GPU Zen 3: Virtual Shadow Maps

[sweet nimbus]

yeah it's just supposed to show scale I guess

bistro is where actual detail is

[hazy steppe]

im super proud to know you guys 🙂

[sweet nimbus]

hmm

are these pics fair if I use TAA in both of them

[wooden jolt]

as long as we mention TAA is fine

[sweet nimbus]

ok

idk how they pack figures next to each other in tex so I'm just gonna combine them in paint.net

[wooden jolt]

we can leave formatting and beautification for last methinks

[sweet nimbus]

huh, that terrain has slightly more vertices than bistro

4.1 million vs 3.9 million

[wooden jolt]

@mystic lark [sampling], should we mention both "inserting a new allocation on page fault when sampling" and "mark visible pages in a predefined radius instead of marking only a single page"?

[mystic lark]

Uhhh the new allocation on page fault imo makes no sense

It will flicker no?

Mark in a radius should be good I think

[wooden jolt]

unless you mark those allocations to not be freed somehow, yes

[mystic lark]

Yeah, then I'd just avoid it

[wooden jolt]

ight

[mystic lark]

And maybe say that we have a coarse page set as we were saying in tido with Jaker

And we can use that as a fallback when we get a page fault

[wooden jolt]

epic, aight

I'm scared to change what jaker already wrote

[mystic lark]

You can just add it below and I'll merge it in if you dont want to change stuff

[vestal snow]

Dang this results pic ended up looking really cool though. Nice!

[acoustic bobcat]

I didn't have time to read it again btw

I'll take a look when you get feedback from the real editor

[mystic lark]

I just awoke, I kinda fell asleep right as I got back home

but I see no notes from the editor so far

[sweet nimbus]

They have a lot to go through so I'm not surprised it's taking them a while

[sweet nimbus]

looking through the articles, I don't think this review period has actually started for any of them. However, there is are some very thorough review comments from a few weeks ago and earlier

[acoustic bobcat]

Yours included? I can do another editing pass tomorrow if it's still not been commented on

[sweet nimbus]

yeah nothing yet. everything is in limbo lmao

[acoustic bobcat]

It takes a while

I remember once being assigned to edit 200 pages of a 1000 page manual took me like a month to get through

Very draining process compared to editing just 10 pages or something

[sweet nimbus]

even editing just 10 pages is pain, but I suppose it doesn't help that I'm also helping write it and we're on a tight deadline

[acoustic bobcat]

Normally you have each author go through it separately and write comments which always takes like a month because people forget or whatever so it gets hammered out over a longer period of time

[sweet nimbus]

we were sorta doing that at a hyper pace. like 16 hour periods where we do a full editing pass + some writing

but the dust has kinda settled even though the conclusion isn't completely done and we're still in limbo

[hazy steppe]

https://tenor.com/view/you-have-no-choice-jade-jade-joddle-engvid-you-dont-have-any-other-option-gif-26827864

:motivationalspeech.gif:

[sweet nimbus]

@vestal snow for the perspective & projective aliasing pic, I think simple pics of the issue would suffice. this article has some good pics
https://learn.microsoft.com/en-us/windows/win32/dxtecharts/common-techniques-to-improve-shadow-depth-maps

[sweet nimbus]

huh that 1993 paper (Hierarchical Z-buffer visibility) really does seem like the origin of hi-z culling

but the version we use, being gpu-driven, is certainly a newer thing that derives from it

and their version has some weird quirks

While the basic Z-pyramid test can reject a substantial number
of polygons, it suffers from a similar difficulty to the basic octree
method. Because of the structure of the pyramid regions, a small
polygon covering the center of the image will be compared to the
Z value at the coarsest level of the pyramid. While the test is still
accurate in this case, it is not particularly powerful

[mystic lark]

Uhhh hmmm I kinda just looked up hiz culling, saw paper with 1k citations and assumed that's it

[sweet nimbus]

yeah idk it's probably fine

[prime ice]

they are dummy and dont sample a 2x2 quad, which solves the issue

[sweet nimbus]

I wrote a little in the results section and added a pic to the beginning. We still need a table with a benchmark for the passes

the table should just show how long each pass of VSM took in a specific scene

[mystic lark]

Do we do a fly through and average the time?

It also heavily depends on the sun angle

[sweet nimbus]

I was thinking of an nsight capture because you don't have per-pass metrics in your renderer yet

[mystic lark]

I will add today

Should take like an hour

[sweet nimbus]

ok

something like this except with just one GPU (taken from GI-1.0 paper)

I was thinking of doing table for a static scene, then the same scene but with caching disabled to show the difference

a flythrough would be ideal but idk how hard it would be for you to do. I guess you can just do it by hand

[mystic lark]

I can hack in camera following a predefined path (I have that in another project) and we can then benchmark on multiple gpus

[sweet nimbus]

and ofc a flythrough would require those metrics and some way to average them (which tbf is easy with an exponential weighted average)

we don't need multiple GPUs

just one will get the point across, which is that it's cheap to render

[mystic lark]

Okay, well the path will be good regardless

So that we get consistent results for caching and no caching

[sweet nimbus]

if you want to add the perf counters and camera path in engine, that fine, but I think those are not strictly necessary to get the data we need

I just want to be sure that we can complete this

[mystic lark]

Yea, perf counters I'll definitely add, camera path probably not then

For caching the counters are needed imo

Because the perf jumps up and down a lot with caching

[sweet nimbus]

I was only ever considering using nsight, not looking at the entire frame time

[mystic lark]

0.12ms - 1.2ms

We need to average to see the actual perf impact imo

[sweet nimbus]

ye

[mystic lark]

Which would be hard to do with nsight

[sweet nimbus]

nsight can actually capture and average multiple frames

[mystic lark]

Okay nevermind I'm just skill issued then

I'll look into it

[sweet nimbus]

before you launch your app, you set it to capture multiple frames

then you check the "aggregate frames" box in the timeline after capturing

btw I do not find the "batch 10 task 0" prefix very useful in this

[mystic lark]

One issue I've had with that is the capture actually freezes the frame making the frame delta time spike up and teleport the camera

[sweet nimbus]

yeah it won't be useful unless the camera is stationary

[mystic lark]

There should be a name of the task afterwards no?

[sweet nimbus]

there is

[mystic lark]

It's DAXA thingy I can talk to Patrick about that

[sweet nimbus]

it gets cut off unless I zoom in a lot

I don't care about that stuff as the user 😄

(and also the source language)

[mystic lark]

I feel like this can be misleading, as caching will be nuclear speed with stationary camera

[sweet nimbus]

yeah

[mystic lark]

But once you move the perf will obv drop

[sweet nimbus]

that's why I suggested testing without caching as well

but a path with averaged numbers would be ideal

[mystic lark]

Yeah good point @prime ice

[sweet nimbus]

my suggestion is to move the "slang" and "batch N task M" stuff to the suffix so it doesn't impede what I care about

[mystic lark]

We should make it a task graph option to disable/enable the prefix

[sweet nimbus]

or just make it a suffix

[mystic lark]

The slang draw is the actual name of the task I believe

[sweet nimbus]

ah lol

[mystic lark]

So slang prefix is app side

Or maybe the name of the pipeline

But those are good points, well change it

[sweet nimbus]

yeah it says pipeline name on the bar with the slang stuff

[prime ice]

yea suffix might be better

[mystic lark]

Do you want to put an image of the nsight capture into the article?

