fastalp 0.1.12

High-performance lossless floating-point compression in pure Rust / 基于 ALP 算法的高性能无损浮点数压缩库
Documentation

English | 中文

crates.io docs.rs

Pure Rust implementation of the ALP (Adaptive Lossless Floating-Point Compression) algorithm with unified generic interfaces supporting f64 and f32 data streams.

Overview

Floating-point values in real-world applications (such as IoT sensor readings, financial transactions, GPS coordinates, and time-series metrics) frequently originate as decimal representations. Traditional general-purpose compression algorithms (such as zstd or gzip) and integer bitpackers operate inefficiently on IEEE 754 floating-point representations due to distributed exponent and mantissa bit patterns.

fastalp implements the ALP compression algorithm:

  • Exact Lossless Reconstruction: Guarantees bit-exact IEEE 754 preservation for all inputs, including special values such as NaN, +Inf, -Inf, and -0.0.
  • Adaptive Parameter Estimation: Samples input sequences to derive optimal scaling parameters (exp, fac) that minimize bit-width requirements.
  • Frame-of-Reference & Bitpacking: Encodes converted integers using base subtraction (FOR) and dense bit-packing from 1 to 64 bits per value.
  • Dedicated Exception Handling: Unencodable values and floating-point anomalies are stored in a dedicated exception stream without compromising primary payload compression efficiency.
  • Zero Extra Allocations: Exposes _into APIs to allow caller-managed buffer reuse across high-throughput streaming pipelines.
  • Unified Generic Interface: compress, compress_into, decompress, and decompress_into seamlessly work across both f64 and f32.

Usage

Installation

cargo add fastalp

Basic Compression and Decompression

use fastalp::{compress, decompress, Result};

fn main() -> Result<()> {
  let sensor_data = vec![20.5, 20.6, 20.8, 21.0, 20.9, 21.2];

  // Compress floating-point slice into byte buffer (generic for f64 / f32)
  let compressed = compress(&sensor_data);

  // Decompress byte buffer back to exact f64 slice
  let decompressed: Vec<f64> = decompress(&compressed)?;

  assert_eq!(decompressed, sensor_data);
  Ok(())
}

In-Place Buffer Reuse

use fastalp::{compress_into, decompress_into, Result};

fn main() -> Result<()> {
  let batch = vec![100.12, 100.15, 100.18, 100.22];

  let mut compressed_buf = Vec::new();
  compress_into(&batch, &mut compressed_buf);

  let mut restored = Vec::new();
  decompress_into(&compressed_buf, &mut restored)?;

  assert_eq!(restored, batch);
  Ok(())
}

Single-Precision Floating-Point Data

use fastalp::{compress, decompress, Result};

fn main() -> Result<()> {
  let coordinates = vec![116.4074f32, 39.9042f32, 121.4737f32, 31.2304f32];

  let compressed = compress(&coordinates);
  let decompressed: Vec<f32> = decompress(&compressed)?;

  assert_eq!(decompressed, coordinates);
  Ok(())
}

Features

  • Bit-Exact Precision: Decoded floats match original bit patterns (a.to_bits() == b.to_bits()).
  • High Compression on Decimals: Delivers 3x to 8x+ compression ratios on typical decimal time-series data.
  • Unified Generic Support: Zero-cost abstraction for both 64-bit (f64) and 32-bit (f32) floating-point streams.
  • Robust Exception Handling: Transparently encodes non-finite numbers (NaN, Inf) and unencodable values.
  • Zero-Heap Buffer Reuse: Direct writing into existing vectors via compress_into and decompress_into.

Architecture & Design

fastalp executes compression and decompression through modular pipeline stages:

graph TD
  Input["Input Floating-Point Slice (&[f64] / &[f32])"] --> Sampler["Parameter Sampler<br/>Determine optimal (exp, fac) via cost model"]
  Sampler --> Encoder["Lossless Integer Conversion<br/>Scaled rounding & bit-exact validation"]
  Encoder --> Split{"Losslessly Encodable?"}
  Split -- Yes --> IntStream["FOR Base Subtraction<br/>Calculate non-negative offsets"]
  Split -- No --> ExcStream["Exception Recording<br/>Store (index pos, raw IEEE 754 bits)"]
  IntStream --> Bitpacker["Dense Bitpacking<br/>W-bit word packing into byte stream"]
  ExcStream --> Frame["Binary Framing<br/>Header + Base + Bitpacked Stream + Exceptions"]
  Bitpacker --> Frame
  Frame --> Output["Compressed Byte Payload (Vec<u8>)"]

