oximedia-simd

Hand-written assembly SIMD kernels and high-performance pixel operations for OxiMedia.

Part of the oximedia workspace — a comprehensive pure-Rust media processing framework.

Version: 0.1.1 — 2026-03-10

Overview

This crate provides highly optimized SIMD implementations of critical performance paths in the OxiMedia video codec, including:

DCT Transforms: 4x4, 8x8, 16x16, 32x32 forward and inverse DCT
Interpolation: Bilinear, bicubic, and 8-tap filters for motion compensation
Motion Estimation: Sum of Absolute Differences (SAD) for 16x16, 32x32, 64x64 blocks
Pixel Operations: Blending, color conversion, YUV operations, histogram
Audio Operations: SIMD-accelerated audio processing primitives
Math Operations: Vector math, dot products, matrix operations

Performance

These optimized implementations provide 2-5x speedup over compiler-generated code:

AVX2 (x86-64): ~3-4x faster than auto-vectorization
AVX-512 (x86-64): ~4-5x faster for SAD operations
NEON (AArch64): ~2-3x faster than auto-vectorization

Architecture Support

x86-64

AVX2: Full support for all operations
AVX-512: Optimized SAD implementations using 512-bit registers

AArch64

NEON: Full support for DCT, interpolation, and SAD

Fallback

Scalar implementations for all operations when SIMD not available

Usage

Add to your Cargo.toml:

[dependencies]
oximedia-simd = "0.1.1"

use oximedia_simd::{detect_cpu_features, forward_dct, DctSize};

let features = detect_cpu_features();
println!("AVX2 available: {}", features.avx2);
println!("NEON available: {}", features.neon);

// API automatically selects best implementation
let input = vec![0i16; 64];
let mut output = vec![0i16; 64];
forward_dct(&input, &mut output, DctSize::Dct8x8).unwrap();

API Overview

detect_cpu_features() — Runtime CPU feature detection
CpuFeatures — Detected feature flags: avx2, avx512f, avx512bw, neon
forward_dct() / inverse_dct() — DCT transforms with automatic SIMD dispatch
interpolate() — Motion compensation interpolation
sad() — Sum of Absolute Differences for motion estimation
DctSize — Dct4x4, Dct8x8, Dct16x16, Dct32x32
InterpolationFilter — Bilinear, Bicubic, EightTap
BlockSize — Block16x16, Block32x32, Block64x64
SimdError / Result — Error and result types
validate_avx2_alignment() / validate_avx512_alignment() / validate_neon_alignment() — Buffer alignment validation
Modules: accumulator, alpha_premul, audio_ops, bitwise_ops, blend, blend_simd, color_convert_simd, color_space, convolution, filter, fixed_point, gather_scatter, histogram, interleave, lookup_table, math_ops, matrix, min_max, pack_unpack, pixel_ops, prefix_sum, reduce, saturate, threshold, transpose, vector_math, yuv_ops

oximedia-simd 0.1.2