English | 中文

---

fastalp : Adaptive Lossless Floating-Point Compression in Rust

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

---

• Overview
• Usage
  • Installation
  • Basic Compression and Decompression
  • In-Place Buffer Reuse
  • Single-Precision Floating-Point Data
• Features
• Architecture & Design
  • Compression Pipeline
  • Decompression Pipeline
• Tech Stack
• Directory Structure
• Benchmarks & C++ Comparison
  • Benchmark Environment & Toolchain
  • Side-by-Side Throughput Comparison
  • Real-World Datasets Compression Ratio
  • Real-World Physical Telemetry Benchmark (NOAA & USGS All 64 Series)
• Architecture Comparison & Engineering Optimizations
  • Algorithmic Innovations on Compression Ratio (vs Reference C++)
    • 1. Decimal Division Exact Mode — Eliminating Multiplication Rounding Exceptions
    • 2. Adaptive Delta-ALP Differential Encoding — Breaking Global Dynamic Range Bounds
    • 3. Outlier Smoothing Isolation — Safeguarding Delta Bit-Width from Noise
    • 4. Raw Fallback Safeguard Against Negative Compression
  • Engineering Micro-Architecture & Throughput Optimizations (vs Reference C++)
    • Constant Sequence Fast Detection & Zero-Heap Allocation
    • Zero-Heap Direct Streaming Decompression
    • Pure-Register SIMD Vectorized Decompression & Hybrid Local Table Acceleration
    • Two-Pass SIMD Vectorized Encoding & Early-Exit Sampling
    • Pure 128-bit Register Bitpacker
    • Sample-Space Cost Lower-Bound Pruning
    • Branchless Arithmetic & Precomputed Constants

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 and integer bitpackers operate inefficiently on IEEE 754 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 Delta Differential Encoding (Delta-ALP): Automatically evaluates smooth and continuous physical time series (weather, hydrology, telemetry), adaptively applying first-order differences and branchless prefix sum accumulation to reduce bit widths by 15% to 38%.

• Decimal Division Exact Mode: Completely eliminates IEEE 754 multiplication roundoff errors (such as * 0.1) by reconstructing via exact decimal division, driving outlier exception counts to zero on real-world telemetry.

• Stack-Allocated LUT & SIMD Hybrid Decompression: Utilizes 256-entry stack lookup tables for division modes to eliminate hardware division latencies, coupled with pure-register SIMD auto-vectorization for linear arithmetic exceeding 55+ GB/s throughput.

• Adaptive Parameter Estimation: Samples input sequences to derive optimal scaling parameters (exp, fac, use_div) that minimize bit-width requirements.

• Frame-of-Reference & Bitpacking: Encodes converted integers using base subtraction (FOR / Delta) 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.

• Raw Fallback Protection: Automatically falls back to uncompressed raw mode when noise or extreme precision values would cause negative compression.

• 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 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: 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

• Constant Detection & Fallback Filter (encoder.rs): Quickly evaluates bit-exact identical sequences (v.is_exact_same(first)). When identical, writes a 5-byte header and base value with zero heap allocation. When estimated payload exceeds raw size plus header overhead, switches to 3-byte raw mode to guarantee zero data inflation.

• 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, float.rs): Multiplies float by $10^{{\text{exp}} \times 10}{-\text{fac}}$, rounds via constants, and verifies exact inverse equality against raw IEEE 754 bit representations.

• Base Offset & Bitpacking (bitpack/pack.rs, encoder.rs): Computes minimum integer value as base, subtracts base from valid integers, determines required bit width, and writes dense packed bits via a 128-bit register accumulator.

• Exception Stream (encoder.rs): Appends position and raw bits for values that fail exact integer roundtrip.

Decompression Pipeline

• Header Parsing (decoder.rs): Reads compact header, extracting format type and element count. For raw fallback chunks, performs direct zero-copy slice restoration. For ALP chunks, extracts packed (exp, fac, bit_width) parameters and base value.

• Bit Unpacking & SIMD Register Reconstruction (bitpack/unpack.rs): Bit-widths of 8, 16, 32, and 64 bits employ pure register SIMD auto-vectorization, eliminating gather lookups and cache stalls; Ultra-small bit-widths (1, 2, 4 bits) leverage compact register-resident tables for rapid reconstruction.

• 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/        # Modular bit-level packing and unpacking
│   │   ├── mod.rs      # Module facade and re-exports
│   │   ├── pack.rs     # Dense bitpacking with 128-bit register accumulator
│   │   └── unpack.rs   # Direct bit unpacking with stack LUT acceleration
│   ├── constants.rs    # Precomputed static power tables and format constants
│   ├── decoder/        # Generic decompression pipeline & decimal division reconstruction
│   │   ├── mod.rs      # Decompression facade and mode dispatch
│   │   ├── standard.rs # Standard FOR reconstruction
│   │   └── delta.rs    # Delta first-order difference reconstruction
│   ├── delta/          # First-order difference estimation & prefix sum
│   │   └── mod.rs
│   ├── encoder/        # Generic compression pipeline & raw fallback protection
│   │   ├── mod.rs      # Compression facade & auto-vectorized stream
│   │   ├── standard.rs # Standard FOR encoding pipeline
│   │   └── delta.rs    # Delta differential encoding pipeline
│   ├── error.rs        # Error definitions and Result type alias
│   ├── float/          # AlpFloat abstraction trait and f32/f64 zero-cost implementation
│   │   ├── mod.rs      # AlpFloat trait and lookup table generator
│   │   ├── f32.rs      # Single-precision f32 implementation
│   │   └── f64.rs      # Double-precision f64 implementation
│   ├── lib.rs          # Public crate exports and high-level API
│   ├── params.rs       # Compact bitfield parameter packing and bit-width utilities
│   └── sampler.rs      # Adaptive parameter optimization and lossless roundtrip verification
├── test.sh             # Test execution script
└── tests/              # Integration and stress tests
    ├── test_alp_dataset.rs # ALP paper 31 real-world datasets roundtrip & ratio tests
    ├── test_delta.rs       # Delta differential time series test suite
    └── test_roundtrip.rs   # Roundtrip integrity and boundary tests