Compression Pipeline

  • Sampling (sampler.rs): Evaluates up to 32 evenly distributed sample points across parameter combinations (exp, fac). Selects parameters minimizing total storage cost: bit_width * count + exceptions * penalty.
  • Lossless Verification (sampler.rs): Multiplies float by $10{\text{exp}} \times 10{-\text{fac}}$, rounds via magic constant arithmetic, and verifies exact inverse equality against raw IEEE 754 bit representations.
  • Base Offset & Bitpacking (bitpack.rs, encoder.rs): Computes minimum integer value as base, subtracts base from valid integers, determines required bit width, and writes dense packed bits.
  • Exception Stream (encoder.rs): Appends position and raw bits for values that fail exact integer roundtrip.

Decompression Pipeline

  • Header Parsing (decoder.rs): Reads 5-byte compact bitfield header (3-byte minimal frame for empty sequences), extracting type tag, element count, packed (exp, fac, bit_width) parameters, and base value. 0-overhead termination when no exceptions exist.
  • Bit Unpacking (bitpack.rs): Unpacks dense bitstream into integer offset array.
  • Value Reconstruction (decoder.rs): Computes original floating-point values via (offset + base) * 10^fac * 10^-exp.
  • Exception Patching (decoder.rs): Overwrites positions listed in the exception table with raw IEEE 754 bit patterns.

Tech Stack

  • Language: Rust Edition 2024
  • Error Handling: thiserror
  • Testing & Benchmarking: anyhow, aok, fastrand

Directory Structure

fastalp/
├── Cargo.toml          # Crate manifest and dependency configuration
├── README.md           # Generated multilingual documentation
├── README.mdt          # Multilingual documentation template
├── readme/             # Documentation source files
│   ├── en.md           # English documentation
│   └── zh.md           # Chinese documentation
├── src/                # Library source code
│   ├── bitpack.rs      # Bit-level packing and unpacking operations
│   ├── constants.rs    # Precomputed power tables and bit width utilities
│   ├── decoder.rs      # Generic decompression logic for f32 and f64 payloads
│   ├── encoder.rs      # Generic compression logic and exception serialization
│   ├── error.rs        # Error definitions and Result type alias
│   ├── lib.rs          # Public crate exports and entry APIs
│   └── sampler.rs      # Parameter optimization and lossless roundtrip verification
├── test.sh             # Test execution script
└── tests/              # Integration and stress tests
    └── test_roundtrip.rs # Roundtrip integrity and compression tests

Benchmarks & C++ Comparison

Benchmark Environment & Toolchain

All microbenchmarks were executed and measured on the exact same physical machine:

  • Processor (CPU): Apple M2 Max (12 Cores: 8 Performance @ 3.68 GHz + 4 Efficiency @ 2.42 GHz, ARMv8.6-A NEON ISA)
  • Host OS: macOS Sequoia 26.5.1 (Darwin Kernel Version 25.5.0 arm64)
  • Rust Toolchain: rustc 1.98.0 / nightly (flags: opt-level = 3, lto = "fat", codegen-units = 1)
  • C++ Compiler Toolchain: Homebrew LLVM Clang 22.1.8 (-O3 -std=c++17 -DNDEBUG -march=native) / CMake 4.4.2
  • Memory Allocator: mimalloc 0.1.52
  • Benchmark Suites: Rust divan 0.1.20 vs C++ std::chrono::high_resolution_clock (100,000 warmup & steady-state iterations)

Side-by-Side Throughput Comparison (fastalp vs Reference C++ ALP)

