D822 - A fragment module reads its quad's neighbours through the host's quad

Status: decided Date: 2026-10-13

Amends D716. A fragment module still simulates one lane, one pixel per invocation. An instruction that reads another lane of its quad - a DPP16 quad permute of a one- or two-source vector instruction's first source - reads it from the host invocation at that place in the host's own 2x2 quad, by OpGroupNonUniformQuadBroadcast, and finds its own place from the low two bits of its subgroup index. The source lane's execution bit is read the same way, so the inactive-lane rules (FETCH_INACTIVE, BOUND_CTRL) hold. Only a module that has such an instruction declares the quad capability and the subgroup index, found by a scan before the entry point is written.

The permuted source is staged in a 257th vector register, one past every register an instruction can name, and the instruction translated as though it named that register. Every model shares that path; one that holds every lane reads the neighbour by index.

Why: a derivative is a quad permute - compilers emit v_mov_b32 with quad_perm:[0,0,2,2] then a subtract with [1,1,3,3] - and D716's premise, that no translated instruction reads another lane, ends there. The guest's quad and the host's are the same object: Vulkan numbers a fragment quad top-left, top-right, bottom-left, bottom-right, as the guest does, and runs helper invocations so every place in it exists. gl2-probe (GLTP00001) checks its own pixels, and its derivatives case passes once this is in.

Rejected:

• Simulating four lanes per fragment invocation: each would need its own interpolated inputs, which the host computes only for the pixel it is.
• Declaring the quad capability on every fragment module: it would make every fragment shader need subgroup support in the fragment stage for the few that permute.
• Rewriting the source operand in place: the instruction set has no spare register to name, and the permuted value must not overwrite one the guest reads later.

Assumed: that the host device supports quad operations in the fragment stage. Desktop drivers do; nothing here checks it, and a device without it would fail to create the pipeline.