Sabitlenmiş Tweet
The future of Blender Hair + Node workflow?😉 Been playing around with geometry nodes to get some hair interpolation, clumping and curls going!
Watch the full length demo: youtu.be/2M8F8PvWiA4
#b3d #geometrynodes
YouTube
English
Bert Van den Bosch
711 posts
Bert Van den Bosch
@bertvdbosch5
CS grad @KULeuven | Engine- & Graphics-Engineer @Cyborn3D
Belgium Katılım Aralık 2016
582 Takip Edilen114 Takipçiler
Bert Van den Bosch retweetledi
wow @garrytan just exposed Anthropic as total frauds
Claude Code was ONLY 512K LOC ☹️
Gary is shipping 37K LOCs PER DAY
so Gary could recreate all of Claude Code in ONLY 13 days!
a supposedly $380 billion is big trouble
English
@duhonedd @SebAaltonen Interesting take, although most renderers running on Adreno 7xx are probably already saturating texture pipe hw because of the material textures?
English
Maybe a compromise. Positions stay available for hardware vert fetch. Rest change to load. Since positions probably benefit the most from hardware fetch, binning pass. I'd be curious to see a performance comparison on Adreno. If hit is small enough, I might take the simplicity instead. Especially since adreno will dump the fetch units at some point anyway, if they haven't already in 800 or newer
English
Setting specialization constants in our shader PSO create API. You provide a span of constants. In the example the span is filled from initializer list (all stack objects, no allocs or copy).
.specializationConstants = {
gfx::shader_constant_bool(alpha_clip_enabled)
}
English
@SebAaltonen Also in the HypeHype renderer? Still lots of mobile gpus using texture hw instead of raw buffer loads when doing vertex pulling... Looking at you adreno
English
I will be nuking the vertex buffer layout soon. 100% vertex pulling in the future. I don't like having to describe my data layouts to the graphics API. It should not need to know my data layout.
English
@SebAaltonen Wait, what is the difference between the two shader variants (old) and the new 2 hardcoded shaders?😅
English
Older results (ubershader without spec constants):
x.com/SebAaltonen/st…
Sebastian Aaltonen@SebAaltonen
Alpha clip timings (6K res): G-buffer: Discard in main shader: 2.67ms Two shader variants: 1.86ms (70%) Binned (discard last): 1.44ms (54%) Shadows: 2 variants: 0.17ms, 0.10ms, 0.29ms, 0.22ms Binned: 0.10ms, 0.07ms, 0.27ms, 0.07ms (65%) Thread...
English
Specialization constant discard timings (6K res):
G-buffer:
2 hardcoded shaders: 1.44ms
Spec constants: 1.45ms
Shadows:
2 hardcoded shaders: 0.10ms, 0.07ms, 0.27ms, 0.07ms
Spec constants: 0.11ms, 0.05ms, 0.27ms, 0.07ms
Identical performance. Some measurement noise of course.
English
@SebAaltonen How is the order of the drawcalls influencing the TBDR binning, I was under the impression the binning pass only executed the vertex shader?
English
If you run all shaders that could perform discard last, you guarantee that rest of the scene doesn't suffer from worse Z-compression / early-Z / Hi-Z performance. And TBDR doesn't need to do extra partial tile evaluations.
English
Alpha clip timings (6K res):
G-buffer:
Discard in main shader: 2.67ms
Two shader variants: 1.86ms (70%)
Binned (discard last): 1.44ms (54%)
Shadows:
2 variants: 0.17ms, 0.10ms, 0.29ms, 0.22ms
Binned: 0.10ms, 0.07ms, 0.27ms, 0.07ms (65%)
Thread...
English
SSAO bilateral blur:
- Codex initial version = 0.085ms
- With gather4 optimizations = 0.03ms
fp16 optimizations didn't make it any faster, but hopefully save power.