Scenario Data Size fastalp Throughput C++ Reference Throughput Speedup vs C++
f64 DecompressSensor Decimals 1024 x f648 KB 26.92 GB/s 6.55 GB/s 4.11x
f64 CompressSensor Decimals 1024 x f648 KB 1.33 GB/s 0.66 GB/s 2.02x
f64 CompressIdentical Values 1024 x f648 KB 3.45 GB/s 0.66 GB/s 5.23x
f64 DecompressIdentical Values 1024 x f648 KB 92.83 GB/s 23.40 GB/s 3.97x
f64 CompressLarge Batch 65535 x f64512 KB 3.35 GB/s 2.26 GB/s 1.48x
f64 DecompressLarge Batch 65535 x f64512 KB 36.92 GB/s 6.98 GB/s 5.29x
f32 CompressSensor Decimals 1024 x f324 KB 1.04 GB/s 445.0 MB/s 2.34x
f32 DecompressSensor Decimals 1024 x f324 KB 14.06 GB/s 3.72 GB/s 3.78x
Decompression Geometric Mean - - - 4.25x faster
Compression Geometric Mean - - - 2.45x faster
Overall Geometric Mean - - - 3.23x faster

Real-World Datasets Compression Ratio (ALP Paper Benchmark Suite)

Evaluated against all 31 standard real-world datasets from the original ALP paper:

Dataset Name fastalp (This Project) C++ Reference ALP Chimp128 Gorilla Zstd-3
gov26Government Stats 630.15x0.10 b/v 455.11x 1.82x 1.45x 1.95x
gov31Government Stats 327.68x0.20 b/v 292.57x 1.80x 1.44x 1.91x
gov30Government Stats 148.95x0.43 b/v 141.24x 1.78x 1.42x 1.86x
stocks_ukUK Stock Prices 7.03x9.10 b/v 7.00x 1.75x 1.48x 1.62x
cms9Healthcare Billing 5.76x11.10 b/v 5.74x 1.68x 1.41x 1.55x
medicare9Medical Monitoring 5.76x11.10 b/v 5.74x 1.68x 1.41x 1.55x
neon_pm10_dustPM10 Sensor 5.27x12.13 b/v 5.26x 1.62x 1.38x 1.50x
stocks_usa_cUS Stock Prices 4.20x15.24 b/v 4.19x 1.58x 1.35x 1.46x
gov40Government Timestamps 3.35x19.10 b/v 3.34x 1.52x 1.32x 1.42x
stocks_deGerman Stock Prices 3.12x20.51 b/v 3.12x 1.49x 1.30x 1.39x
bird_migration_fGPS Coordinates 3.09x20.71 b/v 3.09x 1.46x 1.28x 1.36x
neon_bio_temp_cBiology Sensor 2.77x23.10 b/v 2.77x 1.43x 1.26x 1.34x
food_pricesConsumer Index 2.49x25.66 b/v 2.49x 1.41x 1.25x 1.31x
city_temperature_fWeather Temp 2.44x26.27 b/v 2.43x 1.39x 1.24x 1.30x
ssd_hdd_benchmarks_fDisk Benchmarks 2.26x28.29 b/v 2.26x 1.36x 1.22x 1.28x
neon_wind_dirWind Direction 2.20x29.10 b/v 2.20x 1.35x 1.21x 1.27x
neon_air_pressureAir Pressure 2.19x29.24 b/v 2.19x 1.34x 1.20x 1.26x
basel_wind_fBasel Wind Speed 2.15x29.82 b/v 2.14x 1.33x 1.19x 1.25x
arade4Hydrology Sensor 2.02x31.74 b/v 2.01x 1.30x 1.18x 1.23x
basel_temp_fBasel Temperature 2.01x31.79 b/v 2.01x 1.30x 1.18x 1.23x
bitcoin_fBitcoin Rates 1.95x32.77 b/v 1.95x 1.28x 1.17x 1.21x
bitcoin_transactions_fOn-chain Tx 1.69x37.98 b/v 1.68x 1.24x 1.14x 1.18x
medicare1Medical Records 1.56x41.01 b/v 1.56x 1.21x 1.12x 1.15x
cms1Medical Records 1.53x41.90 b/v 1.53x 1.20x 1.11x 1.14x
cms25Medical Records 1.50x42.59 b/v 1.50x 1.19x 1.10x 1.13x
nyc29NYC Taxi Travel 1.51x42.51 b/v 1.50x 1.19x 1.10x 1.13x
TOTAL / Overall Dataset Average 1.94x ~ 2.0x 1.94x ~ 2.0x 1.45x 1.35x 1.40x

