Expand description
Graph acceleration: GPU-first integer kernels over NQuins / u64 keys, with a CPU radix / scan / sort-merge floor that always works.
Chain: GPU (shared wgpu) → CPU. NPU is a reserved AccelPath slot;
platform::npu_ffi today is wgpu (sieve.wgsl), not a third backend.
Override: QUALIA_GRAPH_ACCEL=cpu|gpu|auto (default auto).
Structs§
- GpuQuin
Sieve - GPU warp-transposed
uvec4equality sieve over 48-byteNQuinWirebuffers. - GpuRadix
Sorter - Join
Outcome - NQuin
Wire - 48-byte Super-Quin wire layout. Six 64-bit integer fields aligned to 16 bytes.
- Sieve
Outcome - Sort
Outcome - Result of a sort: which path ran. Items are mutated in place.
Enums§
- Accel
Path - Best-path that produced a result. Callers never fail because GPU is missing.
- Accel
Policy - How the dispatcher should choose.
QUALIA_GRAPH_ACCEL=cpu|gpu|autooverrides. - Quin
Field
Constants§
- GPU_
JOIN_ MIN - GPU_
SIEVE_ MIN - GPU_
SORT_ MIN - Below this many records, PCIe + dispatch lose to the CPU radix floor.
Functions§
- gpu_
available - Prefer GPU when a shared device exists and policy allows it.
- hash_
join_ u64 - Join
probetobuildon equal keys. Writes(probe_idx, build_idx). - hash_
join_ u64_ cpu - CPU-only oracle (tests).
- npu_
available - Distinct NPU integer backend (DirectML/Hexagon/CoreML as compute, not wgpu).
Today’s
platform::npu_ffidispatchessieve.wgslon wgpu — that isAccelPath::Gpu. - sieve_
eq - Compact Quins whose
field == needleintoout. Returns count + path. - sieve_
eq_ indices - Indices only (caller gathers). Same fallback contract.
- sieve_
volume_ file - Read all Quins from
path(single-file volume) intoscratch, then compact matches intoout. Fails closed if the volume is larger thanscratch. - sort_
quins_ by_ object - Sort
quinsbyobject(ingest run order). Never fails. - sort_
quins_ by_ object_ cpu_ only - CPU-only radix (tests / oracles).
- sort_
u64_ indices - Sort
indicessokeys[indices[i]]is non-decreasing.
Type Aliases§
- GpuN
Quin Sieve - Alias matching Runbook 05 naming (
GpuNQuinSieve).