---

Benchmarks & C++ Comparison

Benchmark Environment & Toolchain

All microbenchmarks were executed and measured side-by-side on the same physical host:

• 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 (steady-state median sampling)

Side-by-Side Throughput Comparison

Scenario	Data Size	fastalp Throughput	C++ Reference Throughput	Throughput Ratio (fastalp / C++)
f64 Compress (Identical Values)	1024 x f64 (8 KB)	23.15 GB/s	7.02 GB/s	3.30x
f64 Compress (Sensor Decimals)	1024 x f64 (8 KB)	6.10 GB/s	0.84 GB/s	7.26x
f64 Compress (Large Batch)	65535 x f64 (512 KB)	6.57 GB/s	5.85 GB/s	1.12x
f32 Compress (Sensor Decimals)	1024 x f32 (4 KB)	3.52 GB/s	2.46 GB/s	1.43x
f64 Decompress (Identical Values)	1024 x f64 (8 KB)	77.01 GB/s	21.85 GB/s	3.52x
f64 Decompress (Sensor Decimals)	1024 x f64 (8 KB)	57.32 GB/s	21.85 GB/s	2.62x
f64 Decompress (Large Batch)	65535 x f64 (512 KB)	55.93 GB/s	18.42 GB/s	3.04x
f32 Decompress (Sensor Decimals)	1024 x f32 (4 KB)	57.45 GB/s	32.77 GB/s	1.75x

Real-World Datasets Compression Ratio

Evaluated against all 31 standard real-world datasets from the original ALP paper (253,952 bytes of raw 64-bit doubles):

Dataset Name	Raw Size	fastalp Compressed Size	fastalp Ratio	C++ Ref ALP Ratio
gov26Government Stats	8192 B	13 B	630.15x(0.10 b/v)	455.11x
gov31Government Stats	8192 B	25 B	327.68x(0.20 b/v)	292.57x
gov30Government Stats	8192 B	55 B	148.95x(0.43 b/v)	141.24x
stocks_ukUK Stock Prices	8192 B	1165 B	7.03x(9.10 b/v)	7.00x
cms9Healthcare Billing	8192 B	1421 B	5.76x(11.10 b/v)	5.74x
medicare9Medical Monitoring	8192 B	1421 B	5.76x(11.10 b/v)	5.74x
neon_pm10_dustPM10 Sensor	8192 B	1553 B	5.27x(12.13 b/v)	5.26x
stocks_usa_cUS Stock Prices	8192 B	1951 B	4.20x(15.24 b/v)	4.19x
gov40Government Timestamps	8192 B	2445 B	3.35x(19.10 b/v)	3.34x
stocks_deGerman Stock Prices	8192 B	2625 B	3.12x(20.51 b/v)	3.12x
bird_migration_fGPS Coordinates	8192 B	2651 B	3.09x(20.71 b/v)	3.09x
neon_bio_temp_cBiology Sensor	8192 B	2957 B	2.77x(23.10 b/v)	2.77x
food_pricesConsumer Index	8192 B	3285 B	2.49x(25.66 b/v)	2.49x
city_temperature_fWeather Temp	8192 B	3363 B	2.44x(26.27 b/v)	2.43x
ssd_hdd_benchmarks_fDisk Benchmarks	8192 B	3621 B	2.26x(28.29 b/v)	2.26x
neon_wind_dirWind Direction	8192 B	3725 B	2.20x(29.10 b/v)	2.20x
neon_air_pressureAir Pressure	8192 B	3743 B	2.19x(29.24 b/v)	2.19x
basel_wind_fBasel Wind Speed	8192 B	3817 B	2.15x(29.82 b/v)	2.14x
arade4Hydrology Sensor	8192 B	4063 B	2.02x(31.74 b/v)	2.01x
basel_temp_fBasel Temperature	8192 B	4069 B	2.01x(31.79 b/v)	2.01x
bitcoin_fBitcoin Rates	8192 B	4195 B	1.95x(32.77 b/v)	1.95x
bitcoin_transactions_fOn-chain Tx	8192 B	4861 B	1.69x(37.98 b/v)	1.68x
medicare1Medical Records	8192 B	5249 B	1.56x(41.01 b/v)	1.56x
cms1Medical Records	8192 B	5363 B	1.53x(41.90 b/v)	1.53x
cms25Medical Records	8192 B	5451 B	1.50x(42.59 b/v)	1.50x
nyc29NYC Taxi Travel	8192 B	5441 B	1.51x(42.51 b/v)	1.50x
air_sensor_fAir Sensor Data	8192 B	8195 B (Fallback)	1.00x(Guaranteed)	0.52x (Expansion)
poi_latHigh-Precision Lat	8192 B	8195 B (Fallback)	1.00x(Guaranteed)	0.51x (Expansion)
poi_lonHigh-Precision Lon	8192 B	8195 B (Fallback)	1.00x(Guaranteed)	0.64x (Expansion)
TOTAL / Overall Average	253,952 B	110,773 B	2.29x	1.94x

Thanks to the raw fallback safeguard, fastalp completely eliminates negative compression on difficult datasets, reducing overall storage from 130,597 B to 110,773 B and elevating average compression ratio to 2.29x.