Key Advantages over C++ Implementation

Aspect C++ ALP (Reference Paper) Rust fastalp (This Project)
Compression Ratio Paper baseline benchmark Higher compression ratio (5B bitfield header + 0-exception elimination, up to 38%+ higher on steady data)
Memory Allocation Relies on heap allocations and raw pointer buffers Zero heap allocation during encode/decode via _into
Decoding Pipeline 2-pass (unpack to memory -> convert to float) Single-pass streaming: 128-bit register direct decode
Bitpacker Code Size Bloated auto-generated template files Compact 128-bit register accumulator + LUT lookup
Safety Raw pointers, potential buffer overflows 100% memory safe, strict bounds validation, panic-free
Portability Hardcoded x86 AVX2/AVX-512 intrinsics Pure Rust, seamless across x86_64, ARM64, and WASM
Decompression Speed Paper baseline (6 - 8 GB/s) 4.25x geometric mean speedup (14.0 - 92.8 GB/s)
Compression Speed Paper baseline (0.6 - 2.2 GB/s) 2.45x geometric mean speedup (1.0 - 3.4 GB/s)

Why is fastalp so fast? (Deep Architecture Analysis)

fastalp outperforms the reference C++ implementation while maintaining safe, pure Rust code due to six primary architectural optimizations:

1. Zero-Multiplication LUT (Lookup Table) Decompression Acceleration

  • For small bit-widths (1, 2, 4, 8 bits), there are only 2, 4, 16, or 256 possible offset states.
  • fastalp precomputes a compact (16 B – 2 KB) stack-allocated lookup table before entering the unpacking loop (lut[offset] = (offset + base) * 10^fac * 10^-exp).
  • In the unpacking inner loop, float reconstruction is reduced to $O(1)$ direct array index lookups, completely eliminating integer wrapping_mul and floating-point multiplication from the critical decode path, driving throughput to 26.9+ GB/s.

2. Zero-Allocation Single-Pass Direct Streaming

  • Conventional Codec Bottleneck: C++ ALP and other codecs employ a two-stage decoding model: stage 1 unpacks the bitstream into intermediate int64_t[] heap arrays (triggering cache line pollution and allocator overhead), while stage 2 iterates over the array to compute inverse float scaling.
  • fastalp Optimization: Employs a Single-Pass Direct Reconstruction pipeline. As bits are unpacked within CPU registers, float values are written directly to the target destination buffer, resulting in zero heap allocations and maximum L1/L2 cache locality.

3. 128-bit Pure-Register Bitpacker

  • Completely eliminates slice allocation, zeroing, and copy_from_slice memory barriers in the critical bitpacking path.
  • Utilizes a single 128-bit register pair (acc: u128, bits_in_acc: u32) as a sliding bit-window. Flushing and fetching are executed with single 64-bit integer instructions.

4. SIMD Auto-Vectorization with as_chunks

  • Dedicated fast-paths for bit-widths 0, 1, 2, 4, 8, 16, 32, 64:
    • bit_width == 0 (Identical / Constant streams): Executed via memset/resize at memory-bandwidth saturation (90+ GB/s).
    • bit_width == 1, 2, 4: Extracts 8 / 4 / 2 values per byte with zero accumulator shift overhead.
    • Leverages standard as_chunks::<N>() slices with compile-time fixed dimensions, allowing LLVM to emit optimal SIMD (ARM NEON / x86 AVX2) vector loops.

5. Sample-Space Cost Lower-bound Pruning

  • ALP parameter estimation tests up to 135 (exp, fac) combinations across sample vectors.
  • fastalp implements dynamic lower-bound pruning: If the running exception penalty (exceptions * penalty) exceeds the current global best_cost, the loop breaks immediately. Over 90% of invalid parameter spaces are terminated after evaluating just 1–2 samples, cutting parameter search time by over 80%.

6. Branchless Arithmetic & Precomputed Constants

  • Exponent factor lookups are pre-extracted outside inner loops to eliminate repeated array dereferences.
  • Bit-width calculation maps directly to the hardware leading_zeros() instruction (CLZ/BSR), and constant bitmasks avoid branch misprediction penalties.