Or just the values in a table

[sweet nimbus]

just in a table

but if you do the benchmark with in-app counters then you can skip the nsight thing

[mystic lark]

Benchmarks will be done after you wake up

(if you don't wake up earlier than in 5 hours )

[sweet nimbus]

well I do have a meeting in 5 hours so

[mystic lark]

Fug

Welp

[sweet nimbus]

maybe potrick can help you

[mystic lark]

He is in Poland

I believe

[sweet nimbus]

I am in poland, but without the la

[prime ice]

nice

[sweet nimbus]

we still need people to review changes in the article btw. there's so much green

I cleared some just now, but 1. a bunch of edits are written by me and I want at least one other person to see them; and 2. I'm about to sleep

@wooden jolt do you think you can find some time to review edits in the article

[wooden jolt]

yes

[mystic lark]

I'm forced to wait for irl shit

[wooden jolt]

I'll be home soon™️

[sweet nimbus]

alright cool

[wooden jolt]

ETA 30 mins or something

[mystic lark]

So will work late again

[sweet nimbus]

no need to actually write much (though that would also be appreciated lol). just make sure things are sane

[mystic lark]

I went through the entire article on my phone. I accepted the changes and made some minor corrections. The later sections (sampling and conclusion) still need a bit of cooking, I'll work on them once back on PC

Someone still needs to review the sections I added, they are still pretty rough and require rewording and polishing (I can also do that if no one does it before I get back)

[vestal snow]

Ok I’ll try to put a similar picture together today

Unless you have an idea for it already and plan to upload

[mystic lark]

Tido now has timings

[sweet nimbus]

I already put such an image in

Shortly after I said that hehe

[vestal snow]

Awesome

I’ll delete the picture thing I uploaded yesterday

[mystic lark]

okay so caching matters a lot lol

left no caching right caching

Is this enough for you Jaker, or how do you imagine the benchmarking process to go?

(I pushed so you should also be able to benchmark yourself if you want)

[sweet nimbus]

yeah that's perfect

[wooden jolt]

we could do fancy shit by writing the timestamps to a CSV file and making a graph out of it

[sweet nimbus]

csvprofile my beloved

[hazy steppe]

the chance to make afps a concrete term too

[mystic lark]

I did ewa_exec_time = 0.99 * ewa_exec_time + 0.01 * curr_exec_time for all of these, hopefully I understood correctly

I can also cook up some graph if desired, but I think it's bikeshed?

[sweet nimbus]

graphs would be cool but I don't think are necessary

also imo they would be more useful if you had a fixed camera flight path

[mystic lark]

okay, I'll focus on text for now

Btw, when fixing the sampling phase text I started to think

when sampling for PCF for example, wouldn't it be better to just figure out the cascade index directly by the same heuristic we use for the main pixel, instead of attempting to "blindly" look into the same page + page in clip above + page in clip below?

[sweet nimbus]

maybe, but doing that N times per pixel could get pretty expensive

I haven't tried it though

[mystic lark]

uh hmmm

yeah

I think we can combine it actually, attempt to look into the same page (we can trivially determine if the sample still lies in the same page) and if it crosses to another page, we can use the heuristic

[sweet nimbus]

I actually forgot that my impl blindly looks up and down one level lol

[mystic lark]

I saw that in LVSTRIs, I don't know if yours does that too 🕵️‍♂️

[sweet nimbus]

mine does 😎

[prime ice]

now this is epic

[mystic lark]

I added timestamps into Tido vsm pass

and I think I already have an idea on how to integrate these into TaskGraph

[prime ice]

nice

god mode

[mystic lark]

Page faults also occur when sampling the VSM for shadows in participating media and translucent objects. - do we know how to fix traslucent objects?

run the marking pass without translucents, and then while drawing every translucent pixel?

[sweet nimbus]

coarse pages or the thingy where page faults create allocation requests

[prime ice]

oh hod

cant you simply run the mask code within trandlucent pixel shader?

[mystic lark]

yeah that is what I'm suggesting

[sweet nimbus]

I think that's also the second idea I wrote if I understood right

unless you mean drawing translucent stuff an extra time just to mark pages

[mystic lark]

aha, I assumed you meant while sampling the shadows (IE the same issue occuring during PCF) not while drawing

[sweet nimbus]

well translucent stuff would be forward so it would be both, no?

unless you are doing dithered transparency

[mystic lark]

I don't follow

[sweet nimbus]

translucent/transparent geo would be drawn in a forward rendering pass in which shadows would be applied

i.e. they wouldn't be deferred as you seem to imply

[mystic lark]

oh I see, yeah in that case I'd just draw them twice

[sweet nimbus]

JS said that he does the allocate-if-page-fault thing for that and volumetrics

idk how bad the artifacts are, but I bet it's pretty easy to implement

[mystic lark]

yeah no, that will be simple to do

maybe the TAA can eat it away

(shame that there is no TAA yet in tido)

[mystic lark]

I rewrote and expanded the Sampling section, so it is ready for a review from someone else

[prime ice]

ah i expected a prepass for mbt

[sweet nimbus]

mboit?

[mystic lark]

I assume Patrick means moment transparency

https://briansharpe.wordpress.com/wp-content/uploads/2018/07/moment-transparency-av.pdf

[sweet nimbus]

that's the same thing as mboit, just from some different people

apparently they were researched independently in parallel

[mystic lark]

ah I didn't know that is also a term for it

we probably should add some images into the sampling phase to break up the wall of text we have rn

[wooden jolt]

I think I'm having some trouble understanding a sentence

[mystic lark]

send

[wooden jolt]

Then one of three actions is taken, depending on the state of the page stored in the page table. If the page has not yet been allocated, its coordinates are added to the buffer storing allocation requests for this frame. If the page is already allocated, but not yet marked as visited, we mark it as visited this frame. Lastly, if the page is allocated \textit{and} marked as visible, we do nothing. in the marking visible pages section

why exactly is the allocation state of the page relevant here

is it caching related

[mystic lark]

well yeah, if the page is already allocated from one of the previous frames you do not want to allocate it again

you only want to allocate the pages that are not yet backed by the physical memory (the ones that became visible this frame)

[wooden jolt]

yeah but this is the the marking visible pages section no?

unless I'm misunderstanding

[mystic lark]

sus name yeah

it should be called something different

uhhh

[wooden jolt]

in my impl marking visible pages is a shrimple unconditional imageStore

the allocation comes later

[mystic lark]

how do you know which pages to allocate?

[acoustic bobcat]

It's pages that were active last frame that are being discussed here right

[mystic lark]

they could have been active any frame in the past, not just the last frame

but yeah

maybe the pass should be called something like classify pages?

[wooden jolt]

if the current page is not backed but visible, an allocation request is emplaced, otherwise if the current page is not visible but backed, it is freed

[mystic lark]

ye I don't free, but you have the same logic

[wooden jolt]

it should be noted that the virtual page table and active page table are different entities in my thing

but I suppose it doesn't matter what I do

if you say it's fine then you get an

[mystic lark]

maybe the diagram will make it more clear

[wooden jolt]

ye so you access the virtual page table immediately during the visible page marking

which makes sense tbh

[mystic lark]

In my impl I don't free the pages unless the memory pool is full and new pages request allocation

so what I do in the first pass is mark all pages that are visible this frame (so I don't free them later) and store all pages that are visible but not yet allocated (cached)

[wooden jolt]

ye this is fine

I was biased

[mystic lark]

I see, so confusion cleared?

