feat: Migrate codebase from x86 AVX2 to ARM64 NEON

[JoeStech]

Summary

This PR migrates the matrix operations codebase from x86-only (AVX2) to support ARM64 (NEON) architecture, enabling deployment on ARM-based infrastructure like AWS Graviton, Ampere Altra, Azure Cobalt, and Apple Silicon.

Changes Made

1. Docker Configuration

• Base Image: Replaced centos:6 (x86-only, EOL) with ubuntu:22.04 (multi-arch supported)
• Build System: Added TARGETARCH build argument for multi-architecture Docker builds
• Compiler Flags:
  • ARM64: -march=armv8-a+simd for NEON SIMD
  • x86-64: -mavx2 for AVX2 (unchanged)

2. Code Changes

matrix_operations.cpp

x86 AVX2	ARM NEON	Description
_mm256_setzero_pd()	vdupq_n_f64(0.0)	Zero vector initialization
_mm256_loadu_pd()	vld1q_f64()	Unaligned vector load
_mm256_mul_pd() + _mm256_add_pd()	vfmaq_f64()	Fused multiply-add (better perf)
_mm256_extractf128_pd() + horizontal add	vpaddd_f64()	Horizontal sum reduction

main.cpp

• Removed #error that blocked non-x86 compilation
• Added ARM64 architecture detection and messaging
• Added scalar fallback for other architectures

3. Architecture Support

The code now supports three compilation modes:

1. ARM64 (__aarch64__): NEON SIMD (128-bit, 2 doubles)
2. x86-64 (__x86_64__): AVX2 SIMD (256-bit, 4 doubles)
3. Generic: Scalar fallback

Performance Predictions

ARM64 (NEON) vs x86 (AVX2)

• SIMD Width: NEON processes 2 doubles vs AVX2's 4 doubles per instruction
• Compensating Factors:
  • ARM's vfmaq_f64 fused multiply-add reduces instruction count
  • Modern ARM cores (Graviton 3/4, M3) have excellent memory subsystems
  • Higher clock efficiency on ARM

Expected Performance

For 200x200 matrix multiplication:

• x86-64 (AVX2): Baseline performance
• ARM64 (NEON): ~80-120% of x86 performance depending on CPU
• Graviton 3/4: Likely comparable or faster due to FMA optimization

Cost Savings (AWS)

Migrating from x86 to ARM (Graviton) typically provides:

• ~20% cost reduction for equivalent performance
• Better price/performance for memory-bound workloads
• Lower power consumption per operation

Tools Used

Tool	Purpose
migrate-ease-scan (cpp)	Detected 10 x86 intrinsic issues requiring migration
skopeo	Verified ubuntu:22.04 ARM64 support, centos:6 incompatibility
knowledge_base_search	Found NEON intrinsic equivalents (vfmaq_f64, vpaddd_f64, etc.)

Validation Steps

1. Build Test (ARM64):

   docker buildx build --platform linux/arm64 -t benchmark:arm64 .

2. Build Test (x86-64):

   docker buildx build --platform linux/amd64 -t benchmark:amd64 .

3. Multi-arch Build:

   docker buildx build --platform linux/amd64,linux/arm64 -t benchmark:multi .

4. Runtime Verification:

   docker run --rm benchmark:arm64
   # Should output: "Running on ARM64 architecture with NEON optimizations"

Migration Scan Results

Issues Found: 10 total
- IncompatibleHeaderFileIssue: 1 (immintrin.h)
- PreprocessorErrorIssue: 1 (#error directive)
- IntrinsicIssue: 8 (AVX2 intrinsics)

All issues resolved ✅

Breaking Changes

None - the code maintains full backward compatibility with x86-64 systems.

Future Enhancements

• Consider SVE/SVE2 for Graviton 3+ (variable-length vectors)
• Add runtime CPU feature detection
• Benchmark comparison between architectures