huh the fuck
#Iris - A Journey through OpenGL and beyond to learn Graphics
buoyant summit
that doesn't work like that in all companies 💀
wicked notch
for me per primitive atteibutes work fine
they decorate everything with perprimitiveext BUT the builtin variables
buoyant summit
if slang was intel and intel had a driver bug they'd probably add a workaround to the compiler
wicked notch
so gl_PrimitiveID and gl_CullPrimitiveEXT are not marked
glass sphinx
they decorate everything with perprimitiveext BUT the builtin variables
ahaaa
what is primitive id
wicked notch
gl_MeshPrimitivesEXT[id].gl_PrimitiveID
I use this to output primitive id to visbuffer
wicked notch
it's probably the same thing, one is per vertex and one is per primitive
idk
gl_CullPrimitiveEXT tho 
glass sphinx
gl_CullPrimitiveEXT tho <:bleakekw:1082598350303539240>
yea that is a problem
glass sphinx
it's probably the same thing, one is per vertex and one is per primitive
from reading opengl spec it seems you are not supoosed to erite that
its just an input for some shader stages
wicked notch
from reading opengl spec it seems you are not supoosed to erite that
perprimitive?
wicked notch
oh you mean gl_PrimitiveID
glass sphinx
yee
wicked notch
I mean
perprimitiveEXT out gl_MeshPerPrimitiveEXT {
int gl_PrimitiveID;
int gl_Layer;
int gl_ViewportIndex;
bool gl_CullPrimitiveEXT;
int gl_PrimitiveShadingRateEXT;
} gl_MeshPrimitivesEXT[];``` vk spec says this is writeonlytbf I also don't really get the point of mesh shader gl_PrimitiveID beside some weird compat with frag shaders that were consuming/expecting gl_PrimitiveID
glass sphinx
yea
buoyant summit
btw I found a use case for secondary command buffers
you can defer decisions like loadOp, storeOp and image layouts until later point in time
glass sphinx
but cant you do that with normal cmd buffers too
buoyant summit
tbh my abstraction is just kinda weird
but secondary cmdbufs turned out to be rather handy for it
delicate rain
I never understood the secondary command buffers
I thought they were what work graphs are meant to be
buoyant summit
do you mean indirect commands
secondary command buffers are like normal command buffers except they need less information before you can record a draw
delicate rain
Ah okay I was heavily misinformed then
delicate rain
I thought secondary command buffers could be dispatched from primary ones
delicate rain
Hmmmm
wicked notch
command list better
the microsoft employee side of you is leaking
buoyant summit
me on my way to rename descriptor_buffer to descriptor_list
glass sphinx
ok good
buoyant summit
tbh one thing I'd like if memory management for cmdbufs was more explicit
CommandPool feels ugh
buoyant summit
can I just please provide a callback that driver will call every now and then 
buoyant summit
bikeshed tho
real
btw
VK_EXT_nested_command_buffer is a thing apparently
so you can recurse with command buffers very deeply
idk who asked for that, but
glass sphinx
lool
wispy spear
John Daxa probably
buoyant summit
Jane Daxa
wispy spear
my bad i meant Jérôme Daxa
wicked notch
https://github.com/KhronosGroup/SPIRV-Tools/issues/4919 turns out, there was already an issue
it's been sitting there for 2 years
thank you John Khronos
loud crag
in the default implementation yeah but I have a custom one which dynamically creates a set with all required images
hmm no that works
all my textures are SRGB generally
frank sail
did you see this already lvstri
https://themaister.net/blog/2024/01/17/modernizing-granites-mesh-rendering/
pale horizon
all my textures are SRGB generally
Well, you have a custom backend, so you probably do some sRGB trickery
But also try drawing some ColorEdit and see if displayed color matches there
Also you can open ImGui's style editor in DemoWindow and color pick the colors to see if they match
Simple sRGB-cope implementations usually have subtly wrong colors
(especially FrameBg, FrameBgHovered etc.)
loud crag
Well, you have a custom backend, so you probably do some sRGB trickery
But also ...
I just convert the vertex colors to linear
primal shadow
@wicked notch have you started software raster? Do you know why nanite forces HW raster for triangles that need to be clipped to the screen?
wicked notch
clipping is faster in HW
primal shadow
Isint it just a min/max of each screen space vertex to clamp to the screen bounds?
glass sphinx
Isint it just a min/max of each screen space vertex to clamp to the screen bound...
no
wheat haven
Isint it just a min/max of each screen space vertex to clamp to the screen bound...
if it was just a min/max clamp you'd change the shape of the triangle
delicate rain
Isint it just a min/max of each screen space vertex to clamp to the screen bound...
I don't even think this gives you clipped triangle
yeah exactly
primal shadow
Ah I see
Oh wait I misunderstood what this tutorial I'm reading is doing
They were clipping the aabb for the triangle, not the triangle vertices
delicate rain
https://cent.felk.cvut.cz/courses/PGR2/lectures/10-architekturyII.pdf
Look this if you're interested
Starting Slide 46
primal shadow
Thanks!
wispy spear
https://cent.felk.cvut.cz/courses/PGR2/lectures/10-architekturyII.pdf
Look this ...
why are paper filenames always ass 😦
delicate rain
it is a lecture slides thingy
delicate rain
hmm I don't see the issue here. It is a lecture 10 - which is about GPU architectures (architektury in Czech) part II
frank sail
why are paper filenames always ass 😦
I thought you liked funny spellings
primal shadow
Does anyone understand what nanite talks about when it comes to solving for the x interval for software raster?
I suppose I should read the PDF they link as reference [71] 😛
primal shadow
Nope I read through cure's GitHub project and can't actually find what they're doing
primal shadow
@wicked notch how do you decide between SW/HW raster per cluster? Nanite presentation says all 3 triangle edges < 32 pixels long. But uhh, how do you determine that at the cluster level? Compute some kind of average triangle edge length per cluster?
primal shadow
And then what, SW rasterize anything with a small viewport area?
wicked notch
yes
primal shadow
Makes sense, thanks
wicked notch
And then what, SW rasterize anything with a small viewport area?
if the aabb is less than 32 pixels in each direction
primal shadow
I'm almost ready to do software raster, I just need some final wgpu changes to do image atomics. Going to use buffers in the meantime.
