NSAlloc

Arena allocator with bulk-reset optimization

g++ -O3 -march=native -std=c++17 alloc_bench_final.cpp -o bench && ./bench

NSBVH

Bounding volume hierarchy for ray tracing

1.37x-1.44x Embree on clustered/mixed scenes. Win condition: structured/clustered geometry: the majority of production workloads.

g++ -O3 -march=native -std=c++17 nsbvh_bench.cpp -o bench && ./bench

NSCache

LRU cache with hot tier protection

g++ -O3 -march=native -std=c++17 bench_nscache.cpp -o bench && ./bench

NSComp

Schema-aware game data compression

g++ -O3 -march=native -std=c++17 nscomp_bench.cpp -o bench && ./bench

NSFFT

Sparse FFT with frequency recovery, comparing against FFTW3

N=1,048,576 (1M samples). 1000/1000 recovery rate at all K.

g++ -O3 -march=native -std=c++17 nsfft.cpp -o bench && ./bench

NSFix

FIX protocol parser with AVX2 SOH/EQ dual-scan

g++ -O3 -march=native -mavx2 -std=c++17 bench_nsfix.cpp -o bench && ./bench

NSGraph

Community-structured graph reordering for locality

g++ -O3 -march=native -std=c++17 nsgraph_bench.cpp -o bench && ./bench

NSHash

Open-addressing hash map optimized for structured keys

11.76x vs absl::flat_hash_map (timestamp lookup). Cold-run medians, 1M keys.

Sequential insert: 4.26x absl | 101x fph | 3.1x std, 142.0M ops/s.

g++ -O3 -march=native -std=c++17 bench.cpp -o bench && ./bench

NSIndex

Spatial indexing with hybrid prediction dispatch

g++ -O3 -march=native -std=c++17 nsindex_bench.cpp -o bench && ./bench

NSLock

Reader-writer lock with AVX2 SIMD writer scan

100M ops, pinned cores 0-7. Headline: 2.29x under pure-read load.

g++ -O3 -march=native -mavx2 -std=c++17 lock_bench.cpp -o bench && ./bench

NSMatrix

Block-diagonal matrix operations with AVX-512

3.3x Eigen at B=1,000 blocks. Win condition: block-diagonal structure declared by caller.

g++ -O3 -march=native -std=c++17 bench_nsmatrix.cpp -o bench && ./bench

NSMesh

3D mesh compression with parallelogram prediction

g++ -O3 -march=native -std=c++17 mesh_bench2.cpp -o bench && ./bench

NSOptimize

Hierarchical TSP solver for routing problems

g++ -O3 -march=native -std=c++17 optimize_bench.cpp -o bench && ./bench

NSPack

Transparent filesystem compression for mixed file types

nspack mount ~/assets ~/assets-store && nspack status ~/assets

Linux / macOS (FUSE3), Windows (WinFSP)

NSQCD

Analytic QCD from first principles: no lattice, no GPU, no free parameters

python3 nsqcd_cli.py --benchmark

NSQueue

Lock-free MPMC queue with batch operations

g++ -O3 -march=native -std=c++17 queue_bench.cpp -o bench && ./bench

NSSort

Distribution-aware sort with counting sort fast path

Dataset: n=100M elements, int64_t. 8/10 wins vs IPS4o. Losses on structureless data only.

g++ -O3 -march=native -fopenmp -std=c++17 NSSort.cpp bench_v10.cpp -o bench && ./bench

NSSort GPU

GPU-accelerated sort with two-level MSD radix

Dataset: n=100M elements, int64_t. 1570x on zero distribution vs cub::DeviceRadixSort (RTX 2000 Ada).

nvcc -O3 -arch=sm_70 -std=c++17 NSSort_GPU.cu -o bench && ./bench

NSStringIndex

String search index for amortized queries

g++ -O3 -march=native -std=c++17 nsstringindex_bench.cpp -o bench && ./bench

NSAttend

Sparse attention with fixed-window and variable-length sequence support

Fixed window size: 512

g++ -O3 -march=native -std=c++17 nsattend.cpp -o bench && ./bench

NSInfer

CPU inference engine with energy-ordered neuron selection

3.04× MLP throughput · TinyLlama-1.1B · standalone

72B real-world: +70% prompt eval with --ns-infer on Qwen2.5-72B (A40)

Cosine similarity: 0.9693

g++ -O3 -march=native -std=c++17 nsinfer.cpp -o bench && ./bench

NSKVCache

KV cache compression with cross-layer grouping and semantic deduplication

g++ -O3 -march=native -std=c++17 kv_box.cpp -o bench && ./bench

NSQuant

Variable-rate per-cluster quantization with repacking for GEMV inference

Qwen2.5-3B · 6.18 GiB F16 → 1.61 GiB · 3.8× · coherent

Qwen2.5-72B · 140 GiB F16 → 19.4 GiB · 7.2× · coherent

Profiling: Wanda single-pass importance, 128 calibration samples. 72B profiles in ~6 minutes.

./nsquant <input.gguf> <output.nsm> [n_prompts [prompts.txt]]