Parallel Algorithms Ranges and oneDPL

on YouTube ↗54m

" summary

Three proposals (P2300, P2500, P3179) add concrete machinery: senders/receivers with factory/adapter/consumer APIs, policy-aware scheduler binds "where" and "how", and parallel range algorithms that preserve serial semantics while requiring random‑access for parallel policies. Examples show piping a multi‑stage pipeline that moves between a CPU threadpool and CUDA streams, fusion of multiple range ops to cut overhead, and execute_on transitions host↔device. Practical backends include oneDPL and Thrust examples running on GPUs and accelerators and the patterns demonstrate working offload and synchronization.

AI-written summary · about

" chapters (10)

1. 0:00Parallel algorithms and oneDPL
2. 1:47C++ parallelism history
3. 6:27Motivation behind P2300
4. 18:30Senders and receivers in P2300
5. 25:36P2500 policy-aware schedulers
6. 33:24P2500 design and integration
7. 38:43P3179 parallel range algorithms
8. 46:47Thrust and oneDPL use cases
9. 52:12Challenges and future directions
10. 53:36Q&A

" related talks

• 52mC++ Standard ParallelismDavid Olsen · Meeting C++ 2022
• 01h 07mGeneric Parallel ProgrammingGuntram Berti · Meeting C++ 2014
• 10mGraph Based Pipeline ParallelismSanket Purandare, Simon Fan · PyTorch Conference Europe 2026
• 16mRevisiting Pipeline Parallelism for LLM ServingSoonjae Hwang, Jeongseob Ahn · 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26)
• 22mCustom Pipeline Parallelism for Frontier-Scale Model ServingAbin Shahab, Michael Lee · Ray Summit 2026
• 10mPyTorch Distributed Data Parallel (DDP)Pritam Damania · PyTorch Developer Day 2020
• 09mHyperParallel: A SuperPoD-Aware Distributed Acceleration LibraryTeng Su, Shendi Wang · KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026
• 44mFuture: Friendly Parallel and Distributed Processing in RHenrik Bengtsson · satRday Los Angeles 2019