[wooden jolt]

yes

[mystic lark]

hmm maybe I should reword this part to make it clearer?

what do you think made you confused?

[wooden jolt]

the diagram is good as well so we confirmed that it works

[mystic lark]

BTW we really should unify the naming, sometimes we use visited and sometimes visible

[wooden jolt]

I think we could mention "VPT" instead of page table generic here

[mystic lark]

I suggest visible as it is more in lines of what actually happens

[wooden jolt]

ye visible is better imo as well

[mystic lark]

@vestal snow this change (visited -> visible) should also be reflected in the diagram/s 🙏

[vestal snow]

Sounds good I should be at the computer soon and can push diagram changes

[wooden jolt]

one more nitpick

should we really mention implementation details such as mesh shaders in the HPB culling section?

I personally think it's fine the way it's written

[mystic lark]

Uh oh yeah good point, I was also thinking about that and I'm not sure

Perhaps I go way into detail about the mapping and such

And simply saying "we use meshlets and mesh shaders" would suffice?

[wooden jolt]

the reason I nitpicked about this is because in other papers you would usually see impl details in the "Our Results" section or something like that

so this is definitely not wrong per se, just probably uhh misplaced a bit?

[wooden jolt]

perhaps this is best, but the "removed" part should definitely go into the results section

"In our implementation [...] we performed HPB culling against meshlet bounding boxes, giving us good culling granularity [...]"

[mystic lark]

Good idea, yeah, we can also move the section where we describe our selected vsm resolutions and stuff down into the results section

[wooden jolt]

sampling is gud btw

imma try to fix the yellowy thingies

[mystic lark]

What do you think about removing the split into bookkeeping, drawing and sampling all together and instead just having the individual steps?

I feel like the split is a bit disproportional in content as is rn

Especially if we remove the implementation details from the drawing

It will become quite plain compared to the bookkeeping part

[wooden jolt]

wym disproportional?

too much/little text between sections?

[mystic lark]

Yeah

[wooden jolt]

hmm

[mystic lark]

Not between but in the sections themselves

Look how long the bookkeeping part is

And how long the drawing part will be

[wooden jolt]

inherent flaw of VSM: it's all bookkeeping

I think the split is useful personally even if the content is very disproportional

@sweet nimbus what do you think

[vestal snow]

I have a question

In the heuristic we mention the frustum side length of the first cascade. Is it true that the first cascade side length determines all other side lengths in the Tido impl?

like second one is 2* first side length, second is double again, etc.

[mystic lark]

yep

[vestal snow]

ok cool

diagrams should be updated
perspective/projection one was deleted since Jaker's replaces it
cascade heuristic diagram is still in the review section at the end

[mystic lark]

I'll take a peeksie at it later today/early tomorrow

[sweet nimbus]

I think clip space should be two words

[sweet nimbus]

I think it's okay

[mystic lark]

Okay, let's leave it as is

[sweet nimbus]

Hmm maybe removing it would be fine lol

How many sub-steps are there in the bookkeeping section anyway

[mystic lark]

5

So we would instead make it 7 steps where one would be culling and drawing and the other would be sampling

[sweet nimbus]

Hmm maybe we keep it then

If there were like 2 substeps then it would be more obvious to merge them

[mystic lark]

I didn't mean to merge the sub steps

I meant to make drawing and sampling sub steps themselves

But it is probably bikeshed

[sweet nimbus]

Yeah that's what I meant but wasn't clear in my wording

[vestal snow]

Should we have some kind of showcase of a few scenes with really nice shadows to put at the very start

I just realized our first image is of artifacting shadows lol

[sweet nimbus]

lmao I had that thought too. I think it would be a good idea

[mystic lark]

Here is our article - let us start with the artifacts

[sweet nimbus]

the picture of projective aliasing is with default vsm settings in my renderer too

I mean we can just be clear about that in the results: "vsm does not fully solve projective aliasing" etc.

there are some nice pics in the results section already fwiw

[mystic lark]

In reviews I've always been told to use assertive language in papers so "VSM is a flawless technique solving all issues other shadowing techniques struggle with"

[wooden jolt]

btw VSM does not prevent you from using perspective skewing so you could advertise it with LiPSM or TSM

[mystic lark]

I already feel kind of weird talking about sampling for participating media and translucent objects when none of us implemented either

[wooden jolt]

but then pages are non uniform

[wooden jolt]

real

[vestal snow]

participating media like volumetrics?

[sweet nimbus]

that should go in future work

[vestal snow]

let me upload the VSM volumetrics test v1

VSM + Volumetrics + TAA

[mystic lark]

Oki, we can move that part or remove it if you want

[sweet nimbus]

I'm not familiar with these techniques but I'm assuming this is a joke suggestion

[wooden jolt]

I'm half joking

[sweet nimbus]

I think we should still discuss it since it's an important feature of a renderer

[wooden jolt]

mathematically VSM + LiPSM is possible

but this is bikeshed

it could go into the future work section

[mystic lark]

You can read what I wrote and decide

I should be able to implement it in Tido till may or whatever the deadline is

Eh probably not transparent objects

These will probably take more work to integrate with the rest of the pipeline

[sweet nimbus]

yeah just leave it in future work. you already have enough on your plate as it is

[mystic lark]

True dat

[vestal snow]

Is it allowed for us to mention certain things and just make it clear that we aren't bundling it in the example code, but either talk about what we tried or what UE does for it?

[mystic lark]

I don't think we should talk about what UE does

[vestal snow]

Ok, what about things we tried but aren't bundling?

[mystic lark]

If one of us tried it I think we can say that

[vestal snow]

such as volumetrics and the depth culling test you guys did

[mystic lark]

Ideally we should bundle everything we talk about with the example (Tido)

There is some time to implement stuff still at least from what I understood

But I'd avoid speculating about stuff we are not sure will work

Or just leave it to the future work section and Jaker said

[vestal snow]

Alright that makes sense to me
My volumetrics test is still kind of in a meme state so idk if I will be able to try and contribute it to Tido before the deadline

[mystic lark]

Do you use the same approach described in the sampling section of the article?

Actually I really should sleep now, I'll be back tomorrow

[vestal snow]

I read through it and I think my example code is using pretty much the same steps

[mystic lark]

Alright, then we at least know it should work

Btw are you doing the "project frustum" part or are you requesting pages to be allocated when you encounter them missing while sampling?

