СЪДЪРЖАНИЕ
- изкуствен интелект
Г. Петров, Битови манипулации и хардуерни оптимизации за бързи трансцендентни функции в Edge и IoT устройства - Н. Делийски, Л. Дзуренда, Д. Ангелски, П. Вичев, К. Атанасова, Изчисляване на топлинния баланс и КПД на бетонни ями при пропарване на фурнирни трупи в тях
- П. Благов, П. Русков, Европейските проекти EBSI-NE (Node Expansion) и OnePass: рамка за доверие между стартиращи компании, инвеститори и оператори на услуги
Key Words: Bit manipulations; hardware optimizations; Edge devices; lookup tables (LUT); CORDIC; fast inverse square root; embedded machine learning.
Abstract. Fast and energy-efficient evaluation of transcendental functions (for example, logarithm, exponential, square root, and normalization) is critical for real-time signal processing and on-device AI inference in Edge and IoT systems. This paper surveys and experimentally validates a set of lightweight techniques – primarily IEEE-754 based bit-level manipulations (bit tricks) – alongside code- and hardware-level optimizations (loop unrolling, DMA usage, and other low-level optimizations) that dramatically accelerate these functions on resource-constrained microcontrollers. We describe simple but effective bit hacks for fast approximations of log₂(x), 2ˣ, √x and 1/√x (including the well-known “Quake” inverse square root), explain their theoretical basis in floating-point representation, and show how modest corrections to the mantissa (linear and low-order polynomial) can substantially improve accuracy while retaining most performance gains. Benchmarking across representative platforms (AVR, Cortex-M variants with/without FPU, ESP32, and RISC-V boards) demonstrates up to 10× speedups relative to standard library implementations, with typical accuracy in the range of 5-6 bits for raw bit-trick estimates and much improved error after correction terms. We discuss practical integration patterns, trade-offs between latency, memory footprint, and numerical fidelity, and recommend when to combine bit tricks with LUTs or CORDIC for broader function coverage. Finally, we outline use cases (embedded ML feature scaling, entropy computation, normalization in DSP and robotics) and caveats (unsuitability for high-precision scientific/financial calculations and for periodic functions like sin/cos), providing guidelines for safe deployment in production Edge/IoT projects.