Real-World Physical Telemetry Benchmark (NOAA & USGS All 64 Series)

Evaluated side-by-side on 64 continuous industrial, marine, and meteorological observation series (NOAA ISD-Lite weather, NOAA CO-OPS tide gauge, USGS NWIS river discharge, comprising 467,550 double-precision points):

Variable	Series Count	Points	fastalp Ratio	C++ Ref Ratio	Space Saved	fastalp Enc	C++ Enc	fastalp Dec	C++ Dec	Dec Speedup
air_temperature	10	79,807	8.07x(7.93 b/v)	7.80x	-3.4%	2.55 GB/s	0.48 GB/s	12.99 GB/s	0.59 GB/s	22.0x
dew_point	10	79,772	8.30x(7.71 b/v)	7.95x	-4.2%	2.48 GB/s	0.49 GB/s	8.58 GB/s	0.60 GB/s	14.3x
sea_level_pressure	10	72,857	9.24x(6.93 b/v)	7.42x	-19.6%	2.32 GB/s	0.46 GB/s	6.40 GB/s	0.58 GB/s	11.0x
wind_direction	9	69,384	7.10x(9.01 b/v)	7.04x	-0.8%	2.10 GB/s	0.45 GB/s	6.48 GB/s	0.61 GB/s	10.6x
wind_speed	9	71,298	7.07x(9.05 b/v)	7.83x	-	2.21 GB/s	0.47 GB/s	23.57 GB/s	0.62 GB/s	38.0x
water_level	4	29,760	8.51x(7.52 b/v)	5.32x	-37.4%	2.41 GB/s	0.42 GB/s	10.74 GB/s	0.54 GB/s	19.9x
water_level_sigma	4	29,760	6.27x(10.20 b/v)	9.36x	-	2.15 GB/s	0.52 GB/s	12.10 GB/s	0.64 GB/s	18.9x
discharge	4	17,452	5.36x(11.93 b/v)	4.34x	-19.2%	1.85 GB/s	0.38 GB/s	4.91 GB/s	0.49 GB/s	10.0x
gage_height	4	17,460	9.72x(6.58 b/v)	7.75x	-20.3%	2.20 GB/s	0.44 GB/s	6.66 GB/s	0.55 GB/s	12.1x
【64 Series Total】	64	467,550	7.72x(8.29 b/v)	7.30x	-5.54%	2.35 GB/s	0.47 GB/s	11.20 GB/s	0.58 GB/s	19.3x

• Compression Ratio Breakthrough: By coupling Decimal Division Exact Mode with adaptive Delta differencing, fastalp compresses real physical telemetry to 8.29 b/v on average, outperforming the C++ reference by 5.54% overall and by 20% to 37% on tidal and gage-height signals.
• Overwhelming Throughput Advantage: Decompression throughput reaches 11.20 GB/s on a single core, surpassing the C++ reference (0.58 GB/s) by 19.3x. Compression throughput reaches 2.35 GB/s (5.0x faster than C++).

---

Architecture Comparison & Engineering Optimizations

Compared with the reference C++ implementation, fastalp not only multiplies throughput performance, but also introduces major algorithmic innovations that break through the compression ratio limitations of the reference implementation.

Algorithmic Innovations on Compression Ratio (vs Reference C++)

1. Decimal Division Exact Mode — Eliminating Multiplication Rounding Exceptions

• Reference C++ Limitation: C++ ALP exclusively uses multiplication for inverse reconstruction: v = (encoded * frac_exp) / fac. Because binary IEEE 754 cannot represent decimal fractions like 0.1 exactly, multiplication introduces unavoidable roundoff discrepancies (e.g. 123 * 0.1 evaluates to 12.30000000000000071... != 12.3). This causes the reference implementation to misclassify clean decimal measurements as "Exceptions", requiring an 8-byte original float plus a 2-byte index (an 80-bit penalty per outlier in standard 1024-element blocks). In real-world weather and oceanographic telemetry, this inflated exception count severely degrades compression ratios.
• fastalp Algorithmic Innovation: fastalp introduces Decimal Division Exact Mode (TYPE_F64_DEC / TYPE_F32_DEC). During parameter evaluation and decoding, dividing by powers of ten (e.g. / 10.0) reconstructs the exact original IEEE 754 bit pattern without truncation error.
  • Compression Gain: On real-world time-series (such as NOAA tidal heights and surface temperatures), the exception count drops from hundreds in C++ ALP to zero. Eliminating the exception dictionary shrinks compressed size by an additional 20% to 38% (e.g. ocean tidal series compression improves from 5.32x to 8.51x).
  • Zero-Latency Decoding: To prevent hardware division latency from impacting decompression throughput, fastalp couples this with a 256-entry stack-allocated lookup table (LUT), preserving maximum compression ratio while sustaining 55+ GB/s throughput.

2. Adaptive Delta-ALP Differential Encoding — Breaking Global Dynamic Range Bounds

• Reference C++ Limitation: C++ ALP relies solely on Frame-of-Reference (FOR) base subtraction: stored = encoded - min_encoded. For continuous physical time-series (such as diurnal temperature cycles, tidal oscillations, river stage ramps, and monotonic metric counters), the dynamic range (max - min) across a 1024-element block is wide (often spanning thousands of integer units, requiring 12 to 16 bits per element). FOR cannot exploit the strong local correlation between adjacent points.
• fastalp Algorithmic Innovation: fastalp introduces Adaptive Delta-ALP (TYPE_F64_DELTA / TYPE_F32_DEC_DELTA):
  1. Adaptive Benefit Evaluation: Dynamically evaluates the bit-width required by FOR versus first-order differences (delta[i] = encoded[i] - encoded[i-1]), activating Delta-ALP only when differences yield measurable savings;
  2. Tightly Packed Adjacent Differences: Stores the initial element as a baseline, bitpacking (delta - min_delta). Physical sensor delta bit-widths routinely collapse to 1 to 6 bits;
  3. Compression Gain: Yields a 50% to 90% size reduction on smooth and monotonic time-series. In monotonic ramp benchmarks, bits per value fall from 68.09 b/v in the reference implementation to 0.16 b/v, elevating compression ratio to 390x (compared to 0.94x in C++ ALP).

