Struct

`gpu::metal::emulated::DispatchShape`

cheatah-gpu v0.5.1-alpha — Biome Standard 0.6.3-alpha

The full shape of one dispatch, as a 3-D-aware stand-in kernel receives it.

threads is the TOTAL thread grid per axis for BOTH dispatch forms — dispatchThreads passes its grid through; dispatchThreadgroups is normalized to groups × threadsPerThreadgroup per axis (what real Metal launches) — so a kernel loops threads without caring which form dispatched it. There is no threadgroup-memory emulation: a C++ stand-in shares the whole address space and needs none (MSL threadgroup tiles are an Apple-hardware concern).

Unit testMetalCompute.Grid3d MetalCompute.Threadgroups

Constants & variables

var Extent threads #

total threads per axis (normalized, both dispatch forms).

var Extent threads_per_threadgroup #

the second MTL::Size the dispatch call passed.