https://www.jaelgareau.com/en/tags/simd/SIMDHugoBlox Kit (https://hugoblox.com)en-USWed, 06 May 2026 00:00:00 +0000https://www.jaelgareau.com/media/icon_hu_e954ec168906b2f9.pnghttps://www.jaelgareau.com/en/tags/simd/https://www.jaelgareau.com/en/publication/gareau_lemire-spe26/Wed, 06 May 2026 00:00:00 +0000https://www.jaelgareau.com/en/publication/gareau_lemire-spe26/https://www.jaelgareau.com/en/project/numeric/Wed, 06 May 2026 00:00:00 +0000https://www.jaelgareau.com/en/project/numeric/<p>Converting numbers to decimal strings is a fundamental yet often overlooked operation. It occurs whenever numerical data is displayed, logged, exported to text formats such as JSON or CSV, or exchanged between systems. Even when computations are performed in binary, we typically need to produce a readable, compact, and exact decimal representation.</p> <p>This project brings together work on two complementary families of conversions. The first concerns <strong>binary floating-point numbers</strong>, where the goal is to produce a short decimal string that recovers the exact original value when read back. We empirically compare modern algorithms such as Schubfach and Dragonbox against more classical approaches such as Dragon4 across several hardware architectures.</p> <p>The second concerns <strong>binary integers</strong>, where the main challenge is conversion speed. We study methods that exploit SIMD parallelism, especially AVX-512 IFMA instructions, to convert integers to decimal strings in under two nanoseconds in some scenarios. These techniques avoid large lookup tables and compute several digits in parallel.</p> <p>In both cases, the goal is to better understand the trade-offs between accuracy, the length of the produced strings, portability, and performance. This work aims to guide the optimization of numerical libraries, programming languages, and systems where number serialization is performance-critical.</p>