[mystic lark]

Btw funny observation, caching perf is 4-5 times better as long as you don't move very close to the surface

[sweet nimbus]

ah yes I need to test that

[mystic lark]

I mean it makes sense as when I fly close to the surface pretty much everything cached gets cleared and invalidated each frame, but id still expect mainly drawing a single cascade to be faster than 1400us

[sweet nimbus]

nsight would probably tell you something

[mystic lark]

But maybe the atomic min contention is going ham

[mystic lark]

Yeah I should take a capture

[sweet nimbus]

we don't need to root cause it atm

[mystic lark]

True

[wooden jolt]

I'm 99% sure of that too

have you seen the mesh for the trees and bushes in bistro

[mystic lark]

I feel like the vsm journey is me and LVSTRI being like "mhmmmm what iff" and Jaker standing with a ruler above us and slapping our fingers "No! Not needed, write the article"

[wooden jolt]

it's all self intersecting quads

[mystic lark]

Yeah I'm leaning closer and closer to not caching the lower cascades at all

[sweet nimbus]

btw I don't see the caching artifacts anymore

I think I triggered it just by enabling caching and moving around while close to a surface

but ye it's good now

[sweet nimbus]

how about we put that in future work

[mystic lark]

Yep, this is all hypothetical

It just feels bad, as our impl could have been so much better/tried out many more thingies

[sweet nimbus]

perfection is the enemy of good enough

we did what we could, given the circumstances

[mystic lark]

https://tenor.com/view/we-did-everything-we-could-do-captain-kareem-moore-mayor-of-kingstown-we-did-what-we-could-we-did-our-best-gif-24147651

[sweet nimbus]

I accepted a bunch of edits and made a few of my own. it seems like the list just keeps getting longer

https://tenor.com/view/subida-gif-18379274

[mystic lark]

Good changes! I reviewed/accepted all of them and made one or two tiny adjustments - asymptotically this surely must be converging to 0 changes

There is still a bunch written by me that needs to be reviewed though

[sweet nimbus]

made some adjustments
it's joever

[vestal snow]

I double checked and current experiment code has both

Hmm I’m reading unreal’s docs and the volumetrics test almost definitely will need more time than we have. I found this: “In a future release, a more elegant solution would be where some of these effects can be marked for localized pages directly in advance rather than the current, overly conservative coarse pages being used.”

Seems like they’re still experimenting and want to move away from having to mark coarse pages?

[sweet nimbus]

@mystic lark do you mean for this to say "light frustum", then "cascade frustum"?

As the light frustum moves to follow the main camera, new pages, previously located on the edge just outside of the cascade frustum, might need to be drawn.

I think it should be "light frustum" both times

[mystic lark]

Yes, cascades frustum for both

[sweet nimbus]

went through the stuff in the drawing phase

now someone just needs to review the sampling phase

Shit I forgot. Someone should copy the contents of the article to the actual book just in case. We've made quite a few changes since then

[prime ice]

how many pages do you have now

is there a link i can read?

[mystic lark]

Currently is 17 pages

[sweet nimbus]

I can add you as a collaborator

Just dm me your email

[acoustic bobcat]

Have you guys gotten any editor feedback yet?

[mystic lark]

Nothing in the offical document so far

[sweet nimbus]

@prime ice have you read the article yet?

[prime ice]

no

[hazy steppe]

https://tenor.com/view/sinnlos-im-weltraum-fiese-fresse-siw-gif-27283330

(wpotrick knows what is about, just in case anyone is wondering wtf that is :D) @prime ice go read it.

[mystic lark]

Lazy German, bad

[prime ice]

yes

[sweet nimbus]

@mystic lark there is a part of the sampling section that I might rewrite

there is a paragraph about participating media and another for translucent stuff, but really they suffer from the same problem

the solution of drawing translucent stuff twice can still be mentioned

[mystic lark]

I initialy had them as a single paragraph and then I rewrote it to be two

but you can definitely merge it

[sweet nimbus]

I will try, inshallah

[mystic lark]

I need to be bullied into writing the final section

I'm procrastinating on it so hard

[sweet nimbus]

you don't have to write it

your only task atm is collecting them 🅱️erf numbers

[mystic lark]

ok, that sounds better

[sweet nimbus]

readers will only care about having a rough idea about the perf of VSM or any other technique

[sweet nimbus]

I realized that the implementation doesn't even support these (afaik), so maybe this info should go in the future work?

[mystic lark]

yeah either that or we rely on JS implementing them

[sweet nimbus]

lol ok then I'll move it to the future work for now

[mystic lark]

I mean we rely on the fact that this works because he implemented them

[sweet nimbus]

I know they work, but they don't exist in your impl (the demo we are shipping)

[mystic lark]

ye(t)

[sweet nimbus]

we have theoretically about a week to prepare the demo

[mystic lark]

A weeek????

uhhhhh

[mystic lark]

well

okay

[sweet nimbus]

is that too little time?

[mystic lark]

nah it's fine

just gotta get to it

[sweet nimbus]

the main things is getting a tiny benchmark (which you seem to almost have) for a table and the invalidation for moving objects

[vestal snow]

Did you guys hack in the moving objects invalidation to the demo?

[mystic lark]

It's in the workings

Should be ready tomorrow

[vestal snow]

Nice that’s awesome

[sweet nimbus]

!remindme 10 hours finish the conclusion

[lost jay]

Alright gpgpu, I'll remind you about finish the conclusion in 10 hours. ID: 67610154

[sweet nimbus]

I need to be whipped into shape

[hazy steppe]

[sweet nimbus]

I got too distracted with the vsmisms in my own renderer

[hazy steppe]

!remindme 9.5 hours whip jaker back onto track

[lost jay]

Alright deccer, I'll remind you about whip jaker back onto track in 9 hours and 30 minutes. ID: 67610452

[hazy steppe]

dogjiffalternativeversion.gif comes to mind

[mystic lark]

!remindme 10 hours write dynamic object invalidation mask

[lost jay]

Alright m_saky, I'll remind you about write dynamic object invalidation mask in 10 hours. ID: 67610485

[sweet nimbus]

https://tenor.com/view/dog-dogs-husky-slap-bitchslap-gif-10529184

That was for myself

[hazy steppe]

doggos' eyes are the best

[prime ice]

lol

[sweet nimbus]

we need some pictures of nice, c r i s p shadows for the VSM article header still

or not lmao. I'm not too opposed to having the first pictures be commonly-encountered shadow mapping artifacts

[sweet nimbus]

I wrote some more stuff at the end of the article

[sweet nimbus]

https://tenor.com/view/lebowski-money-toilet-wheres-the-money-lebowski-gif-24292554

except it's "where's the benchmark"

[hazy steppe]

when it doe be?

24th?

[sweet nimbus]

I dunno, the timelines are all wack now 😭

[hazy steppe]

@wooden jolt @prime ice @mystic lark @vestal snow

https://tenor.com/view/nanka-lazy-work-paint-dont-work-too-hard-gif-17699381

[mystic lark]

Oh uhhh I thought the ones I sent were all you needed

The two screenshots with timestamps that is

And then I forgor

What exactly is required of me?

I'll do it first thing tomorrow, sorry for my recent inactivity, uni is taking it's toll

[hazy steppe]

just checking because the conversation all of a sudden stopped here

and there was no "ALL STUFF SENT, REVIEWT, WE GOOD"

[mystic lark]

I've still been checking the main article each day - no reviews yet ☹️

[hazy steppe]

ah its wolfgangs turn now?

[mystic lark]

Since it's our first (and only ) feedback it's kinda important for us

[mystic lark]

Ye

[prime ice]

me rn

[mystic lark]

I've been mainly fixing Tido bugs over the evenings, so that it doesn't crash on every second user action

And is actually presentable

[prime ice]

the crashes are the spice

[mystic lark]

Half of the bugs were yours mister Bingus

And I'm not even taking into account that you tried to gaslight me into adding some arcane swapchain extension to fix them

[hazy steppe]

just say a word, ill ban potti for incompetency

kissing teeth

[mystic lark]

It's fine, he is a slow learner, but he does learn

(ily Patrick don't hurt me please)

[sweet nimbus]

I must have dementia

when did you send that

[prime ice]

hahahaha

nearlz worked

[prime ice]

i will kiss zou

[mystic lark]

Here

[prime ice]

hte day saky leaves me is when i end my minecraft world

[sweet nimbus]

@hazy steppe can you pin so I don't forget again
#1168692074447642664 message

[prime ice]

https://tenor.com/view/star-wars-warp-millenium-falcon-gif-13239597

[mystic lark]

The day you write a coherent sentence is the day I cry with happiness

[prime ice]

dont push me like that

you know that wont happen

[mystic lark]

Hehehe yeah I know

[prime ice]

i have a broken word quota to achieve

[mystic lark]

Parsing your sentences has been a worthy investment of my skill points

Btw Jaker it should be noted that the "mark required pages pass" heavily depends on the page size (128x128) in those measurements

It almost doubles when making pages smaller

Not sure if it's relevant, but just so that you know

[sweet nimbus]

damn maybe we should do benchmarks for a few more vsm/page size combinations

small/large page and small/large vsm maybe (so just four total)?

