webgpu

Start from a resource and pass graph, define exact WGSL struct layouts, compile shader modules with getCompilationInfo diagnostics, then create pipelines and stable bind groups. The bundled Vite starter demonstrates a complete compute-and-render particle field with ping-pong buffers.

Calculate AlignOf, SizeOf, member offsets, and array strides from WGSL layout rules rather than assuming 16-byte alignment for all structs. Pack mixed scalar types with DataView at explicit byte offsets, and remember uniform arrays require 16-byte strides.

Blank output usually comes from format mismatches, missing getCurrentTexture calls, or silent validation errors. Flicker on resize comes from reassigning equal canvas dimensions; size the backing store from CSS pixels and DPR, and skip no-op assignments.

Yes, if you build a portable core path and gate optional features like shader-f16, timestamp queries, and subgroups behind adapter feature checks. Request only the features and raised limits the workload actually uses, and provide a fallback or clear failure UI.

Avoid full-buffer readback in animation loops. Compact results on the GPU into a bounded output using atomics, copy a small staging buffer, and map it asynchronously with mapAsync, rotating staging buffers for recurring reads.

Treat device loss as a lifecycle transition: listen to device.lost, stop animation loops and observers, and rebuild from a new adapter or switch to a fallback. Resources from a lost device cannot be reused, so keep CPU-side descriptors for recovery.