3. Outlier Smoothing Isolation — Safeguarding Delta Bit-Width from Noise

• Reference C++ Limitation: In traditional differential encoding pipelines, a single sensor spike or outlier pollutes two adjacent deltas (the jump upward and the drop downward), causing the maximum delta range across the entire block to explode and degrading bit-width for all points.
• fastalp Algorithmic Innovation: fastalp designs an Outlier Smoothing Isolation mechanism for Delta encoding: when a floating-point exception occurs, the delta stream records an increment of 0 (carrying forward the previous valid integer), while the true outlier float is isolated in the patch dictionary. On decompression, a branchless register prefix-sum pass reconstructs the base integers before patching outliers in-place. This preserves minimal delta bit-widths without compromising 100% bit-exact lossless fidelity.

4. Raw Fallback Safeguard Against Negative Compression

• Reference C++ Limitation: On high-entropy unstructured floats or high-precision geographic coordinates (such as POI latitude/longitude), exception tables expand beyond uncompressed payload size, leading to negative compression (down to 0.51x, doubling storage size).
• fastalp Algorithmic Innovation: At the end of the encoding pipeline, fastalp checks the compressed payload size. If it exceeds the raw data plus a 3-byte metadata header, the encoder instantly falls back to TYPE_RAW mode, storing raw bytes directly. This eliminates negative compression entirely, lifting overall dataset compression from 1.94x to 2.29x across the 31 paper datasets.

---

Engineering Micro-Architecture & Throughput Optimizations (vs Reference C++)

Constant Sequence Fast Detection & Zero-Heap Allocation

• Reference C++ Implementation: Executes full parameter sampling, intermediate integer transformation, and bit-width analysis even on completely constant sequences, requiring 9.25 µs end-to-end.
• fastalp Optimization: Inspects raw IEEE 754 bits at compression entry (v.is_exact_same(first)), strictly differentiating +0.0 and -0.0 sign bits; Directly emits a 5-byte header and base value (bit_width = 0) upon match, skipping parameter search and vector allocation, reducing compression time to 351 ns (26x speedup).

Zero-Heap Direct Streaming Decompression

• Reference C++ Implementation: Employs a two-stage decompression pipeline: stage 1 unpacks bitstream to an intermediate heap array, and stage 2 iterates over the array to compute float unscaling and patch exceptions, incurring 8 B/elem heap allocation and cache pressure.
• fastalp Optimization: Executes a single-pass direct streaming reconstruction pipeline. Bits are unpacked within CPU registers and written directly to the caller destination slice, keeping L1/L2 caches hot and providing compress_into and decompress_into zero-allocation APIs.

Pure-Register SIMD Vectorized Decompression & Hybrid Local Table Acceleration

• Reference C++ Implementation: Inner loop relies on two-stage heap buffering and scalar arithmetic, failing to saturate modern SIMD execution pipelines.
• fastalp Optimization: Eliminates large stack tables that induce indirect gather memory stalls; bit-widths of 8, 16, 32, and 64 bits execute pure linear register arithmetic with a dedicated fac1 path (omitting integer multiplication), enabling LLVM to emit optimal SIMD vector instructions; 1, 2, and 4 bit-widths utilize tiny register-resident tables, driving single-core decode throughput up to 57+ GB/s.

Two-Pass SIMD Vectorized Encoding & Early-Exit Sampling

• Reference C++ Implementation: Complex multi-level sampling logic with dense conditional branches inside the encoding loop, fragmenting basic blocks.
• fastalp Optimization: Introduces an EARLY_EXIT_BIT_WIDTH threshold during sampling to halt immediately once a high-compression model is identified, bypassing wasteful checks across 135 parameter combinations; adopts a Two-Pass decoupled encoding architecture (Pass 1 branchless register-level float-to-int rounding, Pass 2 centralized exception verification), eliminating per-element pipeline stalls and driving batch compression throughput up to 5.4+ GB/s.

Pure 128-bit Register Bitpacker

• Reference C++ Implementation: Generates extensive template code across multiple compilation units, creating large binaries with architecture-specific intrinsics.
• fastalp Optimization: Maintains a sliding bit window with a single 128-bit register accumulator (acc: u128, bits_in_acc: u32), executing 64-bit word writes and reads in single instructions; Pure safe Rust with zero external C++ toolchain dependencies, cross-compiling seamlessly for x86_64, ARM64, and WebAssembly.

Sample-Space Cost Lower-Bound Pruning

• Reference C++ Implementation: Evaluates all samples across 135 (exp, fac) parameter combinations unconditionally.
• fastalp Optimization: Applies dynamic lower-bound pruning: breaks inner verification immediately once running exception penalty (exceptions * penalty) surpasses current global best_cost, skipping unnecessary parameter iterations.

Branchless Arithmetic & Precomputed Constants

• Pre-extracts exponent factor tables outside inner loops to eliminate repeated array lookups;
• Calculates bit-width using hardware CLZ instructions and applies compile-time bitmasks to eliminate conditional branch mispredictions.

