gistfile1.txt

Get it running

1. Build
cd ~/projects/rhi-zig/examples
zig build -Dtarget=wasm32-freestanding

2. Serve (WebGPU needs a secure context — file:// won't work, localhost counts)
cd zig-out/bin && python3 -m http.server 8000

3. Launch Chrome with WebGPU enabled
google-chrome --enable-unsafe-webgpu http://localhost:8000/00_clear.html
Or persistently: chrome://flags/#enable-unsafe-webgpu → Enabled → relaunch.

Why it failed

Chrome 151 on Linux still gates WebGPU behind that flag. I confirmed it directly on your machine — same page, same build:

┌────────────────────────┬───────────────────┐
│         Flags          │ requestAdapter()  │
├────────────────────────┼───────────────────┤
│ none                   │ null ← your error │
├────────────────────────┼───────────────────┤
│ --enable-unsafe-webgpu │ adapter OK        │
└────────────────────────┴───────────────────┘

Your GPU is fine (RX 7800 XT, RADV, Vulkan 1.4.335). Firefox needs dom.webgpu.enabled in about:config instead.

One thing I have not verified — don't assume it renders yet

With the flag, I confirmed on your hardware: adapter ✓, device ✓, glue loads ✓, wasm instantiates ✓, boot() resolves ✓, rAF frames run ✓, and zero WGSL or pipeline-validation errors from Chrome — so the Slang→WGSL output and both render pipelines are valid on a real implementation.

But when I read the canvas pixels back, it was entirely black — 00_clear should be rgb(26,51,102). No console errors, no WebGPU validation complaints. So work is submitted and accepted, yet nothing reaches the canvas.

That's either a real bug in my frame path (likely candidates: the surface-configure/getCurrentTexture handshake, or the swapchain-rebuild path when the initial size is 0) or an artifact of reading pixels from a headless/occluded window. I was mid-investigation — adding call counters to the glue to find which stage isn't running — when you interrupted.

So: open it yourself and tell me what you see. If it's black for you too, I'll finish tracking it down; if it renders, the black readback was my measurement artifact.