[mystic lark]

okay, I can do that

!remindme 12 hours MEASURE VSM LAZY

[lost jay]

Alright m_saky, I'll remind you about MEASURE VSM LAZY in 12 hours. ID: 67661631

[sweet nimbus]

sorry for bothering you when you already did the original benchmark

[mystic lark]

oh no worries at all, I'm glad I can help

[sweet nimbus]

epic table

I think we need to merge all the pre-draw passes lol

for now I will just sum them but it will be somewhat incorrect

perfect fit

whoops here's context

I'm not sure what the debug passes are exactly

In my thing the debug visualizations are just in the shading pass

[mystic lark]

They can be omitted I don't actually know why I even timed them

They just generate an image for imgui

[sweet nimbus]

Ah lol

I'll split the bookkeeping a little then

[mystic lark]

Should I also include sampling measurements

[sweet nimbus]

Wdym

[mystic lark]

As in filtering etc

Measure the pass that does the shadow lookup during shading

[sweet nimbus]

Do you have filtering

[mystic lark]

I have pcf

[sweet nimbus]

Are you doing the radius marking thing

[mystic lark]

Crappy one but that shouldn't matter

[sweet nimbus]

How does it affect perf

[mystic lark]

I'm not, I'm just looking for allocated page

So cascade chosen by heuristic, one above and one below

Uh 8 samples is like 0.6ms in 1440p I think

16 samples with super sampling were around 4-5ms though

But that's an extreme case

[sweet nimbus]

my friend suggested using a stacked bar graph instead of a boring table

[sweet nimbus]

trying to make this look good is cancer. maybe that's why no one else does it

[prime ice]

those are pretty numbers

there should be some testing for the sampling perf as well

[sweet nimbus]

yeah we discussed it

I'm just trying to make it look presentable

[humble shadow]

You need better colors 😄

[sweet nimbus]

ye

that is obvious

[sweet nimbus]

I am looking at other papers and some of them don't have conclusions (they just end on results + limitations + future work)

the 1-sentence conclusion I wrote sucks so I just commented it out for now

anywayyyyyyy Someone™️ should review what I have at the end currently

I need fresh eyes on that part because I've been looking at it for too long and my neuron has calcified

[sweet nimbus]

@wooden jolt I nominate you, as saky is busy profiling the sampling pass

[wooden jolt]

as you wish boss

[sweet nimbus]

basically I'd just like for you to pls review the text in Results and thereafter

• is it missing any important information?
• is there unnecessary information (I'm afraid much of the future work might be)?
• basic editing concerns like poorly-worded sentences, misspellings, etc.

I've lost the ability to tell what it needs

My watchful gaze will fall upon the white sheets once more tomorrow

[wooden jolt]

@sweet nimbus I reviewed the new sections btw, they look good but I think we should at least mention HZB culling right?

I managed to implement it (unoptimally) and it works wonders

what saky explained is probably how to correctly do it

or 🅱️erhaps we just shove all of that and performance optimizations into the future work & improvements section

[vestal snow]

What’s the font we’re using? I’ll switch the diagrams to use it today if we know it

Hopefully we get some Wolfgang comments soon

[sweet nimbus]

I think Peter will be reviewing our section

[mystic lark]

Okay I'm 95% there, I improved the timing collection and everything, but I am super tired, is it a big issue if I give the timings tomorrow?

[sweet nimbus]

Nope it's fine

[sweet nimbus]

@wooden jolt I didn't look at your feedback yet, but did you have anything to mention about the section with one sentence (local lights)

[wooden jolt]

oh yeah

local lights

[mystic lark]

I've been thinking about local lights, does Unreal really doo VSMs for them too?

[sweet nimbus]

huh

[mystic lark]

sorry

haha

[sweet nimbus]

I think we should expand on the section or something lmao

idk what to say though

[mystic lark]

Feels like a horrible idea to do VSM for local lights

don't you want to mainly cache these and agressively downsample?

[sweet nimbus]

UE5 does VSM for local lights

[mystic lark]

how wtf

[sweet nimbus]

the clipmaps behave differently

[mystic lark]

it's just a mip map no?

[sweet nimbus]

ye

so it's a dynamic resolution system for local lights essentially

but it still uses 99% of the vsm logic

[mystic lark]

but still, you draw the scene additional 6 * 3(active visible mip range) per local point light

[sweet nimbus]

yes but you get sick culling

also you only need to do it for lights within view

[mystic lark]

I guess hiz culling will be hyper good if you store occluders per local light

hmmm

[sweet nimbus]

I don't think you need hiz for local lights, just cull against a cube "frustum"

[wooden jolt]

you really just cull against a cube yeah

[mystic lark]

oh because local lights have a range

[wooden jolt]

and they also cull the VSMs themselves for local lights

[mystic lark]

I'm stupit

[wooden jolt]

they have a feedback pass where they just cull local lights that are not in view or do not contribute to the lighting

(or if adding more local lights would go overbudget)

[sweet nimbus]

hmm I'll just add that local lights are feasible with VSMs, but we "chose" not to implement them for simplicity

[wooden jolt]

ye just mentioning is fine

VSM are a great fit for local lights for the same reasons they're a good fit for directional lights

but can't implement them because, you know 🕐

[vestal snow]

it looks like Computer Modern is the default. I'm not sure if they overwrote the font for the book

Latin Modern is a similar one you can download to use with diagram tools etc.

[mystic lark]

okay here are the measurements

I always included only changes from the original two measurements

so for example changing page size from 128x128 -> 64x64 changed no values when I was not caching and only changed the bookkeeping when I was caching

For sampling the "no search" means that I only attempt to sample the current pixel page and don't try to go one cascade below and above on page miss

[sweet nimbus]

Berfect

[sweet nimbus]

for practical purposes, I think I will just use the mean numbers in this even though they don't paint a perfect picture

I do want to show how vsm/page/sample counts vary the perf though

so I'll probably make three tables

or one big table

or two medium and one small

@wooden jolt can you elaborate on the comment you made on the optimization section? What do you mean by compute

I meant to have asked this yesterday but I forgor

[wooden jolt]

I mean compute-based software raster

it's kind of a key point in VSM to make it quicc

[sweet nimbus]

ah

have we tried it though

it seems like a good optimization in theory at least

[wooden jolt]

it's all theory sadly

but you can logically draw that conclusion pretty easily

bound by PROP -> remove PROP from the equation -> enjoy speedup

[sweet nimbus]

but who's to say that whatever we make will be faster than PROP

[wooden jolt]

by the same argument that primitive culling in compute is faster than fixed function culling