---

fastalp : 基于 ALP 算法的无损浮点数压缩引擎

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。

---

• 功能特性
• 使用示例
  • 添加依赖
  • 基础压缩与解压
  • 内存缓冲区复用
  • 单精度浮点数据处理
• 核心特性
• 架构设计
  • 压缩流程
  • 解压流程
• 技术栈
• 目录结构
• 性能评测与 C++ 原版对比
  • 测试环境与编译配置
  • 同机实测吞吐量对比
  • 真实公开数据集压缩率对比
  • 真实物理观测时序数据集对比 (NOAA & USGS 全量 64 时序)
• 架构对比与工程优化设计
  • 核心压缩率算法优化对比 (vs C++ 原版)
    • 1. 十进制精确除法重构模式 (Decimal Division Mode) —— 根治乘法舍入伪异常
    • 2. 自适应一阶差分时序编码 (Delta-ALP) —— 突破全局极值跨度瓶颈
    • 3. 异常点前值平滑隔离补丁 —— 杜绝尖峰噪声污染差分流
    • 4. 原始保底机制消除负压缩 (Raw Fallback Safeguard)
  • 工程微架构与吞吐性能优化 (vs C++ 原版)
    • 全等序列常数探测与零堆分配
    • 零堆内存分配与单遍流式解码
    • 纯寄存器 SIMD 向量化解压与局部查表混合加速
    • Two-Pass 向量化编码转换与采样早期退出
    • 纯寄存器 128 位累加器与紧凑位打包
    • 采样搜索代价下界剪枝
    • 编译期常量提取与无分支位运算

功能特性

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，通用压缩算法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。

• 时序差分自适应编码 (Delta-ALP)： 自动评估连续平滑的时序物理波形（气象、水文、传感器），自适应采用一阶相邻差分与前缀和递推，位宽进一步收窄 15% ~ 38%。

• 十进制精确除法重构 (Decimal Division Mode)： 彻底消除 IEEE 754 浮点乘法（如 * 0.1）引起的无限循环二进制尾数截断误差，以十进制除法精确重构，将观测时序异常点直接归零。

• 栈上 LUT 查表与 SIMD 混合加速： 小位宽利用 256 项栈上查找表（L1D 缓存命中）彻底消除循环内硬件除法延迟；对直接模式采用纯寄存器 SIMD 向量化计算，吞吐高达 55+ GB/s。

• 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, use_div)。

• 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 / Delta），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。

• 原始保底模式： 当随机噪声或不可压缩数据导致编码后体积膨胀时，自动回退至原始保底模式，杜绝负压缩。

• 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。

---

使用示例

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(())
}

---

核心特性

• 位级精确无损： 解码浮点数与原始输入在二进制位层面保持一致（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。

---

架构设计

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

• 全等探测与保底分流 (encoder.rs)： 先对数据进行常数序列快速校验；若全等且可编码，直接写入 5 字节头与基准值； 若为不可压缩随机数据且编码体积超过原始大小，则直接写入 3 字节头并以原始字节流存储。

• 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, float.rs)： 将浮点数乘以 $10^{{\text{exp}} \times 10}{-\text{fac}}$，利用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/pack.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 128 位寄存器滑动窗口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)： 读取紧凑头部，提取类型标识与元素数量； 若类型为原始保底数据，通过内存复制直出恢复；若为 ALP 压缩数据，提取 (exp, fac) 缩放参数、位宽以及基准值。

• 位流解包与 SIMD 寄存器流水重构 (bitpack/unpack.rs)： 针对 8/16/32/64 bit 采用纯寄存器 SIMD 自动向量化计算，彻底消除堆栈查表与内存间接 gather 寻址延迟；针对 1/2/4 bit 采用微型局部表快速还原。

• 异常值覆盖 (decoder.rs)： 若存在尾部异常表，读取对应索引位置的数值并覆盖为原始 IEEE 754 浮点值。

---

技术栈

• 开发语言：Rust Edition 2024
• 错误处理：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/        # 模块化位打包与位解包
│   │   ├── mod.rs      # 门面导出
│   │   ├── pack.rs     # 128 位累加器位打包算子
│   │   └── unpack.rs   # 局部查表与直接位解包算子
│   ├── constants.rs    # 静态幂次表与格式常量
│   ├── decoder/        # 泛型流式解压与除法重构
│   │   ├── mod.rs      # 解压门面与模式派发
│   │   ├── standard.rs # 标准 FOR 还原解压
│   │   └── delta.rs    # Delta 一阶差分解码
│   ├── delta/          # 一阶差分自适应收益评估与前缀和
│   │   └── mod.rs
│   ├── encoder/        # 泛型压缩流水线与保底回退
│   │   ├── mod.rs      # 编码门面与向量化流
│   │   ├── standard.rs # 标准 FOR 编码流水线
│   │   └── delta.rs    # Delta 一阶差分编码流水线
│   ├── error.rs        # 错误枚举定义与 Result 类型别名
│   ├── float/          # AlpFloat 浮点抽象特征与泛型无损转换
│   │   ├── mod.rs      # AlpFloat trait 定义与查表构建
│   │   ├── f32.rs      # 单精度 f32 乘法/除法编解码实现
│   │   └── f64.rs      # 双精度 f64 乘法/除法编解码实现
│   ├── lib.rs          # 导出接口与高层封装
│   ├── params.rs       # 紧凑位域参数打包与位宽计算
│   └── 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_alp_dataset.rs # ALP 论文 31 真实数据集往返与压缩比评测
    ├── test_delta.rs       # Delta 差分时序专项与异常测试
    └── test_roundtrip.rs   # 往返无损与边界测试

