Costless Software Abstractions for Parallel Architectures

on YouTube ↗01h 22m

" summary

Claims a C++ generative/meta programming approach yields near zero overhead, using expression templates and Boost/Proto to generate skeletons bound to runtimes (OpenMP, TBB, SIMD, CUDA/MAGMA). Builds the NT2 numeric DSL, adds futures-based pipelining and loop fusion, and prototypes GPU and distributed backends. Benchmarks on Sigma‑Delta motion detection, Black‑Scholes and LU show competitive results (reference C ~16–18 cycles/point for the motion kernel, achieved similar performance and multix speedups up to about 7× on SIMD/backends), so the system works end to end.

AI-written summary · about

" chapters (13)

1. 0:00Start
2. 8:56Mapreduce
3. 10:48Generative Programming
4. 16:13Advantages
5. 25:15Matlab Syntax
6. 33:33Synchronous Pipelines
7. 36:50Examples
8. 37:48Sigma Delta Algorithm
9. 41:26Openmp
10. 44:27Apply Loop Fusion Operation
11. 58:44Prototype for a Single Sole Gpu Support
12. 1:08:19Julian Language
13. 1:13:02Compilers

" related talks

• 37mC++ for multi-accelerator, multi-vendor systemsGuy Tamir · Meeting C++ 2022
• 56mboost.tmp: your DSL for MetaprogrammingOdin Holmes · Meeting C++ 2018
• 01h 03mThe Landscape of ParallelismMichael S Wong · Meeting C++ 2015
• 01h 10mThe Three Little Dots and the Big Bad LambdasJoel Falcou · Meeting C++ 2017
• 01h 00mJust-in-Time CompilationJF Bastien · CppCon 2020
• 31mGPU Performance Portability Using Standard C++ with SYCLHugh Delaney, Rod Burns · CppCon 2022
• 01h 08mAn implementer's view of Boost.SIMDJoel Falcou · Meeting C++ 2016
• 55mC++ 20 Innovations: High-Performance Cross-Platform Architecture in C++Noah Stein · CppCon 2024