[jowens]

Stacked on #82, which is stacked on #81. This finishes moving every Gridwise primitive to declared buffers and UniformStruct.

• DLDFScan: ScanParameters is a UniformStruct, and its six buffers are declared with buffers. The output type is array<vec4<T>> for a scan and array<T> for a reduce.
• Core, UniformStruct slots: OneSweep gives each kernel its own parameters, packed into one uniform buffer at aligned offsets. That's now a core capability:
  • allocateSlots(label, valuesList, alignment) packs the sets;
  • slot(label, i, alignment) is the binding for set i.
• OneSweep:
  • SortParameters is a UniformStruct with one slot per kernel.
  • Its buffers are declared with buffers, in two bind groups. Keys and payloads swap roles on odd passes.
  • The long layout comments for hist and passHist moved along with the declarations.
  • The last use of WriteGPUBuffer in a primitive is gone. It's still tested by test_wgsl_functions.mjs.

Checked

• WGSL, layout and bindings compared with main:

  • DLDFScan, 4 configurations (including emulated subgroups): identical apart from where the declarations sit.

  • OneSweep, 26 kernels over 4 configurations (keys u32, key-value i32 descending, u64, KEY_BITS 12): identical apart from three things:

    • where the declarations sit;
    • the struct's padding fields;
    • each slot's binding now having an explicit size (32 bytes) instead of running to the end of the buffer.

    The offsets (0, 256, 512, …) are unchanged.

• Simulated RDNA4 (gfx1201): 83 of 83 tests pass, including all sorts, both scans, permutations, histograms and the new core test.

With this, no Gridwise primitive declares a binding by hand any more.

---

Generated by Claude Code