English
@SebAaltonen @CUDAHandbook But you do want a single position vertex stream for the tiling pass on TBDR gpus
English
@CUDAHandbook Vertices are not in SoA layout on purpose, because index buffer is not contiguous. You want AoS memory layout for vertices. In SoA layout, you end up with N cache misses instead of 1 whenever the index is not contiguous. Not good.
English
This reminded me that I should kill vertex buffers and vertex input assembly setup APIs in our RHI. We no longer need GLES backwards compatibility.
Vertex vertex = buffer.vertices[gl_VertexID];
Tech Bro Memes@techbromemes
English
Bert Van den Bosch retweetledi
@kimmonismus you aren’t though.
none of these things are games, none of them have been shipped, people are not playing them, they are twitter demos.
if this guy can prompt all these games into existence, why is he selling courses instead of selling games?
English
@arno_coomans These results look very promising! Are there also any comparisons between probes and this technique in a more 'open-world/outdoor' scenario where there is possibly less variance in the GI?
English
Our Neural Irradiance Volume (Eurographics 2026) permits real-time rendering of large scenes with dynamic objects and moving lights, while providing a higher quality at a given memory budget (10x improvement over probe grids!). Project page: arnocoomans.be/eg2026/.
English
Bert Van den Bosch retweetledi
@BattleAxeVR What about using vulkan texture blitting? Provides a hw fast path on most mobile gpus struggling with vram bw
English
Single-pass MIPGEN in CS are certainly beneficial and I will implement it in my own engine, however very low priority. I profiled dynamic mipgen on post-DLSS upscaled images for lens unwarping on my RTX 5090 and it was very, very fast. But lower VRAM bw GPUs wd benefit, no doubt.
Kiriakos Gavras@Kiiiri7
Wrote a blog post about the Single Pass Downsampler I implemented for my depth pyramid. Code included. syllogi-graphikon.vercel.app/posts/metal-si…
English
Upcoming Niagara stream! This *Sunday* (Feb 15), at 11 AM PST (7 PM GMT), we will embark on a journey to replace all descriptor uses within the renderer with descriptor heaps. youtube.com/live/VXN4Gewjk…
YouTube
English
Bert Van den Bosch retweetledi
If you're a bottleneck to AI doing its job it, you're doing menial tasks, glue code, low complexity work.
Look for work that would require your taste, aesthetics, sense, knowledge, hunches, heuristics, etc.
Thorsten Ball@thorstenball
"I am the bottleneck now" Few more thoughts
English
Bert Van den Bosch retweetledi
Daily reminder that Bungie made every Halo level a .ASS file and every object you put inside the .ASS is known as a poop.
This is hard-coded, you cannot change it.
English
Bert Van den Bosch retweetledi
Late 60's / early 70's was peak music. It's not really up for debate.
Crazy Vibes@CrazyVibes_1
Be honest, who still actually listens to music from the 1960s through the 1980s in 2025?
English
@gdechichi @panoskarabelas Personally I batch by material e.g. large indirect draw per material type. Still have to figure out some compaction if all draws get culled indirectly… This still leaves the user’s option to use a single master material and use a single batch!
English
@panoskarabelas how you are handling material uniform buffers for different material types? is the layout determined by the largest material struct, or is there some indirection?
English
I got GPU driven rendering working thanks to the single geometry buffer, bindless textures and samplers, and an uber shader approach that keeps PSOs down. Very low state permutations, the engine lives up to the Spartan name, and it made the implementation much easier, leaving it for last was the right call. Draw calls dropped a bit more too, at this point most of them are probably ImGui.
Panos Karabelas@panoskarabelas
I switched from per mesh vertex and index buffers to a single global geometry buffer that holds everything. This is the same idea id Software used in id Tech for Doom, and you can see the vertex and index binding counts drop, which opens the road for easier GPU driven rendering.
English
@SheriefFYI Wait what? This is some unexpected amazing news 😅
English
wtf (I still play Uncharted 3 split screen to this day)
Freako@Freako1
Factions Map running in Uncharted 3 update: I have spawn/respawn locations implemented. Making entire matches possible. Gotta get a friend over for LAN.
English