GPU numerical accuracy: compensated sums, norm-aware results, optional f64

Aperta
  • Qualcosa non va?
#13 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
4/5
Tempo stimato
3-5 giorni
Idoneità per principianti
52/100
Tipo di issue
Bug
Chiarezza
Abbastanza chiara
Stato di attività
Attiva
Stack tecnologico
rust
Ambito
backend, performance

Direzione di ricerca

Start with qubit_probability.wgsl, src/backend.rs around measure_qubit, and src/pauli.rs to trace probability and expectation calculations. Check how shaders are compiled and how the GPU adapter requests features before assessing the optional f64 path. Done means branch renormalisation uses each branch's measured norm, expectation values account for state norm, and f64 remains optional on supported adapters.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

enhancement
What you want to do

The GPU backend works in f32. Three things make its error bigger than it needs to be:

  1. qubit_probability.wgsl adds up to about 32 thousand terms per thread one after another in f32
    (at 30 qubits), and measure_qubit (src/backend.rs:153) renormalises by 1/√p from that sum.
    Errors add up over many mid-circuit measurements in one shot.
  2. The outcome-0 branch is scaled by 1/√(1 − p1), which is only right if the state's norm is
    exactly 1.
  3. Pauli::expectation (src/pauli.rs) assumes ‖ψ‖ = 1 and does not divide by the norm, so norm
    drift becomes a bias in the expectation value.
Implementation notes
  • Use Kahan (compensated) summation in each thread's loop in qubit_probability.wgsl. Return both
    the probability of 1 and the total norm from the same pass, and renormalise each branch by its
    own sum.
  • Divide expectation values by Σ|a|², which is a cheap extra sum in the same loop.
  • f64: request wgpu::Features::SHADER_F64 when the adapter offers it and compile a second
    shader. Most consumer GPUs run f64 at 1/16–1/64 of f32 speed, so it should only be opt-in.
What it could look like
No API change for 1–3. Optionally `backend="gpu64"` on adapters that support `SHADER_F64`.
Contributing
  • I would like to open a pull request for this (see CONTRIBUTING.md).
Lingua principale
Rust
Stelle
1
Fork
0
Merge medio
12m
PR unite (30g)
4

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di kaiwave/qsengine

Tutte le issue di kaiwave/qsengine

Issue simili

Altre issue su Rust

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.