Fusing the band column's kernels and coalescing the split's transposes on devices are unrecorded and untried

Aperta
  • Qualcosa non va?
#352 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
5/5
Tempo stimato
Più di una settimana
Idoneità per principianti
8/100
Tipo di issue
Funzionalità
Chiarezza
Da chiarire
Stato di attività
Attiva
Stack tecnologico
cpp, python
Ambito
devtools, performance

Direzione di ricerca

This is a performance investigation requiring physical GPU hardware (NVIDIA A40 and a data-centre GPU) and CUDA/HIP builds that CI does not cover. Start from the band column kernels and the split-step transposes referenced in docs/dev/benchmarks.md, then run python -m benchmarks ab --device on the named cases. Done means ratios of 0.97 or better are recorded or the changes are marked not adopted in benchmarks.md.

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

Descrizione

area: gpu area: native performance priority: low
What is missing

#59 ended with two device options it did not try, which its pull request says were left in the benchmark ledger. The ledger holds only the measurements around them, and no issue or roadmap entry has them:

  • Fusing the band column's kernels. On an NVIDIA A40 the band plasma column's four tiled kernels take 235 µs of an rdw trial. Neither half-tile teams (0.985 on rdw, 1.002 on he_rdw_averaged) nor 512-sample tiles (1.010, 1.019) moved them, and the card gave no performance counters, so what bounds them is unknown.
  • Coalescing the split's transposes. On kerr_beam_384 earlier work took about 100 ms a step out of the split's transverse pair and transposes, within a fall in kernel time from 827 to 651 ms. Coalescing them has not been measured.

"The split" here is the free-space split-step collar on a device.

Why it matters

GPU runs are launch-bound: on these cases each kernel launch costs more than its arithmetic.

Context
  • #103 is other device work (per-trial mirrors, the accept pass, plasma buffers). The roadmap's launch and copy overhead is the wider picture, on GPUs with full-rate binary64.
  • Not to repeat: band column tilings, which #59 measured and did not adopt.

Read, not run.

To do
  • Prototype fusing the band column's kernels on a device, and compare on rdw, ctc and argon_blueshift.
  • Measure coalescing the split's transposes on Cartesian device runs.

Device tests must pass on a CUDA and a HIP build, which CI does not build.

Done when

python -m benchmarks ab --device ratios on an NVIDIA A40 and one data-centre GPU (H200 or MI325X) are recorded for each. Each change lands at 0.97 or better on its cases with no loss elsewhere, or benchmarks.md records it as measured and not adopted.

Checked against commit 8a9f764.

Lingua principale
Python
Stelle
3
Fork
3
Merge medio
4h 42m
PR unite (30g)
91

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 LupoLab/solprop

Tutte le issue di LupoLab/solprop

Issue simili

Altre issue su Python

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.