[sweet nimbus]

it's not faster in mine

[wooden jolt]

zeux benchmarked it

lemme find ze article

[sweet nimbus]

that argument is just "hardware slow" bruh

[wooden jolt]

https://zeux.io/2023/04/28/triangle-backface-culling/

[wooden jolt]

yes

[sweet nimbus]

certainly we can beat the hardware in some cases

[wooden jolt]

but ye either way we really shouldn't try to state this as fact

and only just in theory

[sweet nimbus]

game theory

[mystic lark]

oh damn this reminds me

there is no backface culling in Tido yet

uhhhhhhhhhhhhhhhhhh

I'll have to fix that tomorrow

[hazy steppe]

i fixed the thread's initial emojis btw 🙂

[mystic lark]

What are unreal guys cooking

Why is the tile smaller towards the shadow edge, rather than when closer to the camera

[wooden jolt]

forbidden witchcraft

really though wtf

they're probably using the shadow mask to find edges and create frustum dynamically?

it's probably the best solution to projective aliasing isn't it

there's a million different ways I can see this dynamic frustum partitioning failing tho

perhaps some kind of feedback to figure out how "full" the frustum is and the ratio between the expected size and the maximum size for said frustum

and why am I pondering this when it's 3AM

[wooden jolt]

why is the frustum allocation so irregular tho what the hecc

how does that make sense? do they have multiple clip N projection matrices?

that's nuts though

[mystic lark]

Yeah I have no clue, or maybe they can afford a per page projection matrix, uhh I don't understand

Will need more pondering on the background threads, hopefully they come up with something while I eep

[wooden jolt]

I have run out of background threads

can you spare some

I need to invest in more brain wrinkles

[mystic lark]

I'm freeing 30% of shared neuron power

May it serve you well

[sweet nimbus]

U should download unreal yourself

[mystic lark]

Keep the enemy close, I like it

[sweet nimbus]

Or I can take a video for you lol

[mystic lark]

Nono it's fine haha, I will download

They also do some magic with SMRT where they dynamically determine their ray count or smth

More food for the neuron ig

[sweet nimbus]

I guess you can use some heuristic to tell if you're on an edge and to therefore allocate more rays

I think the final fantasy shadow paper discussed something kinda like that maybe

Actually I think that was just doing the shadow test, then seeing if the light needed to be shaded if the occlusion was less than 100%

[vestal snow]

this is giving me intrusive thoughts doubts about our whole strategy

actually I feel better now

[sweet nimbus]

it seems like they are a little smarter about clipmap selection- note how the boundaries are square

compared to this

some of the texels here overlap two valid pages, but choose the less detailed one from the heuristic

it barely matters, but yeah

[prime ice]

magic

[prime ice]

intresting

maybe qadtree lodding

[sweet nimbus]

they probably just have slighter smarter math

or they just check 1 page down

[sweet nimbus]

there is some data missing here

I'm sure you're aware

just the draw and bookkeeping for some stuff on the right

[sweet nimbus]

I didn't want to make it three tables. I'm taking suggestions on how to make it look better

[mystic lark]

For caching, the sampling timings remained the same as the original one

Das why I didn't copy them over again

[sweet nimbus]

ye I understand that

[mystic lark]

Ah this is what you mean

Okay I'll remeasure for these three setups you have

And fill in the values

[sweet nimbus]

btw didn't you add some optimizations that improve these numbers

[mystic lark]

Yeah exactly

Just the draw was affected though

But ye, I'll modify the table once I have the values again

[sweet nimbus]

I think I won't include the "sampling with search" results because we didn't discuss that

hmm

does anyone (not saky because he's a busy boi) want to take some pics of the various vsm debugging features in timberdoodle for the appendix

I also think we should each read the entire article again this weekend as a final pass to make sure things are good

then I will email the editor for our section and ask what's up with the timeline

[sweet nimbus]

it would also be useful to see numbers for 1 sample and 8 samples (no search for both)

ah and I guess the sampling numbers are for 4k, 128x128 pages?

[mystic lark]

Ye but it doesn't affect perf at all (the resolution or the page size)

Hmm but the page size should affect it

Uhh hmm

[mystic lark]

Okay, I'll also add that

[sweet nimbus]

that's quite odd tbh

If it didn't affect perf then I'll mention it

[mystic lark]

I'll remeasure again

[sweet nimbus]

I hope it doesn't take too much time

I tell designated people at work to do perf testing like this and it usually takes them a while to complete it

[prime ice]

saky is an amd eployee confirmed

[mystic lark]

All good, I'll get to it this evening (after I do dynamic objects)

My perf testing is not super rigorous, I mainly use the timestamps + sparse nsight captures to validate

Hope that's enough

[sweet nimbus]

perfectly fine

nsight is overkill imo unless your timestamps don't capture something

we just need to give the reader a rough idea of perf

[mystic lark]

Nsight is more of a sanity check thing

To catch bugs in my perf measure code (if there are some)

[prime ice]

the trick is to exclude barriers in the timestamps so the oerf looks better

[mystic lark]

I probably also should wrap each task with a full barrier to get the true exec time

(only for perf measurements ofc)

[prime ice]

yea

[vestal snow]

If I can get it running on my machine I’ll give it a shot

I think saky mentioned he has new build steps posted unless I made that up

[sweet nimbus]

You might need to copy something into the cmakelists.txt to make it fetch vcpkg

[mystic lark]

oh I forgot about that sorryy

[sweet nimbus]

I remember building it for me was roughly:

1. Copy vcpkg lines into cmakelists.txt
2. Generate a VS solution or open the root directory in VS
3. Build
4. Run the exe from the root directory

[mystic lark]

adding it in a bit

[sweet nimbus]

JS might be able to do it with my crappy instructions if he's familiar with cmake

[mystic lark]

yeye I'll just add the vcpkg thingy into the Cmake

[sweet nimbus]

Well, with the vcpkg stuff

[mystic lark]

automatic download of vcpgk is now in, so all you need should be cmake and ninja

[vestal snow]

Awesome

Thanks guys I’ll give it a try soon

[vestal snow]

I've run into a few issues. I might have to continue tomorrow

it seems like the mirrors vcpkg is trying to use I can't access and I'm not too sure why

[mystic lark]

I'll need an error or something to go off of

[vestal snow]

I wasn’t able to work on this yesterday. I’ll post error info in #1166523139468034068

[sweet nimbus]

I'll stop spamming it with my stuff too

[hazy steppe]

https://youtu.be/DDi62bPbWHw?t=885 you guys trying to build a car

[prime ice]

guys i have an idea

i think i can speed the vsms up a lot

maybe im dellusional

[wooden jolt]

shoot

[prime ice]

i ll make my own early z

[wooden jolt]

transient Z buffer is cope btw

you know what would be very helpful

some sort of function to feed the rasterizer stage that transforms output coords

which basically does this:

physicalMemoryCoord = f(primitiveCoord.xyz)```

and physicalMemoryCoord becomes gl_FragCoord

[mystic lark]

visbuffer for directional VSM I want

64bit unreal style

[wooden jolt]

there was someone who did visbuffer shadows

idk who though

[prime ice]

tear note

but it didnt work out well

tear rt'd the tris from the shadow visbuffer

[sweet nimbus]

Correct hybrid ray traced shadows require a linked list of primitives per texel

[prime ice]

but its holey

[sweet nimbus]

And you need conservative raster

[hazy steppe]

just another rabbithole

[wooden jolt]

what hw could we ask for to make hybrid raytraced shadows faster

(i.e not using a per pixel linked list )

[wooden jolt]

the sad thing is that it's a short rabbithole 😢

figure out per pixel linked lists
rip 🅱️erf
give up because there's actually no way for optimization

[sweet nimbus]

hw accelerated linked lists

[hazy steppe]

https://tenor.com/view/angry-beating-gif-24531604

short like this?