I also need to figure out how to rework my culling code to feed into the SW/HW raster. Probably switch to two atomic lists of cluster IDs.
And then indirect dispatched on the SW raster / hw index buffer write.
wicked notch
glass sphinx
yes
wispy spear
thats how we will get LVSTRINV soon
frank sail
LPSTRI
wispy spear
luigi? i think you should try to get into novideo
after school
just like pixel went to valve
having a frog at important spots in the industry helps to fight John Khronos
wicked notch
you're asking me to go to the dark side
wispy spear
explicitly
buoyant summit
why is nv dark side
wicked notch
true that
frank sail
my megacorp is better than your megacorp
delicate rain
@wicked notch I move here to not pollute Vulkan
I don't get it
do you do one dispatch per task shader workgroup?
or per subgroup?
(my actual knowledge of mesh shaders is limited)
wicked notch
workgroup is 128
subgroup is 32
when I call vkCmdDrawMeshTasksEXT(1, 1, 1) I dispatch one task shader to cull 128 meshlets
if all of those survive I will have 4 mesh shader commands (EmitMeshTasksEXT()) each with a workgroup dispatch size of 32
what I'm trying to do is reduce the number of EmitMeshTasksEXT calls so that NV command processor doesn't commit sudoku
delicate rain
okay so you still cull 32 meshlets per task subgroup?
uh
I'll just tell you what we do in Tido
delicate rain
each task shader invoc culls 32 meshletts, we get base offset, subgroup offset (I think) and survivor bitmask
then we dispatch mesh shader (32 threads) for each surviving meshlet
loud crag
yeah survivor bitmask sounds like a good idea
wicked notch
what's the survivor bitmask for
delicate rain
to know what meshlet mesh shader works on
wicked notch
you already know that don't you
loud crag
replaces your deltaIDs from your payload
delicate rain
nth mesh dispatch looks for nth set bit in the bitmask
wicked notch
so you do an exclusive workgroup bitcount?
if there even is an intrinsic for that
maybe it's only subgroup bitcount
delicate rain
we cooked with Patrick
wait
func wave32_find_nth_set_bit(uint mask, uint bit) -> uint
{
// Each thread tests a bit in the mask.
// The nth bit is the nth thread.
let wave_lane_bit_mask = 1u << WaveGetLaneIndex();
let is_nth_bit_set = ((mask & wave_lane_bit_mask) != 0) ? 1u : 0u;
let set_bits_prefix_sum = WavePrefixSum(is_nth_bit_set) + is_nth_bit_set;
let does_nth_bit_match_group = set_bits_prefix_sum == (bit + 1);
uint ret;
uint4 mask = WaveActiveBallot(does_nth_bit_match_group);
uint first_set_bit = WaveActiveMin((mask.x & wave_lane_bit_mask) != 0 ? WaveGetLaneIndex() : 100);
return first_set_bit;
}
returns the position of nth set bit in a bitmask, uniform for entire subgroup
wicked notch
why would you do this instead of just using subgroupBallotExclusiveBitCount
loud crag
I think it's not in slang
no spirv intrinsics?
delicate rain
why would you do this instead of just using subgroupBallotExclusiveBitCount <:bl...
but even if you did exclusive bit count
how do you know what thread to broadcast from
like the thread itself knows sure
but how do the others know
wicked notch
interesting question
intuitively, each subgroup invocation calculates its own exclusive bit count
so instead of a ballot exclusive bit count, you do a regular bit count on the mask
delicate rain
well you want a subgroup per meshlet
so all 32 threads should agree on the meshlet
no?
or do you have a different mapping
delicate rain
the exclusive/inclusive bit count would just replace the WavePrefixSum and two lines above
wicked notch
so it's a workgroup bit count
delicate rain
the ballot and min you still need
glass sphinx
workgroup is 128
subgroup is 32
when I call `vkCmdDrawMeshTasksEXT(1, 1, 1)` I d...
why do you use such large workgroups
nv recommends 32 still
i also always lost perf on larger groups
the survivor list reduced the lsb pressure quite a bit that was very nice
larger groups are probably better when you have larger payload
i see now why you did it
wicked notch
the idea was to reduce the number of invocations needed to cull and draw
buoyant summit
i also always lost perf on larger groups
mesh workgroups are fake on nv, it's just a single hw subgroup cycling through api subgroups, compiler munches your shader
wicked notch
mesh workgroups are fake on nv, it's just a single hw subgroup cycling through a...
wait wat
can I at least expect gl_SubgroupID to work as I would expect it to work?
when workgroup size is multiple of subgroup size
glass sphinx
mesh workgroups are fake on nv, it's just a single hw subgroup cycling through a...
what bird dir you switcher this forbidden knowledge
buoyant summit
can I at least expect gl_SubgroupID to work as I would expect it to work?
yes, that's just an impl detail
buoyant summit
what bird dir you switcher this forbidden knowledge
nvk people
glass sphinx
nice
loud crag
what bird dir you switcher this forbidden knowledge
well its kinda obvious given the NV extension is always one dimensional
glass sphinx
https://github.com/shader-slang/slang/pull/3907
they really are sweet hearts
they just implement it
i love them
glass sphinx
slang your beloved (NOT MINE THOUGH)
cope cope cope
^ you when a slang release drops (at least this one)
wicked notch
@glass sphinx do you have docs on spirv intrinsics for slang
buoyant summit
trying to review a slang patch and it makes my brain boil slightly
glass sphinx
<@238694102437330944> do you have docs on spirv intrinsics for slang
afaik there are only tests
frank sail
I wonder how nano copes with email
buoyant summit
I wonder how nano copes with email
I just re-read my thingy many times before sending
frank sail
Same
buoyant summit
but also by not sending any emails often
frank sail
I hate email
wispy spear
yeah it sucks
buoyant summit
agree but I like receiving notifications on my issues/MRs/whatever in a centralized box
very handy
frank sail
Sure but I mean for direct communication
buoyant summit
yeah for direct communication it kinda sucks
frank sail
Especially when mfs splinter threads and then the inbox becomes impossible to navigate (might be a client/skill issue idk)
buoyant summit
yes, skill issue
wispy spear
plus all the signature junk polluting every thread
buoyant summit
Especially when mfs splinter threads and then the inbox becomes impossible to na...