---

性能评测与 C++ 原版对比

测试环境与编译配置

所有基准测试均在同一物理机上执行并进行同机对比测试：

• 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（稳态中位数采样）

同机实测吞吐量对比

测试场景	数据规模	fastalp 吞吐	C++ 原版 吞吐	吞吐比 (fastalp / C++)
f64 压缩 (常数同值序列)	1024 个 f64 (8 KB)	23.15 GB/s	7.02 GB/s	3.30x
f64 压缩 (传感器十进制)	1024 个 f64 (8 KB)	6.10 GB/s	0.84 GB/s	7.26x
f64 压缩 (大块批量)	65535 个 f64 (512 KB)	6.57 GB/s	5.85 GB/s	1.12x
f32 压缩 (传感器十进制)	1024 个 f32 (4 KB)	3.52 GB/s	2.46 GB/s	1.43x
f64 解压 (同值序列)	1024 个 f64 (8 KB)	77.01 GB/s	21.85 GB/s	3.52x
f64 解压 (传感器十进制)	1024 个 f64 (8 KB)	57.32 GB/s	21.85 GB/s	2.62x
f64 解压 (大块批量)	65535 个 f64 (512 KB)	55.93 GB/s	18.42 GB/s	3.04x
f32 解压 (传感器十进制)	1024 个 f32 (4 KB)	57.45 GB/s	32.77 GB/s	1.75x

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（共 253,952 字节原始浮点数据）进行精确到 bit 的无损往返验证与压缩率评测：

数据集名称	原始大小	fastalp 压缩大小	fastalp 压缩率	C++ 原版 压缩率
gov26政府公开统计	8192 B	13 B	630.15x(0.10 b/v)	455.11x
gov31政府公开统计	8192 B	25 B	327.68x(0.20 b/v)	292.57x
gov30政府公开统计	8192 B	55 B	148.95x(0.43 b/v)	141.24x
stocks_uk英国股票时序	8192 B	1165 B	7.03x(9.10 b/v)	7.00x
cms9医疗报销监测	8192 B	1421 B	5.76x(11.10 b/v)	5.74x
medicare9医疗就诊监测	8192 B	1421 B	5.76x(11.10 b/v)	5.74x
neon_pm10_dustPM10粉尘传感	8192 B	1553 B	5.27x(12.13 b/v)	5.26x
stocks_usa_c美股时序数据	8192 B	1951 B	4.20x(15.24 b/v)	4.19x
gov40政府时序数据	8192 B	2445 B	3.35x(19.10 b/v)	3.34x
stocks_de德国股票时序	8192 B	2625 B	3.12x(20.51 b/v)	3.12x
bird_migration_f鸟类迁徙GPS	8192 B	2651 B	3.09x(20.71 b/v)	3.09x
neon_bio_temp_c生物温度传感	8192 B	2957 B	2.77x(23.10 b/v)	2.77x
food_prices食品价格指数	8192 B	3285 B	2.49x(25.66 b/v)	2.49x
city_temperature_f城市气温数据	8192 B	3363 B	2.44x(26.27 b/v)	2.43x
ssd_hdd_benchmarks_f硬盘性能	8192 B	3621 B	2.26x(28.29 b/v)	2.26x
neon_wind_dir风向角度传感	8192 B	3725 B	2.20x(29.10 b/v)	2.20x
neon_air_pressure气压传感	8192 B	3743 B	2.19x(29.24 b/v)	2.19x
basel_wind_f巴塞尔风速	8192 B	3817 B	2.15x(29.82 b/v)	2.14x
arade4水文传感器	8192 B	4063 B	2.02x(31.74 b/v)	2.01x
basel_temp_f巴塞尔气温	8192 B	4069 B	2.01x(31.79 b/v)	2.01x
bitcoin_f比特币行情	8192 B	4195 B	1.95x(32.77 b/v)	1.95x
bitcoin_transactions_f链上交易	8192 B	4861 B	1.69x(37.98 b/v)	1.68x
medicare1医疗门诊统计	8192 B	5249 B	1.56x(41.01 b/v)	1.56x
cms1医疗报销记录	8192 B	5363 B	1.53x(41.90 b/v)	1.53x
cms25医疗处方记录	8192 B	5451 B	1.50x(42.59 b/v)	1.50x
nyc29纽约出租车数据	8192 B	5441 B	1.51x(42.51 b/v)	1.50x
air_sensor_f高频空气传感	8192 B	8195 B (保底)	1.00x(回退)	0.52x (膨胀)
poi_latPOI高精度纬度	8192 B	8195 B (保底)	1.00x(回退)	0.51x (膨胀)
poi_lonPOI高精度经度	8192 B	8195 B (保底)	1.00x(回退)	0.64x (膨胀)
总计 / 全数据集平均	253,952 B	110,773 B	2.29x	1.94x

得益于原始保底机制，fastalp 彻底消除了高精双精度浮点数在 ALP 模型下的负压缩现象，总压缩体积由 130,597 字节降至 110,773 字节，平均压缩率提升至 2.29x。

真实物理观测时序数据集对比 (NOAA & USGS 全量 64 时序)

针对现实生产中最核心的物理传感与环境监测场景（包括 NOAA ISD-Lite 气象、NOAA CO-OPS 海洋潮位、USGS NWIS 河流流量水文监测，共 64 条真实长时序，467,550 个 64 位浮点数），进行端到端同机实测对比：

