do that much modeling since I just make intentionally shitty placeholder art for my project with very rudimentary texturing

I have never really tried to turn a high poly hard surface modeling project into a game asset

[slender ember]

Small world. I make intentionally shitty low-effort final art for my project with very rudimentary texturing.

[true moon]

No lol your assets are beautiful

Mine are like this

#1147711082702589973 message

[cloud rivet]

I like your assets, they are on brand for your game

shitty is what my enderman thing from my last game looked like

[slender ember]

Well, I mean I don't even have skeletal animation in the engine, so I couldn't do that...

[true moon]

#art-discussion message

My philosophy

I started out doing high poly stuff long before I got into graphics

But learning to actually move quickly and make game assets required a very deliberate change in mindset

[cloud rivet]

yeah I plan on not doing anything high poly

I think high poly is hard to do when you're building your own engine and everything else

[true moon]

It's a good skill to practice, just separately

[cloud rivet]

I'm gonna be starting real soon

[true moon]

You need to practice both to have a well rounded set of skills but you can skip over it at first

[cloud rivet]

yeah true

[true moon]

Making stuff that is intentionally, comically low quality like that airplane in the message I linked helps stay on course and not get lost in the sauce

[cloud rivet]

I'm thinking about what my character will be like

it's an obsolete worker bot

so I feel like the model should sort of be something associated with whatever it was doing for a job

like if it's a mobile trash compactor or some kind of welder bot or somethign I haven't figured it out yet

kinda funny to be playing a mobile trash compactor trying to escape a place trying to reclaim you for parts

[slender ember]

Strong Wall-E vibes

[cloud rivet]

yeah, I gotta do something original

[slender ember]

Oh, there's place in there to do something original.

Wall-E turned into some sort of philosophical movie about the human condition half-way through...

[cloud rivet]

as pixar is wont to do

hrm

it's not going to be trash related

the obsolete robot escape room game idea is kind of fixed, I keep thinking all the small robots that are already in media, in star wars, pixar, etc and it's probably the wrong approach

it'll come to me

ok I got some not very good tangents I guess I should do some tangent math then, let me yoink what I had in my opengl project and see if that works so I can go to bed at a reasonable time

[true moon]

You don't have to figure it out all up front

You just need enough for the next few months of work and it will evolve organically as you go

Just keep a notes.txt in your project files and write down all you ideas and you'll probably find yourself accruing more ideas than you'll ever have time to implement

[cloud rivet]

yes notes are great

well

this isn't even remotely correct

but it looks cool

that just doesn't even make any sense

there's some ub or something

oh

the black area are positive coords and the rest are negative or have at least one negative component

I have to look at my math tomorrow, this didn't work

shortcuts never work smh

I can just fix this in my asset manager I think the rest of this is probably ok

[brisk chasm]

gltf supports tangents naturally

ah discord, thanks for scrolling

eh just make sure you export them in blender when you export to gltf

[steady venture]

Nice, I'll read this first

[steady venture]

It really is, I watched the video last week. Although some stuff flew over my head, it was still nice to watch, even as I am now.

[steady venture]

Yeah I naively tonemapped the color value after the loop. I presume MJP talks about the correct way to do this on the blog you guys mentioned.

Awesome progress as usual by the way Bjorn

[cloud rivet]

[wraith urchin]

Or use mikktspace

https://github.com/mmikk/MikkTSpace

[hushed creek]

Mikuspace

[brisk chasm]

re HDR

the inotherpeopleswallslivver did a great deal of hdrisms in frogfood (powered by john's venerable kunilingus)

[cloud rivet]

nice

I will be sure to take a look thank you

[cloud rivet]

part of really wants to spend some time on the stupid broken slang visual studio extension and fix all the problems with it

[astral hinge]

hey did you have an issue with putting the vs and fs inside the same file with slang?