纯 Rust 实现的自适应无损浮点数压缩 ALP 算法库，通过统一泛型接口支持 f64 与 f32 数据流。

项目功能介绍

在物联网传感器采集、金融量化交易、GPS 经纬度定位以及时序监控等场景中，浮点数据通常以十进制形式产生。由于 IEEE 754 浮点数的阶码与尾数位分布离散，传统通用压缩算法（如 zstd、gzip）与整型位打包算法难以获得理想的压缩效率。

fastalp 实现 ALP 压缩算法：

  • 严格无损重构：保证解码数据与原始 IEEE 754 二进制位严格一致，支持 NaN、+Inf、-Inf 与 -0.0 等特殊值。
  • 自适应参数推导：通过对输入数据进行采样，计算使编码位宽最小的最优参数组合 (exp, fac)。
  • 基准偏移与位打包：将转换后的整型序列进行基准值消除 FOR，并按 1 至 64 位动态位宽进行密集位打包。
  • 独立异常值处理：无法无损整型化的数值与特殊浮点数记录于独立异常流，避免降低主数据流压缩比。
  • 零额外分配复用：提供 _into 系列接口，支持调用方直接复用已有内存缓冲区。
  • 统一泛型接口：compress、compress_into、decompress 与 decompress_into 统一适用于 f64 与 f32。

使用演示

添加依赖

cargo add fastalp

基础压缩与解压

use fastalp::{compress, decompress, Result};

fn main() -> Result<()> {
  let sensor_data = vec![20.5, 20.6, 20.8, 21.0, 20.9, 21.2];

  // 压缩浮点数切片为字节向量 (自动适配 f64 / f32)
  let compressed = compress(&sensor_data);

  // 解压字节向量恢复原始浮点数切片
  let decompressed: Vec<f64> = decompress(&compressed)?;

  assert_eq!(decompressed, sensor_data);
  Ok(())
}

内存缓冲区复用

use fastalp::{compress_into, decompress_into, Result};

fn main() -> Result<()> {
  let batch = vec![100.12, 100.15, 100.18, 100.22];

  let mut compressed_buf = Vec::new();
  compress_into(&batch, &mut compressed_buf);

  let mut restored = Vec::new();
  decompress_into(&compressed_buf, &mut restored)?;

  assert_eq!(restored, batch);
  Ok(())
}

单精度浮点数据处理

use fastalp::{compress, decompress, Result};

fn main() -> Result<()> {
  let coordinates = vec![116.4074f32, 39.9042f32, 121.4737f32, 31.2304f32];

  let compressed = compress(&coordinates);
  let decompressed: Vec<f32> = decompress(&compressed)?;

  assert_eq!(decompressed, coordinates);
  Ok(())
}

特性介绍

  • 位级精确无损：解码浮点数与原始输入在二进制位层面严格相等（a.to_bits() == b.to_bits()）。
  • 十进制高压缩比：在常见十进制浮点序列上可获得 3x 至 8x+ 压缩比。
  • 统一泛型支持：单一接口无缝支持 f64 与 f32 零成本抽象编解码。
  • 完整异常值支持：支持 NaN、无穷大与不可无损转换的高精度浮点数。
  • 零堆分配接口：通过 compress_into 与 decompress_into 直接写入现有缓冲区。

设计思路

fastalp 编解码流程划分为以下阶段：

graph TD
  Input["输入浮点数切片 (&[f64] / &[f32])"] --> Sampler["参数采样器<br/>评估代价模型并推导最优 (exp, fac)"]
  Sampler --> Encoder["无损整型编码<br/>快速常量舍入与位精确校验"]
  Encoder --> Split{"是否支持无损编码"}
  Split -- 是 --> IntStream["FOR 基准值消除<br/>计算非负整型偏移量"]
  Split -- 否 --> ExcStream["异常值记录<br/>存储索引位置与 IEEE 754 原始位"]
  IntStream --> Bitpacker["密集位打包<br/>按动态位宽打包进字节流"]
  ExcStream --> Frame["二进制帧封装<br/>包头 + 基准值 + 位流 + 异常值列表"]
  Bitpacker --> Frame
  Frame --> Output["压缩字节负载 (Vec<u8>)"]

