simd-lib

Assembly math library for C, featuring SIMD (AVX2) optimized array operations.

Functions

Function	Description
add(a, b)	Add two integers
subtract(a, b)	Subtract two integers
multiply(a, b)	Multiply via repeated addition
divide(a, b)	Divide via repeated subtraction
add_arrays(a, b, result, len)	Add two arrays element-wise
simd_add_arrays(a, b, result, len)	Add two arrays using AVX2 SIMD (4 by 4)
subtract_arrays(a, b, result, len)	Subtract two arrays element-wise
simd_subtract_arrays(a, b, result, len)	Subtract two arrays using AVX2 SIMD (4 by 4)
dot_product(a, b, len)	Find dot product of given two arrays element-wise
simd_dot_product(a, b, len)	Find dot product of given two arrays using AVX2 SIMD

Benchmark Results

Tested on x86-64 Linux (AVX2), 100 million iterations:

Implementation	Time	Speed
Scalar add_arrays	0.183s	1x
SIMD simd_add_arrays	0.092s	1.98x

Array Size	Speedup	Bottleneck
8 elements	2.24x	CPU bound
10,000	1.49x	Cache bound
8,000,000	1.07x	Memory bound

Observations

SIMD is most effective when array size is a multiple of 4. AVX2 processes 4 x 64-bit integers per instruction. When the array size is not divisible by 4, the remainder is handled by a scalar cleanup loop, reducing overall speedup.

Array Size	Divisible by 4	Speedup
7 elements	No	1.18x
8 elements	Yes	2.24x

Build

make
./bin/main
./bin/bench

How it works

Functions are written in NASM assembly (src/math.asm) and linked with a C driver (src/main.c) using the System V AMD64 calling convention:

• Arguments passed in rdi, rsi, rdx, rcx
• Return value in rax
• SIMD uses 256-bit ymm registers to process 4 x 64-bit integers per instruction
• SIMD uses 128-bit xmm register to process 2 x 64-bit doubles per instruction
• Return value in xmm0