[cloud rivet]

nope

works fine

just gotta use the "main" as the name

for both

for some reason

[astral hinge]

@harsh ginkgo

[cloud rivet]

oh no I am being raided

btw I am using shader objects idk if that matters

[astral hinge]

he was having an issue but he called the functions something else

[cloud rivet]

not that weak sauce shader module stuff

yeah don't do that

[harsh ginkgo]

well you gotta compile shaders anyways

OH

[harsh ginkgo]

LAME

[cloud rivet]

also

I compile with

https://github.com/btipling/Rosy/blob/master/src/shaders/script_compile.ps1#L22-L26

[astral hinge]

it seems quite cursed to have two mains in a file, but I think that's because my brain is corrupted by c++

[cloud rivet]

no no

they're not called "main" in the slang

[astral hinge]

oh

[cloud rivet]

that's what I put on the shader object

[astral hinge]

oh does slang just generate entry points with main

[cloud rivet]

https://github.com/btipling/Rosy/blob/master/src/shaders/basic.slang

you can give it a cli argument to use the name

but it's more fun to just give it a file imo

and then it figures it out

it's just when you do that I think you have to use the string "main" on the SO thing

let me find that code

https://github.com/btipling/Rosy/blob/master/src/Engine/Graphics.cpp#L2844

I am pretty sure if I call it something else it freaks out let me try it

[astral hinge]

right so slang just epically hardcoded "main" as the entry point

not too bad

[cloud rivet]

2025-01-30 03:05:06.5875887] [ERROR] [Validation] Validation Error: [ VUID-VkShaderCreateInfoEXT-pName-08440 ] | MessageID = 0x571dc788 | vkCreateShadersEXT(): pCreateInfos[1].pName `wtfmain` entrypoint not found for stage VK_SHADER_STAGE_FRAGMENT_BIT.
The Vulkan spec states: If codeType is VK_SHADER_CODE_TYPE_SPIRV_EXT, pName must be the name of an OpEntryPoint in pCode with an execution model that matches stage (https://vulkan.lunarg.com/doc/view/1.4.304.0/windows/1.4-extensions/vkspec.html#VUID-VkShaderCreateInfoEXT-pName-08440)
[2025-01-30 03:05:06.5887832] [WARN] [Performance] CreatePipeline: failed to parse SPIR-V: 2
[2025-01-30 03:05:06.5890384] [WARN] [Performance] CreatePipeline: unexpected compilation failure
[2025-01-30 03:05:06.5892772] [WARN] [Performance] CreateGraphicsPipeline: unexpected failure compiling SPIR-V shader: 0x3b28c47a8529802c
[2025-01-30 03:05:06.5896968] [WARN] [Performance] CreatePipeline: failed to parse SPIR-V: 2
[2025-01-30 03:05:06.5899360] [WARN] [Performance] CreatePipeline: unexpected compilation failure
[2025-01-30 03:05:06.5901774] [WARN] [Performance] CreateGraphicsPipeline: unexpected failure compiling SPIR-V shader: 0x3b28c47a8529802c
[2025-01-30 03:05:06.5904759] [ERROR] Error creating shaders: 243
[2025-01-30 03:05:06.5907179] [ERROR] Asset setting on graphics failed: 1
[2025-01-30 03:05:06.5927839] [INFO] Engine deinit start
[2025-01-30 03:05:06.5930920] [INFO] Graphics deinit start
[2025-01-30 03:05:06.7344592] [INFO] Graphics deinit end

yah

I looked out the output in renderdoc once

and it's like literally "main" in there, even though it also has the actual function name in spirv too

it's so weird

pCreateInfos[1].pName wtfmain entrypoint not found for stage VK_SHADER_STAGE_FRAGMENT_BIT.

[shader("fragment")]
Fragment fragmentMain(BasicVertex basicVertex: BasicVertex, CSM csm: CSM, float4 fragCoord: SV_Position) : SV_Target

so it's not wtfmain and it's not main

I called my shader basic that's because what it is, basic

I should call the whole project basic

I think my lightdir is fucked

idk

this is in tangent space

like wtf is this

[astral hinge]

sponza just has kinda fucked tangents

[cloud rivet]

uh

well blender doesn't have any issues with it

if blender can render it so should I right

[harsh ginkgo]

You have saved me so much trouble figuring that out.
The triangle doesn't show up, but I'm sure I made a booboo somewhere else.
Being lazy and using "main" for both vertex and fragment was something I thought about , but noooo I assumed I had to use different names...
I wouldn't of figured that out for a long while.