frank sail
what client is this
buoyant summit
thunderbird
buoyant summit
never used outlook
frank sail
I use it for work unfortunately
I did use Thunderbird once. Perhaps I should use it again
buoyant summit
considering outlook is supposed to be backed by a big chungus corporation it probably has comparable functionality
so mayhaps just do a bit of googling
buoyant summit
do you use folders
or do you just look at your inbox
create folder + setup a filter by something
delicate rain
Damn email powerusers
frank sail
It's necessary when you get dozens of emails every day
buoyant summit
delicate rain
I contribute to one repo and they have the GitHub actions setup to try and build after each commit for like 15 different setups. So after each commit you receive 15 emails of the build succeeding or failing
frank sail
hmm I only get emails when CI fails in my thing (Fwog)
delicate rain
Or maybe it's that, but their gh actions are broken so everything fails 🥸
I don't really read them, because it was just garbage every time I did
900k+ lines of compiler errors
finite quartz
I hate email
Yeah, takes me hours to write anything, I hate it too
glass sphinx
I wonder how nano copes with email
i always send cryptic shit
if end to end encription gets outruled at some point im safe
i trained all the ppl i know to decypher my shit
noone else will understand
primal shadow
My depth down sampling is missing some pixels on npot textures
And therefore the way I project bounding spheres to screen space is wrong. Ahhh.
delicate rain
https://github.com/LVSTRI/Retina/blob/508dfb0ba11e6ffaba7633c07d2f5613e06472b9/src/Retina/Sandbox/Shaders/GBufferResolve.frag.glsl#L145 @wicked notch mister what is this cope?
wicked notch
it's what one does to avoid pagefaulting
though tbh I don't think adding a page request is too bad
delicate rain
hmm with PCF this is easy, you just request the PCF region to be in the same clip, but I have no clue about how to do it with pcss
wicked notch
same thing except it's max(blockerSearchRadius, maxFilterRadius)
delicate rain
right, but those can be massive no?
delicate rain
I don't follow
wicked notch
when filtering, if a page is not allocated you just emplace an allocation request in the list
delicate rain
to be drawn next frame?
wicked notch
ye
faint crane
wicked notch
@primal shadow you're building your own version of SPD right?
can you pin me the discussions you had with devsh, I can't find them
primal shadow
<@145540119141679105> you're building your own version of SPD right?
No I'm not working on anything ATM. I'm on vacation and don't have my desktop, and my laptop is broken.
wicked notch
rip
primal shadow
can you pin me the discussions you had with devsh, I can't find them
#graphics-techniques message
Depth down sampling for the initial pass is hard though. I messed up and need to fix it.
wicked notch
halving time for hzb copy + reduce, 'ery nice
150us -> 80us
now I can spam this for all 16 clipmaps
glass sphinx
that will be 1.3 ms
wicked notch
yes
wicked notch
because amd cannot comprehend the existence of operating systems other than windows
this is kinda sorta single pass
except it's 3 pass 
glass sphinx
oooff
wait but you can just copy the tido single pass downsampler
it just doesnt properly work on the bottom and right edge if you have a non pod res
glass sphinx
yea
wicked notch
nice
wicked notch
like 2048x1024
how long does it take to downshrimple this image?
assume it's just single channel float
glass sphinx
i was thinking of rescaling the lowest mip to some pot size
glass sphinx
how long does it take to downshrimple this image?
~20 microseconds
on 4080
fully memory bound
wicked notch
hmm
wicked notch
for me it takes 30 microseconds to copy time image and 30 microseconds to fully downsample it
wicked notch
does that include the initial reduction
glass sphinx
no
wicked notch
ok so your thing is 2.9x faster
glass sphinx
arguably i should remove the writeout for the lowest 2-3 mips not just mip0
culling on that fine scale is unnecessary
also slower even potebtially
i need to test it again
delicate rain
ok so your thing is 2.9x faster
on a nuclear GPU
wicked notch
btw saky
hear me out
when drawing the vsms
I bind temporary virtual resolution depth texture to get the sweet early Z
primal shadow
@glass sphinx do you have code to share? I realized my down sampling does not work for npot textures :(.
wicked notch
now let's say I build HZB out of this temporary depth texture
in the initialization step, I query for the active page table, and write 0 if page is inactive, or the depth if it is active
wicked notch
then I reduce the thing and cull
culling outputs a bitmask of visible meshlets
this repeats for each clipmap
downside is subchannel switch
delicate rain
uhh
primal shadow
<https://github.com/Sunset-Flock/Timberdoodle/blob/main/src/rendering/rasterize_...
Ty, I will steal and port to wgsl 🙂
delicate rain
then I reduce the thing and cull
so you cull your lower clips with the deph written out by the higher clips?
I don't fully follow
also it makes it so that you have to barrier between each clip draw
16 barriers each frame
yucky
wicked notch
no I don't need barriers
or yes I do
but only for the depth texture
transitioning from undefined because I discard the results
primal shadow
culling outputs a bitmask of visible meshlets
I switched to 2 bits per meshlet btw. 0 = not eligible (wrong lod, instance culled, failed frustum cull), 1 = first pass draw (passed last frame occlusion cull), 2 = second pass candidate (failed last frame occlusion cull), 3 = second pass draw (passed current frame occlusion cull)
delicate rain
transitioning from undefined because I discard the results
well you need a full clear no?
wicked notch
yes
delicate rain
yeah
wicked notch
but it's accelerated so it doesn't count
wicked notch
they don't but I'm hoping for hiz culling to make up for the difference
primal shadow
Also I asked this question earlier and didn't get a response, do we really need to do an explicit frustum cull for each meshlet? We have to occlusion cull anyways, which involves projecting the culling sphere to a screen space aabb. We can just check if it's valid, no?
glass sphinx
<@238694102437330944> do you have code to share? I realized my down sampling doe...