[acoustic bobcat]

Just allocate like 20 framebuffers

Boom, no linked list necessary

[wooden jolt]

VK_FORMAT_R32_UINT_ARRAY_10000

[sweet nimbus]

@mystic lark I can't remember if you got those new benchmarks

[mystic lark]

Ah I forgot

I'll try to get them today (have some urgent uni work I need to do first)

[sweet nimbus]

ah

I could do the benchmarks instead actually

but I need the ball bistro scene

[mystic lark]

https://drive.google.com/file/d/1WQ8BDlKJhCvh6QPxIc6yKwOhflNqqKRm/view?usp=sharing

[sweet nimbus]

downloading

[mystic lark]

btw patrick was testing something and apparently if you run it through nsight you get much more consistent results

also discord and other electron apps cause random spikes in the frame times nothing major, just letting you know

[sweet nimbus]

I just need ruff numbers

[mystic lark]

what I did to measure was

1. hit Stop Gathering and reset timings in Render Statistics
2. hit Use preset camera and press I to override the keyframe in Shader debug menu
3. maunally set 0th keyframe and 0.0 keyframe progress in Shader debug menu
4. hit start gathering and press I at the same time to start the gather and start moving the camera
5. hit stop gathering once the loop is complete (8th keyframe 1.0 keyframe progress)

its a bit cope, but I am really pressed for time unfortunately ☹️

(also make sure to pull latest Tido I made some small changes)

[sweet nimbus]

it's all good. it doesn't need to be perfect 😄

I pulled, dw

[mystic lark]

perfect, thank you for doing it instead of me

[sweet nimbus]

I'm just trying to figure out why it's not loading the model (probably my working directory being fooked)

and now I'm trying to figure out how to use the VS debugger but have it set the right working dir

[mystic lark]

I'm a bit confused, so it loads the model now?

[sweet nimbus]

1. after I pulled, the application would instantly close when running it
2. I'm trying to use the debugger now, which I previously wasn't because it would set the wrong working directory (so the assets would load)
3. I'm trying to set the correct working directory for the debugger so I can use it

[mystic lark]

Try to run it from the command line so that you get error message

Ah I know why

[sweet nimbus]

I did, the error message is nothing

[mystic lark]

I try to load the wrong file I think

[mystic lark]

Uhhh

[sweet nimbus]

I think I know why

it's crashing when dereferencing the ball

because the file didn't load

[mystic lark]

https://github.com/Sunset-Flock/Timberdoodle/blob/main/src%2Fapplication.cpp#L131

This should be uncommented

[sweet nimbus]

ah

yeah I was fiddling with this stuff

I typed it myself and had
std::filesystem::path const DEFAULT_HARDCODED_FILE = "bistro_fix_ball_compressed\\bistro_c.gltf";

[mystic lark]

Ye I told Patrick to wrap it in an if block it but ig he forgot to push

Meant to answer to the message above ups

[sweet nimbus]

yay I can debug it now

[mystic lark]

Good good

[sweet nimbus]

and no more random device lost after I updated drivers and restarted pc

ballin

[vestal snow]

Very nice, that’s a cool demo

[sweet nimbus]

I should give saky a raise

[hazy steppe]

it would be cooler with a certain cube model : > (not really)

[sweet nimbus]

hmm why is there no deccer cubes emoji or sticker

[mystic lark]

I'll see what I can do

[wooden jolt]

that's a good ball right there

very round

[sweet nimbus]

hehe

yo this is sick

[wooden jolt]

it do be

saky is magical

[mystic lark]

I needed that because I had a bug

so I had to have something to reason about what I'm seeing

[sweet nimbus]

📸 yep, that's going into the appendix (it's a good debug view)

[wooden jolt]

we need a debug view ranking btw

I like the projected overdraw the most

(because it gives me nightmares)

[mystic lark]

btw there is a bug in my heuristic somewhere that I've not yet had time to fix

you can see it when you have visualize clip levels on

sometimes (for some reason) the shade opaque pass requests pages that have no been allocated by the mark required pages pass

[sweet nimbus]

that doesn't seem too bad tbh

[mystic lark]

better image

[sweet nimbus]

I did find an odd flickering bug though (might not show up in this vid because my monitor is 240hz)

[mystic lark]

the search during sampling makes it go away

[mystic lark]

oh god what is that

[sweet nimbus]

actually it might be a caching thing

[wooden jolt]

jaker it's 3am

[mystic lark]

how low is the sun?

[sweet nimbus]

80.5

but with my changes

[mystic lark]

ah

uh

yeah that I have no idea about

[sweet nimbus]

ALSO I might've moved the sun a tiny bit by accident which would cause the cache to be wrong

actually nah it looks obvoiusly wrong when I do that

[mystic lark]

oh yeah I also need to add invalidation on sun movement

I need a checklist

[sweet nimbus]

it's hard enough to trigger my bug that it doesn't really matter tbh

[sweet nimbus]

I gotta do this

[mystic lark]

I have a theory for this btw

I think it might be my unprojection from depth to world as I don't account for the near plane there

actually no, that is not it

eh idk, I'll solve it later

[sweet nimbus]

it's no biggie fr

[wooden jolt]

it's time to rest soldier

[sweet nimbus]

I think tido is basically shippable for the demo

well maybe the cache invalidation when sun moves but eh

[mystic lark]

I have some Java profiling that needs to be done

[sweet nimbus]

java?

[mystic lark]

uni garbage

knocking at my door

due till tomorrow midnight

and tomorrow I'm not here half the day

[wooden jolt]

I am so happy I did all the java related exams (DB and UI)

it's in the top 3 worst exams

[sweet nimbus]

I don't think I can PR my function without making a new fork and going through a bunch of ceremony

it would be simpler if you copied this
#1166523139468034068 message

[mystic lark]

okay

I can also ask Patrick (I assume he will have no issue with that) and give you access

[sweet nimbus]

that would work too

this will be my contribution to tido, then I'll keel over and die as my life's work is complete

[wooden jolt]

lol

lmao even

[mystic lark]

'tis but the beginning of a great journey

[wooden jolt]

you won't die so easily

I will drag you all into nanite one way or the other

or maybe GI

hmm so many things to bikeshed

[sweet nimbus]

if you make a nanite library then I'll happily use it

GI is something I want to try though. the weeds of geometry processing do not interest me as much

[wooden jolt]

I'm very interested in surfel GI

but you can't escape tracing rays

I am so sad

[sweet nimbus]

I need to start with shrimpler stuff first

like getting the vk version of frogfood working

after this article

[wooden jolt]

when's the deadline btw

[sweet nimbus]

idk at this point

[wooden jolt]

tomorrow's lectures are cancelled because uni exploded so I'll do some reviewing I guess

also I have to get tido working

[sweet nimbus]

yesterday was when we were supposed to submit the code

three weeks ago is when the draft review period was supposed to start

I'm gonna email our contact later tonight and ask what's up

[mystic lark]

we have not received any review which is kinda sus

[vestal snow]

[sweet nimbus]

I'm guessing the deadlines are a bit more... "flexible" than they initially seemed

[vestal snow]

also fyi: I think my hardware is too old for Tido. Would anyone else be able to take screenshots of the debug stuff?
I have another GPU laying around but it apparently also doesnt support mesh shaders

[sweet nimbus]

Can do 🫡

What's your GPU?