[cloud rivet]

I should try that mittick thingy

[cloud rivet]

I honestly don't remember how I came to use those strings? I think I may have had them in separate files and then learned I could put them in the same file and forgot about it and that just happened to work 💀

[harsh ginkgo]

and this is why I won't recommend using slang for the time being

[cloud rivet]

I love slang though

modules are awesome

I am so glad to not write glsl and have to declare a bunch of glsl extensions and deal with glsl scope

slang has actual scope

[harsh ginkgo]

sure sure, I just mean that as an "authority" I can't earnestly say its 100% ready for primetime

[cloud rivet]

right right

I agree

this is just a hobby for me so I just wanted all the fancy new stuff

ok so this light dir is a clue

ohh

I should just render the tangent

[harsh ginkgo]

same same, this is for 'research' and not production after all

[cloud rivet]

the visual studio slang extension is missing a bunch of quality of life features

[harsh ginkgo]

how do I get vertexID in slang?

[cloud rivet]

BasicVertexStageOutput vertexMain(uint uiVertexId: SV_VertexID)

it's just the same as HLSL iirc

I read both slang and hlsl docs

I have never used HLSL so not sure

[harsh ginkgo]

oh okay, I had :SV_position which was wrong

I don't get output but I do get no validation errors

[cloud rivet]

sv position is what you output I think

[harsh ginkgo]

ah weird, well lemme renderdoc this up

[cloud rivet]

public struct VertexStageOutput
{
    public CoarseVertex coarseVertex : CoarseVertex;
    public CSM csm : CSM;
    public float4 sv_position : SV_Position;
};

I am not a reliable source of information about slang though

or anything

[astral hinge]

but you are a source of information

[cloud rivet]

yes mostly

[harsh ginkgo]

I'm happy with sources of information. Most of mine are the docs and those don't respond to questions.

[cloud rivet]

well I at least have code that seems to work

ok it's the tangents

that's the tangents I exported from blender

hrm

let me look at renderdoc

that's the tangent there for that side

that's the other side

the w is negated on one side

so

float3 bitangent = normalize(cross(normal, tangent)) * v.tangent.w;

so that w

I thought I am supposed to do that

because the cross product

ehhh

those tangents should be the same and they are except for the w

oh actually the tangent is negated on the z too

this looks better

nope

I want to see the tangents in blender

why I got a line like this I guess that's the next thing to look at

[cloud rivet]

blender tangent export is just garbage I think

I will look at MikkTSpace

The math error which occurs from this mismatch between the normal map baker and the pixel shader used for rendering results in shading seams. These are unwanted hard edges which become visible when the model is lit/shaded. They are a source of wasted time and frustration to an artist/designer.

yup

The implementation is designed, specifically, to make the generation of tangent space as resilient as possible to a 3D model being moved from one application to another. That is generate the same tangent spaces even if there is a change in index list(s), ordering of faces/vertices of a face, and/or the removal of degenerate primitives. Both triangles and quads are supported.

hrmmmm

[dry apex]

i gave up here lol

[cloud rivet]

I think I am learning things right now, I was transforming my normals with my inverse transpose and normalizing the results but reading mittick code and lengyel I see they aren't doing that

and that already made it a tiny bit better

I still have that hard edge on the top half circle banners

idk

I'm gonna try and go full lengyel with these findings

the rings of shame man

I'm gonna pause on this, I think I need the fancy lighting to better understand how the normal maps are doing stuff, I am just seeing the diffuse lighting right now

I think:

1. My ingested tangents are correct, they are directly imported from blender
2. My tangent space is correct now, it directly correlates to math I have found in three different sources.
3. My asset to engine pipeline is sending all the correct values.