mine also doesnt. I still need to add a variant that respects borders
delicate rain
they don't but I'm hoping for hiz culling to make up for the difference
I don't understand how you'd use the temp z for hiz cull
like you draw the biggest clip, and cull the lower one against hiz from the biggest clip?
ie draw clip 8 - build hiz of 8 - cull clip 7 agains 8 - draw clip 7 - build hiz 7 ....
primal shadow
mine also doesnt. I still need to add a variant that respects borders
There's https://miketuritzin.com/post/hierarchical-depth-buffers, but idk how optimized it is.
wicked notch
for clipmap in clipmaps {
Clear(tempDepthBuffer);
Draw(tempDepthBuffer, previousFrameVisibilityMask[clipmap.Index]);
Copy(tempDepthBuffer, clipmap.HZB); // also performs an initial reduction, checks whether the current depth tile
// is active in the page table, if it isn't active the depth written is going to be 0
Reduce(clipmap.HZB);
Cull(clipmap.HZB, currentFrameVisibilityMask[clipmap.Index]);
Barrier();
}```right, but I don't see how this is different than doing it without temp hiz?
you can just as well dras previousFrameVisibilityMask into the VSM itself
reduce the VSM
and cull against that
you don't really need the temp z for this no?
or am I missing somehting again
wicked notch
now that I think about it you're right I don't
I can just make a dispatch over the full virtual resolution and each thread checks for the virtual page table, gets physical texel and write that (or 0 if page isn't active)
this way draws overlap too
delicate rain
and we can do that inplace in the physical memory itself, (we just make the physical memory have mips, that way you can reduce over the physical memory)
wicked notch
ye but I don't understand how to cull against the actual VSM with mips
delicate rain
I mean there will be some index math but I don't see how it would not be doable
conceptually it is all the same no?
delicate rain
It will be a bit more involved yeah
wicked notch
thinking about it, a quad could span 4 different physical pages
delicate rain
yes
wicked notch
(one per corner)
delicate rain
so you need to go from NDC -> VSM_PAGE -> PHYSICAL_STORAGE
or, over some mip cutoff just NDC -> VSM_HIZ
it's just an indirection
(I'm saying it super nonchalantely, as like I won't run into 50 bugs trying to implement this)
frank sail
btw I found the button to enable this view in outlook (show as conversations) 
distant lodge
look what came in the mail today https://jglrxavpok.github.io/2024/04/02/recreating-nanite-runtime-lod-selection.html
distant lodge
please make a blog, I don't wanna read twitter clones
buoyant summit
posting short messages can be easier for the author to write
hallow cedar
jokes on you i procrastinate doing either
buoyant summit
so it's a choice between lustri not writing anything (and you thus not reading) vs reading something that's perhaps not very quality
distant lodge
one weird trick
distant lodge
true, I should probably give in and make a memestodon account so I can continue passively trawling for GP info
buoyant summit
true, I should probably give in and make a memestodon account so I can continue ...
just pick an instance that doesn't tolerate and/or enable morons
or other instances will block the instance you've picked which might reduce the quality of your experience
buoyant summit
@dull oyster 
fully but why do you do this:
vk::DeviceAddress vertexBufferAddress = (vk::DeviceAddress)-1;
vk::DeviceAddress indexBufferAddress = (vk::DeviceAddress)-1;```why not just 0
0 is null & invalid
frank sail
or at least a <https://mastodon.gamedev.place/> account
not Twitter
primal shadow
I use my own blog on GitHub pages. Took like an hour to setup.
dull oyster
<@163242187084005377> <:didnotread:1198383955037134979>fully but why do you do t...
Honestly? Don't really know why I did that
wicked notch
I really should sleep because my eyes are failing me but the vsm HiZ expiment is almost done
wispy spear
allez dodo! 🦤
wicked notch
const uint cornerX = index & log2(position);
const uint cornerY = index >> log2(position);```mmm yes
frank sail
are you doing two-pass hiz
wicked notch
just single pass rn
frank sail
because I think you either have to do that or make pages active for two frames so geometry can render properly
wicked notch
that does make sense
frank sail
I tried one-pass hi-z and didn't think of that issue, then it was broken and I deleted it all
wicked notch
ye I'm willing to accept some artifacts for now
I'll do two pass after this works properly and I get the perf boost I'm expecting
wicked notch
buckle up because I'm about to create a 2048x2048x16x4 image
delicate rain
Does it at least go brr when drawing?
wicked notch
I will test soon
meanwhile
float SampleVirtualShadow(in uvec2 position) {
const uint power = findMSB(VIRTUAL_SHADOW_PAGE_SIZE);
const uvec2 virtualPagePosition = position >> power;
const uint virtualPage = imageLoad(g_VirtualShadowVirtualPageTableImage, ivec3(virtualPagePosition, gl_WorkGroupID.z)).x;
if (!VirtualShadowIsPageBacked(virtualPage)) {
return 0;
}
const ivec2 physicalTexelCorner = VirtualShadowCalculatePhysicalTexelCorner(virtualPage);
const ivec2 physicalTexel = physicalTexelCorner + ivec2(position & uvec2(VIRTUAL_SHADOW_PHYSICAL_PAGE_SIZE - 1));
return uintBitsToFloat(imageLoad(g_VirtualShadowPhysicalMemoryImage, physicalTexel).x);
}
RetinaGroupSize(16, 16, 1)
void main() {
const uvec2 position = gl_GlobalInvocationID.xy;
const uvec2 virtualPosition = position << 1;
const vec4 samples = vec4(
SampleVirtualShadow(virtualPosition + uvec2(0, 0)),
SampleVirtualShadow(virtualPosition + uvec2(0, 1)),
SampleVirtualShadow(virtualPosition + uvec2(1, 0)),
SampleVirtualShadow(virtualPosition + uvec2(1, 1))
);
const float v = min(min(samples.x, samples.y), min(samples.z, samples.w));
imageStore(g_OutputImage, ivec3(position, gl_WorkGroupID.z), vec4(v));
}
``` can I make this better in any wayI stall on TEXTHR though so I doubt that
delicate rain
On phone
Will check when I'm back on pc
2ms shadowmap draw seems fine no?