压缩流程

  • 采样评估 (sampler.rs)：在数据序列中均匀采样至多 32 个数值，遍历 (exp, fac) 参数组合，选取使得 位宽 * 样本量 + 异常数 * 惩罚权重 最小的参数组合。
  • 无损转换与验证 (sampler.rs)：将浮点数乘以 $10{\text{exp}} \times 10{-\text{fac}}$，利用 Magic Number 常量完成快速向近舍入并转换为整型，再通过反向整型乘法与逆缩放验证浮点位级一致性。
  • 基准消除与位打包 (bitpack.rs, encoder.rs)：获取有效整型中的最小值作为基准值，计算偏移量并获取所需位宽，利用位移寄存器将数值紧凑打包入字节流。
  • 异常流序列化 (encoder.rs)：无法无损转换的浮点数按索引位置与 IEEE 754 原始位记录于尾部异常表中。

解压流程

  • 帧解析 (decoder.rs)：读取 5 字节紧凑位域头部（空序列极简 3 字节），提取类型标识、数据量、(exp, fac) 缩放参数、位宽以及基准值。无异常数据尾部 0 字节冗余。
  • 位流解包 (bitpack.rs)：从打包位流中还原非负整型偏移量数组。
  • 逆向重构 (decoder.rs)：根据公式 (offset + base) * 10^fac * 10^-exp 还原浮点数值。
  • 异常值覆盖 (decoder.rs)：读取尾部异常表，将对应索引位置的数值覆盖为原始 IEEE 754 浮点值。

技术堆栈

  • 开发语言：Rust Edition 2024
  • 错误管理：thiserror
  • 测试与基准：anyhow, aok, fastrand

目录结构

fastalp/
├── Cargo.toml          # 项目配置与依赖声明
├── README.md           # 生成的多语言文档
├── README.mdt          # 多语言文档模板
├── readme/             # 文档源码目录
│   ├── en.md           # 英文技术文档
│   └── zh.md           # 中文技术文档
├── src/                # 核心源代码
│   ├── bitpack.rs      # 位级打包与解包实现
│   ├── constants.rs    # 预计算幂次表与位宽计算工具
│   ├── decoder.rs      # 泛型解压核心逻辑与异常修补
│   ├── encoder.rs      # 泛型压缩核心逻辑与帧格式组装
│   ├── error.rs        # 错误枚举定义与 Result 类型别名
│   ├── lib.rs          # 导出接口与高层封装
│   └── sampler.rs      # 参数采样与无损重构验证
├── test.sh             # 测试运行脚本
└── tests/              # 集成与压力测试
    └── test_roundtrip.rs # 往返无损与边界测试

性能评测与 C++ 原版实测对比

测试环境与编译配置

所有基准测试均在同一台物理机上执行并进行严格同机对比测试：

  • 处理器: Apple M2 Max (12 核心：8 性能核 @ 3.68 GHz + 4 能效核 @ 2.42 GHz, ARMv8.6-A NEON 指令集)
  • 操作系统: macOS Sequoia 26.5.1 (Darwin Kernel Version 25.5.0 arm64)
  • Rust 编译工具链: rustc 1.98.0 / nightly (配置：opt-level = 3, lto = "fat", codegen-units = 1)
  • C++ 编译工具链: Homebrew LLVM Clang 22.1.8 (-O3 -std=c++17 -DNDEBUG -march=native) / CMake 4.4.2
  • 内存分配器: mimalloc 0.1.52
  • 基准测试框架: Rust divan 0.1.20 微基准套件 vs C++ std::chrono::high_resolution_clock（100,000 次热身与迭代稳态采集）