观测变量 (Variable)	序列数量	数据点数	fastalp 压缩率	C++ 原版 压缩率	体积缩减率	fastalp 压缩吞吐	C++ 压缩吞吐	fastalp 解压吞吐	C++ 解压吞吐	解压加速比
air_temperature (地面气温)	10	79,807	8.07x(7.93 b/v)	7.80x	-3.4%	2.55 GB/s	0.48 GB/s	12.99 GB/s	0.59 GB/s	22.0x
dew_point (露点温度)	10	79,772	8.30x(7.71 b/v)	7.95x	-4.2%	2.48 GB/s	0.49 GB/s	8.58 GB/s	0.60 GB/s	14.3x
sea_level_pressure (海平面气压)	10	72,857	9.24x(6.93 b/v)	7.42x	-19.6%	2.32 GB/s	0.46 GB/s	6.40 GB/s	0.58 GB/s	11.0x
wind_direction (气象风向)	9	69,384	7.10x(9.01 b/v)	7.04x	-0.8%	2.10 GB/s	0.45 GB/s	6.48 GB/s	0.61 GB/s	10.6x
wind_speed (观测风速)	9	71,298	7.07x(9.05 b/v)	7.83x	-	2.21 GB/s	0.47 GB/s	23.57 GB/s	0.62 GB/s	38.0x
water_level (海洋潮位)	4	29,760	8.51x(7.52 b/v)	5.32x	-37.4%	2.41 GB/s	0.42 GB/s	10.74 GB/s	0.54 GB/s	19.9x
water_level_sigma (潮位标准差)	4	29,760	6.27x(10.20 b/v)	9.36x	-	2.15 GB/s	0.52 GB/s	12.10 GB/s	0.64 GB/s	18.9x
discharge (河流水量流量)	4	17,452	5.36x(11.93 b/v)	4.34x	-19.2%	1.85 GB/s	0.38 GB/s	4.91 GB/s	0.49 GB/s	10.0x
gage_height (水文水尺高度)	4	17,460	9.72x(6.58 b/v)	7.75x	-20.3%	2.20 GB/s	0.44 GB/s	6.66 GB/s	0.55 GB/s	12.1x
【物理时序 64 序列 总计】	64	467,550	7.72x(8.29 b/v)	7.30x	-5.54%	2.35 GB/s	0.47 GB/s	11.20 GB/s	0.58 GB/s	19.3x

• 压缩体积突破：得益于十进制精确除法重构与 Delta 自适应差分，在真实物理观测时序中，fastalp 将每点平均占用压缩至 8.29 bits，较 C++ 原版体积进一步缩减 5.54%，在潮位与水尺高度上体积缩减达 20%~37%。
• 吞吐量压倒性领先：单核解压吞吐达 11.20 GB/s，达到 C++ 原版（0.58 GB/s）的 19.3 倍；压缩吞吐达 2.35 GB/s，达到 C++ 原版的 5.0 倍。

---

架构对比与工程优化设计

相比 C++ 原版实现，fastalp 不仅在吞吐性能上实现成倍飞跃，更在底层算法上做出了重大革新，彻底突破了 C++ 原版在特定场景下的压缩率瓶颈。

核心压缩率算法优化对比 (vs C++ 原版)

1. 十进制精确除法重构模式 (Decimal Division Mode) —— 根治乘法舍入伪异常

• C++ 原版局限： C++ ALP 仅采用基于乘法的逆变换 v = (encoded * frac_exp) / fac。由于二进制无法完全精确表示 0.1，浮点乘法会引入不可逆的微小截断误差（例如 123 * 0.1 的实际计算结果为 12.30000000000000071...，并不等于十进制输入的 12.3）。这导致原版 ALP 将大量原本为严格十进制的真实物理数据误判为“无法无损还原的异常点（Exceptions）”，每个异常点在标准 1024 块中必须单独记录 8 字节原始浮点数 + 2 字节位置索引（单个异常代价高达 80 比特！）。在真实气象与潮位观测时序中，异常点比例居高不下，极大地蚕食了压缩率。
• fastalp 算法革新： fastalp 引入十进制精确除法重构模式 (TYPE_F64_DEC / TYPE_F32_DEC)。在采样与重构中，通过直接除以精确的十进制整数幂（如 / 10.0），使截断误差彻底归零。
  • 压缩收益：在 NOAA 海洋潮位、地表气温等真实时序中，异常点数量直接由 C++ 原版的数百个骤降至 0 个！省去庞大的异常字典开销，压缩体积直接额外缩减 20% ~ 38%（如潮位数据压缩比从 C++ 的 5.32x 提升至 8.51x）。
  • 性能解耦：为防止硬件浮点除法的高延迟拖累解压速度，fastalp 创新性地结合了 256 项栈上微型查表（LUT），在享受 0 异常最高压缩率的同时，维持 55+ GB/s 的极限吞吐。

2. 自适应一阶差分时序编码 (Delta-ALP) —— 突破全局极值跨度瓶颈

• C++ 原版局限： C++ ALP 仅采用全局基准偏移（Frame-of-Reference, FOR）：stored = encoded - min_encoded。对于真实世界具有物理连续性的时序数据（如气温昼夜周期、水尺高度起伏、河流流量平缓变化、单调递增计数器），1024 个点内的数据极值跨度 (max - min) 往往很大（如温度跨越数十度，对应整型跨度达数千），这强制该数据块内所有 1024 个点必须采用 12~16 位的打包位宽，完全无法利用时序数据相邻点极度接近的强局部相关性。
• fastalp 算法革新： fastalp 首创自适应 Delta-ALP (TYPE_F64_DELTA / TYPE_F32_DEC_DELTA) 算法：
  1. 自适应收益决策：在采样阶段动态比对 FOR 全局位宽与一阶差分位宽（delta[i] = encoded[i] - encoded[i-1]），仅在差分确有收益时才启用，兼顾乱序与有序时序；
  2. 相邻差分紧凑打包：以首元素为基准，对相邻差分序列 (delta - min_delta) 进行位打包。物理传感序列的差分位宽通常直接骤降至 1~6 位；
  3. 压缩收益：在单调与平滑时序上，压缩体积较 C++ 原版缩减 50% ~ 90%。在单调递增时序测试中，每点占用由原版的 68.09 bits 暴降至 0.16 bits，压缩比跃升至 390x（C++ 原版仅为 0.94x）。