(from the capture )
wicked notch
that's without hzb, I will make culling go brr now
delicate rain
This is bistro 16 clips?
wicked notch
nono sponza
because it loads faster so I can test without waiting 
mfw no async load
delicate rain
Would be pog if it was bistro
delicate rain
mfw no async load
Ye it's pain to do
But fun when it works
Launch times are instant
Hmm, how much faster is Sponza compared to bistro?
wicked notch
a lot
but it's not comparable to regular raster either way
I think I can make the copy go faster if I abuse the fact that page size is 128
I just allocate a ton of shared memory and put the samples there
delicate rain
Okok, was just wonderin
wicked notch
I reduce memory traffic
memory subsystem is happy
I am happy
copy takes literally 3x less time
hell yeah
delicate rain
At the mere cost of 380 megabytes of VRAM we get hiz
wicked notch
less than what bistro would take in an acceleration structure
plus this is the dumb shitty naive method
if we reduce the VSM itself we get infinite power
delicate rain
The if is the important part
wicked notch
I believe in you
wicked notch
ight moment of truth
delicate rain
Speed
wicked notch
I'm rendering at the blistering pace of VK_ERROR_DEVICE_LOST millseconds per frame
delicate rain
I'm curious btw, are validation layers useful to you?
wicked notch
somewhat
delicate rain
Recently I can't really do anything other than shader print with them
delicate rain
Yeah, GPU based just crashes or misdiagnoses so the app doesn't run properly
Sync doesn't get bindless
And everything else is done by daxa
wicked notch
holy mother of all pog
it is working
ladies and gentlemen
800 microseconds raster on bistro
wicked notch
it takes longer to build the HZB
delicate rain
If we combine it with hpb cull and caching it will be light speed
wicked notch
this is already hpb
delicate rain
Do you know if it actually works?
wicked notch
I write far depth into non backed pages
delicate rain
Like do you have Shadows?
delicate rain
this is already hpb
Aha I see
delicate rain
Crossing my fingers
wicked notch
nop
delicate rain
Hmm weird
delicate rain
Btw is it two pass?
wicked notch
one pass
wicked notch
do you query memory budget every frame for those statistics
ye
delicate rain
So you just keep depth from previous frame and cull against that or?
wicked notch
no I use previous frame visibility mask
I do that because it's easier to switch to two pass this way
wicked notch
previous frame visibility mask to cull current frame meshlets
it's a bogus approach
but it makes it easier to do two pass
Like I only have to do another culling pass and xor the visibility masks together
delicate rain
Hmm okay what is visibility mask? I thought it's just a bit mask of visible/notvisible for each cascade?
wicked notch
something to mark meshlets as visible/not visible
each bit encodes visibility of a single meshlet
wicked notch
bool IsMeshletVisible(uint meshletInstanceIndex) {
const uint maskIndex = meshletInstanceIndex >> 6u;
const uint bitIndex = meshletInstanceIndex & 0x3fu;
const uint64_t mask = RetinaDereference(g_VisibleMeshletBuffer)[maskIndex];
return (mask & (uint64_t(1) << bitIndex)) != 0;
}```this is in the task shaderino
wispy spear
3x fu 😛
delicate rain
Tido has super sampling
delicate rain
```glsl
bool IsMeshletVisible(uint meshletInstanceIndex) {
const uint maskInde...
How do you set this mask?
wicked notch
when I cull
void main() {
const uint meshletInstanceIndex = gl_GlobalInvocationID.x;
if (meshletInstanceIndex >= u_MeshletCount) {
return;
}
...
const uint meshletVisibilityMaskIndex = meshletInstanceIndex >> 6u;
const uint meshletVisibilityMaskBit = meshletInstanceIndex & 0x3fu;
if (IsMeshletVisible(aabb, viewInfo, transform)) {
atomicOr(RetinaDereference(g_VisibleMeshletBuffer)[meshletVisibilityMaskIndex], uint64_t(1) << meshletVisibilityMaskBit);
} else {
atomicAnd(RetinaDereference(g_VisibleMeshletBuffer)[meshletVisibilityMaskIndex], ~(uint64_t(1) << meshletVisibilityMaskBit));
}
}
let me take a moment to curse at shaderc
Failed to compile shader 'D:/Dev/CLion/Retina/src/Retina/Sandbox
/Shaders/GBufferResolve.frag.glsl': shaderc: internal error: compilation succeeded but failed to optimize: ID '128[%g_Vi
rtualShadowInfoBuffer]' defined in block '9[%9]' does not dominate its use in block '258[%258]'
%258 = OpLabel``` fuck youi understand 0, but its utterly fascinating, and im not trying to be sarcastic or anything. its
maybe we lure the slang peeps onto our server and chain them somewhere in the basement, so that they can immediately fix those things
delicate rain
```glsl
void main() {
const uint meshletInstanceIndex = gl_GlobalInvocationID....
So if a meshlet stops being visible, how can it become visible again?
wicked notch
shadows do work with culling
and they're blazing fast to raster
VSM has been conquered
wicked notch
So if a meshlet stops being visible, how can it become visible again?
the mask is cleared every frame
if meshlet passes both frustum and hiz test it is set as visible
delicate rain
and they're blazing fast to raster
POG
okay so you draw meshlets visible last frame, build hiz, draw everything again and cull against hiz?
wicked notch
correct, minus the "draw everything again"
delicate rain
draw meshelts visible last frame, build hiz and...??
wicked notch
cull for next frame
delicate rain
when do you draw what was not visible last frame
wicked notch
don't try too hard to understand this method, it's garbage
wicked notch
when do you draw what was not visible last frame
in theory after you build the hiz yes
I'm just not doing that rn for shrimplicity
loud crag
shadows do work with culling
damn, same scene for me but just using four csm cascades and I have 40fps instead of >200 :(
delicate rain
in theory after you build the hiz yes
but that means that you can miss stuff that was not visible last frame and just became visible this frame no?
wicked notch
ye that's when you xor the masks
delicate rain
for a frame I guess, because it is visible this frame so you draw it next frame
wicked notch
two pass occlusion culling works like this
Draw(visibleLastFrame);
BuildHZB();
CullAndXor();
Draw(disoccludedCurrentFrame);```I do this right now
Draw(visibleLastFrame);
BuildHZB();
Cull(); // for next frame```right so you do have one frame of delay
delicate rain
yeah this is awesome
wicked notch
1ms
delicate rain
what gpu btw?
