ARM Matrix Multiplication (NEON Edition)

Описание и применение

Руководство по реализации умножения матриц на ARM с использованием NEON‑ассемблера, показывающее, как вручную оптимизировать код и измерять производительность.

Задача

Implement fast matrix multiplication on ARM CPUs using NEON assembly instead of relying on GPU or high-level libraries.

Результат

Hand‑written NEON assembly slightly outperforms the C compiler but remains a small fraction of the M2’s theoretical peak. SIMD vectorization yields modest gains, highlighting the need for more advanced optimizations (e.g., SME/SVE2).

Рабочий процесс

• Write reference C matrix multiplication
• Compile with -O2 and benchmark
• Translate naive C loops to ARM assembly
• Benchmark assembly version
• Add SIMD vectorization (load 4 B values, broadcast A, use fmadd)
• Benchmark SIMD version

Доказательства

At -O2 this runs at 2.09 Gflops, about 1.9% of our theoretical maximum.

Источник

alright we have 1.98 Gflops (or about 1.8% of the theoretical max)... the c compiler is smarter than me.

Источник