Clast.Alp

ALP (Adaptive Lossless floating-Point) compression for .NET. Compresses arrays of double values into a compact byte representation, using a multi-stage pipeline that exploits the fact that real-world floating-point data (prices, sensor readings, scientific measurements) is often decimal in origin and uses limited precision. Compression is fully lossless.

Part of the clast-project.

How it works

The compression pipeline has four stages:

1. ALP encode — search for an (exponent, factor) pair such that value * 10^(exponent - factor) rounds to a small integer. The search picks the pair that produces the fewest exceptions — values that can't be losslessly recovered. Exceptions are stored verbatim and patched in on decode.
2. Frame-of-Reference (FOR) — subtract the minimum encoded integer so all deltas are non-negative.
3. Bit-pack — write each delta using exactly ceil(log2(max + 1)) bits.
4. Serialize — emit a 24-byte little-endian header followed by the packed deltas and any exceptions. The full layout is documented in src/Alp/AlpCompressor.cs.

Decompression reverses each step.

Based on the ALP algorithm originally published by researchers at CWI Amsterdam.

Usage

using Clast.Alp;

double[] prices = [99.99, 100.01, 42.50, 0.01, 1234.56];

byte[] compressed = AlpCompressor.Compress(prices);
double[] decoded   = AlpCompressor.Decompress(compressed);
// decoded.SequenceEqual(prices) == true

For pooled-buffer or pipeline scenarios, an IBufferWriter<byte> overload avoids the output byte[] allocation:

AlpCompressor.Compress(prices, writer);  // any IBufferWriter<byte>

For pre-allocated output buffers, use the Span<double> overload paired with GetDecompressedLength:

int n = AlpCompressor.GetDecompressedLength(compressed);
double[] buffer = ArrayPool<double>.Shared.Rent(n);
try
{
    AlpCompressor.Decompress(compressed, buffer);
    // ... consume buffer.AsSpan(0, n) ...
}
finally { ArrayPool<double>.Shared.Return(buffer); }

A few low-level utilities are also public for callers who want to integrate ALP encoding into their own data formats without going through the byte serialization:

• AlpEncoder.FindBestFactorExponent(samples) — returns the best (exponent, factor) pair for a sample, without performing the encoding
• AlpEncoder.EncodeValue(value, exponent, factor) — encode a single double to int64
• AlpDecoder.DecodeValue(encoded, exponent, factor) — decode a single int64 to double

The encode/FOR/bit-pack pipeline stages and the AlpEncodedData intermediate form are internal implementation details.

Target frameworks

The library targets net8.0 and netstandard2.0. The netstandard2.0 build supplies polyfills for BitOperations.LeadingZeroCount, double.IsFinite, and the double overloads of BinaryPrimitives, and relies on PolySharp for Index / Range / init / required attribute polyfills.

Project layout

Folder	Project	Targets
src/Alp/	Clast.Alp	net8.0, netstandard2.0
tests/Alp.Tests/	xunit test suite	net8.0, net472
benchmarks/Alp.Benchmarks/	BenchmarkDotNet	net8.0, net472

The test and benchmark projects target net472 in addition to net8.0 specifically to exercise the netstandard2.0 build of the library on the .NET Framework runtime.

Building, testing, benchmarking

dotnet build alp.sln
dotnet test tests/Alp.Tests
dotnet run -c Release --project benchmarks/Alp.Benchmarks

The benchmark project is configured with both a net8.0 and a net472 runtime job, so a single invocation will compare performance across both runtimes.

License

Apache License, Version 2.0. See LICENSE.