wicked notch
3070
wicked notch
now potrick must figure out a way to make hzb build go brrr
because currently that takes more than VSM raster
delicate rain
I keep running out of resolution on the 0th clip on a 2k monitor
wicked notch
with my heuristic?
delicate rain
I need to turn my 0th clip world size to 8 meters to work
glass sphinx
eeeeviiiiill
delicate rain
I wonder if better culling makes it faster to not cache at all
wispy spear
<:froge_yeehaw:1180194428607008908>
time to get back into the saddle
glass sphinx
@wicked notch we can prob just not erite the lowest mips for hiz
maybe even only on page level
if we do it at page level it will be like 10 mics for all clips
idk how fucked the culling will be with that tho
wicked notch
ye I think the real strat is just doing reduction on the VSM itself
per page, that is
wicked notch
I think it can work, because it's effectively the same thing, just the sampling gets more convoluted
wispy spear
astral field
maybe we lure the slang peeps onto our server and chain them somewhere in the ba...
solution: make an issue
issues fix everything
wispy spear
solution: make an issue
is that you maisonbleu?
astral field
is that you maisonbleu?
I'm not a blue house?
wispy spear
: )
astral field
oh, it wasn't french
wispy spear
i dont remember your nickname but you fly the team-effort-d3d11 role 🙂
or you are the 8bit guy
astral field
I'm 8 bit, yea
wispy spear
ah
astral field
OH you changed your name
wispy spear
you changed your whole account
wispy spear
nibble is 4 bits
astral field
half the intelligence
wispy spear
fair
astral field
i had no idea you changed user, thought you disappeared
wispy spear
ill never disappear
wispy spear
im not a gpu dev though
astral field
wicked notch
I ponder
frank sail
wicked notch
if the first clipmap is smol
frametime goes up a lot in that goddamned spot
near the bushes that is
it's fine literally everywhere else
I need to inspecc more
renderdoc really hates my 2k^2 x16 layers x12 mips image though
so here is a partially reduced version of the first clipmap in that area
wispy spear
could that be overdraw in the bushes killing the perf?
wicked notch
it should be mitigated by hiz
wispy spear
maybe debug draw the bushes like potti did few months ago
frozen frustum and fly around the bush
to see if it do hiz or hiznt
wicked notch
oh you sparked an idea in me
page heatmap
brb
renderdoc is sometimes bogus though smh
frank sail
page heatmap
ahh cool idea
delicate rain
Why do we come up with the coolest of ideas when the articles dealing stares us down???
wispy spear
some people work better when there is pressure
delicate rain
I like to blame my weak discipline
wispy spear
out comes ze diamonds : )
astral field
found the problem at least
frank sail
tbf that's where most of the detail actually is
wispy spear
is it possible that the geometry there is just simply fucked (read not ideal) and they really couldnt just be bothered
mayhaps needs some work in blender
wicked notch
nah I see now what's going on
wispy spear
ah
wicked notch
and yes it is the geometry being horrid
but it's something worse
big meshlets
foliage is a completely disconnected mesh and meshoptimizer already doesn't care about spatial locality
so it just grabs whatever triangles it has available to build the meshlets, this makes the AABB huge
hence hiz can't cope
wispy spear
that means you need a carefully crafted scene to properly generate meshlets for all the newfangledisms
wicked notch
..or
wispy spear
fix meshoptimizer
frank sail
you were never obligated to help with the writing
wicked notch
I must uphold that duty
frank sail
I invite you to read what we have so far on overleaf
wicked notch
I shall
wicked notch
gud
delicate rain
we use tido yes?
I wrote what we do in tido
wispy spear
looks like LVSTRI blogs bout his college setup
its an italian keyboard layout but he tinkered with some keys
wispy spear
lustri i also just noticed your cmake installing deps thingy, thats neat 🙂
also also, you dont need to tell cmake about your headers, cpp is enough, when declaring a target add_executable/library
wicked notch
also also, you dont need to tell cmake about your headers, cpp is enough, when d...
ik, do I do that schtuff somewhere
I think I do target_sources everywhere, maybe there's some stray header
wispy spear
ik, do I do that schtuff somewhere
root CMakeLists.txt
set(RETINA_HEADERS
and you still use cgltf, i will tell @loud crag
wicked notch
set(RETINA_HEADERS
I dun see dis 
wispy spear
oh did you perchance remove it already :3
wicked notch
mayhaps
I don't remember ever having that but 🅱️erhaps my brain is failing me
as it usually does
wispy spear
oh
looks like i was on an old commit 
.Entry is still called .Entry though hehe
man your code is so readable
wicked notch
is it
I still need to make rg to separate the passes
rn everything is in application
wispy spear
i fink it is, still not sure about the C for classes, but i very much stole the S and E prefix too now 🙂 and the auto Foo() -> ism
wispy spear
dlss commit is outdated it seems
make] fatal: Fetched in submodule path 'NVIDIAImageScaling', but it did not contain 35e13ba316c98eeecf16f37eae70ce88019911f6. Direct fetching of that commit failed.
[cmake] CMake Error at nvdia_dlss-subbuild/nvdia_dlss-populate-prefix/tmp/nvdia_dlss-populate-gitclone.cmake:62 (message):
it should be disabled on lunix anyway i suppose, perhaps with some autodetection and a message?
wicked notch
ye
wispy spear
trying clang 17.0.6
ah lol?
dlss cloning worked now
it cant find #include <vulkan/vk_enum_string_helper.h>
wicked notch
wot
wispy spear
hopefully not a new vksdk ;c
wispy spear
eh hang on
wicked notch
the string helper has been there since 1.2 at least
wispy spear
this is weird
vulkan/vulkan.hpp is coming from /usr/include/vulkan wtf
[deccer@rootfs ~]$ sudo pacman -R vulkan-devel
checking dependencies...
