Finally found the cause: override doesn't seem to work with compute shaders in Safari.

This sucks. None of my prior WebGPU sketches work on the new Safari…but works perfectly fine on Chrome…

Also, I’ll sound very harsh: do yourselves a favour and drop any projects or ponder on solutions that are based on WebGL. It is outdated, performs horribly and requires tedious hacks to MAYBE get what you want. With WebGPU, you can think about and solve problems in a more reasonable fashion

Not that I would consider myself an expert at this point (yet), but PBR is surprisingly straightforward! The hassle is generating the assets needed for IBL.

All right: time to learn what is actually going on behind a PBR shader and make one myself.

Marching cubes experiment made with my own WebGPU framework 🗿🌊 Hope you'll find it fun and inspiring (and satisfyingly disgusting 🫠)! dougllab.netlify.app/marchingcubes #webgpu #creativecoding

Very nice. Removed an awkward copyBufferToTexture and now my 3D blurring is even faster! Guess I'm ready to start diving in to marching cubes... #webgpu

Have to try this, but after reading some fine print…storage textures can eventually be used with samplers depending on the bindings. This is awesome, but this means i have some hefty refactoring to do on my fluids and shadow volumes…

Raymarching volumetric lighting is a bit slow… :(((

My voxelizer is working! 128^3 voxels < ~150us. 256^3 around < ~800us (on my M1Pro) Ironically, visualising the voxels is heavier than the voxelizer itself... *cackles manically* #webgpu

Too much at work. Have some fun concepts I want to explore for a few days during my vacation.

Cool accident while updating a previously SD 1.5 based workflow to SDXL / HotshotXL. Not a fan of how all the motion models are aiming for "perfection" and in my workflow's case, I need AnimateDiff + frame interpolators still! #comfyui #aiart

If there is anything that worries me, it is not “AI taking our jobs”. In the end, an experienced dev will know what to prompt compared to one with less experience. I’m more concerned about project deadlines becoming more and more unrealistic due to the false belief that “AI can do everything”

125k flocking boids particle volume shadows 3d fluids (32^3) using hand tracker as cursor 120 fps on M1 pro Learned a ton doing this project, but I'm bored of particles at this point 😅 Hopefully this will be fun and inspiring to whomever tries this out! dougllab.netlify.app/birds #webgpu

Kindoff getting there! Using pressure makes it more interesting... Maybe a habit thing, but interacting in a 3D space feels like you need to "master" the interaction which isn't too fun. More parameter tweaking! 🪛 #webgpu

Ah! You *can* use samplers in compute shaders! Just have to make sure you are using textureSampleLevel! #webgpu

After this practice project I'll be going back to soft-bodies. Particles are getting quite repetitive now...

Back at it! Integrating my 3D fluids into my original boid simulation. Lots of finetuning is needed and scaled up the particles to avoid compression 😅 Doing so reveals a fair bit of aliasing... Also learned that using samplers is not allowed in compute shaders :((((( #webgpu #creativecoding

Few subtle error fixes later and now it looks even better! 🌊 #webgpu #creativecoding

With gradient subtraction (familiar fluid look) 🌊 and with my setup it is easy to toggle between the two behaviors  💨 #webgpu

Just to show I'm not bluffing 🤓 #webgpu

Just need some calibration, but the handtracking is doing wonders! For the Z, I compute a 2D bounding box and use the area as the Z position (far away == sm0l hand, big hand == close) #webgpu

A few months back I made an attempt to recreate the hiss warp effect from the game 'Control'. Curious to move this exploration to WebGPU... #webgl #creativecoding

My hunch was correct. What I thought was working was just seeing my density being diffused...but never advected. *Now* it is properly working 🌊 Time to set up hand tracking as using a regular trackpad is not too exciting...and conveying a higher dimension on a 2D surface makes no sense. #webgpu

Troubleshooting. Maybe it's my OCD, but had a feeling that my fluids never got properly advected... Believe this is how it should look like...will continue at a later point. #webgpu

Yeah it totally works 🔥 Will make an optional projection step as well 🌊 The idea is to use this as a driver for applying forces to stuff #webgpu

Think I got it working! #webgpu

More debugging... Anyhow, 32x32x32 sim at 120 fps on m1 pro mac and no blasting fans...same sim in WebGL would be toasty... #webgpu

debugging #webgpu

Believe me. This is 3D fluid. #webgpu

getting there! #webgpu

⚠️ strobes and flickers ⚠️ Well...something is happening hah A bit of unlearning to do as I have never done a fluidsim with compute shaders before... Following along Jos Stam's paper on fluid sims #webgpu

bleeeghh...sucks that atomics are not supported on storage textures....

Think I have a new favorite Björk song...

All right. Fluid sims it is. Position based fluids is something I'll look at some other time...mainly because I don't have to think too much about working around WebGL's limitations.

Cool. MSAA done.

All right...MSAA next and then time to decide if I want to port a prior experiment or try making 3D fluid simulations... #webgpu

Yay! Render targets and the ability to create render attachments (MRT's) are working! And yes, all buffers are destroyed and created anew when resizing. #webgpu

Building a framework is oddly satisfying...

🎉 proper (and manual) filtering + blurring did the trick! Time to make pretty things #webgpu

Before that...I'll have a look at temporal filtering. Or at least a decent implementation (ping-pong between previous and current state of particle shadow map based on particle velocity)

3D fluids or see how much easier it will be to implement shape-matching while I'm still drunk on power now that compute shaders enables so much

I *think* the banding is fixed...

Hmmm, guess it isn't enough to blur values in a storage buffer, followed by a transfer. Anyhow, got a bit more familiar with workgroupBarriers and workgroup storages...good stuff! #webgpu

Sped things up by packing light values along Z (per the paper) and a few subtle quirks later, I'm now *almost* finished! Currently now at a point where there is a clear staircase aliasing as opposed to a noisy aliasing.

Albeit powerful as hell, race condition managment with compute shaders is an interesting type of problem...still better than figuring out data based rasterization nonsense

What is liberating with WebGPU is not having to stop in a thought process due to “oh yeah I don’t have compute” or “might be possible, but will run and look terrible”

Almost there! The light propagation is now working correctly (apparently for loops is peanuts for compute shaders). Now it's time to make things smoother. The obvious fix is to account for all particles within a voxel and not just one (which is why we see the blocky aliasing) #webgpu

Progress! Artifacts was due to compute shaders being less tolerant than fragment shaders in terms of sampling. Now that I know conceptually that the approach works in WGSL: time to to look in to packing 4 particles in each voxel! TBD if I need atomics (which also caused issues) #webgpu

Embarking on particle shadows! Trying an approach proposed by Crytek: d29g4g2dyqv443.cloudfront.net/sites/defaul... The aliasing I can understand as I'm currently stuck with 64 Z layers due to WebGPU only allows for 64 workgroups along Z...but no idea what is causing the propagation lol