3. 异常点前值平滑隔离补丁 —— 杜绝尖峰噪声污染差分流

• C++ 原版局限： 在传统的差分压缩流水线中，若遭遇单点仪器噪声或异常尖峰，该异常点会同时污染相邻的两个差分值（突变上升与突变下跌），导致整块数据的最大差分跨度剧烈膨胀，使得整个数据块的打包位宽急剧恶化。
• fastalp 算法革新： fastalp 在 Delta 编码时设计了前值平滑隔离补丁机制：当遇到无法缩放的浮点异常点时，差分流中记录增量为 0（平滑沿用前一个有效整数），将真实的异常浮点数独立提取进 Patch 字典。解码时，先以单遍无分支前缀和流水线还原基准整型，再就地修补异常点。既保护了差分位流不受尖峰干扰、维持最低位宽，又确保了 100% 比特精确无损。

4. 原始保底机制消除负压缩 (Raw Fallback Safeguard)

• C++ 原版局限： 遇到高熵非结构化浮点或高精度随机坐标（如 POI 经纬度、高频电磁传感）时，原版 C++ ALP 的异常表膨胀导致体积比原始数据还要大（负压缩，压缩比低至 0.51x，产生 2 倍空间膨胀）。
• fastalp 算法革新： 在编码流水线末端精确计算压缩产物大小；一旦发现压缩大小超过原始数据加上 3 字节元数据头部，自动瞬间回退至 TYPE_RAW 模式，将数据按原样存储。在标准 31 数据集评测中，彻底杜绝了 0.51x 的膨胀，全量数据集平均压缩比从 C++ 的 1.94x 提高到了 2.29x。

---

工程微架构与吞吐性能优化 (vs C++ 原版)

全等序列常数探测与零堆分配

• C++ 原版实现： 面对全量常数序列时，依然需要执行完整的样本采集、临时整型数组转换与位宽分析，端到端耗时达 9.25 微秒。
• fastalp 优化： 在压缩入口通过底层原始比特比对（v.is_exact_same(first)，严格区分 +0.0 与 -0.0 符号位）； 命中后直接写入 5 字节紧凑头部与基准值（bit_width = 0），跳过所有采样与中间数组分配，压缩耗时降至 351 纳秒，相对提速 26 倍。

零堆内存分配与单遍流式解码

• C++ 原版实现： 采用两阶段解码架构：阶段一解包位流到中间堆数组，阶段二遍历中间数组计算浮点逆缩放并修补异常，引发 8 字节/元素的堆分配与 L1/L2 缓存挤占。
• fastalp 优化： 采用单遍直解流式架构；位流在 CPU 寄存器中解包的同时直接计算并写入目标切片，消除中间堆分配与内存往返传输，保持 CPU 缓存高效命中；对外提供 compress_into 与 decompress_into 零分配接口。

纯寄存器 SIMD 向量化解压与局部查表混合加速

• C++ 原版实现： 解包内层循环依赖两阶段堆缓冲传递与标量乘除运算，在非连续加载下难以充分饱和向量单元。
• fastalp 优化： 摒弃会引发间接 gather 寻址与缓存停顿的大尺寸表；针对 8、16、32、64 位宽直接采用纯寄存器线性算术指令流，配合 fac1 路径消除整数乘法，使 LLVM 自动生成 SIMD 矢量流水；针对 1、2、4 超小位宽采用微型寄存器局部表快速解包，单核解压吞吐跃升至 57+ GB/s。

Two-Pass 向量化编码转换与采样早期退出

• C++ 原版实现： 多层采样逻辑复杂度高，编码循环混合了密集条件分支，导致基本块碎片化。
• fastalp 优化： 在压缩采样中引入 EARLY_EXIT_BIT_WIDTH 优质参数即停机制，避免对 135 种组合的盲目遍历；在数据编码阶段采用 Two-Pass 分离架构（Pass 1 纯寄存器无分支舍入转换整型，Pass 2 集中校验异常），彻底消除单元素内的多重分支停顿，批量压缩吞吐飙升至 5.4+ GB/s。

纯寄存器 128 位累加器与紧凑位打包

• C++ 原版实现： 采用多层宏与模板元编程生成大量打包函数，编译生成的目标代码体积庞大，且高度耦合特定硬件平台的指令扩展。
• fastalp 优化： 采用单一 u128 寄存器作为滑动窗口（acc: u128 与 bits_in_acc: u32），单指令 64 位写入或读取； 纯安全 Rust 实现，不依赖外部 C++ 编译链，天然跨平台支持 x86_64、ARM64 以及 WebAssembly。

采样搜索代价下界剪枝

• C++ 原版实现： 参数搜索时遍历 135 种 (exp, fac) 组合的全部样本，遍历开销较高。
• fastalp 优化： 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，跳过剩余的所有样本测试，显著降低参数搜索耗时。

编译期常量提取与无分支位运算

• Exponent factor 预先在外层提取，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（CLZ）计算位宽，利用常量位掩码替代分支判断，减少流水线损耗。