error: failed to prepare transaction (could not satisfy dependencies)
:: removing spirv-tools breaks dependency 'spirv-tools' required by glslang
:: removing vulkan-headers breaks dependency 'vulkan-headers' required by qt6-base
:: removing spirv-tools breaks dependency 'spirv-tools' required by shaderc
``` ;Cvulkan-headers methinks
wispy spear
i have no memories installing anything qt6y
wicked notch
kde I think is qt
wicked notch
issa ok, I'll fix linux building once and for all now
wispy spear
haha
ah
its gnuplot
and that stupid patchpanel for thingy pipewire
and obs : )
removing ffmpeg breaks dependency 'ffmpeg' required by firefox
nice
wicked notch
don't nuke your os
wispy spear
: )
good thing is its ez to reinstall
but why would firefox have a hard dependency to ffmpeg
thats new
important question is, if i remove firefox and reinstall it, will it rember my current 250 open tabs 😛
ok i have to clean the tabs first anyway... will do that first
and i should also start working on local build containers
wispy spear
down to 150 tabs : >
wispy spear
down to 14
loud crag
IDEs or analyzer tools want that and might break or fail to analyse the headers if you dont add the headers
frank sail
perhap
wispy spear
had no trouble with clion or vscode-cpp so far at least
in my cmake/openglstarted and other cppisms for now
wicked notch
IDEs or analyzer tools want that and might break or fail to analyse the headers ...
clion werks fine without
primal shadow
Finished revamping my 2pass occlusion culling to work with LODs, and be a lot cleaner/better in general!
Reusing the previous frame depth pyramid, instead of explicitly tracking cluster visibility between frames
loud crag
@wicked notch you know if nvpro_pyramid works well on non-nvidia gpus? i’m thinking of using it, too, but reading comments in the source like „subgroupSize other than 32 is not tested, should work, message [email protected] if not“ is throwing me off a little
perhaps ffx spd is a better fit for me
primal shadow
<@320895822394818561> you know if nvpro_pyramid works well on non-nvidia gpus? i...
Neither SPD or nvpro_pyramid work for non-power2 textures either, which is a huge pain. I haven't seen anyone handle it.
loud crag
huh, are you sure?
the nvpro dispatcher seems like it can handle anything that is a multiple of 4 by default
that number is templated, though, so perhaps it supports other factors, too
primal shadow
It probably doesn't enforce power of 2 for the textures or generates
Which would prevent you from using mips of a single texture
glass sphinx
Reusing the previous frame depth pyramid, instead of explicitly tracking cluster...
not tracking visibility reduced perf a lot im my tests
maybe i missunderstand what you do
but if you use last frames depth i assume you draw twice, one culled against last frame once culling against partial new frame
primal shadow
but if you use last frames depth i assume you draw twice, one culled against las...
Yes
I'm doing exactly what nanite does
primal shadow
Started porting SPD to bevy. Non power of 2 is actually easy. The samples out of bounds will just return 0, which ends up ensuring conservative depth, at the cost of a poorer culling on the edges of the screen. Totally fine imo.
wicked notch
nvpro_pyramid does support NPOT
primal shadow
Almost finished with porting SPD!
loud crag
dlss be working beautifully
wispy spear
you can use the already exiting short cut alt+f4 to reload the thing
buoyant summit
astral field
c++ classes 
buoyant summit
slang has crappy classes
well it doesn't have classes, only structs, and those aren't crappy themselves, but the semantics for this in methods is hlsl-tier garbage
because this is basically inout
:barf:
astral field
make an issue 
primal shadow
Ok finished SPD port to Bevy/WGSL and using it in my nanite-impl 🙂
Much better perf
glass sphinx
because `this` is basically `inout`
ah wtf
glass sphinx
spirv doesnt have a concept of classes at all
glass sphinx
does spirv have storage class - generic pointers?
buoyant summit
I don't like spir-v generic pointer tbh
or rather
I'm kinda suspicious about just throwing it into vk as is
astral field
they're good if good tooling could use them (vcc & slang one day)
although I don't know the extent to which they can be used (never checked up deeply), I only really know about restricted ptr's
glass sphinx
buoyant summit
<:froge_love:1105211408255295624>
it's not a thing in vk so don't get too excited (though slang probably easily could be taught to emit that and then fed into shady to lower generic ptrs to tags 
)
and the generic pointers issue I opened a while ago is on a really slow cook
primal shadow
I'm hitting the match dispatch limit for culling workgroups 😬
memory also continues to be really annoying to allocate large amounts of
frank sail
I'm hitting the match dispatch limit for culling workgroups 😬
how big
primal shadow
Caused by:
In a ComputePass
note: encoder = `<CommandBuffer-(3, 3, Vulkan)>`
In a dispatch command, indirect:false
note: compute pipeline = `meshlet_culling_first_pipeline`
Each current dispatch group size dimension ([103039, 1, 1]) must be less or equal to 65536
I can do the same dumb trick I did for the other pass and make it a 3d dispatch I then remap to 1d in order to get more workgroups 😛
Or use bigger workgroup sizes ig. Currently it's 64x1x1.
wheat haven
I can do the same dumb trick I did for the other pass and make it a 3d dispatch ...
exactly what I did when I hit that snag
primal shadow
exactly what I did when I hit that snag
Why must I though :/. Why can't drivers be smart.
craggy shale
because actual magic doesn't exist and "magic" means heuristics and heuristics suck
wispy spear
im also confused about cmake 🙂
one cmakelists does if (RETINA_ENABLE_PROFILER) the other one does if (${RETINA_ENABLE_PROFILER})
wispy spear
i cant it to work either
i stole the profiler one
no matter what i set or option with ON and with or without CACHE and or CACHE BOOL it wont pick it up
unless i really have to yoink build/ physically and not just Delete Cache & Reconficture
loud crag
I think it's equivalent
yeah in this context it’s the same but it does have a different meaning
astral field
```
Caused by:
In a ComputePass
note: encoder = `<CommandBuffer-(3, 3,...
