규칙 8.21

A host call passes and returns host values by value: an f32, f64, i32 or u32 is a number, a bool a boolean, a vector a readonly tuple of its components as an argument and a tuple as a result, a matrix a flat column-major array of its components, an array<T, N> an array of N host values of T, a struct an object of its fields, and an enum member its number. A fieldless class is {} on the host and CPU. Its reflected memory footprint accounts for the GPU’s internal carrier: four bytes under natural layout, sixteen under std140, with no visible fields. Nested fields and array strides follow that footprint. Each argument is checked and converted: an ArrayLike of the right length becomes a fresh array, an f32 is rounded as a buffer write rounds it, and a value that does not fit is refused with a TypeError naming the function, the parameter and its TypeShade type; the result aliases no argument and nothing the module keeps (Rule 8.8), and an exported constant is a frozen copy. A call of a function this rule and Rule 8.20 admit is synchronous, and no later tier changes that; as a helper’s parameter, a runtime-sized array, an atomic, a texture, a sampler and a binding have no host value. A binding of an entry a host calls (Rule 8.24) takes a host value too: a uniform<T> or a sized storage<T> takes T’s, a runtime-sized storage<array<T>> of scalars or vectors takes the scalar’s typed array (Float32Array, Int32Array, Uint32Array, or Float64Array for an f64 and a vecNf64) with the vector’s components one after another, which the call pads to the element’s stride, one of structs takes an array of objects, and either may be a Resident of it (Rule 11.8), an atomic takes its integer’s, a texture_2d<f32> takes an image source (ImageBitmap, ImageData, HTMLImageElement, HTMLCanvasElement, HTMLVideoElement or OffscreenCanvas), uploaded at the call, and a sampler takes { filter?, address? } ('nearest' or 'linear'; 'clamp', 'repeat' or 'mirror'), or nothing for linear and clamp. An emulated f64 (change 0013’s split) is held by the GPU as two f32s, hi and lo, and a vecNf64 as a plane of each: the runtime splits a double into them and joins them back, and binds the _fp64 guard of a module that emulates one, which the host never passes; a matNxN<f64> has no host value yet. A storage binding the entry writes is read back into the caller’s value in place, so a written binding whose type is one scalar is passed as a typed array of length one. A kernel function (Rule 8.22) is called asynchronously: it returns Promise<void>, or Promise<R> for a result R; its value parameters take the host values above, its arrays with no size take the scalar’s typed array (Float32Array, Int32Array, Uint32Array, a vector’s components one after another) or, for a struct, an array of objects, and each array it writes is read back into the caller’s in place; each array may be a Resident of it instead (Rule 11.8), and a call whose written arrays are all resident and that returns nothing has a second signature, returning void, since nothing waits on it; before anything runs, the call checks each array against the indices a loop writes at a*i + c and refuses one too short with a TypeError naming the function, the parameter and both numbers.