Write Fast Code Like a Native

on YouTube ↗56m

" summary

Walks through microbenchmarks using std::chrono, rdtsc/rdtscp and RDPMC, plus Godbolt and cycle counters, to build a latency intuition. Shows a naive matrix transpose costing ~250 ns (720 cycles) for tiny sizes and scaling to thousands of times slower when memory spills caches; a summation that the optimizer turns into a closed form (with O2 off it matches expected ~2,000,000 ns); XGBoost prediction costs ~60 assembly instructions per shallow tree (~0.1 ns per tree evaluation), and IO stays in millisecond land (disk spin ~8 ms, 10 Gb/s ≈ 1 s/GB). Advises measure with RDPMC, avoid syscalls on the fast path, use templates for zero-overhead abstractions.

AI-written summary · about

" chapters (11)

1. 0:00Introduction and background
2. 3:48Why write fast code
3. 7:01Understanding CPU latencies
4. 11:20Profiling tools and techniques
5. 15:11Case study: Matrix transpose
6. 22:42Case study: Summing numbers
7. 27:42I/O and real-world operations
8. 28:02Case study: XGBoost performance
9. 33:19Code constructs and performance
10. 39:58Philosophy of optimization
11. 48:52Q&A session

" related talks

• 01h 00m“When a Microsecond Is an Eternity: High Performance Trading Systems in C++”Carl Cook · CppCon 2017
• 01h 00mProgramming with less effort in C++...Sylvain Jubertie · CppCon 2015
• 49mWriting better code with help from Qt and the compilerThiago Macieira · Qt World Summit 2015
• 01h 02mFastwareAndrei Alexandrescu · ACCU 2016
• 40mLessons from two decades of writing bad codeJames Mitchell · GCAP 2025
• 01h 00mFastwareAndrei Alexandrescu · NDC London 2017
• 44mFastwareAndrei Alexandrescu · NDC TechTown 2017
• 23mBeating the Benchmarks: Co-evolving the ISA and Development ToolsPhilipp Tomsich · RISC-V Summit 2022