[vestal snow]

in this machine it's a gtx 1060, and in another one it's an rx 5700

[sweet nimbus]

I see

Mesh shaders aren't supported until RDNA 2 (6000 series) and idk about nvidia

[vestal snow]

nvidia it looks like their 16 series did

but 10 series no

[sweet nimbus]

So turing+

[vestal snow]

yeah

apparently apple supports them, but trying to get things running on mac may be cursed

I might still try lol

[mystic lark]

Oh yea, we committed to the mesh shader pipeline when I brought over VSM, it was a nightmare maintaining both compute and mesh paths ☹️ so mesh shaders are now unfortunately required

[ruby trellis]

only on M3 GPUs though

for M1 and M2 its emulated in compute and might not actually be faster than normal vertex shaders

[vestal snow]

That makes sense

I managed to get it to compile on M1. It doesn't run (symbol missing - maybe arch mismatch somewhere), but it compiles

[vestal snow]

hmm I didn't know that

I think M3 is also getting raytracing hardware

[sweet nimbus]

are these good enough for the appendix? I plan on having two more pics of other debug stuff

if no one says anything in the next 30 seconds it will be added

[mystic lark]

Those are some strict deadlines you have

[ruby trellis]

yeah M3 was the big GPU update, it got that dynamic register allocation thingy, hardware mesh shading and hardware RT

the apis for which have existed since M1, but both things are emulated in compute on M1 and M2

no idea about performance numbers on M3 though, but I do have one

[sweet nimbus]

hmm, what if I were to capture a debug view from my engine and disguise it as saky's cooking

delightfully devilish I must say

(I just want a pic of the overdraw view)

[wooden jolt]

smart

[mystic lark]

Overdraw view?

(Just add it to Tido hehehe)

[wooden jolt]

#1128020727380054046 message

this beauty

[sweet nimbus]

comment within the next 5 picoseconds if you don't like this

hmm maybe the colors need to be hotter

[wooden jolt]

remove the missing alpha placeholder and ship it

[mystic lark]

Ohhhhh this is the shadow overdraw

Ohhh I need this

@prime ice add it

[vestal snow]

Then we just need one from lvstri and we’ll have snuck a screenshot from all 4 implementations into the article lol

[mystic lark]

Pleeeasee 🥹 🙏

[wooden jolt]

it's just an imageAtomicAdd to VSM physmem sized image

[sweet nimbus]

yeah this is a screengrab from paint.net lmao

[mystic lark]

If I didn't have homework due today I would immediately add it

[wooden jolt]

ok I add it

[sweet nimbus]

I wonder why there is overdraw on triangle edges

[wooden jolt]

hmm you mean overdraw from helper invocations?

[sweet nimbus]

it can't be helpers because those don't participate in any output

here you can see overdraw even though it's one triangle thick (you can see near plane clipping at the bottom)

[mystic lark]

If you are serious follow my example in vsm_state to create the image, then cull_and_draw_pages task head in vsm.inl will need to be modified and the call itself too and finally the fragment shader of cull and draw pagesv(both opaque and transparent)

It should actually not be that hard to add, daxa makes stuff very easy

You don't need to cope with the special views I do for the memory block

[wooden jolt]

gotcha

[mystic lark]

Regardless we can use Jakers pretty views and I'll try to work on Tido a bit more tomorrow to finish up the remaining bits

[sweet nimbus]

perfect fit

[wooden jolt]

beautiful

[sweet nimbus]

I decided against having an explainer saying why debug visualizations are good and just kept it at two figures

[humble shadow]

Seconded

[sweet nimbus]

the text may need some light editing, but I'll let someone else deal with that

all the article needs is for me to do those damn benchies (aaaaand perhaps one more light editing pass)

given the current pace of things, it seems I am fine to put them off until tomorrow

[humble shadow]

So did you folks get any feedback on the article ?

[mystic lark]

I can do them if they are only required till tomorrow

(the timings)

[sweet nimbus]

nah you've had a disproportionate amount of stuff on your plate lately, let me take some load off

I also think you deserve to have the first author name since you did the most writing and contributed the whole-ass demo hehe

[prime ice]

saky is an absolute machine recently

i added a single comment to the article btw

[prime ice]

ok

ill add debug views as my next task

[hazy steppe]

man this is so cool what you guys made 🙂

[vestal snow]

Yeah good job everyone! It turned out really nice

will we have a place in the article where we can add something like "special thanks to" for anyone who helped read/comment?

[sweet nimbus]

We can have an acknowledgements section

[sweet nimbus]

VSM is mentioned in this presentation near the end. should we reference it?
https://advances.realtimerendering.com/s2015/aaltonenhaar_siggraph2015_combined_final_footer_220dpi.pdf

[wooden jolt]

I think we already do mention this no?

either way yes

[sweet nimbus]

we do not

I'll just add a sentence

fug idk how to cite presentations like this

eh whatever I'll try

you guys can clean up my mess

ok I hate how all the info about VSM is packed into two slides in this presentation

128k VSM

how do they render all pages at once

[wooden jolt]

I think they mean the "theoretical max resolution"

[sweet nimbus]

are they doing sw raster

[wooden jolt]

uhh idk

it's not mentioned anywhere at least

[sweet nimbus]

idk how the vsm could be so big and they draw all the pages at once

unless they only draw a small portion of it at a time

idk how to even mention this now

[wooden jolt]

someone with a shitter account could probably send a msg to seb?

[sweet nimbus]

I don't

[wooden jolt]

it's way too ambiguous ye

[mystic lark]

The only issue with this is how do you rasterize it

At 128k resolution lmao

[sweet nimbus]

sw raster is one way

the other way is that you are only seeing a tiny part of the shadow map at a time

so you just draw a box around it

[mystic lark]

I guess you could construct per page projection also

But that seems nightmarishly slow

[wooden jolt]

you'd need one drawlist per page

[mystic lark]

Yeah

[wooden jolt]

which is bonkers

[sweet nimbus]

most importantly, how do we relate this to our article?

"virtual shadow mapping has been explored before, ...?"

[mystic lark]

Yeah exactly

[sweet nimbus]

that is unsatisfactory imo

what were the lessons of the previous exploration of vsm

(I mean obviously we didn't use them as we just learned of this thing but yeah)

[mystic lark]

Well the issue is we have no idea what they do in the presentation

I'd tie it with the unreal citation

But eh idk, the presentation information is extremely unsatisfactory

[sweet nimbus]

just enough to say that they did it first

I think it may also be worth mentioning the other virtual texturing stuff that has been explored somewhere around that part too

I have to eep now but I'll think about it

[mystic lark]

Is it okay to just dm Sebastian?

[sweet nimbus]

Of course

But I nominate someone else to do it

[wooden jolt]

does he have any other contacts that are not social media

maybe email?

[sweet nimbus]

Does he have a website

You can get emails from git commits
https://github.com/sebbbi

[mystic lark]

Welp

[wooden jolt]

twittershitter™️

[sweet nimbus]

There was a thing you could do, like putting .patch at the end of a commit url on GitHub to see the raw commit info

[mystic lark]

Their page sizes increase