I feel like it looks better when I remove all my normalize calls in my lighting code

if these were my assets I would just fix the tangents

man my rendering doesn't look anything like

[true moon]

Make the shadows darker

[cloud rivet]

my flowers don't look anything like this either

[cloud rivet]

oh let me try that

[true moon]

A lot of looking good is just the values, just like painting a picture

I.e. "art direction" in the gaming world

[wraith urchin]

AO would also add alot of detail

@cloud rivet my sponza tangents for reference

in viewspace

+z forward left handed

[cloud rivet]

oh same

except

you don't have those rings

how come you don't have those rings :(

like my circel banners have a hard edge down the middle

how did you get your tangents?

[wraith urchin]

I don't use the tangents from the file, I get my tangents from mikktspace

[cloud rivet]

is there documentation for how to use it, or just read all the comments in the file? I started on that

I mean the comments are fantastic

[wraith urchin]

The comments should do it

deccer actually had a good example somewhere

for reference here is my sponza

[cloud rivet]

damn

looks incredible

[wraith urchin]

thanks

AO does alot of heavy lifting

without AO:

[cloud rivet]

yes that makes a big difference thank you

we're all just posting the same picture over and over again

[wraith urchin]

wait are we using slightly different sponzas?

[cloud rivet]

I used the one you linked

the khronos one

[wraith urchin]

the color of the drapes looks different

[cloud rivet]

uh could be a me doing something wrong issue

[wraith urchin]

Maybe I'm not using the khronos one lol

[true moon]

Intel one?

[cloud rivet]

this is what mine looks like in blender

so maybe I am just doing it wrong

oh

maybe it is different directions?

[wraith urchin]

Oh no I thing I'm just looking from the wrong side lmao

[cloud rivet]

yeah

[wraith urchin]

yeah it matches now lol

[cloud rivet]

so I just like

[wraith urchin]

It wouldn't load this fast lol

[cloud rivet]

give one mesh at a time to mikkt?

[wraith urchin]

This is GP in a nutshell

[wraith urchin]

thats what I do

Damn it looks like deccer nuked the HowToScript repo

he had a decent mikktspace example in there

[cloud rivet]

oh wow

that must have been recent

I saw it like a few days ago

[wraith urchin]

regarding your rings, note that the texture only has one side of the pattern, this means there must be a UV seam running exactly along the ring

[wraith urchin]

Are you sure you didn't mix up your tangents and bitangents?

[cloud rivet]

    float3 normal = v.normal;
    float3 tangent = v.tangent.xyz;
    float3 bitangent = cross(normal, tangent) * v.tangent.w;

    float3 view = sd.cameraPosition.xyz - v.position.xyz;
    view = float3(dot(tangent, view), dot(bitangent, view), dot(normal, view));

    float3 light = float3(dot(tangent, sd.sunlight.xyz), dot(bitangent, sd.sunlight.xyz), dot(normal, sd.sunlight.xyz));

^^ this is from lengyel

[wraith urchin]

yeah this is the shader, but how does that data get to the shader

[cloud rivet]

it looks the same with a matrix

just export from gltf

I don't modify it

[wraith urchin]

yeah but there is obviously code in between

[cloud rivet]

it's just writes std::array<float,N> and sends it to the gpu

it gets saved to disk that way, gets pulled out from disk the same way

I literally do no math on this, they stay std::array

[wraith urchin]

did you transform the normals & tangents by the object transform?

[cloud rivet]

no

lengyel's math doesn't do that

[wraith urchin]

you should lol, the 3x3 component

[cloud rivet]

if I do that it looks completely broken

no

[astral hinge]

I assume v contains world- or view-space vertex info already

[cloud rivet]

this is in tangent space

it's not in world space

yeah

my normals/tangents etc never go into world space

and nothing is being scaled or changed in any way

I think if I were doing any of that wrong more things woudl be broken besides those rings

[wraith urchin]

In the vertex buffer these vectors would be in object space

right?

you need to transform them into world space I think

actually can I just see your vertex shader?

the whole thing