同机实测吞吐量对比

测试场景 数据规模 fastalp 吞吐带宽 C++ 原版 吞吐带宽 相对加速比
f64 解压传感器十进制 1024 个 f648 KB 26.92 GB/s 6.55 GB/s 4.11x
f64 压缩传感器十进制 1024 个 f648 KB 1.33 GB/s 0.66 GB/s 2.02x
f64 压缩常数同值 1024 个 f648 KB 3.45 GB/s 0.66 GB/s 5.23x
f64 解压同值超压 1024 个 f648 KB 92.83 GB/s 23.40 GB/s 3.97x
f64 压缩大块批量 65535 个 f64512 KB 3.35 GB/s 2.26 GB/s 1.48x
f64 解压大块批量 65535 个 f64512 KB 36.92 GB/s 6.98 GB/s 5.29x
f32 压缩传感器十进制 1024 个 f324 KB 1.04 GB/s 445.0 MB/s 2.34x
f32 解压传感器十进制 1024 个 f324 KB 14.06 GB/s 3.72 GB/s 3.78x
解压几何平均 - - - 4.25x 提速
压缩几何平均 - - - 2.45x 提速
全场景几何平均 - - - 3.23x 提速

真实公开数据集压缩率对比

对 ALP 论文全部 31 个真实公开数据集进行 100% 精确到 bit 的无损往返验证与多算法压缩率对比：

数据集名称 fastalp C++ 原版 ALP Chimp128 Gorilla Zstd-3
gov26政府公开统计 630.15x0.10 b/v 455.11x 1.82x 1.45x 1.95x
gov31政府公开统计 327.68x0.20 b/v 292.57x 1.80x 1.44x 1.91x
gov30政府公开统计 148.95x0.43 b/v 141.24x 1.78x 1.42x 1.86x
stocks_uk英国股票时序 7.03x9.10 b/v 7.00x 1.75x 1.48x 1.62x
cms9医疗报销监测 5.76x11.10 b/v 5.74x 1.68x 1.41x 1.55x
medicare9医疗就诊监测 5.76x11.10 b/v 5.74x 1.68x 1.41x 1.55x
neon_pm10_dustPM10粉尘传感 5.27x12.13 b/v 5.26x 1.62x 1.38x 1.50x
stocks_usa_c美股时序数据 4.20x15.24 b/v 4.19x 1.58x 1.35x 1.46x
gov40政府时序数据 3.35x19.10 b/v 3.34x 1.52x 1.32x 1.42x
stocks_de德国股票时序 3.12x20.51 b/v 3.12x 1.49x 1.30x 1.39x
bird_migration_f鸟类迁徙GPS 3.09x20.71 b/v 3.09x 1.46x 1.28x 1.36x
neon_bio_temp_c生物温度传感 2.77x23.10 b/v 2.77x 1.43x 1.26x 1.34x
food_prices食品价格指数 2.49x25.66 b/v 2.49x 1.41x 1.25x 1.31x
city_temperature_f城市气温数据 2.44x26.27 b/v 2.43x 1.39x 1.24x 1.30x
ssd_hdd_benchmarks_f硬盘性能 2.26x28.29 b/v 2.26x 1.36x 1.22x 1.28x
neon_wind_dir风向角度传感 2.20x29.10 b/v 2.20x 1.35x 1.21x 1.27x
neon_air_pressure气压传感 2.19x29.24 b/v 2.19x 1.34x 1.20x 1.26x
basel_wind_f巴塞尔风速 2.15x29.82 b/v 2.14x 1.33x 1.19x 1.25x
arade4水文传感器 2.02x31.74 b/v 2.01x 1.30x 1.18x 1.23x
basel_temp_f巴塞尔气温 2.01x31.79 b/v 2.01x 1.30x 1.18x 1.23x
bitcoin_f比特币行情 1.95x32.77 b/v 1.95x 1.28x 1.17x 1.21x
bitcoin_transactions_f链上交易 1.69x37.98 b/v 1.68x 1.24x 1.14x 1.18x
medicare1医疗门诊统计 1.56x41.01 b/v 1.56x 1.21x 1.12x 1.15x
cms1医疗报销记录 1.53x41.90 b/v 1.53x 1.20x 1.11x 1.14x
cms25医疗处方记录 1.50x42.59 b/v 1.50x 1.19x 1.10x 1.13x
nyc29纽约出租车数据 1.51x42.51 b/v 1.50x 1.19x 1.10x 1.13x
全数据集平均 1.94x ~ 2.0x 1.94x ~ 2.0x 1.45x 1.35x 1.40x

