ecs: struct components in GPU accessors, packed u64/u128 fields, and the cost of staged archetype moves

Aberta
  • Algo errado?
#271 0 comentários 0 reações 0 responsáveis Ver no GitHub

Mantenedores costumam responder em até 1 dia

Ninguém assumiu esta issue ainda.

Avaliação

Dificuldade
5/5
Tempo estimado
Mais de uma semana
Facilidade para iniciantes
38/100
Tipo de issue
Funcionalidade
Clareza
Razoavelmente clara
Status de atividade
Ativa
Stack de tecnologia
typescript, wasm
Domínio
backend, performance

Direção de pesquisa

Start with src/interface/gpu/accessors.ts and layoutOf to see how struct layouts are handled, then trace row and kernel access to component values and their change stamping. Check the existing benchmark and profiling entry points, including src/perftest/birth.ts, for the packed-field and command-encoding costs. Done means documenting current u64/u128 support and lane change behavior, with measurements that compare packed and separate fields and define the requested accessor support.

Escrita pelo modelo de indexação a partir do texto da issue.

Descrição

Found while consolidating a voxel renderer's per-entity data (the mecs app, @adobe/ecs 0.3.2). Two things to verify or add, and the measurements that motivate them.

Why

Each voxel carries denormalized, derived values as separate components: a look (color 4×f32, metallic/roughness 2×f32, emission 3×f32, fade f32, refraction 2×f32) and a shade (exposure u32, occlusion 2×u32, sunlight u32). They are written by different systems at different times, so a voxel is born in four stages (spawn, then +cell, +look, +shade), each a move to another archetype, and each wakes the changes finalizers that watch those components. They are read only by render systems.

Measured with real components, 14,000 voxels, no finalizers or indexes (src/perftest/birth.ts in mecs):

birth per voxel
spawn, then three add stages ~1.7 µs
spawn with everything at once ~0.95 µs

So the three extra moves cost about 0.75 µs a voxel before any finalizer runs. Packing each group into one component removes two of the moves and 40 to 48 bytes of columns a voxel.

1. GPU accessors for struct components

accessorsOf (src/interface/gpu/accessors.ts) accepts only a scalar or a tuple of 2 to 4 same-typed scalars. A struct component, or a mixed tuple, is rejected ("isn't a scalar or a tuple of two to four scalars of one type"), although layoutOf already places structs for the storage rules. We can pack by hand into Vec4.U32 and unpack in WGSL, but the fields then have no names or types on either side.

Request: let an accessor name a struct component, either as a WGSL struct return or as one accessor per field (fn name_field(row) -> T), reading only the words of the fields used.

2. Packed bit fields: u64 and u128, and their cost

For the shade group (exposure 6 bits, occlusion 48, sunlight 30: 84 bits) and the look group, the natural representation is a few adjacent u32 words with bit fields. Please check, and state what holds today:

  1. Whether a U64 scalar (or a documented pair/quad of adjacent u32 lanes treated as a U128) can be declared, and how it reads and writes without BigInt in a row system, in a kernel: true system, and in the wasm kernel.
  2. The cost of a bit-field read-modify-write of one field of a 3×u32 component in a kernel row system, against the same field as its own U32 component. If the packed form is not about as fast, where does it lose (extra loads, a write-back of all lanes, change stamping of the whole component)?
  3. Whether a write to one lane of a tuple component stamps changed for the whole component (it seems to), and what that means for a GPU upload of a chunk when only one field changed.

3. Aside: command encoding

In a profile of a 14,000-voxel fall (add fallPosition, spawn voxels, despawn fallers) encodeValue/leafValue/element, the encoding of each command's JS value object into the command buffer, was the largest single line in the ECS's own time (~15 ms of ~176). A bulk path (spawnMany(archetype, count, fill), or per-column writes from a row system for components the entity already has) would avoid the per-entity object. Related: #259.

Happy to share the benchmarks and profiles.

Linguagem predominante
TypeScript
Estrelas
12
Forks
8
Merge médio
9h 2min
PRs com merge (30d)
25

Preparar o ambiente

Primeiros passos

  1. Leia a issue inteira e depois o guia de contribuição do projeto.
  2. Comente na issue dizendo que vai assumir — evita que duas pessoas façam o mesmo trabalho.
  3. Faça um fork do repositório e trabalhe em uma branch.
  4. Abra um pull request que referencie o número da issue.

Mais de adobe/data

Todas as issues de adobe/data

Issues semelhantes

Mais issues de TypeScript

Receba novas issues na sua caixa de entrada

Um resumo curto de issues do GitHub para quem está começando.