2^16 probably not 100% random
delicate rain
Mister mister @wicked notch please show your shadows
delicate rain
Thank you
wicked notch
clang is a requirement
frank sail
you can install clang through the visual studio installer if you want to use VS
primal shadow
Yes. Here's a breakdown of problems:
- The two green vkCmdCopyBuffer()s (4ms each) are staging buffer copies. I'll probably want to use ReBAR and directly map buffers
- The write_index_buffer pass (10.5ms each, 2 per view) are slooooow and have horrendous occupancy. Couple of things I can try:
- Have culling pass write out a list of all visible clusters, indirect draw to spawn write_index_buffer pass workgroups, instead of spawning 1 wg per cluster and just using early exit if they were culled
- Try to merge it with the culling pass
- Add software raster and hope it reduces the need for hardware raster
- Add mesh shaders to wgpu
- Raster (2 per view, 3.3/4.3 for main/shadow view in the first one, and then second is basically free because occlusion culling)
- I'm 83% PES+VPC throughput, which I think means primitive assembler limited? Nothing I can do to fix this really. Software raster + mesh shaders might help, again.
wicked notch
I don't remember our method for writing meshlet index buffers to be that slow tbh
idk how it came to be like this
primal shadow
This is the shader https://github.com/bevyengine/bevy/blob/0ebd414dbc13ef77b84627ca5d0e82b72cd50262/crates/bevy_pbr/src/meshlet/write_index_buffer.wgsl. 1 workgroup per instanced cluster in the entire scene regardless of culling.
primal shadow
Hold on had to go walk somewhere. Give me 15m to get back to my PC.
I'm suspicious that the spawn 1 wg per cluster and early exit if culled is too slow
Because notice both are slow, first and second pass
Even though the second pass should be doing nothing
glass sphinx
Yes. Here's a breakdown of problems:
* The two green vkCmdCopyBuffer()s (4ms eac...
pes + vpc probably means hw culling is the limiting factor
like backface and frustum
also im pretty sure nanite uses no index buffer
writing it is just too slow
glass sphinx
what is the write index buffer doing then
primal shadow
pes + vpc probably means hw culling is the limiting factor
I swear I turned off backface culling in my pipeline descriptor... I'll have to chdck
primal shadow
also im pretty sure nanite uses no index buffer
In my testing using a whole indirect draw args per cluster was wayyy worse. Maybe it's better when almost all of your clusters are software rasterizef though...
glass sphinx
yes it is
idk what your index buffer is
writing out the meshlets should be at most 1/10th of the draw tome
time
primal shadow
idk what your index buffer is
1 u32 (cluster id + triangle id) per triangle, so that the vertex shader knows what to draw
glass sphinx
so your indey buffer is really a primitive buffer i see
hmmm intresting that thats so slow
glass sphinx
yes
ah btw
remove the wg barrier
the atomic will be optimized by the hw if you are on nv
the barrier will most likely make it slower
primal shadow
Will do when I'm home, 5m
glass sphinx
amd and nv bith have heavy hw optimizations once you have extreme contention
so if all threads in a warp use the same address
it will catch that and do one atomic + warp prefix sum instead
instead of warp size n atomics
as soon as one thread diverges you die
wicked notch
in that case there is only one thread in the wg doing the atomic tho
glass sphinx
yea i believe this will be slower
the gain is lower then the cost of the barrier
so what im suggesting is doing the atomic in all threads and let the he catch it and make it fast
this way you don't pay for the barrier
maybe that will make no difference tho
it can make a bug difference
but its hit or miss
glass sphinx
Yes. Here's a breakdown of problems:
* The two green vkCmdCopyBuffer()s (4ms eac...
what i did is write out a bitmask per cluster instead and early out in zhe vertex shader instead
this way you write way less
that was fast af
wicked notch
so what you're saying is that the driver just does this
const uint local_offset = subgroupExclusiveAdd(meshlet_primitive_count);
uint global_offset = 0;
if (subgroupElect()) {
global_offset = atomicAdd(..., meshlet_primitive_count * 3) / 3;
}```shrimple as dat
wicked notch
hardware accelerated subgroup memes
glass sphinx
so what you're saying is that the driver just does this
```cpp
const uint local_...
yes
its the same with branches btw
if all threads have same condition only one path is taken
there are more insteuction slike this where the hw will check if all threads use the same x to go faster
glass sphinx
sc?
frank sail
the hardware also does stuff
glass sphinx
shader compiler
wicked notch
imagine trusting any compiler for glsl
frank sail
imagine trusting any compiler for glsl
it's not a glsl compiler 
glass sphinx
so it can do it even if its unknown
primal shadow
Ok I'm back, let me see
primal shadow
what i did is write out a bitmask per cluster instead and early out in zhe verte...
This is in reference to what?
glass sphinx
the mask masks tris in a meshlet
so one bit is one triangle
i had a limit of up to 64 tris per cluster so it was just 2 uinzs
primal shadow
remove the wg barrier
Here?
// Reserve space in the buffer for this meshlet's triangles, and broadcast the start of that slice to all threads
if triangle_id == 0u {
draw_index_buffer_start_workgroup = atomicAdd(&draw_indirect_args.vertex_count, meshlet.triangle_count * 3u);
draw_index_buffer_start_workgroup /= 3u;
}
workgroupBarrier();
How is removing it safe? Other threads may read the value before thread 0 writes it no?
wicked notch
i had a limit of up to 64 tris per cluster so it was just 2 uinzs
did you try the prefix sum method btw
I wanted to try it but I then just hopped on the mesh shader train
wicked notch
and the tri mask was the best?
glass sphinx
on a 4080
wicked notch
well
ye I would imagine blasting 100 million vertices for a 4080 is just another tuesday
glass sphinx
well, on a 1080ti ir was kinda the same
primal shadow
i had a limit of up to 64 tris per cluster so it was just 2 uinzs
So instead of writing a buffer of cluster|triangle IDs so each vertex invocation knows what data to fetch, you wrote out a list of just cluster IDs (1 per cluster), and then hardcoded the draw size to 64 * clusters? And then each vertex can find it's cluster via vertex_id % 64, and then you just output NaN for excess triangles?
glass sphinx
i should have used nan as amd fats paths those
glass sphinx
So instead of writing a buffer of cluster|triangle IDs so each vertex invocation...
yes
primal shadow
I tried that. It was slow...
glass sphinx
what was slow
primal shadow
The draws. All the extra vertex invocations were expensive.
glass sphinx
yea but now your compute is slow
primal shadow
hrmm maybe it's worth a second test...