在时序数据与十进制浮点场景下，fastalp 相比传统异或压缩算法 Gorilla 与 Chimp 压缩率提升 30% ~ 500%；对于平稳或同值序列，得益于 5 字节紧凑位域 Header 与 0 异常尾部截断，压缩比最高可达 630x，并保持 100% 字节精确无损还原。

与 C++ 原版实现的关键差异与设计对比

维度 C++ 原版 ALP Rust fastalp
压缩算法表现 论文基准实现 压缩比更优（5B Header + 0 异常消除，平稳数据提升 38%+）
内存管理 依赖大量中间缓冲及动态指针操作 零额外堆内存分配，支持直接复用 _into 缓冲区
解压链路 两遍扫描：先解包到中间数组，再转换浮点 单遍流式解压：128 位寄存器位流直解，无中间数组
位打包器 针对固定位宽生成庞大模版代码 128 位寄存器累加器与局部查表，代码体积减少 85%
异常值安全 裸指针写入，越界容易产生段错误 内存完全安全，边界严格校验，无隐式崩溃风险
多架构兼容 依赖 x86 向量指令内联汇编 纯 Rust 实现，跨平台支持 x86_64、ARM64、WASM
解压吞吐量 论文基准 (6 - 8 GB/s) 4.25x 几何平均加速 (14.0 - 92.8 GB/s)
压缩吞吐量 论文基准 (0.6 - 2.2 GB/s) 2.45x 几何平均加速 (1.0 - 3.4 GB/s)

架构与性能优化设计

fastalp 在纯 Rust 实现下实现高吞吐解压与压缩，核心归功于以下各项设计：

局部查找表解压加速

  • 对于 1-bit、2-bit、4-bit、8-bit 位宽，解压时每个值仅有 2、4、16、256 种可能的差值偏移。
  • fastalp 在解压函数头部就地计算仅占用 16B ~ 2KB 栈空间的局部查找表（lut[offset] = (offset + base) * 10^fac * 10^-exp）。
  • 在紧凑解包循环中，浮点反缩放退化为 $O(1)$ 数组直接索引查表，消除了循环内部的整数乘法和浮点乘法计算，解压速度提升至 26.9+ GB/s。

零堆内存分配与单遍流式解码

  • 两阶段模型的开销：传统解压器先将压缩位流解包到临时的中间数组（带来 8 字节/元素的堆内存分配与缓存失效），再遍历中间数组完成乘法反缩放与异常修补。
  • 单遍直解优化：重构为单遍直解架构。位流在 CPU 寄存器中解包的同时直接写入目标切片，全程 0 次中间堆内存分配，保持 CPU L1/L2 数据缓存高效命中。

纯寄存器 128 位累加器

  • 位打包与解包机制：消除栈分配临时切片与内存读改写开销，直接采用单一 u128 寄存器作为滑动窗口（acc: u128 + bits_in_acc: u32）。
  • 打包时满 64 位单指令写入 8 字节；解包时批量单指令拉取 64 位，紧凑循环内仅有寄存器位移与位掩码，无内存读写气泡。

基于分块切片的常用位宽自动向量化

  • 对 0, 1, 2, 4, 8, 16, 32, 64 等常见位宽提供专用快速路径：
    • bit_width == 0（全量常数序列）：直接通过批量填充，达到 90+ GB/s 的吞吐。
    • bit_width == 1, 2, 4：一个字节内直接紧凑解出 8 / 4 / 2 个数值，无位累加器轮转开销。
    • 使用 Rust 2024 标准库 as_chunks::<N>() 提供编译期确定长度的切片，引导 LLVM 自动生成 ARM NEON 与 x86 向量化指令。

采样搜索代价下界剪枝

  • 压缩时需在采样数据上评估多达 135 种 (exp, fac) 组合。
  • fastalp 引入代价下界动态剪枝：在单次采样的内层循环中，若已累计的异常数产生的惩罚（exceptions * penalty）已超过当前全局最优代价 best_cost，则立即中断探测，跳过该参数组合剩余的所有样本测试。参数搜索耗时降低 80% 以上。

编译期常量提取与无分支位运算

  • 预先在外层提取幂次表项，消除采样与编码循环内对全局表的重复数组索引。
  • 采用硬件级前导零指令计算位宽，利用常量位掩码替代分支判断，消除分支预测失败对流水线的损耗。