David Neto

> As it is, I'm not confident we can reasonably port our engine to WebGPU. I would consider that an utter failure of WebGPU. > I would argue the problems...

Related #2321 which discussed an escape hatch for marking a load from a storage buffer as uniform. ----- Ah, added that before I saw/read @raphlinus comment.... :-) I was triaging...

> It's going to be tricky to design a targeted escape hatch that can be used by translation chains, because the upstream languages won't carry any information that could indicate...

I took a closer look at @raphlinus' example. There's a function-scope variable `ready_ix` that triggers the uniformity violation. The loop exit depends on `ready_ix`, near line 2135: ``` let x_2197...

> Yes, that's very much in line with my thoughts. Ok, so some kind of big switch. > > I'm a little concerned that making it useful for your second...

Is this critical for v1? I've never seen HLSL code use namespaces. I didn't even know HLSL had namespaces until I dug quite far. I didn't see usable documentation about...

The direction discussed above is that things that are nameable from the API side must have unique names across the entire source text. These are entry point names *and* override...

I still don't buy this is necessary for V1. Very large C programs don't need namespacing and scoping like this. For example, the Linux kernel is a very large program....

When I was digging into that, I went to the best reference I know. The Vulkan spec goes into all the excruciating but necessary detail about how coordinates are transformed...

For example, Metal keeps it pithy, to a fault. https://developer.apple.com/documentation/metal/mtlsampleraddressmode/clamptoedge > Texture coordinates are clamped between 0.0 and 1.0, inclusive. The D3D11 Functional Spec is fully explicit, with many details...